brintos

brintos / llvm-project-archived public Read only

0
0
Text · 403.7 KiB · e12e31b Raw
9144 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck %s --check-prefixes=GCN3; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck %s --check-prefixes=TAHITI4; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tonga < %s | FileCheck %s --check-prefixes=TONGA5; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=r600 -mcpu=redwood < %s | FileCheck %s --check-prefixes=EG6 7define amdgpu_kernel void @srem_i16_7(ptr addrspace(1) %out, ptr addrspace(1) %in) {8; GCN-LABEL: srem_i16_7:9; GCN:       ; %bb.0:10; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2411; GCN-NEXT:    v_mov_b32_e32 v0, 012; GCN-NEXT:    s_waitcnt lgkmcnt(0)13; GCN-NEXT:    global_load_ushort v1, v0, s[2:3]14; GCN-NEXT:    s_waitcnt vmcnt(0)15; GCN-NEXT:    v_readfirstlane_b32 s2, v116; GCN-NEXT:    s_sext_i32_i16 s2, s217; GCN-NEXT:    s_mulk_i32 s2, 0x492518; GCN-NEXT:    s_lshr_b32 s3, s2, 3119; GCN-NEXT:    s_ashr_i32 s2, s2, 1720; GCN-NEXT:    s_add_i32 s2, s2, s321; GCN-NEXT:    s_mul_i32 s2, s2, 722; GCN-NEXT:    v_subrev_u32_e32 v1, s2, v123; GCN-NEXT:    global_store_short v0, v1, s[0:1]24; GCN-NEXT:    s_endpgm25;26; TAHITI-LABEL: srem_i16_7:27; TAHITI:       ; %bb.0:28; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x929; TAHITI-NEXT:    s_mov_b32 s7, 0xf00030; TAHITI-NEXT:    s_mov_b32 s6, -131; TAHITI-NEXT:    s_mov_b32 s10, s632; TAHITI-NEXT:    s_mov_b32 s11, s733; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)34; TAHITI-NEXT:    s_mov_b32 s8, s235; TAHITI-NEXT:    s_mov_b32 s9, s336; TAHITI-NEXT:    buffer_load_sshort v0, off, s[8:11], 037; TAHITI-NEXT:    s_mov_b32 s4, s038; TAHITI-NEXT:    s_mov_b32 s5, s139; TAHITI-NEXT:    s_waitcnt vmcnt(0)40; TAHITI-NEXT:    v_readfirstlane_b32 s0, v041; TAHITI-NEXT:    s_mulk_i32 s0, 0x492542; TAHITI-NEXT:    s_lshr_b32 s1, s0, 3143; TAHITI-NEXT:    s_ashr_i32 s0, s0, 1744; TAHITI-NEXT:    s_add_i32 s0, s0, s145; TAHITI-NEXT:    s_mul_i32 s0, s0, 746; TAHITI-NEXT:    v_subrev_i32_e32 v0, vcc, s0, v047; TAHITI-NEXT:    buffer_store_short v0, off, s[4:7], 048; TAHITI-NEXT:    s_endpgm49;50; TONGA-LABEL: srem_i16_7:51; TONGA:       ; %bb.0:52; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2453; TONGA-NEXT:    s_waitcnt lgkmcnt(0)54; TONGA-NEXT:    v_mov_b32_e32 v0, s255; TONGA-NEXT:    v_mov_b32_e32 v1, s356; TONGA-NEXT:    flat_load_ushort v2, v[0:1]57; TONGA-NEXT:    v_mov_b32_e32 v0, s058; TONGA-NEXT:    v_mov_b32_e32 v1, s159; TONGA-NEXT:    s_waitcnt vmcnt(0)60; TONGA-NEXT:    v_readfirstlane_b32 s0, v261; TONGA-NEXT:    s_sext_i32_i16 s0, s062; TONGA-NEXT:    s_mulk_i32 s0, 0x492563; TONGA-NEXT:    s_lshr_b32 s1, s0, 3164; TONGA-NEXT:    s_ashr_i32 s0, s0, 1765; TONGA-NEXT:    s_add_i32 s0, s0, s166; TONGA-NEXT:    s_mul_i32 s0, s0, 767; TONGA-NEXT:    v_subrev_u32_e32 v2, vcc, s0, v268; TONGA-NEXT:    flat_store_short v[0:1], v269; TONGA-NEXT:    s_endpgm70;71; EG-LABEL: srem_i16_7:72; EG:       ; %bb.0:73; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]74; EG-NEXT:    TEX 0 @675; EG-NEXT:    ALU 22, @9, KC0[CB0:0-32], KC1[]76; EG-NEXT:    MEM_RAT MSKOR T0.XW, T1.X77; EG-NEXT:    CF_END78; EG-NEXT:    PAD79; EG-NEXT:    Fetch clause starting at 6:80; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 0, #181; EG-NEXT:    ALU clause starting at 8:82; EG-NEXT:     MOV * T0.X, KC0[2].Z,83; EG-NEXT:    ALU clause starting at 9:84; EG-NEXT:     BFE_INT * T0.W, T0.X, 0.0, literal.x,85; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)86; EG-NEXT:     MULLO_INT * T0.Y, PV.W, literal.x,87; EG-NEXT:    18725(2.623931e-41), 0(0.000000e+00)88; EG-NEXT:     ASHR T0.W, PS, literal.x,89; EG-NEXT:     LSHR * T1.W, PS, literal.y,90; EG-NEXT:    17(2.382207e-44), 31(4.344025e-44)91; EG-NEXT:     ADD_INT * T0.W, PV.W, PS,92; EG-NEXT:     MULLO_INT * T0.Y, PV.W, literal.x,93; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)94; EG-NEXT:     AND_INT T0.W, KC0[2].Y, literal.x,95; EG-NEXT:     SUB_INT * T1.W, T0.X, PS,96; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)97; EG-NEXT:     AND_INT T1.W, PS, literal.x,98; EG-NEXT:     LSHL * T0.W, PV.W, literal.y,99; EG-NEXT:    65535(9.183409e-41), 3(4.203895e-45)100; EG-NEXT:     LSHL T0.X, PV.W, PS,101; EG-NEXT:     LSHL * T0.W, literal.x, PS,102; EG-NEXT:    65535(9.183409e-41), 0(0.000000e+00)103; EG-NEXT:     MOV T0.Y, 0.0,104; EG-NEXT:     MOV * T0.Z, 0.0,105; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,106; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)107  %num = load i16, ptr addrspace(1) %in108  %result = srem i16 %num, 7109  store i16 %result, ptr addrspace(1) %out110  ret void111}112 113define amdgpu_kernel void @srem_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {114; GCN-LABEL: srem_i32:115; GCN:       ; %bb.0:116; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24117; GCN-NEXT:    v_mov_b32_e32 v2, 0118; GCN-NEXT:    s_waitcnt lgkmcnt(0)119; GCN-NEXT:    global_load_dwordx2 v[0:1], v2, s[2:3]120; GCN-NEXT:    s_waitcnt vmcnt(0)121; GCN-NEXT:    v_readfirstlane_b32 s2, v1122; GCN-NEXT:    s_abs_i32 s2, s2123; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s2124; GCN-NEXT:    v_readfirstlane_b32 s3, v0125; GCN-NEXT:    s_sub_i32 s5, 0, s2126; GCN-NEXT:    s_ashr_i32 s4, s3, 31127; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1128; GCN-NEXT:    s_abs_i32 s3, s3129; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v1130; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0131; GCN-NEXT:    v_readfirstlane_b32 s6, v0132; GCN-NEXT:    s_mul_i32 s5, s5, s6133; GCN-NEXT:    s_mul_hi_u32 s5, s6, s5134; GCN-NEXT:    s_add_i32 s6, s6, s5135; GCN-NEXT:    s_mul_hi_u32 s5, s3, s6136; GCN-NEXT:    s_mul_i32 s5, s5, s2137; GCN-NEXT:    s_sub_i32 s3, s3, s5138; GCN-NEXT:    s_sub_i32 s5, s3, s2139; GCN-NEXT:    s_cmp_ge_u32 s3, s2140; GCN-NEXT:    s_cselect_b32 s3, s5, s3141; GCN-NEXT:    s_sub_i32 s5, s3, s2142; GCN-NEXT:    s_cmp_ge_u32 s3, s2143; GCN-NEXT:    s_cselect_b32 s2, s5, s3144; GCN-NEXT:    s_xor_b32 s2, s2, s4145; GCN-NEXT:    s_sub_i32 s2, s2, s4146; GCN-NEXT:    v_mov_b32_e32 v0, s2147; GCN-NEXT:    global_store_dword v2, v0, s[0:1]148; GCN-NEXT:    s_endpgm149;150; TAHITI-LABEL: srem_i32:151; TAHITI:       ; %bb.0:152; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9153; TAHITI-NEXT:    s_mov_b32 s7, 0xf000154; TAHITI-NEXT:    s_mov_b32 s6, -1155; TAHITI-NEXT:    s_mov_b32 s10, s6156; TAHITI-NEXT:    s_mov_b32 s11, s7157; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)158; TAHITI-NEXT:    s_mov_b32 s8, s2159; TAHITI-NEXT:    s_mov_b32 s9, s3160; TAHITI-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0161; TAHITI-NEXT:    s_mov_b32 s5, s1162; TAHITI-NEXT:    s_mov_b32 s4, s0163; TAHITI-NEXT:    s_waitcnt vmcnt(0)164; TAHITI-NEXT:    v_readfirstlane_b32 s2, v1165; TAHITI-NEXT:    s_abs_i32 s2, s2166; TAHITI-NEXT:    v_cvt_f32_u32_e32 v1, s2167; TAHITI-NEXT:    s_sub_i32 s3, 0, s2168; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v1, v1169; TAHITI-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1170; TAHITI-NEXT:    v_cvt_u32_f32_e32 v1, v1171; TAHITI-NEXT:    v_mul_lo_u32 v2, s3, v1172; TAHITI-NEXT:    v_readfirstlane_b32 s3, v0173; TAHITI-NEXT:    s_abs_i32 s8, s3174; TAHITI-NEXT:    s_ashr_i32 s0, s3, 31175; TAHITI-NEXT:    v_mul_hi_u32 v2, v1, v2176; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v1, v2177; TAHITI-NEXT:    v_mul_hi_u32 v0, s8, v0178; TAHITI-NEXT:    v_readfirstlane_b32 s1, v0179; TAHITI-NEXT:    s_mul_i32 s1, s1, s2180; TAHITI-NEXT:    s_sub_i32 s1, s8, s1181; TAHITI-NEXT:    s_sub_i32 s3, s1, s2182; TAHITI-NEXT:    s_cmp_ge_u32 s1, s2183; TAHITI-NEXT:    s_cselect_b32 s1, s3, s1184; TAHITI-NEXT:    s_sub_i32 s3, s1, s2185; TAHITI-NEXT:    s_cmp_ge_u32 s1, s2186; TAHITI-NEXT:    s_cselect_b32 s1, s3, s1187; TAHITI-NEXT:    s_xor_b32 s1, s1, s0188; TAHITI-NEXT:    s_sub_i32 s0, s1, s0189; TAHITI-NEXT:    v_mov_b32_e32 v0, s0190; TAHITI-NEXT:    buffer_store_dword v0, off, s[4:7], 0191; TAHITI-NEXT:    s_endpgm192;193; TONGA-LABEL: srem_i32:194; TONGA:       ; %bb.0:195; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24196; TONGA-NEXT:    s_waitcnt lgkmcnt(0)197; TONGA-NEXT:    v_mov_b32_e32 v0, s2198; TONGA-NEXT:    v_mov_b32_e32 v1, s3199; TONGA-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]200; TONGA-NEXT:    s_waitcnt vmcnt(0)201; TONGA-NEXT:    v_readfirstlane_b32 s2, v1202; TONGA-NEXT:    s_abs_i32 s2, s2203; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, s2204; TONGA-NEXT:    s_sub_i32 s3, 0, s2205; TONGA-NEXT:    v_rcp_iflag_f32_e32 v1, v1206; TONGA-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1207; TONGA-NEXT:    v_cvt_u32_f32_e32 v1, v1208; TONGA-NEXT:    v_mul_lo_u32 v2, s3, v1209; TONGA-NEXT:    v_readfirstlane_b32 s3, v0210; TONGA-NEXT:    s_abs_i32 s4, s3211; TONGA-NEXT:    v_mul_hi_u32 v2, v1, v2212; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v1, v2213; TONGA-NEXT:    v_mul_hi_u32 v2, s4, v0214; TONGA-NEXT:    v_mov_b32_e32 v1, s1215; TONGA-NEXT:    v_mov_b32_e32 v0, s0216; TONGA-NEXT:    s_ashr_i32 s0, s3, 31217; TONGA-NEXT:    v_readfirstlane_b32 s1, v2218; TONGA-NEXT:    s_mul_i32 s1, s1, s2219; TONGA-NEXT:    s_sub_i32 s1, s4, s1220; TONGA-NEXT:    s_sub_i32 s3, s1, s2221; TONGA-NEXT:    s_cmp_ge_u32 s1, s2222; TONGA-NEXT:    s_cselect_b32 s1, s3, s1223; TONGA-NEXT:    s_sub_i32 s3, s1, s2224; TONGA-NEXT:    s_cmp_ge_u32 s1, s2225; TONGA-NEXT:    s_cselect_b32 s1, s3, s1226; TONGA-NEXT:    s_xor_b32 s1, s1, s0227; TONGA-NEXT:    s_sub_i32 s0, s1, s0228; TONGA-NEXT:    v_mov_b32_e32 v2, s0229; TONGA-NEXT:    flat_store_dword v[0:1], v2230; TONGA-NEXT:    s_endpgm231;232; EG-LABEL: srem_i32:233; EG:       ; %bb.0:234; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]235; EG-NEXT:    TEX 0 @6236; EG-NEXT:    ALU 23, @9, KC0[CB0:0-32], KC1[]237; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1238; EG-NEXT:    CF_END239; EG-NEXT:    PAD240; EG-NEXT:    Fetch clause starting at 6:241; EG-NEXT:     VTX_READ_64 T0.XY, T0.X, 0, #1242; EG-NEXT:    ALU clause starting at 8:243; EG-NEXT:     MOV * T0.X, KC0[2].Z,244; EG-NEXT:    ALU clause starting at 9:245; EG-NEXT:     SETGT_INT * T0.W, 0.0, T0.Y,246; EG-NEXT:     ADD_INT * T1.W, T0.Y, PV.W,247; EG-NEXT:     XOR_INT * T0.W, PV.W, T0.W,248; EG-NEXT:     SUB_INT T1.W, 0.0, PV.W,249; EG-NEXT:     RECIP_UINT * T0.Y, PV.W,250; EG-NEXT:     SETGT_INT T2.W, 0.0, T0.X,251; EG-NEXT:     MULLO_INT * T0.Z, PV.W, PS,252; EG-NEXT:     ADD_INT T1.W, T0.X, PV.W,253; EG-NEXT:     MULHI * T0.X, T0.Y, PS,254; EG-NEXT:     ADD_INT T3.W, T0.Y, PS,255; EG-NEXT:     XOR_INT * T1.W, PV.W, T2.W,256; EG-NEXT:     MULHI * T0.X, PS, PV.W,257; EG-NEXT:     MULLO_INT * T0.X, PS, T0.W,258; EG-NEXT:     SUB_INT * T1.W, T1.W, PS,259; EG-NEXT:     SETGE_UINT T3.W, PV.W, T0.W,260; EG-NEXT:     SUB_INT * T4.W, PV.W, T0.W,261; EG-NEXT:     CNDE_INT * T1.W, PV.W, T1.W, PS,262; EG-NEXT:     SETGE_UINT T3.W, PV.W, T0.W,263; EG-NEXT:     SUB_INT * T0.W, PV.W, T0.W,264; EG-NEXT:     CNDE_INT * T0.W, PV.W, T1.W, PS,265; EG-NEXT:     XOR_INT * T0.W, PV.W, T2.W,266; EG-NEXT:     SUB_INT T0.X, PV.W, T2.W,267; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,268; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)269  %den_ptr = getelementptr i32, ptr addrspace(1) %in, i32 1270  %num = load i32, ptr addrspace(1) %in271  %den = load i32, ptr addrspace(1) %den_ptr272  %result = srem i32 %num, %den273  store i32 %result, ptr addrspace(1) %out274  ret void275}276 277define amdgpu_kernel void @srem_i32_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {278; GCN-LABEL: srem_i32_4:279; GCN:       ; %bb.0:280; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24281; GCN-NEXT:    v_mov_b32_e32 v0, 0282; GCN-NEXT:    s_waitcnt lgkmcnt(0)283; GCN-NEXT:    global_load_dword v1, v0, s[2:3]284; GCN-NEXT:    s_waitcnt vmcnt(0)285; GCN-NEXT:    v_ashrrev_i32_e32 v2, 31, v1286; GCN-NEXT:    v_lshrrev_b32_e32 v2, 30, v2287; GCN-NEXT:    v_add_u32_e32 v2, v1, v2288; GCN-NEXT:    v_and_b32_e32 v2, -4, v2289; GCN-NEXT:    v_sub_u32_e32 v1, v1, v2290; GCN-NEXT:    global_store_dword v0, v1, s[0:1]291; GCN-NEXT:    s_endpgm292;293; TAHITI-LABEL: srem_i32_4:294; TAHITI:       ; %bb.0:295; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9296; TAHITI-NEXT:    s_mov_b32 s7, 0xf000297; TAHITI-NEXT:    s_mov_b32 s6, -1298; TAHITI-NEXT:    s_mov_b32 s10, s6299; TAHITI-NEXT:    s_mov_b32 s11, s7300; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)301; TAHITI-NEXT:    s_mov_b32 s8, s2302; TAHITI-NEXT:    s_mov_b32 s9, s3303; TAHITI-NEXT:    buffer_load_dword v0, off, s[8:11], 0304; TAHITI-NEXT:    s_mov_b32 s4, s0305; TAHITI-NEXT:    s_mov_b32 s5, s1306; TAHITI-NEXT:    s_waitcnt vmcnt(0)307; TAHITI-NEXT:    v_ashrrev_i32_e32 v1, 31, v0308; TAHITI-NEXT:    v_lshrrev_b32_e32 v1, 30, v1309; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v0310; TAHITI-NEXT:    v_and_b32_e32 v1, -4, v1311; TAHITI-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0312; TAHITI-NEXT:    buffer_store_dword v0, off, s[4:7], 0313; TAHITI-NEXT:    s_endpgm314;315; TONGA-LABEL: srem_i32_4:316; TONGA:       ; %bb.0:317; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24318; TONGA-NEXT:    s_waitcnt lgkmcnt(0)319; TONGA-NEXT:    v_mov_b32_e32 v0, s2320; TONGA-NEXT:    v_mov_b32_e32 v1, s3321; TONGA-NEXT:    flat_load_dword v2, v[0:1]322; TONGA-NEXT:    v_mov_b32_e32 v0, s0323; TONGA-NEXT:    v_mov_b32_e32 v1, s1324; TONGA-NEXT:    s_waitcnt vmcnt(0)325; TONGA-NEXT:    v_ashrrev_i32_e32 v3, 31, v2326; TONGA-NEXT:    v_lshrrev_b32_e32 v3, 30, v3327; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v3, v2328; TONGA-NEXT:    v_and_b32_e32 v3, -4, v3329; TONGA-NEXT:    v_subrev_u32_e32 v2, vcc, v3, v2330; TONGA-NEXT:    flat_store_dword v[0:1], v2331; TONGA-NEXT:    s_endpgm332;333; EG-LABEL: srem_i32_4:334; EG:       ; %bb.0:335; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]336; EG-NEXT:    TEX 0 @6337; EG-NEXT:    ALU 9, @9, KC0[CB0:0-32], KC1[]338; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1339; EG-NEXT:    CF_END340; EG-NEXT:    PAD341; EG-NEXT:    Fetch clause starting at 6:342; EG-NEXT:     VTX_READ_32 T0.X, T0.X, 0, #1343; EG-NEXT:    ALU clause starting at 8:344; EG-NEXT:     MOV * T0.X, KC0[2].Z,345; EG-NEXT:    ALU clause starting at 9:346; EG-NEXT:     ASHR * T0.W, T0.X, literal.x,347; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)348; EG-NEXT:     LSHR * T0.W, PV.W, literal.x,349; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)350; EG-NEXT:     ADD_INT * T0.W, T0.X, PV.W,351; EG-NEXT:     AND_INT * T0.W, PV.W, literal.x,352; EG-NEXT:    -4(nan), 0(0.000000e+00)353; EG-NEXT:     SUB_INT T0.X, T0.X, PV.W,354; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,355; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)356  %num = load i32, ptr addrspace(1) %in357  %result = srem i32 %num, 4358  store i32 %result, ptr addrspace(1) %out359  ret void360}361 362; FIXME: uniform i16 srem should not use VALU instructions363define amdgpu_kernel void @srem_i32_7(ptr addrspace(1) %out, ptr addrspace(1) %in) {364; GCN-LABEL: srem_i32_7:365; GCN:       ; %bb.0:366; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24367; GCN-NEXT:    v_mov_b32_e32 v0, 0368; GCN-NEXT:    s_waitcnt lgkmcnt(0)369; GCN-NEXT:    global_load_dword v1, v0, s[2:3]370; GCN-NEXT:    s_mov_b32 s2, 0x92492493371; GCN-NEXT:    s_waitcnt vmcnt(0)372; GCN-NEXT:    v_mul_hi_i32 v2, v1, s2373; GCN-NEXT:    v_add_u32_e32 v2, v2, v1374; GCN-NEXT:    v_lshrrev_b32_e32 v3, 31, v2375; GCN-NEXT:    v_ashrrev_i32_e32 v2, 2, v2376; GCN-NEXT:    v_add_u32_e32 v2, v2, v3377; GCN-NEXT:    v_mul_lo_u32 v2, v2, 7378; GCN-NEXT:    v_sub_u32_e32 v1, v1, v2379; GCN-NEXT:    global_store_dword v0, v1, s[0:1]380; GCN-NEXT:    s_endpgm381;382; TAHITI-LABEL: srem_i32_7:383; TAHITI:       ; %bb.0:384; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9385; TAHITI-NEXT:    s_mov_b32 s7, 0xf000386; TAHITI-NEXT:    s_mov_b32 s6, -1387; TAHITI-NEXT:    s_mov_b32 s10, s6388; TAHITI-NEXT:    s_mov_b32 s11, s7389; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)390; TAHITI-NEXT:    s_mov_b32 s8, s2391; TAHITI-NEXT:    s_mov_b32 s9, s3392; TAHITI-NEXT:    buffer_load_dword v0, off, s[8:11], 0393; TAHITI-NEXT:    s_mov_b32 s2, 0x92492493394; TAHITI-NEXT:    s_mov_b32 s4, s0395; TAHITI-NEXT:    s_mov_b32 s5, s1396; TAHITI-NEXT:    s_waitcnt vmcnt(0)397; TAHITI-NEXT:    v_mul_hi_i32 v1, v0, s2398; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v0399; TAHITI-NEXT:    v_lshrrev_b32_e32 v2, 31, v1400; TAHITI-NEXT:    v_ashrrev_i32_e32 v1, 2, v1401; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v2402; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, 7403; TAHITI-NEXT:    v_subrev_i32_e32 v0, vcc, v1, v0404; TAHITI-NEXT:    buffer_store_dword v0, off, s[4:7], 0405; TAHITI-NEXT:    s_endpgm406;407; TONGA-LABEL: srem_i32_7:408; TONGA:       ; %bb.0:409; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24410; TONGA-NEXT:    s_waitcnt lgkmcnt(0)411; TONGA-NEXT:    v_mov_b32_e32 v0, s2412; TONGA-NEXT:    v_mov_b32_e32 v1, s3413; TONGA-NEXT:    flat_load_dword v2, v[0:1]414; TONGA-NEXT:    s_mov_b32 s2, 0x92492493415; TONGA-NEXT:    s_waitcnt vmcnt(0)416; TONGA-NEXT:    v_mul_hi_i32 v0, v2, s2417; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v2418; TONGA-NEXT:    v_lshrrev_b32_e32 v1, 31, v0419; TONGA-NEXT:    v_ashrrev_i32_e32 v0, 2, v0420; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v1421; TONGA-NEXT:    v_mul_lo_u32 v3, v0, 7422; TONGA-NEXT:    v_mov_b32_e32 v0, s0423; TONGA-NEXT:    v_mov_b32_e32 v1, s1424; TONGA-NEXT:    v_subrev_u32_e32 v2, vcc, v3, v2425; TONGA-NEXT:    flat_store_dword v[0:1], v2426; TONGA-NEXT:    s_endpgm427;428; EG-LABEL: srem_i32_7:429; EG:       ; %bb.0:430; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]431; EG-NEXT:    TEX 0 @6432; EG-NEXT:    ALU 11, @9, KC0[CB0:0-32], KC1[]433; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1434; EG-NEXT:    CF_END435; EG-NEXT:    PAD436; EG-NEXT:    Fetch clause starting at 6:437; EG-NEXT:     VTX_READ_32 T0.X, T0.X, 0, #1438; EG-NEXT:    ALU clause starting at 8:439; EG-NEXT:     MOV * T0.X, KC0[2].Z,440; EG-NEXT:    ALU clause starting at 9:441; EG-NEXT:     MULHI_INT * T0.Y, T0.X, literal.x,442; EG-NEXT:    -1840700269(-6.346950e-28), 0(0.000000e+00)443; EG-NEXT:     ADD_INT * T0.W, PS, T0.X,444; EG-NEXT:     ASHR T1.W, PV.W, literal.x,445; EG-NEXT:     LSHR * T0.W, PV.W, literal.y,446; EG-NEXT:    2(2.802597e-45), 31(4.344025e-44)447; EG-NEXT:     ADD_INT * T0.W, PV.W, PS,448; EG-NEXT:     MULLO_INT * T0.Y, PV.W, literal.x,449; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)450; EG-NEXT:     SUB_INT T0.X, T0.X, PS,451; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,452; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)453  %num = load i32, ptr addrspace(1) %in454  %result = srem i32 %num, 7455  store i32 %result, ptr addrspace(1) %out456  ret void457}458 459define amdgpu_kernel void @srem_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {460; GCN-LABEL: srem_v2i32:461; GCN:       ; %bb.0:462; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24463; GCN-NEXT:    v_mov_b32_e32 v4, 0464; GCN-NEXT:    s_waitcnt lgkmcnt(0)465; GCN-NEXT:    global_load_dwordx4 v[0:3], v4, s[2:3]466; GCN-NEXT:    s_waitcnt vmcnt(0)467; GCN-NEXT:    v_readfirstlane_b32 s2, v2468; GCN-NEXT:    s_abs_i32 s2, s2469; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s2470; GCN-NEXT:    v_readfirstlane_b32 s4, v0471; GCN-NEXT:    s_sub_i32 s6, 0, s2472; GCN-NEXT:    s_ashr_i32 s5, s4, 31473; GCN-NEXT:    v_rcp_iflag_f32_e32 v2, v2474; GCN-NEXT:    s_abs_i32 s4, s4475; GCN-NEXT:    v_readfirstlane_b32 s3, v3476; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v2477; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0478; GCN-NEXT:    v_readfirstlane_b32 s7, v0479; GCN-NEXT:    s_mul_i32 s6, s6, s7480; GCN-NEXT:    s_mul_hi_u32 s6, s7, s6481; GCN-NEXT:    s_add_i32 s7, s7, s6482; GCN-NEXT:    s_mul_hi_u32 s6, s4, s7483; GCN-NEXT:    s_mul_i32 s6, s6, s2484; GCN-NEXT:    s_sub_i32 s4, s4, s6485; GCN-NEXT:    s_sub_i32 s6, s4, s2486; GCN-NEXT:    s_cmp_ge_u32 s4, s2487; GCN-NEXT:    s_cselect_b32 s4, s6, s4488; GCN-NEXT:    s_sub_i32 s6, s4, s2489; GCN-NEXT:    s_cmp_ge_u32 s4, s2490; GCN-NEXT:    s_cselect_b32 s2, s6, s4491; GCN-NEXT:    s_abs_i32 s3, s3492; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s3493; GCN-NEXT:    s_xor_b32 s2, s2, s5494; GCN-NEXT:    s_sub_i32 s7, 0, s3495; GCN-NEXT:    s_sub_i32 s2, s2, s5496; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v0497; GCN-NEXT:    v_readfirstlane_b32 s4, v1498; GCN-NEXT:    s_ashr_i32 s6, s4, 31499; GCN-NEXT:    s_abs_i32 s4, s4500; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0501; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v0502; GCN-NEXT:    v_readfirstlane_b32 s5, v0503; GCN-NEXT:    s_mul_i32 s7, s7, s5504; GCN-NEXT:    s_mul_hi_u32 s7, s5, s7505; GCN-NEXT:    s_add_i32 s5, s5, s7506; GCN-NEXT:    s_mul_hi_u32 s5, s4, s5507; GCN-NEXT:    s_mul_i32 s5, s5, s3508; GCN-NEXT:    s_sub_i32 s4, s4, s5509; GCN-NEXT:    s_sub_i32 s5, s4, s3510; GCN-NEXT:    s_cmp_ge_u32 s4, s3511; GCN-NEXT:    s_cselect_b32 s4, s5, s4512; GCN-NEXT:    s_sub_i32 s5, s4, s3513; GCN-NEXT:    s_cmp_ge_u32 s4, s3514; GCN-NEXT:    s_cselect_b32 s3, s5, s4515; GCN-NEXT:    s_xor_b32 s3, s3, s6516; GCN-NEXT:    s_sub_i32 s3, s3, s6517; GCN-NEXT:    v_mov_b32_e32 v0, s2518; GCN-NEXT:    v_mov_b32_e32 v1, s3519; GCN-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]520; GCN-NEXT:    s_endpgm521;522; TAHITI-LABEL: srem_v2i32:523; TAHITI:       ; %bb.0:524; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x9525; TAHITI-NEXT:    s_mov_b32 s3, 0xf000526; TAHITI-NEXT:    s_mov_b32 s2, -1527; TAHITI-NEXT:    s_mov_b32 s10, s2528; TAHITI-NEXT:    s_mov_b32 s11, s3529; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)530; TAHITI-NEXT:    s_mov_b32 s8, s6531; TAHITI-NEXT:    s_mov_b32 s9, s7532; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0533; TAHITI-NEXT:    s_waitcnt vmcnt(0)534; TAHITI-NEXT:    v_readfirstlane_b32 s0, v2535; TAHITI-NEXT:    s_abs_i32 s0, s0536; TAHITI-NEXT:    v_cvt_f32_u32_e32 v2, s0537; TAHITI-NEXT:    s_sub_i32 s1, 0, s0538; TAHITI-NEXT:    v_readfirstlane_b32 s7, v3539; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v2, v2540; TAHITI-NEXT:    v_mul_f32_e32 v2, 0x4f7ffffe, v2541; TAHITI-NEXT:    v_cvt_u32_f32_e32 v2, v2542; TAHITI-NEXT:    v_mul_lo_u32 v4, s1, v2543; TAHITI-NEXT:    v_readfirstlane_b32 s1, v0544; TAHITI-NEXT:    s_abs_i32 s6, s1545; TAHITI-NEXT:    s_ashr_i32 s8, s1, 31546; TAHITI-NEXT:    v_mul_hi_u32 v4, v2, v4547; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v2, v4548; TAHITI-NEXT:    v_mul_hi_u32 v0, s6, v0549; TAHITI-NEXT:    v_readfirstlane_b32 s1, v0550; TAHITI-NEXT:    s_mul_i32 s1, s1, s0551; TAHITI-NEXT:    s_sub_i32 s1, s6, s1552; TAHITI-NEXT:    s_sub_i32 s6, s1, s0553; TAHITI-NEXT:    s_cmp_ge_u32 s1, s0554; TAHITI-NEXT:    s_cselect_b32 s1, s6, s1555; TAHITI-NEXT:    s_sub_i32 s6, s1, s0556; TAHITI-NEXT:    s_cmp_ge_u32 s1, s0557; TAHITI-NEXT:    s_cselect_b32 s6, s6, s1558; TAHITI-NEXT:    s_abs_i32 s7, s7559; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, s7560; TAHITI-NEXT:    s_sub_i32 s0, 0, s7561; TAHITI-NEXT:    s_mov_b32 s1, s5562; TAHITI-NEXT:    s_xor_b32 s6, s6, s8563; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v0564; TAHITI-NEXT:    s_sub_i32 s6, s6, s8565; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0566; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v0567; TAHITI-NEXT:    v_mul_lo_u32 v2, s0, v0568; TAHITI-NEXT:    s_mov_b32 s0, s4569; TAHITI-NEXT:    v_readfirstlane_b32 s4, v1570; TAHITI-NEXT:    s_abs_i32 s5, s4571; TAHITI-NEXT:    v_mul_hi_u32 v2, v0, v2572; TAHITI-NEXT:    s_ashr_i32 s4, s4, 31573; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v2574; TAHITI-NEXT:    v_mul_hi_u32 v0, s5, v0575; TAHITI-NEXT:    v_readfirstlane_b32 s8, v0576; TAHITI-NEXT:    s_mul_i32 s8, s8, s7577; TAHITI-NEXT:    s_sub_i32 s5, s5, s8578; TAHITI-NEXT:    s_sub_i32 s8, s5, s7579; TAHITI-NEXT:    s_cmp_ge_u32 s5, s7580; TAHITI-NEXT:    s_cselect_b32 s5, s8, s5581; TAHITI-NEXT:    s_sub_i32 s8, s5, s7582; TAHITI-NEXT:    s_cmp_ge_u32 s5, s7583; TAHITI-NEXT:    s_cselect_b32 s5, s8, s5584; TAHITI-NEXT:    s_xor_b32 s5, s5, s4585; TAHITI-NEXT:    s_sub_i32 s4, s5, s4586; TAHITI-NEXT:    v_mov_b32_e32 v0, s6587; TAHITI-NEXT:    v_mov_b32_e32 v1, s4588; TAHITI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0589; TAHITI-NEXT:    s_endpgm590;591; TONGA-LABEL: srem_v2i32:592; TONGA:       ; %bb.0:593; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24594; TONGA-NEXT:    s_waitcnt lgkmcnt(0)595; TONGA-NEXT:    v_mov_b32_e32 v0, s2596; TONGA-NEXT:    v_mov_b32_e32 v1, s3597; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]598; TONGA-NEXT:    s_waitcnt vmcnt(0)599; TONGA-NEXT:    v_readfirstlane_b32 s2, v2600; TONGA-NEXT:    s_abs_i32 s2, s2601; TONGA-NEXT:    v_cvt_f32_u32_e32 v2, s2602; TONGA-NEXT:    s_sub_i32 s3, 0, s2603; TONGA-NEXT:    v_readfirstlane_b32 s5, v3604; TONGA-NEXT:    v_mov_b32_e32 v3, s1605; TONGA-NEXT:    v_rcp_iflag_f32_e32 v2, v2606; TONGA-NEXT:    v_mul_f32_e32 v2, 0x4f7ffffe, v2607; TONGA-NEXT:    v_cvt_u32_f32_e32 v2, v2608; TONGA-NEXT:    v_mul_lo_u32 v4, s3, v2609; TONGA-NEXT:    v_readfirstlane_b32 s3, v0610; TONGA-NEXT:    s_abs_i32 s4, s3611; TONGA-NEXT:    s_ashr_i32 s3, s3, 31612; TONGA-NEXT:    v_mul_hi_u32 v4, v2, v4613; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v2, v4614; TONGA-NEXT:    v_mul_hi_u32 v0, s4, v0615; TONGA-NEXT:    v_mov_b32_e32 v2, s0616; TONGA-NEXT:    v_readfirstlane_b32 s0, v1617; TONGA-NEXT:    v_readfirstlane_b32 s6, v0618; TONGA-NEXT:    s_mul_i32 s6, s6, s2619; TONGA-NEXT:    s_sub_i32 s4, s4, s6620; TONGA-NEXT:    s_sub_i32 s6, s4, s2621; TONGA-NEXT:    s_cmp_ge_u32 s4, s2622; TONGA-NEXT:    s_cselect_b32 s4, s6, s4623; TONGA-NEXT:    s_sub_i32 s6, s4, s2624; TONGA-NEXT:    s_cmp_ge_u32 s4, s2625; TONGA-NEXT:    s_cselect_b32 s2, s6, s4626; TONGA-NEXT:    s_abs_i32 s4, s5627; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s4628; TONGA-NEXT:    s_sub_i32 s5, 0, s4629; TONGA-NEXT:    s_abs_i32 s1, s0630; TONGA-NEXT:    s_xor_b32 s2, s2, s3631; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v0632; TONGA-NEXT:    s_sub_i32 s2, s2, s3633; TONGA-NEXT:    s_ashr_i32 s0, s0, 31634; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0635; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v0636; TONGA-NEXT:    v_mul_lo_u32 v4, s5, v0637; TONGA-NEXT:    v_mul_hi_u32 v4, v0, v4638; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v4639; TONGA-NEXT:    v_mul_hi_u32 v0, s1, v0640; TONGA-NEXT:    v_readfirstlane_b32 s3, v0641; TONGA-NEXT:    s_mul_i32 s3, s3, s4642; TONGA-NEXT:    s_sub_i32 s1, s1, s3643; TONGA-NEXT:    s_sub_i32 s3, s1, s4644; TONGA-NEXT:    s_cmp_ge_u32 s1, s4645; TONGA-NEXT:    s_cselect_b32 s1, s3, s1646; TONGA-NEXT:    s_sub_i32 s3, s1, s4647; TONGA-NEXT:    s_cmp_ge_u32 s1, s4648; TONGA-NEXT:    s_cselect_b32 s1, s3, s1649; TONGA-NEXT:    s_xor_b32 s1, s1, s0650; TONGA-NEXT:    s_sub_i32 s0, s1, s0651; TONGA-NEXT:    v_mov_b32_e32 v0, s2652; TONGA-NEXT:    v_mov_b32_e32 v1, s0653; TONGA-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]654; TONGA-NEXT:    s_endpgm655;656; EG-LABEL: srem_v2i32:657; EG:       ; %bb.0:658; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]659; EG-NEXT:    TEX 0 @6660; EG-NEXT:    ALU 45, @9, KC0[CB0:0-32], KC1[]661; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1662; EG-NEXT:    CF_END663; EG-NEXT:    PAD664; EG-NEXT:    Fetch clause starting at 6:665; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #1666; EG-NEXT:    ALU clause starting at 8:667; EG-NEXT:     MOV * T0.X, KC0[2].Z,668; EG-NEXT:    ALU clause starting at 9:669; EG-NEXT:     SETGT_INT * T1.W, 0.0, T0.W,670; EG-NEXT:     ADD_INT T0.W, T0.W, PV.W,671; EG-NEXT:     SETGT_INT * T2.W, 0.0, T0.Z,672; EG-NEXT:     XOR_INT * T0.W, PV.W, T1.W,673; EG-NEXT:     SUB_INT T1.Z, 0.0, PV.W,674; EG-NEXT:     ADD_INT T1.W, T0.Z, T2.W,675; EG-NEXT:     RECIP_UINT * T0.Z, PV.W,676; EG-NEXT:     XOR_INT T1.W, PV.W, T2.W,677; EG-NEXT:     MULLO_INT * T1.X, PV.Z, PS,678; EG-NEXT:     SUB_INT T2.W, 0.0, PV.W,679; EG-NEXT:     RECIP_UINT * T1.Y, PV.W,680; EG-NEXT:     SETGT_INT T3.W, 0.0, T0.X,681; EG-NEXT:     MULLO_INT * T1.Z, PV.W, PS,682; EG-NEXT:     SETGT_INT T2.Z, 0.0, T0.Y,683; EG-NEXT:     ADD_INT T2.W, T0.X, PV.W,684; EG-NEXT:     MULHI * T0.X, T1.Y, PS,685; EG-NEXT:     ADD_INT T1.Y, T1.Y, PS,686; EG-NEXT:     XOR_INT T1.Z, PV.W, T3.W,687; EG-NEXT:     ADD_INT T2.W, T0.Y, PV.Z, BS:VEC_120/SCL_212688; EG-NEXT:     MULHI * T0.X, T0.Z, T1.X,689; EG-NEXT:     ADD_INT T0.Z, T0.Z, PS,690; EG-NEXT:     XOR_INT T2.W, PV.W, T2.Z,691; EG-NEXT:     MULHI * T0.X, PV.Z, PV.Y,692; EG-NEXT:     MULHI * T0.Y, PV.W, PV.Z,693; EG-NEXT:     MULLO_INT * T0.Y, PS, T0.W,694; EG-NEXT:     SUB_INT T2.W, T2.W, PS,695; EG-NEXT:     MULLO_INT * T0.X, T0.X, T1.W,696; EG-NEXT:     SUB_INT T0.Z, T1.Z, PS,697; EG-NEXT:     SETGE_UINT T4.W, PV.W, T0.W,698; EG-NEXT:     SUB_INT * T5.W, PV.W, T0.W,699; EG-NEXT:     CNDE_INT T1.Z, PV.W, T2.W, PS, BS:VEC_021/SCL_122700; EG-NEXT:     SETGE_UINT T2.W, PV.Z, T1.W,701; EG-NEXT:     SUB_INT * T4.W, PV.Z, T1.W,702; EG-NEXT:     CNDE_INT T0.Z, PV.W, T0.Z, PS,703; EG-NEXT:     SETGE_UINT T2.W, PV.Z, T0.W,704; EG-NEXT:     SUB_INT * T0.W, PV.Z, T0.W,705; EG-NEXT:     CNDE_INT T1.Z, PV.W, T1.Z, PS,706; EG-NEXT:     SETGE_UINT T0.W, PV.Z, T1.W,707; EG-NEXT:     SUB_INT * T1.W, PV.Z, T1.W,708; EG-NEXT:     CNDE_INT T0.W, PV.W, T0.Z, PS, BS:VEC_021/SCL_122709; EG-NEXT:     XOR_INT * T1.W, PV.Z, T2.Z,710; EG-NEXT:     SUB_INT T0.Y, PS, T2.Z,711; EG-NEXT:     XOR_INT * T0.W, PV.W, T3.W,712; EG-NEXT:     SUB_INT T0.X, PV.W, T3.W,713; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,714; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)715  %den_ptr = getelementptr <2 x i32>, ptr addrspace(1) %in, i32 1716  %num = load <2 x i32>, ptr addrspace(1) %in717  %den = load <2 x i32>, ptr addrspace(1) %den_ptr718  %result = srem <2 x i32> %num, %den719  store <2 x i32> %result, ptr addrspace(1) %out720  ret void721}722 723define amdgpu_kernel void @srem_v2i32_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {724; GCN-LABEL: srem_v2i32_4:725; GCN:       ; %bb.0:726; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24727; GCN-NEXT:    v_mov_b32_e32 v2, 0728; GCN-NEXT:    s_waitcnt lgkmcnt(0)729; GCN-NEXT:    global_load_dwordx2 v[0:1], v2, s[2:3]730; GCN-NEXT:    s_waitcnt vmcnt(0)731; GCN-NEXT:    v_readfirstlane_b32 s2, v0732; GCN-NEXT:    v_readfirstlane_b32 s3, v1733; GCN-NEXT:    s_ashr_i32 s4, s2, 31734; GCN-NEXT:    s_ashr_i32 s5, s3, 31735; GCN-NEXT:    s_lshr_b32 s4, s4, 30736; GCN-NEXT:    s_lshr_b32 s5, s5, 30737; GCN-NEXT:    s_add_i32 s4, s2, s4738; GCN-NEXT:    s_add_i32 s5, s3, s5739; GCN-NEXT:    s_and_b32 s4, s4, -4740; GCN-NEXT:    s_and_b32 s5, s5, -4741; GCN-NEXT:    s_sub_i32 s2, s2, s4742; GCN-NEXT:    s_sub_i32 s3, s3, s5743; GCN-NEXT:    v_mov_b32_e32 v0, s2744; GCN-NEXT:    v_mov_b32_e32 v1, s3745; GCN-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]746; GCN-NEXT:    s_endpgm747;748; TAHITI-LABEL: srem_v2i32_4:749; TAHITI:       ; %bb.0:750; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9751; TAHITI-NEXT:    s_mov_b32 s7, 0xf000752; TAHITI-NEXT:    s_mov_b32 s6, -1753; TAHITI-NEXT:    s_mov_b32 s10, s6754; TAHITI-NEXT:    s_mov_b32 s11, s7755; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)756; TAHITI-NEXT:    s_mov_b32 s8, s2757; TAHITI-NEXT:    s_mov_b32 s9, s3758; TAHITI-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0759; TAHITI-NEXT:    s_mov_b32 s4, s0760; TAHITI-NEXT:    s_mov_b32 s5, s1761; TAHITI-NEXT:    s_waitcnt vmcnt(0)762; TAHITI-NEXT:    v_readfirstlane_b32 s0, v0763; TAHITI-NEXT:    v_readfirstlane_b32 s1, v1764; TAHITI-NEXT:    s_ashr_i32 s2, s0, 31765; TAHITI-NEXT:    s_ashr_i32 s3, s1, 31766; TAHITI-NEXT:    s_lshr_b32 s2, s2, 30767; TAHITI-NEXT:    s_lshr_b32 s3, s3, 30768; TAHITI-NEXT:    s_add_i32 s2, s0, s2769; TAHITI-NEXT:    s_add_i32 s3, s1, s3770; TAHITI-NEXT:    s_and_b32 s2, s2, -4771; TAHITI-NEXT:    s_and_b32 s3, s3, -4772; TAHITI-NEXT:    s_sub_i32 s0, s0, s2773; TAHITI-NEXT:    s_sub_i32 s1, s1, s3774; TAHITI-NEXT:    v_mov_b32_e32 v0, s0775; TAHITI-NEXT:    v_mov_b32_e32 v1, s1776; TAHITI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0777; TAHITI-NEXT:    s_endpgm778;779; TONGA-LABEL: srem_v2i32_4:780; TONGA:       ; %bb.0:781; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24782; TONGA-NEXT:    s_waitcnt lgkmcnt(0)783; TONGA-NEXT:    v_mov_b32_e32 v0, s2784; TONGA-NEXT:    v_mov_b32_e32 v1, s3785; TONGA-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]786; TONGA-NEXT:    v_mov_b32_e32 v2, s0787; TONGA-NEXT:    v_mov_b32_e32 v3, s1788; TONGA-NEXT:    s_waitcnt vmcnt(0)789; TONGA-NEXT:    v_readfirstlane_b32 s0, v0790; TONGA-NEXT:    v_readfirstlane_b32 s1, v1791; TONGA-NEXT:    s_ashr_i32 s2, s0, 31792; TONGA-NEXT:    s_ashr_i32 s3, s1, 31793; TONGA-NEXT:    s_lshr_b32 s2, s2, 30794; TONGA-NEXT:    s_lshr_b32 s3, s3, 30795; TONGA-NEXT:    s_add_i32 s2, s0, s2796; TONGA-NEXT:    s_add_i32 s3, s1, s3797; TONGA-NEXT:    s_and_b32 s2, s2, -4798; TONGA-NEXT:    s_and_b32 s3, s3, -4799; TONGA-NEXT:    s_sub_i32 s0, s0, s2800; TONGA-NEXT:    s_sub_i32 s1, s1, s3801; TONGA-NEXT:    v_mov_b32_e32 v0, s0802; TONGA-NEXT:    v_mov_b32_e32 v1, s1803; TONGA-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]804; TONGA-NEXT:    s_endpgm805;806; EG-LABEL: srem_v2i32_4:807; EG:       ; %bb.0:808; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]809; EG-NEXT:    TEX 0 @6810; EG-NEXT:    ALU 16, @9, KC0[CB0:0-32], KC1[]811; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1812; EG-NEXT:    CF_END813; EG-NEXT:    PAD814; EG-NEXT:    Fetch clause starting at 6:815; EG-NEXT:     VTX_READ_64 T0.XY, T0.X, 0, #1816; EG-NEXT:    ALU clause starting at 8:817; EG-NEXT:     MOV * T0.X, KC0[2].Z,818; EG-NEXT:    ALU clause starting at 9:819; EG-NEXT:     ASHR T0.W, T0.Y, literal.x,820; EG-NEXT:     ASHR * T1.W, T0.X, literal.x,821; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)822; EG-NEXT:     LSHR * T0.W, PV.W, literal.x,823; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)824; EG-NEXT:     ADD_INT T0.W, T0.Y, PV.W,825; EG-NEXT:     LSHR * T1.W, T1.W, literal.x,826; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)827; EG-NEXT:     ADD_INT T1.W, T0.X, PS,828; EG-NEXT:     AND_INT * T0.W, PV.W, literal.x,829; EG-NEXT:    -4(nan), 0(0.000000e+00)830; EG-NEXT:     SUB_INT T0.Y, T0.Y, PS,831; EG-NEXT:     AND_INT * T0.W, PV.W, literal.x,832; EG-NEXT:    -4(nan), 0(0.000000e+00)833; EG-NEXT:     SUB_INT T0.X, T0.X, PV.W,834; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,835; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)836  %num = load <2 x i32>, ptr addrspace(1) %in837  %result = srem <2 x i32> %num, <i32 4, i32 4>838  store <2 x i32> %result, ptr addrspace(1) %out839  ret void840}841 842define amdgpu_kernel void @srem_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {843; GCN-LABEL: srem_v4i32:844; GCN:       ; %bb.0:845; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24846; GCN-NEXT:    v_mov_b32_e32 v0, 0847; GCN-NEXT:    s_waitcnt lgkmcnt(0)848; GCN-NEXT:    global_load_dwordx4 v[1:4], v0, s[2:3] offset:16849; GCN-NEXT:    global_load_dwordx4 v[5:8], v0, s[2:3]850; GCN-NEXT:    s_waitcnt vmcnt(1)851; GCN-NEXT:    v_readfirstlane_b32 s2, v1852; GCN-NEXT:    s_abs_i32 s2, s2853; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s2854; GCN-NEXT:    s_sub_i32 s6, 0, s2855; GCN-NEXT:    s_waitcnt vmcnt(0)856; GCN-NEXT:    v_readfirstlane_b32 s4, v5857; GCN-NEXT:    s_ashr_i32 s5, s4, 31858; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1859; GCN-NEXT:    s_abs_i32 s4, s4860; GCN-NEXT:    v_readfirstlane_b32 s3, v2861; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1862; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1863; GCN-NEXT:    v_readfirstlane_b32 s7, v1864; GCN-NEXT:    s_mul_i32 s6, s6, s7865; GCN-NEXT:    s_mul_hi_u32 s6, s7, s6866; GCN-NEXT:    s_add_i32 s7, s7, s6867; GCN-NEXT:    s_mul_hi_u32 s6, s4, s7868; GCN-NEXT:    s_mul_i32 s6, s6, s2869; GCN-NEXT:    s_sub_i32 s4, s4, s6870; GCN-NEXT:    s_sub_i32 s6, s4, s2871; GCN-NEXT:    s_cmp_ge_u32 s4, s2872; GCN-NEXT:    s_cselect_b32 s4, s6, s4873; GCN-NEXT:    s_sub_i32 s6, s4, s2874; GCN-NEXT:    s_cmp_ge_u32 s4, s2875; GCN-NEXT:    s_cselect_b32 s2, s6, s4876; GCN-NEXT:    s_abs_i32 s3, s3877; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s3878; GCN-NEXT:    s_xor_b32 s2, s2, s5879; GCN-NEXT:    s_sub_i32 s8, 0, s3880; GCN-NEXT:    s_sub_i32 s2, s2, s5881; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1882; GCN-NEXT:    v_readfirstlane_b32 s6, v6883; GCN-NEXT:    s_ashr_i32 s7, s6, 31884; GCN-NEXT:    s_abs_i32 s6, s6885; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1886; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1887; GCN-NEXT:    v_readfirstlane_b32 s4, v3888; GCN-NEXT:    v_readfirstlane_b32 s5, v1889; GCN-NEXT:    s_mul_i32 s8, s8, s5890; GCN-NEXT:    s_mul_hi_u32 s8, s5, s8891; GCN-NEXT:    s_add_i32 s5, s5, s8892; GCN-NEXT:    s_mul_hi_u32 s5, s6, s5893; GCN-NEXT:    s_mul_i32 s5, s5, s3894; GCN-NEXT:    s_sub_i32 s5, s6, s5895; GCN-NEXT:    s_sub_i32 s6, s5, s3896; GCN-NEXT:    s_cmp_ge_u32 s5, s3897; GCN-NEXT:    s_cselect_b32 s5, s6, s5898; GCN-NEXT:    s_sub_i32 s6, s5, s3899; GCN-NEXT:    s_cmp_ge_u32 s5, s3900; GCN-NEXT:    s_cselect_b32 s3, s6, s5901; GCN-NEXT:    s_abs_i32 s4, s4902; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s4903; GCN-NEXT:    s_xor_b32 s3, s3, s7904; GCN-NEXT:    s_sub_i32 s9, 0, s4905; GCN-NEXT:    s_sub_i32 s3, s3, s7906; GCN-NEXT:    v_rcp_iflag_f32_e32 v1, v1907; GCN-NEXT:    v_readfirstlane_b32 s6, v7908; GCN-NEXT:    s_ashr_i32 s8, s6, 31909; GCN-NEXT:    s_abs_i32 s6, s6910; GCN-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v1911; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v1912; GCN-NEXT:    v_readfirstlane_b32 s5, v4913; GCN-NEXT:    v_readfirstlane_b32 s7, v1914; GCN-NEXT:    s_mul_i32 s9, s9, s7915; GCN-NEXT:    s_mul_hi_u32 s9, s7, s9916; GCN-NEXT:    s_add_i32 s7, s7, s9917; GCN-NEXT:    s_mul_hi_u32 s7, s6, s7918; GCN-NEXT:    s_mul_i32 s7, s7, s4919; GCN-NEXT:    s_sub_i32 s6, s6, s7920; GCN-NEXT:    s_sub_i32 s7, s6, s4921; GCN-NEXT:    s_cmp_ge_u32 s6, s4922; GCN-NEXT:    s_cselect_b32 s6, s7, s6923; GCN-NEXT:    s_sub_i32 s7, s6, s4924; GCN-NEXT:    s_cmp_ge_u32 s6, s4925; GCN-NEXT:    s_cselect_b32 s4, s7, s6926; GCN-NEXT:    s_abs_i32 s5, s5927; GCN-NEXT:    v_cvt_f32_u32_e32 v2, s5928; GCN-NEXT:    v_readfirstlane_b32 s6, v8929; GCN-NEXT:    v_mov_b32_e32 v1, s2930; GCN-NEXT:    s_ashr_i32 s2, s6, 31931; GCN-NEXT:    v_rcp_iflag_f32_e32 v3, v2932; GCN-NEXT:    v_mov_b32_e32 v2, s3933; GCN-NEXT:    s_abs_i32 s3, s6934; GCN-NEXT:    s_sub_i32 s6, 0, s5935; GCN-NEXT:    v_mul_f32_e32 v3, 0x4f7ffffe, v3936; GCN-NEXT:    v_cvt_u32_f32_e32 v3, v3937; GCN-NEXT:    s_xor_b32 s4, s4, s8938; GCN-NEXT:    s_sub_i32 s4, s4, s8939; GCN-NEXT:    v_readfirstlane_b32 s7, v3940; GCN-NEXT:    s_mul_i32 s6, s6, s7941; GCN-NEXT:    s_mul_hi_u32 s6, s7, s6942; GCN-NEXT:    s_add_i32 s7, s7, s6943; GCN-NEXT:    s_mul_hi_u32 s6, s3, s7944; GCN-NEXT:    s_mul_i32 s6, s6, s5945; GCN-NEXT:    s_sub_i32 s3, s3, s6946; GCN-NEXT:    s_sub_i32 s6, s3, s5947; GCN-NEXT:    s_cmp_ge_u32 s3, s5948; GCN-NEXT:    s_cselect_b32 s3, s6, s3949; GCN-NEXT:    s_sub_i32 s6, s3, s5950; GCN-NEXT:    s_cmp_ge_u32 s3, s5951; GCN-NEXT:    s_cselect_b32 s3, s6, s3952; GCN-NEXT:    s_xor_b32 s3, s3, s2953; GCN-NEXT:    s_sub_i32 s2, s3, s2954; GCN-NEXT:    v_mov_b32_e32 v3, s4955; GCN-NEXT:    v_mov_b32_e32 v4, s2956; GCN-NEXT:    global_store_dwordx4 v0, v[1:4], s[0:1]957; GCN-NEXT:    s_endpgm958;959; TAHITI-LABEL: srem_v4i32:960; TAHITI:       ; %bb.0:961; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x9962; TAHITI-NEXT:    s_mov_b32 s3, 0xf000963; TAHITI-NEXT:    s_mov_b32 s2, -1964; TAHITI-NEXT:    s_mov_b32 s10, s2965; TAHITI-NEXT:    s_mov_b32 s11, s3966; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)967; TAHITI-NEXT:    s_mov_b32 s8, s6968; TAHITI-NEXT:    s_mov_b32 s9, s7969; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 0 offset:16970; TAHITI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0971; TAHITI-NEXT:    s_waitcnt vmcnt(1)972; TAHITI-NEXT:    v_readfirstlane_b32 s0, v0973; TAHITI-NEXT:    s_abs_i32 s0, s0974; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, s0975; TAHITI-NEXT:    s_sub_i32 s1, 0, s0976; TAHITI-NEXT:    v_readfirstlane_b32 s7, v1977; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v0978; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v0979; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v0980; TAHITI-NEXT:    v_mul_lo_u32 v8, s1, v0981; TAHITI-NEXT:    s_waitcnt vmcnt(0)982; TAHITI-NEXT:    v_readfirstlane_b32 s1, v4983; TAHITI-NEXT:    s_abs_i32 s6, s1984; TAHITI-NEXT:    s_ashr_i32 s1, s1, 31985; TAHITI-NEXT:    v_mul_hi_u32 v8, v0, v8986; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v8987; TAHITI-NEXT:    v_mul_hi_u32 v0, s6, v0988; TAHITI-NEXT:    v_readfirstlane_b32 s8, v0989; TAHITI-NEXT:    s_mul_i32 s8, s8, s0990; TAHITI-NEXT:    s_sub_i32 s6, s6, s8991; TAHITI-NEXT:    s_sub_i32 s8, s6, s0992; TAHITI-NEXT:    s_cmp_ge_u32 s6, s0993; TAHITI-NEXT:    s_cselect_b32 s6, s8, s6994; TAHITI-NEXT:    s_sub_i32 s8, s6, s0995; TAHITI-NEXT:    s_cmp_ge_u32 s6, s0996; TAHITI-NEXT:    s_cselect_b32 s0, s8, s6997; TAHITI-NEXT:    s_abs_i32 s6, s7998; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, s6999; TAHITI-NEXT:    s_sub_i32 s7, 0, s61000; TAHITI-NEXT:    v_readfirstlane_b32 s8, v51001; TAHITI-NEXT:    s_abs_i32 s9, s81002; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v01003; TAHITI-NEXT:    s_xor_b32 s0, s0, s11004; TAHITI-NEXT:    s_sub_i32 s10, s0, s11005; TAHITI-NEXT:    s_ashr_i32 s8, s8, 311006; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01007; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v01008; TAHITI-NEXT:    v_mul_lo_u32 v1, s7, v01009; TAHITI-NEXT:    v_readfirstlane_b32 s7, v21010; TAHITI-NEXT:    v_mul_hi_u32 v1, v0, v11011; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v11012; TAHITI-NEXT:    v_mul_hi_u32 v0, s9, v01013; TAHITI-NEXT:    v_readfirstlane_b32 s0, v01014; TAHITI-NEXT:    s_mul_i32 s0, s0, s61015; TAHITI-NEXT:    s_sub_i32 s0, s9, s01016; TAHITI-NEXT:    s_sub_i32 s1, s0, s61017; TAHITI-NEXT:    s_cmp_ge_u32 s0, s61018; TAHITI-NEXT:    s_cselect_b32 s0, s1, s01019; TAHITI-NEXT:    s_sub_i32 s1, s0, s61020; TAHITI-NEXT:    s_cmp_ge_u32 s0, s61021; TAHITI-NEXT:    s_cselect_b32 s0, s1, s01022; TAHITI-NEXT:    s_abs_i32 s1, s71023; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, s11024; TAHITI-NEXT:    s_sub_i32 s6, 0, s11025; TAHITI-NEXT:    v_readfirstlane_b32 s7, v61026; TAHITI-NEXT:    s_abs_i32 s9, s71027; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v01028; TAHITI-NEXT:    s_xor_b32 s0, s0, s81029; TAHITI-NEXT:    s_sub_i32 s8, s0, s81030; TAHITI-NEXT:    s_ashr_i32 s7, s7, 311031; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01032; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v01033; TAHITI-NEXT:    v_mul_lo_u32 v1, s6, v01034; TAHITI-NEXT:    v_readfirstlane_b32 s6, v31035; TAHITI-NEXT:    v_mul_hi_u32 v1, v0, v11036; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v11037; TAHITI-NEXT:    v_mul_hi_u32 v0, s9, v01038; TAHITI-NEXT:    v_readfirstlane_b32 s0, v01039; TAHITI-NEXT:    s_mul_i32 s0, s0, s11040; TAHITI-NEXT:    s_sub_i32 s0, s9, s01041; TAHITI-NEXT:    s_sub_i32 s9, s0, s11042; TAHITI-NEXT:    s_cmp_ge_u32 s0, s11043; TAHITI-NEXT:    s_cselect_b32 s0, s9, s01044; TAHITI-NEXT:    s_sub_i32 s9, s0, s11045; TAHITI-NEXT:    s_cmp_ge_u32 s0, s11046; TAHITI-NEXT:    s_cselect_b32 s9, s9, s01047; TAHITI-NEXT:    s_abs_i32 s6, s61048; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, s61049; TAHITI-NEXT:    s_sub_i32 s1, 0, s61050; TAHITI-NEXT:    s_mov_b32 s0, s41051; TAHITI-NEXT:    v_readfirstlane_b32 s4, v71052; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v01053; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01054; TAHITI-NEXT:    v_cvt_u32_f32_e32 v2, v01055; TAHITI-NEXT:    v_mov_b32_e32 v0, s101056; TAHITI-NEXT:    v_mul_lo_u32 v1, s1, v21057; TAHITI-NEXT:    s_mov_b32 s1, s51058; TAHITI-NEXT:    s_abs_i32 s5, s41059; TAHITI-NEXT:    s_ashr_i32 s4, s4, 311060; TAHITI-NEXT:    v_mul_hi_u32 v3, v2, v11061; TAHITI-NEXT:    v_mov_b32_e32 v1, s81062; TAHITI-NEXT:    s_xor_b32 s8, s9, s71063; TAHITI-NEXT:    s_sub_i32 s7, s8, s71064; TAHITI-NEXT:    v_add_i32_e32 v2, vcc, v2, v31065; TAHITI-NEXT:    v_mul_hi_u32 v2, s5, v21066; TAHITI-NEXT:    v_readfirstlane_b32 s8, v21067; TAHITI-NEXT:    s_mul_i32 s8, s8, s61068; TAHITI-NEXT:    s_sub_i32 s5, s5, s81069; TAHITI-NEXT:    s_sub_i32 s8, s5, s61070; TAHITI-NEXT:    s_cmp_ge_u32 s5, s61071; TAHITI-NEXT:    s_cselect_b32 s5, s8, s51072; TAHITI-NEXT:    s_sub_i32 s8, s5, s61073; TAHITI-NEXT:    s_cmp_ge_u32 s5, s61074; TAHITI-NEXT:    s_cselect_b32 s5, s8, s51075; TAHITI-NEXT:    s_xor_b32 s5, s5, s41076; TAHITI-NEXT:    s_sub_i32 s4, s5, s41077; TAHITI-NEXT:    v_mov_b32_e32 v2, s71078; TAHITI-NEXT:    v_mov_b32_e32 v3, s41079; TAHITI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 01080; TAHITI-NEXT:    s_endpgm1081;1082; TONGA-LABEL: srem_v4i32:1083; TONGA:       ; %bb.0:1084; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x241085; TONGA-NEXT:    s_waitcnt lgkmcnt(0)1086; TONGA-NEXT:    s_add_u32 s4, s2, 161087; TONGA-NEXT:    s_addc_u32 s5, s3, 01088; TONGA-NEXT:    v_mov_b32_e32 v0, s41089; TONGA-NEXT:    v_mov_b32_e32 v1, s51090; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]1091; TONGA-NEXT:    v_mov_b32_e32 v5, s31092; TONGA-NEXT:    v_mov_b32_e32 v4, s21093; TONGA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]1094; TONGA-NEXT:    s_waitcnt vmcnt(1)1095; TONGA-NEXT:    v_readfirstlane_b32 s2, v01096; TONGA-NEXT:    s_abs_i32 s2, s21097; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s21098; TONGA-NEXT:    s_sub_i32 s3, 0, s21099; TONGA-NEXT:    v_readfirstlane_b32 s5, v11100; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v01101; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01102; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v01103; TONGA-NEXT:    v_mul_lo_u32 v8, s3, v01104; TONGA-NEXT:    s_waitcnt vmcnt(0)1105; TONGA-NEXT:    v_readfirstlane_b32 s3, v41106; TONGA-NEXT:    s_abs_i32 s4, s31107; TONGA-NEXT:    s_ashr_i32 s3, s3, 311108; TONGA-NEXT:    v_mul_hi_u32 v8, v0, v81109; TONGA-NEXT:    v_mov_b32_e32 v4, s01110; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v81111; TONGA-NEXT:    v_mul_hi_u32 v0, s4, v01112; TONGA-NEXT:    v_readfirstlane_b32 s6, v01113; TONGA-NEXT:    s_mul_i32 s6, s6, s21114; TONGA-NEXT:    s_sub_i32 s4, s4, s61115; TONGA-NEXT:    s_sub_i32 s6, s4, s21116; TONGA-NEXT:    s_cmp_ge_u32 s4, s21117; TONGA-NEXT:    s_cselect_b32 s4, s6, s41118; TONGA-NEXT:    s_sub_i32 s6, s4, s21119; TONGA-NEXT:    s_cmp_ge_u32 s4, s21120; TONGA-NEXT:    s_cselect_b32 s2, s6, s41121; TONGA-NEXT:    s_abs_i32 s4, s51122; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s41123; TONGA-NEXT:    s_sub_i32 s5, 0, s41124; TONGA-NEXT:    v_readfirstlane_b32 s6, v51125; TONGA-NEXT:    s_abs_i32 s7, s61126; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v01127; TONGA-NEXT:    s_xor_b32 s2, s2, s31128; TONGA-NEXT:    s_sub_i32 s2, s2, s31129; TONGA-NEXT:    s_ashr_i32 s6, s6, 311130; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01131; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v01132; TONGA-NEXT:    v_mov_b32_e32 v5, s11133; TONGA-NEXT:    v_mul_lo_u32 v1, s5, v01134; TONGA-NEXT:    v_readfirstlane_b32 s5, v21135; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v11136; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v11137; TONGA-NEXT:    v_mul_hi_u32 v0, s7, v01138; TONGA-NEXT:    v_readfirstlane_b32 s3, v01139; TONGA-NEXT:    s_mul_i32 s3, s3, s41140; TONGA-NEXT:    s_sub_i32 s3, s7, s31141; TONGA-NEXT:    s_sub_i32 s7, s3, s41142; TONGA-NEXT:    s_cmp_ge_u32 s3, s41143; TONGA-NEXT:    s_cselect_b32 s3, s7, s31144; TONGA-NEXT:    s_sub_i32 s7, s3, s41145; TONGA-NEXT:    s_cmp_ge_u32 s3, s41146; TONGA-NEXT:    s_cselect_b32 s3, s7, s31147; TONGA-NEXT:    s_abs_i32 s4, s51148; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s41149; TONGA-NEXT:    s_sub_i32 s5, 0, s41150; TONGA-NEXT:    v_readfirstlane_b32 s7, v61151; TONGA-NEXT:    s_abs_i32 s8, s71152; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v01153; TONGA-NEXT:    s_xor_b32 s3, s3, s61154; TONGA-NEXT:    s_sub_i32 s3, s3, s61155; TONGA-NEXT:    s_ashr_i32 s7, s7, 311156; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01157; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v01158; TONGA-NEXT:    v_mul_lo_u32 v1, s5, v01159; TONGA-NEXT:    v_readfirstlane_b32 s5, v31160; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v11161; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v11162; TONGA-NEXT:    v_mul_hi_u32 v0, s8, v01163; TONGA-NEXT:    v_readfirstlane_b32 s6, v01164; TONGA-NEXT:    s_mul_i32 s6, s6, s41165; TONGA-NEXT:    s_sub_i32 s6, s8, s61166; TONGA-NEXT:    s_sub_i32 s8, s6, s41167; TONGA-NEXT:    s_cmp_ge_u32 s6, s41168; TONGA-NEXT:    s_cselect_b32 s6, s8, s61169; TONGA-NEXT:    s_sub_i32 s8, s6, s41170; TONGA-NEXT:    s_cmp_ge_u32 s6, s41171; TONGA-NEXT:    s_cselect_b32 s4, s8, s61172; TONGA-NEXT:    s_abs_i32 s5, s51173; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s51174; TONGA-NEXT:    s_sub_i32 s0, 0, s51175; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v01176; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01177; TONGA-NEXT:    v_cvt_u32_f32_e32 v2, v01178; TONGA-NEXT:    v_mov_b32_e32 v0, s21179; TONGA-NEXT:    s_xor_b32 s2, s4, s71180; TONGA-NEXT:    s_sub_i32 s2, s2, s71181; TONGA-NEXT:    v_mul_lo_u32 v1, s0, v21182; TONGA-NEXT:    v_readfirstlane_b32 s0, v71183; TONGA-NEXT:    s_abs_i32 s1, s01184; TONGA-NEXT:    s_ashr_i32 s0, s0, 311185; TONGA-NEXT:    v_mul_hi_u32 v3, v2, v11186; TONGA-NEXT:    v_mov_b32_e32 v1, s31187; TONGA-NEXT:    v_add_u32_e32 v2, vcc, v2, v31188; TONGA-NEXT:    v_mul_hi_u32 v2, s1, v21189; TONGA-NEXT:    v_readfirstlane_b32 s3, v21190; TONGA-NEXT:    s_mul_i32 s3, s3, s51191; TONGA-NEXT:    s_sub_i32 s1, s1, s31192; TONGA-NEXT:    s_sub_i32 s3, s1, s51193; TONGA-NEXT:    s_cmp_ge_u32 s1, s51194; TONGA-NEXT:    s_cselect_b32 s1, s3, s11195; TONGA-NEXT:    s_sub_i32 s3, s1, s51196; TONGA-NEXT:    s_cmp_ge_u32 s1, s51197; TONGA-NEXT:    s_cselect_b32 s1, s3, s11198; TONGA-NEXT:    s_xor_b32 s1, s1, s01199; TONGA-NEXT:    s_sub_i32 s0, s1, s01200; TONGA-NEXT:    v_mov_b32_e32 v2, s21201; TONGA-NEXT:    v_mov_b32_e32 v3, s01202; TONGA-NEXT:    flat_store_dwordx4 v[4:5], v[0:3]1203; TONGA-NEXT:    s_endpgm1204;1205; EG-LABEL: srem_v4i32:1206; EG:       ; %bb.0:1207; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]1208; EG-NEXT:    TEX 1 @61209; EG-NEXT:    ALU 89, @11, KC0[CB0:0-32], KC1[]1210; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T2.XYZW, T0.X, 11211; EG-NEXT:    CF_END1212; EG-NEXT:    PAD1213; EG-NEXT:    Fetch clause starting at 6:1214; EG-NEXT:     VTX_READ_128 T1.XYZW, T0.X, 16, #11215; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #11216; EG-NEXT:    ALU clause starting at 10:1217; EG-NEXT:     MOV * T0.X, KC0[2].Z,1218; EG-NEXT:    ALU clause starting at 11:1219; EG-NEXT:     SETGT_INT * T2.W, 0.0, T1.Z,1220; EG-NEXT:     ADD_INT T3.W, T1.Z, PV.W,1221; EG-NEXT:     SETGT_INT * T4.W, 0.0, T1.Y,1222; EG-NEXT:     XOR_INT * T2.W, PV.W, T2.W,1223; EG-NEXT:     SUB_INT T1.Z, 0.0, PV.W,1224; EG-NEXT:     ADD_INT T3.W, T1.Y, T4.W,1225; EG-NEXT:     RECIP_UINT * T1.Y, PV.W,1226; EG-NEXT:     XOR_INT T3.W, PV.W, T4.W,1227; EG-NEXT:     MULLO_INT * T1.Z, PV.Z, PS,1228; EG-NEXT:     SUB_INT T4.W, 0.0, PV.W,1229; EG-NEXT:     RECIP_UINT * T2.X, PV.W,1230; EG-NEXT:     SETGT_INT T5.W, 0.0, T0.Y,1231; EG-NEXT:     MULLO_INT * T2.Y, PV.W, PS,1232; EG-NEXT:     SETGT_INT T2.Z, 0.0, T0.Z,1233; EG-NEXT:     ADD_INT T4.W, T0.Y, PV.W,1234; EG-NEXT:     MULHI * T0.Y, T2.X, PS,1235; EG-NEXT:     ADD_INT T2.X, T2.X, PS,1236; EG-NEXT:     XOR_INT T0.Y, PV.W, T5.W,1237; EG-NEXT:     SETGT_INT T3.Z, 0.0, T1.W, BS:VEC_021/SCL_1221238; EG-NEXT:     ADD_INT T4.W, T0.Z, PV.Z,1239; EG-NEXT:     MULHI * T0.Z, T1.Y, T1.Z,1240; EG-NEXT:     ADD_INT T1.Y, T1.Y, PS,1241; EG-NEXT:     XOR_INT T0.Z, PV.W, T2.Z,1242; EG-NEXT:     ADD_INT T1.W, T1.W, PV.Z,1243; EG-NEXT:     MULHI * T1.Z, PV.Y, PV.X,1244; EG-NEXT:     XOR_INT T1.W, PV.W, T3.Z,1245; EG-NEXT:     MULHI * T1.Y, PV.Z, PV.Y,1246; EG-NEXT:     SUB_INT T4.W, 0.0, PV.W,1247; EG-NEXT:     RECIP_UINT * T2.X, PV.W,1248; EG-NEXT:     SETGT_INT T6.W, 0.0, T0.W,1249; EG-NEXT:     MULLO_INT * T2.Y, PV.W, PS,1250; EG-NEXT:     ADD_INT T0.W, T0.W, PV.W,1251; EG-NEXT:     MULHI * T2.Y, T2.X, PS,1252; EG-NEXT:     ADD_INT T2.Y, T2.X, PS,1253; EG-NEXT:     XOR_INT T3.Z, PV.W, T6.W, BS:VEC_021/SCL_1221254; EG-NEXT:     SETGT_INT T0.W, 0.0, T1.X,1255; EG-NEXT:     MULLO_INT * T1.Y, T1.Y, T2.W,1256; EG-NEXT:     ADD_INT T4.W, T1.X, PV.W,1257; EG-NEXT:     MULHI * T1.X, PV.Z, PV.Y,1258; EG-NEXT:     XOR_INT T0.W, PV.W, T0.W, BS:VEC_021/SCL_1221259; EG-NEXT:     MULLO_INT * T1.X, PS, T1.W,1260; EG-NEXT:     SUB_INT T4.W, 0.0, PV.W,1261; EG-NEXT:     RECIP_UINT * T2.X, PV.W,1262; EG-NEXT:     SETGT_INT T7.W, 0.0, T0.X,1263; EG-NEXT:     MULLO_INT * T2.Y, PV.W, PS,1264; EG-NEXT:     ADD_INT T4.W, T0.X, PV.W,1265; EG-NEXT:     MULHI * T0.X, T2.X, PS,1266; EG-NEXT:     ADD_INT T0.X, T2.X, PS,1267; EG-NEXT:     XOR_INT T2.Y, PV.W, T7.W, BS:VEC_021/SCL_1221268; EG-NEXT:     SUB_INT T3.Z, T3.Z, T1.X,1269; EG-NEXT:     SUB_INT T4.W, T0.Z, T1.Y, BS:VEC_120/SCL_2121270; EG-NEXT:     MULLO_INT * T0.Z, T1.Z, T3.W,1271; EG-NEXT:     SETGE_UINT T1.X, PV.W, T2.W,1272; EG-NEXT:     SUB_INT T0.Y, T0.Y, PS,1273; EG-NEXT:     SETGE_UINT T0.Z, PV.Z, T1.W, BS:VEC_021/SCL_1221274; EG-NEXT:     SUB_INT T8.W, PV.Z, T1.W, BS:VEC_021/SCL_1221275; EG-NEXT:     MULHI * T0.X, PV.Y, PV.X,1276; EG-NEXT:     SUB_INT T2.X, T4.W, T2.W,1277; EG-NEXT:     CNDE_INT T1.Y, PV.Z, T3.Z, PV.W,1278; EG-NEXT:     SETGE_UINT T0.Z, PV.Y, T3.W, BS:VEC_021/SCL_1221279; EG-NEXT:     SUB_INT * T8.W, PV.Y, T3.W, BS:VEC_021/SCL_1221280; EG-NEXT:     MULLO_INT * T0.X, T0.X, T0.W,1281; EG-NEXT:     CNDE_INT T3.X, T0.Z, T0.Y, T8.W,1282; EG-NEXT:     SETGE_UINT T0.Y, T1.Y, T1.W,1283; EG-NEXT:     SUB_INT T0.Z, T1.Y, T1.W,1284; EG-NEXT:     CNDE_INT T1.W, T1.X, T4.W, T2.X, BS:VEC_102/SCL_2211285; EG-NEXT:     SUB_INT * T4.W, T2.Y, PS,1286; EG-NEXT:     SETGE_UINT T0.X, PS, T0.W,1287; EG-NEXT:     SUB_INT T2.Y, PS, T0.W,1288; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T2.W, BS:VEC_021/SCL_1221289; EG-NEXT:     SUB_INT T2.W, PV.W, T2.W, BS:VEC_021/SCL_1221290; EG-NEXT:     CNDE_INT * T8.W, PV.Y, T1.Y, PV.Z,1291; EG-NEXT:     XOR_INT T1.X, PS, T6.W,1292; EG-NEXT:     CNDE_INT T0.Y, PV.Z, T1.W, PV.W, BS:VEC_021/SCL_1221293; EG-NEXT:     CNDE_INT * T0.Z, PV.X, T4.W, PV.Y, BS:VEC_102/SCL_2211294; EG-NEXT:     SETGE_UINT T1.W, T3.X, T3.W,1295; EG-NEXT:     SUB_INT * T2.W, T3.X, T3.W,1296; EG-NEXT:     CNDE_INT T0.X, PV.W, T3.X, PS,1297; EG-NEXT:     SETGE_UINT T1.Y, T0.Z, T0.W, BS:VEC_021/SCL_1221298; EG-NEXT:     SUB_INT T1.Z, T0.Z, T0.W, BS:VEC_021/SCL_1221299; EG-NEXT:     XOR_INT T0.W, T0.Y, T2.Z,1300; EG-NEXT:     SUB_INT * T2.W, T1.X, T6.W,1301; EG-NEXT:     SUB_INT T2.Z, PV.W, T2.Z,1302; EG-NEXT:     CNDE_INT T0.W, PV.Y, T0.Z, PV.Z, BS:VEC_021/SCL_1221303; EG-NEXT:     XOR_INT * T1.W, PV.X, T5.W,1304; EG-NEXT:     SUB_INT T2.Y, PS, T5.W,1305; EG-NEXT:     XOR_INT * T0.W, PV.W, T7.W, BS:VEC_021/SCL_1221306; EG-NEXT:     SUB_INT T2.X, PV.W, T7.W,1307; EG-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,1308; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)1309  %den_ptr = getelementptr <4 x i32>, ptr addrspace(1) %in, i32 11310  %num = load <4 x i32>, ptr addrspace(1) %in1311  %den = load <4 x i32>, ptr addrspace(1) %den_ptr1312  %result = srem <4 x i32> %num, %den1313  store <4 x i32> %result, ptr addrspace(1) %out1314  ret void1315}1316 1317define amdgpu_kernel void @srem_v4i32_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {1318; GCN-LABEL: srem_v4i32_4:1319; GCN:       ; %bb.0:1320; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x241321; GCN-NEXT:    v_mov_b32_e32 v4, 01322; GCN-NEXT:    s_waitcnt lgkmcnt(0)1323; GCN-NEXT:    global_load_dwordx4 v[0:3], v4, s[2:3]1324; GCN-NEXT:    s_waitcnt vmcnt(0)1325; GCN-NEXT:    v_readfirstlane_b32 s2, v01326; GCN-NEXT:    v_readfirstlane_b32 s3, v11327; GCN-NEXT:    v_readfirstlane_b32 s4, v21328; GCN-NEXT:    v_readfirstlane_b32 s5, v31329; GCN-NEXT:    s_ashr_i32 s6, s2, 311330; GCN-NEXT:    s_ashr_i32 s7, s3, 311331; GCN-NEXT:    s_ashr_i32 s8, s4, 311332; GCN-NEXT:    s_ashr_i32 s9, s5, 311333; GCN-NEXT:    s_lshr_b32 s6, s6, 301334; GCN-NEXT:    s_lshr_b32 s7, s7, 301335; GCN-NEXT:    s_lshr_b32 s8, s8, 301336; GCN-NEXT:    s_lshr_b32 s9, s9, 301337; GCN-NEXT:    s_add_i32 s6, s2, s61338; GCN-NEXT:    s_add_i32 s7, s3, s71339; GCN-NEXT:    s_add_i32 s8, s4, s81340; GCN-NEXT:    s_add_i32 s9, s5, s91341; GCN-NEXT:    s_and_b32 s6, s6, -41342; GCN-NEXT:    s_and_b32 s7, s7, -41343; GCN-NEXT:    s_and_b32 s8, s8, -41344; GCN-NEXT:    s_and_b32 s9, s9, -41345; GCN-NEXT:    s_sub_i32 s2, s2, s61346; GCN-NEXT:    s_sub_i32 s3, s3, s71347; GCN-NEXT:    s_sub_i32 s4, s4, s81348; GCN-NEXT:    s_sub_i32 s5, s5, s91349; GCN-NEXT:    v_mov_b32_e32 v0, s21350; GCN-NEXT:    v_mov_b32_e32 v1, s31351; GCN-NEXT:    v_mov_b32_e32 v2, s41352; GCN-NEXT:    v_mov_b32_e32 v3, s51353; GCN-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]1354; GCN-NEXT:    s_endpgm1355;1356; TAHITI-LABEL: srem_v4i32_4:1357; TAHITI:       ; %bb.0:1358; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x91359; TAHITI-NEXT:    