1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s 3; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s 4; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10 %s 5; FIXME: promotion not handled without f16 insts 6 7define half @v_constained_fsub_f16_fpexcept_strict(half %x, half %y) #0 { 8; GCN-LABEL: v_constained_fsub_f16_fpexcept_strict: 9; GCN: ; %bb.0: 10; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 11; GCN-NEXT: v_sub_f16_e32 v0, v0, v1 12; GCN-NEXT: s_setpc_b64 s[30:31] 13; 14; GFX10-LABEL: v_constained_fsub_f16_fpexcept_strict: 15; GFX10: ; %bb.0: 16; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 17; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 18; GFX10-NEXT: v_sub_f16_e32 v0, v0, v1 19; GFX10-NEXT: s_setpc_b64 s[30:31] 20 %val = call half @llvm.experimental.constrained.fsub.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 21 ret half %val 22} 23 24define half @v_constained_fsub_f16_fpexcept_ignore(half %x, half %y) #0 { 25; GCN-LABEL: v_constained_fsub_f16_fpexcept_ignore: 26; GCN: ; %bb.0: 27; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 28; GCN-NEXT: v_sub_f16_e32 v0, v0, v1 29; GCN-NEXT: s_setpc_b64 s[30:31] 30; 31; GFX10-LABEL: v_constained_fsub_f16_fpexcept_ignore: 32; GFX10: ; %bb.0: 33; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 34; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 35; GFX10-NEXT: v_sub_f16_e32 v0, v0, v1 36; GFX10-NEXT: s_setpc_b64 s[30:31] 37 %val = call half @llvm.experimental.constrained.fsub.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.ignore") 38 ret half %val 39} 40 41define half @v_constained_fsub_f16_fpexcept_maytrap(half %x, half %y) #0 { 42; GCN-LABEL: v_constained_fsub_f16_fpexcept_maytrap: 43; GCN: ; %bb.0: 44; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 45; GCN-NEXT: v_sub_f16_e32 v0, v0, v1 46; GCN-NEXT: s_setpc_b64 s[30:31] 47; 48; GFX10-LABEL: v_constained_fsub_f16_fpexcept_maytrap: 49; GFX10: ; %bb.0: 50; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 51; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 52; GFX10-NEXT: v_sub_f16_e32 v0, v0, v1 53; GFX10-NEXT: s_setpc_b64 s[30:31] 54 %val = call half @llvm.experimental.constrained.fsub.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.maytrap") 55 ret half %val 56} 57 58define <2 x half> @v_constained_fsub_v2f16_fpexcept_strict(<2 x half> %x, <2 x half> %y) #0 { 59; GFX9-LABEL: v_constained_fsub_v2f16_fpexcept_strict: 60; GFX9: ; %bb.0: 61; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 62; GFX9-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 63; GFX9-NEXT: v_sub_f16_e32 v0, v0, v1 64; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 65; GFX9-NEXT: v_lshl_or_b32 v0, v2, 16, v0 66; GFX9-NEXT: s_setpc_b64 s[30:31] 67; 68; GFX8-LABEL: v_constained_fsub_v2f16_fpexcept_strict: 69; GFX8: ; %bb.0: 70; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 71; GFX8-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 72; GFX8-NEXT: v_sub_f16_e32 v0, v0, v1 73; GFX8-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 74; GFX8-NEXT: s_setpc_b64 s[30:31] 75; 76; GFX10-LABEL: v_constained_fsub_v2f16_fpexcept_strict: 77; GFX10: ; %bb.0: 78; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 79; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 80; GFX10-NEXT: v_sub_f16_e32 v2, v0, v1 81; GFX10-NEXT: v_sub_f16_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 82; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v2 83; GFX10-NEXT: v_lshl_or_b32 v0, v0, 16, v1 84; GFX10-NEXT: s_setpc_b64 s[30:31] 85 %val = call <2 x half> @llvm.experimental.constrained.fsub.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 86 ret <2 x half> %val 87} 88 89define <2 x half> @v_constained_fsub_v2f16_fpexcept_ignore(<2 x half> %x, <2 x half> %y) #0 { 90; GFX9-LABEL: v_constained_fsub_v2f16_fpexcept_ignore: 91; GFX9: ; %bb.0: 92; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 93; GFX9-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 94; GFX9-NEXT: v_sub_f16_e32 v0, v0, v1 95; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 96; GFX9-NEXT: v_lshl_or_b32 v0, v2, 16, v0 97; GFX9-NEXT: s_setpc_b64 s[30:31] 98; 99; GFX8-LABEL: v_constained_fsub_v2f16_fpexcept_ignore: 100; GFX8: ; %bb.0: 101; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 102; GFX8-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 103; GFX8-NEXT: v_sub_f16_e32 v0, v0, v1 104; GFX8-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 105; GFX8-NEXT: s_setpc_b64 s[30:31] 106; 107; GFX10-LABEL: v_constained_fsub_v2f16_fpexcept_ignore: 108; GFX10: ; %bb.0: 109; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 110; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 111; GFX10-NEXT: v_sub_f16_e32 v2, v0, v1 112; GFX10-NEXT: v_sub_f16_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 113; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v2 114; GFX10-NEXT: v_lshl_or_b32 v0, v0, 16, v1 115; GFX10-NEXT: s_setpc_b64 s[30:31] 116 %val = call <2 x half> @llvm.experimental.constrained.fsub.