1234 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -O0 -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false < %s | FileCheck -check-prefix=GFX9-O0 %s3; RUN: llc -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false < %s | FileCheck -check-prefix=GFX9-O3 %s4 5define amdgpu_gfx void @strict_wwm_no_cfg(ptr addrspace(8) inreg %tmp14) {6; GFX9-O0-LABEL: strict_wwm_no_cfg:7; GFX9-O0: ; %bb.0:8; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)9; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -110; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill11; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill12; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill13; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]14; GFX9-O0-NEXT: s_mov_b32 s40, s615; GFX9-O0-NEXT: s_mov_b32 s34, s416; GFX9-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr4117; GFX9-O0-NEXT: s_mov_b32 s41, s718; GFX9-O0-NEXT: s_mov_b32 s42, s4119; GFX9-O0-NEXT: s_mov_b32 s43, s4020; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr3521; GFX9-O0-NEXT: s_mov_b32 s35, s522; GFX9-O0-NEXT: s_mov_b32 s44, s3523; GFX9-O0-NEXT: s_mov_b32 s36, s3424; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr3925; GFX9-O0-NEXT: s_mov_b32 s37, s4426; GFX9-O0-NEXT: s_mov_b32 s38, s4327; GFX9-O0-NEXT: s_mov_b32 s39, s4228; GFX9-O0-NEXT: s_mov_b32 s34, 029; GFX9-O0-NEXT: buffer_load_dwordx2 v[5:6], off, s[36:39], s3430; GFX9-O0-NEXT: s_waitcnt vmcnt(0)31; GFX9-O0-NEXT: v_mov_b32_e32 v4, v532; GFX9-O0-NEXT: v_mov_b32_e32 v3, v633; GFX9-O0-NEXT: ; implicit-def: $sgpr40_sgpr4134; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -135; GFX9-O0-NEXT: v_mov_b32_e32 v0, s3436; GFX9-O0-NEXT: v_cndmask_b32_e64 v0, v0, v4, s[40:41]37; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]38; GFX9-O0-NEXT: ; implicit-def: $sgpr40_sgpr4139; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -140; GFX9-O0-NEXT: v_mov_b32_e32 v1, s3441; GFX9-O0-NEXT: v_cndmask_b32_e64 v1, v1, v3, s[40:41]42; GFX9-O0-NEXT: v_mov_b32_e32 v2, s3443; GFX9-O0-NEXT: s_nop 144; GFX9-O0-NEXT: v_mov_b32_dpp v2, v0 row_bcast:31 row_mask:0xc bank_mask:0xf45; GFX9-O0-NEXT: v_add_u32_e64 v0, v0, v246; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]47; GFX9-O0-NEXT: v_mov_b32_e32 v3, v048; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -149; GFX9-O0-NEXT: v_mov_b32_e32 v0, s3450; GFX9-O0-NEXT: s_nop 151; GFX9-O0-NEXT: v_mov_b32_dpp v0, v1 row_bcast:31 row_mask:0xc bank_mask:0xf52; GFX9-O0-NEXT: v_add_u32_e64 v0, v1, v053; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]54; GFX9-O0-NEXT: v_mov_b32_e32 v4, v055; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[40:41], v3, v456; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[40:41]57; GFX9-O0-NEXT: s_mov_b32 s35, 158; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s35, v359; GFX9-O0-NEXT: s_mov_b32 s35, 260; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s3561; GFX9-O0-NEXT: buffer_store_dword v3, off, s[36:39], s34 offset:462; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -163; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload64; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload65; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload66; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]67; GFX9-O0-NEXT: s_waitcnt vmcnt(0)68; GFX9-O0-NEXT: s_setpc_b64 s[30:31]69;70; GFX9-O3-LABEL: strict_wwm_no_cfg:71; GFX9-O3: ; %bb.0:72; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)73; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -174; GFX9-O3-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill75; GFX9-O3-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill76; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill77; GFX9-O3-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill78; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]79; GFX9-O3-NEXT: buffer_load_dwordx2 v[4:5], off, s[4:7], 080; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -181; GFX9-O3-NEXT: v_mov_b32_e32 v0, 082; GFX9-O3-NEXT: v_mov_b32_e32 v1, 083; GFX9-O3-NEXT: s_waitcnt vmcnt(0)84; GFX9-O3-NEXT: v_cndmask_b32_e64 v2, 0, v4, s[34:35]85; GFX9-O3-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[34:35]86; GFX9-O3-NEXT: s_nop 087; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf88; GFX9-O3-NEXT: v_mov_b32_dpp v0, v3 row_bcast:31 row_mask:0xc bank_mask:0xf89; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v190; GFX9-O3-NEXT: v_add_u32_e32 v0, v3, v091; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]92; GFX9-O3-NEXT: v_mov_b32_e32 v4, v193; GFX9-O3-NEXT: v_mov_b32_e32 v5, v094; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v4, v595; GFX9-O3-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc96; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 1, v497; GFX9-O3-NEXT: v_and_b32_e32 v4, 2, v498; GFX9-O3-NEXT: buffer_store_dword v4, off, s[4:7], 0 offset:499; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1100; GFX9-O3-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload101; GFX9-O3-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload102; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload103; GFX9-O3-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload104; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]105; GFX9-O3-NEXT: s_waitcnt vmcnt(0)106; GFX9-O3-NEXT: s_setpc_b64 s[30:31]107 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)108 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>109 %tmp102 = extractelement <2 x i32> %tmp101, i32 0110 %tmp103 = extractelement <2 x i32> %tmp101, i32 1111 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)112 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)113 114 115 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)116 %tmp121 = add i32 %tmp105, %tmp120117 %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)118 119 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)120 %tmp136 = add i32 %tmp107, %tmp135121 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)122 123 %tmp138 = icmp eq i32 %tmp122, %tmp137124 %tmp139 = sext i1 %tmp138 to i32125 %tmp140 = shl nsw i32 %tmp139, 1126 %tmp141 = and i32 %tmp140, 2127 %tmp145 = bitcast i32 %tmp141 to float128 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)129 ret void130}131 132define amdgpu_gfx void @strict_wwm_cfg(ptr addrspace(8) inreg %tmp14, i32 %arg) {133; GFX9-O0-LABEL: strict_wwm_cfg:134; GFX9-O0: ; %bb.0: ; %entry135; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)136; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1137; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill138; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill139; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill140; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]141; GFX9-O0-NEXT: s_mov_b32 s40, s6142; GFX9-O0-NEXT: s_mov_b32 s34, s4143; GFX9-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41144; GFX9-O0-NEXT: s_mov_b32 s41, s7145; GFX9-O0-NEXT: s_mov_b32 s42, s41146; GFX9-O0-NEXT: s_mov_b32 s43, s40147; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35148; GFX9-O0-NEXT: s_mov_b32 s35, s5149; GFX9-O0-NEXT: s_mov_b32 s44, s35150; GFX9-O0-NEXT: s_mov_b32 s36, s34151; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39152; GFX9-O0-NEXT: s_mov_b32 s37, s44153; GFX9-O0-NEXT: s_mov_b32 s38, s43154; GFX9-O0-NEXT: s_mov_b32 s39, s42155; GFX9-O0-NEXT: ; implicit-def: $vgpr5 : SGPR spill to VGPR lane156; GFX9-O0-NEXT: v_writelane_b32 v5, s40, 0157; GFX9-O0-NEXT: v_writelane_b32 v5, s41, 1158; GFX9-O0-NEXT: v_writelane_b32 v5, s34, 2159; GFX9-O0-NEXT: v_writelane_b32 v5, s35, 3160; GFX9-O0-NEXT: s_mov_b32 s34, 0161; GFX9-O0-NEXT: s_nop 2162; GFX9-O0-NEXT: buffer_load_dwordx2 v[3:4], off, s[36:39], s34163; GFX9-O0-NEXT: s_waitcnt vmcnt(0)164; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill165; GFX9-O0-NEXT: s_nop 0166; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill167; GFX9-O0-NEXT: ; implicit-def: $sgpr36_sgpr37168; GFX9-O0-NEXT: s_or_saveexec_b64 s[36:37], -1169; GFX9-O0-NEXT: v_mov_b32_e32 v1, s34170; GFX9-O0-NEXT: v_cndmask_b32_e64 v1, v1, v3, s[36:37]171; GFX9-O0-NEXT: v_mov_b32_e32 v2, s34172; GFX9-O0-NEXT: s_nop 1173; GFX9-O0-NEXT: v_mov_b32_dpp v2, v1 row_bcast:31 row_mask:0xc bank_mask:0xf174; GFX9-O0-NEXT: v_add_u32_e64 v1, v1, v2175; GFX9-O0-NEXT: s_mov_b64 exec, s[36:37]176; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1177; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill178; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v0, s34179; GFX9-O0-NEXT: v_mov_b32_e32 v0, s34180; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill181; GFX9-O0-NEXT: s_mov_b64 s[34:35], exec182; GFX9-O0-NEXT: v_writelane_b32 v5, s34, 4183; GFX9-O0-NEXT: v_writelane_b32 v5, s35, 5184; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1185; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill186; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]187; GFX9-O0-NEXT: s_and_b64 s[34:35], s[34:35], s[36:37]188; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]189; GFX9-O0-NEXT: s_cbranch_execz .LBB1_2190; GFX9-O0-NEXT: ; %bb.1: ; %if191; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload192; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload193; GFX9-O0-NEXT: s_waitcnt vmcnt(0)194; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4195; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1196; GFX9-O0-NEXT: v_mov_b32_e32 v1, 0197; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]198; GFX9-O0-NEXT: ; implicit-def: $sgpr34_sgpr35199; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1200; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v1, v0, s[34:35]201; GFX9-O0-NEXT: s_nop 1202; GFX9-O0-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf203; GFX9-O0-NEXT: v_add_u32_e64 v1, v2, v1204; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]205; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1206; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill207; GFX9-O0-NEXT: .LBB1_2: ; %merge208; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1209; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload210; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]211; GFX9-O0-NEXT: s_waitcnt vmcnt(0)212; GFX9-O0-NEXT: v_readlane_b32 s36, v5, 4213; GFX9-O0-NEXT: v_readlane_b32 s37, v5, 5214; GFX9-O0-NEXT: s_or_b64 exec, exec, s[36:37]215; GFX9-O0-NEXT: v_readlane_b32 s38, v5, 0216; GFX9-O0-NEXT: v_readlane_b32 s39, v5, 1217; GFX9-O0-NEXT: v_readlane_b32 s34, v5, 2218; GFX9-O0-NEXT: v_readlane_b32 s35, v5, 3219; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload220; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload221; GFX9-O0-NEXT: s_waitcnt vmcnt(0)222; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v0, v3223; GFX9-O0-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[36:37]224; GFX9-O0-NEXT: s_mov_b32 s36, 1225; GFX9-O0-NEXT: v_lshlrev_b32_e64 v0, s36, v0226; GFX9-O0-NEXT: s_mov_b32 s36, 2227; GFX9-O0-NEXT: v_and_b32_e64 v0, v0, s36228; GFX9-O0-NEXT: s_mov_b32 s40, s35229; GFX9-O0-NEXT: s_mov_b32 s36, s34230; GFX9-O0-NEXT: s_mov_b32 s34, s39231; GFX9-O0-NEXT: s_mov_b32 s35, s38232; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39233; GFX9-O0-NEXT: s_mov_b32 s37, s40234; GFX9-O0-NEXT: s_mov_b32 s38, s35235; GFX9-O0-NEXT: s_mov_b32 s39, s34236; GFX9-O0-NEXT: s_mov_b32 s34, 0237; GFX9-O0-NEXT: buffer_store_dword v0, off, s[36:39], s34 offset:4238; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1239; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload240; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload241; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload242; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]243; GFX9-O0-NEXT: s_waitcnt vmcnt(0)244; GFX9-O0-NEXT: s_setpc_b64 s[30:31]245;246; GFX9-O3-LABEL: strict_wwm_cfg:247; GFX9-O3: ; %bb.0: ; %entry248; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)249; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1250; GFX9-O3-NEXT: buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill251; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill252; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]253; GFX9-O3-NEXT: buffer_load_dwordx2 v[3:4], off, s[4:7], 0254; GFX9-O3-NEXT: v_mov_b32_e32 v5, 0255; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -1256; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0257; GFX9-O3-NEXT: s_waitcnt vmcnt(0)258; GFX9-O3-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[34:35]259; GFX9-O3-NEXT: s_nop 1260; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf261; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1262; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]263; GFX9-O3-NEXT: v_mov_b32_e32 v3, v1264; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0265; GFX9-O3-NEXT: s_and_saveexec_b64 s[34:35], vcc266; GFX9-O3-NEXT: ; %bb.1: ; %if267; GFX9-O3-NEXT: s_or_saveexec_b64 s[36:37], -1268; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0269; GFX9-O3-NEXT: v_cndmask_b32_e64 v2, 0, v4, s[36:37]270; GFX9-O3-NEXT: s_nop 1271; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf272; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1273; GFX9-O3-NEXT: s_mov_b64 exec, s[36:37]274; GFX9-O3-NEXT: v_mov_b32_e32 v5, v1275; GFX9-O3-NEXT: ; %bb.2: ; %merge276; GFX9-O3-NEXT: s_or_b64 exec, exec, s[34:35]277; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5278; GFX9-O3-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc279; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 1, v0280; GFX9-O3-NEXT: v_and_b32_e32 v0, 2, v0281; GFX9-O3-NEXT: buffer_store_dword v0, off, s[4:7], 0 offset:4282; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1283; GFX9-O3-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload284; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload285; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]286; GFX9-O3-NEXT: s_waitcnt vmcnt(0)287; GFX9-O3-NEXT: s_setpc_b64 s[30:31]288entry:289 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)290 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>291 %tmp102 = extractelement <2 x i32> %tmp101, i32 0292 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)293 294 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)295 %tmp121 = add i32 %tmp105, %tmp120296 %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)297 298 %cond = icmp eq i32 %arg, 0299 br i1 %cond, label %if, label %merge300if:301 %tmp103 = extractelement <2 x i32> %tmp101, i32 1302 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)303 304 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)305 %tmp136 = add i32 %tmp107, %tmp135306 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)307 br label %merge308 309merge:310 %merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]311 %tmp138 = icmp eq i32 %tmp122, %merge_value312 %tmp139 = sext i1 %tmp138 to i32313 %tmp140 = shl nsw i32 %tmp139, 1314 %tmp141 = and i32 %tmp140, 2315 %tmp145 = bitcast i32 %tmp141 to float316 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)317 ret void318}319 320define hidden amdgpu_gfx i32 @strict_wwm_called(i32 %a) noinline {321; GFX9-O0-LABEL: strict_wwm_called:322; GFX9-O0: ; %bb.0:323; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)324; GFX9-O0-NEXT: v_add_u32_e64 v1, v0, v0325; GFX9-O0-NEXT: v_mul_lo_u32 v0, v1, v0326; GFX9-O0-NEXT: v_sub_u32_e64 v0, v0, v1327; GFX9-O0-NEXT: s_setpc_b64 s[30:31]328;329; GFX9-O3-LABEL: strict_wwm_called:330; GFX9-O3: ; %bb.0:331; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)332; GFX9-O3-NEXT: v_add_u32_e32 v1, v0, v0333; GFX9-O3-NEXT: v_mul_lo_u32 v0, v1, v0334; GFX9-O3-NEXT: v_sub_u32_e32 v0, v0, v1335; GFX9-O3-NEXT: s_setpc_b64 s[30:31]336 %add = add i32 %a, %a337 %mul = mul i32 %add, %a338 %sub = sub i32 %mul, %add339 ret i32 %sub340}341 342define amdgpu_gfx void @strict_wwm_call(ptr addrspace(8) inreg %tmp14, i32 inreg %arg) {343; GFX9-O0-LABEL: strict_wwm_call:344; GFX9-O0: ; %bb.0:345; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)346; GFX9-O0-NEXT: s_mov_b32 s48, s33347; GFX9-O0-NEXT: s_mov_b32 s33, s32348; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1349; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill350; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill351; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill352; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]353; GFX9-O0-NEXT: s_add_i32 s32, s32, 0x400354; GFX9-O0-NEXT: v_writelane_b32 v3, s30, 0355; GFX9-O0-NEXT: v_writelane_b32 v3, s31, 1356; GFX9-O0-NEXT: s_mov_b32 s40, s6357; GFX9-O0-NEXT: s_mov_b32 s34, s4358; GFX9-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41359; GFX9-O0-NEXT: s_mov_b32 s41, s7360; GFX9-O0-NEXT: s_mov_b32 s42, s41361; GFX9-O0-NEXT: s_mov_b32 s43, s40362; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35363; GFX9-O0-NEXT: s_mov_b32 s35, s5364; GFX9-O0-NEXT: s_mov_b32 s44, s35365; GFX9-O0-NEXT: s_mov_b32 s36, s34366; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39367; GFX9-O0-NEXT: s_mov_b32 s37, s44368; GFX9-O0-NEXT: s_mov_b32 s38, s43369; GFX9-O0-NEXT: s_mov_b32 s39, s42370; GFX9-O0-NEXT: s_mov_b32 s34, 0371; GFX9-O0-NEXT: ; implicit-def: $sgpr40_sgpr41372; GFX9-O0-NEXT: v_mov_b32_e32 v0, s8373; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -1374; GFX9-O0-NEXT: v_mov_b32_e32 v1, s34375; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v1, v0, s[40:41]376; GFX9-O0-NEXT: s_getpc_b64 s[42:43]377; GFX9-O0-NEXT: s_add_u32 s42, s42, strict_wwm_called@rel32@lo+4378; GFX9-O0-NEXT: s_addc_u32 s43, s43, strict_wwm_called@rel32@hi+12379; GFX9-O0-NEXT: s_mov_b64 s[46:47], s[2:3]380; GFX9-O0-NEXT: s_mov_b64 s[44:45], s[0:1]381; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[44:45]382; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[46:47]383; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2384; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[42:43]385; GFX9-O0-NEXT: v_mov_b32_e32 v1, v0386; GFX9-O0-NEXT: v_add_u32_e64 v1, v1, v2387; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]388; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1389; GFX9-O0-NEXT: buffer_store_dword v0, off, s[36:39], s34 offset:4390; GFX9-O0-NEXT: v_readlane_b32 s31, v3, 1391; GFX9-O0-NEXT: v_readlane_b32 s30, v3, 0392; GFX9-O0-NEXT: s_mov_b32 s32, s33393; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1394; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload395; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload396; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload397; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]398; GFX9-O0-NEXT: s_mov_b32 s33, s48399; GFX9-O0-NEXT: s_waitcnt vmcnt(0)400; GFX9-O0-NEXT: s_setpc_b64 s[30:31]401;402; GFX9-O3-LABEL: strict_wwm_call:403; GFX9-O3: ; %bb.0:404; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)405; GFX9-O3-NEXT: s_mov_b32 s38, s33406; GFX9-O3-NEXT: s_mov_b32 s33, s32407; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1408; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill409; GFX9-O3-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill410; GFX9-O3-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill411; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]412; GFX9-O3-NEXT: v_writelane_b32 v3, s30, 0413; GFX9-O3-NEXT: s_addk_i32 s32, 0x400414; GFX9-O3-NEXT: v_writelane_b32 v3, s31, 1415; GFX9-O3-NEXT: v_mov_b32_e32 v0, s8416; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -1417; GFX9-O3-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[34:35]418; GFX9-O3-NEXT: s_getpc_b64 s[36:37]419; GFX9-O3-NEXT: s_add_u32 s36, s36, strict_wwm_called@rel32@lo+4420; GFX9-O3-NEXT: s_addc_u32 s37, s37, strict_wwm_called@rel32@hi+12421; GFX9-O3-NEXT: v_mov_b32_e32 v0, v2422; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[36:37]423; GFX9-O3-NEXT: v_mov_b32_e32 v1, v0424; GFX9-O3-NEXT: v_add_u32_e32 v1, v1, v2425; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]426; GFX9-O3-NEXT: v_mov_b32_e32 v0, v1427; GFX9-O3-NEXT: buffer_store_dword v0, off, s[4:7], 0 offset:4428; GFX9-O3-NEXT: v_readlane_b32 s31, v3, 1429; GFX9-O3-NEXT: v_readlane_b32 s30, v3, 0430; GFX9-O3-NEXT: s_mov_b32 s32, s33431; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1432; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload433; GFX9-O3-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload434; GFX9-O3-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload435; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]436; GFX9-O3-NEXT: s_mov_b32 s33, s38437; GFX9-O3-NEXT: s_waitcnt vmcnt(0)438; GFX9-O3-NEXT: s_setpc_b64 s[30:31]439 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)440 %tmp134 = call amdgpu_gfx i32 @strict_wwm_called(i32 %tmp107)441 %tmp136 = add i32 %tmp134, %tmp107442 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)443 call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)444 ret void445}446 447define amdgpu_gfx i64 @strict_wwm_called_i64(i64 %a) noinline {448; GFX9-O0-LABEL: strict_wwm_called_i64:449; GFX9-O0: ; %bb.0:450; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)451; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0452; GFX9-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec453; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1454; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2455; GFX9-O0-NEXT: v_mov_b32_e32 v5, v2456; GFX9-O0-NEXT: v_mov_b32_e32 v0, v3457; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3458; GFX9-O0-NEXT: v_add_co_u32_e64 v4, s[34:35], v4, v5459; GFX9-O0-NEXT: v_addc_co_u32_e64 v0, s[34:35], v0, v1, s[34:35]460; GFX9-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec461; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0462; GFX9-O0-NEXT: s_mov_b32 s34, 32463; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2464; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3465; GFX9-O0-NEXT: v_lshrrev_b64 v[0:1], s34, v[0:1]466; GFX9-O0-NEXT: v_mov_b32_e32 v1, v0467; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4468; GFX9-O0-NEXT: v_mul_lo_u32 v1, v0, v1469; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2470; GFX9-O0-NEXT: v_lshrrev_b64 v[6:7], s34, v[4:5]471; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6472; GFX9-O0-NEXT: v_mul_lo_u32 v2, v2, v3473; GFX9-O0-NEXT: v_mad_u64_u32 v[6:7], s[36:37], v0, v3, 0474; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7475; GFX9-O0-NEXT: v_add3_u32 v0, v0, v1, v2476; GFX9-O0-NEXT: ; implicit-def: $sgpr35477; GFX9-O0-NEXT: ; implicit-def: $sgpr36478; GFX9-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec479; GFX9-O0-NEXT: v_mov_b32_e32 v1, s35480; GFX9-O0-NEXT: v_lshlrev_b64 v[1:2], s34, v[0:1]481; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2482; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec483; GFX9-O0-NEXT: v_mov_b32_e32 v0, 0484; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec485; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0486; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7487; GFX9-O0-NEXT: v_or_b32_e64 v0, v0, v3488; GFX9-O0-NEXT: v_mov_b32_e32 