805 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX9,NOT-GFX12 %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX10,NOT-GFX12 %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX11,GFX1100,NOT-GFX12 %s5; RUN: llc -mtriple=amdgcn -mcpu=gfx1150 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX11,GFX1150,NOT-GFX12 %s6; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX1200 %s7; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX1250 %s8 9declare float @llvm.fabs.f32(float)10declare float @llvm.fma.f32(float, float, float)11 12define float @v_mul_f32_vop2(float %x, float %y) {13; GFX9-LABEL: v_mul_f32_vop2:14; GFX9: ; %bb.0:15; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]16; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x0a]17; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]18;19; GFX10-LABEL: v_mul_f32_vop2:20; GFX10: ; %bb.0:21; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]22; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x10]23; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]24;25; GFX11-LABEL: v_mul_f32_vop2:26; GFX11: ; %bb.0:27; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]28; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x10]29; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]30;31; GFX1200-LABEL: v_mul_f32_vop2:32; GFX1200: ; %bb.0:33; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]34; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]35; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]36; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]37; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]38; GFX1200-NEXT: v_mul_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x10]39; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]40;41; GFX1250-LABEL: v_mul_f32_vop2:42; GFX1250: ; %bb.0:43; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]44; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]45; GFX1250-NEXT: v_mul_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x10]46; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]47 %mul = fmul float %x, %y48 ret float %mul49}50; NOT-GFX12: codeLenInByte = 1251; GFX1200: codeLenInByte = 2852; GFX1250: codeLenInByte = 1653 54define float @v_mul_f32_vop2_inline_imm(float %x) {55; GFX9-LABEL: v_mul_f32_vop2_inline_imm:56; GFX9: ; %bb.0:57; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]58; GFX9-NEXT: v_mul_f32_e32 v0, 4.0, v0 ; encoding: [0xf6,0x00,0x00,0x0a]59; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]60;61; GFX10-LABEL: v_mul_f32_vop2_inline_imm:62; GFX10: ; %bb.0:63; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]64; GFX10-NEXT: v_mul_f32_e32 v0, 4.0, v0 ; encoding: [0xf6,0x00,0x00,0x10]65; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]66;67; GFX11-LABEL: v_mul_f32_vop2_inline_imm:68; GFX11: ; %bb.0:69; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]70; GFX11-NEXT: v_mul_f32_e32 v0, 4.0, v0 ; encoding: [0xf6,0x00,0x00,0x10]71; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]72;73; GFX1200-LABEL: v_mul_f32_vop2_inline_imm:74; GFX1200: ; %bb.0:75; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]76; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]77; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]78; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]79; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]80; GFX1200-NEXT: v_mul_f32_e32 v0, 4.0, v0 ; encoding: [0xf6,0x00,0x00,0x10]81; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]82;83; GFX1250-LABEL: v_mul_f32_vop2_inline_imm:84; GFX1250: ; %bb.0:85; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]86; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]87; GFX1250-NEXT: v_mul_f32_e32 v0, 4.0, v0 ; encoding: [0xf6,0x00,0x00,0x10]88; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]89 %mul = fmul float %x, 4.090 ret float %mul91}92; NOT-GFX12: codeLenInByte = 1293; GFX1200: codeLenInByte = 2894; GFX1250: codeLenInByte = 1695 96define float @v_mul_f32_vop2_literal(float %x) {97; GFX9-LABEL: v_mul_f32_vop2_literal:98; GFX9: ; %bb.0:99; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]100; GFX9-NEXT: v_mul_f32_e32 v0, 0x42f60000, v0 ; encoding: [0xff,0x00,0x00,0x0a,0x00,0x00,0xf6,0x42]101; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]102;103; GFX10-LABEL: v_mul_f32_vop2_literal:104; GFX10: ; %bb.0:105; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]106; GFX10-NEXT: v_mul_f32_e32 v0, 0x42f60000, v0 ; encoding: [0xff,0x00,0x00,0x10,0x00,0x00,0xf6,0x42]107; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]108;109; GFX11-LABEL: v_mul_f32_vop2_literal:110; GFX11: ; %bb.0:111; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]112; GFX11-NEXT: v_mul_f32_e32 v0, 0x42f60000, v0 ; encoding: [0xff,0x00,0x00,0x10,0x00,0x00,0xf6,0x42]113; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]114;115; GFX1200-LABEL: v_mul_f32_vop2_literal:116; GFX1200: ; %bb.0:117; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]118; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]119; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]120; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]121; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]122; GFX1200-NEXT: v_mul_f32_e32 v0, 0x42f60000, v0 ; encoding: [0xff,0x00,0x00,0x10,0x00,0x00,0xf6,0x42]123; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]124;125; GFX1250-LABEL: v_mul_f32_vop2_literal:126; GFX1250: ; %bb.0:127; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]128; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]129; GFX1250-NEXT: v_mul_f32_e32 v0, 0x42f60000, v0 ; encoding: [0xff,0x00,0x00,0x10,0x00,0x00,0xf6,0x42]130; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]131 %mul = fmul float %x, 123.0132 ret float %mul133}134; NOT-GFX12: codeLenInByte = 16135; GFX1200: codeLenInByte = 32136; GFX1250: codeLenInByte = 20137 138define float @v_mul_f32_vop3_src_mods(float %x, float %y) {139; GFX9-LABEL: v_mul_f32_vop3_src_mods:140; GFX9: ; %bb.0:141; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]142; GFX9-NEXT: v_mul_f32_e64 v0, |v0|, v1 ; encoding: [0x00,0x01,0x05,0xd1,0x00,0x03,0x02,0x00]143; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]144;145; GFX10-LABEL: v_mul_f32_vop3_src_mods:146; GFX10: ; %bb.0:147; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]148; GFX10-NEXT: v_mul_f32_e64 v0, |v0|, v1 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0x03,0x02,0x00]149; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]150;151; GFX11-LABEL: v_mul_f32_vop3_src_mods:152; GFX11: ; %bb.0:153; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]154; GFX11-NEXT: v_mul_f32_e64 v0, |v0|, v1 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0x03,0x02,0x00]155; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]156;157; GFX1200-LABEL: v_mul_f32_vop3_src_mods:158; GFX1200: ; %bb.0:159; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]160; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]161; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]162; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]163; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]164; GFX1200-NEXT: v_mul_f32_e64 v0, |v0|, v1 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0x03,0x02,0x00]165; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]166;167; GFX1250-LABEL: v_mul_f32_vop3_src_mods:168; GFX1250: ; %bb.0:169; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]170; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]171; GFX1250-NEXT: v_mul_f32_e64 v0, |v0|, v1 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0x03,0x02,0x00]172; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]173 %fabs.x = call float @llvm.fabs.f32(float %x)174 %mul = fmul float %fabs.x, %y175 ret float %mul176}177; NOT-GFX12: codeLenInByte = 16178; GFX1200: codeLenInByte = 32179; GFX1250: codeLenInByte = 20180 181define float @v_mul_f32_vop3_src_mods_inline_imm(float %x, float %y) {182; GFX9-LABEL: v_mul_f32_vop3_src_mods_inline_imm:183; GFX9: ; %bb.0:184; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]185; GFX9-NEXT: v_mul_f32_e64 v0, |v0|, 4.0 ; encoding: [0x00,0x01,0x05,0xd1,0x00,0xed,0x01,0x00]186; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]187;188; GFX10-LABEL: v_mul_f32_vop3_src_mods_inline_imm:189; GFX10: ; %bb.0:190; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]191; GFX10-NEXT: v_mul_f32_e64 v0, |v0|, 4.0 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0xed,0x01,0x00]192; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]193;194; GFX11-LABEL: v_mul_f32_vop3_src_mods_inline_imm:195; GFX11: ; %bb.0:196; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]197; GFX11-NEXT: v_mul_f32_e64 v0, |v0|, 4.0 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0xed,0x01,0x00]198; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]199;200; GFX1200-LABEL: v_mul_f32_vop3_src_mods_inline_imm:201; GFX1200: ; %bb.0:202; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]203; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]204; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]205; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]206; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]207; GFX1200-NEXT: v_mul_f32_e64 v0, |v0|, 4.0 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0xed,0x01,0x00]208; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]209;210; GFX1250-LABEL: v_mul_f32_vop3_src_mods_inline_imm:211; GFX1250: ; %bb.0:212; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]213; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]214; GFX1250-NEXT: v_mul_f32_e64 v0, |v0|, 4.0 ; encoding: [0x00,0x01,0x08,0xd5,0x00,0xed,0x01,0x00]215; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]216 %fabs.x = call float @llvm.fabs.f32(float %x)217 %mul = fmul float %fabs.x, 4.0218 ret float %mul219}220 221; NOT-GFX12: codeLenInByte = 16222; GFX1200: codeLenInByte = 32223; GFX1250: codeLenInByte = 20224 225define float @v_mul_f32_vop3_src_mods_literal(float %x, float %y) {226; GFX9-LABEL: v_mul_f32_vop3_src_mods_literal:227; GFX9: ; %bb.0:228; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]229; GFX9-NEXT: s_mov_b32 s4, 0x42f60000 ; encoding: [0xff,0x00,0x84,0xbe,0x00,0x00,0xf6,0x42]230; GFX9-NEXT: v_mul_f32_e64 v0, |v0|, s4 ; encoding: [0x00,0x01,0x05,0xd1,0x00,0x09,0x00,0x00]231; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]232;233; GFX10-LABEL: v_mul_f32_vop3_src_mods_literal:234; GFX10: ; %bb.0:235; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]236; GFX10-NEXT: v_mul_f32_e64 v0, 0x42f60000, |v0| ; encoding: [0x00,0x02,0x08,0xd5,0xff,0x00,0x02,0x00,0x00,0x00,0xf6,0x42]237; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]238;239; GFX11-LABEL: v_mul_f32_vop3_src_mods_literal:240; GFX11: ; %bb.0:241; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]242; GFX11-NEXT: v_mul_f32_e64 v0, 0x42f60000, |v0| ; encoding: [0x00,0x02,0x08,0xd5,0xff,0x00,0x02,0x00,0x00,0x00,0xf6,0x42]243; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]244;245; GFX1200-LABEL: v_mul_f32_vop3_src_mods_literal:246; GFX1200: ; %bb.0:247; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]248; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]249; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]250; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]251; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]252; GFX1200-NEXT: v_mul_f32_e64 v0, 0x42f60000, |v0| ; encoding: [0x00,0x02,0x08,0xd5,0xff,0x00,0x02,0x00,0x00,0x00,0xf6,0x42]253; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]254;255; GFX1250-LABEL: v_mul_f32_vop3_src_mods_literal:256; GFX1250: ; %bb.0:257; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]258; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]259; GFX1250-NEXT: v_mul_f32_e64 v0, 0x42f60000, |v0| ; encoding: [0x00,0x02,0x08,0xd5,0xff,0x00,0x02,0x00,0x00,0x00,0xf6,0x42]260; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]261 %fabs.x = call float @llvm.fabs.f32(float %x)262 %mul = fmul float %fabs.x, 123.0263 ret float %mul264}265 266; GFX9: codeLenInByte = 24267; GFX10: codeLenInByte = 20268; GFX11: codeLenInByte = 20269; GFX1200: codeLenInByte = 36270; GFX1250: codeLenInByte = 24271 272define float @v_mul_f32_vop2_frame_index(float %x) {273; GFX9-LABEL: v_mul_f32_vop2_frame_index:274; GFX9: ; %bb.0:275; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]276; GFX9-NEXT: v_lshrrev_b32_e64 v1, 6, s32 ; encoding: [0x01,0x00,0x10,0xd1,0x86,0x40,0x00,0x00]277; GFX9-NEXT: v_mul_f32_e32 v0, v1, v0 ; encoding: [0x01,0x01,0x00,0x0a]278; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]279;280; GFX10-LABEL: v_mul_f32_vop2_frame_index:281; GFX10: ; %bb.0:282; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]283; GFX10-NEXT: v_lshrrev_b32_e64 v1, 5, s32 ; encoding: [0x01,0x00,0x16,0xd5,0x85,0x40,0x00,0x00]284; GFX10-NEXT: v_mul_f32_e32 v0, v1, v0 ; encoding: [0x01,0x01,0x00,0x10]285; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]286;287; GFX11-LABEL: v_mul_f32_vop2_frame_index:288; GFX11: ; %bb.0:289; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]290; GFX11-NEXT: v_mul_f32_e32 v0, s32, v0 ; encoding: [0x20,0x00,0x00,0x10]291; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]292;293; GFX1200-LABEL: v_mul_f32_vop2_frame_index:294; GFX1200: ; %bb.0:295; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]296; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]297; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]298; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]299; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]300; GFX1200-NEXT: v_mul_f32_e32 v0, s32, v0 ; encoding: [0x20,0x00,0x00,0x10]301; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]302;303; GFX1250-LABEL: v_mul_f32_vop2_frame_index:304; GFX1250: ; %bb.0:305; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]306; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]307; GFX1250-NEXT: v_mul_f32_e32 v0, s32, v0 ; encoding: [0x20,0x00,0x00,0x10]308; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]309 %alloca = alloca i32, addrspace(5)310 %ptrtoint = ptrtoint ptr addrspace(5) %alloca to i32311 %cast = bitcast i32 %ptrtoint to float312 %mul = fmul float %x, %cast313 ret float %mul314}315 316; GFX9: codeLenInByte = 20317; GFX10: codeLenInByte = 20318; GFX11: codeLenInByte = 12319; GFX1200: codeLenInByte = 28320; GFX1250: codeLenInByte = 16321 322define float @v_fma_f32(float %x, float %y, float %z) {323; GFX9-LABEL: v_fma_f32:324; GFX9: ; %bb.0:325; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]326; GFX9-NEXT: v_fma_f32 v0, v0, v1, v2 ; encoding: [0x00,0x00,0xcb,0xd1,0x00,0x03,0x0a,0x04]327; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]328;329; GFX10-LABEL: v_fma_f32:330; GFX10: ; %bb.0:331; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]332; GFX10-NEXT: v_fma_f32 v0, v0, v1, v2 ; encoding: [0x00,0x00,0x4b,0xd5,0x00,0x03,0x0a,0x04]333; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]334;335; GFX11-LABEL: v_fma_f32:336; GFX11: ; %bb.0:337; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]338; GFX11-NEXT: v_fma_f32 v0, v0, v1, v2 ; encoding: [0x00,0x00,0x13,0xd6,0x00,0x03,0x0a,0x04]339; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]340;341; GFX1200-LABEL: v_fma_f32:342; GFX1200: ; %bb.0:343; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]344; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]345; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]346; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]347; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]348; GFX1200-NEXT: v_fma_f32 v0, v0, v1, v2 ; encoding: [0x00,0x00,0x13,0xd6,0x00,0x03,0x0a,0x04]349; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]350;351; GFX1250-LABEL: v_fma_f32:352; GFX1250: ; %bb.0:353; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]354; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]355; GFX1250-NEXT: v_fma_f32 v0, v0, v1, v2 ; encoding: [0x00,0x00,0x13,0xd6,0x00,0x03,0x0a,0x04]356; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]357 %fma = call float @llvm.fma.f32(float %x, float %y, float %z)358 ret float %fma359}360 361; NOT-GFX12: codeLenInByte = 16362; GFX1200: codeLenInByte = 32363; GFX1250: codeLenInByte = 20364 365define float @v_fma_f32_src_mods(float %x, float %y, float %z) {366; GFX9-LABEL: v_fma_f32_src_mods:367; GFX9: ; %bb.0:368; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]369; GFX9-NEXT: v_fma_f32 v0, |v0|, v1, v2 ; encoding: [0x00,0x01,0xcb,0xd1,0x00,0x03,0x0a,0x04]370; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]371;372; GFX10-LABEL: v_fma_f32_src_mods:373; GFX10: ; %bb.0:374; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]375; GFX10-NEXT: v_fma_f32 v0, |v0|, v1, v2 ; encoding: [0x00,0x01,0x4b,0xd5,0x00,0x03,0x0a,0x04]376; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]377;378; GFX11-LABEL: v_fma_f32_src_mods:379; GFX11: ; %bb.0:380; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]381; GFX11-NEXT: v_fma_f32 v0, |v0|, v1, v2 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0x0a,0x04]382; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]383;384; GFX1200-LABEL: v_fma_f32_src_mods:385; GFX1200: ; %bb.0:386; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]387; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]388; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]389; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]390; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]391; GFX1200-NEXT: v_fma_f32 v0, |v0|, v1, v2 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0x0a,0x04]392; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]393;394; GFX1250-LABEL: v_fma_f32_src_mods:395; GFX1250: ; %bb.0:396; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]397; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]398; GFX1250-NEXT: v_fma_f32 v0, |v0|, v1, v2 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0x0a,0x04]399; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]400 %fabs.x = call float @llvm.fabs.f32(float %x)401 %fma = call float @llvm.fma.f32(float %fabs.x, float %y, float %z)402 ret float %fma403}404 405; NOT-GFX12: codeLenInByte = 16406; GFX1200: codeLenInByte = 32407; GFX1250: codeLenInByte = 20408 409define float @v_fmac_f32(float %x, float %y) {410; GFX9-LABEL: v_fmac_f32:411; GFX9: ; %bb.0:412; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]413; GFX9-NEXT: v_fma_f32 v0, v0, v1, v0 ; encoding: [0x00,0x00,0xcb,0xd1,0x00,0x03,0x02,0x04]414; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]415;416; GFX10-LABEL: v_fmac_f32:417; GFX10: ; %bb.0:418; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]419; GFX10-NEXT: v_fmac_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x56]420; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]421;422; GFX11-LABEL: v_fmac_f32:423; GFX11: ; %bb.0:424; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]425; GFX11-NEXT: v_fmac_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x56]426; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]427;428; GFX1200-LABEL: v_fmac_f32:429; GFX1200: ; %bb.0:430; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]431; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]432; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]433; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]434; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]435; GFX1200-NEXT: v_fmac_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x56]436; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]437;438; GFX1250-LABEL: v_fmac_f32:439; GFX1250: ; %bb.0:440; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]441; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]442; GFX1250-NEXT: v_fmac_f32_e32 v0, v0, v1 ; encoding: [0x00,0x03,0x00,0x56]443; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]444 %fma = call float @llvm.fma.f32(float %x, float %y, float %x)445 ret float %fma446}447 448; GFX9: codeLenInByte = 16449; GFX10: codeLenInByte = 12450; GFX11: codeLenInByte = 12451; GFX1200: codeLenInByte = 28452; GFX1250: codeLenInByte = 16453 454define float @v_fmaak_f32(float %x, float %y) {455; GFX9-LABEL: v_fmaak_f32:456; GFX9: ; %bb.0:457; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]458; GFX9-NEXT: s_mov_b32 s4, 0x43800000 ; encoding: [0xff,0x00,0x84,0xbe,0x00,0x00,0x80,0x43]459; GFX9-NEXT: v_fma_f32 v0, v0, v1, s4 ; encoding: [0x00,0x00,0xcb,0xd1,0x00,0x03,0x12,0x00]460; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]461;462; GFX10-LABEL: v_fmaak_f32:463; GFX10: ; %bb.0:464; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]465; GFX10-NEXT: v_fmaak_f32 v0, v0, v1, 0x43800000 ; encoding: [0x00,0x03,0x00,0x5a,0x00,0x00,0x80,0x43]466; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]467;468; GFX11-LABEL: v_fmaak_f32:469; GFX11: ; %bb.0:470; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]471; GFX11-NEXT: v_fmaak_f32 v0, v0, v1, 0x43800000 ; encoding: [0x00,0x03,0x00,0x5a,0x00,0x00,0x80,0x43]472; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]473;474; GFX1200-LABEL: v_fmaak_f32:475; GFX1200: ; %bb.0:476; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]477; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]478; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]479; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]480; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]481; GFX1200-NEXT: v_fmaak_f32 v0, v0, v1, 0x43800000 ; encoding: [0x00,0x03,0x00,0x5a,0x00,0x00,0x80,0x43]482; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]483;484; GFX1250-LABEL: v_fmaak_f32:485; GFX1250: ; %bb.0:486; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]487; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]488; GFX1250-NEXT: v_fmaak_f32 v0, v0, v1, 0x43800000 ; encoding: [0x00,0x03,0x00,0x5a,0x00,0x00,0x80,0x43]489; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]490 %fma = call float @llvm.fma.f32(float %x, float %y, float 256.0)491 ret float %fma492}493 494; GFX9: codeLenInByte = 24495; GFX10: codeLenInByte = 16496; GFX11: codeLenInByte = 16497; GFX1200: codeLenInByte = 32498; GFX1250: codeLenInByte = 20499 500define float @v_fma_k_f32_src_mods(float %x, float %y) {501; GFX9-LABEL: v_fma_k_f32_src_mods:502; GFX9: ; %bb.0:503; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]504; GFX9-NEXT: s_mov_b32 s4, 0x43800000 ; encoding: [0xff,0x00,0x84,0xbe,0x00,0x00,0x80,0x43]505; GFX9-NEXT: v_fma_f32 v0, |v0|, v1, s4 ; encoding: [0x00,0x01,0xcb,0xd1,0x00,0x03,0x12,0x00]506; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]507;508; GFX10-LABEL: v_fma_k_f32_src_mods:509; GFX10: ; %bb.0:510; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]511; GFX10-NEXT: v_fma_f32 v0, |v0|, v1, 0x43800000 ; encoding: [0x00,0x01,0x4b,0xd5,0x00,0x03,0xfe,0x03,0x00,0x00,0x80,0x43]512; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]513;514; GFX11-LABEL: v_fma_k_f32_src_mods:515; GFX11: ; %bb.0:516; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]517; GFX11-NEXT: v_fma_f32 v0, |v0|, v1, 0x43800000 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0xfe,0x03,0x00,0x00,0x80,0x43]518; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]519;520; GFX1200-LABEL: v_fma_k_f32_src_mods:521; GFX1200: ; %bb.0:522; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]523; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]524; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]525; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]526; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]527; GFX1200-NEXT: v_fma_f32 v0, |v0|, v1, 0x43800000 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0xfe,0x03,0x00,0x00,0x80,0x43]528; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]529;530; GFX1250-LABEL: v_fma_k_f32_src_mods:531; GFX1250: ; %bb.0:532; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]533; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]534; GFX1250-NEXT: v_fma_f32 v0, |v0|, v1, 0x43800000 ; encoding: [0x00,0x01,0x13,0xd6,0x00,0x03,0xfe,0x03,0x00,0x00,0x80,0x43]535; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]536 %fabs.x = call float @llvm.fabs.f32(float %x)537 %fma = call float @llvm.fma.f32(float %fabs.x, float %y, float 256.0)538 ret float %fma539}540 541; GFX9: codeLenInByte = 24542; GFX10: codeLenInByte = 20543; GFX11: codeLenInByte = 20544; GFX1200: codeLenInByte = 36545; GFX1250: codeLenInByte = 24546 547define amdgpu_ps float @s_fmaak_f32(float inreg %x, float inreg %y) {548; GFX9-LABEL: s_fmaak_f32:549; GFX9: ; %bb.0:550; GFX9-NEXT: v_mov_b32_e32 v0, s1 ; encoding: [0x01,0x02,0x00,0x7e]551; GFX9-NEXT: v_mov_b32_e32 v1, 0x43800000 ; encoding: [0xff,0x02,0x02,0x7e,0x00,0x00,0x80,0x43]552; GFX9-NEXT: v_fma_f32 v0, s0, v0, v1 ; encoding: [0x00,0x00,0xcb,0xd1,0x00,0x00,0x06,0x04]553; GFX9-NEXT: ; return to shader part epilog554;555; GFX10-LABEL: s_fmaak_f32:556; GFX10: ; %bb.0:557; GFX10-NEXT: v_mov_b32_e32 v0, s1 ; encoding: [0x01,0x02,0x00,0x7e]558; GFX10-NEXT: v_fmaak_f32 v0, s0, v0, 0x43800000 ; encoding: [0x00,0x00,0x00,0x5a,0x00,0x00,0x80,0x43]559; GFX10-NEXT: ; return to shader part epilog560;561; GFX1100-LABEL: s_fmaak_f32:562; GFX1100: ; %bb.0:563; GFX1100-NEXT: v_mov_b32_e32 v0, s1 ; encoding: [0x01,0x02,0x00,0x7e]564; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) ; encoding: [0x01,0x00,0x87,0xbf]565; GFX1100-NEXT: v_fmaak_f32 v0, s0, v0, 0x43800000 ; encoding: [0x00,0x00,0x00,0x5a,0x00,0x00,0x80,0x43]566; GFX1100-NEXT: ; return to shader part epilog567;568; GFX1150-LABEL: s_fmaak_f32:569; GFX1150: ; %bb.0:570; GFX1150-NEXT: s_fmaak_f32 s0, s0, s1, 0x43800000 ; encoding: [0x00,0x01,0x80,0xa2,0x00,0x00,0x80,0x43]571; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; encoding: [0x0b,0x00,0x87,0xbf]572; GFX1150-NEXT: v_mov_b32_e32 v0, s0 ; encoding: [0x00,0x02,0x00,0x7e]573; GFX1150-NEXT: ; return to shader part epilog574;575; GFX1200-LABEL: s_fmaak_f32:576; GFX1200: ; %bb.0:577; GFX1200-NEXT: s_fmaak_f32 s0, s0, s1, 0x43800000 ; encoding: [0x00,0x01,0x80,0xa2,0x00,0x00,0x80,0x43]578; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; encoding: [0x0b,0x00,0x87,0xbf]579; GFX1200-NEXT: v_mov_b32_e32 v0, s0 ; encoding: [0x00,0x02,0x00,0x7e]580; GFX1200-NEXT: ; return to shader part epilog581;582; GFX1250-LABEL: s_fmaak_f32:583; GFX1250: ; %bb.0:584; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; encoding: [0x41,0x06,0x80,0xb9,0x01,0x00,0x00,0x00]585; GFX1250-NEXT: s_fmaak_f32 s0, s0, s1, 0x43800000 ; encoding: [0x00,0x01,0x80,0xa2,0x00,0x00,0x80,0x43]586; GFX1250-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; encoding: [0x0b,0x00,0x87,0xbf]587; GFX1250-NEXT: v_mov_b32_e32 v0, s0 ; encoding: [0x00,0x02,0x00,0x7e]588; GFX1250-NEXT: ; return to shader part epilog589 %fma = call float @llvm.fma.f32(float %x, float %y, float 256.0)590 ret float %fma591}592 593; GFX9: codeLenInByte = 20594; GFX10: codeLenInByte = 12595; GFX1100: codeLenInByte = 16596; GFX1150: codeLenInByte = 16597; GFX1200: codeLenInByte = 16598; GFX1250: codeLenInByte = 24599 600define double @v_mul_f64_vop2_literal_32(double %x) {601; GFX9-LABEL: v_mul_f64_vop2_literal_32:602; GFX9: ; %bb.0:603; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]604; GFX9-NEXT: s_mov_b32 s4, 0 ; encoding: [0x80,0x00,0x84,0xbe]605; GFX9-NEXT: s_mov_b32 s5, 0x405ec000 ; encoding: [0xff,0x00,0x85,0xbe,0x00,0xc0,0x5e,0x40]606; GFX9-NEXT: v_mul_f64 v[0:1], v[0:1], s[4:5] ; encoding: [0x00,0x00,0x81,0xd2,0x00,0x09,0x00,0x00]607; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]608;609; GFX10-LABEL: v_mul_f64_vop2_literal_32:610; GFX10: ; %bb.0:611; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]612; GFX10-NEXT: v_mul_f64 v[0:1], 0x405ec000, v[0:1] ; encoding: [0x00,0x00,0x65,0xd5,0xff,0x00,0x02,0x00,0x00,0xc0,0x5e,0x40]613; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]614;615; GFX11-LABEL: v_mul_f64_vop2_literal_32:616; GFX11: ; %bb.0:617; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]618; GFX11-NEXT: v_mul_f64 v[0:1], 0x405ec000, v[0:1] ; encoding: [0x00,0x00,0x28,0xd7,0xff,0x00,0x02,0x00,0x00,0xc0,0x5e,0x40]619; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]620;621; GFX1200-LABEL: v_mul_f64_vop2_literal_32:622; GFX1200: ; %bb.0:623; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]624; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]625; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]626; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]627; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]628; GFX1200-NEXT: v_mul_f64_e32 v[0:1], 0x405ec000, v[0:1] ; encoding: [0xff,0x00,0x00,0x0c,0x00,0xc0,0x5e,0x40]629; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]630;631; GFX1250-LABEL: v_mul_f64_vop2_literal_32:632; GFX1250: ; %bb.0:633; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]634; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]635; GFX1250-NEXT: v_mul_f64_e32 v[0:1], 0x405ec000, v[0:1] ; encoding: [0xff,0x00,0x00,0x0c,0x00,0xc0,0x5e,0x40]636; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]637 %mul = fmul double %x, 123.0638 ret double %mul639}640 641; GFX9: codeLenInByte = 28642; GFX10: codeLenInByte = 20643; GFX1100: codeLenInByte = 20644; GFX1150: codeLenInByte = 20645; GFX1250: codeLenInByte = 20646 647define double @v_mul_f64_vop2_literal_64(double %x) {648; GFX9-LABEL: v_mul_f64_vop2_literal_64:649; GFX9: ; %bb.0:650; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]651; GFX9-NEXT: s_mov_b32 s4, 0x66666666 ; encoding: [0xff,0x00,0x84,0xbe,0x66,0x66,0x66,0x66]652; GFX9-NEXT: s_mov_b32 s5, 0x405ec666 ; encoding: [0xff,0x00,0x85,0xbe,0x66,0xc6,0x5e,0x40]653; GFX9-NEXT: v_mul_f64 v[0:1], v[0:1], s[4:5] ; encoding: [0x00,0x00,0x81,0xd2,0x00,0x09,0x00,0x00]654; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]655;656; GFX10-LABEL: v_mul_f64_vop2_literal_64:657; GFX10: ; %bb.0:658; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]659; GFX10-NEXT: s_mov_b32 s4, 0x66666666 ; encoding: [0xff,0x03,0x84,0xbe,0x66,0x66,0x66,0x66]660; GFX10-NEXT: s_mov_b32 s5, 0x405ec666 ; encoding: [0xff,0x03,0x85,0xbe,0x66,0xc6,0x5e,0x40]661; GFX10-NEXT: v_mul_f64 v[0:1], v[0:1], s[4:5] ; encoding: [0x00,0x00,0x65,0xd5,0x00,0x09,0x00,0x00]662; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]663;664; GFX11-LABEL: v_mul_f64_vop2_literal_64:665; GFX11: ; %bb.0:666; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]667; GFX11-NEXT: s_mov_b32 s0, 0x66666666 ; encoding: [0xff,0x00,0x80,0xbe,0x66,0x66,0x66,0x66]668; GFX11-NEXT: s_mov_b32 s1, 0x405ec666 ; encoding: [0xff,0x00,0x81,0xbe,0x66,0xc6,0x5e,0x40]669; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; encoding: [0x09,0x00,0x87,0xbf]670; GFX11-NEXT: v_mul_f64 v[0:1], v[0:1], s[0:1] ; encoding: [0x00,0x00,0x28,0xd7,0x00,0x01,0x00,0x00]671; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]672;673; GFX1200-LABEL: v_mul_f64_vop2_literal_64:674; GFX1200: ; %bb.0:675; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]676; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]677; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]678; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]679; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]680; GFX1200-NEXT: s_mov_b32 s0, 0x66666666 ; encoding: [0xff,0x00,0x80,0xbe,0x66,0x66,0x66,0x66]681; GFX1200-NEXT: s_mov_b32 s1, 0x405ec666 ; encoding: [0xff,0x00,0x81,0xbe,0x66,0xc6,0x5e,0x40]682; GFX1200-NEXT: s_wait_alu depctr_sa_sdst(0) ; encoding: [0x9e,0xff,0x88,0xbf]683; GFX1200-NEXT: v_mul_f64_e32 v[0:1], s[0:1], v[0:1] ; encoding: [0x00,0x00,0x00,0x0c]684; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]685;686; GFX1250-LABEL: v_mul_f64_vop2_literal_64:687; GFX1250: ; %bb.0:688; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]689; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]690; GFX1250-NEXT: v_mul_f64_e32 v[0:1], 0x405ec66666666666, v[0:1] ; encoding: [0xfe,0x00,0x00,0x0c,0x66,0x66,0x66,0x66,0x66,0xc6,0x5e,0x40]691; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]692 %mul = fmul double %x, 123.1693 ret double %mul694}695 696; GFX9: codeLenInByte = 32697; GFX10: codeLenInByte = 32698; GFX1100: codeLenInByte = 36699; GFX1150: codeLenInByte = 36700; GFX1250: codeLenInByte = 24701 702define i64 @v_add_u64_vop2_literal_32(i64 %x) {703; GFX9-LABEL: v_add_u64_vop2_literal_32:704; GFX9: ; %bb.0:705; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]706; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 0x7b, v0 ; encoding: [0xff,0x00,0x00,0x32,0x7b,0x00,0x00,0x00]707; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; encoding: [0x80,0x02,0x02,0x38]708; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]709;710; GFX10-LABEL: v_add_u64_vop2_literal_32:711; GFX10: ; %bb.0:712; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]713; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7b, v0 ; encoding: [0x00,0x6a,0x0f,0xd7,0xff,0x00,0x02,0x00,0x7b,0x00,0x00,0x00]714; GFX10-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo ; encoding: [0x01,0x7d,0x28,0xd5,0x80,0x02,0xaa,0x01]715; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]716;717; GFX11-LABEL: v_add_u64_vop2_literal_32:718; GFX11: ; %bb.0:719; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]720; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0x7b, v0 ; encoding: [0x00,0x6a,0x00,0xd7,0xff,0x00,0x02,0x00,0x7b,0x00,0x00,0x00]721; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) ; encoding: [0x01,0x00,0x87,0xbf]722; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo ; encoding: [0x01,0x7c,0x20,0xd5,0x80,0x02,0xaa,0x01]723; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]724;725; GFX1200-LABEL: v_add_u64_vop2_literal_32:726; GFX1200: ; %bb.0:727; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]728; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]729; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]730; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]731; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]732; GFX1200-NEXT: v_add_co_u32 v0, vcc_lo, 0x7b, v0 ; encoding: [0x00,0x6a,0x00,0xd7,0xff,0x00,0x02,0x00,0x7b,0x00,0x00,0x00]733; GFX1200-NEXT: s_wait_alu depctr_va_vcc(0) ; encoding: [0x9d,0xff,0x88,0xbf]734; GFX1200-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo ; encoding: [0x01,0x7c,0x20,0xd5,0x80,0x02,0xaa,0x01]735; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]736;737; GFX1250-LABEL: v_add_u64_vop2_literal_32:738; GFX1250: ; %bb.0:739; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]740; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]741; GFX1250-NEXT: v_add_nc_u64_e32 v[0:1], 0x7b, v[0:1] ; encoding: [0xff,0x00,0x00,0x50,0x7b,0x00,0x00,0x00]742; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]743 %add = add i64 %x, 123744 ret i64 %add745}746 747; GFX9: codeLenInByte = 20748; GFX10: codeLenInByte = 28749; GFX1100: codeLenInByte = 32750; GFX1150: codeLenInByte = 32751; GFX1250: codeLenInByte = 20752 753define i64 @v_add_u64_vop2_literal_64(i64 %x) {754; GFX9-LABEL: v_add_u64_vop2_literal_64:755; GFX9: ; %bb.0:756; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]757; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, 0x12345678, v0 ; encoding: [0xff,0x00,0x00,0x32,0x78,0x56,0x34,0x12]758; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, 1, v1, vcc ; encoding: [0x81,0x02,0x02,0x38]759; GFX9-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x1d,0x80,0xbe]760;761; GFX10-LABEL: v_add_u64_vop2_literal_64:762; GFX10: ; %bb.0:763; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x8c,0xbf]764; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x12345678, v0 ; encoding: [0x00,0x6a,0x0f,0xd7,0xff,0x00,0x02,0x00,0x78,0x56,0x34,0x12]765; GFX10-NEXT: v_add_co_ci_u32_e64 v1, null, 1, v1, vcc_lo ; encoding: [0x01,0x7d,0x28,0xd5,0x81,0x02,0xaa,0x01]766; GFX10-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x20,0x80,0xbe]767;768; GFX11-LABEL: v_add_u64_vop2_literal_64:769; GFX11: ; %bb.0:770; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; encoding: [0x00,0x00,0x89,0xbf]771; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0x12345678, v0 ; encoding: [0x00,0x6a,0x00,0xd7,0xff,0x00,0x02,0x00,0x78,0x56,0x34,0x12]772; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) ; encoding: [0x01,0x00,0x87,0xbf]773; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, 1, v1, vcc_lo ; encoding: [0x01,0x7c,0x20,0xd5,0x81,0x02,0xaa,0x01]774; GFX11-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]775;776; GFX1200-LABEL: v_add_u64_vop2_literal_64:777; GFX1200: ; %bb.0:778; GFX1200-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]779; GFX1200-NEXT: s_wait_expcnt 0x0 ; encoding: [0x00,0x00,0xc4,0xbf]780; GFX1200-NEXT: s_wait_samplecnt 0x0 ; encoding: [0x00,0x00,0xc2,0xbf]781; GFX1200-NEXT: s_wait_bvhcnt 0x0 ; encoding: [0x00,0x00,0xc3,0xbf]782; GFX1200-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]783; GFX1200-NEXT: v_add_co_u32 v0, vcc_lo, 0x12345678, v0 ; encoding: [0x00,0x6a,0x00,0xd7,0xff,0x00,0x02,0x00,0x78,0x56,0x34,0x12]784; GFX1200-NEXT: s_wait_alu depctr_va_vcc(0) ; encoding: [0x9d,0xff,0x88,0xbf]785; GFX1200-NEXT: v_add_co_ci_u32_e64 v1, null, 1, v1, vcc_lo ; encoding: [0x01,0x7c,0x20,0xd5,0x81,0x02,0xaa,0x01]786; GFX1200-NEXT: s_setpc_b64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]787;788; GFX1250-LABEL: v_add_u64_vop2_literal_64:789; GFX1250: ; %bb.0:790; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 ; encoding: [0x00,0x00,0xc8,0xbf]791; GFX1250-NEXT: s_wait_kmcnt 0x0 ; encoding: [0x00,0x00,0xc7,0xbf]792; GFX1250-NEXT: v_add_nc_u64_e32 v[0:1], 0x112345678, v[0:1] ; encoding: [0xfe,0x00,0x00,0x50,0x78,0x56,0x34,0x12,0x01,0x00,0x00,0x00]793; GFX1250-NEXT: s_set_pc_i64 s[30:31] ; encoding: [0x1e,0x48,0x80,0xbe]794 %add = add i64 %x, 4600387192795 ret i64 %add796}797 798; GFX9: codeLenInByte = 20799; GFX10: codeLenInByte = 28800; GFX1100: codeLenInByte = 32801; GFX1150: codeLenInByte = 32802; GFX1250: codeLenInByte = 24803;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:804; NOT-GFX12: {{.*}}805