168 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 42; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250 %s3 4define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo(bfloat %src0, bfloat %src1, bfloat %src2) #0 {5; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo:6; GFX1250: ; %bb.0:7; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x08; GFX1250-NEXT: s_wait_kmcnt 0x09; GFX1250-NEXT: v_fma_mixhi_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1]10; GFX1250-NEXT: s_set_pc_i64 s[30:31]11 %src0.ext = fpext bfloat %src0 to float12 %src1.ext = fpext bfloat %src1 to float13 %src2.ext = fpext bfloat %src2 to float14 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)15 %cvt.result = fptrunc float %result to bfloat16 %vec.result = insertelement <2 x bfloat> undef, bfloat %cvt.result, i32 117 ret <2 x bfloat> %vec.result18}19 20define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_constlo(bfloat %src0, bfloat %src1, bfloat %src2) #0 {21; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_constlo:22; GFX1250: ; %bb.0:23; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x024; GFX1250-NEXT: s_wait_kmcnt 0x025; GFX1250-NEXT: v_mov_b32_e32 v3, 0x3f8026; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)27; GFX1250-NEXT: v_fma_mixhi_bf16 v3, v0, v1, v2 op_sel_hi:[1,1,1]28; GFX1250-NEXT: v_mov_b32_e32 v0, v329; GFX1250-NEXT: s_set_pc_i64 s[30:31]30 %src0.ext = fpext bfloat %src0 to float31 %src1.ext = fpext bfloat %src1 to float32 %src2.ext = fpext bfloat %src2 to float33 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)34 %cvt.result = fptrunc float %result to bfloat35 %vec.result = insertelement <2 x bfloat> <bfloat 1.0, bfloat undef>, bfloat %cvt.result, i32 136 ret <2 x bfloat> %vec.result37}38 39define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_reglo(bfloat %src0, bfloat %src1, bfloat %src2, bfloat %lo) #0 {40; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_reglo:41; GFX1250: ; %bb.0:42; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x043; GFX1250-NEXT: s_wait_kmcnt 0x044; GFX1250-NEXT: v_fma_mixhi_bf16 v3, v0, v1, v2 op_sel_hi:[1,1,1]45; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)46; GFX1250-NEXT: v_mov_b32_e32 v0, v347; GFX1250-NEXT: s_set_pc_i64 s[30:31]48 %src0.ext = fpext bfloat %src0 to float49 %src1.ext = fpext bfloat %src1 to float50 %src2.ext = fpext bfloat %src2 to float51 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)52 %cvt.result = fptrunc float %result to bfloat53 %vec = insertelement <2 x bfloat> undef, bfloat %lo, i32 054 %vec.result = insertelement <2 x bfloat> %vec, bfloat %cvt.result, i32 155 ret <2 x bfloat> %vec.result56}57 58define i32 @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_intpack(bfloat %src0, bfloat %src1, bfloat %src2) #0 {59; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_intpack:60; GFX1250: ; %bb.0:61; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x062; GFX1250-NEXT: s_wait_kmcnt 0x063; GFX1250-NEXT: v_fma_mixlo_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1]64; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)65; GFX1250-NEXT: v_lshlrev_b32_e32 v0, 16, v066; GFX1250-NEXT: s_set_pc_i64 s[30:31]67 %src0.ext = fpext bfloat %src0 to float68 %src1.ext = fpext bfloat %src1 to float69 %src2.ext = fpext bfloat %src2 to float70 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)71 %cvt.result = fptrunc float %result to bfloat72 %bc = bitcast bfloat %cvt.result to i1673 %ext = zext i16 %bc to i3274 %shr = shl i32 %ext, 1675 ret i32 %shr76}77 78define i32 @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_intpack_sext(bfloat %src0, bfloat %src1, bfloat %src2) #0 {79; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_intpack_sext:80; GFX1250: ; %bb.0:81; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x082; GFX1250-NEXT: s_wait_kmcnt 0x083; GFX1250-NEXT: v_fma_mixlo_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1]84; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)85; GFX1250-NEXT: v_lshlrev_b32_e32 v0, 