1119 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn-- -mcpu=tahiti < %s | FileCheck -enable-var-scope --check-prefix=SI %s3; RUN: llc -mtriple=amdgcn-- -mcpu=fiji -mattr=-flat-for-global < %s | FileCheck -enable-var-scope --check-prefix=VI %s4; RUN: llc -mtriple=amdgcn-- -mcpu=gfx900 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope --check-prefix=GFX9 %s5; RUN: llc -mtriple=amdgcn-- -mcpu=gfx1010 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope --check-prefix=GFX10 %s6; RUN: llc -mtriple=amdgcn-- -mcpu=gfx1100 -mattr=-flat-for-global,+real-true16 < %s | FileCheck -enable-var-scope --check-prefixes=GFX11,GFX11-TRUE16 %s7; RUN: llc -mtriple=amdgcn-- -mcpu=gfx1100 -mattr=-flat-for-global,-real-true16 < %s | FileCheck -enable-var-scope --check-prefixes=GFX11,GFX11-FAKE16 %s8 9declare half @llvm.maxnum.f16(half %a, half %b)10declare <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)11declare <3 x half> @llvm.maxnum.v3f16(<3 x half> %a, <3 x half> %b)12declare <4 x half> @llvm.maxnum.v4f16(<4 x half> %a, <4 x half> %b)13 14define amdgpu_kernel void @maxnum_f16(15; SI-LABEL: maxnum_f16:16; SI: ; %bb.0: ; %entry17; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x918; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd19; SI-NEXT: s_mov_b32 s7, 0xf00020; SI-NEXT: s_mov_b32 s6, -121; SI-NEXT: s_mov_b32 s14, s622; SI-NEXT: s_waitcnt lgkmcnt(0)23; SI-NEXT: s_mov_b32 s12, s224; SI-NEXT: s_mov_b32 s13, s325; SI-NEXT: s_mov_b32 s15, s726; SI-NEXT: s_mov_b32 s10, s627; SI-NEXT: s_mov_b32 s11, s728; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc29; SI-NEXT: s_waitcnt vmcnt(0)30; SI-NEXT: buffer_load_ushort v1, off, s[8:11], 0 glc31; SI-NEXT: s_waitcnt vmcnt(0)32; SI-NEXT: s_mov_b32 s4, s033; SI-NEXT: s_mov_b32 s5, s134; SI-NEXT: v_cvt_f32_f16_e32 v0, v035; SI-NEXT: v_cvt_f32_f16_e32 v1, v136; SI-NEXT: v_max_f32_e32 v0, v0, v137; SI-NEXT: v_cvt_f16_f32_e32 v0, v038; SI-NEXT: buffer_store_short v0, off, s[4:7], 039; SI-NEXT: s_endpgm40;41; VI-LABEL: maxnum_f16:42; VI: ; %bb.0: ; %entry43; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2444; VI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x3445; VI-NEXT: s_mov_b32 s7, 0xf00046; VI-NEXT: s_mov_b32 s6, -147; VI-NEXT: s_mov_b32 s14, s648; VI-NEXT: s_waitcnt lgkmcnt(0)49; VI-NEXT: s_mov_b32 s12, s250; VI-NEXT: s_mov_b32 s13, s351; VI-NEXT: s_mov_b32 s15, s752; VI-NEXT: s_mov_b32 s10, s653; VI-NEXT: s_mov_b32 s11, s754; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc55; VI-NEXT: s_waitcnt vmcnt(0)56; VI-NEXT: buffer_load_ushort v1, off, s[8:11], 0 glc57; VI-NEXT: s_waitcnt vmcnt(0)58; VI-NEXT: s_mov_b32 s4, s059; VI-NEXT: s_mov_b32 s5, s160; VI-NEXT: v_max_f16_e32 v0, v0, v061; VI-NEXT: v_max_f16_e32 v1, v1, v162; VI-NEXT: v_max_f16_e32 v0, v0, v163; VI-NEXT: buffer_store_short v0, off, s[4:7], 064; VI-NEXT: s_endpgm65;66; GFX9-LABEL: maxnum_f16:67; GFX9: ; %bb.0: ; %entry68; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2469; GFX9-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x3470; GFX9-NEXT: s_mov_b32 s7, 0xf00071; GFX9-NEXT: s_mov_b32 s6, -172; GFX9-NEXT: s_mov_b32 s14, s673; GFX9-NEXT: s_waitcnt lgkmcnt(0)74; GFX9-NEXT: s_mov_b32 s12, s275; GFX9-NEXT: s_mov_b32 s13, s376; GFX9-NEXT: s_mov_b32 s15, s777; GFX9-NEXT: s_mov_b32 s10, s678; GFX9-NEXT: s_mov_b32 s11, s779; GFX9-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc80; GFX9-NEXT: s_waitcnt vmcnt(0)81; GFX9-NEXT: buffer_load_ushort v1, off, s[8:11], 0 glc82; GFX9-NEXT: s_waitcnt vmcnt(0)83; GFX9-NEXT: s_mov_b32 s4, s084; GFX9-NEXT: s_mov_b32 s5, s185; GFX9-NEXT: v_max_f16_e32 v0, v0, v086; GFX9-NEXT: v_max_f16_e32 v1, v1, v187; GFX9-NEXT: v_max_f16_e32 v0, v0, v188; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 089; GFX9-NEXT: s_endpgm90;91; GFX10-LABEL: maxnum_f16:92; GFX10: ; %bb.0: ; %entry93; GFX10-NEXT: s_clause 0x194; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2495; GFX10-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x3496; GFX10-NEXT: s_mov_b32 s6, -197; GFX10-NEXT: s_mov_b32 s7, 0x3101600098; GFX10-NEXT: s_mov_b32 s14, s699; GFX10-NEXT: s_mov_b32 s15, s7100; GFX10-NEXT: s_mov_b32 s10, s6101; GFX10-NEXT: s_mov_b32 s11, s7102; GFX10-NEXT: s_waitcnt lgkmcnt(0)103; GFX10-NEXT: s_mov_b32 s12, s2104; GFX10-NEXT: s_mov_b32 s13, s3105; GFX10-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc dlc106; GFX10-NEXT: s_waitcnt vmcnt(0)107; GFX10-NEXT: buffer_load_ushort v1, off, s[8:11], 0 glc dlc108; GFX10-NEXT: s_waitcnt vmcnt(0)109; GFX10-NEXT: s_mov_b32 s4, s0110; GFX10-NEXT: s_mov_b32 s5, s1111; GFX10-NEXT: v_max_f16_e32 v0, v0, v0112; GFX10-NEXT: v_max_f16_e32 v1, v1, v1113; GFX10-NEXT: v_max_f16_e32 v0, v0, v1114; GFX10-NEXT: buffer_store_short v0, off, s[4:7], 0115; GFX10-NEXT: s_endpgm116;117; GFX11-TRUE16-LABEL: maxnum_f16:118; GFX11-TRUE16: ; %bb.0: ; %entry119; GFX11-TRUE16-NEXT: s_clause 0x1120; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24121; GFX11-TRUE16-NEXT: s_load_b64 s[4:5], s[4:5], 0x34122; GFX11-TRUE16-NEXT: s_mov_b32 s10, -1123; GFX11-TRUE16-NEXT: s_mov_b32 s11, 0x31016000124; GFX11-TRUE16-NEXT: s_mov_b32 s14, s10125; GFX11-TRUE16-NEXT: s_mov_b32 s15, s11126; GFX11-TRUE16-NEXT: s_mov_b32 s6, s10127; GFX11-TRUE16-NEXT: s_mov_b32 s7, s11128; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)129; GFX11-TRUE16-NEXT: s_mov_b32 s12, s2130; GFX11-TRUE16-NEXT: s_mov_b32 s13, s3131; GFX11-TRUE16-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc132; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)133; GFX11-TRUE16-NEXT: buffer_load_u16 v1, off, s[4:7], 0 glc dlc134; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)135; GFX11-TRUE16-NEXT: s_mov_b32 s8, s0136; GFX11-TRUE16-NEXT: s_mov_b32 s9, s1137; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, v0.l, v0.l138; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v1.l, v1.l139; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)140; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, v0.l, v0.h141; GFX11-TRUE16-NEXT: buffer_store_b16 v0, off, s[8:11], 0142; GFX11-TRUE16-NEXT: s_endpgm143;144; GFX11-FAKE16-LABEL: maxnum_f16:145; GFX11-FAKE16: ; %bb.0: ; %entry146; GFX11-FAKE16-NEXT: s_clause 0x1147; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24148; GFX11-FAKE16-NEXT: s_load_b64 s[4:5], s[4:5], 0x34149; GFX11-FAKE16-NEXT: s_mov_b32 s10, -1150; GFX11-FAKE16-NEXT: s_mov_b32 s11, 0x31016000151; GFX11-FAKE16-NEXT: s_mov_b32 s14, s10152; GFX11-FAKE16-NEXT: s_mov_b32 s15, s11153; GFX11-FAKE16-NEXT: s_mov_b32 s6, s10154; GFX11-FAKE16-NEXT: s_mov_b32 s7, s11155; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)156; GFX11-FAKE16-NEXT: s_mov_b32 s12, s2157; GFX11-FAKE16-NEXT: s_mov_b32 s13, s3158; GFX11-FAKE16-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc159; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)160; GFX11-FAKE16-NEXT: buffer_load_u16 v1, off, s[4:7], 0 glc dlc161; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)162; GFX11-FAKE16-NEXT: s_mov_b32 s8, s0163; GFX11-FAKE16-NEXT: s_mov_b32 s9, s1164; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, v0, v0165; GFX11-FAKE16-NEXT: v_max_f16_e32 v1, v1, v1166; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)167; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, v0, v1168; GFX11-FAKE16-NEXT: buffer_store_b16 v0, off, s[8:11], 0169; GFX11-FAKE16-NEXT: s_endpgm170 ptr addrspace(1) %r,171 ptr addrspace(1) %a,172 ptr addrspace(1) %b) #0 {173entry:174 %a.val = load volatile half, ptr addrspace(1) %a175 %b.val = load volatile half, ptr addrspace(1) %b176 %r.val = call half @llvm.maxnum.f16(half %a.val, half %b.val)177 store half %r.val, ptr addrspace(1) %r178 ret void179}180 181define amdgpu_kernel void @maxnum_f16_imm_a(182; SI-LABEL: maxnum_f16_imm_a:183; SI: ; %bb.0: ; %entry184; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9185; SI-NEXT: s_mov_b32 s7, 0xf000186; SI-NEXT: s_mov_b32 s6, -1187; SI-NEXT: s_mov_b32 s10, s6188; SI-NEXT: s_mov_b32 s11, s7189; SI-NEXT: s_waitcnt lgkmcnt(0)190; SI-NEXT: s_mov_b32 s8, s2191; SI-NEXT: s_mov_b32 s9, s3192; SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0193; SI-NEXT: s_mov_b32 s4, s0194; SI-NEXT: s_mov_b32 s5, s1195; SI-NEXT: s_waitcnt vmcnt(0)196; SI-NEXT: v_cvt_f32_f16_e32 v0, v0197; SI-NEXT: v_max_f32_e32 v0, 0x40400000, v0198; SI-NEXT: v_cvt_f16_f32_e32 v0, v0199; SI-NEXT: buffer_store_short v0, off, s[4:7], 0200; SI-NEXT: s_endpgm201;202; VI-LABEL: maxnum_f16_imm_a:203; VI: ; %bb.0: ; %entry204; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24205; VI-NEXT: s_mov_b32 s7, 0xf000206; VI-NEXT: s_mov_b32 s6, -1207; VI-NEXT: s_mov_b32 s10, s6208; VI-NEXT: s_mov_b32 s11, s7209; VI-NEXT: s_waitcnt lgkmcnt(0)210; VI-NEXT: s_mov_b32 s8, s2211; VI-NEXT: s_mov_b32 s9, s3212; VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0213; VI-NEXT: s_mov_b32 s4, s0214; VI-NEXT: s_mov_b32 s5, s1215; VI-NEXT: s_waitcnt vmcnt(0)216; VI-NEXT: v_max_f16_e32 v0, v0, v0217; VI-NEXT: v_max_f16_e32 v0, 0x4200, v0218; VI-NEXT: buffer_store_short v0, off, s[4:7], 0219; VI-NEXT: s_endpgm220;221; GFX9-LABEL: maxnum_f16_imm_a:222; GFX9: ; %bb.0: ; %entry223; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24224; GFX9-NEXT: s_mov_b32 s7, 0xf000225; GFX9-NEXT: s_mov_b32 s6, -1226; GFX9-NEXT: s_mov_b32 s10, s6227; GFX9-NEXT: s_mov_b32 s11, s7228; GFX9-NEXT: s_waitcnt lgkmcnt(0)229; GFX9-NEXT: s_mov_b32 s8, s2230; GFX9-NEXT: s_mov_b32 s9, s3231; GFX9-NEXT: buffer_load_ushort v0, off, s[8:11], 0232; GFX9-NEXT: s_mov_b32 s4, s0233; GFX9-NEXT: s_mov_b32 s5, s1234; GFX9-NEXT: s_waitcnt vmcnt(0)235; GFX9-NEXT: v_max_f16_e32 v0, v0, v0236; GFX9-NEXT: v_max_f16_e32 v0, 0x4200, v0237; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0238; GFX9-NEXT: s_endpgm239;240; GFX10-LABEL: maxnum_f16_imm_a:241; GFX10: ; %bb.0: ; %entry242; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24243; GFX10-NEXT: s_mov_b32 s6, -1244; GFX10-NEXT: s_mov_b32 s7, 0x31016000245; GFX10-NEXT: s_mov_b32 s10, s6246; GFX10-NEXT: s_mov_b32 s11, s7247; GFX10-NEXT: s_waitcnt lgkmcnt(0)248; GFX10-NEXT: s_mov_b32 s8, s2249; GFX10-NEXT: s_mov_b32 s9, s3250; GFX10-NEXT: s_mov_b32 s4, s0251; GFX10-NEXT: buffer_load_ushort v0, off, s[8:11], 0252; GFX10-NEXT: s_mov_b32 s5, s1253; GFX10-NEXT: s_waitcnt vmcnt(0)254; GFX10-NEXT: v_max_f16_e32 v0, v0, v0255; GFX10-NEXT: v_max_f16_e32 v0, 0x4200, v0256; GFX10-NEXT: buffer_store_short v0, off, s[4:7], 0257; GFX10-NEXT: s_endpgm258;259; GFX11-TRUE16-LABEL: maxnum_f16_imm_a:260; GFX11-TRUE16: ; %bb.0: ; %entry261; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24262; GFX11-TRUE16-NEXT: s_mov_b32 s6, -1263; GFX11-TRUE16-NEXT: s_mov_b32 s7, 0x31016000264; GFX11-TRUE16-NEXT: s_mov_b32 s10, s6265; GFX11-TRUE16-NEXT: s_mov_b32 s11, s7266; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)267; GFX11-TRUE16-NEXT: s_mov_b32 s8, s2268; GFX11-TRUE16-NEXT: s_mov_b32 s9, s3269; GFX11-TRUE16-NEXT: s_mov_b32 s4, s0270; GFX11-TRUE16-NEXT: buffer_load_u16 