s_mov_b32 s3, 0xf0001360; TAHITI-NEXT:    s_mov_b32 s2, -11361; TAHITI-NEXT:    s_mov_b32 s10, s21362; TAHITI-NEXT:    s_mov_b32 s11, s31363; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)1364; TAHITI-NEXT:    s_mov_b32 s8, s61365; TAHITI-NEXT:    s_mov_b32 s9, s71366; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 01367; TAHITI-NEXT:    s_mov_b32 s0, s41368; TAHITI-NEXT:    s_mov_b32 s1, s51369; TAHITI-NEXT:    s_waitcnt vmcnt(0)1370; TAHITI-NEXT:    v_readfirstlane_b32 s4, v01371; TAHITI-NEXT:    v_readfirstlane_b32 s5, v11372; TAHITI-NEXT:    v_readfirstlane_b32 s6, v21373; TAHITI-NEXT:    v_readfirstlane_b32 s7, v31374; TAHITI-NEXT:    s_ashr_i32 s8, s4, 311375; TAHITI-NEXT:    s_ashr_i32 s9, s5, 311376; TAHITI-NEXT:    s_ashr_i32 s10, s6, 311377; TAHITI-NEXT:    s_ashr_i32 s11, s7, 311378; TAHITI-NEXT:    s_lshr_b32 s8, s8, 301379; TAHITI-NEXT:    s_lshr_b32 s9, s9, 301380; TAHITI-NEXT:    s_lshr_b32 s10, s10, 301381; TAHITI-NEXT:    s_lshr_b32 s11, s11, 301382; TAHITI-NEXT:    s_add_i32 s8, s4, s81383; TAHITI-NEXT:    s_add_i32 s9, s5, s91384; TAHITI-NEXT:    s_add_i32 s10, s6, s101385; TAHITI-NEXT:    s_add_i32 s11, s7, s111386; TAHITI-NEXT:    s_and_b32 s8, s8, -41387; TAHITI-NEXT:    s_and_b32 s9, s9, -41388; TAHITI-NEXT:    s_and_b32 s10, s10, -41389; TAHITI-NEXT:    s_and_b32 s11, s11, -41390; TAHITI-NEXT:    s_sub_i32 s4, s4, s81391; TAHITI-NEXT:    s_sub_i32 s5, s5, s91392; TAHITI-NEXT:    s_sub_i32 s6, s6, s101393; TAHITI-NEXT:    s_sub_i32 s7, s7, s111394; TAHITI-NEXT:    v_mov_b32_e32 v0, s41395; TAHITI-NEXT:    v_mov_b32_e32 v1, s51396; TAHITI-NEXT:    v_mov_b32_e32 v2, s61397; TAHITI-NEXT:    v_mov_b32_e32 v3, s71398; TAHITI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 01399; TAHITI-NEXT:    s_endpgm1400;1401; TONGA-LABEL: srem_v4i32_4:1402; TONGA:       ; %bb.0:1403; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x241404; TONGA-NEXT:    s_waitcnt lgkmcnt(0)1405; TONGA-NEXT:    v_mov_b32_e32 v0, s21406; TONGA-NEXT:    v_mov_b32_e32 v1, s31407; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]1408; TONGA-NEXT:    v_mov_b32_e32 v4, s01409; TONGA-NEXT:    v_mov_b32_e32 v5, s11410; TONGA-NEXT:    s_waitcnt vmcnt(0)1411; TONGA-NEXT:    v_readfirstlane_b32 s0, v01412; TONGA-NEXT:    v_readfirstlane_b32 s1, v11413; TONGA-NEXT:    v_readfirstlane_b32 s2, v21414; TONGA-NEXT:    v_readfirstlane_b32 s3, v31415; TONGA-NEXT:    s_ashr_i32 s4, s0, 311416; TONGA-NEXT:    s_ashr_i32 s5, s1, 311417; TONGA-NEXT:    s_ashr_i32 s6, s2, 311418; TONGA-NEXT:    s_ashr_i32 s7, s3, 311419; TONGA-NEXT:    s_lshr_b32 s4, s4, 301420; TONGA-NEXT:    s_lshr_b32 s5, s5, 301421; TONGA-NEXT:    s_lshr_b32 s6, s6, 301422; TONGA-NEXT:    s_lshr_b32 s7, s7, 301423; TONGA-NEXT:    s_add_i32 s4, s0, s41424; TONGA-NEXT:    s_add_i32 s5, s1, s51425; TONGA-NEXT:    s_add_i32 s6, s2, s61426; TONGA-NEXT:    s_add_i32 s7, s3, s71427; TONGA-NEXT:    s_and_b32 s4, s4, -41428; TONGA-NEXT:    s_and_b32 s5, s5, -41429; TONGA-NEXT:    s_and_b32 s6, s6, -41430; TONGA-NEXT:    s_and_b32 s7, s7, -41431; TONGA-NEXT:    s_sub_i32 s0, s0, s41432; TONGA-NEXT:    s_sub_i32 s1, s1, s51433; TONGA-NEXT:    s_sub_i32 s2, s2, s61434; TONGA-NEXT:    s_sub_i32 s3, s3, s71435; TONGA-NEXT:    v_mov_b32_e32 v0, s01436; TONGA-NEXT:    v_mov_b32_e32 v1, s11437; TONGA-NEXT:    v_mov_b32_e32 v2, s21438; TONGA-NEXT:    v_mov_b32_e32 v3, s31439; TONGA-NEXT:    flat_store_dwordx4 v[4:5], v[0:3]1440; TONGA-NEXT:    s_endpgm1441;1442; EG-LABEL: srem_v4i32_4:1443; EG:       ; %bb.0:1444; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]1445; EG-NEXT:    TEX 0 @61446; EG-NEXT:    ALU 29, @9, KC0[CB0:0-32], KC1[]1447; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 11448; EG-NEXT:    CF_END1449; EG-NEXT:    PAD1450; EG-NEXT:    Fetch clause starting at 6:1451; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #11452; EG-NEXT:    ALU clause starting at 8:1453; EG-NEXT:     MOV * T0.X, KC0[2].Z,1454; EG-NEXT:    ALU clause starting at 9:1455; EG-NEXT:     ASHR * T1.W, T0.W, literal.x,1456; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)1457; EG-NEXT:     LSHR T1.W, PV.W, literal.x,1458; EG-NEXT:     ASHR * T2.W, T0.Z, literal.y,1459; EG-NEXT:    30(4.203895e-44), 31(4.344025e-44)1460; EG-NEXT:     ASHR T1.Z, T0.Y, literal.x,1461; EG-NEXT:     LSHR T2.W, PS, literal.y,1462; EG-NEXT:     ADD_INT * T1.W, T0.W, PV.W,1463; EG-NEXT:    31(4.344025e-44), 30(4.203895e-44)1464; EG-NEXT:     AND_INT T1.Y, PS, literal.x,1465; EG-NEXT:     ADD_INT T2.Z, T0.Z, PV.W,1466; EG-NEXT:     ASHR T1.W, T0.X, literal.y,1467; EG-NEXT:     LSHR * T2.W, PV.Z, literal.z,1468; EG-NEXT:    -4(nan), 31(4.344025e-44)1469; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)1470; EG-NEXT:     ADD_INT T2.Y, T0.Y, PS,1471; EG-NEXT:     LSHR T1.Z, PV.W, literal.x,1472; EG-NEXT:     AND_INT T1.W, PV.Z, literal.y,1473; EG-NEXT:     SUB_INT * T0.W, T0.W, PV.Y,1474; EG-NEXT:    30(4.203895e-44), -4(nan)1475; EG-NEXT:     SUB_INT T0.Z, T0.Z, PV.W,1476; EG-NEXT:     ADD_INT T1.W, T0.X, PV.Z,1477; EG-NEXT:     AND_INT * T2.W, PV.Y, literal.x,1478; EG-NEXT:    -4(nan), 0(0.000000e+00)1479; EG-NEXT:     SUB_INT T0.Y, T0.Y, PS,1480; EG-NEXT:     AND_INT * T1.W, PV.W, literal.x,1481; EG-NEXT:    -4(nan), 0(0.000000e+00)1482; EG-NEXT:     SUB_INT T0.X, T0.X, PV.W,1483; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,1484; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)1485  %num = load <4 x i32>, ptr addrspace(1) %in1486  %result = srem <4 x i32> %num, <i32 4, i32 4, i32 4, i32 4>1487  store <4 x i32> %result, ptr addrspace(1) %out1488  ret void1489}1490 1491define amdgpu_kernel void @srem_i64(ptr addrspace(1) %out, ptr addrspace(1) %in) {1492; GCN-LABEL: srem_i64:1493; GCN:       ; %bb.0:1494; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x241495; GCN-NEXT:    v_mov_b32_e32 v0, 01496; GCN-NEXT:    s_waitcnt lgkmcnt(0)1497; GCN-NEXT:    global_load_dwordx4 v[0:3], v0, s[2:3]1498; GCN-NEXT:    s_waitcnt vmcnt(0)1499; GCN-NEXT:    v_readfirstlane_b32 s5, v11500; GCN-NEXT:    v_readfirstlane_b32 s4, v01501; GCN-NEXT:    v_readfirstlane_b32 s3, v31502; GCN-NEXT:    v_readfirstlane_b32 s2, v21503; GCN-NEXT:    s_or_b64 s[6:7], s[4:5], s[2:3]1504; GCN-NEXT:    s_mov_b32 s6, 01505; GCN-NEXT:    s_cmp_lg_u64 s[6:7], 01506; GCN-NEXT:    s_cbranch_scc0 .LBB8_41507; GCN-NEXT:  ; %bb.1:1508; GCN-NEXT:    s_ashr_i32 s6, s3, 311509; GCN-NEXT:    s_add_u32 s8, s2, s61510; GCN-NEXT:    s_mov_b32 s7, s61511; GCN-NEXT:    s_addc_u32 s9, s3, s61512; GCN-NEXT:    s_xor_b64 s[8:9], s[8:9], s[6:7]1513; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s81514; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s91515; GCN-NEXT:    s_sub_u32 s3, 0, s81516; GCN-NEXT:    s_subb_u32 s10, 0, s91517; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v01518; GCN-NEXT:    v_rcp_f32_e32 v0, v01519; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v01520; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v01521; GCN-NEXT:    v_trunc_f32_e32 v1, v11522; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v01523; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v11524; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v01525; GCN-NEXT:    v_readfirstlane_b32 s11, v11526; GCN-NEXT:    v_readfirstlane_b32 s12, v01527; GCN-NEXT:    s_mul_i32 s13, s3, s111528; GCN-NEXT:    s_mul_hi_u32 s15, s3, s121529; GCN-NEXT:    s_mul_i32 s14, s10, s121530; GCN-NEXT:    s_add_i32 s13, s15, s131531; GCN-NEXT:    s_add_i32 s13, s13, s141532; GCN-NEXT:    s_mul_i32 s16, s3, s121533; GCN-NEXT:    s_mul_i32 s15, s12, s131534; GCN-NEXT:    s_mul_hi_u32 s17, s12, s161535; GCN-NEXT:    s_mul_hi_u32 s14, s12, s131536; GCN-NEXT:    s_add_u32 s15, s17, s151537; GCN-NEXT:    s_addc_u32 s14, 0, s141538; GCN-NEXT:    s_mul_hi_u32 s18, s11, s161539; GCN-NEXT:    s_mul_i32 s16, s11, s161540; GCN-NEXT:    s_add_u32 s15, s15, s161541; GCN-NEXT:    s_mul_hi_u32 s17, s11, s131542; GCN-NEXT:    s_addc_u32 s14, s14, s181543; GCN-NEXT:    s_addc_u32 s15, s17, 01544; GCN-NEXT:    s_mul_i32 s13, s11, s131545; GCN-NEXT:    s_add_u32 s13, s14, s131546; GCN-NEXT:    s_addc_u32 s14, 0, s151547; GCN-NEXT:    s_add_u32 s12, s12, s131548; GCN-NEXT:    s_addc_u32 s11, s11, s141549; GCN-NEXT:    s_mul_i32 s13, s3, s111550; GCN-NEXT:    s_mul_hi_u32 s14, s3, s121551; GCN-NEXT:    s_add_i32 s13, s14, s131552; GCN-NEXT:    s_mul_i32 s10, s10, s121553; GCN-NEXT:    s_add_i32 s13, s13, s101554; GCN-NEXT:    s_mul_i32 s3, s3, s121555; GCN-NEXT:    s_mul_hi_u32 s14, s11, s31556; GCN-NEXT:    s_mul_i32 s15, s11, s31557; GCN-NEXT:    s_mul_i32 s17, s12, s131558; GCN-NEXT:    s_mul_hi_u32 s3, s12, s31559; GCN-NEXT:    s_mul_hi_u32 s16, s12, s131560; GCN-NEXT:    s_add_u32 s3, s3, s171561; GCN-NEXT:    s_addc_u32 s16, 0, s161562; GCN-NEXT:    s_add_u32 s3, s3, s151563; GCN-NEXT:    s_mul_hi_u32 s10, s11, s131564; GCN-NEXT:    s_addc_u32 s3, s16, s141565; GCN-NEXT:    s_addc_u32 s10, s10, 01566; GCN-NEXT:    s_mul_i32 s13, s11, s131567; GCN-NEXT:    s_add_u32 s3, s3, s131568; GCN-NEXT:    s_addc_u32 s10, 0, s101569; GCN-NEXT:    s_add_u32 s3, s12, s31570; GCN-NEXT:    s_addc_u32 s14, s11, s101571; GCN-NEXT:    s_ashr_i32 s10, s5, 311572; GCN-NEXT:    s_add_u32 s12, s4, s101573; GCN-NEXT:    s_mov_b32 s11, s101574; GCN-NEXT:    s_addc_u32 s13, s5, s101575; GCN-NEXT:    s_xor_b64 s[12:13], s[12:13], s[10:11]1576; GCN-NEXT:    s_mul_i32 s15, s12, s141577; GCN-NEXT:    s_mul_hi_u32 s16, s12, s31578; GCN-NEXT:    s_mul_hi_u32 s5, s12, s141579; GCN-NEXT:    s_add_u32 s15, s16, s151580; GCN-NEXT:    s_addc_u32 s5, 0, s51581; GCN-NEXT:    s_mul_hi_u32 s17, s13, s31582; GCN-NEXT:    s_mul_i32 s3, s13, s31583; GCN-NEXT:    s_add_u32 s3, s15, s31584; GCN-NEXT:    s_mul_hi_u32 s16, s13, s141585; GCN-NEXT:    s_addc_u32 s3, s5, s171586; GCN-NEXT:    s_addc_u32 s5, s16, 01587; GCN-NEXT:    s_mul_i32 s14, s13, s141588; GCN-NEXT:    s_add_u32 s3, s3, s141589; GCN-NEXT:    s_addc_u32 s5, 0, s51590; GCN-NEXT:    s_mul_i32 s5, s8, s51591; GCN-NEXT:    s_mul_hi_u32 s14, s8, s31592; GCN-NEXT:    s_add_i32 s5, s14, s51593; GCN-NEXT:    s_mul_i32 s14, s9, s31594; GCN-NEXT:    s_add_i32 s5, s5, s141595; GCN-NEXT:    s_sub_i32 s16, s13, s51596; GCN-NEXT:    s_mul_i32 s3, s8, s31597; GCN-NEXT:    s_sub_u32 s3, s12, s31598; GCN-NEXT:    s_cselect_b64 s[14:15], -1, 01599; GCN-NEXT:    s_subb_u32 s12, s16, s91600; GCN-NEXT:    s_sub_u32 s18, s3, s81601; GCN-NEXT:    s_cselect_b64 s[16:17], -1, 01602; GCN-NEXT:    s_subb_u32 s19, s12, 01603; GCN-NEXT:    s_cmp_ge_u32 s19, s91604; GCN-NEXT:    s_cselect_b32 s20, -1, 01605; GCN-NEXT:    s_cmp_ge_u32 s18, s81606; GCN-NEXT:    s_cselect_b32 s21, -1, 01607; GCN-NEXT:    s_cmp_eq_u32 s19, s91608; GCN-NEXT:    s_cselect_b32 s20, s21, s201609; GCN-NEXT:    s_cmp_lg_u64 s[16:17], 01610; GCN-NEXT:    s_subb_u32 s12, s12, s91611; GCN-NEXT:    s_sub_u32 s16, s18, s81612; GCN-NEXT:    s_subb_u32 s12, s12, 01613; GCN-NEXT:    s_cmp_lg_u32 s20, 01614; GCN-NEXT:    s_cselect_b32 s16, s16, s181615; GCN-NEXT:    s_cselect_b32 s12, s12, s191616; GCN-NEXT:    s_cmp_lg_u64 s[14:15], 01617; GCN-NEXT:    s_subb_u32 s5, s13, s51618; GCN-NEXT:    s_cmp_ge_u32 s5, s91619; GCN-NEXT:    s_cselect_b32 s13, -1, 01620; GCN-NEXT:    s_cmp_ge_u32 s3, s81621; GCN-NEXT:    s_cselect_b32 s8, -1, 01622; GCN-NEXT:    s_cmp_eq_u32 s5, s91623; GCN-NEXT:    s_cselect_b32 s8, s8, s131624; GCN-NEXT:    s_cmp_lg_u32 s8, 01625; GCN-NEXT:    s_cselect_b32 s9, s12, s51626; GCN-NEXT:    s_cselect_b32 s8, s16, s31627; GCN-NEXT:    s_xor_b64 s[8:9], s[8:9], s[10:11]1628; GCN-NEXT:    s_sub_u32 s8, s8, s101629; GCN-NEXT:    s_subb_u32 s9, s9, s101630; GCN-NEXT:    s_cbranch_execnz .LBB8_31631; GCN-NEXT:  .LBB8_2:1632; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s21633; GCN-NEXT:    s_sub_i32 s3, 0, s21634; GCN-NEXT:    s_mov_b32 s9, 01635; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v01636; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01637; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v01638; GCN-NEXT:    v_readfirstlane_b32 s5, v01639; GCN-NEXT:    s_mul_i32 s3, s3, s51640; GCN-NEXT:    s_mul_hi_u32 s3, s5, s31641; GCN-NEXT:    s_add_i32 s5, s5, s31642; GCN-NEXT:    s_mul_hi_u32 s3, s4, s51643; GCN-NEXT:    s_mul_i32 s3, s3, s21644; GCN-NEXT:    s_sub_i32 s3, s4, s31645; GCN-NEXT:    s_sub_i32 s4, s3, s21646; GCN-NEXT:    s_cmp_ge_u32 s3, s21647; GCN-NEXT:    s_cselect_b32 s3, s4, s31648; GCN-NEXT:    s_sub_i32 s4, s3, s21649; GCN-NEXT:    s_cmp_ge_u32 s3, s21650; GCN-NEXT:    s_cselect_b32 s8, s4, s31651; GCN-NEXT:  .LBB8_3:1652; GCN-NEXT:    v_mov_b32_e32 v0, s81653; GCN-NEXT:    v_mov_b32_e32 v2, 01654; GCN-NEXT:    v_mov_b32_e32 v1, s91655; GCN-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]1656; GCN-NEXT:    s_endpgm1657; GCN-NEXT:  .LBB8_4:1658; GCN-NEXT:    ; implicit-def: $sgpr8_sgpr91659; GCN-NEXT:    s_branch .LBB8_21660;1661; TAHITI-LABEL: srem_i64:1662; TAHITI:       ; %bb.0:1663; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x91664; TAHITI-NEXT:    s_mov_b32 s3, 0xf0001665; TAHITI-NEXT:    s_mov_b32 s2, -11666; TAHITI-NEXT:    v_mov_b32_e32 v4, 01667; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)1668; TAHITI-NEXT:    s_mov_b32 s0, s61669; TAHITI-NEXT:    s_mov_b32 s1, s71670; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 01671; TAHITI-NEXT:    s_waitcnt vmcnt(0)1672; TAHITI-NEXT:    v_or_b32_e32 v5, v1, v31673; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[4:5]1674; TAHITI-NEXT:    s_cbranch_vccz .LBB8_41675; TAHITI-NEXT:  ; %bb.1:1676; TAHITI-NEXT:    v_ashrrev_i32_e32 v5, 31, v31677; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v2, v51678; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, v3, v5, vcc1679; TAHITI-NEXT:    v_xor_b32_e32 v4, v4, v51680; TAHITI-NEXT:    v_xor_b32_e32 v3, v3, v51681; TAHITI-NEXT:    v_cvt_f32_u32_e32 v5, v41682; TAHITI-NEXT:    v_cvt_f32_u32_e32 v6, v31683; TAHITI-NEXT:    v_sub_i32_e32 v7, vcc, 0, v41684; TAHITI-NEXT:    v_subb_u32_e32 v8, vcc, 0, v3, vcc1685; TAHITI-NEXT:    v_madmk_f32 v5, v6, 0x4f800000, v51686; TAHITI-NEXT:    v_rcp_f32_e32 v5, v51687; TAHITI-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v51688; TAHITI-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v51689; TAHITI-NEXT:    v_trunc_f32_e32 v6, v61690; TAHITI-NEXT:    v_madmk_f32 v5, v6, 0xcf800000, v51691; TAHITI-NEXT:    v_cvt_u32_f32_e32 v6, v61692; TAHITI-NEXT:    v_cvt_u32_f32_e32 v5, v51693; TAHITI-NEXT:    v_mul_lo_u32 v10, v7, v61694; TAHITI-NEXT:    v_mul_hi_u32 v9, v7, v51695; TAHITI-NEXT:    v_mul_lo_u32 v11, v8, v51696; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v10, v91697; TAHITI-NEXT:    v_mul_lo_u32 v10, v7, v51698; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v111699; TAHITI-NEXT:    v_mul_lo_u32 v11, v5, v91700; TAHITI-NEXT:    v_mul_hi_u32 v12, v5, v101701; TAHITI-NEXT:    v_mul_hi_u32 v13, v5, v91702; TAHITI-NEXT:    v_mul_hi_u32 v14, v6, v91703; TAHITI-NEXT:    v_mul_lo_u32 v9, v6, v91704; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v111705; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc1706; TAHITI-NEXT:    v_mul_lo_u32 v13, v6, v101707; TAHITI-NEXT:    v_mul_hi_u32 v10, v6, v101708; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v131709; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, v12, v10, vcc1710; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v14, vcc1711; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v10, v91712; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v11, vcc1713; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v5, v91714; TAHITI-NEXT:    v_addc_u32_e32 v6, vcc, v6, v10, vcc1715; TAHITI-NEXT:    v_mul_lo_u32 v9, v7, v61716; TAHITI-NEXT:    v_mul_hi_u32 v10, v7, v51717; TAHITI-NEXT:    v_mul_lo_u32 v8, v8, v51718; TAHITI-NEXT:    v_mul_lo_u32 v7, v7, v51719; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v101720; TAHITI-NEXT:    v_add_i32_e32 v8, vcc, v8, v91721; TAHITI-NEXT:    v_mul_lo_u32 v11, v5, v81722; TAHITI-NEXT:    v_mul_hi_u32 v12, v5, v71723; TAHITI-NEXT:    v_mul_hi_u32 v13, v5, v81724; TAHITI-NEXT:    v_mul_hi_u32 v10, v6, v71725; TAHITI-NEXT:    v_mul_lo_u32 v7, v6, v71726; TAHITI-NEXT:    v_mul_hi_u32 v9, v6, v81727; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v111728; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc1729; TAHITI-NEXT:    v_mul_lo_u32 v8, v6, v81730; TAHITI-NEXT:    v_add_i32_e32 v7, vcc, v11, v71731; TAHITI-NEXT:    v_addc_u32_e32 v7, vcc, v12, v10, vcc1732; TAHITI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc1733; TAHITI-NEXT:    v_add_i32_e32 v7, vcc, v7, v81734; TAHITI-NEXT:    v_addc_u32_e32 v8, vcc, 0, v9, vcc1735; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v5, v71736; TAHITI-NEXT:    v_addc_u32_e32 v6, vcc, v6, v8, vcc1737; TAHITI-NEXT:    v_ashrrev_i32_e32 v7, 31, v11738; TAHITI-NEXT:    v_add_i32_e32 v8, vcc, v0, v71739; TAHITI-NEXT:    v_xor_b32_e32 v8, v8, v71740; TAHITI-NEXT:    v_mul_lo_u32 v9, v8, v61741; TAHITI-NEXT:    v_mul_hi_u32 v10, v8, v51742; TAHITI-NEXT:    v_mul_hi_u32 v11, v8, v61743; TAHITI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v7, vcc1744; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v71745; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v10, v91746; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v11, vcc1747; TAHITI-NEXT:    v_mul_lo_u32 v11, v1, v51748; TAHITI-NEXT:    v_mul_hi_u32 v5, v1, v51749; TAHITI-NEXT:    v_mul_hi_u32 v12, v1, v61750; TAHITI-NEXT:    v_mul_lo_u32 v6, v1, v61751; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v111752; TAHITI-NEXT:    v_addc_u32_e32 v5, vcc, v10, v5, vcc1753; TAHITI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v12, vcc1754; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v5, v61755; TAHITI-NEXT:    v_addc_u32_e32 v6, vcc, 0, v9, vcc1756; TAHITI-NEXT:    v_mul_lo_u32 v6, v4, v61757; TAHITI-NEXT:    v_mul_hi_u32 v9, v4, v51758; TAHITI-NEXT:    v_mul_lo_u32 v10, v3, v51759; TAHITI-NEXT:    v_mul_lo_u32 v5, v4, v51760; TAHITI-NEXT:    v_add_i32_e32 v6, vcc, v6, v91761; TAHITI-NEXT:    v_add_i32_e32 v6, vcc, v10, v61762; TAHITI-NEXT:    v_sub_i32_e32 v9, vcc, v1, v61763; TAHITI-NEXT:    v_sub_i32_e32 v5, vcc, v8, v51764; TAHITI-NEXT:    v_subb_u32_e64 v8, s[0:1], v9, v3, vcc1765; TAHITI-NEXT:    v_sub_i32_e64 v9, s[0:1], v5, v41766; TAHITI-NEXT:    v_subbrev_u32_e64 v10, s[2:3], 0, v8, s[0:1]1767; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v10, v31768; TAHITI-NEXT:    v_cndmask_b32_e64 v11, 0, -1, s[2:3]1769; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v9, v41770; TAHITI-NEXT:    v_subb_u32_e32 v1, vcc, v1, v6, vcc1771; TAHITI-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[2:3]1772; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v10, v31773; TAHITI-NEXT:    v_subb_u32_e64 v8, s[0:1], v8, v3, s[0:1]1774; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v31775; TAHITI-NEXT:    v_cndmask_b32_e64 v11, v11, v12, s[2:3]1776; TAHITI-NEXT:    v_sub_i32_e64 v12, s[0:1], v9, v41777; TAHITI-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc1778; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v41779; TAHITI-NEXT:    v_subbrev_u32_e64 v8, s[0:1], 0, v8, s[0:1]1780; TAHITI-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc1781; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v31782; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v111783; TAHITI-NEXT:    v_cndmask_b32_e32 v3, v6, v4, vcc1784; TAHITI-NEXT:    v_cndmask_b32_e64 v9, v9, v12, s[0:1]1785; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31786; TAHITI-NEXT:    v_cndmask_b32_e64 v8, v10, v8, s[0:1]1787; TAHITI-NEXT:    v_cndmask_b32_e32 v3, v5, v9, vcc1788; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc1789; TAHITI-NEXT:    v_xor_b32_e32 v3, v3, v71790; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v71791; TAHITI-NEXT:    v_sub_i32_e32 v3, vcc, v3, v71792; TAHITI-NEXT:    v_subb_u32_e32 v4, vcc, v1, v7, vcc1793; TAHITI-NEXT:    s_cbranch_execnz .LBB8_31794; TAHITI-NEXT:  .LBB8_2:1795; TAHITI-NEXT:    v_cvt_f32_u32_e32 v1, v21796; TAHITI-NEXT:    v_sub_i32_e32 v3, vcc, 0, v21797; TAHITI-NEXT:    v_mov_b32_e32 v4, 01798; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v1, v11799; TAHITI-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v11800; TAHITI-NEXT:    v_cvt_u32_f32_e32 v1, v11801; TAHITI-NEXT:    v_mul_lo_u32 v3, v3, v11802; TAHITI-NEXT:    v_mul_hi_u32 v3, v1, v31803; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v31804; TAHITI-NEXT:    v_mul_hi_u32 v1, v0, v11805; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, v21806; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v0, v11807; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, v0, v21808; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v21809; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc1810; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, v0, v21811; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v21812; TAHITI-NEXT:    v_cndmask_b32_e32 v3, v0, v1, vcc1813; TAHITI-NEXT:  .LBB8_3:1814; TAHITI-NEXT:    s_mov_b32 s7, 0xf0001815; TAHITI-NEXT:    s_mov_b32 s6, -11816; TAHITI-NEXT:    buffer_store_dwordx2 v[3:4], off, s[4:7], 01817; TAHITI-NEXT:    s_endpgm1818; TAHITI-NEXT:  .LBB8_4:1819; TAHITI-NEXT:    ; implicit-def: $vgpr3_vgpr41820; TAHITI-NEXT:    s_branch .LBB8_21821;1822; TONGA-LABEL: srem_i64:1823; TONGA:       ; %bb.0:1824; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x241825; TONGA-NEXT:    s_waitcnt lgkmcnt(0)1826; TONGA-NEXT:    v_mov_b32_e32 v0, s21827; TONGA-NEXT:    v_mov_b32_e32 v1, s31828; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]1829; TONGA-NEXT:    s_waitcnt vmcnt(0)1830; TONGA-NEXT:    v_readfirstlane_b32 s5, v11831; TONGA-NEXT:    v_readfirstlane_b32 s4, v01832; TONGA-NEXT:    v_readfirstlane_b32 s3, v31833; TONGA-NEXT:    v_readfirstlane_b32 s2, v21834; TONGA-NEXT:    s_or_b64 s[6:7], s[4:5], s[2:3]1835; TONGA-NEXT:    s_mov_b32 s6, 01836; TONGA-NEXT:    s_cmp_lg_u64 s[6:7], 01837; TONGA-NEXT:    s_cbranch_scc0 .LBB8_31838; TONGA-NEXT:  ; %bb.1:1839; TONGA-NEXT:    s_ashr_i32 s6, s3, 311840; TONGA-NEXT:    s_add_u32 s8, s2, s61841; TONGA-NEXT:    s_mov_b32 s7, s61842; TONGA-NEXT:    s_addc_u32 s9, s3, s61843; TONGA-NEXT:    s_xor_b64 s[6:7], s[8:9], s[6:7]1844; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s61845; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, s71846; TONGA-NEXT:    s_sub_u32 s3, 0, s61847; TONGA-NEXT:    s_subb_u32 s10, 0, s71848; TONGA-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v01849; TONGA-NEXT:    v_rcp_f32_e32 v0, v01850; TONGA-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v01851; TONGA-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v01852; TONGA-NEXT:    v_trunc_f32_e32 v1, v11853; TONGA-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v01854; TONGA-NEXT:    v_cvt_u32_f32_e32 v4, v11855; TONGA-NEXT:    v_cvt_u32_f32_e32 v5, v01856; TONGA-NEXT:    v_mul_lo_u32 v2, s3, v41857; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s3, v5, 01858; TONGA-NEXT:    v_mul_lo_u32 v3, s10, v51859; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v1, v21860; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v1, v31861; TONGA-NEXT:    v_mul_hi_u32 v6, v5, v01862; TONGA-NEXT:    v_mad_u64_u32 v[1:2], s[8:9], v5, v3, 01863; TONGA-NEXT:    v_add_u32_e32 v6, vcc, v6, v11864; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], v4, v0, 01865; TONGA-NEXT:    v_addc_u32_e32 v7, vcc, 0, v2, vcc1866; TONGA-NEXT:    v_mad_u64_u32 v[2:3], s[8:9], v4, v3, 01867; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v6, v01868; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v7, v1, vcc1869; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v3, vcc1870; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v21871; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1872; TONGA-NEXT:    v_add_u32_e32 v6, vcc, v5, v01873; TONGA-NEXT:    v_addc_u32_e32 v7, vcc, v4, v1, vcc1874; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s3, v6, 01875; TONGA-NEXT:    v_mul_lo_u32 v4, s3, v71876; TONGA-NEXT:    v_mul_lo_u32 v5, s10, v61877; TONGA-NEXT:    v_mul_hi_u32 v8, v6, v01878; TONGA-NEXT:    v_mad_u64_u32 v[2:3], s[8:9], v7, v0, 01879; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v4, v11880; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v5, v11881; TONGA-NEXT:    v_mad_u64_u32 v[4:5], s[8:9], v6, v1, 01882; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], v7, v1, 01883; TONGA-NEXT:    v_add_u32_e32 v4, vcc, v8, v41884; TONGA-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc1885; TONGA-NEXT:    v_add_u32_e32 v2, vcc, v4, v21886; TONGA-NEXT:    v_addc_u32_e32 v2, vcc, v5, v3, vcc1887; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1888; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v2, v01889; TONGA-NEXT:    s_ashr_i32 s10, s5, 311890; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc1891; TONGA-NEXT:    s_add_u32 s8, s4, s101892; TONGA-NEXT:    v_add_u32_e32 v2, vcc, v6, v01893; TONGA-NEXT:    s_mov_b32 s11, s101894; TONGA-NEXT:    s_addc_u32 s9, s5, s101895; TONGA-NEXT:    v_addc_u32_e32 v3, vcc, v7, v1, vcc1896; TONGA-NEXT:    s_xor_b64 s[12:13], s[8:9], s[10:11]1897; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s12, v3, 01898; TONGA-NEXT:    v_mul_hi_u32 v4, s12, v21899; TONGA-NEXT:    v_readfirstlane_b32 s3, v11900; TONGA-NEXT:    v_readfirstlane_b32 s5, v01901; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s13, v3, 01902; TONGA-NEXT:    v_mad_u64_u32 v[2:3], s[8:9], s13, v2, 01903; TONGA-NEXT:    v_readfirstlane_b32 s14, v41904; TONGA-NEXT:    s_add_u32 s5, s14, s51905; TONGA-NEXT:    s_addc_u32 s3, 0, s31906; TONGA-NEXT:    v_readfirstlane_b32 s14, v21907; TONGA-NEXT:    v_readfirstlane_b32 s9, v31908; TONGA-NEXT:    s_add_u32 s5, s5, s141909; TONGA-NEXT:    v_readfirstlane_b32 s8, v11910; TONGA-NEXT:    s_addc_u32 s3, s3, s91911; TONGA-NEXT:    s_addc_u32 s5, s8, 01912; TONGA-NEXT:    v_readfirstlane_b32 s8, v01913; TONGA-NEXT:    s_add_u32 s3, s3, s81914; TONGA-NEXT:    v_mov_b32_e32 v0, s31915; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s6, v0, 01916; TONGA-NEXT:    s_addc_u32 s5, 0, s51917; TONGA-NEXT:    s_mul_i32 s5, s6, s51918; TONGA-NEXT:    v_readfirstlane_b32 s14, v11919; TONGA-NEXT:    s_add_i32 s5, s14, s51920; TONGA-NEXT:    s_mul_i32 s3, s7, s31921; TONGA-NEXT:    s_add_i32 s5, s5, s31922; TONGA-NEXT:    s_sub_i32 s3, s13, s51923; TONGA-NEXT:    v_readfirstlane_b32 s14, v01924; TONGA-NEXT:    s_sub_u32 s12, s12, s141925; TONGA-NEXT:    s_cselect_b64 s[14:15], -1, 01926; TONGA-NEXT:    s_subb_u32 s3, s3, s71927; TONGA-NEXT:    s_sub_u32 s18, s12, s61928; TONGA-NEXT:    s_cselect_b64 s[16:17], -1, 01929; TONGA-NEXT:    s_subb_u32 s19, s3, 01930; TONGA-NEXT:    s_cmp_ge_u32 s19, s71931; TONGA-NEXT:    s_cselect_b32 s20, -1, 01932; TONGA-NEXT:    s_cmp_ge_u32 s18, s61933; TONGA-NEXT:    s_cselect_b32 s21, -1, 01934; TONGA-NEXT:    s_cmp_eq_u32 s19, s71935; TONGA-NEXT:    s_cselect_b32 s20, s21, s201936; TONGA-NEXT:    s_cmp_lg_u64 s[16:17], 01937; TONGA-NEXT:    s_subb_u32 s3, s3, s71938; TONGA-NEXT:    s_sub_u32 s16, s18, s61939; TONGA-NEXT:    s_subb_u32 s3, s3, 01940; TONGA-NEXT:    s_cmp_lg_u32 s20, 01941; TONGA-NEXT:    s_cselect_b32 s16, s16, s181942; TONGA-NEXT:    s_cselect_b32 s3, s3, s191943; TONGA-NEXT:    s_cmp_lg_u64 s[14:15], 01944; TONGA-NEXT:    s_subb_u32 s5, s13, s51945; TONGA-NEXT:    s_cmp_ge_u32 s5, s71946; TONGA-NEXT:    s_cselect_b32 s13, -1, 01947; TONGA-NEXT:    s_cmp_ge_u32 s12, s61948; TONGA-NEXT:    s_cselect_b32 s6, -1, 01949; TONGA-NEXT:    s_cmp_eq_u32 s5, s71950; TONGA-NEXT:    s_cselect_b32 s6, s6, s131951; TONGA-NEXT:    s_cmp_lg_u32 s6, 01952; TONGA-NEXT:    s_cselect_b32 s7, s3, s51953; TONGA-NEXT:    s_cselect_b32 s6, s16, s121954; TONGA-NEXT:    s_xor_b64 s[6:7], s[6:7], s[10:11]1955; TONGA-NEXT:    s_sub_u32 s6, s6, s101956; TONGA-NEXT:    s_subb_u32 s7, s7, s101957; TONGA-NEXT:    s_cbranch_execnz .LBB8_41958; TONGA-NEXT:  .LBB8_2:1959; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s21960; TONGA-NEXT:    s_sub_i32 s3, 0, s21961; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v01962; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v01963; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v01964; TONGA-NEXT:    v_mul_lo_u32 v1, s3, v01965; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v11966; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v11967; TONGA-NEXT:    v_mul_hi_u32 v0, s4, v01968; TONGA-NEXT:    v_mul_lo_u32 v0, v0, s21969; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, s4, v01970; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, s2, v01971; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s2, v01972; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc1973; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, s2, v01974; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s2, v01975; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc1976; TONGA-NEXT:    v_mov_b32_e32 v1, 01977; TONGA-NEXT:    s_branch .LBB8_51978; TONGA-NEXT:  .LBB8_3:1979; TONGA-NEXT:    ; implicit-def: $sgpr6_sgpr71980; TONGA-NEXT:    s_branch .LBB8_21981; TONGA-NEXT:  .LBB8_4:1982; TONGA-NEXT:    v_mov_b32_e32 v0, s61983; TONGA-NEXT:    v_mov_b32_e32 v1, s71984; TONGA-NEXT:  .LBB8_5:1985; TONGA-NEXT:    v_mov_b32_e32 v2, s01986; TONGA-NEXT:    v_mov_b32_e32 v3, s11987; TONGA-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]1988; TONGA-NEXT:    s_endpgm1989;1990; EG-LABEL: srem_i64:1991; EG:       ; %bb.0:1992; EG-NEXT:    ALU 0, @14, KC0[CB0:0-32], KC1[]1993; EG-NEXT:    TEX 0 @121994; EG-NEXT:    ALU_PUSH_BEFORE 4, @15, KC0[], KC1[]1995; EG-NEXT:    JUMP @9 POP:11996; EG-NEXT:    ALU 114, @20, KC0[], KC1[]1997; EG-NEXT:    ALU 115, @135, KC0[], KC1[]1998; EG-NEXT:    ALU 115, @251, KC0[], KC1[]1999; EG-NEXT:    ALU 115, @367, KC0[], KC1[]2000; EG-NEXT:    ALU_POP_AFTER 82, @483, KC0[], KC1[]2001; EG-NEXT:    ALU 20, @566, KC0[CB0:0-32], KC1[]2002; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T1.XY, T0.X, 12003; EG-NEXT:    CF_END2004; EG-NEXT:    Fetch clause starting at 12:2005; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #12006; EG-NEXT:    ALU clause starting at 14:2007; EG-NEXT:     MOV * T0.X, KC0[2].Z,2008; EG-NEXT:    ALU clause starting at 15:2009; EG-NEXT:     OR_INT T2.W, T0.Y, T0.W,2010; EG-NEXT:     MOV * T1.W, literal.x,2011; EG-NEXT:    1(1.401298e-45), 0(0.000000e+00)2012; EG-NEXT:     SETNE_INT * T2.W, PV.W, 0.0,2013; EG-NEXT:     PRED_SETNE_INT * ExecMask,PredicateBit (MASKED), PV.W, 0.0,2014; EG-NEXT:    ALU clause starting at 20:2015; EG-NEXT:     ASHR * T3.W, T0.W, literal.x,2016; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2017; EG-NEXT:     ADD_INT * T1.W, T0.Z, PV.W,2018; EG-NEXT:     XOR_INT * T2.W, PV.W, T3.W,2019; EG-NEXT:     SUB_INT T1.Z, 0.0, PV.W,2020; EG-NEXT:     ASHR T1.W, T0.Y, literal.x,2021; EG-NEXT:     RECIP_UINT * T1.X, PV.W,2022; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2023; EG-NEXT:     ADD_INT T2.Z, T0.Y, PV.W,2024; EG-NEXT:     ADDC_UINT T4.W, T0.X, PV.W,2025; EG-NEXT:     MULLO_INT * T0.Y, PV.Z, PS,2026; EG-NEXT:     ADD_INT T4.W, PV.Z, PV.W,2027; EG-NEXT:     MULHI * T0.Y, T1.X, PS,2028; EG-NEXT:     ADD_INT T5.W, T1.X, PS,2029; EG-NEXT:     XOR_INT * T4.W, PV.W, T1.W,2030; EG-NEXT:     MULHI * T0.Y, PS, PV.W,2031; EG-NEXT:     MULLO_INT * T0.Y, PS, T2.W,2032; EG-NEXT:     SUB_INT * T5.W, T4.W, PS,2033; EG-NEXT:     SETGE_UINT T6.W, PV.W, T2.W,2034; EG-NEXT:     SUB_INT * T7.W, PV.W, T2.W,2035; EG-NEXT:     CNDE_INT T1.Z, PV.W, T5.W, PS, BS:VEC_021/SCL_1222036; EG-NEXT:     ADD_INT T0.W, T0.W, T3.W,2037; EG-NEXT:     ADDC_UINT * T5.W, T0.Z, T3.W,2038; EG-NEXT:     ADD_INT T2.Z, PV.W, PS,2039; EG-NEXT:     SETGE_UINT T0.W, PV.Z, T2.W,2040; EG-NEXT:     SUB_INT * T5.W, PV.Z, T2.W,2041; EG-NEXT:     ADD_INT T3.Z, T0.X, T1.W, BS:VEC_021/SCL_1222042; EG-NEXT:     CNDE_INT T5.W, PV.W, T1.Z, PS,2043; EG-NEXT:     XOR_INT * T0.W, PV.Z, T3.W,2044; EG-NEXT:     CNDE_INT T4.W, PS, PV.W, T4.W,2045; EG-NEXT:     XOR_INT * T3.W, PV.Z, T1.W,2046; EG-NEXT:     BIT_ALIGN_INT T5.W, PV.W, PS, literal.x,2047; EG-NEXT:     LSHR * T4.W, PV.W, literal.x,2048; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2049; EG-NEXT:     SETE_INT T1.Z, PS, T0.W, BS:VEC_021/SCL_1222050; EG-NEXT:     SETGE_UINT T6.W, PS, T0.W, BS:VEC_021/SCL_1222051; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T2.W,2052; EG-NEXT:     CNDE_INT T0.Y, PV.Z, PV.W, PS,2053; EG-NEXT:     SUB_INT * T1.Z, T5.W, T2.W,2054; EG-NEXT:     SUB_INT * T6.W, T4.W, T0.W,2055; EG-NEXT:     SUBB_UINT * T7.W, T5.W, T2.W,2056; EG-NEXT:     SUB_INT T6.W, T6.W, PV.W,2057; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2058; EG-NEXT:     LSHL T1.Z, PS, 1,2059; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2060; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2061; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)2062; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2063; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2064; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2065; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222066; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2067; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2068; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2069; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2070; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2071; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2072; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222073; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2074; EG-NEXT:     LSHL T1.Z, PS, 1,2075; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2076; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2077; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)2078; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2079; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2080; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2081; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222082; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2083; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2084; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2085; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2086; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2087; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2088; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222089; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2090; EG-NEXT:     LSHL T1.Z, PS, 1,2091; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2092; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2093; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)2094; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2095; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2096; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2097; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222098; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2099; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2100; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2101; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2102; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2103; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2104; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222105; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2106; EG-NEXT:     LSHL T1.Z, PS, 1,2107; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2108; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2109; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)2110; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2111; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2112; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2113; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222114; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2115; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2116; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2117; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2118; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2119; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2120; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222121; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2122; EG-NEXT:     LSHL T1.Z, PS, 1,2123; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2124; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2125; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)2126; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2127; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2128; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2129; EG-NEXT:     SETGE_UINT * T1.Z, PS, T2.W,2130; EG-NEXT:    ALU clause starting at 135:2131; EG-NEXT:     SETE_INT T6.W, T4.W, T0.W,2132; EG-NEXT:     SETGE_UINT * T7.W, T4.W, T0.W,2133; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, T1.Z,2134; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2135; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2136; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2137; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222138; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2139; EG-NEXT:     LSHL T1.Z, PS, 1,2140; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2141; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2142; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)2143; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2144; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2145; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2146; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222147; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2148; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2149; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2150; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2151; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2152; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2153; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222154; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2155; EG-NEXT:     LSHL T1.Z, PS, 1,2156; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2157; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2158; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)2159; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2160; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2161; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2162; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222163; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2164; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2165; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2166; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2167; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2168; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2169; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222170; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2171; EG-NEXT:     LSHL T1.Z, PS, 1,2172; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2173; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2174; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)2175; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2176; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2177; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2178; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222179; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2180; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2181; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2182; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2183; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2184; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2185; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222186; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2187; EG-NEXT:     LSHL T1.Z, PS, 1,2188; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2189; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2190; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)2191; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2192; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2193; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2194; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222195; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2196; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2197; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2198; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2199; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2200; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2201; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222202; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2203; EG-NEXT:     LSHL T1.Z, PS, 1,2204; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2205; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2206; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)2207; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2208; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2209; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2210; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222211; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2212; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2213; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2214; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2215; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2216; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2217; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222218; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2219; EG-NEXT:     LSHL T1.Z, PS, 1,2220; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2221; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2222; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)2223; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2224; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2225; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2226; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222227; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2228; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2229; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2230; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2231; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2232; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2233; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222234; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2235; EG-NEXT:     LSHL T1.Z, PS, 1,2236; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2237; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2238; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)2239; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2240; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2241; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2242; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222243; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2244; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2245; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2246; EG-NEXT:     SUB_INT * T1.Z, T5.W, T2.W,2247; EG-NEXT:    ALU clause starting at 251:2248; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2249; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2250; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222251; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2252; EG-NEXT:     LSHL T1.Z, PS, 1,2253; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2254; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2255; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)2256; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2257; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2258; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2259; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222260; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2261; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2262; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2263; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2264; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2265; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2266; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222267; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2268; EG-NEXT:     LSHL T1.Z, PS, 1,2269; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2270; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2271; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)2272; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2273; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2274; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2275; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222276; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2277; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2278; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2279; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2280; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2281; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2282; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222283; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2284; EG-NEXT:     LSHL T1.Z, PS, 1,2285; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2286; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2287; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)2288; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2289; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2290; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2291; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222292; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2293; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2294; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2295; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2296; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2297; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2298; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222299; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2300; EG-NEXT:     LSHL T1.Z, PS, 1,2301; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2302; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2303; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)2304; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2305; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2306; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2307; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222308; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2309; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2310; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2311; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2312; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2313; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2314; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222315; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2316; EG-NEXT:     LSHL T1.Z, PS, 1,2317; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2318; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2319; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)2320; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2321; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2322; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2323; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222324; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2325; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2326; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2327; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2328; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2329; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2330; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222331; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2332; EG-NEXT:     LSHL T1.Z, PS, 1,2333; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2334; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2335; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)2336; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2337; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2338; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2339; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222340; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2341; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2342; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2343; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2344; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2345; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2346; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222347; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2348; EG-NEXT:     LSHL T1.Z, PS, 1,2349; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2350; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2351; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)2352; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2353; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2354; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2355; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222356; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2357; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2358; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2359; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2360; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2361; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2362; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222363; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2364; EG-NEXT:    ALU clause starting at 367:2365; EG-NEXT:     LSHL T1.Z, T5.W, 1,2366; EG-NEXT:     BFE_UINT * T7.W, T3.W, literal.x, 1, BS:VEC_120/SCL_2122367; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)2368; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T6.W,2369; EG-NEXT:     BIT_ALIGN_INT T4.W, PV.W, T5.W, literal.x, BS:VEC_021/SCL_1222370; EG-NEXT:     OR_INT * T5.W, T1.Z, T7.W,2371; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2372; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222373; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2374; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2375; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2376; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2377; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2378; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2379; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222380; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2381; EG-NEXT:     LSHL T1.Z, PS, 1,2382; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2383; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2384; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)2385; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2386; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2387; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2388; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222389; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2390; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2391; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2392; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2393; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2394; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2395; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222396; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2397; EG-NEXT:     LSHL T1.Z, PS, 1,2398; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2399; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2400; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)2401; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2402; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2403; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2404; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222405; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2406; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2407; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2408; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2409; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2410; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2411; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222412; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2413; EG-NEXT:     LSHL T1.Z, PS, 1,2414; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2415; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2416; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)2417; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2418; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2419; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2420; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222421; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2422; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2423; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2424; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2425; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2426; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2427; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222428; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2429; EG-NEXT:     LSHL T1.Z, PS, 1,2430; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2431; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2432; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)2433; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2434; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2435; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2436; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222437; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2438; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2439; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2440; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2441; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2442; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2443; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222444; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2445; EG-NEXT:     LSHL T1.Z, PS, 1,2446; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2447; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2448; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)2449; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2450; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2451; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2452; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222453; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2454; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2455; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2456; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2457; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2458; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2459; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222460; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2461; EG-NEXT:     LSHL T1.Z, PS, 1,2462; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2463; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2464; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)2465; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2466; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2467; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2468; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222469; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2470; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2471; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2472; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2473; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2474; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2475; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222476; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2477; EG-NEXT:     LSHL T1.Z, PS, 1,2478; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2479; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2480; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)2481; EG-NEXT:    ALU clause starting at 483:2482; EG-NEXT:     BIT_ALIGN_INT T4.W, T4.W, T5.W, literal.x, BS:VEC_021/SCL_1222483; EG-NEXT:     OR_INT * T5.W, T1.Z, T7.W,2484; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2485; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222486; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2487; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2488; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2489; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2490; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2491; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2492; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222493; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2494; EG-NEXT:     LSHL T1.Z, PS, 1,2495; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2496; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2497; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)2498; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2499; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2500; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2501; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222502; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2503; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2504; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2505; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2506; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2507; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2508; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222509; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2510; EG-NEXT:     