v2, v1489; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6490; GFX9-O0-NEXT: v_or_b32_e64 v6, v1, v2491; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec492; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0493; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6494; GFX9-O0-NEXT: v_mov_b32_e32 v3, v4495; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7496; GFX9-O0-NEXT: v_mov_b32_e32 v2, v5497; GFX9-O0-NEXT: v_sub_co_u32_e64 v1, s[36:37], v1, v3498; GFX9-O0-NEXT: v_subb_co_u32_e64 v0, s[36:37], v0, v2, s[36:37]499; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec500; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0501; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1502; GFX9-O0-NEXT: v_lshrrev_b64 v[1:2], s34, v[1:2]503; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec504; GFX9-O0-NEXT: s_setpc_b64 s[30:31]505;506; GFX9-O3-LABEL: strict_wwm_called_i64:507; GFX9-O3: ; %bb.0:508; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)509; GFX9-O3-NEXT: v_add_co_u32_e32 v2, vcc, v0, v0510; GFX9-O3-NEXT: v_addc_co_u32_e32 v3, vcc, v1, v1, vcc511; GFX9-O3-NEXT: v_mul_lo_u32 v4, v3, v0512; GFX9-O3-NEXT: v_mul_lo_u32 v5, v2, v1513; GFX9-O3-NEXT: v_mad_u64_u32 v[0:1], s[34:35], v2, v0, 0514; GFX9-O3-NEXT: v_add3_u32 v1, v1, v5, v4515; GFX9-O3-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v2516; GFX9-O3-NEXT: v_subb_co_u32_e32 v1, vcc, v1, v3, vcc517; GFX9-O3-NEXT: s_setpc_b64 s[30:31]518 %add = add i64 %a, %a519 %mul = mul i64 %add, %a520 %sub = sub i64 %mul, %add521 ret i64 %sub522}523 524define amdgpu_gfx void @strict_wwm_call_i64(ptr addrspace(8) inreg %tmp14, i64 inreg %arg) {525; GFX9-O0-LABEL: strict_wwm_call_i64:526; GFX9-O0: ; %bb.0:527; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)528; GFX9-O0-NEXT: s_mov_b32 s46, s33529; GFX9-O0-NEXT: s_mov_b32 s33, s32530; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1531; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill532; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill533; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill534; GFX9-O0-NEXT: s_nop 0535; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill536; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill537; GFX9-O0-NEXT: s_nop 0538; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill539; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill540; GFX9-O0-NEXT: s_nop 0541; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill542; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill543; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill544; GFX9-O0-NEXT: buffer_store_dword v10, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill545; GFX9-O0-NEXT: buffer_store_dword v11, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill546; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]547; GFX9-O0-NEXT: s_add_i32 s32, s32, 0x1000548; GFX9-O0-NEXT: v_writelane_b32 v10, s30, 0549; GFX9-O0-NEXT: v_writelane_b32 v10, s31, 1550; GFX9-O0-NEXT: s_mov_b32 s34, s8551; GFX9-O0-NEXT: s_mov_b32 s38, s6552; GFX9-O0-NEXT: s_mov_b32 s36, s4553; GFX9-O0-NEXT: ; kill: def $sgpr38 killed $sgpr38 def $sgpr38_sgpr39554; GFX9-O0-NEXT: s_mov_b32 s39, s7555; GFX9-O0-NEXT: s_mov_b32 s35, s39556; GFX9-O0-NEXT: s_mov_b32 s44, s38557; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37558; GFX9-O0-NEXT: s_mov_b32 s37, s5559; GFX9-O0-NEXT: s_mov_b32 s45, s37560; GFX9-O0-NEXT: s_mov_b32 s40, s36561; GFX9-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41_sgpr42_sgpr43562; GFX9-O0-NEXT: s_mov_b32 s41, s45563; GFX9-O0-NEXT: s_mov_b32 s42, s44564; GFX9-O0-NEXT: s_mov_b32 s43, s35565; GFX9-O0-NEXT: ; implicit-def: $vgpr11 : SGPR spill to VGPR lane566; GFX9-O0-NEXT: v_writelane_b32 v11, s40, 0567; GFX9-O0-NEXT: v_writelane_b32 v11, s41, 1568; GFX9-O0-NEXT: v_writelane_b32 v11, s42, 2569; GFX9-O0-NEXT: v_writelane_b32 v11, s43, 3570; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35571; GFX9-O0-NEXT: s_mov_b32 s35, s9572; GFX9-O0-NEXT: ; kill: def $sgpr36_sgpr37 killed $sgpr34_sgpr35573; GFX9-O0-NEXT: s_mov_b32 s38, s35574; GFX9-O0-NEXT: s_mov_b64 s[36:37], 0575; GFX9-O0-NEXT: s_mov_b32 s40, s37576; GFX9-O0-NEXT: ; implicit-def: $sgpr42_sgpr43577; GFX9-O0-NEXT: v_mov_b32_e32 v0, s38578; GFX9-O0-NEXT: s_or_saveexec_b64 s[38:39], -1579; GFX9-O0-NEXT: v_mov_b32_e32 v2, s40580; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v2, v0, s[38:39]581; GFX9-O0-NEXT: s_mov_b64 exec, s[38:39]582; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 killed $sgpr34_sgpr35583; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37584; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr39585; GFX9-O0-NEXT: v_mov_b32_e32 v0, s34586; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1587; GFX9-O0-NEXT: v_writelane_b32 v11, s34, 4588; GFX9-O0-NEXT: v_writelane_b32 v11, s35, 5589; GFX9-O0-NEXT: v_mov_b32_e32 v3, s36590; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, v3, v0, s[34:35]591; GFX9-O0-NEXT: v_mov_b32_e32 v8, v3592; GFX9-O0-NEXT: v_mov_b32_e32 v9, v2593; GFX9-O0-NEXT: s_mov_b32 s34, 32594; GFX9-O0-NEXT: v_lshrrev_b64 v[4:5], s34, v[8:9]595; GFX9-O0-NEXT: v_mov_b32_e32 v2, v4596; GFX9-O0-NEXT: s_getpc_b64 s[34:35]597; GFX9-O0-NEXT: s_add_u32 s34, s34, strict_wwm_called_i64@gotpcrel32@lo+4598; GFX9-O0-NEXT: s_addc_u32 s35, s35, strict_wwm_called_i64@gotpcrel32@hi+12599; GFX9-O0-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0600; GFX9-O0-NEXT: s_mov_b64 s[38:39], s[2:3]601; GFX9-O0-NEXT: s_mov_b64 s[36:37], s[0:1]602; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[36:37]603; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[38:39]604; GFX9-O0-NEXT: v_mov_b32_e32 v0, v3605; GFX9-O0-NEXT: v_mov_b32_e32 v1, v2606; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)607; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[34:35]608; GFX9-O0-NEXT: v_readlane_b32 s34, v11, 4609; GFX9-O0-NEXT: v_readlane_b32 s35, v11, 5610; GFX9-O0-NEXT: v_readlane_b32 s36, v11, 0611; GFX9-O0-NEXT: v_readlane_b32 s37, v11, 1612; GFX9-O0-NEXT: v_readlane_b32 s38, v11, 2613; GFX9-O0-NEXT: v_readlane_b32 s39, v11, 3614; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0615; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1616; GFX9-O0-NEXT: v_mov_b32_e32 v4, v8617; GFX9-O0-NEXT: v_mov_b32_e32 v5, v9618; GFX9-O0-NEXT: v_add_co_u32_e64 v2, s[40:41], v2, v4619; GFX9-O0-NEXT: v_addc_co_u32_e64 v3, s[40:41], v3, v5, s[40:41]620; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]621; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2622; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3623; GFX9-O0-NEXT: s_mov_b32 s34, 0624; GFX9-O0-NEXT: buffer_store_dwordx2 v[0:1], off, s[36:39], s34 offset:4625; GFX9-O0-NEXT: v_readlane_b32 s31, v10, 1626; GFX9-O0-NEXT: v_readlane_b32 s30, v10, 0627; GFX9-O0-NEXT: s_mov_b32 s32, s33628; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1629; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload630; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload631; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload632; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload633; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload634; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload635; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload636; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload637; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload638; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload639; GFX9-O0-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload640; GFX9-O0-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload641; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]642; GFX9-O0-NEXT: s_mov_b32 s33, s46643; GFX9-O0-NEXT: s_waitcnt