v0, off, s[8:11], 0271; GFX11-TRUE16-NEXT: s_mov_b32 s5, s1272; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)273; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, v0.l, v0.l274; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)275; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, 0x4200, v0.l276; GFX11-TRUE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0277; GFX11-TRUE16-NEXT: s_endpgm278;279; GFX11-FAKE16-LABEL: maxnum_f16_imm_a:280; GFX11-FAKE16: ; %bb.0: ; %entry281; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24282; GFX11-FAKE16-NEXT: s_mov_b32 s6, -1283; GFX11-FAKE16-NEXT: s_mov_b32 s7, 0x31016000284; GFX11-FAKE16-NEXT: s_mov_b32 s10, s6285; GFX11-FAKE16-NEXT: s_mov_b32 s11, s7286; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)287; GFX11-FAKE16-NEXT: s_mov_b32 s8, s2288; GFX11-FAKE16-NEXT: s_mov_b32 s9, s3289; GFX11-FAKE16-NEXT: s_mov_b32 s4, s0290; GFX11-FAKE16-NEXT: buffer_load_u16 v0, off, s[8:11], 0291; GFX11-FAKE16-NEXT: s_mov_b32 s5, s1292; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)293; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, v0, v0294; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)295; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, 0x4200, v0296; GFX11-FAKE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0297; GFX11-FAKE16-NEXT: s_endpgm298 ptr addrspace(1) %r,299 ptr addrspace(1) %b) #0 {300entry:301 %b.val = load half, ptr addrspace(1) %b302 %r.val = call half @llvm.maxnum.f16(half 3.0, half %b.val)303 store half %r.val, ptr addrspace(1) %r304 ret void305}306 307define amdgpu_kernel void @maxnum_f16_imm_b(308; SI-LABEL: maxnum_f16_imm_b:309; SI: ; %bb.0: ; %entry310; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9311; SI-NEXT: s_mov_b32 s7, 0xf000312; SI-NEXT: s_mov_b32 s6, -1313; SI-NEXT: s_mov_b32 s10, s6314; SI-NEXT: s_mov_b32 s11, s7315; SI-NEXT: s_waitcnt lgkmcnt(0)316; SI-NEXT: s_mov_b32 s8, s2317; SI-NEXT: s_mov_b32 s9, s3318; SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0319; SI-NEXT: s_mov_b32 s4, s0320; SI-NEXT: s_mov_b32 s5, s1321; SI-NEXT: s_waitcnt vmcnt(0)322; SI-NEXT: v_cvt_f32_f16_e32 v0, v0323; SI-NEXT: v_max_f32_e32 v0, 4.0, v0324; SI-NEXT: v_cvt_f16_f32_e32 v0, v0325; SI-NEXT: buffer_store_short v0, off, s[4:7], 0326; SI-NEXT: s_endpgm327;328; VI-LABEL: maxnum_f16_imm_b:329; VI: ; %bb.0: ; %entry330; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24331; VI-NEXT: s_mov_b32 s7, 0xf000332; VI-NEXT: s_mov_b32 s6, -1333; VI-NEXT: s_mov_b32 s10, s6334; VI-NEXT: s_mov_b32 s11, s7335; VI-NEXT: s_waitcnt lgkmcnt(0)336; VI-NEXT: s_mov_b32 s8, s2337; VI-NEXT: s_mov_b32 s9, s3338; VI-NEXT: buffer_load_ushort v0, off, s[8:11], 0339; VI-NEXT: s_mov_b32 s4, s0340; VI-NEXT: s_mov_b32 s5, s1341; VI-NEXT: s_waitcnt vmcnt(0)342; VI-NEXT: v_max_f16_e32 v0, v0, v0343; VI-NEXT: v_max_f16_e32 v0, 4.0, v0344; VI-NEXT: buffer_store_short v0, off, s[4:7], 0345; VI-NEXT: s_endpgm346;347; GFX9-LABEL: maxnum_f16_imm_b:348; GFX9: ; %bb.0: ; %entry349; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24350; GFX9-NEXT: s_mov_b32 s7, 0xf000351; GFX9-NEXT: s_mov_b32 s6, -1352; GFX9-NEXT: s_mov_b32 s10, s6353; GFX9-NEXT: s_mov_b32 s11, s7354; GFX9-NEXT: s_waitcnt lgkmcnt(0)355; GFX9-NEXT: s_mov_b32 s8, s2356; GFX9-NEXT: s_mov_b32 s9, s3357; GFX9-NEXT: buffer_load_ushort v0, off, s[8:11], 0358; GFX9-NEXT: s_mov_b32 s4, s0359; GFX9-NEXT: s_mov_b32 s5, s1360; GFX9-NEXT: s_waitcnt vmcnt(0)361; GFX9-NEXT: v_max_f16_e32 v0, v0, v0362; GFX9-NEXT: v_max_f16_e32 v0, 4.0, v0363; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0364; GFX9-NEXT: s_endpgm365;366; GFX10-LABEL: maxnum_f16_imm_b:367; GFX10: ; %bb.0: ; %entry368; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24369; GFX10-NEXT: s_mov_b32 s6, -1370; GFX10-NEXT: s_mov_b32 s7, 0x31016000371; GFX10-NEXT: s_mov_b32 s10, s6372; GFX10-NEXT: s_mov_b32 s11, s7373; GFX10-NEXT: s_waitcnt lgkmcnt(0)374; GFX10-NEXT: s_mov_b32 s8, s2375; GFX10-NEXT: s_mov_b32 s9, s3376; GFX10-NEXT: s_mov_b32 s4, s0377; GFX10-NEXT: buffer_load_ushort v0, off, s[8:11], 0378; GFX10-NEXT: s_mov_b32 s5, s1379; GFX10-NEXT: s_waitcnt vmcnt(0)380; GFX10-NEXT: v_max_f16_e32 v0, v0, v0381; GFX10-NEXT: v_max_f16_e32 v0, 4.0, v0382; GFX10-NEXT: buffer_store_short v0, off, s[4:7], 0383; GFX10-NEXT: s_endpgm384;385; GFX11-TRUE16-LABEL: maxnum_f16_imm_b:386; GFX11-TRUE16: ; %bb.0: ; %entry387; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24388; GFX11-TRUE16-NEXT: s_mov_b32 s6, -1389; GFX11-TRUE16-NEXT: s_mov_b32 s7, 0x31016000390; GFX11-TRUE16-NEXT: s_mov_b32 s10, s6391; GFX11-TRUE16-NEXT: s_mov_b32 s11, s7392; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)393; GFX11-TRUE16-NEXT: s_mov_b32 s8, s2394; GFX11-TRUE16-NEXT: s_mov_b32 s9, s3395; GFX11-TRUE16-NEXT: s_mov_b32 s4, s0396; GFX11-TRUE16-NEXT: buffer_load_u16 v0, off, s[8:11], 0397; GFX11-TRUE16-NEXT: s_mov_b32 s5, s1398; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)399; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, v0.l, v0.l400; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)401; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.l, 4.0, v0.l402; GFX11-TRUE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0403; GFX11-TRUE16-NEXT: s_endpgm404;405; GFX11-FAKE16-LABEL: maxnum_f16_imm_b:406; GFX11-FAKE16: ; %bb.0: ; %entry407; GFX11-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24408; GFX11-FAKE16-NEXT: s_mov_b32 s6, -1409; GFX11-FAKE16-NEXT: s_mov_b32 s7, 