LSHL T1.Z, PS, 1,2511; EG-NEXT:     BFE_UINT T7.W, T3.W, literal.x, 1,2512; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2513; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)2514; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2515; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2516; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2517; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222518; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2519; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2520; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2521; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2522; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2523; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2524; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222525; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2526; EG-NEXT:     LSHL T1.Z, PS, 1,2527; EG-NEXT:     BFE_UINT T7.W, T3.W, 1, 1,2528; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2529; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2530; EG-NEXT:     OR_INT * T5.W, PV.Z, PV.W,2531; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2532; EG-NEXT:     SETGE_UINT T1.Z, PS, T2.W, BS:VEC_021/SCL_1222533; EG-NEXT:     SETE_INT T6.W, PV.W, T0.W,2534; EG-NEXT:     SETGE_UINT * T7.W, PV.W, T0.W,2535; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,2536; EG-NEXT:     SUB_INT T1.Z, T5.W, T2.W,2537; EG-NEXT:     SUBB_UINT * T6.W, T5.W, T2.W,2538; EG-NEXT:     SUB_INT * T7.W, T4.W, T0.W,2539; EG-NEXT:     SUB_INT T6.W, PV.W, T6.W, BS:VEC_021/SCL_1222540; EG-NEXT:     CNDE_INT * T5.W, T0.Y, T5.W, T1.Z,2541; EG-NEXT:     LSHL T1.Z, PS, 1,2542; EG-NEXT:     AND_INT T3.W, T3.W, 1,2543; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, PV.W,2544; EG-NEXT:     BIT_ALIGN_INT T4.W, PS, T5.W, literal.x,2545; EG-NEXT:     OR_INT * T3.W, PV.Z, PV.W,2546; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2547; EG-NEXT:     SETGE_UINT T1.X, PS, T2.W, BS:VEC_021/SCL_1222548; EG-NEXT:     SETE_INT T0.Y, PV.W, T0.W,2549; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T0.W,2550; EG-NEXT:     SUBB_UINT T5.W, PS, T2.W, BS:VEC_021/SCL_1222551; EG-NEXT:     SUB_INT * T0.W, PV.W, T0.W,2552; EG-NEXT:     SUB_INT T2.Z, T3.W, T2.W,2553; EG-NEXT:     SUB_INT T0.W, PS, PV.W,2554; EG-NEXT:     CNDE_INT * T2.W, PV.Y, PV.Z, PV.X,2555; EG-NEXT:     CNDE_INT T0.W, PS, T4.W, PV.W, BS:VEC_021/SCL_1222556; EG-NEXT:     CNDE_INT * T2.W, PS, T3.W, PV.Z,2557; EG-NEXT:     XOR_INT T2.W, PS, T1.W,2558; EG-NEXT:     XOR_INT * T0.W, PV.W, T1.W,2559; EG-NEXT:     SUB_INT T0.W, PS, T1.W,2560; EG-NEXT:     SUBB_UINT * T3.W, PV.W, T1.W,2561; EG-NEXT:     SUB_INT * T1.Y, PV.W, PS,2562; EG-NEXT:     SUB_INT T1.X, T2.W, T1.W,2563; EG-NEXT:     MOV * T1.W, literal.x,2564; EG-NEXT:    0(0.000000e+00), 0(0.000000e+00)2565; EG-NEXT:    ALU clause starting at 566:2566; EG-NEXT:     MOV T0.W, KC0[2].Y,2567; EG-NEXT:     SETE_INT * T1.W, T1.W, 0.0,2568; EG-NEXT:     PRED_SETNE_INT * Pred,PredicateBit (MASKED), PS, 0.0,2569; EG-NEXT:     SUB_INT T1.W, 0.0, T0.Z, Pred_sel_zero2570; EG-NEXT:     RECIP_UINT * T0.Y, T0.Z, Pred_sel_zero2571; EG-NEXT:     MULLO_INT * T1.X, T1.W, T0.Y, Pred_sel_zero2572; EG-NEXT:     MULHI * T1.X, T0.Y, T1.X, Pred_sel_zero2573; EG-NEXT:     ADD_INT * T1.W, T0.Y, T1.X, Pred_sel_zero2574; EG-NEXT:     MULHI * T0.Y, T0.X, T1.W, Pred_sel_zero2575; EG-NEXT:     MULLO_INT * T0.Y, T0.Y, T0.Z, Pred_sel_zero2576; EG-NEXT:     SUB_INT * T1.W, T0.X, T0.Y, Pred_sel_zero2577; EG-NEXT:     SETGE_UINT T2.W, T1.W, T0.Z, Pred_sel_zero2578; EG-NEXT:     SUB_INT * T3.W, T1.W, T0.Z, Pred_sel_zero2579; EG-NEXT:     CNDE_INT * T1.W, T2.W, T1.W, T3.W, Pred_sel_zero2580; EG-NEXT:     SETGE_UINT T2.W, T1.W, T0.Z, Pred_sel_zero2581; EG-NEXT:     SUB_INT * T3.W, T1.W, T0.Z, Pred_sel_zero2582; EG-NEXT:     CNDE_INT T1.X, T2.W, T1.W, T3.W, Pred_sel_zero2583; EG-NEXT:     MOV * T1.Y, literal.x, Pred_sel_zero2584; EG-NEXT:    0(0.000000e+00), 0(0.000000e+00)2585; EG-NEXT:     LSHR * T0.X, T0.W, literal.x,2586; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)2587  %den_ptr = getelementptr i64, ptr addrspace(1) %in, i64 12588  %num = load i64, ptr addrspace(1) %in2589  %den = load i64, ptr addrspace(1) %den_ptr2590  %result = srem i64 %num, %den2591  store i64 %result, ptr addrspace(1) %out2592  ret void2593}2594 2595define amdgpu_kernel void @srem_i64_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {2596; GCN-LABEL: srem_i64_4:2597; GCN:       ; %bb.0:2598; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x242599; GCN-NEXT:    v_mov_b32_e32 v2, 02600; GCN-NEXT:    s_waitcnt lgkmcnt(0)2601; GCN-NEXT:    global_load_dwordx2 v[0:1], v2, s[2:3]2602; GCN-NEXT:    s_waitcnt vmcnt(0)2603; GCN-NEXT:    v_ashrrev_i32_e32 v3, 31, v12604; GCN-NEXT:    v_lshrrev_b32_e32 v3, 30, v32605; GCN-NEXT:    v_add_co_u32_e32 v3, vcc, v0, v32606; GCN-NEXT:    v_addc_co_u32_e32 v4, vcc, 0, v1, vcc2607; GCN-NEXT:    v_and_b32_e32 v3, -4, v32608; GCN-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v32609; GCN-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v4, vcc2610; GCN-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]2611; GCN-NEXT:    s_endpgm2612;2613; TAHITI-LABEL: srem_i64_4:2614; TAHITI:       ; %bb.0:2615; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x92616; TAHITI-NEXT:    s_mov_b32 s7, 0xf0002617; TAHITI-NEXT:    s_mov_b32 s6, -12618; TAHITI-NEXT:    s_mov_b32 s10, s62619; TAHITI-NEXT:    s_mov_b32 s11, s72620; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)2621; TAHITI-NEXT:    s_mov_b32 s8, s22622; TAHITI-NEXT:    s_mov_b32 s9, s32623; TAHITI-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 02624; TAHITI-NEXT:    s_mov_b32 s4, s02625; TAHITI-NEXT:    s_mov_b32 s5, s12626; TAHITI-NEXT:    s_waitcnt vmcnt(0)2627; TAHITI-NEXT:    v_ashrrev_i32_e32 v2, 31, v12628; TAHITI-NEXT:    v_lshrrev_b32_e32 v2, 30, v22629; TAHITI-NEXT:    v_add_i32_e32 v2, vcc, v0, v22630; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc2631; TAHITI-NEXT:    v_and_b32_e32 v2, -4, v22632; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v0, v22633; TAHITI-NEXT:    v_subb_u32_e32 v1, vcc, v1, v3, vcc2634; TAHITI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 02635; TAHITI-NEXT:    s_endpgm2636;2637; TONGA-LABEL: srem_i64_4:2638; TONGA:       ; %bb.0:2639; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x242640; TONGA-NEXT:    s_waitcnt lgkmcnt(0)2641; TONGA-NEXT:    v_mov_b32_e32 v0, s22642; TONGA-NEXT:    v_mov_b32_e32 v1, s32643; TONGA-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]2644; TONGA-NEXT:    v_mov_b32_e32 v2, s02645; TONGA-NEXT:    v_mov_b32_e32 v3, s12646; TONGA-NEXT:    s_waitcnt vmcnt(0)2647; TONGA-NEXT:    v_ashrrev_i32_e32 v4, 31, v12648; TONGA-NEXT:    v_lshrrev_b32_e32 v4, 30, v42649; TONGA-NEXT:    v_add_u32_e32 v4, vcc, v0, v42650; TONGA-NEXT:    v_addc_u32_e32 v5, vcc, 0, v1, vcc2651; TONGA-NEXT:    v_and_b32_e32 v4, -4, v42652; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v0, v42653; TONGA-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc2654; TONGA-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]2655; TONGA-NEXT:    s_endpgm2656;2657; EG-LABEL: srem_i64_4:2658; EG:       ; %bb.0:2659; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]2660; EG-NEXT:    TEX 0 @62661; EG-NEXT:    ALU 13, @9, KC0[CB0:0-32], KC1[]2662; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 12663; EG-NEXT:    CF_END2664; EG-NEXT:    PAD2665; EG-NEXT:    Fetch clause starting at 6:2666; EG-NEXT:     VTX_READ_64 T0.XY, T0.X, 0, #12667; EG-NEXT:    ALU clause starting at 8:2668; EG-NEXT:     MOV * T0.X, KC0[2].Z,2669; EG-NEXT:    ALU clause starting at 9:2670; EG-NEXT:     ASHR * T0.W, T0.Y, literal.x,2671; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)2672; EG-NEXT:     LSHR * T0.W, PV.W, literal.x,2673; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)2674; EG-NEXT:     ADD_INT * T1.W, T0.X, PV.W,2675; EG-NEXT:     AND_INT T1.W, PV.W, literal.x,2676; EG-NEXT:     ADDC_UINT * T0.W, T0.X, T0.W,2677; EG-NEXT:    -4(nan), 0(0.000000e+00)2678; EG-NEXT:     BFE_INT T0.W, PS, 0.0, 1,2679; EG-NEXT:     SUBB_UINT * T2.W, T0.X, PV.W,2680; EG-NEXT:     SUB_INT * T0.Y, PV.W, PS,2681; EG-NEXT:     SUB_INT T0.X, T0.X, T1.W,2682; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,2683; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)2684  %num = load i64, ptr addrspace(1) %in2685  %result = srem i64 %num, 42686  store i64 %result, ptr addrspace(1) %out2687  ret void2688}2689 2690define amdgpu_kernel void @srem_v2i64(ptr addrspace(1) %out, ptr addrspace(1) %in) {2691; GCN-LABEL: srem_v2i64:2692; GCN:       ; %bb.0:2693; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x242694; GCN-NEXT:    v_mov_b32_e32 v8, 02695; GCN-NEXT:    s_waitcnt lgkmcnt(0)2696; GCN-NEXT:    global_load_dwordx4 v[0:3], v8, s[2:3] offset:162697; GCN-NEXT:    global_load_dwordx4 v[4:7], v8, s[2:3]2698; GCN-NEXT:    s_waitcnt vmcnt(1)2699; GCN-NEXT:    v_readfirstlane_b32 s9, v12700; GCN-NEXT:    v_readfirstlane_b32 s8, v02701; GCN-NEXT:    s_waitcnt vmcnt(0)2702; GCN-NEXT:    v_readfirstlane_b32 s11, v52703; GCN-NEXT:    v_readfirstlane_b32 s10, v42704; GCN-NEXT:    s_or_b64 s[6:7], s[10:11], s[8:9]2705; GCN-NEXT:    s_mov_b32 s6, 02706; GCN-NEXT:    v_readfirstlane_b32 s3, v32707; GCN-NEXT:    v_readfirstlane_b32 s2, v22708; GCN-NEXT:    v_readfirstlane_b32 s5, v72709; GCN-NEXT:    s_cmp_lg_u64 s[6:7], 02710; GCN-NEXT:    v_readfirstlane_b32 s4, v62711; GCN-NEXT:    s_cbranch_scc0 .LBB10_62712; GCN-NEXT:  ; %bb.1:2713; GCN-NEXT:    s_ashr_i32 s6, s9, 312714; GCN-NEXT:    s_add_u32 s12, s8, s62715; GCN-NEXT:    s_mov_b32 s7, s62716; GCN-NEXT:    s_addc_u32 s13, s9, s62717; GCN-NEXT:    s_xor_b64 s[6:7], s[12:13], s[6:7]2718; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s62719; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s72720; GCN-NEXT:    s_sub_u32 s9, 0, s62721; GCN-NEXT:    s_subb_u32 s14, 0, s72722; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v02723; GCN-NEXT:    v_rcp_f32_e32 v0, v02724; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v02725; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v02726; GCN-NEXT:    v_trunc_f32_e32 v1, v12727; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v02728; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v12729; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v02730; GCN-NEXT:    v_readfirstlane_b32 s15, v12731; GCN-NEXT:    v_readfirstlane_b32 s16, v02732; GCN-NEXT:    s_mul_i32 s17, s9, s152733; GCN-NEXT:    s_mul_hi_u32 s19, s9, s162734; GCN-NEXT:    s_mul_i32 s18, s14, s162735; GCN-NEXT:    s_add_i32 s17, s19, s172736; GCN-NEXT:    s_add_i32 s17, s17, s182737; GCN-NEXT:    s_mul_i32 s20, s9, s162738; GCN-NEXT:    s_mul_i32 s19, s16, s172739; GCN-NEXT:    s_mul_hi_u32 s21, s16, s202740; GCN-NEXT:    s_mul_hi_u32 s18, s16, s172741; GCN-NEXT:    s_add_u32 s19, s21, s192742; GCN-NEXT:    s_addc_u32 s18, 0, s182743; GCN-NEXT:    s_mul_hi_u32 s22, s15, s202744; GCN-NEXT:    s_mul_i32 s20, s15, s202745; GCN-NEXT:    s_add_u32 s19, s19, s202746; GCN-NEXT:    s_mul_hi_u32 s21, s15, s172747; GCN-NEXT:    s_addc_u32 s18, s18, s222748; GCN-NEXT:    s_addc_u32 s19, s21, 02749; GCN-NEXT:    s_mul_i32 s17, s15, s172750; GCN-NEXT:    s_add_u32 s17, s18, s172751; GCN-NEXT:    s_addc_u32 s18, 0, s192752; GCN-NEXT:    s_add_u32 s16, s16, s172753; GCN-NEXT:    s_addc_u32 s15, s15, s182754; GCN-NEXT:    s_mul_i32 s17, s9, s152755; GCN-NEXT:    s_mul_hi_u32 s18, s9, s162756; GCN-NEXT:    s_add_i32 s17, s18, s172757; GCN-NEXT:    s_mul_i32 s14, s14, s162758; GCN-NEXT:    s_add_i32 s17, s17, s142759; GCN-NEXT:    s_mul_i32 s9, s9, s162760; GCN-NEXT:    s_mul_hi_u32 s18, s15, s92761; GCN-NEXT:    s_mul_i32 s19, s15, s92762; GCN-NEXT:    s_mul_i32 s21, s16, s172763; GCN-NEXT:    s_mul_hi_u32 s9, s16, s92764; GCN-NEXT:    s_mul_hi_u32 s20, s16, s172765; GCN-NEXT:    s_add_u32 s9, s9, s212766; GCN-NEXT:    s_addc_u32 s20, 0, s202767; GCN-NEXT:    s_add_u32 s9, s9, s192768; GCN-NEXT:    s_mul_hi_u32 s14, s15, s172769; GCN-NEXT:    s_addc_u32 s9, s20, s182770; GCN-NEXT:    s_addc_u32 s14, s14, 02771; GCN-NEXT:    s_mul_i32 s17, s15, s172772; GCN-NEXT:    s_add_u32 s9, s9, s172773; GCN-NEXT:    s_addc_u32 s14, 0, s142774; GCN-NEXT:    s_add_u32 s9, s16, s92775; GCN-NEXT:    s_addc_u32 s18, s15, s142776; GCN-NEXT:    s_ashr_i32 s14, s11, 312777; GCN-NEXT:    s_add_u32 s16, s10, s142778; GCN-NEXT:    s_mov_b32 s15, s142779; GCN-NEXT:    s_addc_u32 s17, s11, s142780; GCN-NEXT:    s_xor_b64 s[16:17], s[16:17], s[14:15]2781; GCN-NEXT:    s_mul_i32 s19, s16, s182782; GCN-NEXT:    s_mul_hi_u32 s20, s16, s92783; GCN-NEXT:    s_mul_hi_u32 s11, s16, s182784; GCN-NEXT:    s_add_u32 s19, s20, s192785; GCN-NEXT:    s_addc_u32 s11, 0, s112786; GCN-NEXT:    s_mul_hi_u32 s21, s17, s92787; GCN-NEXT:    s_mul_i32 s9, s17, s92788; GCN-NEXT:    s_add_u32 s9, s19, s92789; GCN-NEXT:    s_mul_hi_u32 s20, s17, s182790; GCN-NEXT:    s_addc_u32 s9, s11, s212791; GCN-NEXT:    s_addc_u32 s11, s20, 02792; GCN-NEXT:    s_mul_i32 s18, s17, s182793; GCN-NEXT:    s_add_u32 s9, s9, s182794; GCN-NEXT:    s_addc_u32 s11, 0, s112795; GCN-NEXT:    s_mul_i32 s11, s6, s112796; GCN-NEXT:    s_mul_hi_u32 s18, s6, s92797; GCN-NEXT:    s_add_i32 s11, s18, s112798; GCN-NEXT:    s_mul_i32 s18, s7, s92799; GCN-NEXT:    s_add_i32 s11, s11, s182800; GCN-NEXT:    s_sub_i32 s20, s17, s112801; GCN-NEXT:    s_mul_i32 s9, s6, s92802; GCN-NEXT:    s_sub_u32 s9, s16, s92803; GCN-NEXT:    s_cselect_b64 s[18:19], -1, 02804; GCN-NEXT:    s_subb_u32 s16, s20, s72805; GCN-NEXT:    s_sub_u32 s22, s9, s62806; GCN-NEXT:    s_cselect_b64 s[20:21], -1, 02807; GCN-NEXT:    s_subb_u32 s23, s16, 02808; GCN-NEXT:    s_cmp_ge_u32 s23, s72809; GCN-NEXT:    s_cselect_b32 s24, -1, 02810; GCN-NEXT:    s_cmp_ge_u32 s22, s62811; GCN-NEXT:    s_cselect_b32 s25, -1, 02812; GCN-NEXT:    s_cmp_eq_u32 s23, s72813; GCN-NEXT:    s_cselect_b32 s24, s25, s242814; GCN-NEXT:    s_cmp_lg_u64 s[20:21], 02815; GCN-NEXT:    s_subb_u32 s16, s16, s72816; GCN-NEXT:    s_sub_u32 s20, s22, s62817; GCN-NEXT:    s_subb_u32 s16, s16, 02818; GCN-NEXT:    s_cmp_lg_u32 s24, 02819; GCN-NEXT:    s_cselect_b32 s20, s20, s222820; GCN-NEXT:    s_cselect_b32 s16, s16, s232821; GCN-NEXT:    s_cmp_lg_u64 s[18:19], 02822; GCN-NEXT:    s_subb_u32 s11, s17, s112823; GCN-NEXT:    s_cmp_ge_u32 s11, s72824; GCN-NEXT:    s_cselect_b32 s17, -1, 02825; GCN-NEXT:    s_cmp_ge_u32 s9, s62826; GCN-NEXT:    s_cselect_b32 s6, -1, 02827; GCN-NEXT:    s_cmp_eq_u32 s11, s72828; GCN-NEXT:    s_cselect_b32 s6, s6, s172829; GCN-NEXT:    s_cmp_lg_u32 s6, 02830; GCN-NEXT:    s_cselect_b32 s7, s16, s112831; GCN-NEXT:    s_cselect_b32 s6, s20, s92832; GCN-NEXT:    s_xor_b64 s[6:7], s[6:7], s[14:15]2833; GCN-NEXT:    s_sub_u32 s6, s6, s142834; GCN-NEXT:    s_subb_u32 s7, s7, s142835; GCN-NEXT:    s_cbranch_execnz .LBB10_32836; GCN-NEXT:  .LBB10_2:2837; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s82838; GCN-NEXT:    s_sub_i32 s6, 0, s82839; GCN-NEXT:    s_mov_b32 s7, 02840; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v02841; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v02842; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v02843; GCN-NEXT:    v_readfirstlane_b32 s9, v02844; GCN-NEXT:    s_mul_i32 s6, s6, s92845; GCN-NEXT:    s_mul_hi_u32 s6, s9, s62846; GCN-NEXT:    s_add_i32 s9, s9, s62847; GCN-NEXT:    s_mul_hi_u32 s6, s10, s92848; GCN-NEXT:    s_mul_i32 s6, s6, s82849; GCN-NEXT:    s_sub_i32 s6, s10, s62850; GCN-NEXT:    s_sub_i32 s9, s6, s82851; GCN-NEXT:    s_cmp_ge_u32 s6, s82852; GCN-NEXT:    s_cselect_b32 s6, s9, s62853; GCN-NEXT:    s_sub_i32 s9, s6, s82854; GCN-NEXT:    s_cmp_ge_u32 s6, s82855; GCN-NEXT:    s_cselect_b32 s6, s9, s62856; GCN-NEXT:  .LBB10_3:2857; GCN-NEXT:    s_or_b64 s[8:9], s[4:5], s[2:3]2858; GCN-NEXT:    s_mov_b32 s8, 02859; GCN-NEXT:    s_cmp_lg_u64 s[8:9], 02860; GCN-NEXT:    s_cbranch_scc0 .LBB10_72861; GCN-NEXT:  ; %bb.4:2862; GCN-NEXT:    s_ashr_i32 s8, s3, 312863; GCN-NEXT:    s_add_u32 s10, s2, s82864; GCN-NEXT:    s_mov_b32 s9, s82865; GCN-NEXT:    s_addc_u32 s11, s3, s82866; GCN-NEXT:    s_xor_b64 s[10:11], s[10:11], s[8:9]2867; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s102868; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s112869; GCN-NEXT:    s_sub_u32 s3, 0, s102870; GCN-NEXT:    s_subb_u32 s12, 0, s112871; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v02872; GCN-NEXT:    v_rcp_f32_e32 v0, v02873; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v02874; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v02875; GCN-NEXT:    v_trunc_f32_e32 v1, v12876; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v02877; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v12878; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v02879; GCN-NEXT:    v_readfirstlane_b32 s13, v12880; GCN-NEXT:    v_readfirstlane_b32 s14, v02881; GCN-NEXT:    s_mul_i32 s15, s3, s132882; GCN-NEXT:    s_mul_hi_u32 s17, s3, s142883; GCN-NEXT:    s_mul_i32 s16, s12, s142884; GCN-NEXT:    s_add_i32 s15, s17, s152885; GCN-NEXT:    s_add_i32 s15, s15, s162886; GCN-NEXT:    s_mul_i32 s18, s3, s142887; GCN-NEXT:    s_mul_i32 s17, s14, s152888; GCN-NEXT:    s_mul_hi_u32 s19, s14, s182889; GCN-NEXT:    s_mul_hi_u32 s16, s14, s152890; GCN-NEXT:    s_add_u32 s17, s19, s172891; GCN-NEXT:    s_addc_u32 s16, 0, s162892; GCN-NEXT:    s_mul_hi_u32 s20, s13, s182893; GCN-NEXT:    s_mul_i32 s18, s13, s182894; GCN-NEXT:    s_add_u32 s17, s17, s182895; GCN-NEXT:    s_mul_hi_u32 s19, s13, s152896; GCN-NEXT:    s_addc_u32 s16, s16, s202897; GCN-NEXT:    s_addc_u32 s17, s19, 02898; GCN-NEXT:    s_mul_i32 s15, s13, s152899; GCN-NEXT:    s_add_u32 s15, s16, s152900; GCN-NEXT:    s_addc_u32 s16, 0, s172901; GCN-NEXT:    s_add_u32 s14, s14, s152902; GCN-NEXT:    s_addc_u32 s13, s13, s162903; GCN-NEXT:    s_mul_i32 s15, s3, s132904; GCN-NEXT:    s_mul_hi_u32 s16, s3, s142905; GCN-NEXT:    s_add_i32 s15, s16, s152906; GCN-NEXT:    s_mul_i32 s12, s12, s142907; GCN-NEXT:    s_add_i32 s15, s15, s122908; GCN-NEXT:    s_mul_i32 s3, s3, s142909; GCN-NEXT:    s_mul_hi_u32 s16, s13, s32910; GCN-NEXT:    s_mul_i32 s17, s13, s32911; GCN-NEXT:    s_mul_i32 s19, s14, s152912; GCN-NEXT:    s_mul_hi_u32 s3, s14, s32913; GCN-NEXT:    s_mul_hi_u32 s18, s14, s152914; GCN-NEXT:    s_add_u32 s3, s3, s192915; GCN-NEXT:    s_addc_u32 s18, 0, s182916; GCN-NEXT:    s_add_u32 s3, s3, s172917; GCN-NEXT:    s_mul_hi_u32 s12, s13, s152918; GCN-NEXT:    s_addc_u32 s3, s18, s162919; GCN-NEXT:    s_addc_u32 s12, s12, 02920; GCN-NEXT:    s_mul_i32 s15, s13, s152921; GCN-NEXT:    s_add_u32 s3, s3, s152922; GCN-NEXT:    s_addc_u32 s12, 0, s122923; GCN-NEXT:    s_add_u32 s3, s14, s32924; GCN-NEXT:    s_addc_u32 s16, s13, s122925; GCN-NEXT:    s_ashr_i32 s12, s5, 312926; GCN-NEXT:    s_add_u32 s14, s4, s122927; GCN-NEXT:    s_mov_b32 s13, s122928; GCN-NEXT:    s_addc_u32 s15, s5, s122929; GCN-NEXT:    s_xor_b64 s[14:15], s[14:15], s[12:13]2930; GCN-NEXT:    s_mul_i32 s17, s14, s162931; GCN-NEXT:    s_mul_hi_u32 s18, s14, s32932; GCN-NEXT:    s_mul_hi_u32 s5, s14, s162933; GCN-NEXT:    s_add_u32 s17, s18, s172934; GCN-NEXT:    s_addc_u32 s5, 0, s52935; GCN-NEXT:    s_mul_hi_u32 s19, s15, s32936; GCN-NEXT:    s_mul_i32 s3, s15, s32937; GCN-NEXT:    s_add_u32 s3, s17, s32938; GCN-NEXT:    s_mul_hi_u32 s18, s15, s162939; GCN-NEXT:    s_addc_u32 s3, s5, s192940; GCN-NEXT:    s_addc_u32 s5, s18, 02941; GCN-NEXT:    s_mul_i32 s16, s15, s162942; GCN-NEXT:    s_add_u32 s3, s3, s162943; GCN-NEXT:    s_addc_u32 s5, 0, s52944; GCN-NEXT:    s_mul_i32 s5, s10, s52945; GCN-NEXT:    s_mul_hi_u32 s16, s10, s32946; GCN-NEXT:    s_add_i32 s5, s16, s52947; GCN-NEXT:    s_mul_i32 s16, s11, s32948; GCN-NEXT:    s_add_i32 s5, s5, s162949; GCN-NEXT:    s_sub_i32 s18, s15, s52950; GCN-NEXT:    s_mul_i32 s3, s10, s32951; GCN-NEXT:    s_sub_u32 s3, s14, s32952; GCN-NEXT:    s_cselect_b64 s[16:17], -1, 02953; GCN-NEXT:    s_subb_u32 s14, s18, s112954; GCN-NEXT:    s_sub_u32 s20, s3, s102955; GCN-NEXT:    s_cselect_b64 s[18:19], -1, 02956; GCN-NEXT:    s_subb_u32 s21, s14, 02957; GCN-NEXT:    s_cmp_ge_u32 s21, s112958; GCN-NEXT:    s_cselect_b32 s22, -1, 02959; GCN-NEXT:    s_cmp_ge_u32 s20, s102960; GCN-NEXT:    s_cselect_b32 s23, -1, 02961; GCN-NEXT:    s_cmp_eq_u32 s21, s112962; GCN-NEXT:    s_cselect_b32 s22, s23, s222963; GCN-NEXT:    s_cmp_lg_u64 s[18:19], 02964; GCN-NEXT:    s_subb_u32 s14, s14, s112965; GCN-NEXT:    s_sub_u32 s18, s20, s102966; GCN-NEXT:    s_subb_u32 s14, s14, 02967; GCN-NEXT:    s_cmp_lg_u32 s22, 02968; GCN-NEXT:    s_cselect_b32 s18, s18, s202969; GCN-NEXT:    s_cselect_b32 s14, s14, s212970; GCN-NEXT:    s_cmp_lg_u64 s[16:17], 02971; GCN-NEXT:    s_subb_u32 s5, s15, s52972; GCN-NEXT:    s_cmp_ge_u32 s5, s112973; GCN-NEXT:    s_cselect_b32 s15, -1, 02974; GCN-NEXT:    s_cmp_ge_u32 s3, s102975; GCN-NEXT:    s_cselect_b32 s10, -1, 02976; GCN-NEXT:    s_cmp_eq_u32 s5, s112977; GCN-NEXT:    s_cselect_b32 s10, s10, s152978; GCN-NEXT:    s_cmp_lg_u32 s10, 02979; GCN-NEXT:    s_cselect_b32 s11, s14, s52980; GCN-NEXT:    s_cselect_b32 s10, s18, s32981; GCN-NEXT:    s_xor_b64 s[10:11], s[10:11], s[12:13]2982; GCN-NEXT:    s_sub_u32 s10, s10, s122983; GCN-NEXT:    s_subb_u32 s11, s11, s122984; GCN-NEXT:    s_cbranch_execnz .LBB10_82985; GCN-NEXT:  .LBB10_5:2986; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s22987; GCN-NEXT:    s_sub_i32 s3, 0, s22988; GCN-NEXT:    v_mov_b32_e32 v3, 02989; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v02990; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v02991; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v02992; GCN-NEXT:    v_mul_lo_u32 v1, s3, v02993; GCN-NEXT:    v_mul_hi_u32 v1, v0, v12994; GCN-NEXT:    v_add_u32_e32 v0, v0, v12995; GCN-NEXT:    v_mul_hi_u32 v0, s4, v02996; GCN-NEXT:    v_mul_lo_u32 v0, v0, s22997; GCN-NEXT:    v_sub_u32_e32 v0, s4, v02998; GCN-NEXT:    v_subrev_u32_e32 v1, s2, v02999; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v03000; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc3001; GCN-NEXT:    v_subrev_u32_e32 v1, s2, v03002; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v03003; GCN-NEXT:    v_cndmask_b32_e32 v2, v0, v1, vcc3004; GCN-NEXT:    s_branch .LBB10_93005; GCN-NEXT:  .LBB10_6:3006; GCN-NEXT:    ; implicit-def: $sgpr6_sgpr73007; GCN-NEXT:    s_branch .LBB10_23008; GCN-NEXT:  .LBB10_7:3009; GCN-NEXT:    ; implicit-def: $sgpr10_sgpr113010; GCN-NEXT:    s_branch .LBB10_53011; GCN-NEXT:  .LBB10_8:3012; GCN-NEXT:    v_mov_b32_e32 v2, s103013; GCN-NEXT:    v_mov_b32_e32 v3, s113014; GCN-NEXT:  .LBB10_9:3015; GCN-NEXT:    v_mov_b32_e32 v4, 03016; GCN-NEXT:    v_mov_b32_e32 v0, s63017; GCN-NEXT:    v_mov_b32_e32 v1, s73018; GCN-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]3019; GCN-NEXT:    s_endpgm3020;3021; TAHITI-LABEL: srem_v2i64:3022; TAHITI:       ; %bb.0:3023; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x93024; TAHITI-NEXT:    s_mov_b32 s3, 0xf0003025; TAHITI-NEXT:    s_mov_b32 s2, -13026; TAHITI-NEXT:    v_mov_b32_e32 v8, 03027; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)3028; TAHITI-NEXT:    s_mov_b32 s0, s63029; TAHITI-NEXT:    s_mov_b32 s1, s73030; TAHITI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[0:3], 03031; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 0 offset:163032; TAHITI-NEXT:    s_waitcnt vmcnt(0)3033; TAHITI-NEXT:    v_or_b32_e32 v9, v5, v13034; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[8:9]3035; TAHITI-NEXT:    s_cbranch_vccz .LBB10_73036; TAHITI-NEXT:  ; %bb.1:3037; TAHITI-NEXT:    v_ashrrev_i32_e32 v9, 31, v13038; TAHITI-NEXT:    v_add_i32_e32 v8, vcc, v0, v93039; TAHITI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v9, vcc3040; TAHITI-NEXT:    v_xor_b32_e32 v8, v8, v93041; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v93042; TAHITI-NEXT:    v_cvt_f32_u32_e32 v9, v83043; TAHITI-NEXT:    v_cvt_f32_u32_e32 v10, v13044; TAHITI-NEXT:    v_sub_i32_e32 v11, vcc, 0, v83045; TAHITI-NEXT:    v_subb_u32_e32 v12, vcc, 0, v1, vcc3046; TAHITI-NEXT:    v_madmk_f32 v9, v10, 0x4f800000, v93047; TAHITI-NEXT:    v_rcp_f32_e32 v9, v93048; TAHITI-NEXT:    v_mul_f32_e32 v9, 0x5f7ffffc, v93049; TAHITI-NEXT:    v_mul_f32_e32 v10, 0x2f800000, v93050; TAHITI-NEXT:    v_trunc_f32_e32 v10, v103051; TAHITI-NEXT:    v_madmk_f32 v9, v10, 0xcf800000, v93052; TAHITI-NEXT:    v_cvt_u32_f32_e32 v10, v103053; TAHITI-NEXT:    v_cvt_u32_f32_e32 v9, v93054; TAHITI-NEXT:    v_mul_lo_u32 v14, v11, v103055; TAHITI-NEXT:    v_mul_hi_u32 v13, v11, v93056; TAHITI-NEXT:    v_mul_lo_u32 v15, v12, v93057; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v14, v133058; TAHITI-NEXT:    v_mul_lo_u32 v14, v11, v93059; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v153060; TAHITI-NEXT:    v_mul_lo_u32 v15, v9, v133061; TAHITI-NEXT:    v_mul_hi_u32 v16, v9, v143062; TAHITI-NEXT:    v_mul_hi_u32 v17, v9, v133063; TAHITI-NEXT:    v_mul_hi_u32 v18, v10, v133064; TAHITI-NEXT:    v_mul_lo_u32 v13, v10, v133065; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v16, v153066; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v17, vcc3067; TAHITI-NEXT:    v_mul_lo_u32 v17, v10, v143068; TAHITI-NEXT:    v_mul_hi_u32 v14, v10, v143069; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v173070; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, v16, v14, vcc3071; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v18, vcc3072; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v14, v133073; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc3074; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v133075; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, v10, v14, vcc3076; TAHITI-NEXT:    v_mul_lo_u32 v13, v11, v103077; TAHITI-NEXT:    v_mul_hi_u32 v14, v11, v93078; TAHITI-NEXT:    v_mul_lo_u32 v12, v12, v93079; TAHITI-NEXT:    v_mul_lo_u32 v11, v11, v93080; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v143081; TAHITI-NEXT:    v_add_i32_e32 v12, vcc, v12, v133082; TAHITI-NEXT:    v_mul_lo_u32 v15, v9, v123083; TAHITI-NEXT:    v_mul_hi_u32 v16, v9, v113084; TAHITI-NEXT:    v_mul_hi_u32 v17, v9, v123085; TAHITI-NEXT:    v_mul_hi_u32 v14, v10, v113086; TAHITI-NEXT:    v_mul_lo_u32 v11, v10, v113087; TAHITI-NEXT:    v_mul_hi_u32 v13, v10, v123088; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v16, v153089; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v17, vcc3090; TAHITI-NEXT:    v_mul_lo_u32 v12, v10, v123091; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v15, v113092; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, v16, v14, vcc3093; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc3094; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v123095; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc3096; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v113097; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, v10, v12, vcc3098; TAHITI-NEXT:    v_ashrrev_i32_e32 v11, 31, v53099; TAHITI-NEXT:    v_add_i32_e32 v12, vcc, v4, v113100; TAHITI-NEXT:    v_xor_b32_e32 v12, v12, v113101; TAHITI-NEXT:    v_mul_lo_u32 v13, v12, v103102; TAHITI-NEXT:    v_mul_hi_u32 v14, v12, v93103; TAHITI-NEXT:    v_mul_hi_u32 v15, v12, v103104; TAHITI-NEXT:    v_addc_u32_e32 v5, vcc, v5, v11, vcc3105; TAHITI-NEXT:    v_xor_b32_e32 v5, v5, v113106; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v14, v133107; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc3108; TAHITI-NEXT:    v_mul_lo_u32 v15, v5, v93109; TAHITI-NEXT:    v_mul_hi_u32 v9, v5, v93110; TAHITI-NEXT:    v_mul_hi_u32 v16, v5, v103111; TAHITI-NEXT:    v_mul_lo_u32 v10, v5, v103112; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v153113; TAHITI-NEXT:    v_addc_u32_e32 v9, vcc, v14, v9, vcc3114; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v16, vcc3115; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v9, v103116; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v13, vcc3117; TAHITI-NEXT:    v_mul_lo_u32 v10, v8, v103118; TAHITI-NEXT:    v_mul_hi_u32 v13, v8, v93119; TAHITI-NEXT:    v_mul_lo_u32 v14, v1, v93120; TAHITI-NEXT:    v_mul_lo_u32 v9, v8, v93121; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v10, v133122; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v14, v103123; TAHITI-NEXT:    v_sub_i32_e32 v13, vcc, v5, v103124; TAHITI-NEXT:    v_sub_i32_e32 v9, vcc, v12, v93125; TAHITI-NEXT:    v_subb_u32_e64 v12, s[0:1], v13, v1, vcc3126; TAHITI-NEXT:    v_sub_i32_e64 v13, s[0:1], v9, v83127; TAHITI-NEXT:    v_subbrev_u32_e64 v14, s[2:3], 0, v12, s[0:1]3128; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v14, v13129; TAHITI-NEXT:    v_cndmask_b32_e64 v15, 0, -1, s[2:3]3130; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v13, v83131; TAHITI-NEXT:    v_subb_u32_e32 v5, vcc, v5, v10, vcc3132; TAHITI-NEXT:    v_cndmask_b32_e64 v16, 0, -1, s[2:3]3133; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v14, v13134; TAHITI-NEXT:    v_subb_u32_e64 v12, s[0:1], v12, v1, s[0:1]3135; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v13136; TAHITI-NEXT:    v_cndmask_b32_e64 v15, v15, v16, s[2:3]3137; TAHITI-NEXT:    v_sub_i32_e64 v16, s[0:1], v13, v83138; TAHITI-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc3139; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v83140; TAHITI-NEXT:    v_subbrev_u32_e64 v12, s[0:1], 0, v12, s[0:1]3141; TAHITI-NEXT:    v_cndmask_b32_e64 v8, 0, -1, vcc3142; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v13143; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v153144; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc3145; TAHITI-NEXT:    v_cndmask_b32_e64 v13, v13, v16, s[0:1]3146; TAHITI-NEXT:    v_cndmask_b32_e64 v12, v14, v12, s[0:1]3147; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v13148; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v5, v12, vcc3149; TAHITI-NEXT:    v_cndmask_b32_e32 v5, v9, v13, vcc3150; TAHITI-NEXT:    v_xor_b32_e32 v5, v5, v113151; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v113152; TAHITI-NEXT:    v_sub_i32_e32 v8, vcc, v5, v113153; TAHITI-NEXT:    v_subb_u32_e32 v9, vcc, v1, v11, vcc3154; TAHITI-NEXT:    s_cbranch_execnz .LBB10_33155; TAHITI-NEXT:  .LBB10_2:3156; TAHITI-NEXT:    v_cvt_f32_u32_e32 v1, v03157; TAHITI-NEXT:    v_sub_i32_e32 v5, vcc, 0, v03158; TAHITI-NEXT:    v_mov_b32_e32 v9, 03159; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v1, v13160; TAHITI-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v13161; TAHITI-NEXT:    v_cvt_u32_f32_e32 v1, v13162; TAHITI-NEXT:    v_mul_lo_u32 v5, v5, v13163; TAHITI-NEXT:    v_mul_hi_u32 v5, v1, v53164; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v53165; TAHITI-NEXT:    v_mul_hi_u32 v1, v4, v13166; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, v03167; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, v4, v13168; TAHITI-NEXT:    v_sub_i32_e32 v4, vcc, v1, v03169; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v03170; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc3171; TAHITI-NEXT:    v_sub_i32_e32 v4, vcc, v1, v03172; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v03173; TAHITI-NEXT:    v_cndmask_b32_e32 v8, v1, v4, vcc3174; TAHITI-NEXT:  .LBB10_3:3175; TAHITI-NEXT:    v_or_b32_e32 v1, v7, v33176; TAHITI-NEXT:    v_mov_b32_e32 v0, 03177; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[0:1]3178; TAHITI-NEXT:    s_cbranch_vccz .LBB10_83179; TAHITI-NEXT:  ; %bb.4:3180; TAHITI-NEXT:    v_ashrrev_i32_e32 v0, 31, v33181; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v2, v03182; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, v3, v0, vcc3183; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v03184; TAHITI-NEXT:    v_xor_b32_e32 v0, v3, v03185; TAHITI-NEXT:    v_cvt_f32_u32_e32 v3, v13186; TAHITI-NEXT:    v_cvt_f32_u32_e32 v4, v03187; TAHITI-NEXT:    v_sub_i32_e32 v5, vcc, 0, v13188; TAHITI-NEXT:    v_subb_u32_e32 v10, vcc, 0, v0, vcc3189; TAHITI-NEXT:    v_madmk_f32 v3, v4, 0x4f800000, v33190; TAHITI-NEXT:    v_rcp_f32_e32 v3, v33191; TAHITI-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v33192; TAHITI-NEXT:    v_mul_f32_e32 v4, 0x2f800000, v33193; TAHITI-NEXT:    v_trunc_f32_e32 v4, v43194; TAHITI-NEXT:    v_madmk_f32 v3, v4, 0xcf800000, v33195; TAHITI-NEXT:    v_cvt_u32_f32_e32 v4, v43196; TAHITI-NEXT:    v_cvt_u32_f32_e32 v3, v33197; TAHITI-NEXT:    v_mul_lo_u32 v12, v5, v43198; TAHITI-NEXT:    v_mul_hi_u32 v11, v5, v33199; TAHITI-NEXT:    v_mul_lo_u32 v13, v10, v33200; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v113201; TAHITI-NEXT:    v_mul_lo_u32 v12, v5, v33202; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v133203; TAHITI-NEXT:    v_mul_lo_u32 v13, v3, v113204; TAHITI-NEXT:    v_mul_hi_u32 v14, v3, v123205; TAHITI-NEXT:    v_mul_hi_u32 v15, v3, v113206; TAHITI-NEXT:    v_mul_hi_u32 v16, v4, v113207; TAHITI-NEXT:    v_mul_lo_u32 v11, v4, v113208; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v14, v133209; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc3210; TAHITI-NEXT:    v_mul_lo_u32 v15, v4, v123211; TAHITI-NEXT:    v_mul_hi_u32 v12, v4, v123212; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v153213; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, v14, v12, vcc3214; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v16, vcc3215; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v113216; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc3217; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v113218; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, v4, v12, vcc3219; TAHITI-NEXT:    v_mul_lo_u32 v11, v5, v43220; TAHITI-NEXT:    v_mul_hi_u32 v12, v5, v33221; TAHITI-NEXT:    v_mul_lo_u32 v10, v10, v33222; TAHITI-NEXT:    v_mul_lo_u32 v5, v5, v33223; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v123224; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v10, v113225; TAHITI-NEXT:    v_mul_lo_u32 v13, v3, v103226; TAHITI-NEXT:    v_mul_hi_u32 v14, v3, v53227; TAHITI-NEXT:    v_mul_hi_u32 v15, v3, v103228; TAHITI-NEXT:    v_mul_hi_u32 v12, v4, v53229; TAHITI-NEXT:    v_mul_lo_u32 v5, v4, v53230; TAHITI-NEXT:    v_mul_hi_u32 v11, v4, v103231; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v14, v133232; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc3233; TAHITI-NEXT:    v_mul_lo_u32 v10, v4, v103234; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v13, v53235; TAHITI-NEXT:    v_addc_u32_e32 v5, vcc, v14, v12, vcc3236; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc3237; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v5, v103238; TAHITI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v11, vcc3239; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v53240; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, v4, v10, vcc3241; TAHITI-NEXT:    v_ashrrev_i32_e32 v5, 31, v73242; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v6, v53243; TAHITI-NEXT:    v_xor_b32_e32 v10, v10, v53244; TAHITI-NEXT:    v_mul_lo_u32 v11, v10, v43245; TAHITI-NEXT:    v_mul_hi_u32 v12, v10, v33246; TAHITI-NEXT:    v_mul_hi_u32 v13, v10, v43247; TAHITI-NEXT:    v_addc_u32_e32 v7, vcc, v7, v5, vcc3248; TAHITI-NEXT:    v_xor_b32_e32 v7, v7, v53249; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v113250; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc3251; TAHITI-NEXT:    v_mul_lo_u32 v13, v7, v33252; TAHITI-NEXT:    v_mul_hi_u32 v3, v7, v33253; TAHITI-NEXT:    v_mul_hi_u32 v14, v7, v43254; TAHITI-NEXT:    v_mul_lo_u32 v4, v7, v43255; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v133256; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, v12, v3, vcc3257; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v14, vcc3258; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v43259; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, 0, v11, vcc3260; TAHITI-NEXT:    v_mul_lo_u32 v4, v1, v43261; TAHITI-NEXT:    v_mul_hi_u32 v11, v1, v33262; TAHITI-NEXT:    v_mul_lo_u32 v12, v0, v33263; TAHITI-NEXT:    v_mul_lo_u32 v3, v1, v33264; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v4, v113265; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v12, v43266; TAHITI-NEXT:    v_sub_i32_e32 v11, vcc, v7, v43267; TAHITI-NEXT:    v_sub_i32_e32 v3, vcc, v10, v33268; TAHITI-NEXT:    v_subb_u32_e64 v10, s[0:1], v11, v0, vcc3269; TAHITI-NEXT:    v_sub_i32_e64 v11, s[0:1], v3, v13270; TAHITI-NEXT:    v_subbrev_u32_e64 v12, s[2:3], 0, v10, s[0:1]3271; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v12, v03272; TAHITI-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[2:3]3273; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v11, v13274; TAHITI-NEXT:    v_subb_u32_e32 v4, vcc, v7, v4, vcc3275; TAHITI-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[2:3]3276; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v12, v03277; TAHITI-NEXT:    v_subb_u32_e64 v10, s[0:1], v10, v0, s[0:1]3278; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v4, v03279; TAHITI-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[2:3]3280; TAHITI-NEXT:    v_sub_i32_e64 v14, s[0:1], v11, v13281; TAHITI-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc3282; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v13283; TAHITI-NEXT:    v_subbrev_u32_e64 v10, s[0:1], 0, v10, s[0:1]3284; TAHITI-NEXT:    v_cndmask_b32_e64 v1, 0, -1, vcc3285; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v03286; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v133287; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v7, v1, vcc3288; TAHITI-NEXT:    v_cndmask_b32_e64 v11, v11, v14, s[0:1]3289; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v03290; TAHITI-NEXT:    v_cndmask_b32_e64 v10, v12, v10, s[0:1]3291; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v3, v11, vcc3292; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v4, v10, vcc3293; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v53294; TAHITI-NEXT:    v_xor_b32_e32 v0, v0, v53295; TAHITI-NEXT:    v_sub_i32_e32 v10, vcc, v1, v53296; TAHITI-NEXT:    v_subb_u32_e32 v11, vcc, v0, v5, vcc3297; TAHITI-NEXT:    s_cbranch_execnz .LBB10_63298; TAHITI-NEXT:  .LBB10_5:3299; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, v23300; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, 0, v23301; TAHITI-NEXT:    v_mov_b32_e32 v11, 03302; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v03303; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v03304; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v03305; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, v03306; TAHITI-NEXT:    v_mul_hi_u32 v1, v0, v13307; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v13308; TAHITI-NEXT:    v_mul_hi_u32 v0, v6, v03309; TAHITI-NEXT:    v_mul_lo_u32 v0, v0, v23310; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v6, v03311; TAHITI-NEXT:    v_subrev_i32_e32 v1, vcc, v2, v03312; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v23313; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc3314; TAHITI-NEXT:    v_subrev_i32_e32 v1, vcc, v2, v03315; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v23316; TAHITI-NEXT:    v_cndmask_b32_e32 v10, v0, v1, vcc3317; TAHITI-NEXT:  .LBB10_6:3318; TAHITI-NEXT:    s_mov_b32 s7, 0xf0003319; TAHITI-NEXT:    s_mov_b32 s6, -13320; TAHITI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 03321; TAHITI-NEXT:    s_endpgm3322; TAHITI-NEXT:  .LBB10_7:3323; TAHITI-NEXT:    ; implicit-def: $vgpr8_vgpr93324; TAHITI-NEXT:    s_branch .LBB10_23325; TAHITI-NEXT:  .LBB10_8:3326; TAHITI-NEXT:    s_branch .LBB10_53327;3328; TONGA-LABEL: srem_v2i64:3329; TONGA:       ; %bb.0:3330; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x243331; TONGA-NEXT:    s_waitcnt lgkmcnt(0)3332; TONGA-NEXT:    s_add_u32 s0, s6, 163333; TONGA-NEXT:    s_addc_u32 s1, s7, 03334; TONGA-NEXT:    v_mov_b32_e32 v0, s03335; TONGA-NEXT:    v_mov_b32_e32 v4, s63336; TONGA-NEXT:    v_mov_b32_e32 v1, s13337; TONGA-NEXT:    v_mov_b32_e32 v5, s73338; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]3339; TONGA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]3340; TONGA-NEXT:    s_waitcnt vmcnt(1)3341; TONGA-NEXT:    v_readfirstlane_b32 s1, v13342; TONGA-NEXT:    v_readfirstlane_b32 s0, v03343; TONGA-NEXT:    s_waitcnt vmcnt(0)3344; TONGA-NEXT:    v_readfirstlane_b32 s3, v53345; TONGA-NEXT:    v_readfirstlane_b32 s2, v43346; TONGA-NEXT:    s_or_b64 s[6:7], s[2:3], s[0:1]3347; TONGA-NEXT:    s_mov_b32 s6, 03348; TONGA-NEXT:    s_cmp_lg_u64 s[6:7], 03349; TONGA-NEXT:    s_cbranch_scc0 .LBB10_33350; TONGA-NEXT:  ; %bb.1:3351; TONGA-NEXT:    s_ashr_i32 s6, s1, 313352; TONGA-NEXT:    s_add_u32 s8, s0, s63353; TONGA-NEXT:    s_mov_b32 s7, s63354; TONGA-NEXT:    s_addc_u32 s9, s1, s63355; TONGA-NEXT:    s_xor_b64 s[6:7], s[8:9], s[6:7]3356; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s63357; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, s73358; TONGA-NEXT:    s_sub_u32 s1, 0, s63359; TONGA-NEXT:    s_subb_u32 s10, 0, s73360; TONGA-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v03361; TONGA-NEXT:    v_rcp_f32_e32 v0, v03362; TONGA-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v03363; TONGA-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v03364; TONGA-NEXT:    v_trunc_f32_e32 v1, v13365; TONGA-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v03366; TONGA-NEXT:    v_cvt_u32_f32_e32 v8, v13367; TONGA-NEXT:    v_cvt_u32_f32_e32 v9, v03368; TONGA-NEXT:    v_mul_lo_u32 v4, s1, v83369; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s1, v9, 03370; TONGA-NEXT:    v_mul_lo_u32 v5, s10, v93371; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v1, v43372; TONGA-NEXT:    v_add_u32_e32 v11, vcc, v1, v53373; TONGA-NEXT:    v_mul_hi_u32 v10, v9, v03374; TONGA-NEXT:    v_mad_u64_u32 v[4:5], s[8:9], v9, v11, 03375; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], v8, v0, 03376; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v10, v43377; TONGA-NEXT:    v_addc_u32_e32 v12, vcc, 0, v5, vcc3378; TONGA-NEXT:    v_mad_u64_u32 v[4:5], s[8:9], v8, v11, 03379; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v10, v03380; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v12, v1, vcc3381; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v5, vcc3382; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v43383; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3384; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v9, v03385; TONGA-NEXT:    v_addc_u32_e32 v11, vcc, v8, v1, vcc3386; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s1, v10, 03387; TONGA-NEXT:    v_mul_lo_u32 v8, s1, v113388; TONGA-NEXT:    v_mul_lo_u32 v9, s10, v103389; TONGA-NEXT:    v_mul_hi_u32 v12, v10, v03390; TONGA-NEXT:    v_mad_u64_u32 v[4:5], s[8:9], v11, v0, 03391; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v8, v13392; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v9, v13393; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], v10, v1, 03394; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], v11, v1, 03395; TONGA-NEXT:    v_add_u32_e32 v8, vcc, v12, v83396; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc3397; TONGA-NEXT:    v_add_u32_e32 v4, vcc, v8, v43398; TONGA-NEXT:    v_addc_u32_e32 v4, vcc, v9, v5, vcc3399; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3400; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v4, v03401; TONGA-NEXT:    s_ashr_i32 s10, s3, 313402; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3403; TONGA-NEXT:    s_add_u32 s8, s2, s103404; TONGA-NEXT:    v_add_u32_e32 v4, vcc, v10, v03405; TONGA-NEXT:    s_mov_b32 s11, s103406; TONGA-NEXT:    s_addc_u32 s9, s3, s103407; TONGA-NEXT:    v_addc_u32_e32 v5, vcc, v11, v1, vcc3408; TONGA-NEXT:    s_xor_b64 s[12:13], s[8:9], s[10:11]3409; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s12, v5, 03410; TONGA-NEXT:    v_mul_hi_u32 v8, s12, v43411; TONGA-NEXT:    v_readfirstlane_b32 s1, v13412; TONGA-NEXT:    v_readfirstlane_b32 s3, v03413; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s13, v5, 03414; TONGA-NEXT:    v_mad_u64_u32 v[4:5], s[8:9], s13, v4, 03415; TONGA-NEXT:    v_readfirstlane_b32 s14, v83416; TONGA-NEXT:    s_add_u32 s3, s14, s33417; TONGA-NEXT:    s_addc_u32 s1, 0, s13418; TONGA-NEXT:    v_readfirstlane_b32 s14, v43419; TONGA-NEXT:    v_readfirstlane_b32 s9, v53420; TONGA-NEXT:    s_add_u32 s3, s3, s143421; TONGA-NEXT:    v_readfirstlane_b32 s8, v13422; TONGA-NEXT:    s_addc_u32 s1, s1, s93423; TONGA-NEXT:    s_addc_u32 s3, s8, 03424; TONGA-NEXT:    v_readfirstlane_b32 s8, v03425; TONGA-NEXT:    s_add_u32 s1, s1, s83426; TONGA-NEXT:    v_mov_b32_e32 v0, s13427; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[8:9], s6, v0, 03428; TONGA-NEXT:    s_addc_u32 s3, 0, s33429; TONGA-NEXT:    s_mul_i32 s3, s6, s33430; TONGA-NEXT:    v_readfirstlane_b32 s14, v13431; TONGA-NEXT:    s_add_i32 s3, s14, s33432; TONGA-NEXT:    s_mul_i32 s1, s7, s13433; TONGA-NEXT:    s_add_i32 s3, s3, s13434; TONGA-NEXT:    s_sub_i32 s1, s13, s33435; TONGA-NEXT:    v_readfirstlane_b32 s14, v03436; TONGA-NEXT:    s_sub_u32 s12, s12, s143437; TONGA-NEXT:    s_cselect_b64 s[14:15], -1, 03438; TONGA-NEXT:    s_subb_u32 s1, s1, s73439; TONGA-NEXT:    s_sub_u32 s18, s12, s63440; TONGA-NEXT:    s_cselect_b64 s[16:17], -1, 03441; TONGA-NEXT:    s_subb_u32 s19, s1, 03442; TONGA-NEXT:    s_cmp_ge_u32 s19, s73443; TONGA-NEXT:    s_cselect_b32 s20, -1, 03444; TONGA-NEXT:    s_cmp_ge_u32 s18, s63445; TONGA-NEXT:    s_cselect_b32 s21, -1, 03446; TONGA-NEXT:    s_cmp_eq_u32 s19, s73447; TONGA-NEXT:    s_cselect_b32 s20, s21, s203448; TONGA-NEXT:    s_cmp_lg_u64 s[16:17], 03449; TONGA-NEXT:    s_subb_u32 s1, s1, s73450; TONGA-NEXT:    s_sub_u32 s16, s18, s63451; TONGA-NEXT:    s_subb_u32 s1, s1, 03452; TONGA-NEXT:    s_cmp_lg_u32 s20, 03453; TONGA-NEXT:    s_cselect_b32 s16, s16, s183454; TONGA-NEXT:    s_cselect_b32 s1, s1, s193455; TONGA-NEXT:    s_cmp_lg_u64 s[14:15], 03456; TONGA-NEXT:    s_subb_u32 s3, s13, s33457; TONGA-NEXT:    s_cmp_ge_u32 s3, s73458; TONGA-NEXT:    s_cselect_b32 s13, -1, 03459; TONGA-NEXT:    s_cmp_ge_u32 s12, s63460; TONGA-NEXT:    s_cselect_b32 s6, -1, 03461; TONGA-NEXT:    s_cmp_eq_u32 s3, s73462; TONGA-NEXT:    s_cselect_b32 s6, s6, s133463; TONGA-NEXT:    s_cmp_lg_u32 s6, 03464; TONGA-NEXT:    s_cselect_b32 s7, s1, s33465; TONGA-NEXT:    s_cselect_b32 s6, s16, s123466; TONGA-NEXT:    s_xor_b64 s[6:7], s[6:7], s[10:11]3467; TONGA-NEXT:    s_sub_u32 s6, s6, s103468; TONGA-NEXT:    s_subb_u32 s7, s7, s103469; TONGA-NEXT:    s_cbranch_execnz .LBB10_43470; TONGA-NEXT:  .LBB10_2:3471; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, s03472; TONGA-NEXT:    s_sub_i32 s1, 0, s03473; TONGA-NEXT:    v_mov_b32_e32 v9, 03474; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v03475; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v03476; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v03477; TONGA-NEXT:    v_mul_lo_u32 v1, s1, v03478; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v13479; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v13480; TONGA-NEXT:    v_mul_hi_u32 v0, s2, v03481; TONGA-NEXT:    v_mul_lo_u32 v0, v0, s03482; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, s2, v03483; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v03484; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s0, v03485; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc3486; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, s0, v03487; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s0, v03488; TONGA-NEXT:    v_cndmask_b32_e32 v8, v0, v1, vcc3489; TONGA-NEXT:    s_branch .LBB10_53490; TONGA-NEXT:  .LBB10_3:3491; TONGA-NEXT:    ; implicit-def: $sgpr6_sgpr73492; TONGA-NEXT:    s_branch .LBB10_23493; TONGA-NEXT:  .LBB10_4:3494; TONGA-NEXT:    v_mov_b32_e32 v9, s73495; TONGA-NEXT:    v_mov_b32_e32 v8, s63496; TONGA-NEXT:  .LBB10_5:3497; TONGA-NEXT:    v_or_b32_e32 v1, v7, v33498; TONGA-NEXT:    v_mov_b32_e32 v0, 03499; TONGA-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[0:1]3500; TONGA-NEXT:    s_cbranch_vccz .LBB10_93501; TONGA-NEXT:  ; %bb.6:3502; TONGA-NEXT:    v_ashrrev_i32_e32 v0, 31, v33503; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v2, v03504; TONGA-NEXT:    v_addc_u32_e32 v3, vcc, v3, v0, vcc3505; TONGA-NEXT:    v_xor_b32_e32 v5, v1, v03506; TONGA-NEXT:    v_xor_b32_e32 v12, v3, v03507; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, v53508; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, v123509; TONGA-NEXT:    v_sub_u32_e32 v13, vcc, 0, v53510; TONGA-NEXT:    v_subb_u32_e32 v14, vcc, 0, v12, vcc3511; TONGA-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v03512; TONGA-NEXT:    v_rcp_f32_e32 v0, v03513; TONGA-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v03514; TONGA-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v03515; TONGA-NEXT:    v_trunc_f32_e32 v1, v13516; TONGA-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v03517; TONGA-NEXT:    v_cvt_u32_f32_e32 v10, v13518; TONGA-NEXT:    v_cvt_u32_f32_e32 v11, v03519; TONGA-NEXT:    v_mul_lo_u32 v3, v13, v103520; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v13, v11, 03521; TONGA-NEXT:    v_mul_lo_u32 v4, v14, v113522; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v1, v33523; TONGA-NEXT:    v_add_u32_e32 v15, vcc, v1, v43524; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v11, v15, 03525; TONGA-NEXT:    v_mul_hi_u32 v1, v11, v03526; TONGA-NEXT:    v_add_u32_e32 v16, vcc, v1, v33527; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v10, v0, 03528; TONGA-NEXT:    v_addc_u32_e32 v17, vcc, 0, v4, vcc3529; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v10, v15, 03530; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v16, v03531; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v17, v1, vcc3532; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v4, vcc3533; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v33534; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3535; TONGA-NEXT:    v_add_u32_e32 v15, vcc, v11, v03536; TONGA-NEXT:    v_addc_u32_e32 v16, vcc, v10, v1, vcc3537; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v13, v15, 03538; TONGA-NEXT:    v_mul_lo_u32 v10, v13, v163539; TONGA-NEXT:    v_mul_lo_u32 v11, v14, v153540; TONGA-NEXT:    v_mul_hi_u32 v13, v15, v03541; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v16, v0, 03542; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v10, v13543; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v11, v13544; TONGA-NEXT:    v_mad_u64_u32 v[10:11], s[0:1], v15, v1, 03545; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v16, v1, 03546; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v13, v103547; TONGA-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc3548; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v10, v33549; TONGA-NEXT:    v_addc_u32_e32 v3, vcc, v11, v4, vcc3550; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3551; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v3, v03552; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc3553; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v15, v03554; TONGA-NEXT:    v_addc_u32_e32 v4, vcc, v16, v1, vcc3555; TONGA-NEXT:    v_ashrrev_i32_e32 v11, 31, v73556; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v6, v113557; TONGA-NEXT:    v_xor_b32_e32 v10, v0, v113558; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v10, v4, 03559; TONGA-NEXT:    v_mul_hi_u32 v13, v10, v33560; TONGA-NEXT:    v_addc_u32_e32 v7, vcc, v7, v11, vcc3561; TONGA-NEXT:    v_xor_b32_e32 v7, v7, v113562; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v03563; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v1, vcc3564; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v7, v3, 03565; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v7, v4, 03566; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v13, v03567; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v14, v1, vcc3568; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v4, vcc3569; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v0, v33570; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, 0, v1, vcc3571; TONGA-NEXT:    v_mul_lo_u32 v4, v5, v03572; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v5, v3, 03573; TONGA-NEXT:    v_mul_lo_u32 v3, v12, v33574; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v4, v13575; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v3, v13576; TONGA-NEXT:    v_sub_u32_e32 v3, vcc, v7, v13577; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v10, v03578; TONGA-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v12, vcc3579; TONGA-NEXT:    v_sub_u32_e64 v4, s[0:1], v0, v53580; TONGA-NEXT:    v_subbrev_u32_e64 v10, s[2:3], 0, v3, s[0:1]3581; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v10, v123582; TONGA-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[2:3]3583; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v4, v53584; TONGA-NEXT:    v_subb_u32_e32 v1, vcc, v7, v1, vcc3585; TONGA-NEXT:    v_cndmask_b32_e64 v14, 0, -1, s[2:3]3586; TONGA-NEXT:    v_cmp_eq_u32_e64 s[2:3], v10, v123587; TONGA-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v12, s[0:1]3588; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v123589; TONGA-NEXT:    v_cndmask_b32_e64 v13, v13, v14, s[2:3]3590; TONGA-NEXT:    v_sub_u32_e64 v14, s[0:1], v4, v53591; TONGA-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc3592; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v53593; TONGA-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]3594; TONGA-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc3595; TONGA-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v123596; TONGA-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v133597; TONGA-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc3598; TONGA-NEXT:    v_cndmask_b32_e64 v4, v4, v14, s[0:1]3599; TONGA-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v53600; TONGA-NEXT:    v_cndmask_b32_e64 v3, v10, v3, s[0:1]3601; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc3602; TONGA-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc3603; TONGA-NEXT:    v_xor_b32_e32 v0, v0, v113604; TONGA-NEXT:    v_xor_b32_e32 v1, v1, v113605; TONGA-NEXT:    v_sub_u32_e32 v10, vcc, v0, v113606; TONGA-NEXT:    v_subb_u32_e32 v11, vcc, v1, v11, vcc3607; TONGA-NEXT:    s_cbranch_execnz .LBB10_83608; TONGA-NEXT:  .LBB10_7:3609; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, v23610; TONGA-NEXT:    v_sub_u32_e32 v1, vcc, 0, v23611; TONGA-NEXT:    v_mov_b32_e32 v11, 03612; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v03613; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v03614; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v03615; TONGA-NEXT:    v_mul_lo_u32 v1, v1, v03616; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v13617; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v13618; TONGA-NEXT:    v_mul_hi_u32 v0, v6, v03619; TONGA-NEXT:    v_mul_lo_u32 v0, v0, v23620; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v6, v03621; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, v2, v03622; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v23623; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc3624; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, v2, v03625; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v23626; TONGA-NEXT:    v_cndmask_b32_e32 v10, v0, v1, vcc3627; TONGA-NEXT:  .LBB10_8:3628; TONGA-NEXT:    v_mov_b32_e32 v0, s43629; TONGA-NEXT:    v_mov_b32_e32 v1, s53630; TONGA-NEXT:    flat_store_dwordx4 v[0:1], v[8:11]3631; TONGA-NEXT:    s_endpgm3632; TONGA-NEXT:  .LBB10_9:3633; TONGA-NEXT:    s_branch .LBB10_73634;3635; EG-LABEL: srem_v2i64:3636; EG:       ; %bb.0:3637; EG-NEXT:    ALU 0, @18, KC0[CB0:0-32], KC1[]3638; EG-NEXT:    TEX 1 @143639; EG-NEXT:    ALU 112, @19, KC0[], KC1[]3640; EG-NEXT:    ALU 111, @132, KC0[], KC1[]3641; EG-NEXT:    ALU 112, @244, KC0[], KC1[]3642; EG-NEXT:    ALU 111, @357, KC0[], KC1[]3643; EG-NEXT:    ALU 112, @469, KC0[], KC1[]3644; EG-NEXT:    ALU 112, @582, KC0[], KC1[]3645; EG-NEXT:    ALU 111, @695, KC0[], KC1[]3646; EG-NEXT:    ALU 112, @807, KC0[], KC1[]3647; EG-NEXT:    ALU 112, @920, KC0[], KC1[]3648; EG-NEXT:    ALU 47, @1033, KC0[CB0:0-32], KC1[]3649; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T0.X, 13650; EG-NEXT:    CF_END3651; EG-NEXT:    Fetch clause starting at 14:3652; EG-NEXT:     VTX_READ_128 T4.XYZW, T0.X, 16, #13653; EG-NEXT:     VTX_READ_128 T1.XYZW, T0.X, 0, #13654; EG-NEXT:    ALU clause starting at 18:3655; EG-NEXT:     MOV * T0.X, KC0[2].Z,3656; EG-NEXT:    ALU clause starting at 19:3657; EG-NEXT:     ASHR * T5.W, T4.Y, literal.x,3658; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3659; EG-NEXT:     ADD_INT T0.W, T4.X, PV.W,3660; EG-NEXT:     ASHR * T6.W, T4.W, literal.x,3661; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3662; EG-NEXT:     XOR_INT * T2.W, PV.W, T5.W,3663; EG-NEXT:     SUB_INT T0.Z, 0.0, PV.W,3664; EG-NEXT:     ADD_INT T0.W, T4.Z, T6.W,3665; EG-NEXT:     RECIP_UINT * T0.X, PV.W,3666; EG-NEXT:     XOR_INT T3.W, PV.W, T6.W,3667; EG-NEXT:     MULLO_INT * T0.Y, PV.Z, PS,3668; EG-NEXT:     SUB_INT T0.Z, 0.0, PV.W,3669; EG-NEXT:     ASHR T0.W, T1.W, literal.x,3670; EG-NEXT:     RECIP_UINT * T2.X, PV.W,3671; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3672; EG-NEXT:     ADD_INT T2.Y, T1.W, PV.W,3673; EG-NEXT:     ADDC_UINT T2.Z, T1.Z, PV.W,3674; EG-NEXT:     ASHR T1.W, T1.Y, literal.x,3675; EG-NEXT:     MULLO_INT * T0.Z, PV.Z, PS,3676; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3677; EG-NEXT:     ADD_INT T1.Y, T1.Y, PV.W,3678; EG-NEXT:     ADDC_UINT T3.Z, T1.X, PV.W,3679; EG-NEXT:     ADD_INT T7.W, PV.Y, PV.Z,3680; EG-NEXT:     MULHI * T0.Z, T2.X, PS,3681; EG-NEXT:     ADD_INT T2.Y, T2.X, PS,3682; EG-NEXT:     XOR_INT T0.Z, PV.W, T0.W,3683; EG-NEXT:     ADD_INT T7.W, PV.Y, PV.Z,3684; EG-NEXT:     MULHI * T0.Y, T0.X, T0.Y,3685; EG-NEXT:     ADD_INT T2.Z, T0.X, PS,3686; EG-NEXT:     XOR_INT T7.W, PV.W, T1.W,3687; EG-NEXT:     MULHI * T0.X, PV.Z, PV.Y,3688; EG-NEXT:     MULHI * T0.Y, PV.W, PV.Z,3689; EG-NEXT:     MULLO_INT * T0.Y, PS, T2.W,3690; EG-NEXT:     SUB_INT T8.W, T7.W, PS,3691; EG-NEXT:     MULLO_INT * T0.X, T0.X, T3.W,3692; EG-NEXT:     SETGE_UINT T2.X, PV.W, T2.W, BS:VEC_021/SCL_1223693; EG-NEXT:     SUB_INT T0.Y, PV.W, T2.W, BS:VEC_021/SCL_1223694; EG-NEXT:     SUB_INT T2.Z, T0.Z, PS,3695; EG-NEXT:     ADD_INT T9.W, T4.Y, T5.W,3696; EG-NEXT:     ADDC_UINT * T10.W, T4.X, T5.W,3697; EG-NEXT:     ADD_INT T1.Y, PV.W, PS,3698; EG-NEXT:     SETGE_UINT T3.Z, PV.Z, T3.W, BS:VEC_021/SCL_1223699; EG-NEXT:     SUB_INT T9.W, PV.Z, T3.W, BS:VEC_021/SCL_1223700; EG-NEXT:     CNDE_INT * T8.W, PV.X, T8.W, PV.Y,3701; EG-NEXT:     SETGE_UINT T0.X, PS, T2.W, BS:VEC_021/SCL_1223702; EG-NEXT:     SUB_INT T0.Y, PS, T2.W, BS:VEC_021/SCL_1223703; EG-NEXT:     CNDE_INT T2.Z, PV.Z, T2.Z, PV.W,3704; EG-NEXT:     ADD_INT T4.W, T4.W, T6.W,3705; EG-NEXT:     ADDC_UINT * T9.W, T4.Z, T6.W,3706; EG-NEXT:     ADD_INT T2.X, PV.W, PS,3707; EG-NEXT:     SETGE_UINT T2.Y, PV.Z, T3.W, BS:VEC_021/SCL_1223708; EG-NEXT:     SUB_INT T3.Z, PV.Z, T3.W, BS:VEC_021/SCL_1223709; EG-NEXT:     CNDE_INT T8.W, PV.X, T8.W, PV.Y, BS:VEC_102/SCL_2213710; EG-NEXT:     XOR_INT * T4.W, T1.Y, T5.W,3711; EG-NEXT:     CNDE_INT T0.X, PS, PV.W, T7.W,3712; EG-NEXT:     ADD_INT T1.Y, T1.Z, T0.W,3713; EG-NEXT:     CNDE_INT T1.Z, PV.Y, T2.Z, PV.Z,3714; EG-NEXT:     XOR_INT * T5.W, PV.X, T6.W, BS:VEC_102/SCL_2213715; EG-NEXT:     ADD_INT * T6.W, T1.X, T1.W,3716; EG-NEXT:     XOR_INT T0.Y, PV.W, T1.W,3717; EG-NEXT:     CNDE_INT T0.Z, T5.W, T1.Z, T0.Z,3718; EG-NEXT:     XOR_INT T6.W, T1.Y, T0.W, BS:VEC_021/SCL_1223719; EG-NEXT:     LSHR * T7.W, T0.X, literal.x,3720; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3721; EG-NEXT:     SETE_INT T1.Y, PS, T4.W,3722; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Z, PV.W, literal.x,3723; EG-NEXT:     BIT_ALIGN_INT T8.W, T0.X, PV.Y, literal.x,3724; EG-NEXT:     LSHR * T9.W, PV.Z, literal.x,3725; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3726; EG-NEXT:     SETGE_UINT T0.X, T7.W, T4.W,3727; EG-NEXT:     SETGE_UINT * T2.Y, PV.W, T2.W, BS:VEC_021/SCL_1223728; EG-NEXT:     SETE_INT T0.Z, T9.W, T5.W, BS:VEC_021/SCL_1223729; EG-NEXT:     SETGE_UINT T10.W, T9.W, T5.W, BS:VEC_021/SCL_1223730; EG-NEXT:     SETGE_UINT * T11.W, T1.Z, T3.W,3731; EG-NEXT:     CNDE_INT T3.Y, PV.Z, PV.W, PS,3732; EG-NEXT:     SUB_INT T0.Z, T1.Z, T3.W, BS:VEC_102/SCL_2213733; EG-NEXT:     CNDE_INT T10.W, T1.Y, T0.X, T2.Y,3734; EG-NEXT:     SUB_INT * T11.W, T8.W, T2.W,3735; EG-NEXT:     SUB_INT * T0.X, T7.W, T4.W,3736; EG-NEXT:     SUBB_UINT * T1.Y, T8.W, T2.W,3737; EG-NEXT:     SUB_INT T2.Z, T9.W, T5.W,3738; EG-NEXT:     SUBB_UINT * T12.W, T1.Z, T3.W, BS:VEC_021/SCL_1223739; EG-NEXT:     CNDE_INT * T8.W, T10.W, T8.W, T11.W,3740; EG-NEXT:     LSHL T1.X, PV.W, 1,3741; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3742; EG-NEXT:     SUB_INT T2.Z, T2.Z, T12.W, BS:VEC_2013743; EG-NEXT:     SUB_INT T11.W, T0.X, T1.Y,3744; EG-NEXT:     CNDE_INT * T12.W, T3.Y, T1.Z, T0.Z,3745; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)3746; EG-NEXT:     LSHL T0.X, PS, 1,3747; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3748; EG-NEXT:     CNDE_INT * T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2123749; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)3750; EG-NEXT:     CNDE_INT T7.W, T3.Y, T9.W, T2.Z,3751; EG-NEXT:     OR_INT * T9.W, T1.X, T2.Y,3752; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3753; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T12.W, literal.x, BS:VEC_021/SCL_1223754; EG-NEXT:     BIT_ALIGN_INT T7.W, T0.Z, T8.W, literal.x, BS:VEC_102/SCL_2213755; EG-NEXT:     OR_INT * T8.W, T0.X, T1.Y,3756; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3757; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223758; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223759; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213760; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,3761; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3762; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3763; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3764; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,3765; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,3766; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,3767; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3768; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3769; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223770; EG-NEXT:    ALU clause starting at 132:3771; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,3772; EG-NEXT:     LSHL T1.X, PV.W, 1,3773; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3774; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3775; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123776; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3777; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)3778; EG-NEXT:     LSHL T0.X, PS, 1,3779; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3780; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2123781; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3782; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,3783; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)3784; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3785; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223786; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2213787; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3788; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3789; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223790; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223791; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213792; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,3793; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3794; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3795; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3796; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,3797; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,3798; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,3799; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3800; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3801; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223802; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,3803; EG-NEXT:     LSHL T1.X, PV.W, 1,3804; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3805; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3806; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123807; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3808; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)3809; EG-NEXT:     LSHL T0.X, PS, 1,3810; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3811; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2123812; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3813; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,3814; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)3815; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3816; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223817; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2213818; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3819; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3820; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223821; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223822; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213823; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,3824; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3825; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3826; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3827; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,3828; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,3829; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,3830; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3831; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3832; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223833; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,3834; EG-NEXT:     LSHL T1.X, PV.W, 1,3835; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3836; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3837; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123838; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3839; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)3840; EG-NEXT:     LSHL T0.X, PS, 1,3841; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3842; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2123843; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3844; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,3845; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)3846; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3847; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223848; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2213849; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3850; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3851; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223852; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223853; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213854; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,3855; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3856; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3857; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3858; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,3859; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,3860; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,3861; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3862; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3863; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223864; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,3865; EG-NEXT:     LSHL T1.X, PV.W, 1,3866; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3867; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3868; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123869; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3870; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)3871; EG-NEXT:     LSHL T0.X, PS, 1,3872; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3873; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2123874; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3875; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,3876; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)3877; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3878; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223879; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2213880; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3881; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3882; EG-NEXT:     SETE_INT * T0.X, PV.W, T4.W,3883; EG-NEXT:    ALU clause starting at 244:3884; EG-NEXT:     SETGE_UINT * T1.Y, T7.W, T4.W,3885; EG-NEXT:     SETGE_UINT T0.Z, T8.W, T3.W, BS:VEC_021/SCL_1223886; EG-NEXT:     SETE_INT T10.W, T1.Z, T5.W,3887; EG-NEXT:     SETGE_UINT * T11.W, T1.Z, T5.W,3888; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3889; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3890; EG-NEXT:     CNDE_INT * T10.W, T0.X, T1.Y, T2.Y,3891; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,3892; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,3893; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3894; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3895; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223896; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,3897; EG-NEXT:     LSHL T1.X, PV.W, 1,3898; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3899; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3900; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123901; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3902; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)3903; EG-NEXT:     LSHL T0.X, PS, 1,3904; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3905; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2123906; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3907; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,3908; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)3909; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3910; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223911; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2213912; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3913; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3914; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223915; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223916; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213917; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,3918; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3919; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3920; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3921; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,3922; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,3923; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,3924; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3925; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3926; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223927; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,3928; EG-NEXT:     LSHL T1.X, PV.W, 1,3929; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3930; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3931; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123932; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3933; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)3934; EG-NEXT:     LSHL T0.X, PS, 1,3935; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3936; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2123937; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3938; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,3939; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)3940; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3941; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223942; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2213943; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3944; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3945; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223946; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223947; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213948; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,3949; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3950; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3951; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3952; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,3953; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,3954; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,3955; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3956; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3957; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223958; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,3959; EG-NEXT:     LSHL T1.X, PV.W, 1,3960; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3961; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3962; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123963; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3964; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)3965; EG-NEXT:     LSHL T0.X, PS, 1,3966; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,3967; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2123968; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,3969; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,3970; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)3971; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,3972; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1223973; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2213974; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,3975; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)3976; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1223977; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1223978; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2213979; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,3980; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,3981; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,3982; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,3983; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,3984; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,3985; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,3986; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,3987; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,3988; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1223989; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,3990; EG-NEXT:     LSHL T1.X, PV.W, 1,3991; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,3992; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,3993; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2123994; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,3995; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)3996; EG-NEXT:     LSHL * T0.X, PS, 1,3997; EG-NEXT:    ALU clause starting at 357:3998; EG-NEXT:     BFE_UINT * T1.Y, T6.W, literal.x, 1,3999; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)4000; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, T11.W,4001; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, T2.Z,4002; EG-NEXT:     OR_INT * T9.W, T1.X, T2.Y,4003; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4004; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224005; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214006; EG-NEXT:     OR_INT * T8.W, T0.X, T1.Y,4007; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4008; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224009; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224010; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214011; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4012; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4013; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4014; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4015; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4016; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4017; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4018; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4019; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4020; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224021; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4022; EG-NEXT:     LSHL T1.X, PV.W, 1,4023; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4024; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4025; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124026; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4027; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)4028; EG-NEXT:     LSHL T0.X, PS, 1,4029; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4030; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124031; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4032; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4033; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)4034; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4035; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224036; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214037; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4038; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4039; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224040; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224041; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214042; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4043; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4044; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4045; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4046; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4047; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4048; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4049; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4050; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4051; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224052; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4053; EG-NEXT:     LSHL T1.X, PV.W, 1,4054; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4055; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4056; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124057; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4058; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)4059; EG-NEXT:     LSHL T0.X, PS, 1,4060; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4061; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124062; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4063; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4064; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)4065; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4066; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224067; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214068; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4069; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4070; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224071; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224072; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214073; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4074; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4075; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4076; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4077; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4078; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4079; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4080; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4081; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4082; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224083; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4084; EG-NEXT:     LSHL T1.X, PV.W, 1,4085; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4086; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4087; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124088; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4089; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)4090; EG-NEXT:     LSHL T0.X, PS, 1,4091; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4092; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124093; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4094; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4095; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)4096; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4097; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224098; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214099; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4100; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4101; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224102; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224103; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214104; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4105; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4106; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4107; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4108; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4109; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4110; EG-NEXT:    ALU clause starting at 469:4111; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4112; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4113; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4114; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224115; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4116; EG-NEXT:     LSHL T1.X, PV.W, 1,4117; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4118; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4119; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124120; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4121; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)4122; EG-NEXT:     LSHL T0.X, PS, 1,4123; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4124; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124125; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4126; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4127; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)4128; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4129; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224130; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214131; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4132; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4133; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224134; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224135; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214136; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4137; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4138; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4139; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4140; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4141; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4142; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4143; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4144; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4145; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224146; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4147; EG-NEXT:     LSHL T1.X, PV.W, 1,4148; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4149; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4150; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124151; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4152; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)4153; EG-NEXT:     LSHL T0.X, PS, 1,4154; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4155; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124156; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4157; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4158; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)4159; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4160; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224161; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214162; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4163; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4164; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224165; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224166; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214167; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4168; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4169; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4170; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4171; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4172; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4173; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4174; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4175; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4176; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224177; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4178; EG-NEXT:     LSHL T1.X, PV.W, 1,4179; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4180; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4181; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124182; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4183; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)4184; EG-NEXT:     LSHL T0.X, PS, 1,4185; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4186; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124187; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4188; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4189; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)4190; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4191; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224192; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214193; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4194; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4195; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224196; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224197; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214198; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4199; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4200; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4201; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4202; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4203; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4204; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4205; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4206; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4207; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224208; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4209; EG-NEXT:     LSHL T1.X, PV.W, 1,4210; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4211; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4212; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124213; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4214; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)4215; EG-NEXT:     LSHL T0.X, PS, 1,4216; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4217; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124218; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4219; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4220; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)4221; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4222; EG-NEXT:     BIT_ALIGN_INT * T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224223; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4224; EG-NEXT:    ALU clause starting at 582:4225; EG-NEXT:     BIT_ALIGN_INT T7.W, T0.Z, T9.W, literal.x,4226; EG-NEXT:     OR_INT * T8.W, T0.X, T1.Y,4227; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4228; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224229; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224230; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214231; EG-NEXT:     SETE_INT T9.W, T1.Z, T5.W,4232; EG-NEXT:     SETGE_UINT * T11.W, T1.Z, T5.W,4233; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4234; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4235; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4236; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4237; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4238; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4239; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4240; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224241; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4242; EG-NEXT:     LSHL T1.X, PV.W, 1,4243; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4244; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4245; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124246; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4247; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)4248; EG-NEXT:     LSHL T0.X, PS, 1,4249; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4250; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124251; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4252; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4253; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)4254; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4255; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224256; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214257; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4258; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4259; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224260; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224261; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214262; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4263; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4264; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4265; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4266; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4267; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4268; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4269; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4270; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4271; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224272; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4273; EG-NEXT:     LSHL T1.X, PV.W, 1,4274; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4275; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4276; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124277; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4278; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)4279; EG-NEXT:     LSHL T0.X, PS, 1,4280; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4281; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124282; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4283; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4284; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)4285; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4286; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224287; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214288; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4289; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4290; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224291; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224292; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214293; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4294; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4295; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4296; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4297; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4298; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4299; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4300; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4301; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4302; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224303; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4304; EG-NEXT:     LSHL T1.X, PV.W, 1,4305; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4306; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4307; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124308; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4309; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)4310; EG-NEXT:     LSHL T0.X, PS, 1,4311; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4312; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124313; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4314; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4315; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)4316; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4317; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224318; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214319; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4320; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4321; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224322; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224323; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214324; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4325; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4326; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4327; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4328; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4329; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4330; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4331; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4332; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4333; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224334; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4335; EG-NEXT:     LSHL T1.X, PV.W, 1,4336; EG-NEXT:     BFE_UINT * T2.Y, T0.Y, literal.x, 1,4337; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)4338; EG-NEXT:    ALU clause starting at 695:4339; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4340; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X,4341; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4342; EG-NEXT:     LSHL T0.X, PS, 1,4343; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4344; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124345; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4346; EG-NEXT:     OR_INT * T10.W, T1.X, T2.Y,4347; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)4348; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4349; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224350; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214351; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4352; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4353; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224354; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224355; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214356; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4357; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4358; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4359; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4360; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4361; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4362; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4363; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4364; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4365; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224366; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4367; EG-NEXT:     LSHL T1.X, PV.W, 1,4368; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4369; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4370; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124371; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4372; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)4373; EG-NEXT:     LSHL T0.X, PS, 1,4374; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4375; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124376; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4377; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4378; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)4379; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4380; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224381; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214382; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4383; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4384; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224385; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224386; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214387; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4388; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4389; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4390; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4391; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4392; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4393; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4394; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4395; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4396; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224397; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4398; EG-NEXT:     LSHL T1.X, PV.W, 1,4399; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4400; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4401; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124402; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4403; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)4404; EG-NEXT:     LSHL T0.X, PS, 1,4405; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4406; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124407; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4408; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4409; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)4410; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4411; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224412; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214413; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4414; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4415; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224416; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224417; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214418; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4419; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4420; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4421; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4422; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4423; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4424; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4425; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4426; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4427; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224428; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4429; EG-NEXT:     LSHL T1.X, PV.W, 1,4430; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4431; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4432; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124433; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4434; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)4435; EG-NEXT:     LSHL T0.X, PS, 1,4436; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4437; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124438; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4439; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4440; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)4441; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4442; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224443; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214444; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4445; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4446; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224447; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224448; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214449; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4450; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4451; EG-NEXT:    ALU clause starting at 807:4452; EG-NEXT:     CNDE_INT * T3.Y, T10.W, T11.W, T0.Z,4453; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4454; EG-NEXT:     CNDE_INT * T10.W, T0.X, T1.Y, T2.Y,4455; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4456; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4457; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4458; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4459; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224460; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4461; EG-NEXT:     LSHL T1.X, PV.W, 1,4462; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4463; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4464; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124465; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4466; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)4467; EG-NEXT:     LSHL T0.X, PS, 1,4468; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4469; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124470; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4471; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4472; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)4473; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4474; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224475; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214476; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4477; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4478; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224479; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224480; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214481; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4482; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4483; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4484; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4485; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4486; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4487; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4488; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4489; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4490; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224491; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4492; EG-NEXT:     LSHL T1.X, PV.W, 1,4493; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4494; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4495; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124496; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4497; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)4498; EG-NEXT:     LSHL T0.X, PS, 1,4499; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4500; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124501; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4502; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4503; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)4504; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4505; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224506; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214507; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4508; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4509; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224510; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224511; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214512; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4513; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4514; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4515; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4516; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4517; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4518; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4519; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4520; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4521; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224522; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4523; EG-NEXT:     LSHL T1.X, PV.W, 1,4524; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4525; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4526; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124527; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4528; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)4529; EG-NEXT:     LSHL T0.X, PS, 1,4530; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4531; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124532; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4533; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4534; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)4535; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4536; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224537; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214538; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4539; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4540; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224541; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224542; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214543; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4544; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4545; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4546; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4547; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4548; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4549; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4550; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4551; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4552; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224553; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4554; EG-NEXT:     LSHL T1.X, PV.W, 1,4555; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4556; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4557; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124558; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4559; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)4560; EG-NEXT:     LSHL T0.X, PS, 1,4561; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4562; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124563; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T1.Z, PV.Z,4564; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)4565; EG-NEXT:    ALU clause starting at 920:4566; EG-NEXT:     OR_INT * T9.W, T1.X, T2.Y,4567; EG-NEXT:     SETGE_UINT T2.Y, PV.W, T2.W,4568; EG-NEXT:     BIT_ALIGN_INT * T1.Z, T7.W, T8.W, literal.x, BS:VEC_021/SCL_1224569; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4570; EG-NEXT:     BIT_ALIGN_INT T7.W, T0.Z, T10.W, literal.x,4571; EG-NEXT:     OR_INT * T8.W, T0.X, T1.Y,4572; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4573; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224574; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224575; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214576; EG-NEXT:     SETE_INT T10.W, T1.Z, T5.W,4577; EG-NEXT:     SETGE_UINT * T11.W, T1.Z, T5.W,4578; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4579; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4580; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4581; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4582; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4583; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4584; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4585; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224586; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4587; EG-NEXT:     LSHL T1.X, PV.W, 1,4588; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4589; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4590; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124591; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4592; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)4593; EG-NEXT:     LSHL T0.X, PS, 1,4594; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4595; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124596; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4597; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4598; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)4599; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4600; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224601; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214602; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4603; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4604; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224605; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224606; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214607; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4608; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4609; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4610; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4611; EG-NEXT:     CNDE_INT * T9.W, PV.X, PV.Y, T2.Y,4612; EG-NEXT:     SUB_INT * T11.W, T10.W, T2.W,4613; EG-NEXT:     SUBB_UINT * T0.X, T10.W, T2.W,4614; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4615; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4616; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224617; EG-NEXT:     CNDE_INT * T10.W, T9.W, T10.W, T11.W,4618; EG-NEXT:     LSHL T1.X, PV.W, 1,4619; EG-NEXT:     BFE_UINT T2.Y, T0.Y, literal.x, 1,4620; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4621; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124622; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4623; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)4624; EG-NEXT:     LSHL T0.X, PS, 1,4625; EG-NEXT:     BFE_UINT T1.Y, T6.W, literal.x, 1,4626; EG-NEXT:     CNDE_INT T0.Z, T9.W, T7.W, PV.W, BS:VEC_120/SCL_2124627; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4628; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4629; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)4630; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4631; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224632; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T10.W, literal.x, BS:VEC_102/SCL_2214633; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4634; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4635; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224636; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224637; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214638; EG-NEXT:     SETE_INT T10.W, PV.Z, T5.W,4639; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4640; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,4641; EG-NEXT:     SUB_INT T0.Z, T8.W, T3.W,4642; EG-NEXT:     CNDE_INT * T10.W, PV.X, PV.Y, T2.Y,4643; EG-NEXT:     SUB_INT * T11.W, T9.W, T2.W,4644; EG-NEXT:     SUBB_UINT * T0.X, T9.W, T2.W,4645; EG-NEXT:     SUB_INT * T1.Y, T7.W, T4.W,4646; EG-NEXT:     SUBB_UINT T2.Z, T8.W, T3.W,4647; EG-NEXT:     SUB_INT * T12.W, T1.Z, T5.W, BS:VEC_021/SCL_1224648; EG-NEXT:     CNDE_INT * T9.W, T10.W, T9.W, T11.W,4649; EG-NEXT:     LSHL T1.X, PV.W, 1,4650; EG-NEXT:     BFE_UINT T2.Y, T0.Y, 1, 1,4651; EG-NEXT:     SUB_INT T2.Z, T12.W, T2.Z,4652; EG-NEXT:     SUB_INT T11.W, T1.Y, T0.X, BS:VEC_120/SCL_2124653; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,4654; EG-NEXT:     LSHL T0.X, PS, 1,4655; EG-NEXT:     BFE_UINT T1.Y, T6.W, 1, 1,4656; EG-NEXT:     CNDE_INT T0.Z, T10.W, T7.W, PV.W, BS:VEC_120/SCL_2124657; EG-NEXT:     CNDE_INT T7.W, T3.Y, T1.Z, PV.Z,4658; EG-NEXT:     OR_INT * T10.W, PV.X, PV.Y,4659; EG-NEXT:     SETGE_UINT T2.Y, PS, T2.W,4660; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1224661; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T9.W, literal.x, BS:VEC_102/SCL_2214662; EG-NEXT:     OR_INT * T8.W, PV.X, PV.Y,4663; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4664; EG-NEXT:     SETE_INT T0.X, PV.W, T4.W, BS:VEC_021/SCL_1224665; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1224666; EG-NEXT:     SETGE_UINT T0.Z, PS, T3.W, BS:VEC_102/SCL_2214667; EG-NEXT:     SETE_INT T9.W, PV.Z, T5.W,4668; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T5.W,4669; EG-NEXT:     CNDE_INT T1.X, PV.W, PS, PV.Z,4670; EG-NEXT:     CNDE_INT T1.Y, PV.X, PV.Y, T2.Y,4671; EG-NEXT:     SUB_INT T0.Z, T10.W, T2.W,4672; EG-NEXT:     SUBB_UINT * T9.W, T10.W, T2.W,4673; EG-NEXT:     SUB_INT * T11.W, T7.W, T4.W,4674; EG-NEXT:     SUB_INT T0.X, T8.W, T3.W,4675; EG-NEXT:     SUBB_UINT T2.Y, T8.W, T3.W,4676; EG-NEXT:     SUB_INT * T2.Z, T1.Z, T5.W, BS:VEC_021/SCL_1224677; EG-NEXT:     SUB_INT T9.W, T11.W, T9.W, BS:VEC_021/SCL_1224678; EG-NEXT:     CNDE_INT * T10.W, T1.Y, T10.W, T0.Z,4679; EG-NEXT:    ALU clause starting at 1033:4680; EG-NEXT:     LSHL T2.X, T10.W, 1,4681; EG-NEXT:     AND_INT T0.Y, T0.Y, 1, BS:VEC_120/SCL_2124682; EG-NEXT:     CNDE_INT T0.Z, T1.Y, T7.W, T9.W,4683; EG-NEXT:     SUB_INT * T7.W, T2.Z, T2.Y, BS:VEC_021/SCL_1224684; EG-NEXT:     CNDE_INT * T8.W, T1.X, T8.W, T0.X,4685; EG-NEXT:     LSHL T0.X, PV.W, 1,4686; EG-NEXT:     AND_INT T1.Y, T6.W, 1,4687; EG-NEXT:     CNDE_INT T1.Z, T1.X, T1.Z, T7.W,4688; EG-NEXT:     BIT_ALIGN_INT T6.W, T0.Z, T10.W, literal.x,4689; EG-NEXT:     OR_INT * T7.W, T2.X, T0.Y,4690; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4691; EG-NEXT:     SETGE_UINT T1.X, PS, T2.W,4692; EG-NEXT:     SETE_INT T0.Y, PV.W, T4.W, BS:VEC_021/SCL_1224693; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T4.W, BS:VEC_021/SCL_1224694; EG-NEXT:     BIT_ALIGN_INT T8.W, PV.Z, T8.W, literal.x, BS:VEC_102/SCL_2214695; EG-NEXT:     OR_INT * T9.W, PV.X, PV.Y,4696; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4697; EG-NEXT:     SETGE_UINT T0.X, PS, T3.W, BS:VEC_021/SCL_1224698; EG-NEXT:     SETE_INT T1.Y, PV.W, T5.W,4699; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T5.W,4700; EG-NEXT:     SUBB_UINT T10.W, PS, T3.W, BS:VEC_021/SCL_1224701; EG-NEXT:     SUB_INT * T5.W, PV.W, T5.W,4702; EG-NEXT:     SUBB_UINT * T2.X, T7.W, T2.W,4703; EG-NEXT:     SUB_INT * T2.Y, T6.W, T4.W,4704; EG-NEXT:     SUB_INT * T2.Z, T9.W, T3.W,4705; EG-NEXT:     SUB_INT T3.W, T5.W, T10.W,4706; EG-NEXT:     CNDE_INT * T4.W, T1.Y, T1.Z, T0.X,4707; EG-NEXT:     CNDE_INT T0.X, PS, T8.W, PV.W,4708; EG-NEXT:     CNDE_INT * T1.Y, PS, T9.W, T2.Z, BS:VEC_021/SCL_1224709; EG-NEXT:     SUB_INT T1.Z, T7.W, T2.W,4710; EG-NEXT:     SUB_INT T2.W, T2.Y, T2.X,4711; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T0.Z, T1.X,4712; EG-NEXT:     CNDE_INT T0.Y, PS, T6.W, PV.W, BS:VEC_021/SCL_1224713; EG-NEXT:     CNDE_INT T0.Z, PS, T7.W, PV.Z, BS:VEC_102/SCL_2214714; EG-NEXT:     XOR_INT T2.W, T1.Y, T0.W,4715; EG-NEXT:     