16, v086; GFX1250-NEXT: s_set_pc_i64 s[30:31]87 %src0.ext = fpext bfloat %src0 to float88 %src1.ext = fpext bfloat %src1 to float89 %src2.ext = fpext bfloat %src2 to float90 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)91 %cvt.result = fptrunc float %result to bfloat92 %bc = bitcast bfloat %cvt.result to i1693 %ext = sext i16 %bc to i3294 %shr = shl i32 %ext, 1695 ret i32 %shr96}97 98define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_precvt(bfloat %src0, bfloat %src1, bfloat %src2) #0 {99; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_precvt:100; GFX1250: ; %bb.0:101; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0102; GFX1250-NEXT: s_wait_kmcnt 0x0103; GFX1250-NEXT: v_fma_mix_f32_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1] clamp104; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)105; GFX1250-NEXT: v_cvt_pk_bf16_f32 v0, v0, s0106; GFX1250-NEXT: v_lshlrev_b32_e32 v0, 16, v0107; GFX1250-NEXT: s_set_pc_i64 s[30:31]108 %src0.ext = fpext bfloat %src0 to float109 %src1.ext = fpext bfloat %src1 to float110 %src2.ext = fpext bfloat %src2 to float111 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)112 %max = call float @llvm.maxnum.f32(float %result, float 0.0)113 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)114 %cvt.result = fptrunc float %clamp to bfloat115 %vec.result = insertelement <2 x bfloat> undef, bfloat %cvt.result, i32 1116 ret <2 x bfloat> %vec.result117}118 119define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_postcvt(bfloat %src0, bfloat %src1, bfloat %src2) #0 {120; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_postcvt:121; GFX1250: ; %bb.0:122; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0123; GFX1250-NEXT: s_wait_kmcnt 0x0124; GFX1250-NEXT: v_fma_mixhi_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1] clamp125; GFX1250-NEXT: s_set_pc_i64 s[30:31]126 %src0.ext = fpext bfloat %src0 to float127 %src1.ext = fpext bfloat %src1 to float128 %src2.ext = fpext bfloat %src2 to float129 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)130 %cvt.result = fptrunc float %result to bfloat131 %max = call bfloat @llvm.maxnum.bf16(bfloat %cvt.result, bfloat 0.0)132 %clamp = call bfloat @llvm.minnum.bf16(bfloat %max, bfloat 1.0)133 %vec.result = insertelement <2 x bfloat> undef, bfloat %clamp, i32 1134 ret <2 x bfloat> %vec.result135}136 137define <2 x bfloat> @v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_postcvt_multi_use(bfloat %src0, bfloat %src1, bfloat %src2) #0 {138; GFX1250-LABEL: v_mad_mixhi_bf16_bf16lo_bf16lo_bf16lo_undeflo_clamp_postcvt_multi_use:139; GFX1250: ; %bb.0:140; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0141; GFX1250-NEXT: s_wait_kmcnt 0x0142; GFX1250-NEXT: v_fma_mixlo_bf16 v3, v0, v1, v2 op_sel_hi:[1,1,1]143; GFX1250-NEXT: v_fma_mixhi_bf16 v0, v0, v1, v2 op_sel_hi:[1,1,1] clamp144; GFX1250-NEXT: global_store_b16 v[0:1], v3, off scope:SCOPE_SYS145; GFX1250-NEXT: s_wait_storecnt 0x0146; GFX1250-NEXT: s_set_pc_i64 s[30:31]147 %src0.ext = fpext bfloat %src0 to float148 %src1.ext = fpext bfloat %src1 to float149 %src2.ext = fpext bfloat %src2 to float150 %result = tail call float @llvm.fmuladd.f32(float %src0.ext, float %src1.ext, float %src2.ext)151 %cvt.result = fptrunc float %result to bfloat152 store volatile bfloat %cvt.result, ptr addrspace(1) undef153 %max = call bfloat @llvm.maxnum.bf16(bfloat %cvt.result, bfloat 0.0)154 %clamp = call bfloat @llvm.minnum.bf16(bfloat %max, bfloat 1.0)155 %vec.result = insertelement <2 x bfloat> undef, bfloat %clamp, i32 1156 ret <2 x bfloat> %vec.result157}158 159declare bfloat @llvm.minnum.bf16(bfloat, bfloat) #1160declare bfloat @llvm.maxnum.bf16(bfloat, bfloat) #1161declare float @llvm.minnum.f32(float, float) #1162declare float @llvm.maxnum.f32(float, float) #1163declare float @llvm.fmuladd.f32(float, float, float) #1164declare <2 x float> @llvm.fmuladd.v2f32(<2 x float>, <2 x float>, <2 x float>) #1165 166attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }167attributes #1 = { nounwind readnone speculatable }168