163 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -fast-isel -mtriple=i386-unknown-unknown -mattr=+avx512f,+avx512vbmi | FileCheck %s --check-prefixes=CHECK,X863; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi | FileCheck %s --check-prefixes=CHECK,X644 5; NOTE: This should use IR equivalent to what is generated by clang/test/CodeGen/avx512vbmi-builtins.c6 7define <8 x i64> @test_mm512_mask2_permutex2var_epi8(<8 x i64> %__A, <8 x i64> %__I, i64 %__U, <8 x i64> %__B) {8; X86-LABEL: test_mm512_mask2_permutex2var_epi8:9; X86: # %bb.0: # %entry10; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k011; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k112; X86-NEXT: kunpckdq %k1, %k0, %k113; X86-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1}14; X86-NEXT: vmovdqa64 %zmm1, %zmm015; X86-NEXT: retl16;17; X64-LABEL: test_mm512_mask2_permutex2var_epi8:18; X64: # %bb.0: # %entry19; X64-NEXT: kmovq %rdi, %k120; X64-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1}21; X64-NEXT: vmovdqa64 %zmm1, %zmm022; X64-NEXT: retq23entry:24 %0 = bitcast <8 x i64> %__A to <64 x i8>25 %1 = bitcast <8 x i64> %__I to <64 x i8>26 %2 = bitcast <8 x i64> %__B to <64 x i8>27 %3 = tail call <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8> %0, <64 x i8> %1, <64 x i8> %2)28 %4 = bitcast i64 %__U to <64 x i1>29 %5 = select <64 x i1> %4, <64 x i8> %3, <64 x i8> %130 %6 = bitcast <64 x i8> %5 to <8 x i64>31 ret <8 x i64> %632}33 34define <8 x i64> @test_mm512_permutex2var_epi8(<8 x i64> %__A, <8 x i64> %__I, <8 x i64> %__B) {35; CHECK-LABEL: test_mm512_permutex2var_epi8:36; CHECK: # %bb.0: # %entry37; CHECK-NEXT: vpermt2b %zmm2, %zmm1, %zmm038; CHECK-NEXT: ret{{[l|q]}}39entry:40 %0 = bitcast <8 x i64> %__A to <64 x i8>41 %1 = bitcast <8 x i64> %__I to <64 x i8>42 %2 = bitcast <8 x i64> %__B to <64 x i8>43 %3 = tail call <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8> %0, <64 x i8> %1, <64 x i8> %2)44 %4 = bitcast <64 x i8> %3 to <8 x i64>45 ret <8 x i64> %446}47 48define <8 x i64> @test_mm512_mask_permutex2var_epi8(<8 x i64> %__A, i64 %__U, <8 x i64> %__I, <8 x i64> %__B) {49; X86-LABEL: test_mm512_mask_permutex2var_epi8:50; X86: # %bb.0: # %entry51; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k052; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k153; X86-NEXT: kunpckdq %k1, %k0, %k154; X86-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1}55; X86-NEXT: retl56;57; X64-LABEL: test_mm512_mask_permutex2var_epi8:58; X64: # %bb.0: # %entry59; X64-NEXT: kmovq %rdi, %k160; X64-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1}61; X64-NEXT: retq62entry:63 %0 = bitcast <8 x i64> %__A to <64 x i8>64 %1 = bitcast <8 x i64> %__I to <64 x i8>65 %2 = bitcast <8 x i64> %__B to <64 x i8>66 %3 = tail call <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8> %0, <64 x i8> %1, <64 x i8> %2)67 %4 = bitcast i64 %__U to <64 x i1>68 %5 = select <64 x i1> %4, <64 x i8> %3, <64 x i8> %069 %6 = bitcast <64 x i8> %5 to <8 x i64>70 ret <8 x i64> %671}72 73define <8 x i64> @test_mm512_maskz_permutex2var_epi8(i64 %__U, <8 x i64> %__A, <8 x i64> %__I, <8 x i64> %__B) {74; X86-LABEL: test_mm512_maskz_permutex2var_epi8:75; X86: # %bb.0: # %entry76; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k077; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k178; X86-NEXT: kunpckdq %k1, %k0, %k179; X86-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} {z}80; X86-NEXT: retl81;82; X64-LABEL: test_mm512_maskz_permutex2var_epi8:83; X64: # %bb.0: # %entry84; X64-NEXT: kmovq %rdi, %k185; X64-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} {z}86; X64-NEXT: retq87entry:88 %0 = bitcast <8 x i64> %__A to <64 x i8>89 %1 = bitcast <8 x i64> %__I to <64 x i8>90 %2 = bitcast <8 x i64> %__B to <64 x i8>91 %3 = tail call <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8> %0, <64 x i8> %1, <64 x i8> %2)92 %4 = bitcast i64 %__U to <64 x i1>93 %5 = select <64 x i1> %4, <64 x i8> %3, <64 x i8> zeroinitializer94 %6 = bitcast <64 x i8> %5 to <8 x i64>95 ret <8 x i64> %696}97 98declare <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>)99 100define <8 x i64> @test_mm512_mask_multishift_epi64_epi8(<8 x i64> %__W, i64 %__M, <8 x i64> %__X, <8 x i64> %__Y) {101; X86-LABEL: test_mm512_mask_multishift_epi64_epi8:102; X86: # %bb.0: # %entry103; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0104; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1105; X86-NEXT: kunpckdq %k1, %k0, %k1106; X86-NEXT: vpmultishiftqb %zmm2, %zmm1, %zmm0 {%k1}107; X86-NEXT: retl108;109; X64-LABEL: test_mm512_mask_multishift_epi64_epi8:110; X64: # %bb.0: # %entry111; X64-NEXT: kmovq %rdi, %k1112; X64-NEXT: vpmultishiftqb %zmm2, %zmm1, %zmm0 {%k1}113; X64-NEXT: retq114entry:115 %0 = bitcast <8 x i64> %__X to <64 x i8>116 %1 = bitcast <8 x i64> %__Y to <64 x i8>117 %2 = tail call <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8> %0, <64 x i8> %1)118 %3 = bitcast <8 x i64> %__W to <64 x i8>119 %4 = bitcast i64 %__M to <64 x i1>120 %5 = select <64 x i1> %4, <64 x i8> %2, <64 x i8> %3121 %6 = bitcast <64 x i8> %5 to <8 x i64>122 ret <8 x i64> %6123}124 125define <8 x i64> @test_mm512_maskz_multishift_epi64_epi8(i64 %__M, <8 x i64> %__X, <8 x i64> %__Y) {126; X86-LABEL: test_mm512_maskz_multishift_epi64_epi8:127; X86: # %bb.0: # %entry128; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0129; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1130; X86-NEXT: kunpckdq %k1, %k0, %k1131; X86-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0 {%k1} {z}132; X86-NEXT: retl133;134; X64-LABEL: test_mm512_maskz_multishift_epi64_epi8:135; X64: # %bb.0: # %entry136; X64-NEXT: kmovq %rdi, %k1137; X64-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0 {%k1} {z}138; X64-NEXT: retq139entry:140 %0 = bitcast <8 x i64> %__X to <64 x i8>141 %1 = bitcast <8 x i64> %__Y to <64 x i8>142 %2 = tail call <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8> %0, <64 x i8> %1)143 %3 = bitcast i64 %__M to <64 x i1>144 %4 = select <64 x i1> %3, <64 x i8> %2, <64 x i8> zeroinitializer145 %5 = bitcast <64 x i8> %4 to <8 x i64>146 ret <8 x i64> %5147}148 149define <8 x i64> @test_mm512_multishift_epi64_epi8(<8 x i64> %__X, <8 x i64> %__Y) {150; CHECK-LABEL: test_mm512_multishift_epi64_epi8:151; CHECK: # %bb.0: # %entry152; CHECK-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0153; CHECK-NEXT: ret{{[l|q]}}154entry:155 %0 = bitcast <8 x i64> %__X to <64 x i8>156 %1 = bitcast <8 x i64> %__Y to <64 x i8>157 %2 = tail call <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8> %0, <64 x i8> %1)158 %3 = bitcast <64 x i8> %2 to <8 x i64>159 ret <8 x i64> %3160}161 162declare <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8>, <64 x i8>)163