XOR_INT * T3.W, T0.X, T0.W,4716; EG-NEXT:     SUB_INT T1.Y, PS, T0.W, BS:VEC_021/SCL_1224717; EG-NEXT:     SUBB_UINT T1.Z, PV.W, T0.W, BS:VEC_021/SCL_1224718; EG-NEXT:     XOR_INT T3.W, PV.Z, T1.W,4719; EG-NEXT:     XOR_INT * T4.W, PV.Y, T1.W,4720; EG-NEXT:     SUB_INT T0.Z, PS, T1.W,4721; EG-NEXT:     SUBB_UINT T4.W, PV.W, T1.W,4722; EG-NEXT:     SUB_INT * T5.W, PV.Y, PV.Z,4723; EG-NEXT:     SUB_INT T5.Y, PV.Z, PV.W,4724; EG-NEXT:     SUB_INT * T5.Z, T2.W, T0.W,4725; EG-NEXT:     SUB_INT T5.X, T3.W, T1.W,4726; EG-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,4727; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)4728  %den_ptr = getelementptr <2 x i64>, ptr addrspace(1) %in, i64 14729  %num = load <2 x i64>, ptr addrspace(1) %in4730  %den = load <2 x i64>, ptr addrspace(1) %den_ptr4731  %result = srem <2 x i64> %num, %den4732  store <2 x i64> %result, ptr addrspace(1) %out4733  ret void4734}4735 4736define amdgpu_kernel void @srem_v2i64_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {4737; GCN-LABEL: srem_v2i64_4:4738; GCN:       ; %bb.0:4739; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x244740; GCN-NEXT:    v_mov_b32_e32 v4, 04741; GCN-NEXT:    s_waitcnt lgkmcnt(0)4742; GCN-NEXT:    global_load_dwordx4 v[0:3], v4, s[2:3]4743; GCN-NEXT:    s_waitcnt vmcnt(0)4744; GCN-NEXT:    v_ashrrev_i32_e32 v5, 31, v14745; GCN-NEXT:    v_lshrrev_b32_e32 v5, 30, v54746; GCN-NEXT:    v_ashrrev_i32_e32 v6, 31, v34747; GCN-NEXT:    v_add_co_u32_e32 v5, vcc, v0, v54748; GCN-NEXT:    v_lshrrev_b32_e32 v6, 30, v64749; GCN-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v1, vcc4750; GCN-NEXT:    v_add_co_u32_e32 v6, vcc, v2, v64751; GCN-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc4752; GCN-NEXT:    v_and_b32_e32 v5, -4, v54753; GCN-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v54754; GCN-NEXT:    v_and_b32_e32 v6, -4, v64755; GCN-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v7, vcc4756; GCN-NEXT:    v_sub_co_u32_e32 v2, vcc, v2, v64757; GCN-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v8, vcc4758; GCN-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]4759; GCN-NEXT:    s_endpgm4760;4761; TAHITI-LABEL: srem_v2i64_4:4762; TAHITI:       ; %bb.0:4763; TAHITI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x94764; TAHITI-NEXT:    s_mov_b32 s7, 0xf0004765; TAHITI-NEXT:    s_mov_b32 s6, -14766; TAHITI-NEXT:    s_mov_b32 s10, s64767; TAHITI-NEXT:    s_mov_b32 s11, s74768; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)4769; TAHITI-NEXT:    s_mov_b32 s8, s24770; TAHITI-NEXT:    s_mov_b32 s9, s34771; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 04772; TAHITI-NEXT:    s_mov_b32 s4, s04773; TAHITI-NEXT:    s_mov_b32 s5, s14774; TAHITI-NEXT:    s_waitcnt vmcnt(0)4775; TAHITI-NEXT:    v_ashrrev_i32_e32 v4, 31, v14776; TAHITI-NEXT:    v_lshrrev_b32_e32 v4, 30, v44777; TAHITI-NEXT:    v_ashrrev_i32_e32 v5, 31, v34778; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v0, v44779; TAHITI-NEXT:    v_lshrrev_b32_e32 v5, 30, v54780; TAHITI-NEXT:    v_addc_u32_e32 v6, vcc, 0, v1, vcc4781; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v2, v54782; TAHITI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v3, vcc4783; TAHITI-NEXT:    v_and_b32_e32 v4, -4, v44784; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v0, v44785; TAHITI-NEXT:    v_and_b32_e32 v5, -4, v54786; TAHITI-NEXT:    v_subb_u32_e32 v1, vcc, v1, v6, vcc4787; TAHITI-NEXT:    v_sub_i32_e32 v2, vcc, v2, v54788; TAHITI-NEXT:    v_subb_u32_e32 v3, vcc, v3, v7, vcc4789; TAHITI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 04790; TAHITI-NEXT:    s_endpgm4791;4792; TONGA-LABEL: srem_v2i64_4:4793; TONGA:       ; %bb.0:4794; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x244795; TONGA-NEXT:    s_waitcnt lgkmcnt(0)4796; TONGA-NEXT:    v_mov_b32_e32 v0, s24797; TONGA-NEXT:    v_mov_b32_e32 v1, s34798; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]4799; TONGA-NEXT:    v_mov_b32_e32 v4, s04800; TONGA-NEXT:    v_mov_b32_e32 v5, s14801; TONGA-NEXT:    s_waitcnt vmcnt(0)4802; TONGA-NEXT:    v_ashrrev_i32_e32 v6, 31, v14803; TONGA-NEXT:    v_lshrrev_b32_e32 v6, 30, v64804; TONGA-NEXT:    v_ashrrev_i32_e32 v7, 31, v34805; TONGA-NEXT:    v_add_u32_e32 v6, vcc, v0, v64806; TONGA-NEXT:    v_lshrrev_b32_e32 v7, 30, v74807; TONGA-NEXT:    v_addc_u32_e32 v8, vcc, 0, v1, vcc4808; TONGA-NEXT:    v_add_u32_e32 v7, vcc, v2, v74809; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v3, vcc4810; TONGA-NEXT:    v_and_b32_e32 v6, -4, v64811; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v0, v64812; TONGA-NEXT:    v_and_b32_e32 v7, -4, v74813; TONGA-NEXT:    v_subb_u32_e32 v1, vcc, v1, v8, vcc4814; TONGA-NEXT:    v_sub_u32_e32 v2, vcc, v2, v74815; TONGA-NEXT:    v_subb_u32_e32 v3, vcc, v3, v9, vcc4816; TONGA-NEXT:    flat_store_dwordx4 v[4:5], v[0:3]4817; TONGA-NEXT:    s_endpgm4818;4819; EG-LABEL: srem_v2i64_4:4820; EG:       ; %bb.0:4821; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]4822; EG-NEXT:    TEX 0 @64823; EG-NEXT:    ALU 24, @9, KC0[CB0:0-32], KC1[]4824; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T3.XYZW, T0.X, 14825; EG-NEXT:    CF_END4826; EG-NEXT:    PAD4827; EG-NEXT:    Fetch clause starting at 6:4828; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #14829; EG-NEXT:    ALU clause starting at 8:4830; EG-NEXT:     MOV * T0.X, KC0[2].Z,4831; EG-NEXT:    ALU clause starting at 9:4832; EG-NEXT:     ASHR T0.W, T0.W, literal.x,4833; EG-NEXT:     ASHR * T1.W, T0.Y, literal.x,4834; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)4835; EG-NEXT:     LSHR * T0.W, PV.W, literal.x,4836; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)4837; EG-NEXT:     ADD_INT T2.W, T0.Z, PV.W,4838; EG-NEXT:     LSHR * T1.W, T1.W, literal.x,4839; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)4840; EG-NEXT:     ADD_INT T1.Z, T0.X, PS,4841; EG-NEXT:     ADDC_UINT T0.W, T0.Z, T0.W,4842; EG-NEXT:     AND_INT * T2.W, PV.W, literal.x,4843; EG-NEXT:    -4(nan), 0(0.000000e+00)4844; EG-NEXT:     SUBB_UINT T0.Y, T0.Z, PS,4845; EG-NEXT:     BFE_INT T2.Z, PV.W, 0.0, 1,4846; EG-NEXT:     ADDC_UINT T0.W, T0.X, T1.W,4847; EG-NEXT:     AND_INT * T1.W, PV.Z, literal.x,4848; EG-NEXT:    -4(nan), 0(0.000000e+00)4849; EG-NEXT:     SUBB_UINT T1.Z, T0.X, PS,4850; EG-NEXT:     BFE_INT T0.W, PV.W, 0.0, 1,4851; EG-NEXT:     SUB_INT * T3.W, PV.Z, PV.Y,4852; EG-NEXT:     SUB_INT T3.Y, PV.W, PV.Z,4853; EG-NEXT:     SUB_INT * T3.Z, T0.Z, T2.W,4854; EG-NEXT:     SUB_INT T3.X, T0.X, T1.W,4855; EG-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,4856; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)4857  %num = load <2 x i64>, ptr addrspace(1) %in4858  %result = srem <2 x i64> %num, <i64 4, i64 4>4859  store <2 x i64> %result, ptr addrspace(1) %out4860  ret void4861}4862 4863define amdgpu_kernel void @srem_v4i64(ptr addrspace(1) %out, ptr addrspace(1) %in) {4864; GCN-LABEL: srem_v4i64:4865; GCN:       ; %bb.0:4866; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x244867; GCN-NEXT:    v_mov_b32_e32 v16, 04868; GCN-NEXT:    s_waitcnt lgkmcnt(0)4869; GCN-NEXT:    global_load_dwordx4 v[0:3], v16, s[2:3] offset:484870; GCN-NEXT:    global_load_dwordx4 v[4:7], v16, s[2:3] offset:324871; GCN-NEXT:    global_load_dwordx4 v[8:11], v16, s[2:3] offset:164872; GCN-NEXT:    global_load_dwordx4 v[12:15], v16, s[2:3]4873; GCN-NEXT:    s_waitcnt vmcnt(3)4874; GCN-NEXT:    v_readfirstlane_b32 s3, v34875; GCN-NEXT:    s_waitcnt vmcnt(2)4876; GCN-NEXT:    v_readfirstlane_b32 s17, v54877; GCN-NEXT:    v_readfirstlane_b32 s16, v44878; GCN-NEXT:    s_waitcnt vmcnt(0)4879; GCN-NEXT:    v_readfirstlane_b32 s19, v134880; GCN-NEXT:    v_readfirstlane_b32 s18, v124881; GCN-NEXT:    s_or_b64 s[6:7], s[18:19], s[16:17]4882; GCN-NEXT:    s_mov_b32 s6, 04883; GCN-NEXT:    v_readfirstlane_b32 s2, v24884; GCN-NEXT:    v_readfirstlane_b32 s9, v14885; GCN-NEXT:    v_readfirstlane_b32 s8, v04886; GCN-NEXT:    v_readfirstlane_b32 s13, v74887; GCN-NEXT:    v_readfirstlane_b32 s12, v64888; GCN-NEXT:    v_readfirstlane_b32 s5, v114889; GCN-NEXT:    v_readfirstlane_b32 s4, v104890; GCN-NEXT:    v_readfirstlane_b32 s11, v94891; GCN-NEXT:    v_readfirstlane_b32 s10, v84892; GCN-NEXT:    v_readfirstlane_b32 s15, v154893; GCN-NEXT:    s_cmp_lg_u64 s[6:7], 04894; GCN-NEXT:    v_readfirstlane_b32 s14, v144895; GCN-NEXT:    s_cbranch_scc0 .LBB12_64896; GCN-NEXT:  ; %bb.1:4897; GCN-NEXT:    s_ashr_i32 s6, s17, 314898; GCN-NEXT:    s_add_u32 s20, s16, s64899; GCN-NEXT:    s_mov_b32 s7, s64900; GCN-NEXT:    s_addc_u32 s21, s17, s64901; GCN-NEXT:    s_xor_b64 s[6:7], s[20:21], s[6:7]4902; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s64903; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s74904; GCN-NEXT:    s_sub_u32 s17, 0, s64905; GCN-NEXT:    s_subb_u32 s22, 0, s74906; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v04907; GCN-NEXT:    v_rcp_f32_e32 v0, v04908; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v04909; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v04910; GCN-NEXT:    v_trunc_f32_e32 v1, v14911; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v04912; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v14913; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v04914; GCN-NEXT:    v_readfirstlane_b32 s23, v14915; GCN-NEXT:    v_readfirstlane_b32 s24, v04916; GCN-NEXT:    s_mul_i32 s25, s17, s234917; GCN-NEXT:    s_mul_hi_u32 s27, s17, s244918; GCN-NEXT:    s_mul_i32 s26, s22, s244919; GCN-NEXT:    s_add_i32 s25, s27, s254920; GCN-NEXT:    s_add_i32 s25, s25, s264921; GCN-NEXT:    s_mul_i32 s28, s17, s244922; GCN-NEXT:    s_mul_i32 s27, s24, s254923; GCN-NEXT:    s_mul_hi_u32 s29, s24, s284924; GCN-NEXT:    s_mul_hi_u32 s26, s24, s254925; GCN-NEXT:    s_add_u32 s27, s29, s274926; GCN-NEXT:    s_addc_u32 s26, 0, s264927; GCN-NEXT:    s_mul_hi_u32 s30, s23, s284928; GCN-NEXT:    s_mul_i32 s28, s23, s284929; GCN-NEXT:    s_add_u32 s27, s27, s284930; GCN-NEXT:    s_mul_hi_u32 s29, s23, s254931; GCN-NEXT:    s_addc_u32 s26, s26, s304932; GCN-NEXT:    s_addc_u32 s27, s29, 04933; GCN-NEXT:    s_mul_i32 s25, s23, s254934; GCN-NEXT:    s_add_u32 s25, s26, s254935; GCN-NEXT:    s_addc_u32 s26, 0, s274936; GCN-NEXT:    s_add_u32 s24, s24, s254937; GCN-NEXT:    s_addc_u32 s23, s23, s264938; GCN-NEXT:    s_mul_i32 s25, s17, s234939; GCN-NEXT:    s_mul_hi_u32 s26, s17, s244940; GCN-NEXT:    s_add_i32 s25, s26, s254941; GCN-NEXT:    s_mul_i32 s22, s22, s244942; GCN-NEXT:    s_add_i32 s25, s25, s224943; GCN-NEXT:    s_mul_i32 s17, s17, s244944; GCN-NEXT:    s_mul_hi_u32 s26, s23, s174945; GCN-NEXT:    s_mul_i32 s27, s23, s174946; GCN-NEXT:    s_mul_i32 s29, s24, s254947; GCN-NEXT:    s_mul_hi_u32 s17, s24, s174948; GCN-NEXT:    s_mul_hi_u32 s28, s24, s254949; GCN-NEXT:    s_add_u32 s17, s17, s294950; GCN-NEXT:    s_addc_u32 s28, 0, s284951; GCN-NEXT:    s_add_u32 s17, s17, s274952; GCN-NEXT:    s_mul_hi_u32 s22, s23, s254953; GCN-NEXT:    s_addc_u32 s17, s28, s264954; GCN-NEXT:    s_addc_u32 s22, s22, 04955; GCN-NEXT:    s_mul_i32 s25, s23, s254956; GCN-NEXT:    s_add_u32 s17, s17, s254957; GCN-NEXT:    s_addc_u32 s22, 0, s224958; GCN-NEXT:    s_add_u32 s17, s24, s174959; GCN-NEXT:    s_addc_u32 s26, s23, s224960; GCN-NEXT:    s_ashr_i32 s22, s19, 314961; GCN-NEXT:    s_add_u32 s24, s18, s224962; GCN-NEXT:    s_mov_b32 s23, s224963; GCN-NEXT:    s_addc_u32 s25, s19, s224964; GCN-NEXT:    s_xor_b64 s[24:25], s[24:25], s[22:23]4965; GCN-NEXT:    s_mul_i32 s27, s24, s264966; GCN-NEXT:    s_mul_hi_u32 s28, s24, s174967; GCN-NEXT:    s_mul_hi_u32 s19, s24, s264968; GCN-NEXT:    s_add_u32 s27, s28, s274969; GCN-NEXT:    s_addc_u32 s19, 0, s194970; GCN-NEXT:    s_mul_hi_u32 s29, s25, s174971; GCN-NEXT:    s_mul_i32 s17, s25, s174972; GCN-NEXT:    s_add_u32 s17, s27, s174973; GCN-NEXT:    s_mul_hi_u32 s28, s25, s264974; GCN-NEXT:    s_addc_u32 s17, s19, s294975; GCN-NEXT:    s_addc_u32 s19, s28, 04976; GCN-NEXT:    s_mul_i32 s26, s25, s264977; GCN-NEXT:    s_add_u32 s17, s17, s264978; GCN-NEXT:    s_addc_u32 s19, 0, s194979; GCN-NEXT:    s_mul_i32 s19, s6, s194980; GCN-NEXT:    s_mul_hi_u32 s26, s6, s174981; GCN-NEXT:    s_add_i32 s19, s26, s194982; GCN-NEXT:    s_mul_i32 s26, s7, s174983; GCN-NEXT:    s_add_i32 s19, s19, s264984; GCN-NEXT:    s_sub_i32 s28, s25, s194985; GCN-NEXT:    s_mul_i32 s17, s6, s174986; GCN-NEXT:    s_sub_u32 s17, s24, s174987; GCN-NEXT:    s_cselect_b64 s[26:27], -1, 04988; GCN-NEXT:    s_subb_u32 s24, s28, s74989; GCN-NEXT:    s_sub_u32 s30, s17, s64990; GCN-NEXT:    s_cselect_b64 s[28:29], -1, 04991; GCN-NEXT:    s_subb_u32 s31, s24, 04992; GCN-NEXT:    s_cmp_ge_u32 s31, s74993; GCN-NEXT:    s_cselect_b32 s33, -1, 04994; GCN-NEXT:    s_cmp_ge_u32 s30, s64995; GCN-NEXT:    s_cselect_b32 s34, -1, 04996; GCN-NEXT:    s_cmp_eq_u32 s31, s74997; GCN-NEXT:    s_cselect_b32 s33, s34, s334998; GCN-NEXT:    s_cmp_lg_u64 s[28:29], 04999; GCN-NEXT:    s_subb_u32 s24, s24, s75000; GCN-NEXT:    s_sub_u32 s28, s30, s65001; GCN-NEXT:    s_subb_u32 s24, s24, 05002; GCN-NEXT:    s_cmp_lg_u32 s33, 05003; GCN-NEXT:    s_cselect_b32 s28, s28, s305004; GCN-NEXT:    s_cselect_b32 s24, s24, s315005; GCN-NEXT:    s_cmp_lg_u64 s[26:27], 05006; GCN-NEXT:    s_subb_u32 s19, s25, s195007; GCN-NEXT:    s_cmp_ge_u32 s19, s75008; GCN-NEXT:    s_cselect_b32 s25, -1, 05009; GCN-NEXT:    s_cmp_ge_u32 s17, s65010; GCN-NEXT:    s_cselect_b32 s6, -1, 05011; GCN-NEXT:    s_cmp_eq_u32 s19, s75012; GCN-NEXT:    s_cselect_b32 s6, s6, s255013; GCN-NEXT:    s_cmp_lg_u32 s6, 05014; GCN-NEXT:    s_cselect_b32 s7, s24, s195015; GCN-NEXT:    s_cselect_b32 s6, s28, s175016; GCN-NEXT:    s_xor_b64 s[6:7], s[6:7], s[22:23]5017; GCN-NEXT:    s_sub_u32 s6, s6, s225018; GCN-NEXT:    s_subb_u32 s7, s7, s225019; GCN-NEXT:    s_cbranch_execnz .LBB12_35020; GCN-NEXT:  .LBB12_2:5021; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s165022; GCN-NEXT:    s_sub_i32 s6, 0, s165023; GCN-NEXT:    s_mov_b32 s7, 05024; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v05025; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v05026; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05027; GCN-NEXT:    v_readfirstlane_b32 s17, v05028; GCN-NEXT:    s_mul_i32 s6, s6, s175029; GCN-NEXT:    s_mul_hi_u32 s6, s17, s65030; GCN-NEXT:    s_add_i32 s17, s17, s65031; GCN-NEXT:    s_mul_hi_u32 s6, s18, s175032; GCN-NEXT:    s_mul_i32 s6, s6, s165033; GCN-NEXT:    s_sub_i32 s6, s18, s65034; GCN-NEXT:    s_sub_i32 s17, s6, s165035; GCN-NEXT:    s_cmp_ge_u32 s6, s165036; GCN-NEXT:    s_cselect_b32 s6, s17, s65037; GCN-NEXT:    s_sub_i32 s17, s6, s165038; GCN-NEXT:    s_cmp_ge_u32 s6, s165039; GCN-NEXT:    s_cselect_b32 s6, s17, s65040; GCN-NEXT:  .LBB12_3:5041; GCN-NEXT:    s_or_b64 s[16:17], s[14:15], s[12:13]5042; GCN-NEXT:    s_mov_b32 s16, 05043; GCN-NEXT:    s_cmp_lg_u64 s[16:17], 05044; GCN-NEXT:    s_cbranch_scc0 .LBB12_75045; GCN-NEXT:  ; %bb.4:5046; GCN-NEXT:    s_ashr_i32 s16, s13, 315047; GCN-NEXT:    s_add_u32 s18, s12, s165048; GCN-NEXT:    s_mov_b32 s17, s165049; GCN-NEXT:    s_addc_u32 s19, s13, s165050; GCN-NEXT:    s_xor_b64 s[18:19], s[18:19], s[16:17]5051; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s185052; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s195053; GCN-NEXT:    s_sub_u32 s13, 0, s185054; GCN-NEXT:    s_subb_u32 s20, 0, s195055; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v05056; GCN-NEXT:    v_rcp_f32_e32 v0, v05057; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v05058; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v05059; GCN-NEXT:    v_trunc_f32_e32 v1, v15060; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v05061; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v15062; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05063; GCN-NEXT:    v_readfirstlane_b32 s21, v15064; GCN-NEXT:    v_readfirstlane_b32 s22, v05065; GCN-NEXT:    s_mul_i32 s23, s13, s215066; GCN-NEXT:    s_mul_hi_u32 s25, s13, s225067; GCN-NEXT:    s_mul_i32 s24, s20, s225068; GCN-NEXT:    s_add_i32 s23, s25, s235069; GCN-NEXT:    s_add_i32 s23, s23, s245070; GCN-NEXT:    s_mul_i32 s26, s13, s225071; GCN-NEXT:    s_mul_i32 s25, s22, s235072; GCN-NEXT:    s_mul_hi_u32 s27, s22, s265073; GCN-NEXT:    s_mul_hi_u32 s24, s22, s235074; GCN-NEXT:    s_add_u32 s25, s27, s255075; GCN-NEXT:    s_addc_u32 s24, 0, s245076; GCN-NEXT:    s_mul_hi_u32 s28, s21, s265077; GCN-NEXT:    s_mul_i32 s26, s21, s265078; GCN-NEXT:    s_add_u32 s25, s25, s265079; GCN-NEXT:    s_mul_hi_u32 s27, s21, s235080; GCN-NEXT:    s_addc_u32 s24, s24, s285081; GCN-NEXT:    s_addc_u32 s25, s27, 05082; GCN-NEXT:    s_mul_i32 s23, s21, s235083; GCN-NEXT:    s_add_u32 s23, s24, s235084; GCN-NEXT:    s_addc_u32 s24, 0, s255085; GCN-NEXT:    s_add_u32 s22, s22, s235086; GCN-NEXT:    s_addc_u32 s21, s21, s245087; GCN-NEXT:    s_mul_i32 s23, s13, s215088; GCN-NEXT:    s_mul_hi_u32 s24, s13, s225089; GCN-NEXT:    s_add_i32 s23, s24, s235090; GCN-NEXT:    s_mul_i32 s20, s20, s225091; GCN-NEXT:    s_add_i32 s23, s23, s205092; GCN-NEXT:    s_mul_i32 s13, s13, s225093; GCN-NEXT:    s_mul_hi_u32 s24, s21, s135094; GCN-NEXT:    s_mul_i32 s25, s21, s135095; GCN-NEXT:    s_mul_i32 s27, s22, s235096; GCN-NEXT:    s_mul_hi_u32 s13, s22, s135097; GCN-NEXT:    s_mul_hi_u32 s26, s22, s235098; GCN-NEXT:    s_add_u32 s13, s13, s275099; GCN-NEXT:    s_addc_u32 s26, 0, s265100; GCN-NEXT:    s_add_u32 s13, s13, s255101; GCN-NEXT:    s_mul_hi_u32 s20, s21, s235102; GCN-NEXT:    s_addc_u32 s13, s26, s245103; GCN-NEXT:    s_addc_u32 s20, s20, 05104; GCN-NEXT:    s_mul_i32 s23, s21, s235105; GCN-NEXT:    s_add_u32 s13, s13, s235106; GCN-NEXT:    s_addc_u32 s20, 0, s205107; GCN-NEXT:    s_add_u32 s13, s22, s135108; GCN-NEXT:    s_addc_u32 s24, s21, s205109; GCN-NEXT:    s_ashr_i32 s20, s15, 315110; GCN-NEXT:    s_add_u32 s22, s14, s205111; GCN-NEXT:    s_mov_b32 s21, s205112; GCN-NEXT:    s_addc_u32 s23, s15, s205113; GCN-NEXT:    s_xor_b64 s[22:23], s[22:23], s[20:21]5114; GCN-NEXT:    s_mul_i32 s25, s22, s245115; GCN-NEXT:    s_mul_hi_u32 s26, s22, s135116; GCN-NEXT:    s_mul_hi_u32 s15, s22, s245117; GCN-NEXT:    s_add_u32 s25, s26, s255118; GCN-NEXT:    s_addc_u32 s15, 0, s155119; GCN-NEXT:    s_mul_hi_u32 s27, s23, s135120; GCN-NEXT:    s_mul_i32 s13, s23, s135121; GCN-NEXT:    s_add_u32 s13, s25, s135122; GCN-NEXT:    s_mul_hi_u32 s26, s23, s245123; GCN-NEXT:    s_addc_u32 s13, s15, s275124; GCN-NEXT:    s_addc_u32 s15, s26, 05125; GCN-NEXT:    s_mul_i32 s24, s23, s245126; GCN-NEXT:    s_add_u32 s13, s13, s245127; GCN-NEXT:    s_addc_u32 s15, 0, s155128; GCN-NEXT:    s_mul_i32 s15, s18, s155129; GCN-NEXT:    s_mul_hi_u32 s24, s18, s135130; GCN-NEXT:    s_add_i32 s15, s24, s155131; GCN-NEXT:    s_mul_i32 s24, s19, s135132; GCN-NEXT:    s_add_i32 s15, s15, s245133; GCN-NEXT:    s_sub_i32 s26, s23, s155134; GCN-NEXT:    s_mul_i32 s13, s18, s135135; GCN-NEXT:    s_sub_u32 s13, s22, s135136; GCN-NEXT:    s_cselect_b64 s[24:25], -1, 05137; GCN-NEXT:    s_subb_u32 s22, s26, s195138; GCN-NEXT:    s_sub_u32 s28, s13, s185139; GCN-NEXT:    s_cselect_b64 s[26:27], -1, 05140; GCN-NEXT:    s_subb_u32 s29, s22, 05141; GCN-NEXT:    s_cmp_ge_u32 s29, s195142; GCN-NEXT:    s_cselect_b32 s30, -1, 05143; GCN-NEXT:    s_cmp_ge_u32 s28, s185144; GCN-NEXT:    s_cselect_b32 s31, -1, 05145; GCN-NEXT:    s_cmp_eq_u32 s29, s195146; GCN-NEXT:    s_cselect_b32 s30, s31, s305147; GCN-NEXT:    s_cmp_lg_u64 s[26:27], 05148; GCN-NEXT:    s_subb_u32 s22, s22, s195149; GCN-NEXT:    s_sub_u32 s26, s28, s185150; GCN-NEXT:    s_subb_u32 s22, s22, 05151; GCN-NEXT:    s_cmp_lg_u32 s30, 05152; GCN-NEXT:    s_cselect_b32 s26, s26, s285153; GCN-NEXT:    s_cselect_b32 s22, s22, s295154; GCN-NEXT:    s_cmp_lg_u64 s[24:25], 05155; GCN-NEXT:    s_subb_u32 s15, s23, s155156; GCN-NEXT:    s_cmp_ge_u32 s15, s195157; GCN-NEXT:    s_cselect_b32 s23, -1, 05158; GCN-NEXT:    s_cmp_ge_u32 s13, s185159; GCN-NEXT:    s_cselect_b32 s18, -1, 05160; GCN-NEXT:    s_cmp_eq_u32 s15, s195161; GCN-NEXT:    s_cselect_b32 s18, s18, s235162; GCN-NEXT:    s_cmp_lg_u32 s18, 05163; GCN-NEXT:    s_cselect_b32 s19, s22, s155164; GCN-NEXT:    s_cselect_b32 s18, s26, s135165; GCN-NEXT:    s_xor_b64 s[18:19], s[18:19], s[20:21]5166; GCN-NEXT:    s_sub_u32 s18, s18, s205167; GCN-NEXT:    s_subb_u32 s19, s19, s205168; GCN-NEXT:    s_cbranch_execnz .LBB12_85169; GCN-NEXT:  .LBB12_5:5170; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s125171; GCN-NEXT:    s_sub_i32 s13, 0, s125172; GCN-NEXT:    v_mov_b32_e32 v3, 05173; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v05174; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v05175; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05176; GCN-NEXT:    v_mul_lo_u32 v1, s13, v05177; GCN-NEXT:    v_mul_hi_u32 v1, v0, v15178; GCN-NEXT:    v_add_u32_e32 v0, v0, v15179; GCN-NEXT:    v_mul_hi_u32 v0, s14, v05180; GCN-NEXT:    v_mul_lo_u32 v0, v0, s125181; GCN-NEXT:    v_sub_u32_e32 v0, s14, v05182; GCN-NEXT:    v_subrev_u32_e32 v1, s12, v05183; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v05184; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc5185; GCN-NEXT:    v_subrev_u32_e32 v1, s12, v05186; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s12, v05187; GCN-NEXT:    v_cndmask_b32_e32 v2, v0, v1, vcc5188; GCN-NEXT:    s_branch .LBB12_95189; GCN-NEXT:  .LBB12_6:5190; GCN-NEXT:    ; implicit-def: $sgpr6_sgpr75191; GCN-NEXT:    s_branch .LBB12_25192; GCN-NEXT:  .LBB12_7:5193; GCN-NEXT:    ; implicit-def: $sgpr18_sgpr195194; GCN-NEXT:    s_branch .LBB12_55195; GCN-NEXT:  .LBB12_8:5196; GCN-NEXT:    v_mov_b32_e32 v2, s185197; GCN-NEXT:    v_mov_b32_e32 v3, s195198; GCN-NEXT:  .LBB12_9:5199; GCN-NEXT:    s_or_b64 s[12:13], s[10:11], s[8:9]5200; GCN-NEXT:    s_mov_b32 s12, 05201; GCN-NEXT:    s_cmp_lg_u64 s[12:13], 05202; GCN-NEXT:    s_cbranch_scc0 .LBB12_125203; GCN-NEXT:  ; %bb.10:5204; GCN-NEXT:    s_ashr_i32 s12, s9, 315205; GCN-NEXT:    s_add_u32 s14, s8, s125206; GCN-NEXT:    s_mov_b32 s13, s125207; GCN-NEXT:    s_addc_u32 s15, s9, s125208; GCN-NEXT:    s_xor_b64 s[14:15], s[14:15], s[12:13]5209; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s145210; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s155211; GCN-NEXT:    s_sub_u32 s9, 0, s145212; GCN-NEXT:    s_subb_u32 s16, 0, s155213; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v05214; GCN-NEXT:    v_rcp_f32_e32 v0, v05215; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v05216; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v05217; GCN-NEXT:    v_trunc_f32_e32 v1, v15218; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v05219; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v15220; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05221; GCN-NEXT:    v_readfirstlane_b32 s17, v15222; GCN-NEXT:    v_readfirstlane_b32 s18, v05223; GCN-NEXT:    s_mul_i32 s19, s9, s175224; GCN-NEXT:    s_mul_hi_u32 s21, s9, s185225; GCN-NEXT:    s_mul_i32 s20, s16, s185226; GCN-NEXT:    s_add_i32 s19, s21, s195227; GCN-NEXT:    s_add_i32 s19, s19, s205228; GCN-NEXT:    s_mul_i32 s22, s9, s185229; GCN-NEXT:    s_mul_i32 s21, s18, s195230; GCN-NEXT:    s_mul_hi_u32 s23, s18, s225231; GCN-NEXT:    s_mul_hi_u32 s20, s18, s195232; GCN-NEXT:    s_add_u32 s21, s23, s215233; GCN-NEXT:    s_addc_u32 s20, 0, s205234; GCN-NEXT:    s_mul_hi_u32 s24, s17, s225235; GCN-NEXT:    s_mul_i32 s22, s17, s225236; GCN-NEXT:    s_add_u32 s21, s21, s225237; GCN-NEXT:    s_mul_hi_u32 s23, s17, s195238; GCN-NEXT:    s_addc_u32 s20, s20, s245239; GCN-NEXT:    s_addc_u32 s21, s23, 05240; GCN-NEXT:    s_mul_i32 s19, s17, s195241; GCN-NEXT:    s_add_u32 s19, s20, s195242; GCN-NEXT:    s_addc_u32 s20, 0, s215243; GCN-NEXT:    s_add_u32 s18, s18, s195244; GCN-NEXT:    s_addc_u32 s17, s17, s205245; GCN-NEXT:    s_mul_i32 s19, s9, s175246; GCN-NEXT:    s_mul_hi_u32 s20, s9, s185247; GCN-NEXT:    s_add_i32 s19, s20, s195248; GCN-NEXT:    s_mul_i32 s16, s16, s185249; GCN-NEXT:    s_add_i32 s19, s19, s165250; GCN-NEXT:    s_mul_i32 s9, s9, s185251; GCN-NEXT:    s_mul_hi_u32 s20, s17, s95252; GCN-NEXT:    s_mul_i32 s21, s17, s95253; GCN-NEXT:    s_mul_i32 s23, s18, s195254; GCN-NEXT:    s_mul_hi_u32 s9, s18, s95255; GCN-NEXT:    s_mul_hi_u32 s22, s18, s195256; GCN-NEXT:    s_add_u32 s9, s9, s235257; GCN-NEXT:    s_addc_u32 s22, 0, s225258; GCN-NEXT:    s_add_u32 s9, s9, s215259; GCN-NEXT:    s_mul_hi_u32 s16, s17, s195260; GCN-NEXT:    s_addc_u32 s9, s22, s205261; GCN-NEXT:    s_addc_u32 s16, s16, 05262; GCN-NEXT:    s_mul_i32 s19, s17, s195263; GCN-NEXT:    s_add_u32 s9, s9, s195264; GCN-NEXT:    s_addc_u32 s16, 0, s165265; GCN-NEXT:    s_add_u32 s9, s18, s95266; GCN-NEXT:    s_addc_u32 s20, s17, s165267; GCN-NEXT:    s_ashr_i32 s16, s11, 315268; GCN-NEXT:    s_add_u32 s18, s10, s165269; GCN-NEXT:    s_mov_b32 s17, s165270; GCN-NEXT:    s_addc_u32 s19, s11, s165271; GCN-NEXT:    s_xor_b64 s[18:19], s[18:19], s[16:17]5272; GCN-NEXT:    s_mul_i32 s21, s18, s205273; GCN-NEXT:    s_mul_hi_u32 s22, s18, s95274; GCN-NEXT:    s_mul_hi_u32 s11, s18, s205275; GCN-NEXT:    s_add_u32 s21, s22, s215276; GCN-NEXT:    s_addc_u32 s11, 0, s115277; GCN-NEXT:    s_mul_hi_u32 s23, s19, s95278; GCN-NEXT:    s_mul_i32 s9, s19, s95279; GCN-NEXT:    s_add_u32 s9, s21, s95280; GCN-NEXT:    s_mul_hi_u32 s22, s19, s205281; GCN-NEXT:    s_addc_u32 s9, s11, s235282; GCN-NEXT:    s_addc_u32 s11, s22, 05283; GCN-NEXT:    s_mul_i32 s20, s19, s205284; GCN-NEXT:    s_add_u32 s9, s9, s205285; GCN-NEXT:    s_addc_u32 s11, 0, s115286; GCN-NEXT:    s_mul_i32 s11, s14, s115287; GCN-NEXT:    s_mul_hi_u32 s20, s14, s95288; GCN-NEXT:    s_add_i32 s11, s20, s115289; GCN-NEXT:    s_mul_i32 s20, s15, s95290; GCN-NEXT:    s_add_i32 s11, s11, s205291; GCN-NEXT:    s_sub_i32 s22, s19, s115292; GCN-NEXT:    s_mul_i32 s9, s14, s95293; GCN-NEXT:    s_sub_u32 s9, s18, s95294; GCN-NEXT:    s_cselect_b64 s[20:21], -1, 05295; GCN-NEXT:    s_subb_u32 s18, s22, s155296; GCN-NEXT:    s_sub_u32 s24, s9, s145297; GCN-NEXT:    s_cselect_b64 s[22:23], -1, 05298; GCN-NEXT:    s_subb_u32 s25, s18, 05299; GCN-NEXT:    s_cmp_ge_u32 s25, s155300; GCN-NEXT:    s_cselect_b32 s26, -1, 05301; GCN-NEXT:    s_cmp_ge_u32 s24, s145302; GCN-NEXT:    s_cselect_b32 s27, -1, 05303; GCN-NEXT:    s_cmp_eq_u32 s25, s155304; GCN-NEXT:    s_cselect_b32 s26, s27, s265305; GCN-NEXT:    s_cmp_lg_u64 s[22:23], 05306; GCN-NEXT:    s_subb_u32 s18, s18, s155307; GCN-NEXT:    s_sub_u32 s22, s24, s145308; GCN-NEXT:    s_subb_u32 s18, s18, 05309; GCN-NEXT:    s_cmp_lg_u32 s26, 05310; GCN-NEXT:    s_cselect_b32 s22, s22, s245311; GCN-NEXT:    s_cselect_b32 s18, s18, s255312; GCN-NEXT:    s_cmp_lg_u64 s[20:21], 05313; GCN-NEXT:    s_subb_u32 s11, s19, s115314; GCN-NEXT:    s_cmp_ge_u32 s11, s155315; GCN-NEXT:    s_cselect_b32 s19, -1, 05316; GCN-NEXT:    s_cmp_ge_u32 s9, s145317; GCN-NEXT:    s_cselect_b32 s14, -1, 05318; GCN-NEXT:    s_cmp_eq_u32 s11, s155319; GCN-NEXT:    s_cselect_b32 s14, s14, s195320; GCN-NEXT:    s_cmp_lg_u32 s14, 05321; GCN-NEXT:    s_cselect_b32 s15, s18, s115322; GCN-NEXT:    s_cselect_b32 s14, s22, s95323; GCN-NEXT:    s_xor_b64 s[14:15], s[14:15], s[16:17]5324; GCN-NEXT:    s_sub_u32 s14, s14, s165325; GCN-NEXT:    s_subb_u32 s15, s15, s165326; GCN-NEXT:    s_cbranch_execnz .LBB12_135327; GCN-NEXT:  .LBB12_11:5328; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s85329; GCN-NEXT:    s_sub_i32 s9, 0, s85330; GCN-NEXT:    v_mov_b32_e32 v5, 05331; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v05332; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v05333; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05334; GCN-NEXT:    v_mul_lo_u32 v1, s9, v05335; GCN-NEXT:    v_mul_hi_u32 v1, v0, v15336; GCN-NEXT:    v_add_u32_e32 v0, v0, v15337; GCN-NEXT:    v_mul_hi_u32 v0, s10, v05338; GCN-NEXT:    v_mul_lo_u32 v0, v0, s85339; GCN-NEXT:    v_sub_u32_e32 v0, s10, v05340; GCN-NEXT:    v_subrev_u32_e32 v1, s8, v05341; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v05342; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc5343; GCN-NEXT:    v_subrev_u32_e32 v1, s8, v05344; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s8, v05345; GCN-NEXT:    v_cndmask_b32_e32 v4, v0, v1, vcc5346; GCN-NEXT:    s_branch .LBB12_145347; GCN-NEXT:  .LBB12_12:5348; GCN-NEXT:    ; implicit-def: $sgpr14_sgpr155349; GCN-NEXT:    s_branch .LBB12_115350; GCN-NEXT:  .LBB12_13:5351; GCN-NEXT:    v_mov_b32_e32 v4, s145352; GCN-NEXT:    v_mov_b32_e32 v5, s155353; GCN-NEXT:  .LBB12_14:5354; GCN-NEXT:    s_or_b64 s[8:9], s[4:5], s[2:3]5355; GCN-NEXT:    s_mov_b32 s8, 05356; GCN-NEXT:    s_cmp_lg_u64 s[8:9], 05357; GCN-NEXT:    s_cbranch_scc0 .LBB12_175358; GCN-NEXT:  ; %bb.15:5359; GCN-NEXT:    s_ashr_i32 s8, s3, 315360; GCN-NEXT:    s_add_u32 s10, s2, s85361; GCN-NEXT:    s_mov_b32 s9, s85362; GCN-NEXT:    s_addc_u32 s11, s3, s85363; GCN-NEXT:    s_xor_b64 s[10:11], s[10:11], s[8:9]5364; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s105365; GCN-NEXT:    v_cvt_f32_u32_e32 v1, s115366; GCN-NEXT:    s_sub_u32 s3, 0, s105367; GCN-NEXT:    s_subb_u32 s12, 0, s115368; GCN-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v05369; GCN-NEXT:    v_rcp_f32_e32 v0, v05370; GCN-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v05371; GCN-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v05372; GCN-NEXT:    v_trunc_f32_e32 v1, v15373; GCN-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v05374; GCN-NEXT:    v_cvt_u32_f32_e32 v1, v15375; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05376; GCN-NEXT:    v_readfirstlane_b32 s13, v15377; GCN-NEXT:    v_readfirstlane_b32 s14, v05378; GCN-NEXT:    s_mul_i32 s15, s3, s135379; GCN-NEXT:    s_mul_hi_u32 s17, s3, s145380; GCN-NEXT:    s_mul_i32 s16, s12, s145381; GCN-NEXT:    s_add_i32 s15, s17, s155382; GCN-NEXT:    s_add_i32 s15, s15, s165383; GCN-NEXT:    s_mul_i32 s18, s3, s145384; GCN-NEXT:    s_mul_i32 s17, s14, s155385; GCN-NEXT:    s_mul_hi_u32 s19, s14, s185386; GCN-NEXT:    s_mul_hi_u32 s16, s14, s155387; GCN-NEXT:    s_add_u32 s17, s19, s175388; GCN-NEXT:    s_addc_u32 s16, 0, s165389; GCN-NEXT:    s_mul_hi_u32 s20, s13, s185390; GCN-NEXT:    s_mul_i32 s18, s13, s185391; GCN-NEXT:    s_add_u32 s17, s17, s185392; GCN-NEXT:    s_mul_hi_u32 s19, s13, s155393; GCN-NEXT:    s_addc_u32 s16, s16, s205394; GCN-NEXT:    s_addc_u32 s17, s19, 05395; GCN-NEXT:    s_mul_i32 s15, s13, s155396; GCN-NEXT:    s_add_u32 s15, s16, s155397; GCN-NEXT:    s_addc_u32 s16, 0, s175398; GCN-NEXT:    s_add_u32 s14, s14, s155399; GCN-NEXT:    s_addc_u32 s13, s13, s165400; GCN-NEXT:    s_mul_i32 s15, s3, s135401; GCN-NEXT:    s_mul_hi_u32 s16, s3, s145402; GCN-NEXT:    s_add_i32 s15, s16, s155403; GCN-NEXT:    s_mul_i32 s12, s12, s145404; GCN-NEXT:    s_add_i32 s15, s15, s125405; GCN-NEXT:    s_mul_i32 s3, s3, s145406; GCN-NEXT:    s_mul_hi_u32 s16, s13, s35407; GCN-NEXT:    s_mul_i32 s17, s13, s35408; GCN-NEXT:    s_mul_i32 s19, s14, s155409; GCN-NEXT:    s_mul_hi_u32 s3, s14, s35410; GCN-NEXT:    s_mul_hi_u32 s18, s14, s155411; GCN-NEXT:    s_add_u32 s3, s3, s195412; GCN-NEXT:    s_addc_u32 s18, 0, s185413; GCN-NEXT:    s_add_u32 s3, s3, s175414; GCN-NEXT:    s_mul_hi_u32 s12, s13, s155415; GCN-NEXT:    s_addc_u32 s3, s18, s165416; GCN-NEXT:    s_addc_u32 s12, s12, 05417; GCN-NEXT:    s_mul_i32 s15, s13, s155418; GCN-NEXT:    s_add_u32 s3, s3, s155419; GCN-NEXT:    s_addc_u32 s12, 0, s125420; GCN-NEXT:    s_add_u32 s3, s14, s35421; GCN-NEXT:    s_addc_u32 s16, s13, s125422; GCN-NEXT:    s_ashr_i32 s12, s5, 315423; GCN-NEXT:    s_add_u32 s14, s4, s125424; GCN-NEXT:    s_mov_b32 s13, s125425; GCN-NEXT:    s_addc_u32 s15, s5, s125426; GCN-NEXT:    s_xor_b64 s[14:15], s[14:15], s[12:13]5427; GCN-NEXT:    s_mul_i32 s17, s14, s165428; GCN-NEXT:    s_mul_hi_u32 s18, s14, s35429; GCN-NEXT:    s_mul_hi_u32 s5, s14, s165430; GCN-NEXT:    s_add_u32 s17, s18, s175431; GCN-NEXT:    s_addc_u32 s5, 0, s55432; GCN-NEXT:    s_mul_hi_u32 s19, s15, s35433; GCN-NEXT:    s_mul_i32 s3, s15, s35434; GCN-NEXT:    s_add_u32 s3, s17, s35435; GCN-NEXT:    s_mul_hi_u32 s18, s15, s165436; GCN-NEXT:    s_addc_u32 s3, s5, s195437; GCN-NEXT:    s_addc_u32 s5, s18, 05438; GCN-NEXT:    s_mul_i32 s16, s15, s165439; GCN-NEXT:    s_add_u32 s3, s3, s165440; GCN-NEXT:    s_addc_u32 s5, 0, s55441; GCN-NEXT:    s_mul_i32 s5, s10, s55442; GCN-NEXT:    s_mul_hi_u32 s16, s10, s35443; GCN-NEXT:    s_add_i32 s5, s16, s55444; GCN-NEXT:    s_mul_i32 s16, s11, s35445; GCN-NEXT:    s_add_i32 s5, s5, s165446; GCN-NEXT:    s_sub_i32 s18, s15, s55447; GCN-NEXT:    s_mul_i32 s3, s10, s35448; GCN-NEXT:    s_sub_u32 s3, s14, s35449; GCN-NEXT:    s_cselect_b64 s[16:17], -1, 05450; GCN-NEXT:    s_subb_u32 s14, s18, s115451; GCN-NEXT:    s_sub_u32 s20, s3, s105452; GCN-NEXT:    s_cselect_b64 s[18:19], -1, 05453; GCN-NEXT:    s_subb_u32 s21, s14, 05454; GCN-NEXT:    s_cmp_ge_u32 s21, s115455; GCN-NEXT:    s_cselect_b32 s22, -1, 05456; GCN-NEXT:    s_cmp_ge_u32 s20, s105457; GCN-NEXT:    s_cselect_b32 s23, -1, 05458; GCN-NEXT:    s_cmp_eq_u32 s21, s115459; GCN-NEXT:    s_cselect_b32 s22, s23, s225460; GCN-NEXT:    s_cmp_lg_u64 s[18:19], 05461; GCN-NEXT:    s_subb_u32 s14, s14, s115462; GCN-NEXT:    s_sub_u32 s18, s20, s105463; GCN-NEXT:    s_subb_u32 s14, s14, 05464; GCN-NEXT:    s_cmp_lg_u32 s22, 05465; GCN-NEXT:    s_cselect_b32 s18, s18, s205466; GCN-NEXT:    s_cselect_b32 s14, s14, s215467; GCN-NEXT:    s_cmp_lg_u64 s[16:17], 05468; GCN-NEXT:    s_subb_u32 s5, s15, s55469; GCN-NEXT:    s_cmp_ge_u32 s5, s115470; GCN-NEXT:    s_cselect_b32 s15, -1, 05471; GCN-NEXT:    s_cmp_ge_u32 s3, s105472; GCN-NEXT:    s_cselect_b32 s10, -1, 05473; GCN-NEXT:    s_cmp_eq_u32 s5, s115474; GCN-NEXT:    s_cselect_b32 s10, s10, s155475; GCN-NEXT:    s_cmp_lg_u32 s10, 05476; GCN-NEXT:    s_cselect_b32 s11, s14, s55477; GCN-NEXT:    s_cselect_b32 s10, s18, s35478; GCN-NEXT:    s_xor_b64 s[10:11], s[10:11], s[12:13]5479; GCN-NEXT:    s_sub_u32 s10, s10, s125480; GCN-NEXT:    s_subb_u32 s11, s11, s125481; GCN-NEXT:    s_cbranch_execnz .LBB12_185482; GCN-NEXT:  .LBB12_16:5483; GCN-NEXT:    v_cvt_f32_u32_e32 v0, s25484; GCN-NEXT:    s_sub_i32 s3, 0, s25485; GCN-NEXT:    v_mov_b32_e32 v7, 05486; GCN-NEXT:    v_rcp_iflag_f32_e32 v0, v05487; GCN-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v05488; GCN-NEXT:    v_cvt_u32_f32_e32 v0, v05489; GCN-NEXT:    v_mul_lo_u32 v1, s3, v05490; GCN-NEXT:    v_mul_hi_u32 v1, v0, v15491; GCN-NEXT:    v_add_u32_e32 v0, v0, v15492; GCN-NEXT:    v_mul_hi_u32 v0, s4, v05493; GCN-NEXT:    v_mul_lo_u32 v0, v0, s25494; GCN-NEXT:    v_sub_u32_e32 v0, s4, v05495; GCN-NEXT:    v_subrev_u32_e32 v1, s2, v05496; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v05497; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc5498; GCN-NEXT:    v_subrev_u32_e32 v1, s2, v05499; GCN-NEXT:    v_cmp_le_u32_e32 vcc, s2, v05500; GCN-NEXT:    v_cndmask_b32_e32 v6, v0, v1, vcc5501; GCN-NEXT:    s_branch .LBB12_195502; GCN-NEXT:  .LBB12_17:5503; GCN-NEXT:    ; implicit-def: $sgpr10_sgpr115504; GCN-NEXT:    s_branch .LBB12_165505; GCN-NEXT:  .LBB12_18:5506; GCN-NEXT:    v_mov_b32_e32 v6, s105507; GCN-NEXT:    v_mov_b32_e32 v7, s115508; GCN-NEXT:  .LBB12_19:5509; GCN-NEXT:    v_mov_b32_e32 v8, 05510; GCN-NEXT:    v_mov_b32_e32 v0, s65511; GCN-NEXT:    v_mov_b32_e32 v1, s75512; GCN-NEXT:    global_store_dwordx4 v8, v[4:7], s[0:1] offset:165513; GCN-NEXT:    global_store_dwordx4 v8, v[0:3], s[0:1]5514; GCN-NEXT:    s_endpgm5515;5516; TAHITI-LABEL: srem_v4i64:5517; TAHITI:       ; %bb.0:5518; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x95519; TAHITI-NEXT:    s_mov_b32 s3, 0xf0005520; TAHITI-NEXT:    s_mov_b32 s2, -15521; TAHITI-NEXT:    v_mov_b32_e32 v8, 05522; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)5523; TAHITI-NEXT:    s_mov_b32 s0, s65524; TAHITI-NEXT:    s_mov_b32 s1, s75525; TAHITI-NEXT:    buffer_load_dwordx4 v[10:13], off, s[0:3], 0 offset:325526; TAHITI-NEXT:    buffer_load_dwordx4 v[14:17], off, s[0:3], 05527; TAHITI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[0:3], 0 offset:165528; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[0:3], 0 offset:485529; TAHITI-NEXT:    s_waitcnt vmcnt(2)5530; TAHITI-NEXT:    v_or_b32_e32 v9, v15, v115531; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[8:9]5532; TAHITI-NEXT:    s_cbranch_vccz .LBB12_135533; TAHITI-NEXT:  ; %bb.1:5534; TAHITI-NEXT:    v_ashrrev_i32_e32 v8, 31, v115535; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v10, v85536; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, v11, v8, vcc5537; TAHITI-NEXT:    v_xor_b32_e32 v9, v9, v85538; TAHITI-NEXT:    v_xor_b32_e32 v8, v11, v85539; TAHITI-NEXT:    v_cvt_f32_u32_e32 v11, v95540; TAHITI-NEXT:    v_cvt_f32_u32_e32 v18, v85541; TAHITI-NEXT:    v_sub_i32_e32 v19, vcc, 0, v95542; TAHITI-NEXT:    v_subb_u32_e32 v20, vcc, 0, v8, vcc5543; TAHITI-NEXT:    v_madmk_f32 v11, v18, 0x4f800000, v115544; TAHITI-NEXT:    v_rcp_f32_e32 v11, v115545; TAHITI-NEXT:    v_mul_f32_e32 v11, 0x5f7ffffc, v115546; TAHITI-NEXT:    v_mul_f32_e32 v18, 0x2f800000, v115547; TAHITI-NEXT:    v_trunc_f32_e32 v18, v185548; TAHITI-NEXT:    v_madmk_f32 v11, v18, 0xcf800000, v115549; TAHITI-NEXT:    v_cvt_u32_f32_e32 v18, v185550; TAHITI-NEXT:    v_cvt_u32_f32_e32 v11, v115551; TAHITI-NEXT:    v_mul_lo_u32 v22, v19, v185552; TAHITI-NEXT:    v_mul_hi_u32 v21, v19, v115553; TAHITI-NEXT:    v_mul_lo_u32 v23, v20, v115554; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v22, v215555; TAHITI-NEXT:    v_mul_lo_u32 v22, v19, v115556; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v21, v235557; TAHITI-NEXT:    v_mul_lo_u32 v23, v11, v215558; TAHITI-NEXT:    v_mul_hi_u32 v24, v11, v225559; TAHITI-NEXT:    v_mul_hi_u32 v25, v11, v215560; TAHITI-NEXT:    v_mul_hi_u32 v26, v18, v215561; TAHITI-NEXT:    v_mul_lo_u32 v21, v18, v215562; TAHITI-NEXT:    v_add_i32_e32 v23, vcc, v24, v235563; TAHITI-NEXT:    v_addc_u32_e32 v24, vcc, 0, v25, vcc5564; TAHITI-NEXT:    v_mul_lo_u32 v25, v18, v225565; TAHITI-NEXT:    v_mul_hi_u32 v22, v18, v225566; TAHITI-NEXT:    v_add_i32_e32 v23, vcc, v23, v255567; TAHITI-NEXT:    v_addc_u32_e32 v22, vcc, v24, v22, vcc5568; TAHITI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v26, vcc5569; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v22, v215570; TAHITI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v23, vcc5571; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v215572; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, v18, v22, vcc5573; TAHITI-NEXT:    v_mul_lo_u32 v21, v19, v185574; TAHITI-NEXT:    v_mul_hi_u32 v22, v19, v115575; TAHITI-NEXT:    v_mul_lo_u32 v20, v20, v115576; TAHITI-NEXT:    v_mul_lo_u32 v19, v19, v115577; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v21, v225578; TAHITI-NEXT:    v_add_i32_e32 v20, vcc, v20, v215579; TAHITI-NEXT:    v_mul_lo_u32 v23, v11, v205580; TAHITI-NEXT:    v_mul_hi_u32 v24, v11, v195581; TAHITI-NEXT:    v_mul_hi_u32 v25, v11, v205582; TAHITI-NEXT:    v_mul_hi_u32 v22, v18, v195583; TAHITI-NEXT:    v_mul_lo_u32 v19, v18, v195584; TAHITI-NEXT:    v_mul_hi_u32 v21, v18, v205585; TAHITI-NEXT:    v_add_i32_e32 v23, vcc, v24, v235586; TAHITI-NEXT:    v_addc_u32_e32 v24, vcc, 0, v25, vcc5587; TAHITI-NEXT:    v_mul_lo_u32 v20, v18, v205588; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v23, v195589; TAHITI-NEXT:    v_addc_u32_e32 v19, vcc, v24, v22, vcc5590; TAHITI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc5591; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v19, v205592; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v21, vcc5593; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v195594; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, v18, v20, vcc5595; TAHITI-NEXT:    v_ashrrev_i32_e32 v19, 31, v155596; TAHITI-NEXT:    v_add_i32_e32 v20, vcc, v14, v195597; TAHITI-NEXT:    v_xor_b32_e32 v20, v20, v195598; TAHITI-NEXT:    v_mul_lo_u32 v21, v20, v185599; TAHITI-NEXT:    v_mul_hi_u32 v22, v20, v115600; TAHITI-NEXT:    v_mul_hi_u32 v23, v20, v185601; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, v15, v19, vcc5602; TAHITI-NEXT:    v_xor_b32_e32 v15, v15, v195603; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v22, v215604; TAHITI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v23, vcc5605; TAHITI-NEXT:    v_mul_lo_u32 v23, v15, v115606; TAHITI-NEXT:    v_mul_hi_u32 v11, v15, v115607; TAHITI-NEXT:    v_mul_hi_u32 v24, v15, v185608; TAHITI-NEXT:    v_mul_lo_u32 v18, v15, v185609; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v21, v235610; TAHITI-NEXT:    v_addc_u32_e32 v11, vcc, v22, v11, vcc5611; TAHITI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v24, vcc5612; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v11, v185613; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v21, vcc5614; TAHITI-NEXT:    v_mul_lo_u32 v18, v9, v185615; TAHITI-NEXT:    v_mul_hi_u32 v21, v9, v115616; TAHITI-NEXT:    v_mul_lo_u32 v22, v8, v115617; TAHITI-NEXT:    v_mul_lo_u32 v11, v9, v115618; TAHITI-NEXT:    v_add_i32_e32 v18, vcc, v18, v215619; TAHITI-NEXT:    v_add_i32_e32 v18, vcc, v22, v185620; TAHITI-NEXT:    v_sub_i32_e32 v21, vcc, v15, v185621; TAHITI-NEXT:    v_sub_i32_e32 v11, vcc, v20, v115622; TAHITI-NEXT:    v_subb_u32_e64 v20, s[0:1], v21, v8, vcc5623; TAHITI-NEXT:    v_sub_i32_e64 v21, s[0:1], v11, v95624; TAHITI-NEXT:    v_subbrev_u32_e64 v22, s[2:3], 0, v20, s[0:1]5625; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v22, v85626; TAHITI-NEXT:    v_cndmask_b32_e64 v23, 0, -1, s[2:3]5627; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v21, v95628; TAHITI-NEXT:    v_subb_u32_e32 v15, vcc, v15, v18, vcc5629; TAHITI-NEXT:    v_cndmask_b32_e64 v24, 0, -1, s[2:3]5630; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v22, v85631; TAHITI-NEXT:    v_subb_u32_e64 v20, s[0:1], v20, v8, s[0:1]5632; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v15, v85633; TAHITI-NEXT:    v_cndmask_b32_e64 v23, v23, v24, s[2:3]5634; TAHITI-NEXT:    v_sub_i32_e64 v24, s[0:1], v21, v95635; TAHITI-NEXT:    v_cndmask_b32_e64 v18, 0, -1, vcc5636; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v11, v95637; TAHITI-NEXT:    v_subbrev_u32_e64 v20, s[0:1], 0, v20, s[0:1]5638; TAHITI-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc5639; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v15, v85640; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v235641; TAHITI-NEXT:    v_cndmask_b32_e32 v8, v18, v9, vcc5642; TAHITI-NEXT:    v_cndmask_b32_e64 v21, v21, v24, s[0:1]5643; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v85644; TAHITI-NEXT:    v_cndmask_b32_e64 v20, v22, v20, s[0:1]5645; TAHITI-NEXT:    v_cndmask_b32_e32 v9, v11, v21, vcc5646; TAHITI-NEXT:    v_cndmask_b32_e32 v8, v15, v20, vcc5647; TAHITI-NEXT:    v_xor_b32_e32 v9, v9, v195648; TAHITI-NEXT:    v_xor_b32_e32 v11, v8, v195649; TAHITI-NEXT:    v_sub_i32_e32 v8, vcc, v9, v195650; TAHITI-NEXT:    v_subb_u32_e32 v9, vcc, v11, v19, vcc5651; TAHITI-NEXT:    s_cbranch_execnz .LBB12_35652; TAHITI-NEXT:  .LBB12_2:5653; TAHITI-NEXT:    v_cvt_f32_u32_e32 v8, v105654; TAHITI-NEXT:    v_sub_i32_e32 v9, vcc, 0, v105655; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v8, v85656; TAHITI-NEXT:    v_mul_f32_e32 v8, 0x4f7ffffe, v85657; TAHITI-NEXT:    v_cvt_u32_f32_e32 v8, v85658; TAHITI-NEXT:    v_mul_lo_u32 v9, v9, v85659; TAHITI-NEXT:    v_mul_hi_u32 v9, v8, v95660; TAHITI-NEXT:    v_add_i32_e32 v8, vcc, v8, v95661; TAHITI-NEXT:    v_mul_hi_u32 v8, v14, v85662; TAHITI-NEXT:    v_mul_lo_u32 v8, v8, v105663; TAHITI-NEXT:    v_sub_i32_e32 v8, vcc, v14, v85664; TAHITI-NEXT:    v_sub_i32_e32 v9, vcc, v8, v105665; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v105666; TAHITI-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc5667; TAHITI-NEXT:    v_sub_i32_e32 v9, vcc, v8, v105668; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v105669; TAHITI-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc5670; TAHITI-NEXT:    v_mov_b32_e32 v9, 05671; TAHITI-NEXT:  .LBB12_3:5672; TAHITI-NEXT:    v_or_b32_e32 v11, v17, v135673; TAHITI-NEXT:    v_mov_b32_e32 v10, 05674; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[10:11]5675; TAHITI-NEXT:    s_cbranch_vccz .LBB12_145676; TAHITI-NEXT:  ; %bb.4:5677; TAHITI-NEXT:    v_ashrrev_i32_e32 v10, 31, v135678; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v12, v105679; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, v13, v10, vcc5680; TAHITI-NEXT:    v_xor_b32_e32 v11, v11, v105681; TAHITI-NEXT:    v_xor_b32_e32 v10, v13, v105682; TAHITI-NEXT:    v_cvt_f32_u32_e32 v13, v115683; TAHITI-NEXT:    v_cvt_f32_u32_e32 v14, v105684; TAHITI-NEXT:    v_sub_i32_e32 v15, vcc, 0, v115685; TAHITI-NEXT:    v_subb_u32_e32 v18, vcc, 0, v10, vcc5686; TAHITI-NEXT:    v_madmk_f32 v13, v14, 0x4f800000, v135687; TAHITI-NEXT:    v_rcp_f32_e32 v13, v135688; TAHITI-NEXT:    v_mul_f32_e32 v13, 0x5f7ffffc, v135689; TAHITI-NEXT:    v_mul_f32_e32 v14, 0x2f800000, v135690; TAHITI-NEXT:    v_trunc_f32_e32 v14, v145691; TAHITI-NEXT:    v_madmk_f32 v13, v14, 0xcf800000, v135692; TAHITI-NEXT:    v_cvt_u32_f32_e32 v14, v145693; TAHITI-NEXT:    v_cvt_u32_f32_e32 v13, v135694; TAHITI-NEXT:    v_mul_lo_u32 v20, v15, v145695; TAHITI-NEXT:    v_mul_hi_u32 v19, v15, v135696; TAHITI-NEXT:    v_mul_lo_u32 v21, v18, v135697; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v20, v195698; TAHITI-NEXT:    v_mul_lo_u32 v20, v15, v135699; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v19, v215700; TAHITI-NEXT:    v_mul_lo_u32 v21, v13, v195701; TAHITI-NEXT:    v_mul_hi_u32 v22, v13, v205702; TAHITI-NEXT:    v_mul_hi_u32 v23, v13, v195703; TAHITI-NEXT:    v_mul_hi_u32 v24, v14, v195704; TAHITI-NEXT:    v_mul_lo_u32 v19, v14, v195705; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v22, v215706; TAHITI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v23, vcc5707; TAHITI-NEXT:    v_mul_lo_u32 v23, v14, v205708; TAHITI-NEXT:    v_mul_hi_u32 v20, v14, v205709; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v21, v235710; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, v22, v20, vcc5711; TAHITI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v24, vcc5712; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v20, v195713; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v21, vcc5714; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v195715; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, v14, v20, vcc5716; TAHITI-NEXT:    v_mul_lo_u32 v19, v15, v145717; TAHITI-NEXT:    v_mul_hi_u32 v20, v15, v135718; TAHITI-NEXT:    v_mul_lo_u32 v18, v18, v135719; TAHITI-NEXT:    v_mul_lo_u32 v15, v15, v135720; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v19, v205721; TAHITI-NEXT:    v_add_i32_e32 v18, vcc, v18, v195722; TAHITI-NEXT:    v_mul_lo_u32 v21, v13, v185723; TAHITI-NEXT:    v_mul_hi_u32 v22, v13, v155724; TAHITI-NEXT:    v_mul_hi_u32 v23, v13, v185725; TAHITI-NEXT:    v_mul_hi_u32 v20, v14, v155726; TAHITI-NEXT:    v_mul_lo_u32 v15, v14, v155727; TAHITI-NEXT:    v_mul_hi_u32 v19, v14, v185728; TAHITI-NEXT:    v_add_i32_e32 v21, vcc, v22, v215729; TAHITI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v23, vcc5730; TAHITI-NEXT:    v_mul_lo_u32 v18, v14, v185731; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v21, v155732; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, v22, v20, vcc5733; TAHITI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc5734; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v185735; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v19, vcc5736; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v155737; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, v14, v18, vcc5738; TAHITI-NEXT:    v_ashrrev_i32_e32 v15, 31, v175739; TAHITI-NEXT:    v_add_i32_e32 v18, vcc, v16, v155740; TAHITI-NEXT:    v_xor_b32_e32 v18, v18, v155741; TAHITI-NEXT:    v_mul_lo_u32 v19, v18, v145742; TAHITI-NEXT:    v_mul_hi_u32 v20, v18, v135743; TAHITI-NEXT:    v_mul_hi_u32 v21, v18, v145744; TAHITI-NEXT:    v_addc_u32_e32 v17, vcc, v17, v15, vcc5745; TAHITI-NEXT:    v_xor_b32_e32 v17, v17, v155746; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v20, v195747; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v21, vcc5748; TAHITI-NEXT:    v_mul_lo_u32 v21, v17, v135749; TAHITI-NEXT:    v_mul_hi_u32 v13, v17, v135750; TAHITI-NEXT:    v_mul_hi_u32 v22, v17, v145751; TAHITI-NEXT:    v_mul_lo_u32 v14, v17, v145752; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v19, v215753; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, v20, v13, vcc5754; TAHITI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v22, vcc5755; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v145756; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v19, vcc5757; TAHITI-NEXT:    v_mul_lo_u32 v14, v11, v145758; TAHITI-NEXT:    v_mul_hi_u32 v19, v11, v135759; TAHITI-NEXT:    v_mul_lo_u32 v20, v10, v135760; TAHITI-NEXT:    v_mul_lo_u32 v13, v11, v135761; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v14, v195762; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v20, v145763; TAHITI-NEXT:    v_sub_i32_e32 v19, vcc, v17, v145764; TAHITI-NEXT:    v_sub_i32_e32 v13, vcc, v18, v135765; TAHITI-NEXT:    v_subb_u32_e64 v18, s[0:1], v19, v10, vcc5766; TAHITI-NEXT:    v_sub_i32_e64 v19, s[0:1], v13, v115767; TAHITI-NEXT:    v_subbrev_u32_e64 v20, s[2:3], 0, v18, s[0:1]5768; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v20, v105769; TAHITI-NEXT:    v_cndmask_b32_e64 v21, 0, -1, s[2:3]5770; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v19, v115771; TAHITI-NEXT:    v_subb_u32_e32 v14, vcc, v17, v14, vcc5772; TAHITI-NEXT:    v_cndmask_b32_e64 v22, 0, -1, s[2:3]5773; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v20, v105774; TAHITI-NEXT:    v_subb_u32_e64 v18, s[0:1], v18, v10, s[0:1]5775; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v14, v105776; TAHITI-NEXT:    v_cndmask_b32_e64 v21, v21, v22, s[2:3]5777; TAHITI-NEXT:    v_sub_i32_e64 v22, s[0:1], v19, v115778; TAHITI-NEXT:    v_cndmask_b32_e64 v17, 0, -1, vcc5779; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v13, v115780; TAHITI-NEXT:    v_subbrev_u32_e64 v18, s[0:1], 0, v18, s[0:1]5781; TAHITI-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc5782; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v14, v105783; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v215784; TAHITI-NEXT:    v_cndmask_b32_e32 v10, v17, v11, vcc5785; TAHITI-NEXT:    v_cndmask_b32_e64 v19, v19, v22, s[0:1]5786; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v105787; TAHITI-NEXT:    v_cndmask_b32_e64 v18, v20, v18, s[0:1]5788; TAHITI-NEXT:    v_cndmask_b32_e32 v11, v13, v19, vcc5789; TAHITI-NEXT:    v_cndmask_b32_e32 v10, v14, v18, vcc5790; TAHITI-NEXT:    v_xor_b32_e32 v11, v11, v155791; TAHITI-NEXT:    v_xor_b32_e32 v13, v10, v155792; TAHITI-NEXT:    v_sub_i32_e32 v10, vcc, v11, v155793; TAHITI-NEXT:    v_subb_u32_e32 v11, vcc, v13, v15, vcc5794; TAHITI-NEXT:    s_cbranch_execnz .LBB12_65795; TAHITI-NEXT:  .LBB12_5:5796; TAHITI-NEXT:    v_cvt_f32_u32_e32 v10, v125797; TAHITI-NEXT:    v_sub_i32_e32 v11, vcc, 0, v125798; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v10, v105799; TAHITI-NEXT:    v_mul_f32_e32 v10, 0x4f7ffffe, v105800; TAHITI-NEXT:    v_cvt_u32_f32_e32 v10, v105801; TAHITI-NEXT:    v_mul_lo_u32 v11, v11, v105802; TAHITI-NEXT:    v_mul_hi_u32 v11, v10, v115803; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v10, v115804; TAHITI-NEXT:    v_mul_hi_u32 v10, v16, v105805; TAHITI-NEXT:    v_mul_lo_u32 v10, v10, v125806; TAHITI-NEXT:    v_sub_i32_e32 v10, vcc, v16, v105807; TAHITI-NEXT:    v_subrev_i32_e32 v11, vcc, v12, v105808; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v125809; TAHITI-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc5810; TAHITI-NEXT:    v_subrev_i32_e32 v11, vcc, v12, v105811; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v125812; TAHITI-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc5813; TAHITI-NEXT:    v_mov_b32_e32 v11, 05814; TAHITI-NEXT:  .LBB12_6:5815; TAHITI-NEXT:    s_waitcnt vmcnt(0)5816; TAHITI-NEXT:    v_or_b32_e32 v13, v5, v15817; TAHITI-NEXT:    v_mov_b32_e32 v12, 05818; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[12:13]5819; TAHITI-NEXT:    s_cbranch_vccz .LBB12_155820; TAHITI-NEXT:  ; %bb.7:5821; TAHITI-NEXT:    v_ashrrev_i32_e32 v13, 31, v15822; TAHITI-NEXT:    v_add_i32_e32 v12, vcc, v0, v135823; TAHITI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v13, vcc5824; TAHITI-NEXT:    v_xor_b32_e32 v12, v12, v135825; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v135826; TAHITI-NEXT:    v_cvt_f32_u32_e32 v13, v125827; TAHITI-NEXT:    v_cvt_f32_u32_e32 v14, v15828; TAHITI-NEXT:    v_sub_i32_e32 v15, vcc, 0, v125829; TAHITI-NEXT:    v_subb_u32_e32 v16, vcc, 0, v1, vcc5830; TAHITI-NEXT:    v_madmk_f32 v13, v14, 0x4f800000, v135831; TAHITI-NEXT:    v_rcp_f32_e32 v13, v135832; TAHITI-NEXT:    v_mul_f32_e32 v13, 0x5f7ffffc, v135833; TAHITI-NEXT:    v_mul_f32_e32 v14, 0x2f800000, v135834; TAHITI-NEXT:    v_trunc_f32_e32 v14, v145835; TAHITI-NEXT:    v_madmk_f32 v13, v14, 0xcf800000, v135836; TAHITI-NEXT:    v_cvt_u32_f32_e32 v14, v145837; TAHITI-NEXT:    v_cvt_u32_f32_e32 v13, v135838; TAHITI-NEXT:    v_mul_lo_u32 v18, v15, v145839; TAHITI-NEXT:    v_mul_hi_u32 v17, v15, v135840; TAHITI-NEXT:    v_mul_lo_u32 v19, v16, v135841; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v18, v175842; TAHITI-NEXT:    v_mul_lo_u32 v18, v15, v135843; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v17, v195844; TAHITI-NEXT:    v_mul_lo_u32 v19, v13, v175845; TAHITI-NEXT:    v_mul_hi_u32 v20, v13, v185846; TAHITI-NEXT:    v_mul_hi_u32 v21, v13, v175847; TAHITI-NEXT:    v_mul_hi_u32 v22, v14, v175848; TAHITI-NEXT:    v_mul_lo_u32 v17, v14, v175849; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v20, v195850; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v21, vcc5851; TAHITI-NEXT:    v_mul_lo_u32 v21, v14, v185852; TAHITI-NEXT:    v_mul_hi_u32 v18, v14, v185853; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v19, v215854; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, v20, v18, vcc5855; TAHITI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v22, vcc5856; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v18, v175857; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v19, vcc5858; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v175859; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, v14, v18, vcc5860; TAHITI-NEXT:    v_mul_lo_u32 v17, v15, v145861; TAHITI-NEXT:    v_mul_hi_u32 v18, v15, v135862; TAHITI-NEXT:    v_mul_lo_u32 v16, v16, v135863; TAHITI-NEXT:    v_mul_lo_u32 v15, v15, v135864; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v17, v185865; TAHITI-NEXT:    v_add_i32_e32 v16, vcc, v16, v175866; TAHITI-NEXT:    v_mul_lo_u32 v19, v13, v165867; TAHITI-NEXT:    v_mul_hi_u32 v20, v13, v155868; TAHITI-NEXT:    v_mul_hi_u32 v21, v13, v165869; TAHITI-NEXT:    v_mul_hi_u32 v18, v14, v155870; TAHITI-NEXT:    v_mul_lo_u32 v15, v14, v155871; TAHITI-NEXT:    v_mul_hi_u32 v17, v14, v165872; TAHITI-NEXT:    v_add_i32_e32 v19, vcc, v20, v195873; TAHITI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v21, vcc5874; TAHITI-NEXT:    v_mul_lo_u32 v16, v14, v165875; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v19, v155876; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, v20, v18, vcc5877; TAHITI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc5878; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v165879; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v17, vcc5880; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v155881; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, v14, v16, vcc5882; TAHITI-NEXT:    v_ashrrev_i32_e32 v15, 31, v55883; TAHITI-NEXT:    v_add_i32_e32 v16, vcc, v4, v155884; TAHITI-NEXT:    v_xor_b32_e32 v16, v16, v155885; TAHITI-NEXT:    v_mul_lo_u32 v17, v16, v145886; TAHITI-NEXT:    v_mul_hi_u32 v18, v16, v135887; TAHITI-NEXT:    v_mul_hi_u32 v19, v16, v145888; TAHITI-NEXT:    v_addc_u32_e32 v5, vcc, v5, v15, vcc5889; TAHITI-NEXT:    v_xor_b32_e32 v5, v5, v155890; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v18, v175891; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v19, vcc5892; TAHITI-NEXT:    v_mul_lo_u32 v19, v5, v135893; TAHITI-NEXT:    v_mul_hi_u32 v13, v5, v135894; TAHITI-NEXT:    v_mul_hi_u32 v20, v5, v145895; TAHITI-NEXT:    v_mul_lo_u32 v14, v5, v145896; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v17, v195897; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, v18, v13, vcc5898; TAHITI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v20, vcc5899; TAHITI-NEXT:    v_add_i32_e32 v13, vcc, v13, v145900; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v17, vcc5901; TAHITI-NEXT:    v_mul_lo_u32 v14, v12, v145902; TAHITI-NEXT:    v_mul_hi_u32 v17, v12, v135903; TAHITI-NEXT:    v_mul_lo_u32 v18, v1, v135904; TAHITI-NEXT:    v_mul_lo_u32 v13, v12, v135905; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v14, v175906; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v18, v145907; TAHITI-NEXT:    v_sub_i32_e32 v17, vcc, v5, v145908; TAHITI-NEXT:    v_sub_i32_e32 v13, vcc, v16, v135909; TAHITI-NEXT:    v_subb_u32_e64 v16, s[0:1], v17, v1, vcc5910; TAHITI-NEXT:    v_sub_i32_e64 v17, s[0:1], v13, v125911; TAHITI-NEXT:    v_subbrev_u32_e64 v18, s[2:3], 0, v16, s[0:1]5912; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v18, v15913; TAHITI-NEXT:    v_cndmask_b32_e64 v19, 0, -1, s[2:3]5914; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v17, v125915; TAHITI-NEXT:    v_subb_u32_e32 v5, vcc, v5, v14, vcc5916; TAHITI-NEXT:    v_cndmask_b32_e64 v20, 0, -1, s[2:3]5917; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v18, v15918; TAHITI-NEXT:    v_subb_u32_e64 v16, s[0:1], v16, v1, s[0:1]5919; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v15920; TAHITI-NEXT:    v_cndmask_b32_e64 v19, v19, v20, s[2:3]5921; TAHITI-NEXT:    v_sub_i32_e64 v20, s[0:1], v17, v125922; TAHITI-NEXT:    v_cndmask_b32_e64 v14, 0, -1, vcc5923; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v13, v125924; TAHITI-NEXT:    v_subbrev_u32_e64 v16, s[0:1], 0, v16, s[0:1]5925; TAHITI-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc5926; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v15927; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v195928; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v14, v12, vcc5929; TAHITI-NEXT:    