1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2 ; RUN: llc < %s -fast-isel -mtriple=i386-unknown-unknown -mattr=+avx512f,+avx512vbmi | FileCheck %s --check-prefixes=CHECK,X86 3 ; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi | FileCheck %s --check-prefixes=CHECK,X64 4 5 ; NOTE: This should use IR equivalent to what is generated by clang/test/CodeGen/avx512vbmi-builtins.c 6 7 define <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: # %entry 10 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0 11 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 12 ; X86-NEXT: kunpckdq %k1, %k0, %k1 13 ; X86-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1} 14 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 15 ; X86-NEXT: retl 16 ; 17 ; X64-LABEL: test_mm512_mask2_permutex2var_epi8: 18 ; X64: # %bb.0: # %entry 19 ; X64-NEXT: kmovq %rdi, %k1 20 ; X64-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1} 21 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 22 ; X64-NEXT: retq 23 entry: 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> %1 30 %6 = bitcast <64 x i8> %5 to <8 x i64> 31 ret <8 x i64> %6 32 } 33 34 define <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: # %entry 37 ; CHECK-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 38 ; CHECK-NEXT: ret{{[l|q]}} 39 entry: 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> %4 46 } 47 48 define <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: # %entry 51 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0 52 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 53 ; X86-NEXT: kunpckdq %k1, %k0, %k1 54 ; X86-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} 55 ; X86-NEXT: retl 56 ; 57 ; X64-LABEL: test_mm512_mask_permutex2var_epi8: 58 ; X64: # %bb.0: # %entry 59 ; X64-NEXT: kmovq %rdi, %k1 60 ; X64-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} 61 ; X64-NEXT: retq 62 entry: 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> %0 69 %6 = bitcast <64 x i8> %5 to <8 x i64> 70 ret <8 x i64> %6 71 } 72 73 define <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: # %entry 76 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0 77 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 78 ; X86-NEXT: kunpckdq %k1, %k0, %k1 79 ; X86-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} {z} 80 ; X86-NEXT: retl 81 ; 82 ; X64-LABEL: test_mm512_maskz_permutex2var_epi8: 83 ; X64: # %bb.0: # %entry 84 ; X64-NEXT: kmovq %rdi, %k1 85 ; X64-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 {%k1} {z} 86 ; X64-NEXT: retq 87 entry: 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> zeroinitializer 94 %6 = bitcast <64 x i8> %5 to <8 x i64> 95 ret <8 x i64> %6 96 } 97 98 declare <64 x i8> @llvm.x86.avx512.vpermi2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>) 99