0x31016000410; GFX11-FAKE16-NEXT: s_mov_b32 s10, s6411; GFX11-FAKE16-NEXT: s_mov_b32 s11, s7412; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0)413; GFX11-FAKE16-NEXT: s_mov_b32 s8, s2414; GFX11-FAKE16-NEXT: s_mov_b32 s9, s3415; GFX11-FAKE16-NEXT: s_mov_b32 s4, s0416; GFX11-FAKE16-NEXT: buffer_load_u16 v0, off, s[8:11], 0417; GFX11-FAKE16-NEXT: s_mov_b32 s5, s1418; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0)419; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, v0, v0420; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)421; GFX11-FAKE16-NEXT: v_max_f16_e32 v0, 4.0, v0422; GFX11-FAKE16-NEXT: buffer_store_b16 v0, off, s[4:7], 0423; GFX11-FAKE16-NEXT: s_endpgm424 ptr addrspace(1) %r,425 ptr addrspace(1) %a) #0 {426entry:427 %a.val = load half, ptr addrspace(1) %a428 %r.val = call half @llvm.maxnum.f16(half %a.val, half 4.0)429 store half %r.val, ptr addrspace(1) %r430 ret void431}432 433define amdgpu_kernel void @maxnum_v2f16(434; SI-LABEL: maxnum_v2f16:435; SI: ; %bb.0: ; %entry436; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9437; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd438; SI-NEXT: s_waitcnt lgkmcnt(0)439; SI-NEXT: s_load_dword s2, s[2:3], 0x0440; SI-NEXT: s_load_dword s4, s[4:5], 0x0441; SI-NEXT: s_mov_b32 s3, 0xf000442; SI-NEXT: s_waitcnt lgkmcnt(0)443; SI-NEXT: s_lshr_b32 s5, s2, 16444; SI-NEXT: s_lshr_b32 s6, s4, 16445; SI-NEXT: v_cvt_f32_f16_e32 v0, s5446; SI-NEXT: v_cvt_f32_f16_e32 v1, s6447; SI-NEXT: v_cvt_f32_f16_e32 v2, s2448; SI-NEXT: v_cvt_f32_f16_e32 v3, s4449; SI-NEXT: s_mov_b32 s2, -1450; SI-NEXT: v_max_f32_e32 v0, v0, v1451; SI-NEXT: v_cvt_f16_f32_e32 v0, v0452; SI-NEXT: v_max_f32_e32 v1, v2, v3453; SI-NEXT: v_cvt_f16_f32_e32 v1, v1454; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0455; SI-NEXT: v_or_b32_e32 v0, v1, v0456; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0457; SI-NEXT: s_endpgm458;459; VI-LABEL: maxnum_v2f16:460; VI: ; %bb.0: ; %entry461; VI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34462; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24463; VI-NEXT: s_mov_b32 s7, 0xf000464; VI-NEXT: s_mov_b32 s6, -1465; VI-NEXT: s_waitcnt lgkmcnt(0)466; VI-NEXT: s_load_dword s8, s[8:9], 0x0467; VI-NEXT: s_load_dword s2, s[2:3], 0x0468; VI-NEXT: s_mov_b32 s4, s0469; VI-NEXT: s_mov_b32 s5, s1470; VI-NEXT: s_waitcnt lgkmcnt(0)471; VI-NEXT: v_max_f16_e64 v0, s8, s8472; VI-NEXT: v_max_f16_e64 v1, s2, s2473; VI-NEXT: s_lshr_b32 s0, s8, 16474; VI-NEXT: v_max_f16_e32 v0, v1, v0475; VI-NEXT: v_max_f16_e64 v1, s0, s0476; VI-NEXT: s_lshr_b32 s0, s2, 16477; VI-NEXT: v_max_f16_e64 v2, s0, s0478; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD479; VI-NEXT: v_or_b32_e32 v0, v0, v1480; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0481; VI-NEXT: s_endpgm482;483; GFX9-LABEL: maxnum_v2f16:484; GFX9: ; %bb.0: ; %entry485; GFX9-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34486; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24487; GFX9-NEXT: s_mov_b32 s7, 0xf000488; GFX9-NEXT: s_mov_b32 s6, -1489; GFX9-NEXT: s_waitcnt lgkmcnt(0)490; GFX9-NEXT: s_load_dword s10, s[8:9], 0x0491; GFX9-NEXT: s_load_dword s11, s[2:3], 0x0492; GFX9-NEXT: s_mov_b32 s4, s0493; GFX9-NEXT: s_mov_b32 s5, s1494; GFX9-NEXT: s_waitcnt lgkmcnt(0)495; GFX9-NEXT: v_pk_max_f16 v0, s10, s10496; GFX9-NEXT: v_pk_max_f16 v1, s11, s11497; GFX9-NEXT: v_pk_max_f16 v0, v1, v0498; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0499; GFX9-NEXT: s_endpgm500;501; GFX10-LABEL: maxnum_v2f16:502; GFX10: ; %bb.0: ; %entry503; GFX10-NEXT: s_clause 0x1504; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34505; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24506; GFX10-NEXT: s_waitcnt lgkmcnt(0)507; GFX10-NEXT: s_load_dword s4, s[6:7], 0x0508; GFX10-NEXT: s_load_dword s5, s[2:3], 0x0509; GFX10-NEXT: s_mov_b32 s3, 0x31016000510; GFX10-NEXT: s_mov_b32 s2, -1511; GFX10-NEXT: s_waitcnt lgkmcnt(0)512; GFX10-NEXT: v_pk_max_f16 v0, s4, s4513; GFX10-NEXT: v_pk_max_f16 v1, s5, s5514; GFX10-NEXT: v_pk_max_f16 v0, v1, v0515; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], 0516; GFX10-NEXT: s_endpgm517;518; GFX11-LABEL: maxnum_v2f16:519; GFX11: ; %bb.0: ; %entry520; GFX11-NEXT: s_clause 0x1521; GFX11-NEXT: s_load_b64 s[6:7], s[4:5], 0x34522; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24523; GFX11-NEXT: s_waitcnt lgkmcnt(0)524; GFX11-NEXT: s_load_b32 s4, s[6:7], 0x0525; GFX11-NEXT: s_load_b32 s2, s[2:3], 0x0526; GFX11-NEXT: s_mov_b32 s3, 0x31016000527; GFX11-NEXT: s_waitcnt lgkmcnt(0)528; GFX11-NEXT: v_pk_max_f16 v0, s4, s4529; GFX11-NEXT: v_pk_max_f16 v1, s2, s2530; GFX11-NEXT: s_mov_b32 s2, -1531; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)532; GFX11-NEXT: v_pk_max_f16 v0, v1, v0533; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0534; GFX11-NEXT: s_endpgm535 ptr addrspace(1) %r,536 ptr addrspace(1) %a,537 ptr addrspace(1) %b) #0 {538entry:539 %a.val = load <2 x half>, ptr addrspace(1) %a540 %b.val = load <2 x half>, ptr addrspace(1) %b541 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> %b.val)542 store <2 x half> %r.val, ptr addrspace(1) %r543 ret void544}545 546define amdgpu_kernel void @maxnum_v2f16_imm_a(547; SI-LABEL: maxnum_v2f16_imm_a:548; SI: ; %bb.0: ; %entry549; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9550; SI-NEXT: s_waitcnt lgkmcnt(0)551; SI-NEXT: s_load_dword s2, s[2:3], 0x0552; SI-NEXT: s_waitcnt lgkmcnt(0)553; SI-NEXT: s_lshr_b32 s3, s2, 16554; SI-NEXT: v_cvt_f32_f16_e32 