v_cndmask_b32_e64 v17, v17, v20, s[0:1]5930; TAHITI-NEXT:    v_cndmask_b32_e64 v16, v18, v16, s[0:1]5931; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v15932; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v5, v16, vcc5933; TAHITI-NEXT:    v_cndmask_b32_e32 v5, v13, v17, vcc5934; TAHITI-NEXT:    v_xor_b32_e32 v5, v5, v155935; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v155936; TAHITI-NEXT:    v_sub_i32_e32 v12, vcc, v5, v155937; TAHITI-NEXT:    v_subb_u32_e32 v13, vcc, v1, v15, vcc5938; TAHITI-NEXT:    s_cbranch_execnz .LBB12_95939; TAHITI-NEXT:  .LBB12_8:5940; TAHITI-NEXT:    v_cvt_f32_u32_e32 v1, v05941; TAHITI-NEXT:    v_sub_i32_e32 v5, vcc, 0, v05942; TAHITI-NEXT:    v_mov_b32_e32 v13, 05943; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v1, v15944; TAHITI-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v15945; TAHITI-NEXT:    v_cvt_u32_f32_e32 v1, v15946; TAHITI-NEXT:    v_mul_lo_u32 v5, v5, v15947; TAHITI-NEXT:    v_mul_hi_u32 v5, v1, v55948; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v1, v55949; TAHITI-NEXT:    v_mul_hi_u32 v1, v4, v15950; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, v05951; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, v4, v15952; TAHITI-NEXT:    v_subrev_i32_e32 v4, vcc, v0, v15953; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v05954; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc5955; TAHITI-NEXT:    v_subrev_i32_e32 v4, vcc, v0, v15956; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v05957; TAHITI-NEXT:    v_cndmask_b32_e32 v12, v1, v4, vcc5958; TAHITI-NEXT:  .LBB12_9:5959; TAHITI-NEXT:    v_or_b32_e32 v1, v7, v35960; TAHITI-NEXT:    v_mov_b32_e32 v0, 05961; TAHITI-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[0:1]5962; TAHITI-NEXT:    s_cbranch_vccz .LBB12_165963; TAHITI-NEXT:  ; %bb.10:5964; TAHITI-NEXT:    v_ashrrev_i32_e32 v0, 31, v35965; TAHITI-NEXT:    v_add_i32_e32 v1, vcc, v2, v05966; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, v3, v0, vcc5967; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v05968; TAHITI-NEXT:    v_xor_b32_e32 v0, v3, v05969; TAHITI-NEXT:    v_cvt_f32_u32_e32 v3, v15970; TAHITI-NEXT:    v_cvt_f32_u32_e32 v4, v05971; TAHITI-NEXT:    v_sub_i32_e32 v5, vcc, 0, v15972; TAHITI-NEXT:    v_subb_u32_e32 v14, vcc, 0, v0, vcc5973; TAHITI-NEXT:    v_madmk_f32 v3, v4, 0x4f800000, v35974; TAHITI-NEXT:    v_rcp_f32_e32 v3, v35975; TAHITI-NEXT:    v_mul_f32_e32 v3, 0x5f7ffffc, v35976; TAHITI-NEXT:    v_mul_f32_e32 v4, 0x2f800000, v35977; TAHITI-NEXT:    v_trunc_f32_e32 v4, v45978; TAHITI-NEXT:    v_madmk_f32 v3, v4, 0xcf800000, v35979; TAHITI-NEXT:    v_cvt_u32_f32_e32 v4, v45980; TAHITI-NEXT:    v_cvt_u32_f32_e32 v3, v35981; TAHITI-NEXT:    v_mul_lo_u32 v16, v5, v45982; TAHITI-NEXT:    v_mul_hi_u32 v15, v5, v35983; TAHITI-NEXT:    v_mul_lo_u32 v17, v14, v35984; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v16, v155985; TAHITI-NEXT:    v_mul_lo_u32 v16, v5, v35986; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v175987; TAHITI-NEXT:    v_mul_lo_u32 v17, v3, v155988; TAHITI-NEXT:    v_mul_hi_u32 v18, v3, v165989; TAHITI-NEXT:    v_mul_hi_u32 v19, v3, v155990; TAHITI-NEXT:    v_mul_hi_u32 v20, v4, v155991; TAHITI-NEXT:    v_mul_lo_u32 v15, v4, v155992; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v18, v175993; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v19, vcc5994; TAHITI-NEXT:    v_mul_lo_u32 v19, v4, v165995; TAHITI-NEXT:    v_mul_hi_u32 v16, v4, v165996; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v17, v195997; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, v18, v16, vcc5998; TAHITI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v20, vcc5999; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v16, v156000; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v17, vcc6001; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v156002; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, v4, v16, vcc6003; TAHITI-NEXT:    v_mul_lo_u32 v15, v5, v46004; TAHITI-NEXT:    v_mul_hi_u32 v16, v5, v36005; TAHITI-NEXT:    v_mul_lo_u32 v14, v14, v36006; TAHITI-NEXT:    v_mul_lo_u32 v5, v5, v36007; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v166008; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v14, v156009; TAHITI-NEXT:    v_mul_lo_u32 v17, v3, v146010; TAHITI-NEXT:    v_mul_hi_u32 v18, v3, v56011; TAHITI-NEXT:    v_mul_hi_u32 v19, v3, v146012; TAHITI-NEXT:    v_mul_hi_u32 v16, v4, v56013; TAHITI-NEXT:    v_mul_lo_u32 v5, v4, v56014; TAHITI-NEXT:    v_mul_hi_u32 v15, v4, v146015; TAHITI-NEXT:    v_add_i32_e32 v17, vcc, v18, v176016; TAHITI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v19, vcc6017; TAHITI-NEXT:    v_mul_lo_u32 v14, v4, v146018; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v17, v56019; TAHITI-NEXT:    v_addc_u32_e32 v5, vcc, v18, v16, vcc6020; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc6021; TAHITI-NEXT:    v_add_i32_e32 v5, vcc, v5, v146022; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc6023; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v56024; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, v4, v14, vcc6025; TAHITI-NEXT:    v_ashrrev_i32_e32 v5, 31, v76026; TAHITI-NEXT:    v_add_i32_e32 v14, vcc, v6, v56027; TAHITI-NEXT:    v_xor_b32_e32 v14, v14, v56028; TAHITI-NEXT:    v_mul_lo_u32 v15, v14, v46029; TAHITI-NEXT:    v_mul_hi_u32 v16, v14, v36030; TAHITI-NEXT:    v_mul_hi_u32 v17, v14, v46031; TAHITI-NEXT:    v_addc_u32_e32 v7, vcc, v7, v5, vcc6032; TAHITI-NEXT:    v_xor_b32_e32 v7, v7, v56033; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v16, v156034; TAHITI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v17, vcc6035; TAHITI-NEXT:    v_mul_lo_u32 v17, v7, v36036; TAHITI-NEXT:    v_mul_hi_u32 v3, v7, v36037; TAHITI-NEXT:    v_mul_hi_u32 v18, v7, v46038; TAHITI-NEXT:    v_mul_lo_u32 v4, v7, v46039; TAHITI-NEXT:    v_add_i32_e32 v15, vcc, v15, v176040; TAHITI-NEXT:    v_addc_u32_e32 v3, vcc, v16, v3, vcc6041; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v18, vcc6042; TAHITI-NEXT:    v_add_i32_e32 v3, vcc, v3, v46043; TAHITI-NEXT:    v_addc_u32_e32 v4, vcc, 0, v15, vcc6044; TAHITI-NEXT:    v_mul_lo_u32 v4, v1, v46045; TAHITI-NEXT:    v_mul_hi_u32 v15, v1, v36046; TAHITI-NEXT:    v_mul_lo_u32 v16, v0, v36047; TAHITI-NEXT:    v_mul_lo_u32 v3, v1, v36048; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v4, v156049; TAHITI-NEXT:    v_add_i32_e32 v4, vcc, v16, v46050; TAHITI-NEXT:    v_sub_i32_e32 v15, vcc, v7, v46051; TAHITI-NEXT:    v_sub_i32_e32 v3, vcc, v14, v36052; TAHITI-NEXT:    v_subb_u32_e64 v14, s[0:1], v15, v0, vcc6053; TAHITI-NEXT:    v_sub_i32_e64 v15, s[0:1], v3, v16054; TAHITI-NEXT:    v_subbrev_u32_e64 v16, s[2:3], 0, v14, s[0:1]6055; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v16, v06056; TAHITI-NEXT:    v_cndmask_b32_e64 v17, 0, -1, s[2:3]6057; TAHITI-NEXT:    v_cmp_ge_u32_e64 s[2:3], v15, v16058; TAHITI-NEXT:    v_subb_u32_e32 v4, vcc, v7, v4, vcc6059; TAHITI-NEXT:    v_cndmask_b32_e64 v18, 0, -1, s[2:3]6060; TAHITI-NEXT:    v_cmp_eq_u32_e64 s[2:3], v16, v06061; TAHITI-NEXT:    v_subb_u32_e64 v14, s[0:1], v14, v0, s[0:1]6062; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v4, v06063; TAHITI-NEXT:    v_cndmask_b32_e64 v17, v17, v18, s[2:3]6064; TAHITI-NEXT:    v_sub_i32_e64 v18, s[0:1], v15, v16065; TAHITI-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc6066; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v16067; TAHITI-NEXT:    v_subbrev_u32_e64 v14, s[0:1], 0, v14, s[0:1]6068; TAHITI-NEXT:    v_cndmask_b32_e64 v1, 0, -1, vcc6069; TAHITI-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v06070; TAHITI-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v176071; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v7, v1, vcc6072; TAHITI-NEXT:    v_cndmask_b32_e64 v15, v15, v18, s[0:1]6073; TAHITI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v06074; TAHITI-NEXT:    v_cndmask_b32_e64 v14, v16, v14, s[0:1]6075; TAHITI-NEXT:    v_cndmask_b32_e32 v1, v3, v15, vcc6076; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v4, v14, vcc6077; TAHITI-NEXT:    v_xor_b32_e32 v1, v1, v56078; TAHITI-NEXT:    v_xor_b32_e32 v0, v0, v56079; TAHITI-NEXT:    v_sub_i32_e32 v14, vcc, v1, v56080; TAHITI-NEXT:    v_subb_u32_e32 v15, vcc, v0, v5, vcc6081; TAHITI-NEXT:    s_cbranch_execnz .LBB12_126082; TAHITI-NEXT:  .LBB12_11:6083; TAHITI-NEXT:    v_cvt_f32_u32_e32 v0, v26084; TAHITI-NEXT:    v_sub_i32_e32 v1, vcc, 0, v26085; TAHITI-NEXT:    v_mov_b32_e32 v15, 06086; TAHITI-NEXT:    v_rcp_iflag_f32_e32 v0, v06087; TAHITI-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v06088; TAHITI-NEXT:    v_cvt_u32_f32_e32 v0, v06089; TAHITI-NEXT:    v_mul_lo_u32 v1, v1, v06090; TAHITI-NEXT:    v_mul_hi_u32 v1, v0, v16091; TAHITI-NEXT:    v_add_i32_e32 v0, vcc, v0, v16092; TAHITI-NEXT:    v_mul_hi_u32 v0, v6, v06093; TAHITI-NEXT:    v_mul_lo_u32 v0, v0, v26094; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v6, v06095; TAHITI-NEXT:    v_subrev_i32_e32 v1, vcc, v2, v06096; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v26097; TAHITI-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc6098; TAHITI-NEXT:    v_subrev_i32_e32 v1, vcc, v2, v06099; TAHITI-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v26100; TAHITI-NEXT:    v_cndmask_b32_e32 v14, v0, v1, vcc6101; TAHITI-NEXT:  .LBB12_12:6102; TAHITI-NEXT:    s_mov_b32 s7, 0xf0006103; TAHITI-NEXT:    s_mov_b32 s6, -16104; TAHITI-NEXT:    buffer_store_dwordx4 v[12:15], off, s[4:7], 0 offset:166105; TAHITI-NEXT:    buffer_store_dwordx4 v[8:11], off, s[4:7], 06106; TAHITI-NEXT:    s_endpgm6107; TAHITI-NEXT:  .LBB12_13:6108; TAHITI-NEXT:    ; implicit-def: $vgpr8_vgpr96109; TAHITI-NEXT:    s_branch .LBB12_26110; TAHITI-NEXT:  .LBB12_14:6111; TAHITI-NEXT:    s_branch .LBB12_56112; TAHITI-NEXT:  .LBB12_15:6113; TAHITI-NEXT:    ; implicit-def: $vgpr12_vgpr136114; TAHITI-NEXT:    s_branch .LBB12_86115; TAHITI-NEXT:  .LBB12_16:6116; TAHITI-NEXT:    s_branch .LBB12_116117;6118; TONGA-LABEL: srem_v4i64:6119; TONGA:       ; %bb.0:6120; TONGA-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x246121; TONGA-NEXT:    s_waitcnt lgkmcnt(0)6122; TONGA-NEXT:    s_add_u32 s0, s6, 486123; TONGA-NEXT:    v_mov_b32_e32 v0, s66124; TONGA-NEXT:    s_addc_u32 s1, s7, 06125; TONGA-NEXT:    v_mov_b32_e32 v1, s76126; TONGA-NEXT:    s_add_u32 s2, s6, 326127; TONGA-NEXT:    flat_load_dwordx4 v[14:17], v[0:1]6128; TONGA-NEXT:    s_addc_u32 s3, s7, 06129; TONGA-NEXT:    v_mov_b32_e32 v0, s26130; TONGA-NEXT:    v_mov_b32_e32 v1, s36131; TONGA-NEXT:    flat_load_dwordx4 v[10:13], v[0:1]6132; TONGA-NEXT:    v_mov_b32_e32 v0, s06133; TONGA-NEXT:    v_mov_b32_e32 v1, s16134; TONGA-NEXT:    s_add_u32 s0, s6, 166135; TONGA-NEXT:    s_addc_u32 s1, s7, 06136; TONGA-NEXT:    v_mov_b32_e32 v5, s16137; TONGA-NEXT:    v_mov_b32_e32 v4, s06138; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]6139; TONGA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]6140; TONGA-NEXT:    s_waitcnt vmcnt(3)6141; TONGA-NEXT:    v_readfirstlane_b32 s3, v156142; TONGA-NEXT:    v_readfirstlane_b32 s2, v146143; TONGA-NEXT:    s_waitcnt vmcnt(2)6144; TONGA-NEXT:    v_readfirstlane_b32 s1, v116145; TONGA-NEXT:    v_readfirstlane_b32 s0, v106146; TONGA-NEXT:    s_or_b64 s[6:7], s[2:3], s[0:1]6147; TONGA-NEXT:    s_mov_b32 s6, 06148; TONGA-NEXT:    s_cmp_lg_u64 s[6:7], 06149; TONGA-NEXT:    s_cbranch_scc0 .LBB12_36150; TONGA-NEXT:  ; %bb.1:6151; TONGA-NEXT:    s_ashr_i32 s6, s1, 316152; TONGA-NEXT:    s_add_u32 s8, s0, s66153; TONGA-NEXT:    s_mov_b32 s7, s66154; TONGA-NEXT:    s_addc_u32 s9, s1, s66155; TONGA-NEXT:    s_xor_b64 s[6:7], s[8:9], s[6:7]6156; TONGA-NEXT:    v_cvt_f32_u32_e32 v8, s66157; TONGA-NEXT:    v_cvt_f32_u32_e32 v9, s76158; TONGA-NEXT:    s_sub_u32 s1, 0, s66159; TONGA-NEXT:    s_subb_u32 s10, 0, s76160; TONGA-NEXT:    v_madmk_f32 v8, v9, 0x4f800000, v86161; TONGA-NEXT:    v_rcp_f32_e32 v8, v86162; TONGA-NEXT:    v_mul_f32_e32 v8, 0x5f7ffffc, v86163; TONGA-NEXT:    v_mul_f32_e32 v9, 0x2f800000, v86164; TONGA-NEXT:    v_trunc_f32_e32 v9, v96165; TONGA-NEXT:    v_madmk_f32 v8, v9, 0xcf800000, v86166; TONGA-NEXT:    v_cvt_u32_f32_e32 v14, v96167; TONGA-NEXT:    v_cvt_u32_f32_e32 v15, v86168; TONGA-NEXT:    v_mul_lo_u32 v10, s1, v146169; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], s1, v15, 06170; TONGA-NEXT:    v_mul_lo_u32 v11, s10, v156171; TONGA-NEXT:    v_add_u32_e32 v9, vcc, v9, v106172; TONGA-NEXT:    v_add_u32_e32 v11, vcc, v9, v116173; TONGA-NEXT:    v_mul_hi_u32 v18, v15, v86174; TONGA-NEXT:    v_mad_u64_u32 v[9:10], s[8:9], v15, v11, 06175; TONGA-NEXT:    v_add_u32_e32 v18, vcc, v18, v96176; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], v14, v8, 06177; TONGA-NEXT:    v_addc_u32_e32 v19, vcc, 0, v10, vcc6178; TONGA-NEXT:    v_mad_u64_u32 v[10:11], s[8:9], v14, v11, 06179; TONGA-NEXT:    v_add_u32_e32 v8, vcc, v18, v86180; TONGA-NEXT:    v_addc_u32_e32 v8, vcc, v19, v9, vcc6181; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v11, vcc6182; TONGA-NEXT:    v_add_u32_e32 v8, vcc, v8, v106183; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc6184; TONGA-NEXT:    v_add_u32_e32 v18, vcc, v15, v86185; TONGA-NEXT:    v_addc_u32_e32 v19, vcc, v14, v9, vcc6186; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], s1, v18, 06187; TONGA-NEXT:    v_mul_lo_u32 v14, s1, v196188; TONGA-NEXT:    v_mul_lo_u32 v15, s10, v186189; TONGA-NEXT:    v_mul_hi_u32 v20, v18, v86190; TONGA-NEXT:    v_mad_u64_u32 v[10:11], s[8:9], v19, v8, 06191; TONGA-NEXT:    v_add_u32_e32 v9, vcc, v14, v96192; TONGA-NEXT:    v_add_u32_e32 v9, vcc, v15, v96193; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[8:9], v18, v9, 06194; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], v19, v9, 06195; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v20, v146196; TONGA-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc6197; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v14, v106198; TONGA-NEXT:    v_addc_u32_e32 v10, vcc, v15, v11, vcc6199; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc6200; TONGA-NEXT:    v_add_u32_e32 v8, vcc, v10, v86201; TONGA-NEXT:    s_ashr_i32 s10, s3, 316202; TONGA-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc6203; TONGA-NEXT:    s_add_u32 s8, s2, s106204; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v18, v86205; TONGA-NEXT:    s_mov_b32 s11, s106206; TONGA-NEXT:    s_addc_u32 s9, s3, s106207; TONGA-NEXT:    v_addc_u32_e32 v11, vcc, v19, v9, vcc6208; TONGA-NEXT:    s_xor_b64 s[12:13], s[8:9], s[10:11]6209; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], s12, v11, 06210; TONGA-NEXT:    v_mul_hi_u32 v14, s12, v106211; TONGA-NEXT:    v_readfirstlane_b32 s1, v96212; TONGA-NEXT:    v_readfirstlane_b32 s3, v86213; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], s13, v11, 06214; TONGA-NEXT:    v_mad_u64_u32 v[10:11], s[8:9], s13, v10, 06215; TONGA-NEXT:    v_readfirstlane_b32 s14, v146216; TONGA-NEXT:    s_add_u32 s3, s14, s36217; TONGA-NEXT:    s_addc_u32 s1, 0, s16218; TONGA-NEXT:    v_readfirstlane_b32 s14, v106219; TONGA-NEXT:    v_readfirstlane_b32 s9, v116220; TONGA-NEXT:    s_add_u32 s3, s3, s146221; TONGA-NEXT:    v_readfirstlane_b32 s8, v96222; TONGA-NEXT:    s_addc_u32 s1, s1, s96223; TONGA-NEXT:    s_addc_u32 s3, s8, 06224; TONGA-NEXT:    v_readfirstlane_b32 s8, v86225; TONGA-NEXT:    s_add_u32 s1, s1, s86226; TONGA-NEXT:    v_mov_b32_e32 v8, s16227; TONGA-NEXT:    v_mad_u64_u32 v[8:9], s[8:9], s6, v8, 06228; TONGA-NEXT:    s_addc_u32 s3, 0, s36229; TONGA-NEXT:    s_mul_i32 s3, s6, s36230; TONGA-NEXT:    v_readfirstlane_b32 s14, v96231; TONGA-NEXT:    s_add_i32 s3, s14, s36232; TONGA-NEXT:    s_mul_i32 s1, s7, s16233; TONGA-NEXT:    s_add_i32 s3, s3, s16234; TONGA-NEXT:    s_sub_i32 s1, s13, s36235; TONGA-NEXT:    v_readfirstlane_b32 s14, v86236; TONGA-NEXT:    s_sub_u32 s12, s12, s146237; TONGA-NEXT:    s_cselect_b64 s[14:15], -1, 06238; TONGA-NEXT:    s_subb_u32 s1, s1, s76239; TONGA-NEXT:    s_sub_u32 s18, s12, s66240; TONGA-NEXT:    s_cselect_b64 s[16:17], -1, 06241; TONGA-NEXT:    s_subb_u32 s19, s1, 06242; TONGA-NEXT:    s_cmp_ge_u32 s19, s76243; TONGA-NEXT:    s_cselect_b32 s20, -1, 06244; TONGA-NEXT:    s_cmp_ge_u32 s18, s66245; TONGA-NEXT:    s_cselect_b32 s21, -1, 06246; TONGA-NEXT:    s_cmp_eq_u32 s19, s76247; TONGA-NEXT:    s_cselect_b32 s20, s21, s206248; TONGA-NEXT:    s_cmp_lg_u64 s[16:17], 06249; TONGA-NEXT:    s_subb_u32 s1, s1, s76250; TONGA-NEXT:    s_sub_u32 s16, s18, s66251; TONGA-NEXT:    s_subb_u32 s1, s1, 06252; TONGA-NEXT:    s_cmp_lg_u32 s20, 06253; TONGA-NEXT:    s_cselect_b32 s16, s16, s186254; TONGA-NEXT:    s_cselect_b32 s1, s1, s196255; TONGA-NEXT:    s_cmp_lg_u64 s[14:15], 06256; TONGA-NEXT:    s_subb_u32 s3, s13, s36257; TONGA-NEXT:    s_cmp_ge_u32 s3, s76258; TONGA-NEXT:    s_cselect_b32 s13, -1, 06259; TONGA-NEXT:    s_cmp_ge_u32 s12, s66260; TONGA-NEXT:    s_cselect_b32 s6, -1, 06261; TONGA-NEXT:    s_cmp_eq_u32 s3, s76262; TONGA-NEXT:    s_cselect_b32 s6, s6, s136263; TONGA-NEXT:    s_cmp_lg_u32 s6, 06264; TONGA-NEXT:    s_cselect_b32 s7, s1, s36265; TONGA-NEXT:    s_cselect_b32 s6, s16, s126266; TONGA-NEXT:    s_xor_b64 s[6:7], s[6:7], s[10:11]6267; TONGA-NEXT:    s_sub_u32 s6, s6, s106268; TONGA-NEXT:    s_subb_u32 s7, s7, s106269; TONGA-NEXT:    s_cbranch_execnz .LBB12_46270; TONGA-NEXT:  .LBB12_2:6271; TONGA-NEXT:    v_cvt_f32_u32_e32 v8, s06272; TONGA-NEXT:    s_sub_i32 s1, 0, s06273; TONGA-NEXT:    v_rcp_iflag_f32_e32 v8, v86274; TONGA-NEXT:    v_mul_f32_e32 v8, 0x4f7ffffe, v86275; TONGA-NEXT:    v_cvt_u32_f32_e32 v8, v86276; TONGA-NEXT:    v_mul_lo_u32 v9, s1, v86277; TONGA-NEXT:    v_mul_hi_u32 v9, v8, v96278; TONGA-NEXT:    v_add_u32_e32 v8, vcc, v8, v96279; TONGA-NEXT:    v_mul_hi_u32 v8, s2, v86280; TONGA-NEXT:    v_mul_lo_u32 v8, v8, s06281; TONGA-NEXT:    v_sub_u32_e32 v8, vcc, s2, v86282; TONGA-NEXT:    v_subrev_u32_e32 v9, vcc, s0, v86283; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s0, v86284; TONGA-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc6285; TONGA-NEXT:    v_subrev_u32_e32 v9, vcc, s0, v86286; TONGA-NEXT:    v_cmp_le_u32_e32 vcc, s0, v86287; TONGA-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc6288; TONGA-NEXT:    v_mov_b32_e32 v9, 06289; TONGA-NEXT:    s_branch .LBB12_56290; TONGA-NEXT:  .LBB12_3:6291; TONGA-NEXT:    ; implicit-def: $sgpr6_sgpr76292; TONGA-NEXT:    s_branch .LBB12_26293; TONGA-NEXT:  .LBB12_4:6294; TONGA-NEXT:    v_mov_b32_e32 v9, s76295; TONGA-NEXT:    v_mov_b32_e32 v8, s66296; TONGA-NEXT:  .LBB12_5:6297; TONGA-NEXT:    v_or_b32_e32 v11, v17, v136298; TONGA-NEXT:    v_mov_b32_e32 v10, 06299; TONGA-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[10:11]6300; TONGA-NEXT:    s_cbranch_vccz .LBB12_156301; TONGA-NEXT:  ; %bb.6:6302; TONGA-NEXT:    v_ashrrev_i32_e32 v10, 31, v136303; TONGA-NEXT:    v_add_u32_e32 v11, vcc, v12, v106304; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, v13, v10, vcc6305; TONGA-NEXT:    v_xor_b32_e32 v11, v11, v106306; TONGA-NEXT:    v_xor_b32_e32 v10, v13, v106307; TONGA-NEXT:    v_cvt_f32_u32_e32 v13, v116308; TONGA-NEXT:    v_cvt_f32_u32_e32 v14, v106309; TONGA-NEXT:    v_sub_u32_e32 v22, vcc, 0, v116310; TONGA-NEXT:    v_subb_u32_e32 v23, vcc, 0, v10, vcc6311; TONGA-NEXT:    v_madmk_f32 v13, v14, 0x4f800000, v136312; TONGA-NEXT:    v_rcp_f32_e32 v13, v136313; TONGA-NEXT:    v_mul_f32_e32 v13, 0x5f7ffffc, v136314; TONGA-NEXT:    v_mul_f32_e32 v14, 0x2f800000, v136315; TONGA-NEXT:    v_trunc_f32_e32 v14, v146316; TONGA-NEXT:    v_madmk_f32 v13, v14, 0xcf800000, v136317; TONGA-NEXT:    v_cvt_u32_f32_e32 v20, v146318; TONGA-NEXT:    v_cvt_u32_f32_e32 v21, v136319; TONGA-NEXT:    v_mul_lo_u32 v15, v22, v206320; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v22, v21, 06321; TONGA-NEXT:    v_mul_lo_u32 v18, v23, v216322; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v14, v156323; TONGA-NEXT:    v_add_u32_e32 v18, vcc, v14, v186324; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v21, v18, 06325; TONGA-NEXT:    v_mul_hi_u32 v19, v21, v136326; TONGA-NEXT:    v_add_u32_e32 v24, vcc, v19, v146327; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v20, v13, 06328; TONGA-NEXT:    v_mad_u64_u32 v[18:19], s[0:1], v20, v18, 06329; TONGA-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc6330; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v24, v136331; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, v15, v14, vcc6332; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v19, vcc6333; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v186334; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v14, vcc6335; TONGA-NEXT:    v_add_u32_e32 v24, vcc, v21, v136336; TONGA-NEXT:    v_addc_u32_e32 v25, vcc, v20, v14, vcc6337; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v22, v24, 06338; TONGA-NEXT:    v_mul_lo_u32 v15, v22, v256339; TONGA-NEXT:    v_mul_lo_u32 v20, v23, v246340; TONGA-NEXT:    v_mad_u64_u32 v[18:19], s[0:1], v25, v13, 06341; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v15, v146342; TONGA-NEXT:    v_add_u32_e32 v20, vcc, v20, v146343; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v24, v20, 06344; TONGA-NEXT:    v_mul_hi_u32 v13, v24, v136345; TONGA-NEXT:    v_mad_u64_u32 v[20:21], s[0:1], v25, v20, 06346; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v146347; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v15, vcc6348; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v186349; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, v14, v19, vcc6350; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v21, vcc6351; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v206352; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v14, vcc6353; TONGA-NEXT:    v_add_u32_e32 v15, vcc, v24, v136354; TONGA-NEXT:    v_addc_u32_e32 v18, vcc, v25, v14, vcc6355; TONGA-NEXT:    v_ashrrev_i32_e32 v19, 31, v176356; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v16, v196357; TONGA-NEXT:    v_xor_b32_e32 v20, v13, v196358; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v20, v18, 06359; TONGA-NEXT:    v_mul_hi_u32 v21, v20, v156360; TONGA-NEXT:    v_addc_u32_e32 v17, vcc, v17, v19, vcc6361; TONGA-NEXT:    v_xor_b32_e32 v22, v17, v196362; TONGA-NEXT:    v_add_u32_e32 v21, vcc, v21, v136363; TONGA-NEXT:    v_addc_u32_e32 v23, vcc, 0, v14, vcc6364; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v22, v15, 06365; TONGA-NEXT:    v_mad_u64_u32 v[17:18], s[0:1], v22, v18, 06366; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v21, v136367; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, v23, v14, vcc6368; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, 0, v18, vcc6369; TONGA-NEXT:    v_add_u32_e32 v15, vcc, v13, v176370; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v14, vcc6371; TONGA-NEXT:    v_mul_lo_u32 v17, v11, v136372; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v11, v15, 06373; TONGA-NEXT:    v_mul_lo_u32 v15, v10, v156374; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v17, v146375; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v15, v146376; TONGA-NEXT:    v_sub_u32_e32 v15, vcc, v22, v146377; TONGA-NEXT:    v_sub_u32_e32 v13, vcc, v20, v136378; TONGA-NEXT:    v_subb_u32_e64 v15, s[0:1], v15, v10, vcc6379; TONGA-NEXT:    v_sub_u32_e64 v17, s[0:1], v13, v116380; TONGA-NEXT:    v_subbrev_u32_e64 v18, s[2:3], 0, v15, s[0:1]6381; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v18, v106382; TONGA-NEXT:    v_cndmask_b32_e64 v20, 0, -1, s[2:3]6383; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v17, v116384; TONGA-NEXT:    v_cndmask_b32_e64 v21, 0, -1, s[2:3]6385; TONGA-NEXT:    v_cmp_eq_u32_e64 s[2:3], v18, v106386; TONGA-NEXT:    v_subb_u32_e64 v15, s[0:1], v15, v10, s[0:1]6387; TONGA-NEXT:    v_cndmask_b32_e64 v20, v20, v21, s[2:3]6388; TONGA-NEXT:    v_sub_u32_e64 v21, s[0:1], v17, v116389; TONGA-NEXT:    v_subbrev_u32_e64 v15, s[0:1], 0, v15, s[0:1]6390; TONGA-NEXT:    v_subb_u32_e32 v14, vcc, v22, v14, vcc6391; TONGA-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v206392; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v14, v106393; TONGA-NEXT:    v_cndmask_b32_e64 v15, v18, v15, s[0:1]6394; TONGA-NEXT:    v_cndmask_b32_e64 v18, 0, -1, vcc6395; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v13, v116396; TONGA-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc6397; TONGA-NEXT:    v_cmp_eq_u32_e32 vcc, v14, v106398; TONGA-NEXT:    v_cndmask_b32_e32 v10, v18, v11, vcc6399; TONGA-NEXT:    v_cndmask_b32_e64 v17, v17, v21, s[0:1]6400; TONGA-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v106401; TONGA-NEXT:    v_cndmask_b32_e32 v11, v13, v17, vcc6402; TONGA-NEXT:    v_cndmask_b32_e32 v10, v14, v15, vcc6403; TONGA-NEXT:    v_xor_b32_e32 v11, v11, v196404; TONGA-NEXT:    v_xor_b32_e32 v13, v10, v196405; TONGA-NEXT:    v_sub_u32_e32 v10, vcc, v11, v196406; TONGA-NEXT:    v_subb_u32_e32 v11, vcc, v13, v19, vcc6407; TONGA-NEXT:    s_cbranch_execnz .LBB12_86408; TONGA-NEXT:  .LBB12_7:6409; TONGA-NEXT:    v_cvt_f32_u32_e32 v10, v126410; TONGA-NEXT:    v_sub_u32_e32 v11, vcc, 0, v126411; TONGA-NEXT:    v_rcp_iflag_f32_e32 v10, v106412; TONGA-NEXT:    v_mul_f32_e32 v10, 0x4f7ffffe, v106413; TONGA-NEXT:    v_cvt_u32_f32_e32 v10, v106414; TONGA-NEXT:    v_mul_lo_u32 v11, v11, v106415; TONGA-NEXT:    v_mul_hi_u32 v11, v10, v116416; TONGA-NEXT:    v_add_u32_e32 v10, vcc, v10, v116417; TONGA-NEXT:    v_mul_hi_u32 v10, v16, v106418; TONGA-NEXT:    v_mul_lo_u32 v10, v10, v126419; TONGA-NEXT:    v_sub_u32_e32 v10, vcc, v16, v106420; TONGA-NEXT:    v_subrev_u32_e32 v11, vcc, v12, v106421; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v126422; TONGA-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc6423; TONGA-NEXT:    v_subrev_u32_e32 v11, vcc, v12, v106424; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v126425; TONGA-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc6426; TONGA-NEXT:    v_mov_b32_e32 v11, 06427; TONGA-NEXT:  .LBB12_8:6428; TONGA-NEXT:    s_waitcnt vmcnt(0)6429; TONGA-NEXT:    v_or_b32_e32 v13, v5, v16430; TONGA-NEXT:    v_mov_b32_e32 v12, 06431; TONGA-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[12:13]6432; TONGA-NEXT:    s_cbranch_vccz .LBB12_166433; TONGA-NEXT:  ; %bb.9:6434; TONGA-NEXT:    v_ashrrev_i32_e32 v12, 31, v16435; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v0, v126436; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, v1, v12, vcc6437; TONGA-NEXT:    v_xor_b32_e32 v18, v13, v126438; TONGA-NEXT:    v_xor_b32_e32 v1, v1, v126439; TONGA-NEXT:    v_cvt_f32_u32_e32 v12, v186440; TONGA-NEXT:    v_cvt_f32_u32_e32 v13, v16441; TONGA-NEXT:    v_sub_u32_e32 v19, vcc, 0, v186442; TONGA-NEXT:    v_subb_u32_e32 v20, vcc, 0, v1, vcc6443; TONGA-NEXT:    v_madmk_f32 v12, v13, 0x4f800000, v126444; TONGA-NEXT:    v_rcp_f32_e32 v12, v126445; TONGA-NEXT:    v_mul_f32_e32 v12, 0x5f7ffffc, v126446; TONGA-NEXT:    v_mul_f32_e32 v13, 0x2f800000, v126447; TONGA-NEXT:    v_trunc_f32_e32 v13, v136448; TONGA-NEXT:    v_madmk_f32 v12, v13, 0xcf800000, v126449; TONGA-NEXT:    v_cvt_u32_f32_e32 v16, v136450; TONGA-NEXT:    v_cvt_u32_f32_e32 v17, v126451; TONGA-NEXT:    v_mul_lo_u32 v14, v19, v166452; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v19, v17, 06453; TONGA-NEXT:    v_mul_lo_u32 v15, v20, v176454; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v13, v146455; TONGA-NEXT:    v_add_u32_e32 v15, vcc, v13, v156456; TONGA-NEXT:    v_mad_u64_u32 v[13:14], s[0:1], v17, v15, 06457; TONGA-NEXT:    v_mul_hi_u32 v21, v17, v126458; TONGA-NEXT:    v_add_u32_e32 v21, vcc, v21, v136459; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v16, v12, 06460; TONGA-NEXT:    v_addc_u32_e32 v22, vcc, 0, v14, vcc6461; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v16, v15, 06462; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v21, v126463; TONGA-NEXT:    v_addc_u32_e32 v12, vcc, v22, v13, vcc6464; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v15, vcc6465; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v12, v146466; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc6467; TONGA-NEXT:    v_add_u32_e32 v21, vcc, v17, v126468; TONGA-NEXT:    v_addc_u32_e32 v22, vcc, v16, v13, vcc6469; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v19, v21, 06470; TONGA-NEXT:    v_mul_lo_u32 v16, v19, v226471; TONGA-NEXT:    v_mul_lo_u32 v17, v20, v216472; TONGA-NEXT:    v_mul_hi_u32 v19, v21, v126473; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v22, v12, 06474; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v16, v136475; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v17, v136476; TONGA-NEXT:    v_mad_u64_u32 v[16:17], s[0:1], v21, v13, 06477; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v22, v13, 06478; TONGA-NEXT:    v_add_u32_e32 v16, vcc, v19, v166479; TONGA-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc6480; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v16, v146481; TONGA-NEXT:    v_addc_u32_e32 v14, vcc, v17, v15, vcc6482; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc6483; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v14, v126484; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc6485; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v21, v126486; TONGA-NEXT:    v_addc_u32_e32 v15, vcc, v22, v13, vcc6487; TONGA-NEXT:    v_ashrrev_i32_e32 v16, 31, v56488; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v4, v166489; TONGA-NEXT:    v_xor_b32_e32 v17, v12, v166490; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v17, v15, 06491; TONGA-NEXT:    v_mul_hi_u32 v19, v17, v146492; TONGA-NEXT:    v_addc_u32_e32 v5, vcc, v5, v16, vcc6493; TONGA-NEXT:    v_xor_b32_e32 v5, v5, v166494; TONGA-NEXT:    v_add_u32_e32 v19, vcc, v19, v126495; TONGA-NEXT:    v_addc_u32_e32 v20, vcc, 0, v13, vcc6496; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v5, v14, 06497; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v5, v15, 06498; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v19, v126499; TONGA-NEXT:    v_addc_u32_e32 v12, vcc, v20, v13, vcc6500; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v15, vcc6501; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v12, v146502; TONGA-NEXT:    v_addc_u32_e32 v12, vcc, 0, v13, vcc6503; TONGA-NEXT:    v_mul_lo_u32 v15, v18, v126504; TONGA-NEXT:    v_mad_u64_u32 v[12:13], s[0:1], v18, v14, 06505; TONGA-NEXT:    v_mul_lo_u32 v14, v1, v146506; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v15, v136507; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v14, v136508; TONGA-NEXT:    v_sub_u32_e32 v14, vcc, v5, v136509; TONGA-NEXT:    v_sub_u32_e32 v12, vcc, v17, v126510; TONGA-NEXT:    v_subb_u32_e64 v14, s[0:1], v14, v1, vcc6511; TONGA-NEXT:    v_sub_u32_e64 v15, s[0:1], v12, v186512; TONGA-NEXT:    v_subbrev_u32_e64 v17, s[2:3], 0, v14, s[0:1]6513; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v17, v16514; TONGA-NEXT:    v_cndmask_b32_e64 v19, 0, -1, s[2:3]6515; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v15, v186516; TONGA-NEXT:    v_cndmask_b32_e64 v20, 0, -1, s[2:3]6517; TONGA-NEXT:    v_cmp_eq_u32_e64 s[2:3], v17, v16518; TONGA-NEXT:    v_subb_u32_e64 v14, s[0:1], v14, v1, s[0:1]6519; TONGA-NEXT:    v_cndmask_b32_e64 v19, v19, v20, s[2:3]6520; TONGA-NEXT:    v_sub_u32_e64 v20, s[0:1], v15, v186521; TONGA-NEXT:    v_subb_u32_e32 v5, vcc, v5, v13, vcc6522; TONGA-NEXT:    v_subbrev_u32_e64 v14, s[0:1], 0, v14, s[0:1]6523; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v16524; TONGA-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v196525; TONGA-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc6526; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v12, v186527; TONGA-NEXT:    v_cndmask_b32_e64 v14, v17, v14, s[0:1]6528; TONGA-NEXT:    v_cndmask_b32_e64 v17, 0, -1, vcc6529; TONGA-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v16530; TONGA-NEXT:    v_cndmask_b32_e32 v1, v13, v17, vcc6531; TONGA-NEXT:    v_cndmask_b32_e64 v15, v15, v20, s[0:1]6532; TONGA-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v16533; TONGA-NEXT:    v_cndmask_b32_e32 v1, v5, v14, vcc6534; TONGA-NEXT:    v_cndmask_b32_e32 v5, v12, v15, vcc6535; TONGA-NEXT:    v_xor_b32_e32 v5, v5, v166536; TONGA-NEXT:    v_xor_b32_e32 v1, v1, v166537; TONGA-NEXT:    v_sub_u32_e32 v12, vcc, v5, v166538; TONGA-NEXT:    v_subb_u32_e32 v13, vcc, v1, v16, vcc6539; TONGA-NEXT:    s_cbranch_execnz .LBB12_116540; TONGA-NEXT:  .LBB12_10:6541; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, v06542; TONGA-NEXT:    v_sub_u32_e32 v5, vcc, 0, v06543; TONGA-NEXT:    v_mov_b32_e32 v13, 06544; TONGA-NEXT:    v_rcp_iflag_f32_e32 v1, v16545; TONGA-NEXT:    v_mul_f32_e32 v1, 0x4f7ffffe, v16546; TONGA-NEXT:    v_cvt_u32_f32_e32 v1, v16547; TONGA-NEXT:    v_mul_lo_u32 v5, v5, v16548; TONGA-NEXT:    v_mul_hi_u32 v5, v1, v56549; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v1, v56550; TONGA-NEXT:    v_mul_hi_u32 v1, v4, v16551; TONGA-NEXT:    v_mul_lo_u32 v1, v1, v06552; TONGA-NEXT:    v_sub_u32_e32 v1, vcc, v4, v16553; TONGA-NEXT:    v_subrev_u32_e32 v4, vcc, v0, v16554; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v06555; TONGA-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc6556; TONGA-NEXT:    v_subrev_u32_e32 v4, vcc, v0, v16557; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v06558; TONGA-NEXT:    v_cndmask_b32_e32 v12, v1, v4, vcc6559; TONGA-NEXT:  .LBB12_11:6560; TONGA-NEXT:    v_or_b32_e32 v1, v7, v36561; TONGA-NEXT:    v_mov_b32_e32 v0, 06562; TONGA-NEXT:    v_cmp_ne_u64_e32 vcc, 0, v[0:1]6563; TONGA-NEXT:    s_cbranch_vccz .LBB12_176564; TONGA-NEXT:  ; %bb.12:6565; TONGA-NEXT:    v_ashrrev_i32_e32 v0, 31, v36566; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v2, v06567; TONGA-NEXT:    v_addc_u32_e32 v3, vcc, v3, v0, vcc6568; TONGA-NEXT:    v_xor_b32_e32 v5, v1, v06569; TONGA-NEXT:    v_xor_b32_e32 v16, v3, v06570; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, v56571; TONGA-NEXT:    v_cvt_f32_u32_e32 v1, v166572; TONGA-NEXT:    v_sub_u32_e32 v17, vcc, 0, v56573; TONGA-NEXT:    v_subb_u32_e32 v18, vcc, 0, v16, vcc6574; TONGA-NEXT:    v_madmk_f32 v0, v1, 0x4f800000, v06575; TONGA-NEXT:    v_rcp_f32_e32 v0, v06576; TONGA-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v06577; TONGA-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v06578; TONGA-NEXT:    v_trunc_f32_e32 v1, v16579; TONGA-NEXT:    v_madmk_f32 v0, v1, 0xcf800000, v06580; TONGA-NEXT:    v_cvt_u32_f32_e32 v14, v16581; TONGA-NEXT:    v_cvt_u32_f32_e32 v15, v06582; TONGA-NEXT:    v_mul_lo_u32 v3, v17, v146583; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v17, v15, 06584; TONGA-NEXT:    v_mul_lo_u32 v4, v18, v156585; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v1, v36586; TONGA-NEXT:    v_add_u32_e32 v19, vcc, v1, v46587; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v15, v19, 06588; TONGA-NEXT:    v_mul_hi_u32 v1, v15, v06589; TONGA-NEXT:    v_add_u32_e32 v20, vcc, v1, v36590; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v14, v0, 06591; TONGA-NEXT:    v_addc_u32_e32 v21, vcc, 0, v4, vcc6592; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v14, v19, 06593; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v20, v06594; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v21, v1, vcc6595; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v4, vcc6596; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v36597; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc6598; TONGA-NEXT:    v_add_u32_e32 v19, vcc, v15, v06599; TONGA-NEXT:    v_addc_u32_e32 v20, vcc, v14, v1, vcc6600; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v17, v19, 06601; TONGA-NEXT:    v_mul_lo_u32 v14, v17, v206602; TONGA-NEXT:    v_mul_lo_u32 v15, v18, v196603; TONGA-NEXT:    v_mul_hi_u32 v17, v19, v06604; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v20, v0, 06605; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v14, v16606; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v15, v16607; TONGA-NEXT:    v_mad_u64_u32 v[14:15], s[0:1], v19, v1, 06608; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v20, v1, 06609; TONGA-NEXT:    v_add_u32_e32 v14, vcc, v17, v146610; TONGA-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc6611; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v14, v36612; TONGA-NEXT:    v_addc_u32_e32 v3, vcc, v15, v4, vcc6613; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc6614; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v3, v06615; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc6616; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v19, v06617; TONGA-NEXT:    v_addc_u32_e32 v4, vcc, v20, v1, vcc6618; TONGA-NEXT:    v_ashrrev_i32_e32 v15, 31, v76619; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v6, v156620; TONGA-NEXT:    v_xor_b32_e32 v14, v0, v156621; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v14, v4, 06622; TONGA-NEXT:    v_mul_hi_u32 v17, v14, v36623; TONGA-NEXT:    v_addc_u32_e32 v7, vcc, v7, v15, vcc6624; TONGA-NEXT:    v_xor_b32_e32 v7, v7, v156625; TONGA-NEXT:    v_add_u32_e32 v17, vcc, v17, v06626; TONGA-NEXT:    v_addc_u32_e32 v18, vcc, 0, v1, vcc6627; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v7, v3, 06628; TONGA-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], v7, v4, 06629; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v17, v06630; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, v18, v1, vcc6631; TONGA-NEXT:    v_addc_u32_e32 v1, vcc, 0, v4, vcc6632; TONGA-NEXT:    v_add_u32_e32 v3, vcc, v0, v36633; TONGA-NEXT:    v_addc_u32_e32 v0, vcc, 0, v1, vcc6634; TONGA-NEXT:    v_mul_lo_u32 v4, v5, v06635; TONGA-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v5, v3, 06636; TONGA-NEXT:    v_mul_lo_u32 v3, v16, v36637; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v4, v16638; TONGA-NEXT:    v_add_u32_e32 v1, vcc, v3, v16639; TONGA-NEXT:    v_sub_u32_e32 v3, vcc, v7, v16640; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v14, v06641; TONGA-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v16, vcc6642; TONGA-NEXT:    v_sub_u32_e64 v4, s[0:1], v0, v56643; TONGA-NEXT:    v_subbrev_u32_e64 v14, s[2:3], 0, v3, s[0:1]6644; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v14, v166645; TONGA-NEXT:    v_cndmask_b32_e64 v17, 0, -1, s[2:3]6646; TONGA-NEXT:    v_cmp_ge_u32_e64 s[2:3], v4, v56647; TONGA-NEXT:    v_subb_u32_e32 v1, vcc, v7, v1, vcc6648; TONGA-NEXT:    v_cndmask_b32_e64 v18, 0, -1, s[2:3]6649; TONGA-NEXT:    v_cmp_eq_u32_e64 s[2:3], v14, v166650; TONGA-NEXT:    v_subb_u32_e64 v3, s[0:1], v3, v16, s[0:1]6651; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v166652; TONGA-NEXT:    v_cndmask_b32_e64 v17, v17, v18, s[2:3]6653; TONGA-NEXT:    v_sub_u32_e64 v18, s[0:1], v4, v56654; TONGA-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc6655; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v56656; TONGA-NEXT:    v_subbrev_u32_e64 v3, s[0:1], 0, v3, s[0:1]6657; TONGA-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc6658; TONGA-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v166659; TONGA-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v176660; TONGA-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc6661; TONGA-NEXT:    v_cndmask_b32_e64 v4, v4, v18, s[0:1]6662; TONGA-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v56663; TONGA-NEXT:    v_cndmask_b32_e64 v3, v14, v3, s[0:1]6664; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc6665; TONGA-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc6666; TONGA-NEXT:    v_xor_b32_e32 v0, v0, v156667; TONGA-NEXT:    v_xor_b32_e32 v1, v1, v156668; TONGA-NEXT:    v_sub_u32_e32 v14, vcc, v0, v156669; TONGA-NEXT:    v_subb_u32_e32 v15, vcc, v1, v15, vcc6670; TONGA-NEXT:    s_cbranch_execnz .LBB12_146671; TONGA-NEXT:  .LBB12_13:6672; TONGA-NEXT:    v_cvt_f32_u32_e32 v0, v26673; TONGA-NEXT:    v_sub_u32_e32 v1, vcc, 0, v26674; TONGA-NEXT:    v_mov_b32_e32 v15, 06675; TONGA-NEXT:    v_rcp_iflag_f32_e32 v0, v06676; TONGA-NEXT:    v_mul_f32_e32 v0, 0x4f7ffffe, v06677; TONGA-NEXT:    v_cvt_u32_f32_e32 v0, v06678; TONGA-NEXT:    v_mul_lo_u32 v1, v1, v06679; TONGA-NEXT:    v_mul_hi_u32 v1, v0, v16680; TONGA-NEXT:    v_add_u32_e32 v0, vcc, v0, v16681; TONGA-NEXT:    v_mul_hi_u32 v0, v6, v06682; TONGA-NEXT:    v_mul_lo_u32 v0, v0, v26683; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v6, v06684; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, v2, v06685; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v26686; TONGA-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc6687; TONGA-NEXT:    v_subrev_u32_e32 v1, vcc, v2, v06688; TONGA-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v26689; TONGA-NEXT:    v_cndmask_b32_e32 v14, v0, v1, vcc6690; TONGA-NEXT:  .LBB12_14:6691; TONGA-NEXT:    v_mov_b32_e32 v0, s46692; TONGA-NEXT:    v_mov_b32_e32 v1, s56693; TONGA-NEXT:    s_add_u32 s0, s4, 166694; TONGA-NEXT:    flat_store_dwordx4 v[0:1], v[8:11]6695; TONGA-NEXT:    s_addc_u32 s1, s5, 06696; TONGA-NEXT:    v_mov_b32_e32 v0, s06697; TONGA-NEXT:    v_mov_b32_e32 v1, s16698; TONGA-NEXT:    flat_store_dwordx4 v[0:1], v[12:15]6699; TONGA-NEXT:    s_endpgm6700; TONGA-NEXT:  .LBB12_15:6701; TONGA-NEXT:    s_branch .LBB12_76702; TONGA-NEXT:  .LBB12_16:6703; TONGA-NEXT:    ; implicit-def: $vgpr12_vgpr136704; TONGA-NEXT:    s_branch .LBB12_106705; TONGA-NEXT:  .LBB12_17:6706; TONGA-NEXT:    s_branch .LBB12_136707;6708; EG-LABEL: srem_v4i64:6709; EG:       ; %bb.0:6710; EG-NEXT:    ALU 0, @34, KC0[CB0:0-32], KC1[]6711; EG-NEXT:    TEX 1 @266712; EG-NEXT:    ALU 114, @35, KC0[], KC1[]6713; EG-NEXT:    ALU 115, @150, KC0[], KC1[]6714; EG-NEXT:    ALU 115, @266, KC0[], KC1[]6715; EG-NEXT:    ALU 111, @382, KC0[], KC1[]6716; EG-NEXT:    TEX 1 @306717; EG-NEXT:    ALU 114, @494, KC0[], KC1[]6718; EG-NEXT:    ALU 113, @609, KC0[], KC1[]6719; EG-NEXT:    ALU 114, @723, KC0[], KC1[]6720; EG-NEXT:    ALU 113, @838, KC0[], KC1[]6721; EG-NEXT:    ALU 114, @952, KC0[], KC1[]6722; EG-NEXT:    ALU 113, @1067, KC0[], KC1[]6723; EG-NEXT:    ALU 114, @1181, KC0[], KC1[]6724; EG-NEXT:    ALU 113, @1296, KC0[], KC1[]6725; EG-NEXT:    ALU 114, @1410, KC0[], KC1[]6726; EG-NEXT:    ALU 114, @1525, KC0[], KC1[]6727; EG-NEXT:    ALU 114, @1640, KC0[], KC1[]6728; EG-NEXT:    ALU 115, @1755, KC0[], KC1[]6729; EG-NEXT:    ALU 113, @1871, KC0[], KC1[]6730; EG-NEXT:    ALU 112, @1985, KC0[], KC1[]6731; EG-NEXT:    ALU 99, @2098, KC0[CB0:0-32], KC1[]6732; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T1.X, 06733; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T0.X, 16734; EG-NEXT:    CF_END6735; EG-NEXT:    PAD6736; EG-NEXT:    Fetch clause starting at 26:6737; EG-NEXT:     VTX_READ_128 T1.XYZW, T2.X, 32, #16738; EG-NEXT:     VTX_READ_128 T0.XYZW, T2.X, 0, #16739; EG-NEXT:    Fetch clause starting at 30:6740; EG-NEXT:     VTX_READ_128 T9.XYZW, T2.X, 16, #16741; EG-NEXT:     VTX_READ_128 T10.XYZW, T2.X, 48, #16742; EG-NEXT:    ALU clause starting at 34:6743; EG-NEXT:     MOV * T2.X, KC0[2].Z,6744; EG-NEXT:    ALU clause starting at 35:6745; EG-NEXT:     ASHR * T3.W, T1.Y, literal.x,6746; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6747; EG-NEXT:     ADD_INT * T2.W, T1.X, PV.W,6748; EG-NEXT:     XOR_INT * T7.W, PV.W, T3.W,6749; EG-NEXT:     SUB_INT T2.Z, 0.0, PV.W,6750; EG-NEXT:     ASHR T2.W, T0.Y, literal.x,6751; EG-NEXT:     RECIP_UINT * T2.Y, PV.W,6752; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6753; EG-NEXT:     ADD_INT T3.Z, T0.Y, PV.W,6754; EG-NEXT:     ADDC_UINT T4.W, T0.X, PV.W,6755; EG-NEXT:     MULLO_INT * T0.Y, PV.Z, PS,6756; EG-NEXT:     ADD_INT T4.W, PV.Z, PV.W,6757; EG-NEXT:     MULHI * T0.Y, T2.Y, PS,6758; EG-NEXT:     ADD_INT T5.W, T2.Y, PS,6759; EG-NEXT:     XOR_INT * T4.W, PV.W, T2.W,6760; EG-NEXT:     MULHI * T0.Y, PS, PV.W,6761; EG-NEXT:     MULLO_INT * T0.Y, PS, T7.W,6762; EG-NEXT:     SUB_INT * T5.W, T4.W, PS,6763; EG-NEXT:     SETGE_UINT T6.W, PV.W, T7.W,6764; EG-NEXT:     SUB_INT * T8.W, PV.W, T7.W,6765; EG-NEXT:     CNDE_INT T2.Z, PV.W, T5.W, PS, BS:VEC_021/SCL_1226766; EG-NEXT:     ADD_INT T5.W, T1.Y, T3.W,6767; EG-NEXT:     ADDC_UINT * T6.W, T1.X, T3.W,6768; EG-NEXT:     ADD_INT T3.Z, PV.W, PS,6769; EG-NEXT:     SETGE_UINT T5.W, PV.Z, T7.W,6770; EG-NEXT:     SUB_INT * T6.W, PV.Z, T7.W,6771; EG-NEXT:     ADD_INT T4.Z, T0.X, T2.W, BS:VEC_021/SCL_1226772; EG-NEXT:     CNDE_INT T5.W, PV.W, T2.Z, PS,6773; EG-NEXT:     XOR_INT * T6.W, PV.Z, T3.W,6774; EG-NEXT:     CNDE_INT T3.W, PS, PV.W, T4.W,6775; EG-NEXT:     XOR_INT * T8.W, PV.Z, T2.W,6776; EG-NEXT:     BIT_ALIGN_INT T4.W, PV.W, PS, literal.x,6777; EG-NEXT:     LSHR * T3.W, PV.W, literal.x,6778; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6779; EG-NEXT:     SETE_INT T2.Z, PS, T6.W, BS:VEC_021/SCL_1226780; EG-NEXT:     SETGE_UINT T5.W, PS, T6.W, BS:VEC_021/SCL_1226781; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T7.W,6782; EG-NEXT:     CNDE_INT T0.Y, PV.Z, PV.W, PS,6783; EG-NEXT:     SUB_INT * T2.Z, T4.W, T7.W,6784; EG-NEXT:     SUB_INT * T5.W, T3.W, T6.W,6785; EG-NEXT:     SUBB_UINT * T9.W, T4.W, T7.W,6786; EG-NEXT:     SUB_INT T5.W, T5.W, PV.W,6787; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6788; EG-NEXT:     LSHL T2.Z, PS, 1,6789; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6790; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6791; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)6792; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6793; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6794; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6795; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226796; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6797; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6798; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6799; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6800; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6801; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6802; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226803; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6804; EG-NEXT:     LSHL T2.Z, PS, 1,6805; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6806; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6807; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)6808; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6809; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6810; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6811; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226812; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6813; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6814; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6815; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6816; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6817; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6818; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226819; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6820; EG-NEXT:     LSHL T2.Z, PS, 1,6821; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6822; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6823; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)6824; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6825; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6826; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6827; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226828; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6829; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6830; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6831; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6832; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6833; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6834; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226835; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6836; EG-NEXT:     LSHL T2.Z, PS, 1,6837; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6838; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6839; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)6840; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6841; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6842; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6843; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226844; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6845; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6846; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6847; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6848; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6849; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6850; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226851; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6852; EG-NEXT:     LSHL T2.Z, PS, 1,6853; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6854; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6855; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)6856; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6857; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6858; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6859; EG-NEXT:     SETGE_UINT * T2.Z, PS, T7.W,6860; EG-NEXT:    ALU clause starting at 150:6861; EG-NEXT:     SETE_INT T5.W, T3.W, T6.W,6862; EG-NEXT:     SETGE_UINT * T9.W, T3.W, T6.W,6863; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, T2.Z,6864; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6865; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6866; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6867; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226868; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6869; EG-NEXT:     LSHL T2.Z, PS, 1,6870; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6871; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6872; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)6873; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6874; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6875; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6876; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226877; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6878; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6879; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6880; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6881; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6882; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6883; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226884; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6885; EG-NEXT:     LSHL T2.Z, PS, 1,6886; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6887; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6888; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)6889; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6890; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6891; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6892; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226893; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6894; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6895; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6896; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6897; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6898; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6899; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226900; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6901; EG-NEXT:     LSHL T2.Z, PS, 1,6902; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6903; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6904; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)6905; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6906; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6907; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6908; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226909; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6910; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6911; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6912; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6913; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6914; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6915; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226916; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6917; EG-NEXT:     LSHL T2.Z, PS, 1,6918; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6919; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6920; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)6921; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6922; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6923; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6924; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226925; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6926; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6927; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6928; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6929; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6930; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6931; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226932; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6933; EG-NEXT:     LSHL T2.Z, PS, 1,6934; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6935; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6936; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)6937; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6938; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6939; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6940; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226941; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6942; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6943; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6944; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6945; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6946; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6947; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226948; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6949; EG-NEXT:     LSHL T2.Z, PS, 1,6950; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6951; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6952; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)6953; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6954; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6955; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6956; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226957; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6958; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6959; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6960; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6961; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6962; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6963; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226964; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6965; EG-NEXT:     LSHL T2.Z, PS, 1,6966; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6967; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6968; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)6969; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6970; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6971; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6972; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226973; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6974; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6975; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6976; EG-NEXT:     SUB_INT * T2.Z, T4.W, T7.W,6977; EG-NEXT:    ALU clause starting at 266:6978; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6979; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6980; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226981; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6982; EG-NEXT:     LSHL T2.Z, PS, 1,6983; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,6984; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,6985; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)6986; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,6987; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,6988; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)6989; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1226990; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,6991; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,6992; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,6993; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,6994; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,6995; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,6996; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1226997; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,6998; EG-NEXT:     LSHL T2.Z, PS, 1,6999; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7000; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7001; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)7002; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7003; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7004; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7005; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227006; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7007; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7008; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7009; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7010; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7011; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7012; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227013; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7014; EG-NEXT:     LSHL T2.Z, PS, 