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.ignore") 117 ret <2 x half> %val 118} 119 120define <2 x half> @v_constained_fsub_v2f16_fpexcept_maytrap(<2 x half> %x, <2 x half> %y) #0 { 121; GFX9-LABEL: v_constained_fsub_v2f16_fpexcept_maytrap: 122; GFX9: ; %bb.0: 123; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 124; GFX9-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 125; GFX9-NEXT: v_sub_f16_e32 v0, v0, v1 126; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 127; GFX9-NEXT: v_lshl_or_b32 v0, v2, 16, v0 128; GFX9-NEXT: s_setpc_b64 s[30:31] 129; 130; GFX8-LABEL: v_constained_fsub_v2f16_fpexcept_maytrap: 131; GFX8: ; %bb.0: 132; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 133; GFX8-NEXT: v_sub_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 134; GFX8-NEXT: v_sub_f16_e32 v0, v0, v1 135; GFX8-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 136; GFX8-NEXT: s_setpc_b64 s[30:31] 137; 138; GFX10-LABEL: v_constained_fsub_v2f16_fpexcept_maytrap: 139; GFX10: ; %bb.0: 140; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 141; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 142; GFX10-NEXT: v_sub_f16_e32 v2, v0, v1 143; GFX10-NEXT: v_sub_f16_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 144; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v2 145; GFX10-NEXT: v_lshl_or_b32 v0, v0, 16, v1 146; GFX10-NEXT: s_setpc_b64 s[30:31] 147 %val = call <2 x half> @llvm.experimental.constrained.fsub.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.maytrap") 148 ret <2 x half> %val 149} 150 151define <3 x half> @v_constained_fsub_v3f16_fpexcept_strict(<3 x half> %x, <3 x half> %y) #0 { 152; GFX9-LABEL: v_constained_fsub_v3f16_fpexcept_strict: 153; GFX9: ; %bb.0: 154; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 155; GFX9-NEXT: v_sub_f16_sdwa v4, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 156; GFX9-NEXT: v_sub_f16_e32 v0, v0, v2 157; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 158; GFX9-NEXT: v_lshl_or_b32 v0, v4, 16, v0 159; GFX9-NEXT: v_sub_f16_e32 v1, v1, v3 160; GFX9-NEXT: s_setpc_b64 s[30:31] 161; 162; GFX8-LABEL: v_constained_fsub_v3f16_fpexcept_strict: 163; GFX8: ; %bb.0: 164; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 165; GFX8-NEXT: v_sub_f16_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 166; GFX8-NEXT: v_sub_f16_e32 v0, v0, v2 167; GFX8-NEXT: v_or_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 168; GFX8-NEXT: v_sub_f16_e32 v1, v1, v3 169; GFX8-NEXT: s_setpc_b64 s[30:31] 170; 171; GFX10-LABEL: v_constained_fsub_v3f16_fpexcept_strict: 172; GFX10: ; %bb.0: 173; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 174; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 175; GFX10-NEXT: v_sub_f16_e32 v4, v0, v2 176; GFX10-NEXT: v_sub_f16_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 177; GFX10-NEXT: v_sub_f16_e32 v1, v1, v3 178; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v4 179; GFX10-NEXT: v_lshl_or_b32 v0, v0, 16, v2 180; GFX10-NEXT: s_setpc_b64 s[30:31] 181 %val = call <3 x half> @llvm.experimental.constrained.fsub.v3f16(<3 x half> %x, <3 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 182 ret <3 x half> %val 183} 184 185; FIXME: Scalarized 186define <4 x half> @v_constained_fsub_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y) #0 { 187; GFX9-LABEL: v_constained_fsub_v4f16_fpexcept_strict: 188; GFX9: ; %bb.0: 189; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 190; GFX9-NEXT: v_sub_f16_sdwa v4, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 191; GFX9-NEXT: v_sub_f16_sdwa v5, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 192; GFX9-NEXT: v_sub_f16_e32 v0, v0, v2 193; GFX9-NEXT: v_mov_b32_e32 v2, 0xffff 194; GFX9-NEXT: v_sub_f16_e32 v1, v1, v3 195; GFX9-NEXT: v_and_b32_e32 v0, v2, v0 196; GFX9-NEXT: v_and_b32_e32 v1, v2, v1 197; GFX9-NEXT: v_lshl_or_b32 v0, v5, 16, v0 198; GFX9-NEXT: v_lshl_or_b32 v1, v4, 16, v1 199; GFX9-NEXT: s_setpc_b64 s[30:31] 200; 201; GFX8-LABEL: v_constained_fsub_v4f16_fpexcept_strict: 202; GFX8: ; %bb.0: 203; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 204; GFX8-NEXT: v_sub_f16_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 205; GFX8-NEXT: v_sub_f16_e32 v1, v1, v3 206; GFX8-NEXT: v_sub_f16_sdwa v5, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 207; GFX8-NEXT: v_sub_f16_e32 v0, v0, v2 208; GFX8-NEXT: v_or_b32_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 209; GFX8-NEXT: v_or_b32_sdwa v1, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 210; GFX8-NEXT: s_setpc_b64 s[30:31] 211; 212; GFX10-LABEL: v_constained_fsub_v4f16_fpexcept_strict: 213; GFX10: ; %bb.0: 214; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 215; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 216; GFX10-NEXT: v_sub_f16_e32 v4, v0, v2 217; GFX10-NEXT: v_sub_f16_e32 v6, v1, v3 218; GFX10-NEXT: v_mov_b32_e32 v5, 0xffff 219; GFX10-NEXT: v_sub_f16_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 220; GFX10-NEXT: v_sub_f16_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 221; GFX10-NEXT: v_and_b32_e32 v2, v5, v4 222; GFX10-NEXT: v_and_b32_e32 v3, v5, v6 223; GFX10-NEXT: v_lshl_or_b32 v0, v0, 16, v2 224; GFX10-NEXT: v_lshl_or_b32 v1, v1, 16, v3 225; GFX10-NEXT: s_setpc_b64 s[30:31] 226 %val = call <4 x half> @llvm.experimental.constrained.fsub.v4f16(<4 x half> %x, <4 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 227 ret <4 x half> %val 228} 229 230define amdgpu_ps half @s_constained_fsub_f16_fpexcept_strict(half inreg %x, half inreg %y) #0 { 231; GCN-LABEL: s_constained_fsub_f16_fpexcept_strict: 232; GCN: ; %bb.0: 233; GCN-NEXT: v_mov_b32_e32 v0, s3 234; GCN-NEXT: v_sub_f16_e32 v0, s2, v0 235; GCN-NEXT: ; return to shader part epilog 236; 237; GFX10-LABEL: s_constained_fsub_f16_fpexcept_strict: 238; GFX10: ; %bb.0: 239; GFX10-NEXT: v_sub_f16_e64 v0, s2, s3 240; GFX10-NEXT: ; return to shader part epilog 241 %val = call half @llvm.experimental.constrained.fsub.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 242 ret half %val 243} 244 245define amdgpu_ps <2 x half> @s_constained_fsub_v2f16_fpexcept_strict(<2 x half> inreg %x, <2 x half> inreg %y) #0 { 246; GFX9-LABEL: s_constained_fsub_v2f16_fpexcept_strict: 247; GFX9: ; %bb.0: 248; GFX9-NEXT: v_mov_b32_e32 v1, s3 249; GFX9-NEXT: s_lshr_b32 s0, s3, 16 250; GFX9-NEXT: v_sub_f16_e32 v1, s2, v1 251; GFX9-NEXT: s_lshr_b32 s1, s2, 16 252; GFX9-NEXT: v_mov_b32_e32 v0, s0 253; GFX9-NEXT: v_sub_f16_e32 v0, s1, v0 254; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1 255; GFX9-NEXT: v_lshl_or_b32 v0, v0, 16, v1 256; GFX9-NEXT: ; return to shader part epilog 257; 258; GFX8-LABEL: s_constained_fsub_v2f16_fpexcept_strict: 259; GFX8: ; %bb.0: 260; GFX8-NEXT: s_lshr_b32 s0, s3, 16 261; GFX8-NEXT: s_lshr_b32 s1, s2, 16 262; GFX8-NEXT: v_mov_b32_e32 v0, s0 263; GFX8-NEXT: v_mov_b32_e32 v1, s1 264; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 265; GFX8-NEXT: v_mov_b32_e32 v1, s3 266; GFX8-NEXT: v_sub_f16_e32 v1, s2, v1 267; GFX8-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 268; GFX8-NEXT: ; return to shader part epilog 269; 270; GFX10-LABEL: s_constained_fsub_v2f16_fpexcept_strict: 271; GFX10: ; %bb.0: 272; GFX10-NEXT: v_sub_f16_e64 v0, s2, s3 273; GFX10-NEXT: s_lshr_b32 s0, s3, 16 274; GFX10-NEXT: s_lshr_b32 s1, s2, 16 275; GFX10-NEXT: v_sub_f16_e64 v1, s1, s0 276; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0 277; GFX10-NEXT: v_lshl_or_b32 v0, v1, 16, v0 278; GFX10-NEXT: ; return to shader part epilog 279 %val = call <2 x half> @llvm.experimental.constrained.fsub.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict") 280 ret <2 x half> %val 281} 282 283declare half @llvm.experimental.constrained.fsub.f16(half, half, metadata, metadata) #1 284declare <2 x half> @llvm.experimental.constrained.fsub.v2f16(<2 x half>, <2 x half>, metadata, metadata) #1 285declare <3 x half> @llvm.experimental.constrained.fsub.v3f16(<3 x half>, <3 x half>, metadata, metadata) #1 286declare <4 x half> @llvm.experimental.constrained.fsub.v4f16(<4 x half>, <4 x half>, metadata, metadata) #1 287 288attributes #0 = { strictfp } 289attributes #1 = { inaccessiblememonly nounwind willreturn } 290