vmcnt(0)644; GFX9-O0-NEXT: s_setpc_b64 s[30:31]645;646; GFX9-O3-LABEL: strict_wwm_call_i64:647; GFX9-O3: ; %bb.0:648; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)649; GFX9-O3-NEXT: s_mov_b32 s38, s33650; GFX9-O3-NEXT: s_mov_b32 s33, s32651; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1652; GFX9-O3-NEXT: buffer_store_dword v6, off, s[0:3], s33 ; 4-byte Folded Spill653; GFX9-O3-NEXT: buffer_store_dword v7, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill654; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill655; GFX9-O3-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill656; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill657; GFX9-O3-NEXT: s_nop 0658; GFX9-O3-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill659; GFX9-O3-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill660; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]661; GFX9-O3-NEXT: v_writelane_b32 v8, s30, 0662; GFX9-O3-NEXT: s_addk_i32 s32, 0x800663; GFX9-O3-NEXT: v_writelane_b32 v8, s31, 1664; GFX9-O3-NEXT: v_mov_b32_e32 v0, s9665; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -1666; GFX9-O3-NEXT: v_cndmask_b32_e64 v6, 0, v0, s[34:35]667; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]668; GFX9-O3-NEXT: v_mov_b32_e32 v0, s8669; GFX9-O3-NEXT: s_or_saveexec_b64 s[36:37], -1670; GFX9-O3-NEXT: s_getpc_b64 s[34:35]671; GFX9-O3-NEXT: s_add_u32 s34, s34, strict_wwm_called_i64@gotpcrel32@lo+4672; GFX9-O3-NEXT: s_addc_u32 s35, s35, strict_wwm_called_i64@gotpcrel32@hi+12673; GFX9-O3-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0674; GFX9-O3-NEXT: v_cndmask_b32_e64 v7, 0, v0, s[36:37]675; GFX9-O3-NEXT: v_mov_b32_e32 v0, v7676; GFX9-O3-NEXT: v_mov_b32_e32 v1, v6677; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)678; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[34:35]679; GFX9-O3-NEXT: v_mov_b32_e32 v2, v0680; GFX9-O3-NEXT: v_mov_b32_e32 v3, v1681; GFX9-O3-NEXT: v_add_co_u32_e32 v2, vcc, v2, v7682; GFX9-O3-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v6, vcc683; GFX9-O3-NEXT: s_mov_b64 exec, s[36:37]684; GFX9-O3-NEXT: v_mov_b32_e32 v0, v2685; GFX9-O3-NEXT: v_mov_b32_e32 v1, v3686; GFX9-O3-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 offset:4687; GFX9-O3-NEXT: v_readlane_b32 s31, v8, 1688; GFX9-O3-NEXT: v_readlane_b32 s30, v8, 0689; GFX9-O3-NEXT: s_mov_b32 s32, s33690; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1691; GFX9-O3-NEXT: buffer_load_dword v6, off, s[0:3], s33 ; 4-byte Folded Reload692; GFX9-O3-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload693; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload694; GFX9-O3-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload695; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload696; GFX9-O3-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload697; GFX9-O3-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload698; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]699; GFX9-O3-NEXT: s_mov_b32 s33, s38700; GFX9-O3-NEXT: s_waitcnt vmcnt(0)701; GFX9-O3-NEXT: s_setpc_b64 s[30:31]702 %tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)703 %tmp134 = call amdgpu_gfx i64 @strict_wwm_called_i64(i64 %tmp107)704 %tmp136 = add i64 %tmp134, %tmp107705 %tmp137 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp136)706 %tmp138 = bitcast i64 %tmp137 to <2 x i32>707 call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %tmp138, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)708 ret void709}710 711define amdgpu_gfx void @strict_wwm_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {712; GFX9-O0-LABEL: strict_wwm_amdgpu_cs_main:713; GFX9-O0: ; %bb.0:714; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)715; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1716; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill717; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill718; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill719; GFX9-O0-NEXT: s_nop 0720; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill721; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]722; GFX9-O0-NEXT: s_mov_b32 s36, s4723; GFX9-O0-NEXT: ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39724; GFX9-O0-NEXT: s_mov_b32 s37, s5725; GFX9-O0-NEXT: s_mov_b32 s38, s6726; GFX9-O0-NEXT: s_mov_b32 s39, s7727; GFX9-O0-NEXT: ; kill: def $sgpr40_sgpr41_sgpr42_sgpr43 killed $sgpr36_sgpr37_sgpr38_sgpr39728; GFX9-O0-NEXT: s_mov_b32 s34, 5729; GFX9-O0-NEXT: v_lshlrev_b32_e64 v0, s34, v0730; GFX9-O0-NEXT: s_mov_b32 s34, 0731; GFX9-O0-NEXT: buffer_load_dwordx4 v[11:14], v0, s[36:39], s34 offen732; GFX9-O0-NEXT: buffer_load_dwordx2 v[4:5], v0, s[36:39], s34 offen offset:16733; GFX9-O0-NEXT: s_waitcnt vmcnt(1)734; GFX9-O0-NEXT: v_mov_b32_e32 v6, v11735; GFX9-O0-NEXT: v_mov_b32_e32 v7, v12736; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -1737; GFX9-O0-NEXT: s_mov_b32 s35, 0x7fffffff738; GFX9-O0-NEXT: s_mov_b32 s44, -1739; GFX9-O0-NEXT: ; kill: def $sgpr44 killed $sgpr44 def $sgpr44_sgpr45740; GFX9-O0-NEXT: s_mov_b32 s45, s35741; GFX9-O0-NEXT: s_mov_b32 s42, s45742; GFX9-O0-NEXT: ; implicit-def: $sgpr46_sgpr47743; GFX9-O0-NEXT: v_mov_b32_e32 v1, s42744; GFX9-O0-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[40:41]745; GFX9-O0-NEXT: s_mov_b32 s35, s44746; GFX9-O0-NEXT: ; implicit-def: $sgpr44_sgpr45747; GFX9-O0-NEXT: v_mov_b32_e32 v2, s35748; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[40:41]749; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1750; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]751; GFX9-O0-NEXT: v_mov_b32_e32 v9, v2752; GFX9-O0-NEXT: v_mov_b32_e32 v10, v3753; GFX9-O0-NEXT: v_mov_b32_e32 v6, v13754; GFX9-O0-NEXT: v_mov_b32_e32 v7, v14755; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -1756; GFX9-O0-NEXT: ; implicit-def: $sgpr44_sgpr45757; GFX9-O0-NEXT: v_mov_b32_e32 v1, s42758; GFX9-O0-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[40:41]759; GFX9-O0-NEXT: ; implicit-def: $sgpr44_sgpr45760; GFX9-O0-NEXT: v_mov_b32_e32 v2, s35761; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v2, v6, s[40:41]762; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1763; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]764; GFX9-O0-NEXT: v_mov_b32_e32 v7, v2765; GFX9-O0-NEXT: v_mov_b32_e32 v8, v3766; GFX9-O0-NEXT: s_waitcnt vmcnt(0)767; GFX9-O0-NEXT: v_mov_b32_e32 v6, v5768; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -1769; GFX9-O0-NEXT: ; implicit-def: $sgpr44_sgpr45770; GFX9-O0-NEXT: v_mov_b32_e32 v1, s42771; GFX9-O0-NEXT: v_cndmask_b32_e64 v1, v1, v6, s[40:41]772; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]773; GFX9-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec774; GFX9-O0-NEXT: s_or_saveexec_b64 s[40:41], -1775; GFX9-O0-NEXT: ; implicit-def: $sgpr42_sgpr43776; GFX9-O0-NEXT: v_mov_b32_e32 v2, s35777; GFX9-O0-NEXT: v_cndmask_b32_e64 v2, v2, v4, s[40:41]778; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1779; GFX9-O0-NEXT: s_mov_b64 exec, s[40:41]780; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2781; GFX9-O0-NEXT: v_mov_b32_e32 v5, v3782; GFX9-O0-NEXT: v_mov_b32_e32 v12, v10783; GFX9-O0-NEXT: v_mov_b32_e32 v6, v9784; GFX9-O0-NEXT: v_mov_b32_e32 v10, v8785; GFX9-O0-NEXT: v_mov_b32_e32 v11, v7786; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7_vgpr8_vgpr9 killed $exec787; GFX9-O0-NEXT: v_mov_b32_e32 v7, v12788; GFX9-O0-NEXT: v_mov_b32_e32 v8, v11789; GFX9-O0-NEXT: v_mov_b32_e32 v9, v10790; GFX9-O0-NEXT: buffer_store_dwordx4 v[6:9], v0, s[36:39], s34 offen791; GFX9-O0-NEXT: buffer_store_dwordx2 v[4:5], v0, s[36:39], s34 offen