1,7015; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7016; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7017; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)7018; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7019; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7020; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7021; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227022; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7023; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7024; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7025; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7026; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7027; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7028; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227029; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7030; EG-NEXT:     LSHL T2.Z, PS, 1,7031; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7032; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7033; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)7034; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7035; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7036; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7037; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227038; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7039; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7040; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7041; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7042; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7043; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7044; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227045; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7046; EG-NEXT:     LSHL T2.Z, PS, 1,7047; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7048; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7049; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)7050; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7051; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7052; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7053; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227054; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7055; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7056; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7057; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7058; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7059; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7060; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227061; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7062; EG-NEXT:     LSHL T2.Z, PS, 1,7063; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7064; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7065; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)7066; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7067; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7068; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7069; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227070; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7071; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7072; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7073; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7074; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7075; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7076; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227077; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7078; EG-NEXT:     LSHL T2.Z, PS, 1,7079; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7080; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7081; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)7082; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7083; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7084; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7085; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227086; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7087; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7088; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7089; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7090; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7091; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7092; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227093; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7094; EG-NEXT:    ALU clause starting at 382:7095; EG-NEXT:     LSHL T2.Z, T4.W, 1,7096; EG-NEXT:     BFE_UINT * T9.W, T8.W, literal.x, 1, BS:VEC_120/SCL_2127097; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)7098; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, T5.W,7099; EG-NEXT:     BIT_ALIGN_INT T3.W, PV.W, T4.W, literal.x, BS:VEC_021/SCL_1227100; EG-NEXT:     OR_INT * T4.W, T2.Z, T9.W,7101; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7102; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227103; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7104; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7105; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7106; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7107; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7108; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7109; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227110; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7111; EG-NEXT:     LSHL T2.Z, PS, 1,7112; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7113; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7114; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)7115; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7116; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7117; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7118; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227119; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7120; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7121; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7122; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7123; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7124; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7125; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227126; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7127; EG-NEXT:     LSHL T2.Z, PS, 1,7128; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7129; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7130; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)7131; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7132; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7133; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7134; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227135; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7136; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7137; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7138; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7139; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7140; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7141; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227142; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7143; EG-NEXT:     LSHL T2.Z, PS, 1,7144; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7145; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7146; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)7147; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7148; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7149; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7150; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227151; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7152; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7153; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7154; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7155; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7156; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7157; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227158; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7159; EG-NEXT:     LSHL T2.Z, PS, 1,7160; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7161; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7162; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)7163; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7164; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7165; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7166; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227167; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7168; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7169; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7170; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7171; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7172; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7173; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227174; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7175; EG-NEXT:     LSHL T2.Z, PS, 1,7176; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7177; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7178; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)7179; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7180; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7181; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7182; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227183; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7184; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7185; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7186; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7187; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7188; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7189; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227190; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7191; EG-NEXT:     LSHL T2.Z, PS, 1,7192; EG-NEXT:     BFE_UINT T9.W, T8.W, literal.x, 1,7193; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7194; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)7195; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7196; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7197; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7198; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227199; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7200; EG-NEXT:     SETGE_UINT * T9.W, PV.W, T6.W,7201; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7202; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7203; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7204; EG-NEXT:     SUB_INT * T9.W, T3.W, T6.W,7205; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227206; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7207; EG-NEXT:    ALU clause starting at 494:7208; EG-NEXT:     LSHL T2.Z, T4.W, 1,7209; EG-NEXT:     BFE_UINT * T11.W, T8.W, literal.x, 1, BS:VEC_120/SCL_2127210; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)7211; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, T5.W,7212; EG-NEXT:     BIT_ALIGN_INT T3.W, PV.W, T4.W, literal.x, BS:VEC_021/SCL_1227213; EG-NEXT:     OR_INT * T4.W, T2.Z, T11.W,7214; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7215; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227216; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7217; EG-NEXT:     SETGE_UINT * T11.W, PV.W, T6.W,7218; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7219; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7220; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7221; EG-NEXT:     SUB_INT * T11.W, T3.W, T6.W,7222; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227223; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7224; EG-NEXT:     LSHL T2.Z, PS, 1,7225; EG-NEXT:     BFE_UINT T11.W, T8.W, literal.x, 1,7226; EG-NEXT:     CNDE_INT * T3.W, T0.Y, T3.W, PV.W,7227; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)7228; EG-NEXT:     BIT_ALIGN_INT T3.W, PS, T4.W, literal.x,7229; EG-NEXT:     OR_INT * T4.W, PV.Z, PV.W,7230; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7231; EG-NEXT:     SETGE_UINT T2.Z, PS, T7.W, BS:VEC_021/SCL_1227232; EG-NEXT:     SETE_INT T5.W, PV.W, T6.W,7233; EG-NEXT:     SETGE_UINT * T11.W, PV.W, T6.W,7234; EG-NEXT:     CNDE_INT T0.Y, PV.W, PS, PV.Z,7235; EG-NEXT:     SUB_INT T2.Z, T4.W, T7.W,7236; EG-NEXT:     SUBB_UINT * T5.W, T4.W, T7.W,7237; EG-NEXT:     SUB_INT * T11.W, T3.W, T6.W,7238; EG-NEXT:     SUB_INT T5.W, PV.W, T5.W, BS:VEC_021/SCL_1227239; EG-NEXT:     CNDE_INT * T4.W, T0.Y, T4.W, T2.Z,7240; EG-NEXT:     LSHL T1.Y, PS, 1,7241; EG-NEXT:     BFE_UINT T2.Z, T8.W, literal.x, 1,7242; EG-NEXT:     CNDE_INT T3.W, T0.Y, T3.W, PV.W,7243; EG-NEXT:     ASHR * T11.W, T10.Y, literal.y,7244; EG-NEXT:    2(2.802597e-45), 31(4.344025e-44)7245; EG-NEXT:     ADD_INT T3.Z, T10.X, PS,7246; EG-NEXT:     BIT_ALIGN_INT T5.W, PV.W, T4.W, literal.x,7247; EG-NEXT:     OR_INT * T12.W, PV.Y, PV.Z,7248; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7249; EG-NEXT:     SETGE_UINT T0.Y, PS, T7.W, BS:VEC_021/SCL_1227250; EG-NEXT:     SETE_INT T2.Z, PV.W, T6.W, BS:VEC_102/SCL_2217251; EG-NEXT:     SETGE_UINT T3.W, PV.W, T6.W, BS:VEC_102/SCL_2217252; EG-NEXT:     XOR_INT * T4.W, PV.Z, T11.W,7253; EG-NEXT:     SUB_INT T13.W, 0.0, PS,7254; EG-NEXT:     CNDE_INT * T14.W, PV.Z, PV.W, PV.Y,7255; EG-NEXT:     SUB_INT T0.X, T12.W, T7.W,7256; EG-NEXT:     SUBB_UINT * T0.Y, T12.W, T7.W,7257; EG-NEXT:     SUB_INT T2.Z, T5.W, T6.W,7258; EG-NEXT:     ASHR T3.W, T9.Y, literal.x,7259; EG-NEXT:     RECIP_UINT * T1.X, T4.W,7260; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7261; EG-NEXT:     ADD_INT T1.Y, T9.Y, PV.W,7262; EG-NEXT:     SUB_INT T2.Z, PV.Z, T0.Y,7263; EG-NEXT:     CNDE_INT T12.W, T14.W, T12.W, T0.X,7264; EG-NEXT:     MULLO_INT * T0.X, T13.W, PS,7265; EG-NEXT:     ADDC_UINT T2.X, T9.X, T3.W,7266; EG-NEXT:     LSHL T0.Y, PV.W, 1,7267; EG-NEXT:     BFE_UINT * T3.Z, T8.W, 1, 1,7268; EG-NEXT:     CNDE_INT T5.W, T14.W, T5.W, T2.Z,7269; EG-NEXT:     MULHI * T0.X, T1.X, T0.X,7270; EG-NEXT:     ADD_INT T2.Y, T1.X, PS,7271; EG-NEXT:     BIT_ALIGN_INT T4.Z, PV.W, T12.W, literal.x,7272; EG-NEXT:     OR_INT T12.W, T0.Y, T3.Z,7273; EG-NEXT:     ADD_INT * T5.W, T1.Y, T2.X,7274; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7275; EG-NEXT:     XOR_INT T1.X, PS, T3.W,7276; EG-NEXT:     ASHR T0.Y, T10.W, literal.x,7277; EG-NEXT:     SETGE_UINT * T2.Z, PV.W, T7.W, BS:VEC_021/SCL_1227278; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7279; EG-NEXT:     SETE_INT T5.W, T4.Z, T6.W,7280; EG-NEXT:     SETGE_UINT * T13.W, T4.Z, T6.W,7281; EG-NEXT:     CNDE_INT T1.Y, PV.W, PS, T2.Z,7282; EG-NEXT:     SUB_INT T3.Z, T12.W, T7.W,7283; EG-NEXT:     ADD_INT T5.W, T10.Z, T0.Y, BS:VEC_021/SCL_1227284; EG-NEXT:     MULHI * T0.X, T1.X, T2.Y,7285; EG-NEXT:     SUBB_UINT T2.X, T12.W, T7.W,7286; EG-NEXT:     SUB_INT T2.Y, T4.Z, T6.W, BS:VEC_021/SCL_1227287; EG-NEXT:     XOR_INT * T2.Z, PV.W, T0.Y,7288; EG-NEXT:     ASHR T5.W, T9.W, literal.x,7289; EG-NEXT:     MULLO_INT * T0.X, T0.X, T4.W,7290; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7291; EG-NEXT:     ADD_INT T3.X, T9.W, PV.W,7292; EG-NEXT:     SUB_INT T3.Y, 0.0, T2.Z,7293; EG-NEXT:     SUB_INT T5.Z, T2.Y, T2.X,7294; EG-NEXT:     CNDE_INT T9.W, T1.Y, T12.W, T3.Z, BS:VEC_120/SCL_2127295; EG-NEXT:     RECIP_UINT * T2.X, T2.Z,7296; EG-NEXT:     ADDC_UINT T4.X, T9.Z, T5.W,7297; EG-NEXT:     LSHL T2.Y, PV.W, 1,7298; EG-NEXT:     AND_INT T3.Z, T8.W, 1,7299; EG-NEXT:     CNDE_INT T8.W, T1.Y, T4.Z, PV.Z,7300; EG-NEXT:     MULLO_INT * T1.Y, PV.Y, PS,7301; EG-NEXT:     BIT_ALIGN_INT T3.Y, PV.W, T9.W, literal.x,7302; EG-NEXT:     OR_INT T5.Z, PV.Y, PV.Z,7303; EG-NEXT:     ADD_INT T8.W, T3.X, PV.X,7304; EG-NEXT:     MULHI * T1.Y, T2.X, PS,7305; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7306; EG-NEXT:     ADD_INT T2.X, T2.X, PS,7307; EG-NEXT:     XOR_INT T1.Y, PV.W, T5.W,7308; EG-NEXT:     SETGE_UINT T3.Z, PV.Z, T7.W, BS:VEC_021/SCL_1227309; EG-NEXT:     SETE_INT T8.W, PV.Y, T6.W, BS:VEC_102/SCL_2217310; EG-NEXT:     SUB_INT * T9.W, T1.X, T0.X,7311; EG-NEXT:     SETGE_UINT T0.X, T3.Y, T6.W, BS:VEC_021/SCL_1227312; EG-NEXT:     SUBB_UINT T2.Y, T5.Z, T7.W, BS:VEC_102/SCL_2217313; EG-NEXT:     SUB_INT T4.Z, T3.Y, T6.W, BS:VEC_021/SCL_1227314; EG-NEXT:     SETGE_UINT T6.W, PS, T4.W,7315; EG-NEXT:     SUB_INT * T12.W, PS, T4.W,7316; EG-NEXT:     CNDE_INT T3.X, PV.W, T9.W, PS,7317; EG-NEXT:     ADD_INT T4.Y, T10.Y, T11.W, BS:VEC_102/SCL_2217318; EG-NEXT:     ADDC_UINT T6.Z, T10.X, T11.W, BS:VEC_102/SCL_2217319; EG-NEXT:     SUB_INT T6.W, PV.Z, PV.Y,7320; EG-NEXT:     CNDE_INT * T8.W, T8.W, PV.X, T3.Z,7321; EG-NEXT:     CNDE_INT T0.X, PS, T3.Y, PV.W,7322; EG-NEXT:     ADD_INT * T2.Y, PV.Y, PV.Z,7323; EG-NEXT:    ALU clause starting at 609:7324; EG-NEXT:     SETGE_UINT T3.Z, T3.X, T4.W,7325; EG-NEXT:     SUB_INT T6.W, T3.X, T4.W,7326; EG-NEXT:     MULHI * T2.X, T1.Y, T2.X,7327; EG-NEXT:     ASHR T4.X, T1.W, literal.x,7328; EG-NEXT:     CNDE_INT T3.Y, PV.Z, T3.X, PV.W,7329; EG-NEXT:     XOR_INT T3.Z, T2.Y, T11.W,7330; EG-NEXT:     ADD_INT T6.W, T9.Z, T5.W, BS:VEC_021/SCL_1227331; EG-NEXT:     MULLO_INT * T2.X, PS, T2.Z,7332; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7333; EG-NEXT:     XOR_INT T3.X, PV.W, T5.W,7334; EG-NEXT:     CNDE_INT T2.Y, PV.Z, PV.Y, T1.X,7335; EG-NEXT:     ADD_INT T4.Z, T1.Z, PV.X,7336; EG-NEXT:     SUB_INT T9.W, T1.Y, PS,7337; EG-NEXT:     ASHR * T6.W, T0.W, literal.x,7338; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7339; EG-NEXT:     ADD_INT T1.X, T0.W, PS,7340; EG-NEXT:     ADDC_UINT T3.Y, T0.Z, PS,7341; EG-NEXT:     SETGE_UINT T6.Z, PV.W, T2.Z,7342; EG-NEXT:     SUB_INT T11.W, PV.W, T2.Z,7343; EG-NEXT:     XOR_INT * T0.W, PV.Z, T4.X,7344; EG-NEXT:     SUB_INT T2.X, 0.0, PS,7345; EG-NEXT:     CNDE_INT T4.Y, PV.Z, T9.W, PV.W,7346; EG-NEXT:     ADD_INT T4.Z, T10.W, T0.Y,7347; EG-NEXT:     ADDC_UINT T9.W, T10.Z, T0.Y,7348; EG-NEXT:     RECIP_UINT * T5.X, PS,7349; EG-NEXT:     ADD_INT T6.X, PV.Z, PV.W,7350; EG-NEXT:     SETGE_UINT T5.Y, PV.Y, T2.Z,7351; EG-NEXT:     SUB_INT T4.Z, PV.Y, T2.Z,7352; EG-NEXT:     ADD_INT T9.W, T9.X, T3.W,7353; EG-NEXT:     MULLO_INT * T6.Y, PV.X, PS,7354; EG-NEXT:     XOR_INT T2.X, PV.W, T3.W,7355; EG-NEXT:     CNDE_INT T4.Y, PV.Y, T4.Y, PV.Z,7356; EG-NEXT:     XOR_INT T4.Z, PV.X, T0.Y, BS:VEC_021/SCL_1227357; EG-NEXT:     ADD_INT T9.W, T1.X, T3.Y, BS:VEC_102/SCL_2217358; EG-NEXT:     MULHI * T0.Y, T5.X, PS,7359; EG-NEXT:     ADD_INT T1.X, T5.X, PS,7360; EG-NEXT:     XOR_INT T0.Y, PV.W, T6.W,7361; EG-NEXT:     CNDE_INT T6.Z, PV.Z, PV.Y, T1.Y, BS:VEC_021/SCL_1227362; EG-NEXT:     BIT_ALIGN_INT T9.W, T2.Y, PV.X, literal.x,7363; EG-NEXT:     LSHR * T10.W, T2.Y, literal.x,7364; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7365; EG-NEXT:     SETE_INT T5.X, PS, T3.Z,7366; EG-NEXT:     SETGE_UINT T1.Y, PS, T3.Z,7367; EG-NEXT:     SETGE_UINT T7.Z, PV.W, T4.W,7368; EG-NEXT:     LSHR T11.W, PV.Z, literal.x,7369; EG-NEXT:     MULHI * T1.X, PV.Y, PV.X,7370; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7371; EG-NEXT:     SETE_INT T6.X, PV.W, T4.Z,7372; EG-NEXT:     CNDE_INT T1.Y, PV.X, PV.Y, PV.Z,7373; EG-NEXT:     SUB_INT T7.Z, T9.W, T4.W, BS:VEC_021/SCL_1227374; EG-NEXT:     BIT_ALIGN_INT T12.W, T6.Z, T3.X, literal.x,7375; EG-NEXT:     MULLO_INT * T1.X, PS, T0.W,7376; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7377; EG-NEXT:     SETGE_UINT T5.X, T11.W, T4.Z,7378; EG-NEXT:     SETGE_UINT T2.Y, PV.W, T2.Z, BS:VEC_102/SCL_2217379; EG-NEXT:     SUB_INT T6.Z, T0.Y, PS,7380; EG-NEXT:     CNDE_INT T13.W, PV.Y, T9.W, PV.Z, BS:VEC_021/SCL_1227381; EG-NEXT:     SUB_INT * T7.W, T5.Z, T7.W,7382; EG-NEXT:     CNDE_INT T1.X, T8.W, T5.Z, PS,7383; EG-NEXT:     LSHL T3.Y, PV.W, 1,7384; EG-NEXT:     SETGE_UINT T5.Z, PV.Z, T0.W,7385; EG-NEXT:     SUB_INT T7.W, PV.Z, T0.W,7386; EG-NEXT:     CNDE_INT * T8.W, T6.X, PV.X, PV.Y,7387; EG-NEXT:     SUB_INT T5.X, T12.W, T2.Z,7388; EG-NEXT:     SUB_INT T2.Y, T11.W, T4.Z, BS:VEC_102/SCL_2217389; EG-NEXT:     SUBB_UINT T7.Z, T12.W, T2.Z,7390; EG-NEXT:     ADD_INT T1.W, T1.W, T4.X, BS:VEC_2017391; EG-NEXT:     ADDC_UINT * T14.W, T1.Z, T4.X,7392; EG-NEXT:     BFE_UINT T6.X, T2.X, literal.x, 1,7393; EG-NEXT:     ADD_INT T4.Y, PV.W, PS,7394; EG-NEXT:     SUB_INT T1.Z, PV.Y, PV.Z,7395; EG-NEXT:     CNDE_INT T1.W, T8.W, T12.W, PV.X, BS:VEC_120/SCL_2127396; EG-NEXT:     CNDE_INT * T7.W, T5.Z, T6.Z, T7.W,7397; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)7398; EG-NEXT:     SETGE_UINT T5.X, PS, T0.W, BS:VEC_102/SCL_2217399; EG-NEXT:     SUB_INT T2.Y, PS, T0.W, BS:VEC_102/SCL_2217400; EG-NEXT:     LSHL T5.Z, PV.W, 1,7401; EG-NEXT:     BFE_UINT T12.W, T3.X, literal.x, 1,7402; EG-NEXT:     CNDE_INT * T8.W, T8.W, T11.W, PV.Z,7403; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)7404; EG-NEXT:     BIT_ALIGN_INT T7.X, PS, T1.W, literal.x,7405; EG-NEXT:     OR_INT T5.Y, PV.Z, PV.W,7406; EG-NEXT:     ADD_INT T0.Z, T0.Z, T6.W, BS:VEC_021/SCL_1227407; EG-NEXT:     CNDE_INT T7.W, PV.X, T7.W, PV.Y, BS:VEC_102/SCL_2217408; EG-NEXT:     XOR_INT * T1.W, T4.Y, T4.X,7409; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7410; EG-NEXT:     CNDE_INT T4.X, PS, PV.W, T0.Y,7411; EG-NEXT:     XOR_INT T0.Y, PV.Z, T6.W,7412; EG-NEXT:     SETGE_UINT T0.Z, PV.Y, T2.Z, BS:VEC_021/SCL_1227413; EG-NEXT:     SETE_INT T7.W, PV.X, T4.Z,7414; EG-NEXT:     SETGE_UINT * T8.W, PV.X, T4.Z,7415; EG-NEXT:     SUB_INT T5.X, T10.W, T3.Z,7416; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,7417; EG-NEXT:     SUB_INT T0.Z, T5.Y, T2.Z, BS:VEC_021/SCL_1227418; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.X, PV.Y, literal.x,7419; EG-NEXT:     LSHR * T8.W, PV.X, literal.x,7420; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7421; EG-NEXT:     SUBB_UINT T4.X, T5.Y, T2.Z,7422; EG-NEXT:     SUB_INT T4.Y, T7.X, T4.Z, BS:VEC_021/SCL_1227423; EG-NEXT:     SETE_INT T1.Z, PS, T1.W, BS:VEC_021/SCL_1227424; EG-NEXT:     SETGE_UINT T11.W, PS, T1.W, BS:VEC_021/SCL_1227425; EG-NEXT:     SETGE_UINT * T12.W, PV.W, T0.W,7426; EG-NEXT:     SUB_INT T8.X, T8.W, T1.W,7427; EG-NEXT:     CNDE_INT * T6.Y, PV.Z, PV.W, PS,7428; EG-NEXT:     SUB_INT T1.Z, T7.W, T0.W,7429; EG-NEXT:     SUB_INT T11.W, T4.Y, T4.X,7430; EG-NEXT:     CNDE_INT * T12.W, T2.Y, T5.Y, T0.Z,7431; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W, BS:VEC_021/SCL_1227432; EG-NEXT:     LSHL T4.Y, PS, 1,7433; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,7434; EG-NEXT:     CNDE_INT T11.W, T2.Y, T7.X, PV.W,7435; EG-NEXT:     CNDE_INT * T7.W, T6.Y, T7.W, PV.Z,7436; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)7437; EG-NEXT:     SUBB_UINT * T7.X, T9.W, T4.W,7438; EG-NEXT:    ALU clause starting at 723:7439; EG-NEXT:     LSHL T2.Y, T7.W, 1,7440; EG-NEXT:     BIT_ALIGN_INT T1.Z, T11.W, T12.W, literal.x, BS:VEC_120/SCL_2127441; EG-NEXT:     OR_INT T9.W, T4.Y, T0.Z,7442; EG-NEXT:     SUB_INT * T11.W, T8.X, T4.X,7443; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7444; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,7445; EG-NEXT:     CNDE_INT T4.Y, T6.Y, T8.W, PS, BS:VEC_120/SCL_2127446; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1227447; EG-NEXT:     SETE_INT T8.W, PV.Z, T4.Z,7448; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T4.Z,7449; EG-NEXT:    30(4.203895e-44), 0(0.000000e+00)7450; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,7451; EG-NEXT:     SUB_INT T5.Y, T9.W, T2.Z,7452; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T7.W, literal.x,7453; EG-NEXT:     OR_INT T7.W, T2.Y, PV.X,7454; EG-NEXT:     SUB_INT * T8.W, T5.X, T7.X,7455; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7456; EG-NEXT:     CNDE_INT T4.X, T1.Y, T10.W, PS,7457; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1227458; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2217459; EG-NEXT:     SETGE_UINT * T8.W, PV.Z, T1.W, BS:VEC_102/SCL_2217460; EG-NEXT:     CNDE_INT * T10.W, T8.X, T9.W, T5.Y,7461; EG-NEXT:     LSHL T5.X, PV.W, 1,7462; EG-NEXT:     CNDE_INT T1.Y, T5.Z, T8.W, T1.Y,7463; EG-NEXT:     SUB_INT * T5.Z, T7.W, T0.W, BS:VEC_021/SCL_1227464; EG-NEXT:     BIT_ALIGN_INT T8.W, T4.X, T13.W, literal.x,7465; EG-NEXT:     OR_INT * T11.W, T3.Y, T6.X,7466; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7467; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,7468; EG-NEXT:     SUB_INT * T2.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227469; EG-NEXT:     SETGE_UINT T6.Z, T11.W, T4.W,7470; EG-NEXT:     SETE_INT T12.W, T8.W, T3.Z, BS:VEC_2017471; EG-NEXT:     SETGE_UINT * T13.W, T8.W, T3.Z,7472; EG-NEXT:     SUBB_UINT T6.X, T11.W, T4.W, BS:VEC_021/SCL_1227473; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7474; EG-NEXT:     SUB_INT T6.Z, T11.W, T4.W, BS:VEC_021/SCL_1227475; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,7476; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T5.Z,7477; EG-NEXT:     SUB_INT T4.X, T8.W, T3.Z,7478; EG-NEXT:     LSHL T2.Y, PS, 1,7479; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7480; EG-NEXT:     CNDE_INT T12.W, T1.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127481; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,7482; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)7483; EG-NEXT:     BFE_UINT T7.X, T3.X, literal.x, 1,7484; EG-NEXT:     LSHL T1.Y, PS, 1,7485; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.y,7486; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7487; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,7488; EG-NEXT:    28(3.923636e-44), 31(4.344025e-44)7489; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,7490; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T8.W, PS, BS:VEC_021/SCL_1227491; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2217492; EG-NEXT:     SETE_INT T8.W, PV.Z, T1.W,7493; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T1.W,7494; EG-NEXT:    29(4.063766e-44), 0(0.000000e+00)7495; EG-NEXT:     SUBB_UINT T6.X, T9.W, T2.Z,7496; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7497; EG-NEXT:     SUB_INT * T5.Z, T7.W, T0.W, BS:VEC_120/SCL_2127498; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,7499; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7500; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7501; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,7502; EG-NEXT:     SUB_INT * T1.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227503; EG-NEXT:     SETGE_UINT T6.Z, T9.W, T4.W,7504; EG-NEXT:     SETE_INT T11.W, T8.W, T3.Z, BS:VEC_2017505; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T3.Z,7506; EG-NEXT:     SUBB_UINT T9.X, T9.W, T4.W, BS:VEC_021/SCL_1227507; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,7508; EG-NEXT:     SUB_INT T6.Z, T9.W, T4.W, BS:VEC_021/SCL_1227509; EG-NEXT:     SUB_INT T11.W, T1.Y, T4.X,7510; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T5.Z,7511; EG-NEXT:     SUB_INT T4.X, T8.W, T3.Z,7512; EG-NEXT:     LSHL T1.Y, PS, 1,7513; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7514; EG-NEXT:     CNDE_INT T11.W, T3.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127515; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7516; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)7517; EG-NEXT:     SUB_INT T10.X, T1.Z, T4.Z,7518; EG-NEXT:     LSHL T3.Y, PS, 1,7519; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.x,7520; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7521; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,7522; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7523; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,7524; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS, BS:VEC_021/SCL_1227525; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2217526; EG-NEXT:     SETE_INT T8.W, PV.Z, T1.W,7527; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T1.W,7528; EG-NEXT:    28(3.923636e-44), 0(0.000000e+00)7529; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,7530; EG-NEXT:     SUB_INT T2.Y, T7.W, T0.W,7531; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1227532; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,7533; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,7534; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7535; EG-NEXT:     CNDE_INT T4.X, T8.X, T1.Z, PS,7536; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1227537; EG-NEXT:     SETE_INT T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1227538; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T3.Z, BS:VEC_021/SCL_1227539; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,7540; EG-NEXT:     LSHL T6.X, PS, 1,7541; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7542; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,7543; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x, BS:VEC_021/SCL_1227544; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,7545; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7546; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,7547; EG-NEXT:     SUB_INT T2.Y, T5.Z, T3.Z,7548; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1227549; EG-NEXT:     SETE_INT T12.W, T9.W, T4.Z,7550; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T4.Z,7551; EG-NEXT:     SUBB_UINT T5.X, T10.W, T2.Z,7552; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T6.Z,7553; EG-NEXT:     SUB_INT * T6.Z, T10.W, T2.Z,7554; EG-NEXT:    ALU clause starting at 838:7555; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,7556; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T1.Z,7557; EG-NEXT:     SUB_INT T4.X, T9.W, T4.Z,7558; EG-NEXT:     LSHL T2.Y, PS, 1,7559; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,7560; EG-NEXT:     CNDE_INT T12.W, T1.Y, T5.Z, PV.W, BS:VEC_021/SCL_1227561; EG-NEXT:     CNDE_INT * T10.W, T3.Y, T10.W, T6.Z,7562; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)7563; EG-NEXT:     BFE_UINT T7.X, T0.Y, literal.x, 1,7564; EG-NEXT:     LSHL T1.Y, PS, 1,7565; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.y,7566; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7567; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,7568; EG-NEXT:    27(3.783506e-44), 31(4.344025e-44)7569; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,7570; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS,7571; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1227572; EG-NEXT:     SETE_INT T9.W, PV.Z, T3.Z,7573; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T3.Z,7574; EG-NEXT:    27(3.783506e-44), 0(0.000000e+00)7575; EG-NEXT:     SUBB_UINT T5.X, T7.W, T0.W,7576; EG-NEXT:     CNDE_INT * T3.Y, PV.W, PS, PV.Z,7577; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,7578; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x, BS:VEC_021/SCL_1227579; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7580; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7581; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,7582; EG-NEXT:     SUB_INT T1.Y, T5.Z, T3.Z,7583; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1227584; EG-NEXT:     SETE_INT T10.W, T7.W, T4.Z,7585; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T4.Z,7586; EG-NEXT:     SUBB_UINT T8.X, T9.W, T2.Z,7587; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,7588; EG-NEXT:     SUB_INT T6.Z, T9.W, T2.Z,7589; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,7590; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T1.Z,7591; EG-NEXT:     SUB_INT T4.X, T7.W, T4.Z,7592; EG-NEXT:     LSHL T1.Y, PS, 1,7593; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,7594; EG-NEXT:     CNDE_INT T10.W, T3.Y, T5.Z, PV.W, BS:VEC_021/SCL_1227595; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7596; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)7597; EG-NEXT:     SUB_INT T10.X, T0.Z, T1.W,7598; EG-NEXT:     LSHL T3.Y, PS, 1,7599; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1227600; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7601; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,7602; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7603; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,7604; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS,7605; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1227606; EG-NEXT:     SETE_INT T7.W, PV.Z, T3.Z,7607; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T3.Z,7608; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)7609; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,7610; EG-NEXT:     SUB_INT T2.Y, T8.W, T4.W,7611; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1227612; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,7613; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,7614; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7615; EG-NEXT:     CNDE_INT T4.X, T9.X, T0.Z, PS,7616; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T2.Z, BS:VEC_021/SCL_1227617; EG-NEXT:     SETE_INT T0.Z, PV.Z, T4.Z, BS:VEC_102/SCL_2217618; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T4.Z, BS:VEC_102/SCL_2217619; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,7620; EG-NEXT:     LSHL T5.X, PS, 1,7621; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7622; EG-NEXT:     SUB_INT T0.Z, T7.W, T2.Z,7623; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x,7624; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,7625; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7626; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,7627; EG-NEXT:     SUB_INT T2.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1227628; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1227629; EG-NEXT:     SETE_INT T12.W, PV.W, T1.W,7630; EG-NEXT:     SETGE_UINT * T13.W, PV.W, T1.W,7631; EG-NEXT:     SUBB_UINT T6.X, T11.W, T0.W, BS:VEC_021/SCL_1227632; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7633; EG-NEXT:     SUB_INT T6.Z, T11.W, T0.W, BS:VEC_021/SCL_1227634; EG-NEXT:     SUB_INT T12.W, PV.Y, PV.X,7635; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T0.Z,7636; EG-NEXT:     SUB_INT T4.X, T9.W, T1.W, BS:VEC_021/SCL_1227637; EG-NEXT:     LSHL T2.Y, PS, 1,7638; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,7639; EG-NEXT:     CNDE_INT T12.W, T1.Y, T1.Z, PV.W,7640; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,7641; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)7642; EG-NEXT:     BFE_UINT T7.X, T2.X, literal.x, 1,7643; EG-NEXT:     LSHL T1.Y, PS, 1,7644; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T7.W, literal.y,7645; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7646; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,7647; EG-NEXT:    25(3.503246e-44), 31(4.344025e-44)7648; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,7649; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_120/SCL_2127650; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1227651; EG-NEXT:     SETE_INT T9.W, PV.Z, T4.Z,7652; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T4.Z,7653; EG-NEXT:    26(3.643376e-44), 0(0.000000e+00)7654; EG-NEXT:     SUBB_UINT T6.X, T8.W, T4.W,7655; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7656; EG-NEXT:     SUB_INT * T0.Z, T7.W, T2.Z, BS:VEC_2017657; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,7658; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7659; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7660; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,7661; EG-NEXT:     SUB_INT T1.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1227662; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1227663; EG-NEXT:     SETE_INT T11.W, PV.W, T1.W,7664; EG-NEXT:     SETGE_UINT * T12.W, PV.W, T1.W,7665; EG-NEXT:     SUBB_UINT T9.X, T9.W, T0.W,7666; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,7667; EG-NEXT:     SUB_INT T6.Z, T9.W, T0.W,7668; EG-NEXT:     SUB_INT * T11.W, PV.Y, PV.X,7669; EG-NEXT:    ALU clause starting at 952:7670; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T0.Z,7671; EG-NEXT:     SUB_INT T4.X, T8.W, T1.W,7672; EG-NEXT:     LSHL T1.Y, PV.W, 1,7673; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,7674; EG-NEXT:     CNDE_INT * T11.W, T3.Y, T1.Z, T11.W,7675; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)7676; EG-NEXT:     CNDE_INT * T9.W, T2.Y, T9.W, T6.Z,7677; EG-NEXT:     SUB_INT T10.X, T5.Z, T3.Z,7678; EG-NEXT:     LSHL T3.Y, PV.W, 1,7679; EG-NEXT:     BIT_ALIGN_INT T1.Z, T11.W, T7.W, literal.x,7680; EG-NEXT:     OR_INT T7.W, T1.Y, T0.Z, BS:VEC_021/SCL_1227681; EG-NEXT:     SUB_INT * T11.W, T4.X, T9.X,7682; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7683; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,7684; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS, BS:VEC_120/SCL_2127685; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1227686; EG-NEXT:     SETE_INT T8.W, PV.Z, T4.Z,7687; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T4.Z,7688; EG-NEXT:    25(3.503246e-44), 0(0.000000e+00)7689; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,7690; EG-NEXT:     SUB_INT T2.Y, T7.W, T2.Z,7691; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T9.W, literal.x,7692; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,7693; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,7694; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7695; EG-NEXT:     CNDE_INT T4.X, T8.X, T5.Z, PS,7696; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1227697; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2217698; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T1.W, BS:VEC_102/SCL_2217699; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,7700; EG-NEXT:     LSHL T6.X, PS, 1,7701; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7702; EG-NEXT:     SUB_INT T5.Z, T8.W, T0.W,7703; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x, BS:VEC_021/SCL_1227704; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,7705; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7706; EG-NEXT:     SUBB_UINT T4.X, T8.W, T0.W,7707; EG-NEXT:     SUB_INT * T2.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227708; EG-NEXT:     SETGE_UINT T6.Z, T10.W, T4.W,7709; EG-NEXT:     SETE_INT T12.W, T9.W, T3.Z, BS:VEC_2017710; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T3.Z,7711; EG-NEXT:     SUBB_UINT T5.X, T10.W, T4.W, BS:VEC_021/SCL_1227712; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7713; EG-NEXT:     SUB_INT T6.Z, T10.W, T4.W, BS:VEC_021/SCL_1227714; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,7715; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T5.Z,7716; EG-NEXT:     SUB_INT T4.X, T9.W, T3.Z,7717; EG-NEXT:     LSHL T2.Y, PS, 1,7718; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7719; EG-NEXT:     CNDE_INT T12.W, T1.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127720; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,7721; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)7722; EG-NEXT:     BFE_UINT T7.X, T3.X, literal.x, 1,7723; EG-NEXT:     LSHL T1.Y, PS, 1,7724; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T8.W, literal.y,7725; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7726; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,7727; EG-NEXT:    23(3.222986e-44), 31(4.344025e-44)7728; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,7729; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_021/SCL_1227730; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2217731; EG-NEXT:     SETE_INT T9.W, PV.Z, T1.W,7732; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T1.W,7733; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)7734; EG-NEXT:     SUBB_UINT T5.X, T7.W, T2.Z,7735; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7736; EG-NEXT:     SUB_INT * T5.Z, T8.W, T0.W, BS:VEC_120/SCL_2127737; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x,7738; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7739; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7740; EG-NEXT:     SUBB_UINT T4.X, T8.W, T0.W,7741; EG-NEXT:     SUB_INT * T1.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227742; EG-NEXT:     SETGE_UINT T6.Z, T9.W, T4.W,7743; EG-NEXT:     SETE_INT T10.W, T7.W, T3.Z, BS:VEC_2017744; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T3.Z,7745; EG-NEXT:     SUBB_UINT T8.X, T9.W, T4.W, BS:VEC_021/SCL_1227746; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,7747; EG-NEXT:     SUB_INT T6.Z, T9.W, T4.W, BS:VEC_021/SCL_1227748; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,7749; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T5.Z,7750; EG-NEXT:     SUB_INT T4.X, T7.W, T3.Z,7751; EG-NEXT:     LSHL T1.Y, PS, 1,7752; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7753; EG-NEXT:     CNDE_INT T10.W, T3.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127754; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7755; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)7756; EG-NEXT:     SUB_INT T10.X, T1.Z, T4.Z,7757; EG-NEXT:     LSHL T3.Y, PS, 1,7758; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T8.W, literal.x,7759; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7760; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,7761; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7762; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,7763; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS, BS:VEC_021/SCL_1227764; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2217765; EG-NEXT:     SETE_INT T7.W, PV.Z, T1.W,7766; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T1.W,7767; EG-NEXT:    23(3.222986e-44), 0(0.000000e+00)7768; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,7769; EG-NEXT:     SUB_INT T2.Y, T8.W, T0.W,7770; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1227771; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,7772; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,7773; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7774; EG-NEXT:     CNDE_INT T4.X, T9.X, T1.Z, PS,7775; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1227776; EG-NEXT:     SETE_INT T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1227777; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T3.Z, BS:VEC_021/SCL_1227778; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,7779; EG-NEXT:     LSHL T5.X, PS, 1,7780; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7781; EG-NEXT:     SUB_INT T1.Z, T7.W, T4.W,7782; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x, BS:VEC_021/SCL_1227783; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,7784; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7785; EG-NEXT:    ALU clause starting at 1067:7786; EG-NEXT:     SUBB_UINT T4.X, T7.W, T4.W,7787; EG-NEXT:     SUB_INT T2.Y, T5.Z, T3.Z, BS:VEC_021/SCL_1227788; EG-NEXT:     SETGE_UINT * T6.Z, T11.W, T2.Z, BS:VEC_2107789; EG-NEXT:     SETE_INT T12.W, T9.W, T4.Z,7790; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T4.Z,7791; EG-NEXT:     SUBB_UINT T6.X, T11.W, T2.Z,7792; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T6.Z,7793; EG-NEXT:     SUB_INT T6.Z, T11.W, T2.Z,7794; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,7795; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T1.Z,7796; EG-NEXT:     SUB_INT T4.X, T9.W, T4.Z,7797; EG-NEXT:     LSHL T2.Y, PS, 1,7798; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,7799; EG-NEXT:     CNDE_INT T12.W, T1.Y, T5.Z, PV.W, BS:VEC_021/SCL_1227800; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,7801; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)7802; EG-NEXT:     BFE_UINT T7.X, T0.Y, literal.x, 1,7803; EG-NEXT:     LSHL T1.Y, PS, 1,7804; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T7.W, literal.y,7805; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7806; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,7807; EG-NEXT:    22(3.082857e-44), 31(4.344025e-44)7808; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,7809; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS,7810; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1227811; EG-NEXT:     SETE_INT T9.W, PV.Z, T3.Z,7812; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T3.Z,7813; EG-NEXT:    22(3.082857e-44), 0(0.000000e+00)7814; EG-NEXT:     SUBB_UINT T6.X, T8.W, T0.W,7815; EG-NEXT:     CNDE_INT * T3.Y, PV.W, PS, PV.Z,7816; EG-NEXT:     SUB_INT T1.Z, T7.W, T4.W,7817; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x, BS:VEC_021/SCL_1227818; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7819; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7820; EG-NEXT:     SUBB_UINT T4.X, T7.W, T4.W,7821; EG-NEXT:     SUB_INT T1.Y, T5.Z, T3.Z,7822; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1227823; EG-NEXT:     SETE_INT T11.W, T8.W, T4.Z,7824; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T4.Z,7825; EG-NEXT:     SUBB_UINT T9.X, T9.W, T2.Z,7826; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,7827; EG-NEXT:     SUB_INT T6.Z, T9.W, T2.Z,7828; EG-NEXT:     SUB_INT T11.W, T1.Y, T4.X,7829; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T1.Z,7830; EG-NEXT:     SUB_INT T4.X, T8.W, T4.Z,7831; EG-NEXT:     LSHL T1.Y, PS, 1,7832; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,7833; EG-NEXT:     CNDE_INT T11.W, T3.Y, T5.Z, PV.W, BS:VEC_021/SCL_1227834; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7835; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)7836; EG-NEXT:     SUB_INT T10.X, T0.Z, T1.W,7837; EG-NEXT:     LSHL T3.Y, PS, 1,7838; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T7.W, literal.x, BS:VEC_021/SCL_1227839; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7840; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,7841; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7842; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,7843; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS,7844; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1227845; EG-NEXT:     SETE_INT T8.W, PV.Z, T3.Z,7846; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T3.Z,7847; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)7848; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,7849; EG-NEXT:     SUB_INT T2.Y, T7.W, T4.W,7850; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1227851; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,7852; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,7853; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7854; EG-NEXT:     CNDE_INT T4.X, T8.X, T0.Z, PS,7855; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T2.Z, BS:VEC_021/SCL_1227856; EG-NEXT:     SETE_INT T0.Z, PV.Z, T4.Z, BS:VEC_102/SCL_2217857; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T4.Z, BS:VEC_102/SCL_2217858; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,7859; EG-NEXT:     LSHL T6.X, PS, 1,7860; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7861; EG-NEXT:     SUB_INT T0.Z, T8.W, T2.Z,7862; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x,7863; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,7864; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7865; EG-NEXT:     SUBB_UINT T4.X, T8.W, T2.Z,7866; EG-NEXT:     SUB_INT T2.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1227867; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1227868; EG-NEXT:     SETE_INT T12.W, PV.W, T1.W,7869; EG-NEXT:     SETGE_UINT * T13.W, PV.W, T1.W,7870; EG-NEXT:     SUBB_UINT T5.X, T10.W, T0.W, BS:VEC_021/SCL_1227871; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7872; EG-NEXT:     SUB_INT T6.Z, T10.W, T0.W, BS:VEC_021/SCL_1227873; EG-NEXT:     SUB_INT T12.W, PV.Y, PV.X,7874; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T0.Z,7875; EG-NEXT:     SUB_INT T4.X, T9.W, T1.W, BS:VEC_021/SCL_1227876; EG-NEXT:     LSHL T2.Y, PS, 1,7877; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,7878; EG-NEXT:     CNDE_INT T12.W, T1.Y, T1.Z, PV.W,7879; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,7880; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)7881; EG-NEXT:     BFE_UINT T7.X, T2.X, literal.x, 1,7882; EG-NEXT:     LSHL T1.Y, PS, 1,7883; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.y,7884; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7885; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,7886; EG-NEXT:    20(2.802597e-44), 31(4.344025e-44)7887; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,7888; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_120/SCL_2127889; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1227890; EG-NEXT:     SETE_INT T9.W, PV.Z, T4.Z,7891; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T4.Z,7892; EG-NEXT:    21(2.942727e-44), 0(0.000000e+00)7893; EG-NEXT:     SUBB_UINT T5.X, T7.W, T4.W,7894; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7895; EG-NEXT:     SUB_INT * T0.Z, T8.W, T2.Z, BS:VEC_2017896; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x,7897; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7898; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7899; EG-NEXT:     SUBB_UINT * T4.X, T8.W, T2.Z,7900; EG-NEXT:    ALU clause starting at 1181:7901; EG-NEXT:     SUB_INT T1.Y, T1.Z, T4.Z,7902; EG-NEXT:     SETGE_UINT * T6.Z, T9.W, T0.W,7903; EG-NEXT:     SETE_INT T10.W, T7.W, T1.W,7904; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T1.W,7905; EG-NEXT:     SUBB_UINT T8.X, T9.W, T0.W, BS:VEC_021/SCL_1227906; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,7907; EG-NEXT:     SUB_INT T6.Z, T9.W, T0.W, BS:VEC_021/SCL_1227908; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,7909; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,7910; EG-NEXT:     SUB_INT T4.X, T7.W, T1.W, BS:VEC_021/SCL_1227911; EG-NEXT:     LSHL T1.Y, PS, 1,7912; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,7913; EG-NEXT:     CNDE_INT T10.W, T3.Y, T1.Z, PV.W,7914; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7915; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)7916; EG-NEXT:     SUB_INT T10.X, T5.Z, T3.Z,7917; EG-NEXT:     LSHL T3.Y, PS, 1,7918; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x,7919; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,7920; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,7921; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7922; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,7923; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS, BS:VEC_120/SCL_2127924; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1227925; EG-NEXT:     SETE_INT T7.W, PV.Z, T4.Z,7926; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T4.Z,7927; EG-NEXT:    20(2.802597e-44), 0(0.000000e+00)7928; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,7929; EG-NEXT:     SUB_INT T2.Y, T8.W, T2.Z,7930; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T9.W, literal.x,7931; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,7932; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,7933; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7934; EG-NEXT:     CNDE_INT T4.X, T9.X, T5.Z, PS,7935; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1227936; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2217937; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T1.W, BS:VEC_102/SCL_2217938; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,7939; EG-NEXT:     LSHL T5.X, PS, 1,7940; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,7941; EG-NEXT:     SUB_INT T5.Z, T7.W, T0.W,7942; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x, BS:VEC_021/SCL_1227943; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,7944; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7945; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,7946; EG-NEXT:     SUB_INT * T2.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227947; EG-NEXT:     SETGE_UINT T6.Z, T11.W, T4.W,7948; EG-NEXT:     SETE_INT T12.W, T9.W, T3.Z, BS:VEC_2017949; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T3.Z,7950; EG-NEXT:     SUBB_UINT T6.X, T11.W, T4.W, BS:VEC_021/SCL_1227951; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7952; EG-NEXT:     SUB_INT T6.Z, T11.W, T4.W, BS:VEC_021/SCL_1227953; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,7954; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T5.Z,7955; EG-NEXT:     SUB_INT T4.X, T9.W, T3.Z,7956; EG-NEXT:     LSHL T2.Y, PS, 1,7957; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7958; EG-NEXT:     CNDE_INT T12.W, T1.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127959; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,7960; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)7961; EG-NEXT:     BFE_UINT T7.X, T3.X, literal.x, 1,7962; EG-NEXT:     LSHL T1.Y, PS, 1,7963; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.y,7964; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7965; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,7966; EG-NEXT:    18(2.522337e-44), 31(4.344025e-44)7967; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,7968; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_021/SCL_1227969; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2217970; EG-NEXT:     SETE_INT T9.W, PV.Z, T1.W,7971; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T1.W,7972; EG-NEXT:    19(2.662467e-44), 0(0.000000e+00)7973; EG-NEXT:     SUBB_UINT T6.X, T8.W, T2.Z,7974; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,7975; EG-NEXT:     SUB_INT * T5.Z, T7.W, T0.W, BS:VEC_120/SCL_2127976; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,7977; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,7978; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)7979; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,7980; EG-NEXT:     SUB_INT * T1.Y, T0.Z, T1.W, BS:VEC_021/SCL_1227981; EG-NEXT:     SETGE_UINT T6.Z, T9.W, T4.W,7982; EG-NEXT:     SETE_INT T11.W, T8.W, T3.Z, BS:VEC_2017983; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T3.Z,7984; EG-NEXT:     SUBB_UINT T9.X, T9.W, T4.W, BS:VEC_021/SCL_1227985; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,7986; EG-NEXT:     SUB_INT T6.Z, T9.W, T4.W, BS:VEC_021/SCL_1227987; EG-NEXT:     SUB_INT T11.W, T1.Y, T4.X,7988; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T5.Z,7989; EG-NEXT:     SUB_INT T4.X, T8.W, T3.Z,7990; EG-NEXT:     LSHL T1.Y, PS, 1,7991; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,7992; EG-NEXT:     CNDE_INT T11.W, T3.Y, T0.Z, PV.W, BS:VEC_120/SCL_2127993; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,7994; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)7995; EG-NEXT:     SUB_INT T10.X, T1.Z, T4.Z,7996; EG-NEXT:     LSHL T3.Y, PS, 1,7997; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.x,7998; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,7999; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,8000; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8001; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,8002; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS, BS:VEC_021/SCL_1228003; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2218004; EG-NEXT:     SETE_INT T8.W, PV.Z, T1.W,8005; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T1.W,8006; EG-NEXT:    18(2.522337e-44), 0(0.000000e+00)8007; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,8008; EG-NEXT:     SUB_INT T2.Y, T7.W, T0.W,8009; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1228010; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,8011; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,8012; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8013; EG-NEXT:     CNDE_INT T4.X, T8.X, T1.Z, PS,8014; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W,8015; EG-NEXT:     SETE_INT * T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1228016; EG-NEXT:    ALU clause starting at 1296:8017; EG-NEXT:     SETGE_UINT T9.W, T5.Z, T3.Z,8018; EG-NEXT:     CNDE_INT * T11.W, T9.X, T7.W, T2.Y,8019; EG-NEXT:     LSHL T6.X, PS, 1,8020; EG-NEXT:     CNDE_INT T1.Y, T1.Z, PV.W, T1.Y,8021; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,8022; EG-NEXT:     BIT_ALIGN_INT T9.W, T4.X, T10.W, literal.x, BS:VEC_021/SCL_1228023; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,8024; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8025; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,8026; EG-NEXT:     SUB_INT T2.Y, T5.Z, T3.Z,8027; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228028; EG-NEXT:     SETE_INT T12.W, T9.W, T4.Z,8029; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T4.Z,8030; EG-NEXT:     SUBB_UINT T5.X, T10.W, T2.Z,8031; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T6.Z,8032; EG-NEXT:     SUB_INT T6.Z, T10.W, T2.Z,8033; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,8034; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T1.Z,8035; EG-NEXT:     SUB_INT T4.X, T9.W, T4.Z,8036; EG-NEXT:     LSHL T2.Y, PS, 1,8037; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8038; EG-NEXT:     CNDE_INT T12.W, T1.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228039; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,8040; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)8041; EG-NEXT:     BFE_UINT T7.X, T0.Y, literal.x, 1,8042; EG-NEXT:     LSHL T1.Y, PS, 1,8043; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.y,8044; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8045; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,8046; EG-NEXT:    17(2.382207e-44), 31(4.344025e-44)8047; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8048; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS,8049; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228050; EG-NEXT:     SETE_INT T9.W, PV.Z, T3.Z,8051; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T3.Z,8052; EG-NEXT:    17(2.382207e-44), 0(0.000000e+00)8053; EG-NEXT:     SUBB_UINT T5.X, T7.W, T0.W,8054; EG-NEXT:     CNDE_INT * T3.Y, PV.W, PS, PV.Z,8055; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,8056; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x, BS:VEC_021/SCL_1228057; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8058; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8059; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,8060; EG-NEXT:     SUB_INT T1.Y, T5.Z, T3.Z,8061; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228062; EG-NEXT:     SETE_INT T10.W, T7.W, T4.Z,8063; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T4.Z,8064; EG-NEXT:     SUBB_UINT T8.X, T9.W, T2.Z,8065; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,8066; EG-NEXT:     SUB_INT T6.Z, T9.W, T2.Z,8067; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,8068; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T1.Z,8069; EG-NEXT:     SUB_INT T4.X, T7.W, T4.Z,8070; EG-NEXT:     LSHL T1.Y, PS, 1,8071; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8072; EG-NEXT:     CNDE_INT T10.W, T3.