v0, s3555; SI-NEXT: v_cvt_f32_f16_e32 v1, s2556; SI-NEXT: s_mov_b32 s3, 0xf000557; SI-NEXT: s_mov_b32 s2, -1558; SI-NEXT: v_max_f32_e32 v0, 4.0, v0559; SI-NEXT: v_cvt_f16_f32_e32 v0, v0560; SI-NEXT: v_max_f32_e32 v1, 0x40400000, v1561; SI-NEXT: v_cvt_f16_f32_e32 v1, v1562; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0563; SI-NEXT: v_or_b32_e32 v0, v1, v0564; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0565; SI-NEXT: s_endpgm566;567; VI-LABEL: maxnum_v2f16_imm_a:568; VI: ; %bb.0: ; %entry569; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24570; VI-NEXT: v_mov_b32_e32 v2, 0x4400571; VI-NEXT: s_waitcnt lgkmcnt(0)572; VI-NEXT: s_load_dword s4, s[2:3], 0x0573; VI-NEXT: s_mov_b32 s3, 0xf000574; VI-NEXT: s_mov_b32 s2, -1575; VI-NEXT: s_waitcnt lgkmcnt(0)576; VI-NEXT: v_max_f16_e64 v0, s4, s4577; VI-NEXT: s_lshr_b32 s4, s4, 16578; VI-NEXT: v_max_f16_e64 v1, s4, s4579; VI-NEXT: v_max_f16_e32 v0, 0x4200, v0580; VI-NEXT: v_max_f16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD581; VI-NEXT: v_or_b32_e32 v0, v0, v1582; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0583; VI-NEXT: s_endpgm584;585; GFX9-LABEL: maxnum_v2f16_imm_a:586; GFX9: ; %bb.0: ; %entry587; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24588; GFX9-NEXT: s_waitcnt lgkmcnt(0)589; GFX9-NEXT: s_load_dword s4, s[2:3], 0x0590; GFX9-NEXT: s_mov_b32 s3, 0xf000591; GFX9-NEXT: s_mov_b32 s2, -1592; GFX9-NEXT: s_waitcnt lgkmcnt(0)593; GFX9-NEXT: v_pk_max_f16 v0, s4, s4594; GFX9-NEXT: s_mov_b32 s4, 0x44004200595; GFX9-NEXT: v_pk_max_f16 v0, v0, s4596; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0597; GFX9-NEXT: s_endpgm598;599; GFX10-LABEL: maxnum_v2f16_imm_a:600; GFX10: ; %bb.0: ; %entry601; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24602; GFX10-NEXT: s_waitcnt lgkmcnt(0)603; GFX10-NEXT: s_load_dword s2, s[2:3], 0x0604; GFX10-NEXT: s_mov_b32 s3, 0x31016000605; GFX10-NEXT: s_waitcnt lgkmcnt(0)606; GFX10-NEXT: v_pk_max_f16 v0, s2, s2607; GFX10-NEXT: s_mov_b32 s2, -1608; GFX10-NEXT: v_pk_max_f16 v0, 0x44004200, v0609; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], 0610; GFX10-NEXT: s_endpgm611;612; GFX11-LABEL: maxnum_v2f16_imm_a:613; GFX11: ; %bb.0: ; %entry614; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24615; GFX11-NEXT: s_waitcnt lgkmcnt(0)616; GFX11-NEXT: s_load_b32 s2, s[2:3], 0x0617; GFX11-NEXT: s_mov_b32 s3, 0x31016000618; GFX11-NEXT: s_waitcnt lgkmcnt(0)619; GFX11-NEXT: v_pk_max_f16 v0, s2, s2620; GFX11-NEXT: s_mov_b32 s2, -1621; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)622; GFX11-NEXT: v_pk_max_f16 v0, 0x44004200, v0623; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0624; GFX11-NEXT: s_endpgm625 ptr addrspace(1) %r,626 ptr addrspace(1) %b) #0 {627entry:628 %b.val = load <2 x half>, ptr addrspace(1) %b629 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> <half 3.0, half 4.0>, <2 x half> %b.val)630 store <2 x half> %r.val, ptr addrspace(1) %r631 ret void632}633 634define amdgpu_kernel void @maxnum_v2f16_imm_b(635; SI-LABEL: maxnum_v2f16_imm_b:636; SI: ; %bb.0: ; %entry637; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9638; SI-NEXT: s_waitcnt lgkmcnt(0)639; SI-NEXT: s_load_dword s2, s[2:3], 0x0640; SI-NEXT: s_waitcnt lgkmcnt(0)641; SI-NEXT: s_lshr_b32 s3, s2, 16642; SI-NEXT: v_cvt_f32_f16_e32 v0, s3643; SI-NEXT: v_cvt_f32_f16_e32 v1, s2644; SI-NEXT: s_mov_b32 s3, 0xf000645; SI-NEXT: s_mov_b32 s2, -1646; SI-NEXT: v_max_f32_e32 v0, 0x40400000, v0647; SI-NEXT: v_cvt_f16_f32_e32 v0, v0648; SI-NEXT: v_max_f32_e32 v1, 4.0, v1649; SI-NEXT: v_cvt_f16_f32_e32 v1, v1650; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0651; SI-NEXT: v_or_b32_e32 v0, v1, v0652; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0653; SI-NEXT: s_endpgm654;655; VI-LABEL: maxnum_v2f16_imm_b:656; VI: ; %bb.0: ; %entry657; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24658; VI-NEXT: v_mov_b32_e32 v2, 0x4200659; VI-NEXT: s_waitcnt lgkmcnt(0)660; VI-NEXT: s_load_dword s4, s[2:3], 0x0661; VI-NEXT: s_mov_b32 s3, 0xf000662; VI-NEXT: s_mov_b32 s2, -1663; VI-NEXT: s_waitcnt lgkmcnt(0)664; VI-NEXT: v_max_f16_e64 v0, s4, s4665; VI-NEXT: s_lshr_b32 s4, s4, 16666; VI-NEXT: v_max_f16_e64 v1, s4, s4667; VI-NEXT: v_max_f16_e32 v0, 4.0, v0668; VI-NEXT: v_max_f16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD669; VI-NEXT: v_or_b32_e32 v0, v0, v1670; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0671; VI-NEXT: s_endpgm672;673; GFX9-LABEL: maxnum_v2f16_imm_b:674; GFX9: ; %bb.0: ; %entry675; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24676; GFX9-NEXT: s_waitcnt lgkmcnt(0)677; GFX9-NEXT: s_load_dword s4, s[2:3], 0x0678; GFX9-NEXT: s_mov_b32 s3, 0xf000679; GFX9-NEXT: s_mov_b32 s2, -1680; GFX9-NEXT: s_waitcnt lgkmcnt(0)681; GFX9-NEXT: v_pk_max_f16 v0, s4, s4682; GFX9-NEXT: s_mov_b32 s4, 0x42004400683; GFX9-NEXT: v_pk_max_f16 v0, v0, s4684; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0685; GFX9-NEXT: s_endpgm686;687; GFX10-LABEL: maxnum_v2f16_imm_b:688; GFX10: ; %bb.0: ; %entry689; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24690; GFX10-NEXT: s_waitcnt lgkmcnt(0)691; GFX10-NEXT: s_load_dword s2, s[2:3], 0x0692; GFX10-NEXT: s_mov_b32 s3, 0x31016000693; GFX10-NEXT: s_waitcnt lgkmcnt(0)694; GFX10-NEXT: v_pk_max_f16 v0, s2, s2695; GFX10-NEXT: s_mov_b32 s2, -1696; GFX10-NEXT: v_pk_max_f16 v0, 0x42004400, v0697; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], 0698; GFX10-NEXT: s_endpgm699;700; GFX11-LABEL: maxnum_v2f16_imm_b:701; GFX11: ; %bb.0: ; %entry702; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24703; GFX11-NEXT: s_waitcnt lgkmcnt(0)704; GFX11-NEXT: s_load_b32 s2, s[2:3], 0x0705; GFX11-NEXT: s_mov_b32 s3, 0x31016000706; GFX11-NEXT: s_waitcnt lgkmcnt(0)707; GFX11-NEXT: v_pk_max_f16 v0, s2, s2708; GFX11-NEXT: s_mov_b32 s2, -1709; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)710; GFX11-NEXT: v_pk_max_f16 v0, 0x42004400, v0711; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0712; GFX11-NEXT: s_endpgm713 ptr addrspace(1) %r,714 ptr addrspace(1) %a) #0 {715entry:716 %a.val = load <2 x half>, ptr addrspace(1) %a717 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> <half 4.0, half 3.0>)718 store <2 x half> %r.val, ptr addrspace(1) %r719 ret void720}721 722; FIXME: Scalarize with undef half723define amdgpu_kernel void @maxnum_v3f16(724; SI-LABEL: maxnum_v3f16:725; SI: ; %bb.0: ; %entry726; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9727; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd728; SI-NEXT: s_waitcnt lgkmcnt(0)729; SI-NEXT: s_load_dwordx2 s[6:7], s[2:3], 0x0730; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0731; SI-NEXT: s_mov_b32 s3, 0xf000732; SI-NEXT: s_mov_b32 s2, -1733; SI-NEXT: s_waitcnt lgkmcnt(0)734; SI-NEXT: v_cvt_f32_f16_e32 v0, s7735; SI-NEXT: s_lshr_b32 s7, s6, 16736; SI-NEXT: s_lshr_b32 s8, s4, 16737; SI-NEXT: v_cvt_f32_f16_e32 v1, s7738; SI-NEXT: v_cvt_f32_f16_e32 v2, s8739; SI-NEXT: v_cvt_f32_f16_e32 v3, s6740; SI-NEXT: v_cvt_f32_f16_e32 v4, s4741; SI-NEXT: v_cvt_f32_f16_e32 v5, s5742; SI-NEXT: v_max_f32_e32 v1, v1, v2743; SI-NEXT: v_cvt_f16_f32_e32 v1, v1744; SI-NEXT: v_max_f32_e32 v2, v3, v4745; SI-NEXT: v_max_f32_e32 v0, v0, v5746; SI-NEXT: v_cvt_f16_f32_e32 v2, v2747; SI-NEXT: v_cvt_f16_f32_e32 v0, v0748; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1749; SI-NEXT: v_or_b32_e32 v1, v2, v1750; SI-NEXT: buffer_store_short v0, off, s[0:3], 0 offset:4751; SI-NEXT: buffer_store_dword v1, off, s[0:3], 0752; SI-NEXT: s_endpgm753;754; VI-LABEL: maxnum_v3f16:755; VI: ; %bb.0: ; %entry756; VI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34757; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24758; VI-NEXT: s_mov_b32 s7, 0xf000759; VI-NEXT: s_mov_b32 s6, -1760; VI-NEXT: s_waitcnt lgkmcnt(0)761; VI-NEXT: s_load_dwordx2 s[8:9], s[8:9], 0x0762; VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0763; VI-NEXT: s_mov_b32 s4, s0764; VI-NEXT: s_mov_b32 s5, s1765; VI-NEXT: s_waitcnt lgkmcnt(0)766; VI-NEXT: v_max_f16_e64 v0, s8, s8767; VI-NEXT: v_max_f16_e64 v1, s2, s2768; VI-NEXT: s_lshr_b32 s0, s8, 16769; VI-NEXT: v_max_f16_e32 v0, v1, v0770; VI-NEXT: v_max_f16_e64 v1, s0, s0771; VI-NEXT: s_lshr_b32 s0, s2, 16772; VI-NEXT: v_max_f16_e64 v2, s0, s0773; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD774; VI-NEXT: v_or_b32_e32 v0, v0, v1775; VI-NEXT: v_max_f16_e64 v1, s9, s9776; VI-NEXT: v_max_f16_e64 v2, s3, s3777; VI-NEXT: v_max_f16_e32 v1, v2, v1778; VI-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4779; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0780; VI-NEXT: s_endpgm781;782; GFX9-LABEL: maxnum_v3f16:783; GFX9: ; %bb.0: ; %entry784; GFX9-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34785; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24786; GFX9-NEXT: s_mov_b32 s7, 0xf000787; GFX9-NEXT: s_mov_b32 s6, -1788; GFX9-NEXT: s_waitcnt lgkmcnt(0)789; GFX9-NEXT: s_load_dwordx2 s[10:11], s[8:9], 0x0790; GFX9-NEXT: s_load_dwordx2 s[12:13], s[2:3], 0x0791; GFX9-NEXT: s_mov_b32 s4, s0792; GFX9-NEXT: s_mov_b32 s5, s1793; GFX9-NEXT: s_waitcnt lgkmcnt(0)794; GFX9-NEXT: v_pk_max_f16 v0, s10, s10795; GFX9-NEXT: v_pk_max_f16 v1, s12, s12796; GFX9-NEXT: v_pk_max_f16 v2, s11, s11797; GFX9-NEXT: v_pk_max_f16 v0, v1, v0798; GFX9-NEXT: v_pk_max_f16 v1, s13, s13799; GFX9-NEXT: v_pk_max_f16 v1, v1, v2800; GFX9-NEXT: buffer_store_short v1, off, s[4:7], 0 offset:4801; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0802; GFX9-NEXT: s_endpgm803;804; GFX10-LABEL: maxnum_v3f16:805; GFX10: ; %bb.0: ; %entry806; GFX10-NEXT: s_clause 0x1807; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34808; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24809; GFX10-NEXT: s_waitcnt lgkmcnt(0)810; GFX10-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0811; GFX10-NEXT: s_load_dwordx2 s[8:9], s[2:3], 0x0812; GFX10-NEXT: s_mov_b32 s3, 0x31016000813; GFX10-NEXT: s_mov_b32 s2, -1814; GFX10-NEXT: s_waitcnt lgkmcnt(0)815; GFX10-NEXT: v_pk_max_f16 v1, s5, s5816; GFX10-NEXT: v_pk_max_f16 v2, s9, s9817; GFX10-NEXT: v_pk_max_f16 v0, s4, s4818; GFX10-NEXT: v_pk_max_f16 v3, s8, s8819; GFX10-NEXT: v_pk_max_f16 v1, v2, v1820; GFX10-NEXT: v_pk_max_f16 v0, v3, v0821; GFX10-NEXT: buffer_store_short v1, off, s[0:3], 0 offset:4822; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], 0823; GFX10-NEXT: s_endpgm824;825; GFX11-LABEL: maxnum_v3f16:826; GFX11: ; %bb.0: ; %entry827; GFX11-NEXT: s_clause 0x1828; GFX11-NEXT: s_load_b64 s[6:7], s[4:5], 0x34829; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24830; GFX11-NEXT: s_waitcnt lgkmcnt(0)831; GFX11-NEXT: s_load_b64 s[4:5], s[6:7], 0x0832; GFX11-NEXT: s_load_b64 s[2:3], s[2:3], 0x0833; GFX11-NEXT: s_waitcnt lgkmcnt(0)834; GFX11-NEXT: v_pk_max_f16 v1, s5, s5835; GFX11-NEXT: v_pk_max_f16 v2, s3, s3836; GFX11-NEXT: v_pk_max_f16 v0, s4, s4837; GFX11-NEXT: v_pk_max_f16 v3, s2, s2838; GFX11-NEXT: s_mov_b32 s3, 0x31016000839; GFX11-NEXT: s_mov_b32 s2, -1840; GFX11-NEXT: v_pk_max_f16 v1, v2, v1841; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)842; GFX11-NEXT: v_pk_max_f16 v0, v3, v0843; GFX11-NEXT: s_clause 0x1844; GFX11-NEXT: buffer_store_b16 v1, off, s[0:3], 0 offset:4845; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0846; GFX11-NEXT: s_endpgm847 ptr addrspace(1) %r,848 ptr addrspace(1) %a,849 ptr addrspace(1) %b) #0 {850entry:851 %a.val = load <3 x half>, ptr addrspace(1) %a852 %b.val = load <3 x half>, ptr addrspace(1) %b853 %r.val = call <3 x half> @llvm.maxnum.v3f16(<3 x half> %a.val, <3 x half> %b.val)854 store <3 x half> %r.val, ptr addrspace(1) %r855 ret void856}857 858define amdgpu_kernel void @maxnum_v4f16(859; SI-LABEL: maxnum_v4f16:860; SI: ; %bb.0: ; %entry861; SI-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x9862; SI-NEXT: s_mov_b32 s3, 0xf000863; SI-NEXT: s_mov_b32 s2, -1864; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd865; SI-NEXT: s_waitcnt lgkmcnt(0)866; SI-NEXT: s_load_dwordx2 s[6:7], s[10:11], 0x0867; SI-NEXT: s_mov_b32 s0, s8868; SI-NEXT: s_mov_b32 s1, s9869; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0870; SI-NEXT: s_waitcnt lgkmcnt(0)871; SI-NEXT: v_cvt_f32_f16_e32 v0, s6872; SI-NEXT: s_lshr_b32 s6, s6, 16873; SI-NEXT: v_cvt_f32_f16_e32 v2, s6874; SI-NEXT: s_lshr_b32 s6, s7, 16875; SI-NEXT: v_cvt_f32_f16_e32 v3, s6876; SI-NEXT: v_cvt_f32_f16_e32 v4, s4877; SI-NEXT: s_lshr_b32 s6, s5, 16878; SI-NEXT: s_lshr_b32 s4, s4, 16879; SI-NEXT: v_cvt_f32_f16_e32 v5, s6880; SI-NEXT: v_cvt_f32_f16_e32 v7, s4881; SI-NEXT: v_cvt_f32_f16_e32 v1, s7882; SI-NEXT: v_cvt_f32_f16_e32 v6, s5883; SI-NEXT: v_max_f32_e32 v3, v3, v5884; SI-NEXT: v_max_f32_e32 v2, v2, v7885; SI-NEXT: v_cvt_f16_f32_e32 v3, v3886; SI-NEXT: v_max_f32_e32 v1, v1, v6887; SI-NEXT: v_cvt_f16_f32_e32 v2, v2888; SI-NEXT: v_max_f32_e32 v0, v0, v4889; SI-NEXT: v_cvt_f16_f32_e32 v1, v1890; SI-NEXT: v_cvt_f16_f32_e32 v0, v0891; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3892; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2893; SI-NEXT: v_or_b32_e32 v1, v1, v3894; SI-NEXT: v_or_b32_e32 v0, v0, v2895; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0896; SI-NEXT: s_endpgm897;898; VI-LABEL: maxnum_v4f16:899; VI: ; %bb.0: ; %entry900; VI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34901; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24902; VI-NEXT: s_mov_b32 s7, 0xf000903; VI-NEXT: s_mov_b32 s6, -1904; VI-NEXT: s_waitcnt lgkmcnt(0)905; VI-NEXT: s_load_dwordx2 s[8:9], s[8:9], 0x0906; VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0907; VI-NEXT: s_mov_b32 s4, s0908; VI-NEXT: s_mov_b32 s5, s1909; VI-NEXT: s_waitcnt lgkmcnt(0)910; VI-NEXT: v_max_f16_e64 v0, s9, s9911; VI-NEXT: v_max_f16_e64 v1, s3, s3912; VI-NEXT: s_lshr_b32 s0, s9, 16913; VI-NEXT: v_max_f16_e32 v0, v1, v0914; VI-NEXT: v_max_f16_e64 v1, s0, s0915; VI-NEXT: s_lshr_b32 s0, s3, 16916; VI-NEXT: v_max_f16_e64 v2, s0, s0917; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD918; VI-NEXT: v_or_b32_e32 v1, v0, v1919; VI-NEXT: v_max_f16_e64 v0, s8, s8920; VI-NEXT: v_max_f16_e64 v2, s2, s2921; VI-NEXT: s_lshr_b32 s0, s8, 16922; VI-NEXT: v_max_f16_e32 v0, v2, v0923; VI-NEXT: v_max_f16_e64 v2, s0, s0924; VI-NEXT: s_lshr_b32 s0, s2, 16925; VI-NEXT: v_max_f16_e64 v3, s0, s0926; VI-NEXT: v_max_f16_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD927; VI-NEXT: v_or_b32_e32 v0, v0, v2928; VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0929; VI-NEXT: s_endpgm930;931; GFX9-LABEL: maxnum_v4f16:932; GFX9: ; %bb.0: ; %entry933; GFX9-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x34934; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24935; GFX9-NEXT: s_mov_b32 s7, 0xf000936; GFX9-NEXT: s_mov_b32 s6, -1937; GFX9-NEXT: s_waitcnt lgkmcnt(0)938; GFX9-NEXT: s_load_dwordx2 s[10:11], s[8:9], 0x0939; GFX9-NEXT: s_load_dwordx2 s[12:13], s[2:3], 0x0940; GFX9-NEXT: s_mov_b32 s4, s0941; GFX9-NEXT: s_mov_b32 s5, s1942; GFX9-NEXT: s_waitcnt lgkmcnt(0)943; GFX9-NEXT: v_pk_max_f16 v0, s11, s11944; GFX9-NEXT: v_pk_max_f16 v1, s13, s13945; GFX9-NEXT: v_pk_max_f16 v2, s10, s10946; GFX9-NEXT: v_pk_max_f16 v1, v1, v0947; GFX9-NEXT: v_pk_max_f16 v0, s12, s12948; GFX9-NEXT: v_pk_max_f16 v0, v0, v2949; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0950; GFX9-NEXT: s_endpgm951;952; GFX10-LABEL: maxnum_v4f16:953; GFX10: ; %bb.0: ; %entry954; GFX10-NEXT: s_clause 0x1955; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34956; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24957; GFX10-NEXT: s_waitcnt lgkmcnt(0)958; GFX10-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0959; GFX10-NEXT: s_load_dwordx2 s[8:9], s[2:3], 0x0960; GFX10-NEXT: s_mov_b32 s3, 0x31016000961; GFX10-NEXT: s_mov_b32 s2, -1962; GFX10-NEXT: s_waitcnt lgkmcnt(0)963; GFX10-NEXT: v_pk_max_f16 v0, s5, s5964; GFX10-NEXT: v_pk_max_f16 v1, s9, s9965; GFX10-NEXT: v_pk_max_f16 v2, s4, s4966; GFX10-NEXT: v_pk_max_f16 v3, s8, s8967; GFX10-NEXT: v_pk_max_f16 v1, v1, v0968; GFX10-NEXT: v_pk_max_f16 v0, v3, v2969; GFX10-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0970; GFX10-NEXT: s_endpgm971;972; GFX11-LABEL: maxnum_v4f16:973; GFX11: ; %bb.0: ; %entry974; GFX11-NEXT: s_clause 0x1975; GFX11-NEXT: s_load_b64 s[6:7], s[4:5], 0x34976; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24977; GFX11-NEXT: s_waitcnt lgkmcnt(0)978; GFX11-NEXT: s_load_b64 s[4:5], s[6:7], 0x0979; GFX11-NEXT: s_load_b64 s[2:3], s[2:3], 0x0980; GFX11-NEXT: s_waitcnt lgkmcnt(0)981; GFX11-NEXT: v_pk_max_f16 