offset:16792; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1793; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload794; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload795; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload796; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload797; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]798; GFX9-O0-NEXT: s_waitcnt vmcnt(0)799; GFX9-O0-NEXT: s_setpc_b64 s[30:31]800;801; GFX9-O3-LABEL: strict_wwm_amdgpu_cs_main:802; GFX9-O3: ; %bb.0:803; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)804; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1805; GFX9-O3-NEXT: buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill806; GFX9-O3-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill807; GFX9-O3-NEXT: s_nop 0808; GFX9-O3-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill809; GFX9-O3-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill810; GFX9-O3-NEXT: s_nop 0811; GFX9-O3-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill812; GFX9-O3-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill813; GFX9-O3-NEXT: s_nop 0814; GFX9-O3-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill815; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]816; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 5, v0817; GFX9-O3-NEXT: buffer_load_dwordx4 v[8:11], v0, s[4:7], 0 offen818; GFX9-O3-NEXT: buffer_load_dwordx2 v[12:13], v0, s[4:7], 0 offen offset:16819; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -1820; GFX9-O3-NEXT: v_bfrev_b32_e32 v1, -2821; GFX9-O3-NEXT: s_waitcnt vmcnt(1)822; GFX9-O3-NEXT: v_cndmask_b32_e64 v3, v1, v9, s[34:35]823; GFX9-O3-NEXT: v_cndmask_b32_e64 v2, -1, v8, s[34:35]824; GFX9-O3-NEXT: v_cndmask_b32_e64 v5, v1, v11, s[34:35]825; GFX9-O3-NEXT: v_cndmask_b32_e64 v4, -1, v10, s[34:35]826; GFX9-O3-NEXT: s_waitcnt vmcnt(0)827; GFX9-O3-NEXT: v_cndmask_b32_e64 v7, v1, v13, s[34:35]828; GFX9-O3-NEXT: v_cndmask_b32_e64 v6, -1, v12, s[34:35]829; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]830; GFX9-O3-NEXT: v_mov_b32_e32 v8, v2831; GFX9-O3-NEXT: v_mov_b32_e32 v10, v4832; GFX9-O3-NEXT: v_mov_b32_e32 v9, v3833; GFX9-O3-NEXT: v_mov_b32_e32 v11, v5834; GFX9-O3-NEXT: v_mov_b32_e32 v12, v6835; GFX9-O3-NEXT: v_mov_b32_e32 v13, v7836; GFX9-O3-NEXT: buffer_store_dwordx4 v[8:11], v0, s[4:7], 0 offen837; GFX9-O3-NEXT: buffer_store_dwordx2 v[12:13], v0, s[4:7], 0 offen offset:16838; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -1839; GFX9-O3-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload840; GFX9-O3-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload841; GFX9-O3-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload842; GFX9-O3-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload843; GFX9-O3-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload844; GFX9-O3-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload845; GFX9-O3-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload846; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]847; GFX9-O3-NEXT: s_waitcnt vmcnt(0)848; GFX9-O3-NEXT: s_setpc_b64 s[30:31]849 %tmp17 = shl i32 %index, 5850 %tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)851 %.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>852 %tmp19 = or i32 %tmp17, 16853 %tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)854 %.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0855 %tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)856 %tmp97 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp22)857 %.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1858 %tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)859 %tmp174 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp99)860 %.i25 = bitcast <2 x i32> %tmp20 to i64861 %tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)862 %tmp251 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp176)863 %.cast = bitcast i64 %tmp97 to <2 x float>864 %.cast6 = bitcast i64 %tmp174 to <2 x float>865 %.cast7 = bitcast i64 %tmp251 to <2 x float>866 %tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>867 %desc.int = bitcast <4 x i32> %desc to i128868 %desc.ptr = inttoptr i128 %desc.int to ptr addrspace(8)869 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %tmp254, ptr addrspace(8) %desc.ptr, i32 %tmp17, i32 0, i32 0)870 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float> %.cast7, ptr addrspace(8) %desc.ptr, i32 %tmp19, i32 0, i32 0)871 ret void872}873 874; FIXME: This spills v40 and v41 twice, once in whole-wave-mode and once for the active lanes875define amdgpu_gfx <32 x i32> @strict_wwm_callee_saves(<32 x i32> inreg %keep, ptr addrspace(5) %ptr, i64 %a, i64 %b, i64 %c, i64 %d, i64 %e) #0 {876; GFX9-O0-LABEL: strict_wwm_callee_saves:877; GFX9-O0: ; %bb.0:878; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)879; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1880; GFX9-O0-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill881; GFX9-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill882; GFX9-O0-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill883; GFX9-O0-NEXT: s_nop 0884; GFX9-O0-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill885; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]886; GFX9-O0-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill887; GFX9-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill888; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill889; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill890; GFX9-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill891; GFX9-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill892; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill893; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill894; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill895; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill896; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill897; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill898; GFX9-O0-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:20899; GFX9-O0-NEXT: s_waitcnt vmcnt(0)900; GFX9-O0-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill901; GFX9-O0-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:16902; GFX9-O0-NEXT: s_waitcnt vmcnt(0)903; GFX9-O0-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill904; GFX9-O0-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:12905; GFX9-O0-NEXT: s_waitcnt vmcnt(0)906; GFX9-O0-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill907; GFX9-O0-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:8908; GFX9-O0-NEXT: s_waitcnt vmcnt(0)909; GFX9-O0-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill910; GFX9-O0-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:4911; GFX9-O0-NEXT: s_waitcnt vmcnt(0)912; GFX9-O0-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill913; GFX9-O0-NEXT: buffer_load_dword v40, off, s[0:3], s32914; GFX9-O0-NEXT: s_waitcnt vmcnt(0)915; GFX9-O0-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill916; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4917; GFX9-O0-NEXT: v_mov_b32_e32 v1, s5918; GFX9-O0-NEXT: v_mov_b32_e32 v2, s6919; GFX9-O0-NEXT: v_mov_b32_e32 v3, s7920; GFX9-O0-NEXT: v_mov_b32_e32 v4, s8921; GFX9-O0-NEXT: v_mov_b32_e32 v5, s9922; GFX9-O0-NEXT: v_mov_b32_e32 v6, s10923; GFX9-O0-NEXT: v_mov_b32_e32 v7, s11924; GFX9-O0-NEXT: v_mov_b32_e32 v8, s12925; GFX9-O0-NEXT: v_mov_b32_e32 v9, s13926; GFX9-O0-NEXT: v_mov_b32_e32 v10, s14927; GFX9-O0-NEXT: v_mov_b32_e32 v11, s15928; GFX9-O0-NEXT: v_mov_b32_e32 v12, s16929; GFX9-O0-NEXT: v_mov_b32_e32 v13, s17930; GFX9-O0-NEXT: v_mov_b32_e32 v14, s18931; GFX9-O0-NEXT: v_mov_b32_e32 v15, s19932; GFX9-O0-NEXT: v_mov_b32_e32 v16, s20933; GFX9-O0-NEXT: v_mov_b32_e32 v17, s21934; GFX9-O0-NEXT: v_mov_b32_e32 v18, s22935; GFX9-O0-NEXT: v_mov_b32_e32 