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228073; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8074; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)8075; EG-NEXT:     SUB_INT T10.X, T0.Z, T1.W,8076; EG-NEXT:     LSHL T3.Y, PS, 1,8077; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1228078; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8079; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,8080; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8081; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8082; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS,8083; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228084; EG-NEXT:     SETE_INT T7.W, PV.Z, T3.Z,8085; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T3.Z,8086; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)8087; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,8088; EG-NEXT:     SUB_INT T2.Y, T8.W, T4.W,8089; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1228090; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,8091; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,8092; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8093; EG-NEXT:     CNDE_INT T4.X, T9.X, T0.Z, PS,8094; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T2.Z, BS:VEC_021/SCL_1228095; EG-NEXT:     SETE_INT T0.Z, PV.Z, T4.Z, BS:VEC_102/SCL_2218096; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T4.Z, BS:VEC_102/SCL_2218097; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,8098; EG-NEXT:     LSHL T5.X, PS, 1,8099; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8100; EG-NEXT:     SUB_INT T0.Z, T7.W, T2.Z,8101; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x,8102; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,8103; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8104; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,8105; EG-NEXT:     SUB_INT T2.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228106; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228107; EG-NEXT:     SETE_INT T12.W, PV.W, T1.W,8108; EG-NEXT:     SETGE_UINT * T13.W, PV.W, T1.W,8109; EG-NEXT:     SUBB_UINT T6.X, T11.W, T0.W, BS:VEC_021/SCL_1228110; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8111; EG-NEXT:     SUB_INT T6.Z, T11.W, T0.W, BS:VEC_021/SCL_1228112; EG-NEXT:     SUB_INT T12.W, PV.Y, PV.X,8113; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T0.Z,8114; EG-NEXT:     SUB_INT T4.X, T9.W, T1.W, BS:VEC_021/SCL_1228115; EG-NEXT:     LSHL T2.Y, PS, 1,8116; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,8117; EG-NEXT:     CNDE_INT T12.W, T1.Y, T1.Z, PV.W,8118; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,8119; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)8120; EG-NEXT:     BFE_UINT T7.X, T2.X, literal.x, 1,8121; EG-NEXT:     LSHL T1.Y, PS, 1,8122; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T7.W, literal.y,8123; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8124; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,8125; EG-NEXT:    15(2.101948e-44), 31(4.344025e-44)8126; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8127; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_120/SCL_2128128; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z,8129; EG-NEXT:     SETE_INT * T9.W, PV.Z, T4.Z, BS:VEC_021/SCL_1228130; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)8131; EG-NEXT:    ALU clause starting at 1410:8132; EG-NEXT:     SETGE_UINT * T12.W, T1.Z, T4.Z,8133; EG-NEXT:     SUBB_UINT T6.X, T8.W, T4.W,8134; EG-NEXT:     CNDE_INT * T3.Y, T9.W, PV.W, T0.Z, BS:VEC_2018135; EG-NEXT:     SUB_INT T0.Z, T7.W, T2.Z,8136; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,8137; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8138; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8139; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,8140; EG-NEXT:     SUB_INT T1.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228141; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228142; EG-NEXT:     SETE_INT T11.W, PV.W, T1.W,8143; EG-NEXT:     SETGE_UINT * T12.W, PV.W, T1.W,8144; EG-NEXT:     SUBB_UINT T9.X, T9.W, T0.W, BS:VEC_021/SCL_1228145; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,8146; EG-NEXT:     SUB_INT T6.Z, T9.W, T0.W, BS:VEC_021/SCL_1228147; EG-NEXT:     SUB_INT T11.W, PV.Y, PV.X,8148; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T0.Z,8149; EG-NEXT:     SUB_INT T4.X, T8.W, T1.W, BS:VEC_021/SCL_1228150; EG-NEXT:     LSHL T1.Y, PS, 1,8151; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,8152; EG-NEXT:     CNDE_INT T11.W, T3.Y, T1.Z, PV.W,8153; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8154; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)8155; EG-NEXT:     SUB_INT T10.X, T5.Z, T3.Z,8156; EG-NEXT:     LSHL T3.Y, PS, 1,8157; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T7.W, literal.x,8158; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8159; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,8160; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8161; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8162; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS, BS:VEC_120/SCL_2128163; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1228164; EG-NEXT:     SETE_INT T8.W, PV.Z, T4.Z,8165; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T4.Z,8166; EG-NEXT:    15(2.101948e-44), 0(0.000000e+00)8167; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,8168; EG-NEXT:     SUB_INT T2.Y, T7.W, T2.Z,8169; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T9.W, literal.x,8170; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,8171; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,8172; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8173; EG-NEXT:     CNDE_INT T4.X, T8.X, T5.Z, PS,8174; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1228175; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2218176; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T1.W, BS:VEC_102/SCL_2218177; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,8178; EG-NEXT:     LSHL T6.X, PS, 1,8179; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8180; EG-NEXT:     SUB_INT T5.Z, T8.W, T0.W,8181; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x, BS:VEC_021/SCL_1228182; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,8183; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8184; EG-NEXT:     SUBB_UINT T4.X, T8.W, T0.W,8185; EG-NEXT:     SUB_INT * T2.Y, T0.Z, T1.W, BS:VEC_021/SCL_1228186; EG-NEXT:     SETGE_UINT T6.Z, T10.W, T4.W,8187; EG-NEXT:     SETE_INT T12.W, T9.W, T3.Z, BS:VEC_2018188; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T3.Z,8189; EG-NEXT:     SUBB_UINT T5.X, T10.W, T4.W, BS:VEC_021/SCL_1228190; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8191; EG-NEXT:     SUB_INT T6.Z, T10.W, T4.W, BS:VEC_021/SCL_1228192; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,8193; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T5.Z,8194; EG-NEXT:     SUB_INT T4.X, T9.W, T3.Z,8195; EG-NEXT:     LSHL T2.Y, PS, 1,8196; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,8197; EG-NEXT:     CNDE_INT T12.W, T1.Y, T0.Z, PV.W, BS:VEC_120/SCL_2128198; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,8199; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)8200; EG-NEXT:     BFE_UINT T7.X, T3.X, literal.x, 1,8201; EG-NEXT:     LSHL T1.Y, PS, 1,8202; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T8.W, literal.y,8203; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8204; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,8205; EG-NEXT:    13(1.821688e-44), 31(4.344025e-44)8206; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,8207; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_021/SCL_1228208; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2218209; EG-NEXT:     SETE_INT T9.W, PV.Z, T1.W,8210; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T1.W,8211; EG-NEXT:    14(1.961818e-44), 0(0.000000e+00)8212; EG-NEXT:     SUBB_UINT T5.X, T7.W, T2.Z,8213; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8214; EG-NEXT:     SUB_INT * T5.Z, T8.W, T0.W, BS:VEC_120/SCL_2128215; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x,8216; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8217; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8218; EG-NEXT:     SUBB_UINT T4.X, T8.W, T0.W,8219; EG-NEXT:     SUB_INT * T1.Y, T0.Z, T1.W, BS:VEC_021/SCL_1228220; EG-NEXT:     SETGE_UINT T6.Z, T9.W, T4.W,8221; EG-NEXT:     SETE_INT T10.W, T7.W, T3.Z, BS:VEC_2018222; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T3.Z,8223; EG-NEXT:     SUBB_UINT T8.X, T9.W, T4.W, BS:VEC_021/SCL_1228224; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,8225; EG-NEXT:     SUB_INT T6.Z, T9.W, T4.W, BS:VEC_021/SCL_1228226; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,8227; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T5.Z,8228; EG-NEXT:     SUB_INT T4.X, T7.W, T3.Z,8229; EG-NEXT:     LSHL T1.Y, PS, 1,8230; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,8231; EG-NEXT:     CNDE_INT T10.W, T3.Y, T0.Z, PV.W, BS:VEC_120/SCL_2128232; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8233; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)8234; EG-NEXT:     SUB_INT T10.X, T1.Z, T4.Z,8235; EG-NEXT:     LSHL T3.Y, PS, 1,8236; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T8.W, literal.x,8237; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8238; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,8239; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8240; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,8241; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS, BS:VEC_021/SCL_1228242; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2218243; EG-NEXT:     SETE_INT T7.W, PV.Z, T1.W,8244; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T1.W,8245; EG-NEXT:    13(1.821688e-44), 0(0.000000e+00)8246; EG-NEXT:     CNDE_INT * T8.X, PV.W, PS, PV.Z,8247; EG-NEXT:    ALU clause starting at 1525:8248; EG-NEXT:     SUB_INT T2.Y, T8.W, T0.W,8249; EG-NEXT:     BIT_ALIGN_INT T5.Z, T1.Y, T9.W, literal.x, BS:VEC_021/SCL_1228250; EG-NEXT:     OR_INT T7.W, T3.Y, T4.X, BS:VEC_102/SCL_2218251; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,8252; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8253; EG-NEXT:     CNDE_INT T4.X, T9.X, T1.Z, PS,8254; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1228255; EG-NEXT:     SETE_INT T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1228256; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T3.Z, BS:VEC_021/SCL_1228257; EG-NEXT:     CNDE_INT * T10.W, T8.X, T8.W, PV.Y,8258; EG-NEXT:     LSHL T5.X, PS, 1,8259; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8260; EG-NEXT:     SUB_INT T1.Z, T7.W, T4.W,8261; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x, BS:VEC_021/SCL_1228262; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,8263; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8264; EG-NEXT:     SUBB_UINT T4.X, T7.W, T4.W,8265; EG-NEXT:     SUB_INT T2.Y, T5.Z, T3.Z,8266; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228267; EG-NEXT:     SETE_INT T12.W, T9.W, T4.Z,8268; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T4.Z,8269; EG-NEXT:     SUBB_UINT T6.X, T11.W, T2.Z,8270; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T6.Z,8271; EG-NEXT:     SUB_INT T6.Z, T11.W, T2.Z,8272; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,8273; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T1.Z,8274; EG-NEXT:     SUB_INT T4.X, T9.W, T4.Z,8275; EG-NEXT:     LSHL T2.Y, PS, 1,8276; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8277; EG-NEXT:     CNDE_INT T12.W, T1.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228278; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,8279; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)8280; EG-NEXT:     BFE_UINT T7.X, T0.Y, literal.x, 1,8281; EG-NEXT:     LSHL T1.Y, PS, 1,8282; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T7.W, literal.y,8283; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8284; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,8285; EG-NEXT:    12(1.681558e-44), 31(4.344025e-44)8286; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8287; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS,8288; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228289; EG-NEXT:     SETE_INT T9.W, PV.Z, T3.Z,8290; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T3.Z,8291; EG-NEXT:    12(1.681558e-44), 0(0.000000e+00)8292; EG-NEXT:     SUBB_UINT T6.X, T8.W, T0.W,8293; EG-NEXT:     CNDE_INT * T3.Y, PV.W, PS, PV.Z,8294; EG-NEXT:     SUB_INT T1.Z, T7.W, T4.W,8295; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x, BS:VEC_021/SCL_1228296; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8297; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8298; EG-NEXT:     SUBB_UINT T4.X, T7.W, T4.W,8299; EG-NEXT:     SUB_INT T1.Y, T5.Z, T3.Z,8300; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228301; EG-NEXT:     SETE_INT T11.W, T8.W, T4.Z,8302; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T4.Z,8303; EG-NEXT:     SUBB_UINT T9.X, T9.W, T2.Z,8304; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,8305; EG-NEXT:     SUB_INT T6.Z, T9.W, T2.Z,8306; EG-NEXT:     SUB_INT T11.W, T1.Y, T4.X,8307; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T1.Z,8308; EG-NEXT:     SUB_INT T4.X, T8.W, T4.Z,8309; EG-NEXT:     LSHL T1.Y, PS, 1,8310; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8311; EG-NEXT:     CNDE_INT T11.W, T3.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228312; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8313; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)8314; EG-NEXT:     SUB_INT T10.X, T0.Z, T1.W,8315; EG-NEXT:     LSHL T3.Y, PS, 1,8316; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T7.W, literal.x, BS:VEC_021/SCL_1228317; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8318; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,8319; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8320; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8321; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS,8322; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228323; EG-NEXT:     SETE_INT T8.W, PV.Z, T3.Z,8324; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T3.Z,8325; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)8326; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,8327; EG-NEXT:     SUB_INT T2.Y, T7.W, T4.W,8328; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1228329; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,8330; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,8331; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8332; EG-NEXT:     CNDE_INT T4.X, T8.X, T0.Z, PS,8333; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T2.Z, BS:VEC_021/SCL_1228334; EG-NEXT:     SETE_INT T0.Z, PV.Z, T4.Z, BS:VEC_102/SCL_2218335; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T4.Z, BS:VEC_102/SCL_2218336; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,8337; EG-NEXT:     LSHL T6.X, PS, 1,8338; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8339; EG-NEXT:     SUB_INT T0.Z, T8.W, T2.Z,8340; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x,8341; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,8342; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8343; EG-NEXT:     SUBB_UINT T4.X, T8.W, T2.Z,8344; EG-NEXT:     SUB_INT T2.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228345; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228346; EG-NEXT:     SETE_INT T12.W, PV.W, T1.W,8347; EG-NEXT:     SETGE_UINT * T13.W, PV.W, T1.W,8348; EG-NEXT:     SUBB_UINT T5.X, T10.W, T0.W, BS:VEC_021/SCL_1228349; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8350; EG-NEXT:     SUB_INT T6.Z, T10.W, T0.W, BS:VEC_021/SCL_1228351; EG-NEXT:     SUB_INT T12.W, PV.Y, PV.X,8352; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T0.Z,8353; EG-NEXT:     SUB_INT T4.X, T9.W, T1.W, BS:VEC_021/SCL_1228354; EG-NEXT:     LSHL T2.Y, PS, 1,8355; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,8356; EG-NEXT:     CNDE_INT T12.W, T1.Y, T1.Z, PV.W,8357; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,8358; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)8359; EG-NEXT:     BFE_UINT T7.X, T2.X, literal.x, 1,8360; EG-NEXT:     LSHL T1.Y, PS, 1,8361; EG-NEXT:     BIT_ALIGN_INT * T1.Z, PV.W, T8.W, literal.y,8362; EG-NEXT:    10(1.401298e-44), 31(4.344025e-44)8363; EG-NEXT:    ALU clause starting at 1640:8364; EG-NEXT:     OR_INT T8.W, T2.Y, T0.Z,8365; EG-NEXT:     SUB_INT * T12.W, T4.X, T5.X,8366; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8367; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_120/SCL_2128368; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_102/SCL_2218369; EG-NEXT:     SETE_INT T9.W, T1.Z, T4.Z, BS:VEC_2108370; EG-NEXT:     SETGE_UINT * T12.W, T1.Z, T4.Z,8371; EG-NEXT:    11(1.541428e-44), 0(0.000000e+00)8372; EG-NEXT:     SUBB_UINT T5.X, T7.W, T4.W,8373; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8374; EG-NEXT:     SUB_INT * T0.Z, T8.W, T2.Z, BS:VEC_2018375; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x,8376; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8377; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8378; EG-NEXT:     SUBB_UINT T4.X, T8.W, T2.Z,8379; EG-NEXT:     SUB_INT T1.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228380; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228381; EG-NEXT:     SETE_INT T10.W, PV.W, T1.W,8382; EG-NEXT:     SETGE_UINT * T12.W, PV.W, T1.W,8383; EG-NEXT:     SUBB_UINT T8.X, T9.W, T0.W, BS:VEC_021/SCL_1228384; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,8385; EG-NEXT:     SUB_INT T6.Z, T9.W, T0.W, BS:VEC_021/SCL_1228386; EG-NEXT:     SUB_INT T10.W, PV.Y, PV.X,8387; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T0.Z,8388; EG-NEXT:     SUB_INT T4.X, T7.W, T1.W, BS:VEC_021/SCL_1228389; EG-NEXT:     LSHL T1.Y, PS, 1,8390; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,8391; EG-NEXT:     CNDE_INT T10.W, T3.Y, T1.Z, PV.W,8392; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8393; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)8394; EG-NEXT:     SUB_INT T10.X, T5.Z, T3.Z,8395; EG-NEXT:     LSHL T3.Y, PS, 1,8396; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T8.W, literal.x,8397; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8398; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,8399; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8400; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8401; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS, BS:VEC_120/SCL_2128402; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1228403; EG-NEXT:     SETE_INT T7.W, PV.Z, T4.Z,8404; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T4.Z,8405; EG-NEXT:    10(1.401298e-44), 0(0.000000e+00)8406; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,8407; EG-NEXT:     SUB_INT T2.Y, T8.W, T2.Z,8408; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T9.W, literal.x,8409; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,8410; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,8411; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8412; EG-NEXT:     CNDE_INT T4.X, T9.X, T5.Z, PS,8413; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1228414; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2218415; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T1.W, BS:VEC_102/SCL_2218416; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,8417; EG-NEXT:     LSHL T5.X, PS, 1,8418; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8419; EG-NEXT:     SUB_INT T5.Z, T7.W, T0.W,8420; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x, BS:VEC_021/SCL_1228421; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,8422; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8423; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,8424; EG-NEXT:     SUB_INT * T2.Y, T0.Z, T1.W, BS:VEC_021/SCL_1228425; EG-NEXT:     SETGE_UINT T6.Z, T11.W, T4.W,8426; EG-NEXT:     SETE_INT T12.W, T9.W, T3.Z, BS:VEC_2018427; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T3.Z,8428; EG-NEXT:     SUBB_UINT T6.X, T11.W, T4.W, BS:VEC_021/SCL_1228429; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8430; EG-NEXT:     SUB_INT T6.Z, T11.W, T4.W, BS:VEC_021/SCL_1228431; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,8432; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T5.Z,8433; EG-NEXT:     SUB_INT T4.X, T9.W, T3.Z,8434; EG-NEXT:     LSHL T2.Y, PS, 1,8435; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,8436; EG-NEXT:     CNDE_INT T12.W, T1.Y, T0.Z, PV.W, BS:VEC_120/SCL_2128437; EG-NEXT:     CNDE_INT * T11.W, PV.Y, T11.W, PV.Z,8438; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)8439; EG-NEXT:     BFE_UINT T7.X, T3.X, literal.x, 1,8440; EG-NEXT:     LSHL T1.Y, PS, 1,8441; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.y,8442; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8443; EG-NEXT:     SUB_INT * T12.W, PV.X, T6.X,8444; EG-NEXT:    8(1.121039e-44), 31(4.344025e-44)8445; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,8446; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_021/SCL_1228447; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T0.W, BS:VEC_102/SCL_2218448; EG-NEXT:     SETE_INT T9.W, PV.Z, T1.W,8449; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T1.W,8450; EG-NEXT:    9(1.261169e-44), 0(0.000000e+00)8451; EG-NEXT:     SUBB_UINT T6.X, T8.W, T2.Z,8452; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8453; EG-NEXT:     SUB_INT * T5.Z, T7.W, T0.W, BS:VEC_120/SCL_2128454; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,8455; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8456; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8457; EG-NEXT:     SUBB_UINT T4.X, T7.W, T0.W,8458; EG-NEXT:     SUB_INT * T1.Y, T0.Z, T1.W, BS:VEC_021/SCL_1228459; EG-NEXT:     SETGE_UINT T6.Z, T9.W, T4.W,8460; EG-NEXT:     SETE_INT T11.W, T8.W, T3.Z, BS:VEC_2018461; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T3.Z,8462; EG-NEXT:     SUBB_UINT T9.X, T9.W, T4.W, BS:VEC_021/SCL_1228463; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,8464; EG-NEXT:     SUB_INT T6.Z, T9.W, T4.W, BS:VEC_021/SCL_1228465; EG-NEXT:     SUB_INT T11.W, T1.Y, T4.X,8466; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T5.Z,8467; EG-NEXT:     SUB_INT T4.X, T8.W, T3.Z,8468; EG-NEXT:     LSHL T1.Y, PS, 1,8469; EG-NEXT:     BFE_UINT T5.Z, T0.Y, literal.x, 1,8470; EG-NEXT:     CNDE_INT T11.W, T3.Y, T0.Z, PV.W, BS:VEC_120/SCL_2128471; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8472; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)8473; EG-NEXT:     SUB_INT T10.X, T1.Z, T4.Z,8474; EG-NEXT:     LSHL T3.Y, PS, 1,8475; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T7.W, literal.x,8476; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8477; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,8478; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8479; EG-NEXT:    ALU clause starting at 1755:8480; EG-NEXT:     BFE_UINT T4.X, T2.X, literal.x, 1,8481; EG-NEXT:     CNDE_INT * T1.Y, T2.Y, T8.W, T11.W,8482; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)8483; EG-NEXT:     SETGE_UINT T5.Z, T7.W, T0.W, BS:VEC_021/SCL_1228484; EG-NEXT:     SETE_INT T8.W, T0.Z, T1.W,8485; EG-NEXT:     SETGE_UINT * T11.W, T0.Z, T1.W,8486; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,8487; EG-NEXT:     SUB_INT T2.Y, T7.W, T0.W,8488; EG-NEXT:     BIT_ALIGN_INT T5.Z, T1.Y, T9.W, literal.x, BS:VEC_021/SCL_1228489; EG-NEXT:     OR_INT T8.W, T3.Y, T4.X, BS:VEC_102/SCL_2218490; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,8491; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8492; EG-NEXT:     CNDE_INT T4.X, T8.X, T1.Z, PS,8493; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T4.W, BS:VEC_021/SCL_1228494; EG-NEXT:     SETE_INT T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1228495; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T3.Z, BS:VEC_021/SCL_1228496; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,8497; EG-NEXT:     LSHL T6.X, PS, 1,8498; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8499; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,8500; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x, BS:VEC_021/SCL_1228501; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,8502; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8503; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,8504; EG-NEXT:     SUB_INT T2.Y, T5.Z, T3.Z,8505; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228506; EG-NEXT:     SETE_INT T12.W, T9.W, T4.Z,8507; EG-NEXT:     SETGE_UINT * T13.W, T9.W, T4.Z,8508; EG-NEXT:     SUBB_UINT T5.X, T10.W, T2.Z,8509; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T6.Z,8510; EG-NEXT:     SUB_INT T6.Z, T10.W, T2.Z,8511; EG-NEXT:     SUB_INT T12.W, T2.Y, T4.X,8512; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T1.Z,8513; EG-NEXT:     SUB_INT T4.X, T9.W, T4.Z,8514; EG-NEXT:     LSHL T2.Y, PS, 1,8515; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8516; EG-NEXT:     CNDE_INT T12.W, T1.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228517; EG-NEXT:     CNDE_INT * T10.W, PV.Y, T10.W, PV.Z,8518; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)8519; EG-NEXT:     BFE_UINT T7.X, T0.Y, literal.x, 1,8520; EG-NEXT:     LSHL T1.Y, PS, 1,8521; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.y,8522; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8523; EG-NEXT:     SUB_INT * T12.W, PV.X, T5.X,8524; EG-NEXT:    7(9.809089e-45), 31(4.344025e-44)8525; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8526; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS,8527; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228528; EG-NEXT:     SETE_INT T9.W, PV.Z, T3.Z,8529; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T3.Z,8530; EG-NEXT:    7(9.809089e-45), 0(0.000000e+00)8531; EG-NEXT:     SUBB_UINT T5.X, T7.W, T0.W,8532; EG-NEXT:     CNDE_INT * T3.Y, PV.W, PS, PV.Z,8533; EG-NEXT:     SUB_INT T1.Z, T8.W, T4.W,8534; EG-NEXT:     BIT_ALIGN_INT T7.W, T2.Y, T10.W, literal.x, BS:VEC_021/SCL_1228535; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8536; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8537; EG-NEXT:     SUBB_UINT T4.X, T8.W, T4.W,8538; EG-NEXT:     SUB_INT T1.Y, T5.Z, T3.Z,8539; EG-NEXT:     SETGE_UINT * T6.Z, PS, T2.Z, BS:VEC_021/SCL_1228540; EG-NEXT:     SETE_INT T10.W, T7.W, T4.Z,8541; EG-NEXT:     SETGE_UINT * T12.W, T7.W, T4.Z,8542; EG-NEXT:     SUBB_UINT T8.X, T9.W, T2.Z,8543; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, T6.Z,8544; EG-NEXT:     SUB_INT T6.Z, T9.W, T2.Z,8545; EG-NEXT:     SUB_INT T10.W, T1.Y, T4.X,8546; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T1.Z,8547; EG-NEXT:     SUB_INT T4.X, T7.W, T4.Z,8548; EG-NEXT:     LSHL T1.Y, PS, 1,8549; EG-NEXT:     BFE_UINT T1.Z, T2.X, literal.x, 1,8550; EG-NEXT:     CNDE_INT T10.W, T3.Y, T5.Z, PV.W, BS:VEC_021/SCL_1228551; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8552; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)8553; EG-NEXT:     SUB_INT T10.X, T0.Z, T1.W,8554; EG-NEXT:     LSHL T3.Y, PS, 1,8555; EG-NEXT:     BIT_ALIGN_INT T5.Z, PV.W, T8.W, literal.x, BS:VEC_021/SCL_1228556; EG-NEXT:     OR_INT T8.W, PV.Y, PV.Z,8557; EG-NEXT:     SUB_INT * T10.W, PV.X, T8.X,8558; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8559; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8560; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T7.W, PS,8561; EG-NEXT:     SETGE_UINT T1.Z, PV.W, T4.W, BS:VEC_021/SCL_1228562; EG-NEXT:     SETE_INT T7.W, PV.Z, T3.Z,8563; EG-NEXT:     SETGE_UINT * T10.W, PV.Z, T3.Z,8564; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)8565; EG-NEXT:     CNDE_INT T8.X, PV.W, PS, PV.Z,8566; EG-NEXT:     SUB_INT T2.Y, T8.W, T4.W,8567; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T9.W, literal.x, BS:VEC_021/SCL_1228568; EG-NEXT:     OR_INT T7.W, T3.Y, PV.X,8569; EG-NEXT:     SUB_INT * T9.W, T10.X, T5.X,8570; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8571; EG-NEXT:     CNDE_INT T4.X, T9.X, T0.Z, PS,8572; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T2.Z, BS:VEC_021/SCL_1228573; EG-NEXT:     SETE_INT T0.Z, PV.Z, T4.Z, BS:VEC_102/SCL_2218574; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T4.Z, BS:VEC_102/SCL_2218575; EG-NEXT:     CNDE_INT * T10.W, PV.X, T8.W, PV.Y,8576; EG-NEXT:     LSHL T5.X, PS, 1,8577; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8578; EG-NEXT:     SUB_INT T0.Z, T7.W, T2.Z,8579; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T11.W, literal.x,8580; EG-NEXT:     OR_INT * T11.W, T6.X, T7.X,8581; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8582; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,8583; EG-NEXT:     SUB_INT T2.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228584; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228585; EG-NEXT:     SETE_INT T12.W, PV.W, T1.W,8586; EG-NEXT:     SETGE_UINT * T13.W, PV.W, T1.W,8587; EG-NEXT:     SUBB_UINT T6.X, T11.W, T0.W, BS:VEC_021/SCL_1228588; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8589; EG-NEXT:     SUB_INT T6.Z, T11.W, T0.W, BS:VEC_021/SCL_1228590; EG-NEXT:     SUB_INT T12.W, PV.Y, PV.X,8591; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, T0.Z,8592; EG-NEXT:     SUB_INT T4.X, T9.W, T1.W,8593; EG-NEXT:     LSHL T2.Y, PS, 1,8594; EG-NEXT:     BFE_UINT * T0.Z, T3.X, literal.x, 1,8595; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)8596; EG-NEXT:    ALU clause starting at 1871:8597; EG-NEXT:     CNDE_INT T12.W, T1.Y, T1.Z, T12.W,8598; EG-NEXT:     CNDE_INT * T11.W, T3.Y, T11.W, T6.Z,8599; EG-NEXT:     BFE_UINT T7.X, T2.X, literal.x, 1,8600; EG-NEXT:     LSHL T1.Y, PS, 1,8601; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T7.W, literal.y,8602; EG-NEXT:     OR_INT T7.W, T2.Y, T0.Z,8603; EG-NEXT:     SUB_INT * T12.W, T4.X, T6.X,8604; EG-NEXT:    5(7.006492e-45), 31(4.344025e-44)8605; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8606; EG-NEXT:     CNDE_INT T2.Y, T3.Y, T9.W, PS, BS:VEC_120/SCL_2128607; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1228608; EG-NEXT:     SETE_INT T9.W, PV.Z, T4.Z,8609; EG-NEXT:     SETGE_UINT * T12.W, PV.Z, T4.Z,8610; EG-NEXT:    6(8.407791e-45), 0(0.000000e+00)8611; EG-NEXT:     SUBB_UINT T6.X, T8.W, T4.W,8612; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, PV.Z,8613; EG-NEXT:     SUB_INT * T0.Z, T7.W, T2.Z, BS:VEC_2018614; EG-NEXT:     BIT_ALIGN_INT T8.W, T2.Y, T11.W, literal.x,8615; EG-NEXT:     OR_INT * T9.W, T1.Y, T4.X,8616; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8617; EG-NEXT:     SUBB_UINT T4.X, T7.W, T2.Z,8618; EG-NEXT:     SUB_INT T1.Y, T1.Z, T4.Z, BS:VEC_021/SCL_1228619; EG-NEXT:     SETGE_UINT T6.Z, PS, T0.W, BS:VEC_021/SCL_1228620; EG-NEXT:     SETE_INT T11.W, PV.W, T1.W,8621; EG-NEXT:     SETGE_UINT * T12.W, PV.W, T1.W,8622; EG-NEXT:     SUBB_UINT T9.X, T9.W, T0.W, BS:VEC_021/SCL_1228623; EG-NEXT:     CNDE_INT T2.Y, PV.W, PS, PV.Z,8624; EG-NEXT:     SUB_INT T6.Z, T9.W, T0.W, BS:VEC_021/SCL_1228625; EG-NEXT:     SUB_INT T11.W, PV.Y, PV.X,8626; EG-NEXT:     CNDE_INT * T7.W, T3.Y, T7.W, T0.Z,8627; EG-NEXT:     SUB_INT T4.X, T8.W, T1.W, BS:VEC_021/SCL_1228628; EG-NEXT:     LSHL T1.Y, PS, 1,8629; EG-NEXT:     BFE_UINT T0.Z, T3.X, literal.x, 1,8630; EG-NEXT:     CNDE_INT T11.W, T3.Y, T1.Z, PV.W,8631; EG-NEXT:     CNDE_INT * T9.W, PV.Y, T9.W, PV.Z,8632; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)8633; EG-NEXT:     SUB_INT T10.X, T5.Z, T3.Z,8634; EG-NEXT:     LSHL T3.Y, PS, 1,8635; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.W, T7.W, literal.x,8636; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8637; EG-NEXT:     SUB_INT * T11.W, PV.X, T9.X,8638; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8639; EG-NEXT:     BFE_UINT T4.X, T0.Y, literal.x, 1,8640; EG-NEXT:     CNDE_INT T1.Y, T2.Y, T8.W, PS, BS:VEC_120/SCL_2128641; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T2.Z, BS:VEC_021/SCL_1228642; EG-NEXT:     SETE_INT T8.W, PV.Z, T4.Z,8643; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T4.Z,8644; EG-NEXT:    5(7.006492e-45), 0(0.000000e+00)8645; EG-NEXT:     CNDE_INT T9.X, PV.W, PS, PV.Z,8646; EG-NEXT:     SUB_INT T2.Y, T7.W, T2.Z,8647; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.Y, T9.W, literal.x,8648; EG-NEXT:     OR_INT T8.W, T3.Y, PV.X,8649; EG-NEXT:     SUB_INT * T9.W, T10.X, T6.X,8650; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8651; EG-NEXT:     CNDE_INT T4.X, T8.X, T5.Z, PS,8652; EG-NEXT:     SETGE_UINT T1.Y, PV.W, T0.W, BS:VEC_021/SCL_1228653; EG-NEXT:     SETE_INT T5.Z, PV.Z, T1.W, BS:VEC_102/SCL_2218654; EG-NEXT:     SETGE_UINT T9.W, PV.Z, T1.W, BS:VEC_102/SCL_2218655; EG-NEXT:     CNDE_INT * T11.W, PV.X, T7.W, PV.Y,8656; EG-NEXT:     LSHL T6.X, PS, 1,8657; EG-NEXT:     CNDE_INT T1.Y, PV.Z, PV.W, PV.Y,8658; EG-NEXT:     SUB_INT T5.Z, T8.W, T0.W,8659; EG-NEXT:     BIT_ALIGN_INT T9.W, PV.X, T10.W, literal.x, BS:VEC_021/SCL_1228660; EG-NEXT:     OR_INT * T10.W, T5.X, T7.X,8661; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8662; EG-NEXT:     BFE_UINT T4.X, T3.X, literal.x, 1,8663; EG-NEXT:     SETGE_UINT T2.Y, PS, T4.W, BS:VEC_021/SCL_1228664; EG-NEXT:     SETE_INT T6.Z, PV.W, T3.Z,8665; EG-NEXT:     SETGE_UINT T12.W, PV.W, T3.Z,8666; EG-NEXT:     CNDE_INT * T13.W, PV.Y, T8.W, PV.Z,8667; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)8668; EG-NEXT:     LSHL T5.X, PS, 1,8669; EG-NEXT:     CNDE_INT T2.Y, PV.Z, PV.W, PV.Y,8670; EG-NEXT:     SUB_INT T5.Z, T10.W, T4.W,8671; EG-NEXT:     SUBB_UINT T12.W, T10.W, T4.W,8672; EG-NEXT:     SUB_INT * T14.W, T9.W, T3.Z,8673; EG-NEXT:     SUBB_UINT T7.X, T7.W, T2.Z,8674; EG-NEXT:     SUBB_UINT * T3.Y, T8.W, T0.W, BS:VEC_120/SCL_2128675; EG-NEXT:     SUB_INT T6.Z, T0.Z, T1.W,8676; EG-NEXT:     SUB_INT * T7.W, T14.W, T12.W, BS:VEC_021/SCL_1228677; EG-NEXT:     CNDE_INT * T8.W, T2.Y, T10.W, T5.Z,8678; EG-NEXT:     SUB_INT T8.X, T1.Z, T4.Z,8679; EG-NEXT:     LSHL T4.Y, PV.W, 1,8680; EG-NEXT:     BFE_UINT T5.Z, T2.X, literal.x, 1,8681; EG-NEXT:     CNDE_INT T7.W, T2.Y, T9.W, T7.W,8682; EG-NEXT:     SUB_INT * T9.W, T6.Z, T3.Y,8683; EG-NEXT:    4(5.605194e-45), 0(0.000000e+00)8684; EG-NEXT:     BFE_UINT T10.X, T0.Y, literal.x, 1,8685; EG-NEXT:     CNDE_INT T1.Y, T1.Y, T0.Z, PS, BS:VEC_120/SCL_2128686; EG-NEXT:     BIT_ALIGN_INT T0.Z, PV.W, T8.W, literal.y,8687; EG-NEXT:     OR_INT T7.W, PV.Y, PV.Z,8688; EG-NEXT:     SUB_INT * T8.W, PV.X, T7.X,8689; EG-NEXT:    4(5.605194e-45), 31(4.344025e-44)8690; EG-NEXT:     CNDE_INT T7.X, T9.X, T1.Z, PS,8691; EG-NEXT:     SETGE_UINT T2.Y, PV.W, T4.W,8692; EG-NEXT:     SETE_INT T1.Z, PV.Z, T3.Z, BS:VEC_021/SCL_1228693; EG-NEXT:     BIT_ALIGN_INT T8.W, PV.Y, T13.W, literal.x, BS:VEC_021/SCL_1228694; EG-NEXT:     OR_INT * T9.W, T5.X, PV.X,8695; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8696; EG-NEXT:     SETGE_UINT T5.X, T0.Z, T3.Z,8697; EG-NEXT:     SUBB_UINT T1.Y, T7.W, T4.W,8698; EG-NEXT:     SETGE_UINT * T5.Z, PS, T0.W, BS:VEC_021/SCL_1228699; EG-NEXT:     SETE_INT T10.W, T8.W, T1.W,8700; EG-NEXT:     SETGE_UINT * T12.W, T8.W, T1.W,8701; EG-NEXT:     SUB_INT T8.X, T0.Z, T3.Z,8702; EG-NEXT:     CNDE_INT T3.Y, PV.W, PS, T5.Z,8703; EG-NEXT:     SUB_INT T5.Z, T9.W, T0.W,8704; EG-NEXT:     SUBB_UINT * T10.W, T9.W, T0.W,8705; EG-NEXT:     SUB_INT * T12.W, T8.W, T1.W,8706; EG-NEXT:     SUB_INT T9.X, PV.W, T10.W,8707; EG-NEXT:     CNDE_INT * T4.Y, T3.Y, T9.W, T5.Z, BS:VEC_021/SCL_1228708; EG-NEXT:     SUB_INT T5.Z, T7.W, T4.W,8709; EG-NEXT:     SUB_INT T9.W, T8.X, T1.Y,8710; EG-NEXT:     CNDE_INT * T10.W, T1.Z, T5.X, T2.Y,8711; EG-NEXT:    ALU clause starting at 1985:8712; EG-NEXT:     CNDE_INT T5.X, T10.W, T0.Z, T9.W,8713; EG-NEXT:     CNDE_INT T1.Y, T10.W, T7.W, T5.Z,8714; EG-NEXT:     LSHL T0.Z, T4.Y, 1,8715; EG-NEXT:     BFE_UINT * T7.W, T0.Y, literal.x, 1, BS:VEC_120/SCL_2128716; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)8717; EG-NEXT:     CNDE_INT * T8.W, T3.Y, T8.W, T9.X,8718; EG-NEXT:     BIT_ALIGN_INT T8.X, PV.W, T4.Y, literal.x,8719; EG-NEXT:     OR_INT T2.Y, T0.Z, T7.W,8720; EG-NEXT:     BIT_ALIGN_INT T0.Z, T5.X, T1.Y, literal.x, BS:VEC_021/SCL_1228721; EG-NEXT:     BIT_ALIGN_INT * T7.W, T7.X, T11.W, literal.x, BS:VEC_120/SCL_2128722; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8723; EG-NEXT:     OR_INT * T8.W, T6.X, T4.X,8724; EG-NEXT:     SETGE_UINT T4.X, PV.W, T2.Z,8725; EG-NEXT:     SETE_INT T3.Y, T7.W, T4.Z, BS:VEC_021/SCL_1228726; EG-NEXT:     SETGE_UINT T1.Z, T7.W, T4.Z, BS:VEC_021/SCL_1228727; EG-NEXT:     BFE_UINT T9.W, T2.X, literal.x, 1,8728; EG-NEXT:     LSHL * T10.W, T1.Y, 1,8729; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)8730; EG-NEXT:     OR_INT T5.X, PS, PV.W,8731; EG-NEXT:     CNDE_INT T1.Y, PV.Y, PV.Z, PV.X,8732; EG-NEXT:     SUB_INT T1.Z, T8.W, T2.Z, BS:VEC_021/SCL_1228733; EG-NEXT:     SUBB_UINT T9.W, T8.W, T2.Z, BS:VEC_021/SCL_1228734; EG-NEXT:     SUB_INT * T10.W, T7.W, T4.Z,8735; EG-NEXT:     SUB_INT T4.X, PS, PV.W,8736; EG-NEXT:     CNDE_INT T3.Y, PV.Y, T8.W, PV.Z,8737; EG-NEXT:     SETGE_UINT T1.Z, PV.X, T4.W, BS:VEC_021/SCL_1228738; EG-NEXT:     SETE_INT T8.W, T0.Z, T3.Z,8739; EG-NEXT:     SETGE_UINT * T9.W, T0.Z, T3.Z,8740; EG-NEXT:     CNDE_INT T6.X, PV.W, PS, PV.Z,8741; EG-NEXT:     SUB_INT T4.Y, T5.X, T4.W, BS:VEC_021/SCL_1228742; EG-NEXT:     LSHL T1.Z, PV.Y, 1,8743; EG-NEXT:     BFE_UINT T8.W, T3.X, literal.x, 1, BS:VEC_120/SCL_2128744; EG-NEXT:     CNDE_INT * T7.W, T1.Y, T7.W, PV.X,8745; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)8746; EG-NEXT:     BIT_ALIGN_INT T4.X, PS, T3.Y, literal.x,8747; EG-NEXT:     OR_INT T1.Y, PV.Z, PV.W,8748; EG-NEXT:     CNDE_INT T1.Z, PV.X, T5.X, PV.Y,8749; EG-NEXT:     SUBB_UINT T7.W, T2.Y, T0.W, BS:VEC_021/SCL_1228750; EG-NEXT:     SUB_INT * T8.W, T8.X, T1.W,8751; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8752; EG-NEXT:     SUB_INT T7.X, PS, PV.W,8753; EG-NEXT:     LSHL T3.Y, PV.Z, 1,8754; EG-NEXT:     SETGE_UINT T5.Z, PV.Y, T2.Z, BS:VEC_021/SCL_1228755; EG-NEXT:     SETE_INT T7.W, PV.X, T4.Z,8756; EG-NEXT:     SETGE_UINT * T8.W, PV.X, T4.Z,8757; EG-NEXT:     BFE_UINT T9.X, T2.X, literal.x, 1,8758; EG-NEXT:     SETGE_UINT T4.Y, T2.Y, T0.W,8759; EG-NEXT:     SETE_INT T6.Z, T8.X, T1.W, BS:VEC_120/SCL_2128760; EG-NEXT:     CNDE_INT T7.W, PV.W, PS, PV.Z,8761; EG-NEXT:     SUB_INT * T8.W, T1.Y, T2.Z,8762; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)8763; EG-NEXT:     SUBB_UINT T10.X, T1.Y, T2.Z,8764; EG-NEXT:     SUB_INT T5.Y, T4.X, T4.Z, BS:VEC_021/SCL_1228765; EG-NEXT:     SUBB_UINT * T5.Z, T5.X, T4.W, BS:VEC_120/SCL_2128766; EG-NEXT:     SUB_INT T9.W, T0.Z, T3.Z,8767; EG-NEXT:     CNDE_INT * T8.W, T7.W, T1.Y, T8.W,8768; EG-NEXT:     SETGE_UINT T5.X, T8.X, T1.W,8769; EG-NEXT:     LSHL T1.Y, PS, 1,8770; EG-NEXT:     BFE_UINT T7.Z, T3.X, 1, 1, BS:VEC_2018771; EG-NEXT:     SUB_INT T9.W, PV.W, T5.Z,8772; EG-NEXT:     SUB_INT * T10.W, T5.Y, T10.X,8773; EG-NEXT:     CNDE_INT T4.X, T7.W, T4.X, PS,8774; EG-NEXT:     CNDE_INT T5.Y, T6.X, T0.Z, PV.W,8775; EG-NEXT:     OR_INT T0.Z, PV.Y, PV.Z,8776; EG-NEXT:     CNDE_INT T7.W, T6.Z, PV.X, T4.Y, BS:VEC_021/SCL_1228777; EG-NEXT:     SUB_INT * T9.W, T2.Y, T0.W,8778; EG-NEXT:     CNDE_INT T5.X, PV.W, T2.Y, PS,8779; EG-NEXT:     SETGE_UINT T1.Y, PV.Z, T2.Z,8780; EG-NEXT:     BIT_ALIGN_INT T1.Z, PV.Y, T1.Z, literal.x, BS:VEC_021/SCL_1228781; EG-NEXT:     BIT_ALIGN_INT T8.W, PV.X, T8.W, literal.x,8782; EG-NEXT:     OR_INT * T9.W, T3.Y, T9.X,8783; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8784; EG-NEXT:     SETE_INT T4.X, PV.W, T4.Z, BS:VEC_021/SCL_1228785; EG-NEXT:     SETGE_UINT T2.Y, PV.W, T4.Z, BS:VEC_021/SCL_1228786; EG-NEXT:     SETGE_UINT T5.Z, PS, T4.W,8787; EG-NEXT:     SETE_INT T10.W, PV.Z, T3.Z,8788; EG-NEXT:     SETGE_UINT * T11.W, PV.Z, T3.Z,8789; EG-NEXT:     CNDE_INT T6.X, PV.W, PS, PV.Z,8790; EG-NEXT:     CNDE_INT T1.Y, PV.X, PV.Y, T1.Y,8791; EG-NEXT:     LSHL T5.Z, T5.X, 1, BS:VEC_2018792; EG-NEXT:     BFE_UINT T10.W, T0.Y, literal.x, 1,8793; EG-NEXT:     CNDE_INT * T7.W, T7.W, T8.X, T7.X,8794; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)8795; EG-NEXT:     SUB_INT T4.X, T0.Z, T2.Z,8796; EG-NEXT:     SUBB_UINT T2.Y, T0.Z, T2.Z,8797; EG-NEXT:     SUB_INT T6.Z, T8.W, T4.Z, BS:VEC_021/SCL_1228798; EG-NEXT:     BIT_ALIGN_INT T7.W, PS, T5.X, literal.x,8799; EG-NEXT:     OR_INT * T10.W, PV.Z, PV.W,8800; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8801; EG-NEXT:     SETGE_UINT T5.X, PS, T0.W,8802; EG-NEXT:     SETE_INT T3.Y, PV.W, T1.W, BS:VEC_021/SCL_1228803; EG-NEXT:     SETGE_UINT T5.Z, PV.W, T1.W, BS:VEC_021/SCL_1228804; EG-NEXT:     SUB_INT T11.W, PV.Z, PV.Y,8805; EG-NEXT:     CNDE_INT * T12.W, T1.Y, T0.Z, PV.X,8806; EG-NEXT:     SUBB_UINT * T4.X, T10.W, T0.W,8807; EG-NEXT:     SUB_INT T2.Y, T7.W, T1.W,8808; EG-NEXT:     LSHL T0.Z, T12.W, 1, BS:VEC_2018809; EG-NEXT:     AND_INT * T13.W, T3.X, 1,8810; EG-NEXT:     CNDE_INT * T8.W, T1.Y, T8.W, T11.W,8811; EG-NEXT:     SUB_INT T3.X, T10.W, T0.W,8812; EG-NEXT:     BIT_ALIGN_INT * T1.Y, PV.W, T12.W, literal.x, BS:VEC_021/SCL_1228813; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8814; EG-NEXT:     OR_INT T0.Z, T0.Z, T13.W,8815; EG-NEXT:     SUB_INT T8.W, T2.Y, T4.X,8816; EG-NEXT:     CNDE_INT * T11.W, T3.Y, T5.Z, T5.X,8817; EG-NEXT:     SUB_INT T4.X, T9.W, T4.W,8818; EG-NEXT:     CNDE_INT T2.Y, PS, T7.W, PV.W, BS:VEC_021/SCL_1228819; EG-NEXT:     SETGE_UINT T5.Z, PV.Z, T2.Z,8820; EG-NEXT:     SETE_INT * T7.W, T1.Y, T4.Z, BS:VEC_021/SCL_1228821; EG-NEXT:     CNDE_INT * T8.W, T11.W, T10.W, T3.X,8822; EG-NEXT:     SETGE_UINT T3.X, T1.Y, T4.Z,8823; EG-NEXT:     SUBB_UINT T3.Y, T0.Z, T2.Z, BS:VEC_021/SCL_1228824; EG-NEXT:     SUB_INT * T4.Z, T1.Y, T4.Z,8825; EG-NEXT:    ALU clause starting at 2098:8826; EG-NEXT:     BFE_UINT T10.W, T0.Y, 1, 1,8827; EG-NEXT:     LSHL * T11.W, T8.W, 1,8828; EG-NEXT:     SUBB_UINT T5.X, T9.W, T4.W,8829; EG-NEXT:     SUB_INT T4.Y, T1.Z, T3.Z,8830; EG-NEXT:     OR_INT T6.Z, PS, PV.W,8831; EG-NEXT:     SUB_INT * T10.W, T4.Z, T3.Y, BS:VEC_2018832; EG-NEXT:     CNDE_INT * T7.W, T7.W, T3.X, T5.Z,8833; EG-NEXT:     CNDE_INT T3.X, PV.W, T1.Y, T10.W,8834; EG-NEXT:     SETGE_UINT T1.Y, T6.Z, T0.W,8835; EG-NEXT:     SUB_INT T4.Z, T4.Y, T5.X,8836; EG-NEXT:     BIT_ALIGN_INT * T8.W, T2.Y, T8.W, literal.x, BS:VEC_2018837; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8838; EG-NEXT:     CNDE_INT * T9.W, T6.X, T9.W, T4.X,8839; EG-NEXT:     SETE_INT T4.X, T8.W, T1.W,8840; EG-NEXT:     SETGE_UINT T2.Y, T8.W, T1.W,8841; EG-NEXT:     LSHL T5.Z, PV.W, 1,8842; EG-NEXT:     BFE_UINT T10.W, T2.X, 1, 1,8843; EG-NEXT:     CNDE_INT * T11.W, T6.X, T1.Z, T4.Z,8844; EG-NEXT:     BIT_ALIGN_INT T5.X, PS, T9.W, literal.x, BS:VEC_021/SCL_1228845; EG-NEXT:     OR_INT T3.Y, PV.Z, PV.W,8846; EG-NEXT:     CNDE_INT T1.Z, PV.X, PV.Y, T1.Y,8847; EG-NEXT:     SUB_INT T9.W, T6.Z, T0.W, BS:VEC_102/SCL_2218848; EG-NEXT:     XOR_INT * T10.W, T3.X, T5.W,8849; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8850; EG-NEXT:     SUB_INT T3.X, PS, T5.W,8851; EG-NEXT:     CNDE_INT T1.Y, PV.Z, T6.Z, PV.W, BS:VEC_021/SCL_1228852; EG-NEXT:     SETGE_UINT T4.Z, PV.Y, T4.W, BS:VEC_021/SCL_1228853; EG-NEXT:     SETE_INT T9.W, PV.X, T3.Z,8854; EG-NEXT:     SETGE_UINT * T10.W, PV.X, T3.Z,8855; EG-NEXT:     CNDE_INT T4.X, PV.W, PS, PV.Z,8856; EG-NEXT:     LSHL T2.Y, PV.Y, 1,8857; EG-NEXT:     AND_INT T4.Z, T0.Y, 1,8858; EG-NEXT:     SUBB_UINT T9.W, T6.Z, T0.W, BS:VEC_102/SCL_2218859; EG-NEXT:     SUB_INT * T10.W, T8.W, T1.W,8860; EG-NEXT:     SUB_INT T6.X, T3.Y, T4.W,8861; EG-NEXT:     SUB_INT T0.Y, PS, PV.W,8862; EG-NEXT:     SUBB_UINT T5.Z, T3.Y, T4.W,8863; EG-NEXT:     SUB_INT T9.W, T5.X, T3.Z,8864; EG-NEXT:     OR_INT * T10.W, PV.Y, PV.Z,8865; EG-NEXT:     SUB_INT T7.X, PS, T0.W,8866; EG-NEXT:     SUB_INT T2.Y, PV.W, PV.Z,8867; EG-NEXT:     CNDE_INT T1.Z, T1.Z, T8.W, PV.Y, BS:VEC_021/SCL_1228868; EG-NEXT:     CNDE_INT T8.W, T4.X, T3.Y, PV.X,8869; EG-NEXT:     SUB_INT * T9.W, T0.Z, T2.Z,8870; EG-NEXT:     CNDE_INT T6.X, T7.W, T0.Z, PS,8871; EG-NEXT:     LSHL T0.Y, PV.W, 1,8872; EG-NEXT:     AND_INT T0.Z, T2.X, 1,8873; EG-NEXT:     BIT_ALIGN_INT T7.W, PV.Z, T1.Y, literal.x,8874; EG-NEXT:     CNDE_INT * T9.W, T4.X, T5.X, PV.Y,8875; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8876; EG-NEXT:     SETGE_UINT T2.X, T10.W, T0.W,8877; EG-NEXT:     SETE_INT T1.Y, PV.W, T1.W, BS:VEC_021/SCL_1228878; EG-NEXT:     SETGE_UINT * T1.Z, PV.W, T1.W, BS:VEC_021/SCL_1228879; EG-NEXT:     BIT_ALIGN_INT T8.W, T9.W, T8.W, literal.x,8880; EG-NEXT:     OR_INT * T9.W, T0.Y, T0.Z,8881; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)8882; EG-NEXT:     SETGE_UINT T4.X, PS, T4.W,8883; EG-NEXT:     SETE_INT T0.Y, PV.W, T3.Z,8884; EG-NEXT:     SETGE_UINT T0.Z, PV.W, T3.Z,8885; EG-NEXT:     SUBB_UINT T11.W, PS, T4.W,8886; EG-NEXT:     SUB_INT * T12.W, PV.W, T3.Z,8887; EG-NEXT:     SUBB_UINT * T5.X, T10.W, T0.W,8888; EG-NEXT:     SUB_INT * T2.Y, T7.W, T1.W,8889; EG-NEXT:     SUB_INT * T2.Z, T9.W, T4.W,8890; EG-NEXT:     SUB_INT T0.W, T12.W, T11.W,8891; EG-NEXT:     CNDE_INT * T1.W, T0.Y, T0.Z, T4.X,8892; EG-NEXT:     CNDE_INT T4.X, PS, T8.W, PV.W, BS:VEC_021/SCL_1228893; EG-NEXT:     CNDE_INT T0.Y, PS, T9.W, T2.Z, BS:VEC_102/SCL_2218894; EG-NEXT:     SUB_INT T0.Z, T2.Y, T5.X,8895; EG-NEXT:     CNDE_INT T0.W, T1.Y, T1.Z, T2.X, BS:VEC_2108896; EG-NEXT:     XOR_INT * T1.W, T6.X, T5.W,8897; EG-NEXT:     SUBB_UINT T2.X, PS, T5.W,8898; EG-NEXT:     CNDE_INT T1.Y, PV.W, T7.W, PV.Z, BS:VEC_021/SCL_1228899; EG-NEXT:     XOR_INT T0.Z, PV.Y, T3.W, BS:VEC_102/SCL_2218900; EG-NEXT:     XOR_INT * T4.W, PV.X, T3.W, BS:VEC_102/SCL_2218901; EG-NEXT:     CNDE_INT * T0.W, T0.W, T10.W, T7.X,8902; EG-NEXT:     XOR_INT T4.X, PV.W, T6.W,8903; EG-NEXT:     SUB_INT T0.Y, T4.W, T3.W, BS:VEC_021/SCL_1228904; EG-NEXT:     SUBB_UINT T1.Z, T0.Z, T3.W, BS:VEC_021/SCL_1228905; EG-NEXT:     XOR_INT T0.W, T1.Y, T6.W,8906; EG-NEXT:     SUB_INT * T7.W, T3.X, T2.X,8907; EG-NEXT:     SUB_INT T2.X, PV.W, T6.W, BS:VEC_021/SCL_1228908; EG-NEXT:     SUB_INT T7.Y, PV.Y, PV.Z,8909; EG-NEXT:     SUBB_UINT T1.Z, PV.X, T6.W, BS:VEC_021/SCL_1228910; EG-NEXT:     XOR_INT T0.W, T1.X, T2.W,8911; EG-NEXT:     XOR_INT * T4.W, T0.X, T2.W,8912; EG-NEXT:     SUB_INT T0.Y, PS, T2.W,8913; EG-NEXT:     SUB_INT T7.Z, T1.W, T5.W, BS:VEC_021/SCL_1228914; EG-NEXT:     SUBB_UINT T1.W, PV.W, T2.W,8915; EG-NEXT:     SUB_INT * T4.W, PV.X, PV.Z,8916; EG-NEXT:     SUB_INT T7.X, T0.Z, T3.W,8917; EG-NEXT:     SUB_INT T4.Y, PV.Y, PV.W,8918; EG-NEXT:     ADD_INT * T1.W, KC0[2].Y, literal.x,8919; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)8920; EG-NEXT:     LSHR T0.X, PV.W, literal.x,8921; EG-NEXT:     SUB_INT T4.Z, T4.X, T6.W, BS:VEC_102/SCL_2218922; EG-NEXT:     SUB_INT * T4.X, T0.W, T2.W,8923; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)8924; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,8925; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)8926  %den_ptr = getelementptr <4 x i64>, ptr addrspace(1) %in, i64 18927  %num = load <4 x i64>, ptr addrspace(1) %in8928  %den = load <4 x i64>, ptr addrspace(1) %den_ptr8929  %result = srem <4 x i64> %num, %den8930  store <4 x i64> %result, ptr addrspace(1) %out8931  ret void8932}8933 8934define amdgpu_kernel void @srem_v4i64_4(ptr addrspace(1) %out, ptr addrspace(1) %in) {8935; GCN-LABEL: srem_v4i64_4:8936; GCN:       ; %bb.0:8937; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x248938; GCN-NEXT:    v_mov_b32_e32 v8, 08939; GCN-NEXT:    s_waitcnt lgkmcnt(0)8940; GCN-NEXT:    global_load_dwordx4 v[0:3], v8, s[2:3]8941; GCN-NEXT:    global_load_dwordx4 v[4:7], v8, s[2:3] offset:168942; GCN-NEXT:    s_waitcnt vmcnt(1)8943; GCN-NEXT:    v_ashrrev_i32_e32 v9, 31, v18944; GCN-NEXT:    v_lshrrev_b32_e32 v9, 30, v98945; GCN-NEXT:    v_ashrrev_i32_e32 v10, 31, v38946; GCN-NEXT:    v_add_co_u32_e32 v9, vcc, v0, v98947; GCN-NEXT:    v_lshrrev_b32_e32 v10, 30, v108948; GCN-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v1, vcc8949; GCN-NEXT:    s_waitcnt vmcnt(0)8950; GCN-NEXT:    v_ashrrev_i32_e32 v11, 31, v58951; GCN-NEXT:    v_add_co_u32_e32 v10, vcc, v2, v108952; GCN-NEXT:    v_lshrrev_b32_e32 v11, 30, v118953; GCN-NEXT:    v_addc_co_u32_e32 v14, vcc, 0, v3, vcc8954; GCN-NEXT:    v_ashrrev_i32_e32 v12, 31, v78955; GCN-NEXT:    v_add_co_u32_e32 v11, vcc, v4, v118956; GCN-NEXT:    v_lshrrev_b32_e32 v12, 30, v128957; GCN-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v5, vcc8958; GCN-NEXT:    v_add_co_u32_e32 v12, vcc, v6, v128959; GCN-NEXT:    v_addc_co_u32_e32 v16, vcc, 0, v7, vcc8960; GCN-NEXT:    v_and_b32_e32 v9, -4, v98961; GCN-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v98962; GCN-NEXT:    v_and_b32_e32 v10, -4, v108963; GCN-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v13, vcc8964; GCN-NEXT:    v_sub_co_u32_e32 v2, vcc, v2, v108965; GCN-NEXT:    v_and_b32_e32 v11, -4, v118966; GCN-NEXT:    v_subb_co_u32_e32 v3, vcc, v3, v14, vcc8967; GCN-NEXT:    v_sub_co_u32_e32 v4, vcc, v4, v118968; GCN-NEXT:    v_and_b32_e32 v12, -4, v128969; GCN-NEXT:    v_subb_co_u32_e32 v5, vcc, v5, v15, vcc8970; GCN-NEXT:    v_sub_co_u32_e32 v6, vcc, v6, v128971; GCN-NEXT:    v_subb_co_u32_e32 v7, vcc, v7, v16, vcc8972; GCN-NEXT:    global_store_dwordx4 v8, v[4:7], s[0:1] offset:168973; GCN-NEXT:    global_store_dwordx4 v8, v[0:3], s[0:1]8974; GCN-NEXT:    s_endpgm8975;8976; TAHITI-LABEL: srem_v4i64_4:8977; TAHITI:       ; %bb.0:8978; TAHITI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x98979; TAHITI-NEXT:    s_mov_b32 s3, 0xf0008980; TAHITI-NEXT:    s_mov_b32 s2, -18981; TAHITI-NEXT:    s_mov_b32 s10, s28982; TAHITI-NEXT:    s_mov_b32 s11, s38983; TAHITI-NEXT:    s_waitcnt lgkmcnt(0)8984; TAHITI-NEXT:    s_mov_b32 s8, s68985; TAHITI-NEXT:    s_mov_b32 s9, s78986; TAHITI-NEXT:    buffer_load_dwordx4 v[0:3], off, s[8:11], 08987; TAHITI-NEXT:    buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:168988; TAHITI-NEXT:    s_mov_b32 s0, s48989; TAHITI-NEXT:    s_mov_b32 s1, s58990; TAHITI-NEXT:    s_waitcnt vmcnt(1)8991; TAHITI-NEXT:    v_ashrrev_i32_e32 v8, 31, v18992; TAHITI-NEXT:    v_lshrrev_b32_e32 v8, 30, v88993; TAHITI-NEXT:    v_ashrrev_i32_e32 v9, 31, v38994; TAHITI-NEXT:    v_add_i32_e32 v8, vcc, v0, v88995; TAHITI-NEXT:    v_lshrrev_b32_e32 v9, 30, v98996; TAHITI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v1, vcc8997; TAHITI-NEXT:    s_waitcnt vmcnt(0)8998; TAHITI-NEXT:    v_ashrrev_i32_e32 v10, 31, v58999; TAHITI-NEXT:    v_add_i32_e32 v9, vcc, v2, v99000; TAHITI-NEXT:    v_lshrrev_b32_e32 v10, 30, v109001; TAHITI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v3, vcc9002; TAHITI-NEXT:    v_ashrrev_i32_e32 v11, 31, v79003; TAHITI-NEXT:    v_add_i32_e32 v10, vcc, v4, v109004; TAHITI-NEXT:    v_lshrrev_b32_e32 v11, 30, v119005; TAHITI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v5, vcc9006; TAHITI-NEXT:    v_add_i32_e32 v11, vcc, v6, v119007; TAHITI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v7, vcc9008; TAHITI-NEXT:    v_and_b32_e32 v8, -4, v89009; TAHITI-NEXT:    v_sub_i32_e32 v0, vcc, v0, v89010; TAHITI-NEXT:    v_and_b32_e32 v9, -4, v99011; TAHITI-NEXT:    v_subb_u32_e32 v1, vcc, v1, v12, vcc9012; TAHITI-NEXT:    v_sub_i32_e32 v2, vcc, v2, v99013; TAHITI-NEXT:    v_and_b32_e32 v10, -4, v109014; TAHITI-NEXT:    v_subb_u32_e32 v3, vcc, v3, v13, vcc9015; TAHITI-NEXT:    v_sub_i32_e32 v4, vcc, v4, v109016; TAHITI-NEXT:    v_and_b32_e32 v11, -4, v119017; TAHITI-NEXT:    v_subb_u32_e32 v5, vcc, v5, v14, vcc9018; TAHITI-NEXT:    v_sub_i32_e32 v6, vcc, v6, v119019; TAHITI-NEXT:    v_subb_u32_e32 v7, vcc, v7, v15, vcc9020; TAHITI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:169021; TAHITI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 09022; TAHITI-NEXT:    s_endpgm9023;9024; TONGA-LABEL: srem_v4i64_4:9025; TONGA:       ; %bb.0:9026; TONGA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x249027; TONGA-NEXT:    s_waitcnt lgkmcnt(0)9028; TONGA-NEXT:    v_mov_b32_e32 v0, s29029; TONGA-NEXT:    v_mov_b32_e32 v1, s39030; TONGA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]9031; TONGA-NEXT:    s_add_u32 s4, s2, 169032; TONGA-NEXT:    s_addc_u32 s5, s3, 09033; TONGA-NEXT:    v_mov_b32_e32 v4, s49034; TONGA-NEXT:    v_mov_b32_e32 v5, s59035; TONGA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]9036; TONGA-NEXT:    v_mov_b32_e32 v9, s19037; TONGA-NEXT:    v_mov_b32_e32 v8, s09038; TONGA-NEXT:    s_add_u32 s0, s0, 169039; TONGA-NEXT:    s_addc_u32 s1, s1, 09040; TONGA-NEXT:    v_mov_b32_e32 v11, s19041; TONGA-NEXT:    v_mov_b32_e32 v10, s09042; TONGA-NEXT:    s_waitcnt vmcnt(1)9043; TONGA-NEXT:    v_ashrrev_i32_e32 v12, 31, v19044; TONGA-NEXT:    v_lshrrev_b32_e32 v12, 30, v129045; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v0, v129046; TONGA-NEXT:    v_addc_u32_e32 v16, vcc, 0, v1, vcc9047; TONGA-NEXT:    v_and_b32_e32 v12, -4, v129048; TONGA-NEXT:    v_ashrrev_i32_e32 v13, 31, v39049; TONGA-NEXT:    v_sub_u32_e32 v0, vcc, v0, v129050; TONGA-NEXT:    v_lshrrev_b32_e32 v13, 30, v139051; TONGA-NEXT:    v_subb_u32_e32 v1, vcc, v1, v16, vcc9052; TONGA-NEXT:    v_add_u32_e32 v12, vcc, v2, v139053; TONGA-NEXT:    v_addc_u32_e32 v13, vcc, 0, v3, vcc9054; TONGA-NEXT:    v_and_b32_e32 v12, -4, v129055; TONGA-NEXT:    s_waitcnt vmcnt(0)9056; TONGA-NEXT:    v_ashrrev_i32_e32 v14, 31, v59057; TONGA-NEXT:    v_ashrrev_i32_e32 v15, 31, v79058; TONGA-NEXT:    v_sub_u32_e32 v2, vcc, v2, v129059; TONGA-NEXT:    v_lshrrev_b32_e32 v14, 30, v149060; TONGA-NEXT:    v_lshrrev_b32_e32 v15, 30, v159061; TONGA-NEXT:    v_subb_u32_e32 v3, vcc, v3, v13, vcc9062; TONGA-NEXT:    v_add_u32_e64 v12, s[0:1], v4, v149063; TONGA-NEXT:    v_add_u32_e32 v13, vcc, v6, v159064; TONGA-NEXT:    v_addc_u32_e32 v15, vcc, 0, v7, vcc9065; TONGA-NEXT:    v_and_b32_e32 v12, -4, v129066; TONGA-NEXT:    v_addc_u32_e64 v14, s[0:1], 0, v5, s[0:1]9067; TONGA-NEXT:    v_sub_u32_e32 v4, vcc, v4, v129068; TONGA-NEXT:    v_and_b32_e32 v13, -4, v139069; TONGA-NEXT:    v_subb_u32_e32 v5, vcc, v5, v14, vcc9070; TONGA-NEXT:    v_sub_u32_e32 v6, vcc, v6, v139071; TONGA-NEXT:    v_subb_u32_e32 v7, vcc, v7, v15, vcc9072; TONGA-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]9073; TONGA-NEXT:    flat_store_dwordx4 v[10:11], v[4:7]9074; TONGA-NEXT:    s_endpgm9075;9076; EG-LABEL: srem_v4i64_4:9077; EG:       ; %bb.0:9078; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]9079; EG-NEXT:    TEX 1 @69080; EG-NEXT:    ALU 48, @11, KC0[CB0:0-32], KC1[]9081; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T3.XYZW, T0.X, 09082; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T1.X, 19083; EG-NEXT:    CF_END9084; EG-NEXT:    Fetch clause starting at 6:9085; EG-NEXT:     VTX_READ_128 T1.XYZW, T0.X, 16, #19086; EG-NEXT:     VTX_READ_128 T0.XYZW, T0.X, 0, #19087; EG-NEXT:    ALU clause starting at 10:9088; EG-NEXT:     MOV * T0.X, KC0[2].Z,9089; EG-NEXT:    ALU clause starting at 11:9090; EG-NEXT:     ASHR * T1.W, T1.W, literal.x,9091; EG-NEXT:    31(4.344025e-44), 0(0.000000e+00)9092; EG-NEXT:     ASHR T2.Z, T0.Y, literal.x,9093; EG-NEXT:     ASHR T2.W, T1.Y, literal.x, BS:VEC_120/SCL_2129094; EG-NEXT:     LSHR * T1.W, PV.W, literal.y,9095; EG-NEXT:    31(4.344025e-44), 30(4.203895e-44)9096; EG-NEXT:     ADD_INT T0.Y, T1.Z, PS,9097; EG-NEXT:     ASHR T3.Z, T0.W, literal.x,9098; EG-NEXT:     LSHR T0.W, PV.W, literal.y,9099; EG-NEXT:     LSHR * T2.W, PV.Z, literal.y,9100; EG-NEXT:    31(4.344025e-44), 30(4.203895e-44)9101; EG-NEXT:     ADD_INT T2.X, T0.X, PS,9102; EG-NEXT:     ADD_INT T1.Y, T1.X, PV.W, BS:VEC_120/SCL_2129103; EG-NEXT:     LSHR T2.Z, PV.Z, literal.x,9104; EG-NEXT:     ADDC_UINT T1.W, T1.Z, T1.W,9105; EG-NEXT:     AND_INT * T3.W, PV.Y, literal.y,9106; EG-NEXT:    30(4.203895e-44), -4(nan)9107; EG-NEXT:     SUBB_UINT T3.X, T1.Z, PS,9108; EG-NEXT:     BFE_INT T0.Y, PV.W, 0.0, 1,9109; EG-NEXT:     ADD_INT T3.Z, T0.Z, PV.Z, BS:VEC_120/SCL_2129110; EG-NEXT:     ADDC_UINT T0.W, T1.X, T0.W,9111; EG-NEXT:     AND_INT * T1.W, PV.Y, literal.x,9112; EG-NEXT:    -4(nan), 0(0.000000e+00)9113; EG-NEXT:     ADDC_UINT T4.X, T0.Z, T2.Z,9114; EG-NEXT:     SUBB_UINT T1.Y, T1.X, PS,9115; EG-NEXT:     BFE_INT T2.Z, PV.W, 0.0, 1,9116; EG-NEXT:     AND_INT T0.W, PV.Z, literal.x,9117; EG-NEXT:     SUB_INT * T4.W, PV.Y, PV.X,9118; EG-NEXT:    -4(nan), 0(0.000000e+00)9119; EG-NEXT:     SUBB_UINT T3.X, T0.Z, PV.W,9120; EG-NEXT:     SUB_INT T4.Y, PV.Z, PV.Y,9121; EG-NEXT:     BFE_INT T2.Z, PV.X, 0.0, 1,9122; EG-NEXT:     ADDC_UINT T2.W, T0.X, T2.W,9123; EG-NEXT:     AND_INT * T5.W, T2.X, literal.x,9124; EG-NEXT:    -4(nan), 0(0.000000e+00)9125; EG-NEXT:     SUBB_UINT T0.Y, T0.X, PS,9126; EG-NEXT:     SUB_INT T4.Z, T1.Z, T3.W,9127; EG-NEXT:     BFE_INT T2.W, PV.W, 0.0, 1,9128; EG-NEXT:     SUB_INT * T3.W, PV.Z, PV.X,9129; EG-NEXT:     SUB_INT T4.X, T1.X, T1.W,9130; EG-NEXT:     SUB_INT T3.Y, PV.W, PV.Y,9131; EG-NEXT:     ADD_INT * T1.W, KC0[2].Y, literal.x,9132; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)9133; EG-NEXT:     LSHR T1.X, PV.W, literal.x,9134; EG-NEXT:     SUB_INT T3.Z, T0.Z, T0.W, BS:VEC_021/SCL_1229135; EG-NEXT:     SUB_INT * T3.X, T0.X, T5.W,9136; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)9137; EG-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,9138; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)9139  %num = load <4 x i64>, ptr addrspace(1) %in9140  %result = srem <4 x i64> %num, <i64 4, i64 4, i64 4, i64 4>9141  store <4 x i64> %result, ptr addrspace(1) %out9142  ret void9143}9144