1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 2; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,SI %s 3; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx803 < %s | FileCheck -check-prefixes=GCN,GFX89,GFX8 %s 4; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX89,GFX9 %s 5; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GCN,GFX1011,GFX10 %s 6; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GCN,GFX1011,GFX11 %s 7 8define half @v_constrained_fptrunc_f32_to_f16_fpexcept_strict(float %arg) #0 { 9; SI-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict: 10; SI: ; %bb.0: 11; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 12; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 13; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 14; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 15; SI-NEXT: s_setpc_b64 s[30:31] 16; 17; GFX89-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict: 18; GFX89: ; %bb.0: 19; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 20; GFX89-NEXT: v_cvt_f16_f32_e32 v0, v0 21; GFX89-NEXT: s_setpc_b64 s[30:31] 22; 23; GFX1011-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict: 24; GFX1011: ; %bb.0: 25; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 26; GFX1011-NEXT: v_cvt_f16_f32_e32 v0, v0 27; GFX1011-NEXT: s_setpc_b64 s[30:31] 28 %val = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 29 ret half %val 30} 31 32define <2 x half> @v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict(<2 x float> %arg) #0 { 33; SI-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict: 34; SI: ; %bb.0: 35; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 36; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 37; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 38; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 39; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 40; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 41; SI-NEXT: v_cvt_f32_f16_e32 v1, v1 42; SI-NEXT: s_setpc_b64 s[30:31] 43; 44; GFX8-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict: 45; GFX8: ; %bb.0: 46; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 47; GFX8-NEXT: v_cvt_f16_f32_e32 v0, v0 48; GFX8-NEXT: v_cvt_f16_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD 49; GFX8-NEXT: v_or_b32_e32 v0, v0, v1 50; GFX8-NEXT: s_setpc_b64 s[30:31] 51; 52; GFX9-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict: 53; GFX9: ; %bb.0: 54; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 55; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 56; GFX9-NEXT: v_cvt_f16_f32_e32 v1, v1 57; GFX9-NEXT: s_mov_b32 s4, 0x5040100 58; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4 59; GFX9-NEXT: s_setpc_b64 s[30:31] 60; 61; GFX1011-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict: 62; GFX1011: ; %bb.0: 63; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 64; GFX1011-NEXT: v_cvt_f16_f32_e32 v0, v0 65; GFX1011-NEXT: v_cvt_f16_f32_e32 v1, v1 66; GFX1011-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 67; GFX1011-NEXT: s_setpc_b64 s[30:31] 68 %val = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 69 ret <2 x half> %val 70} 71 72define <3 x half> @v_constrained_fptrunc_v3f32_to_v3f16_fpexcept_strict(<3 x float> %arg) #0 { 73; SI-LABEL: v_constrained_fptrunc_v3f32_to_v3f16_fpexcept_strict: 74; SI: ; %bb.0: 75; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 76; SI-NEXT: v_cvt_f16_f32_e32 v2, v2 77; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 78; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 79; SI-NEXT: v_and_b32_e32 v2, 0xffff, v2 80; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 81; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 82; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 83; SI-NEXT: v_cvt_f32_f16_e32 v1, v1 84; SI-NEXT: v_cvt_f32_f16_e32 v2, v2 85; SI-NEXT: s_setpc_b64 s[30:31] 86; 87; GFX8-LABEL: v_constrained_fptrunc_v3f32_to_v3f16_fpexcept_strict: 88; GFX8: ; %bb.0: 89; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 90; GFX8-NEXT: v_cvt_f16_f32_sdwa v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD 91; GFX8-NEXT: v_cvt_f16_f32_e32 v0, v0 92; GFX8-NEXT: v_cvt_f16_f32_e32 v1, v2 93; GFX8-NEXT: v_or_b32_e32 v0, v0, v3 94; GFX8-NEXT: s_setpc_b64 s[30:31] 95; 96; GFX9-LABEL: v_constrained_fptrunc_v3f32_to_v3f16_fpexcept_strict: 97; GFX9: ; %bb.0: 98; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 99; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 100; GFX9-NEXT: v_cvt_f16_f32_e32 v3, v1 101; GFX9-NEXT: v_cvt_f16_f32_e32 v1, v2 102; GFX9-NEXT: s_mov_b32 s4, 0x5040100 103; GFX9-NEXT: v_perm_b32 v0, v3, v0, s4 104; GFX9-NEXT: s_setpc_b64 s[30:31] 105; 106; GFX1011-LABEL: v_constrained_fptrunc_v3f32_to_v3f16_fpexcept_strict: 107; GFX1011: ; %bb.0: 108; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 109; GFX1011-NEXT: v_cvt_f16_f32_e32 v0, v0 110; GFX1011-NEXT: v_cvt_f16_f32_e32 v1, v1 111; GFX1011-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 112; GFX1011-NEXT: v_cvt_f16_f32_e32 v1, v2 113; GFX1011-NEXT: s_setpc_b64 s[30:31] 114 %val = call <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 115 ret <3 x half> %val 116} 117 118define float @v_constrained_fptrunc_f64_to_f32_fpexcept_strict(double %arg) #0 { 119; GCN-LABEL: v_constrained_fptrunc_f64_to_f32_fpexcept_strict: 120; GCN: ; %bb.0: 121; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 122; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] 123; GCN-NEXT: s_setpc_b64 s[30:31] 124 %val = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 125 ret float %val 126} 127 128define <2 x float> @v_constrained_fptrunc_v2f64_to_v2f32_fpexcept_strict(<2 x double> %arg) #0 { 129; GCN-LABEL: v_constrained_fptrunc_v2f64_to_v2f32_fpexcept_strict: 130; GCN: ; %bb.0: 131; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 132; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] 133; GCN-NEXT: v_cvt_f32_f64_e32 v1, v[2:3] 134; GCN-NEXT: s_setpc_b64 s[30:31] 135 %val = call <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(<2 x double> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 136 ret <2 x float> %val 137} 138 139define <3 x float> @v_constrained_fptrunc_v3f64_to_v3f32_fpexcept_strict(<3 x double> %arg) #0 { 140; GCN-LABEL: v_constrained_fptrunc_v3f64_to_v3f32_fpexcept_strict: 141; GCN: ; %bb.0: 142; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 143; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] 144; GCN-NEXT: v_cvt_f32_f64_e32 v1, v[2:3] 145; GCN-NEXT: v_cvt_f32_f64_e32 v2, v[4:5] 146; GCN-NEXT: s_setpc_b64 s[30:31] 147 %val = call <3 x float> @llvm.experimental.constrained.fptrunc.v3f32.v3f64(<3 x double> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 148 ret <3 x float> %val 149} 150 151; FIXME: 152; define half @v_constrained_fptrunc_f64_to_f16_fpexcept_strict(double %arg) #0 { 153; %val = call half @llvm.experimental.constrained.fptrunc.f16.f64(double %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 154; ret half %val 155; } 156 157; define <2 x half> @v_constrained_fptrunc_v2f64_to_v2f16_fpexcept_strict(<2 x double> %arg) #0 { 158; %val = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(<2 x double> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 159; ret <2 x half> %val 160; } 161 162; define <3 x half> @v_constrained_fptrunc_v3f64_to_v3f16_fpexcept_strict(<3 x double> %arg) #0 { 163; %val = call <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f64(<3 x double> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 164; ret <3 x half> %val 165; } 166 167define half @v_constrained_fneg_fptrunc_f32_to_f16_fpexcept_strict(float %arg) #0 { 168; SI-LABEL: v_constrained_fneg_fptrunc_f32_to_f16_fpexcept_strict: 169; SI: ; %bb.0: 170; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 171; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 172; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 173; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 174; SI-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 175; SI-NEXT: s_setpc_b64 s[30:31] 176; 177; GFX89-LABEL: v_constrained_fneg_fptrunc_f32_to_f16_fpexcept_strict: 178; GFX89: ; %bb.0: 179; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 180; GFX89-NEXT: v_cvt_f16_f32_e32 v0, v0 181; GFX89-NEXT: v_xor_b32_e32 v0, 0x8000, v0 182; GFX89-NEXT: s_setpc_b64 s[30:31] 183; 184; GFX1011-LABEL: v_constrained_fneg_fptrunc_f32_to_f16_fpexcept_strict: 185; GFX1011: ; %bb.0: 186; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 187; GFX1011-NEXT: v_cvt_f16_f32_e32 v0, v0 188; GFX1011-NEXT: v_xor_b32_e32 v0, 0x8000, v0 189; GFX1011-NEXT: s_setpc_b64 s[30:31] 190 %val = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 191 %neg.val = fneg half %val 192 ret half %neg.val 193} 194 195define half @v_constrained_fptrunc_fneg_f32_to_f16_fpexcept_strict(float %arg) #0 { 196; SI-LABEL: v_constrained_fptrunc_fneg_f32_to_f16_fpexcept_strict: 197; SI: ; %bb.0: 198; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 199; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0 200; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 201; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 202; SI-NEXT: s_setpc_b64 s[30:31] 203; 204; GFX89-LABEL: v_constrained_fptrunc_fneg_f32_to_f16_fpexcept_strict: 205; GFX89: ; %bb.0: 206; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 207; GFX89-NEXT: v_cvt_f16_f32_e64 v0, -v0 208; GFX89-NEXT: s_setpc_b64 s[30:31] 209; 210; GFX1011-LABEL: v_constrained_fptrunc_fneg_f32_to_f16_fpexcept_strict: 211; GFX1011: ; %bb.0: 212; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 213; GFX1011-NEXT: v_cvt_f16_f32_e64 v0, -v0 214; GFX1011-NEXT: s_setpc_b64 s[30:31] 215 %neg.arg = fneg float %arg 216 %val = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %neg.arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 217 ret half %val 218} 219 220define float @v_constrained_fneg_fptrunc_f64_to_f32_fpexcept_strict(double %arg) #0 { 221; GCN-LABEL: v_constrained_fneg_fptrunc_f64_to_f32_fpexcept_strict: 222; GCN: ; %bb.0: 223; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 224; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] 225; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 226; GCN-NEXT: s_setpc_b64 s[30:31] 227 %val = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 228 %neg.val = fneg float %val 229 ret float %neg.val 230} 231 232define float @v_constrained_fptrunc_fneg_f64_to_f32_fpexcept_strict(double %arg) #0 { 233; GCN-LABEL: v_constrained_fptrunc_fneg_f64_to_f32_fpexcept_strict: 234; GCN: ; %bb.0: 235; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 236; GCN-NEXT: v_cvt_f32_f64_e64 v0, -v[0:1] 237; GCN-NEXT: s_setpc_b64 s[30:31] 238 %neg.arg = fneg double %arg 239 %val = call float @llvm.experimental.constrained.fptrunc.f32.f64(double %neg.arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 240 ret float %val 241} 242 243define void @v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi(float %arg, ptr addrspace(1) %ptr) #0 { 244; SI-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi: 245; SI: ; %bb.0: 246; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 247; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 248; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 249; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 250; SI-NEXT: s_setpc_b64 s[30:31] 251; 252; GFX89-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi: 253; GFX89: ; %bb.0: 254; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 255; GFX89-NEXT: v_cvt_f16_f32_e32 v0, v0 256; GFX89-NEXT: s_setpc_b64 s[30:31] 257; 258; GFX1011-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi: 259; GFX1011: ; %bb.0: 260; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 261; GFX1011-NEXT: v_cvt_f16_f32_e32 v0, v0 262; GFX1011-NEXT: s_setpc_b64 s[30:31] 263 %result = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 264 ret void 265} 266 267define void @v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi(<2 x float> %arg, ptr addrspace(1) %ptr) #0 { 268; SI-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi: 269; SI: ; %bb.0: 270; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 271; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 272; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 273; SI-NEXT: s_mov_b32 s6, 0 274; SI-NEXT: s_mov_b32 s7, 0xf000 275; SI-NEXT: v_and_b32_e32 v1, 0xffff, v1 276; SI-NEXT: v_cvt_f32_f16_e32 v1, v1 277; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 278; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 279; SI-NEXT: s_mov_b32 s4, s6 280; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 281; SI-NEXT: s_mov_b32 s5, s6 282; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 283; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 284; SI-NEXT: v_or_b32_e32 v0, v0, v1 285; SI-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64 286; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) 287; SI-NEXT: s_setpc_b64 s[30:31] 288; 289; GFX8-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi: 290; GFX8: ; %bb.0: 291; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 292; GFX8-NEXT: v_cvt_f16_f32_e32 v0, v0 293; GFX8-NEXT: v_cvt_f16_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD 294; GFX8-NEXT: v_or_b32_e32 v0, v0, v1 295; GFX8-NEXT: flat_store_dword v[2:3], v0 296; GFX8-NEXT: s_waitcnt vmcnt(0) 297; GFX8-NEXT: s_setpc_b64 s[30:31] 298; 299; GFX9-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi: 300; GFX9: ; %bb.0: 301; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 302; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 303; GFX9-NEXT: v_cvt_f16_f32_e32 v1, v1 304; GFX9-NEXT: s_mov_b32 s4, 0x5040100 305; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4 306; GFX9-NEXT: global_store_dword v[2:3], v0, off 307; GFX9-NEXT: s_waitcnt vmcnt(0) 308; GFX9-NEXT: s_setpc_b64 s[30:31] 309; 310; GFX10-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi: 311; GFX10: ; %bb.0: 312; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 313; GFX10-NEXT: v_cvt_f16_f32_e32 v0, v0 314; GFX10-NEXT: v_cvt_f16_f32_e32 v1, v1 315; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 316; GFX10-NEXT: global_store_dword v[2:3], v0, off 317; GFX10-NEXT: s_setpc_b64 s[30:31] 318; 319; GFX11-LABEL: v_constrained_fptrunc_v2f32_to_v2f16_fpexcept_strict_noabi: 320; GFX11: ; %bb.0: 321; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 322; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 323; GFX11-NEXT: v_cvt_f16_f32_e32 v1, v1 324; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x5040100 325; GFX11-NEXT: global_store_b32 v[2:3], v0, off 326; GFX11-NEXT: s_setpc_b64 s[30:31] 327 %result = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 328 store <2 x half> %result, ptr addrspace(1) %ptr 329 ret void 330} 331 332define void @v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fneg(float %arg, ptr addrspace(1) %ptr) #0 { 333; SI-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fneg: 334; SI: ; %bb.0: 335; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 336; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0 337; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 338; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 339; SI-NEXT: s_setpc_b64 s[30:31] 340; 341; GFX89-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fneg: 342; GFX89: ; %bb.0: 343; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 344; GFX89-NEXT: v_cvt_f16_f32_e64 v0, -v0 345; GFX89-NEXT: s_setpc_b64 s[30:31] 346; 347; GFX1011-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fneg: 348; GFX1011: ; %bb.0: 349; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 350; GFX1011-NEXT: v_cvt_f16_f32_e64 v0, -v0 351; GFX1011-NEXT: s_setpc_b64 s[30:31] 352 %neg.arg = fneg float %arg 353 %result = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %neg.arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 354 ret void 355} 356 357define void @v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fabs(float %arg, ptr addrspace(1) %ptr) #0 { 358; SI-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fabs: 359; SI: ; %bb.0: 360; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 361; SI-NEXT: v_cvt_f16_f32_e64 v0, |v0| 362; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0 363; SI-NEXT: v_cvt_f32_f16_e32 v0, v0 364; SI-NEXT: s_setpc_b64 s[30:31] 365; 366; GFX89-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fabs: 367; GFX89: ; %bb.0: 368; GFX89-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 369; GFX89-NEXT: v_cvt_f16_f32_e64 v0, |v0| 370; GFX89-NEXT: s_setpc_b64 s[30:31] 371; 372; GFX1011-LABEL: v_constrained_fptrunc_f32_to_f16_fpexcept_strict_noabi_fabs: 373; GFX1011: ; %bb.0: 374; GFX1011-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 375; GFX1011-NEXT: v_cvt_f16_f32_e64 v0, |v0| 376; GFX1011-NEXT: s_setpc_b64 s[30:31] 377 %abs.arg = call float @llvm.fabs.f32(float %arg) #0 378 %result = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %abs.arg, metadata !"round.tonearest", metadata !"fpexcept.strict") 379 ret void 380} 381 382declare half @llvm.experimental.constrained.fptrunc.f16.f32(float, metadata, metadata) 383declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float>, metadata, metadata) 384declare <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float>, metadata, metadata) 385 386declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata, metadata) 387declare <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(<2 x double>, metadata, metadata) 388declare <3 x float> @llvm.experimental.constrained.fptrunc.v3f32.v3f64(<3 x double>, metadata, metadata) 389 390declare half @llvm.experimental.constrained.fptrunc.f16.f64(double, metadata, metadata) 391declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(<2 x double>, metadata, metadata) 392declare <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f64(<3 x double>, metadata, metadata) 393 394declare float @llvm.fabs.f32(float) 395 396attributes #0 = { strictfp } 397