v19, s23936; GFX9-O0-NEXT: v_mov_b32_e32 v20, s24937; GFX9-O0-NEXT: v_mov_b32_e32 v21, s25938; GFX9-O0-NEXT: v_mov_b32_e32 v22, s26939; GFX9-O0-NEXT: v_mov_b32_e32 v23, s27940; GFX9-O0-NEXT: v_mov_b32_e32 v24, s28941; GFX9-O0-NEXT: v_mov_b32_e32 v25, s29942; GFX9-O0-NEXT: v_mov_b32_e32 v26, v40943; GFX9-O0-NEXT: v_mov_b32_e32 v27, v39944; GFX9-O0-NEXT: v_mov_b32_e32 v28, v38945; GFX9-O0-NEXT: v_mov_b32_e32 v29, v37946; GFX9-O0-NEXT: v_mov_b32_e32 v30, v36947; GFX9-O0-NEXT: ; kill: def $vgpr31 killed $vgpr35 killed $exec948; GFX9-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload949; GFX9-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload950; GFX9-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload951; GFX9-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload952; GFX9-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload953; GFX9-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload954; GFX9-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload955; GFX9-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload956; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload957; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload958; GFX9-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload959; GFX9-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload960; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload961; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload962; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload963; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload964; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload965; GFX9-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec966; GFX9-O0-NEXT: s_waitcnt vmcnt(4)967; GFX9-O0-NEXT: v_mov_b32_e32 v3, v5968; GFX9-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec969; GFX9-O0-NEXT: v_mov_b32_e32 v5, v7970; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec971; GFX9-O0-NEXT: v_mov_b32_e32 v7, v9972; GFX9-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec973; GFX9-O0-NEXT: v_mov_b32_e32 v9, v11974; GFX9-O0-NEXT: ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec975; GFX9-O0-NEXT: s_waitcnt vmcnt(0)976; GFX9-O0-NEXT: v_mov_b32_e32 v11, v0977; GFX9-O0-NEXT: v_mov_b32_e32 v0, v11978; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1979; GFX9-O0-NEXT: s_mov_b64 s[38:39], 0980; GFX9-O0-NEXT: s_mov_b32 s37, s39981; GFX9-O0-NEXT: ; implicit-def: $sgpr40_sgpr41982; GFX9-O0-NEXT: v_mov_b32_e32 v32, s37983; GFX9-O0-NEXT: v_cndmask_b32_e64 v32, v32, v0, s[34:35]984; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]985; GFX9-O0-NEXT: v_mov_b32_e32 v0, v10986; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1987; GFX9-O0-NEXT: s_mov_b32 s36, s38988; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr39989; GFX9-O0-NEXT: v_mov_b32_e32 v33, s36990; GFX9-O0-NEXT: v_cndmask_b32_e64 v33, v33, v0, s[34:35]991; GFX9-O0-NEXT: v_mov_b32_e32 v34, v32992; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]993; GFX9-O0-NEXT: v_mov_b32_e32 v10, v33994; GFX9-O0-NEXT: v_mov_b32_e32 v11, v34995; GFX9-O0-NEXT: v_mov_b32_e32 v0, v9996; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1997; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr39998; GFX9-O0-NEXT: v_mov_b32_e32 v32, s37999; GFX9-O0-NEXT: v_cndmask_b32_e64 v32, v32, v0, s[34:35]1000; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1001; GFX9-O0-NEXT: v_mov_b32_e32 v0, v81002; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11003; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391004; GFX9-O0-NEXT: v_mov_b32_e32 v33, s361005; GFX9-O0-NEXT: v_cndmask_b32_e64 v33, v33, v0, s[34:35]1006; GFX9-O0-NEXT: v_mov_b32_e32 v34, v321007; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1008; GFX9-O0-NEXT: v_mov_b32_e32 v8, v331009; GFX9-O0-NEXT: v_mov_b32_e32 v9, v341010; GFX9-O0-NEXT: v_mov_b32_e32 v0, v71011; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11012; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391013; GFX9-O0-NEXT: v_mov_b32_e32 v32, s371014; GFX9-O0-NEXT: v_cndmask_b32_e64 v32, v32, v0, s[34:35]1015; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1016; GFX9-O0-NEXT: v_mov_b32_e32 v0, v61017; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11018; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391019; GFX9-O0-NEXT: v_mov_b32_e32 v33, s361020; GFX9-O0-NEXT: v_cndmask_b32_e64 v33, v33, v0, s[34:35]1021; GFX9-O0-NEXT: v_mov_b32_e32 v34, v321022; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1023; GFX9-O0-NEXT: v_mov_b32_e32 v6, v331024; GFX9-O0-NEXT: v_mov_b32_e32 v7, v341025; GFX9-O0-NEXT: v_mov_b32_e32 v0, v51026; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11027; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391028; GFX9-O0-NEXT: v_mov_b32_e32 v32, s371029; GFX9-O0-NEXT: v_cndmask_b32_e64 v32, v32, v0, s[34:35]1030; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1031; GFX9-O0-NEXT: v_mov_b32_e32 v0, v41032; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11033; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391034; GFX9-O0-NEXT: v_mov_b32_e32 v33, s361035; GFX9-O0-NEXT: v_cndmask_b32_e64 v33, v33, v0, s[34:35]1036; GFX9-O0-NEXT: v_mov_b32_e32 v34, v321037; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1038; GFX9-O0-NEXT: v_mov_b32_e32 v4, v331039; GFX9-O0-NEXT: v_mov_b32_e32 v5, v341040; GFX9-O0-NEXT: v_mov_b32_e32 v0, v31041; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11042; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391043; GFX9-O0-NEXT: v_mov_b32_e32 v32, s371044; GFX9-O0-NEXT: v_cndmask_b32_e64 v32, v32, v0, s[34:35]1045; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1046; GFX9-O0-NEXT: v_mov_b32_e32 v0, v21047; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -11048; GFX9-O0-NEXT: ; implicit-def: $sgpr38_sgpr391049; GFX9-O0-NEXT: v_mov_b32_e32 v33, s361050; GFX9-O0-NEXT: v_cndmask_b32_e64 v33, v33, v0, s[34:35]1051; GFX9-O0-NEXT: v_mov_b32_e32 v34, v321052; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1053; GFX9-O0-NEXT: v_mov_b32_e32 v2, v331054; GFX9-O0-NEXT: v_mov_b32_e32 v3, v341055; GFX9-O0-NEXT: v_mov_b32_e32 v0, v111056; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:41057; GFX9-O0-NEXT: v_mov_b32_e32 v0, v101058; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen1059; GFX9-O0-NEXT: v_mov_b32_e32 v0, v91060; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:121061; GFX9-O0-NEXT: v_mov_b32_e32 v0, v81062; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:81063; GFX9-O0-NEXT: v_mov_b32_e32 v0, v71064; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:201065; GFX9-O0-NEXT: v_mov_b32_e32 v0, v61066; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:161067; GFX9-O0-NEXT: v_mov_b32_e32 v0, v51068; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:281069; GFX9-O0-NEXT: v_mov_b32_e32 v0, v41070; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:241071; GFX9-O0-NEXT: v_mov_b32_e32 v0, v31072; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:361073; GFX9-O0-NEXT: v_mov_b32_e32 v0, v21074; GFX9-O0-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen offset:321075; GFX9-O0-NEXT: v_mov_b32_e32 v0, s41076; GFX9-O0-NEXT: v_mov_b32_e32 v1, s51077; GFX9-O0-NEXT: v_mov_b32_e32 v2, s61078; GFX9-O0-NEXT: v_mov_b32_e32 v3, s71079; GFX9-O0-NEXT: v_mov_b32_e32 v4, s81080; GFX9-O0-NEXT: v_mov_b32_e32 v5, s91081; GFX9-O0-NEXT: v_mov_b32_e32 v6, s101082; GFX9-O0-NEXT: v_mov_b32_e32 v7, s111083; GFX9-O0-NEXT: v_mov_b32_e32 v8, s121084; GFX9-O0-NEXT: v_mov_b32_e32 v9, s131085; GFX9-O0-NEXT: v_mov_b32_e32 v10, s141086; GFX9-O0-NEXT: v_mov_b32_e32 v11, s151087; GFX9-O0-NEXT: v_mov_b32_e32 v12, s161088; GFX9-O0-NEXT: v_mov_b32_e32 v13, s171089; GFX9-O0-NEXT: v_mov_b32_e32 v14, s181090; GFX9-O0-NEXT: v_mov_b32_e32 v15, s191091; GFX9-O0-NEXT: v_mov_b32_e32 v16, s201092; GFX9-O0-NEXT: v_mov_b32_e32 v17, s211093; GFX9-O0-NEXT: v_mov_b32_e32 v18, s221094; GFX9-O0-NEXT: v_mov_b32_e32 v19, s231095; GFX9-O0-NEXT: v_mov_b32_e32 v20, s241096; GFX9-O0-NEXT: v_mov_b32_e32 v21, s251097; GFX9-O0-NEXT: v_mov_b32_e32 v22, s261098; GFX9-O0-NEXT: v_mov_b32_e32 v23, s271099; GFX9-O0-NEXT: v_mov_b32_e32 v24, s281100; GFX9-O0-NEXT: v_mov_b32_e32 v25, s291101; GFX9-O0-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload1102; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -11103; GFX9-O0-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload1104; GFX9-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload1105; GFX9-O0-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload1106; GFX9-O0-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload1107; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]1108; GFX9-O0-NEXT: s_waitcnt vmcnt(0)1109; GFX9-O0-NEXT: s_setpc_b64 s[30:31]1110;1111; GFX9-O3-LABEL: strict_wwm_callee_saves:1112; GFX9-O3: ; %bb.0:1113; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)1114; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -11115; GFX9-O3-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill1116; GFX9-O3-NEXT: s_nop 01117; GFX9-O3-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill1118; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1119; GFX9-O3-NEXT: buffer_load_dword v26, off, s[0:3], s321120; GFX9-O3-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:41121; GFX9-O3-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:81122; GFX9-O3-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:121123; GFX9-O3-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:161124; GFX9-O3-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:201125; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -11126; GFX9-O3-NEXT: v_cndmask_b32_e64 v33, 0, v2, s[34:35]1127; GFX9-O3-NEXT: v_cndmask_b32_e64 v32, 0, v1, s[34:35]1128; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1129; GFX9-O3-NEXT: v_mov_b32_e32 v1, v321130; GFX9-O3-NEXT: v_mov_b32_e32 v2, v331131; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -11132; GFX9-O3-NEXT: v_cndmask_b32_e64 v33, 0, v4, s[34:35]1133; GFX9-O3-NEXT: v_cndmask_b32_e64 v32, 0, v3, s[34:35]1134; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1135; GFX9-O3-NEXT: v_mov_b32_e32 v3, v321136; GFX9-O3-NEXT: v_mov_b32_e32 v4, v331137; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -11138; GFX9-O3-NEXT: v_cndmask_b32_e64 v33, 0, v6, s[34:35]1139; GFX9-O3-NEXT: v_cndmask_b32_e64 v32, 0, v5, s[34:35]1140; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1141; GFX9-O3-NEXT: v_mov_b32_e32 v5, v321142; GFX9-O3-NEXT: v_mov_b32_e32 v6, v331143; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -11144; GFX9-O3-NEXT: v_cndmask_b32_e64 v33, 0, v8, s[34:35]1145; GFX9-O3-NEXT: v_cndmask_b32_e64 v32, 0, v7, s[34:35]1146; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1147; GFX9-O3-NEXT: v_mov_b32_e32 v7, v321148; GFX9-O3-NEXT: v_mov_b32_e32 v8, v331149; GFX9-O3-NEXT: s_or_saveexec_b64 s[34:35], -11150; GFX9-O3-NEXT: v_cndmask_b32_e64 v33, 0, v10, s[34:35]1151; GFX9-O3-NEXT: v_cndmask_b32_e64 v32, 0, v9, s[34:35]1152; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1153; GFX9-O3-NEXT: v_mov_b32_e32 v9, v321154; GFX9-O3-NEXT: v_mov_b32_e32 v10, v331155; GFX9-O3-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:41156; GFX9-O3-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen1157; GFX9-O3-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:121158; GFX9-O3-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:81159; GFX9-O3-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:201160; GFX9-O3-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:161161; GFX9-O3-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:281162; GFX9-O3-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:241163; GFX9-O3-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:361164; GFX9-O3-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:321165; GFX9-O3-NEXT: v_mov_b32_e32 v0, s41166; GFX9-O3-NEXT: v_mov_b32_e32 v1, s51167; GFX9-O3-NEXT: v_mov_b32_e32 v2, s61168; GFX9-O3-NEXT: v_mov_b32_e32 v3, s71169; GFX9-O3-NEXT: v_mov_b32_e32 v4, s81170; GFX9-O3-NEXT: v_mov_b32_e32 v5, s91171; GFX9-O3-NEXT: v_mov_b32_e32 v6, s101172; GFX9-O3-NEXT: v_mov_b32_e32 v7, s111173; GFX9-O3-NEXT: v_mov_b32_e32 v8, s121174; GFX9-O3-NEXT: v_mov_b32_e32 v9, s131175; GFX9-O3-NEXT: v_mov_b32_e32 v10, s141176; GFX9-O3-NEXT: v_mov_b32_e32 v11, s151177; GFX9-O3-NEXT: v_mov_b32_e32 v12, s161178; GFX9-O3-NEXT: v_mov_b32_e32 v13, s171179; GFX9-O3-NEXT: v_mov_b32_e32 v14, s181180; GFX9-O3-NEXT: v_mov_b32_e32 v15, s191181; GFX9-O3-NEXT: v_mov_b32_e32 v16, s201182; GFX9-O3-NEXT: v_mov_b32_e32 v17, s211183; GFX9-O3-NEXT: v_mov_b32_e32 v18, s221184; GFX9-O3-NEXT: v_mov_b32_e32 v19, s231185; GFX9-O3-NEXT: v_mov_b32_e32 v20, s241186; GFX9-O3-NEXT: v_mov_b32_e32 v21, s251187; GFX9-O3-NEXT: v_mov_b32_e32 v22, s261188; GFX9-O3-NEXT: v_mov_b32_e32 v23, s271189; GFX9-O3-NEXT: v_mov_b32_e32 v24, s281190; GFX9-O3-NEXT: v_mov_b32_e32 v25, s291191; GFX9-O3-NEXT: s_xor_saveexec_b64 s[34:35], -11192; GFX9-O3-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload1193; GFX9-O3-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload1194; GFX9-O3-NEXT: s_mov_b64 exec, s[34:35]1195; GFX9-O3-NEXT: s_waitcnt vmcnt(0)1196; GFX9-O3-NEXT: s_setpc_b64 s[30:31]1197 %a2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %a, i64 0)1198 %a2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %a2.i)1199 %b2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %b, i64 0)1200 %b2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %b2.i)1201 %c2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %c, i64 0)1202 %c2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %c2.i)1203 %d2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %d, i64 0)1204 %d2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %d2.i)1205 %e2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %e, i64 0)1206 %e2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %e2.i)1207 store i64 %a2, ptr addrspace(5) %ptr1208 %ptr_b = getelementptr i64, ptr addrspace(5) %ptr, i32 11209 store i64 %b2, ptr addrspace(5) %ptr_b1210 %ptr_c = getelementptr i64, ptr addrspace(5) %ptr, i32 21211 store i64 %c2, ptr addrspace(5) %ptr_c1212 %ptr_d = getelementptr i64, ptr addrspace(5) %ptr, i32 31213 store i64 %d2, ptr addrspace(5) %ptr_d1214 %ptr_e = getelementptr i64, ptr addrspace(5) %ptr, i32 41215 store i64 %e2, ptr addrspace(5) %ptr_e1216 ret <32 x i32> %keep1217}1218 1219declare i32 @llvm.amdgcn.strict.wwm.i32(i32)1220declare i64 @llvm.amdgcn.strict.wwm.i64(i64)1221declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32)1222declare i64 @llvm.amdgcn.set.inactive.i64(i64, i64)1223declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1)1224declare <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8), i32, i32, i32)1225declare void @llvm.amdgcn.raw.ptr.buffer.store.f32(float, ptr addrspace(8), i32, i32, i32)1226declare void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32, ptr addrspace(8), i32, i32, i32)1227declare void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32>, ptr addrspace(8), i32, i32, i32)1228declare void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float>, ptr addrspace(8), i32, i32, i32)1229declare void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32)1230declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)1231declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)1232 1233attributes #0 = { "amdgpu-waves-per-eu"="5,5" }1234