v0, s5, s5982; GFX11-NEXT: v_pk_max_f16 v1, s3, s3983; GFX11-NEXT: v_pk_max_f16 v2, s4, s4984; GFX11-NEXT: v_pk_max_f16 v3, s2, s2985; GFX11-NEXT: s_mov_b32 s3, 0x31016000986; GFX11-NEXT: s_mov_b32 s2, -1987; GFX11-NEXT: v_pk_max_f16 v1, v1, v0988; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)989; GFX11-NEXT: v_pk_max_f16 v0, v3, v2990; GFX11-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0991; GFX11-NEXT: s_endpgm992 ptr addrspace(1) %r,993 ptr addrspace(1) %a,994 ptr addrspace(1) %b) #0 {995entry:996 %a.val = load <4 x half>, ptr addrspace(1) %a997 %b.val = load <4 x half>, ptr addrspace(1) %b998 %r.val = call <4 x half> @llvm.maxnum.v4f16(<4 x half> %a.val, <4 x half> %b.val)999 store <4 x half> %r.val, ptr addrspace(1) %r1000 ret void1001}1002 1003define amdgpu_kernel void @fmax_v4f16_imm_a(1004; SI-LABEL: fmax_v4f16_imm_a:1005; SI: ; %bb.0: ; %entry1006; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x91007; SI-NEXT: s_waitcnt lgkmcnt(0)1008; SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x01009; SI-NEXT: s_mov_b32 s3, 0xf0001010; SI-NEXT: s_mov_b32 s2, -11011; SI-NEXT: s_waitcnt lgkmcnt(0)1012; SI-NEXT: v_cvt_f32_f16_e32 v1, s51013; SI-NEXT: s_lshr_b32 s5, s5, 161014; SI-NEXT: v_cvt_f32_f16_e32 v0, s41015; SI-NEXT: s_lshr_b32 s4, s4, 161016; SI-NEXT: v_cvt_f32_f16_e32 v2, s51017; SI-NEXT: v_cvt_f32_f16_e32 v3, s41018; SI-NEXT: v_max_f32_e32 v1, 0x40400000, v11019; SI-NEXT: v_max_f32_e32 v0, 0x41000000, v01020; SI-NEXT: v_max_f32_e32 v2, 4.0, v21021; SI-NEXT: v_cvt_f16_f32_e32 v2, v21022; SI-NEXT: v_max_f32_e32 v3, 2.0, v31023; SI-NEXT: v_cvt_f16_f32_e32 v1, v11024; SI-NEXT: v_cvt_f16_f32_e32 v3, v31025; SI-NEXT: v_cvt_f16_f32_e32 v0, v01026; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v21027; SI-NEXT: v_or_b32_e32 v1, v1, v21028; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v31029; SI-NEXT: v_or_b32_e32 v0, v0, v21030; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 01031; SI-NEXT: s_endpgm1032;1033; VI-LABEL: fmax_v4f16_imm_a:1034; VI: ; %bb.0: ; %entry1035; VI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x241036; VI-NEXT: v_mov_b32_e32 v0, 0x44001037; VI-NEXT: s_mov_b32 s7, 0xf0001038; VI-NEXT: s_mov_b32 s6, -11039; VI-NEXT: s_waitcnt lgkmcnt(0)1040; VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x01041; VI-NEXT: s_mov_b32 s4, s01042; VI-NEXT: s_mov_b32 s5, s11043; VI-NEXT: s_waitcnt lgkmcnt(0)1044; VI-NEXT: s_lshr_b32 s0, s3, 161045; VI-NEXT: v_max_f16_e64 v1, s3, s31046; VI-NEXT: v_max_f16_e64 v3, s0, s01047; VI-NEXT: v_max_f16_e64 v2, s2, s21048; VI-NEXT: v_max_f16_e32 v1, 0x4200, v11049; VI-NEXT: v_max_f16_sdwa v0, v3, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD1050; VI-NEXT: s_lshr_b32 s0, s2, 161051; VI-NEXT: v_or_b32_e32 v1, v1, v01052; VI-NEXT: v_max_f16_e32 v0, 0x4800, v21053; VI-NEXT: v_max_f16_e64 v2, s0, s01054; VI-NEXT: v_mov_b32_e32 v3, 0x40001055; VI-NEXT: v_max_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD1056; VI-NEXT: v_or_b32_e32 v0, v0, v21057; VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 01058; VI-NEXT: s_endpgm1059;1060; GFX9-LABEL: fmax_v4f16_imm_a:1061; GFX9: ; %bb.0: ; %entry1062; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x241063; GFX9-NEXT: s_mov_b32 s8, 0x440042001064; GFX9-NEXT: s_mov_b32 s9, 0x400048001065; GFX9-NEXT: s_mov_b32 s7, 0xf0001066; GFX9-NEXT: s_mov_b32 s6, -11067; GFX9-NEXT: s_waitcnt lgkmcnt(0)1068; GFX9-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x01069; GFX9-NEXT: s_mov_b32 s4, s01070; GFX9-NEXT: s_mov_b32 s5, s11071; GFX9-NEXT: s_waitcnt lgkmcnt(0)1072; GFX9-NEXT: v_pk_max_f16 v0, s3, s31073; GFX9-NEXT: v_pk_max_f16 v2, s2, s21074; GFX9-NEXT: v_pk_max_f16 v1, v0, s81075; GFX9-NEXT: v_pk_max_f16 v0, v2, s91076; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 01077; GFX9-NEXT: s_endpgm1078;1079; GFX10-LABEL: fmax_v4f16_imm_a:1080; GFX10: ; %bb.0: ; %entry1081; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x241082; GFX10-NEXT: s_waitcnt lgkmcnt(0)1083; GFX10-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x01084; GFX10-NEXT: s_waitcnt lgkmcnt(0)1085; GFX10-NEXT: v_pk_max_f16 v0, s3, s31086; GFX10-NEXT: v_pk_max_f16 v2, s2, s21087; GFX10-NEXT: s_mov_b32 s3, 0x310160001088; GFX10-NEXT: s_mov_b32 s2, -11089; GFX10-NEXT: v_pk_max_f16 v1, 0x44004200, v01090; GFX10-NEXT: v_pk_max_f16 v0, 0x40004800, v21091; GFX10-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 01092; GFX10-NEXT: s_endpgm1093;1094; GFX11-LABEL: fmax_v4f16_imm_a:1095; GFX11: ; %bb.0: ; %entry1096; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x241097; GFX11-NEXT: s_waitcnt lgkmcnt(0)1098; GFX11-NEXT: s_load_b64 s[2:3], s[2:3], 0x01099; GFX11-NEXT: s_waitcnt lgkmcnt(0)1100; GFX11-NEXT: v_pk_max_f16 v0, s3, s31101; GFX11-NEXT: v_pk_max_f16 v2, s2, s21102; GFX11-NEXT: s_mov_b32 s3, 0x310160001103; GFX11-NEXT: s_mov_b32 s2, -11104; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)1105; GFX11-NEXT: v_pk_max_f16 v1, 0x44004200, v01106; GFX11-NEXT: v_pk_max_f16 v0, 0x40004800, v21107; GFX11-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 01108; GFX11-NEXT: s_endpgm1109 ptr addrspace(1) %r,1110 ptr addrspace(1) %b) #0 {1111entry:1112 %b.val = load <4 x half>, ptr addrspace(1) %b1113 %r.val = call <4 x half> @llvm.maxnum.v4f16(<4 x half> <half 8.0, half 2.0, half 3.0, half 4.0>, <4 x half> %b.val)1114 store <4 x half> %r.val, ptr addrspace(1) %r1115 ret void1116}1117 1118attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" }1119