From 7e37746011cdf0f396f35a7d25fed3d8b7bd405b Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Wed, 20 Apr 2016 05:18:58 +0000 Subject: [PATCH] [AVX512] Update vector-tzcnt-512 test to show how bad v32i16 and v64i8 is with avx512bw enabled. llvm-svn: 266859 --- llvm/test/CodeGen/X86/vector-tzcnt-512.ll | 972 ++++++++++++++++++++++++++---- 1 file changed, 859 insertions(+), 113 deletions(-) diff --git a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll index 9265fad..5d7bcd8 100644 --- a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll +++ b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll @@ -1,5 +1,6 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd | FileCheck %s --check-prefix=ALL --check-prefix=AVX512 --check-prefix=AVX512CD +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd,-avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512CD +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd,+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512CDBW define <8 x i64> @testv8i64(<8 x i64> %in) nounwind { ; ALL-LABEL: testv8i64: @@ -134,133 +135,878 @@ define <16 x i32> @testv16i32u(<16 x i32> %in) nounwind { } define <32 x i16> @testv32i16(<32 x i16> %in) nounwind { -; ALL-LABEL: testv32i16: -; ALL: ## BB#0: -; ALL-NEXT: vpxor %ymm2, %ymm2, %ymm2 -; ALL-NEXT: vpsubw %ymm0, %ymm2, %ymm3 -; ALL-NEXT: vpand %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] -; ALL-NEXT: vpsubw %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm5 -; ALL-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] -; ALL-NEXT: vpshufb %ymm5, %ymm6, %ymm5 -; ALL-NEXT: vpsrlw $4, %ymm0, %ymm0 -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm0 -; ALL-NEXT: vpshufb %ymm0, %ymm6, %ymm0 -; ALL-NEXT: vpaddb %ymm5, %ymm0, %ymm0 -; ALL-NEXT: vpsllw $8, %ymm0, %ymm5 -; ALL-NEXT: vpaddb %ymm0, %ymm5, %ymm0 -; ALL-NEXT: vpsrlw $8, %ymm0, %ymm0 -; ALL-NEXT: vpsubw %ymm1, %ymm2, %ymm2 -; ALL-NEXT: vpand %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsubw %ymm3, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm2 -; ALL-NEXT: vpshufb %ymm2, %ymm6, %ymm2 -; ALL-NEXT: vpsrlw $4, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm1 -; ALL-NEXT: vpshufb %ymm1, %ymm6, %ymm1 -; ALL-NEXT: vpaddb %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsllw $8, %ymm1, %ymm2 -; ALL-NEXT: vpaddb %ymm1, %ymm2, %ymm1 -; ALL-NEXT: vpsrlw $8, %ymm1, %ymm1 -; ALL-NEXT: retq +; AVX512CD-LABEL: testv32i16: +; AVX512CD: ## BB#0: +; AVX512CD-NEXT: vpxor %ymm2, %ymm2, %ymm2 +; AVX512CD-NEXT: vpsubw %ymm0, %ymm2, %ymm3 +; AVX512CD-NEXT: vpand %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] +; AVX512CD-NEXT: vpsubw %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm5 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CD-NEXT: vpshufb %ymm5, %ymm6, %ymm5 +; AVX512CD-NEXT: vpsrlw $4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpshufb %ymm0, %ymm6, %ymm0 +; AVX512CD-NEXT: vpaddb %ymm5, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsllw $8, %ymm0, %ymm5 +; AVX512CD-NEXT: vpaddb %ymm0, %ymm5, %ymm0 +; AVX512CD-NEXT: vpsrlw $8, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsubw %ymm1, %ymm2, %ymm2 +; AVX512CD-NEXT: vpand %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsubw %ymm3, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm2 +; AVX512CD-NEXT: vpshufb %ymm2, %ymm6, %ymm2 +; AVX512CD-NEXT: vpsrlw $4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpshufb %ymm1, %ymm6, %ymm1 +; AVX512CD-NEXT: vpaddb %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsllw $8, %ymm1, %ymm2 +; AVX512CD-NEXT: vpaddb %ymm1, %ymm2, %ymm1 +; AVX512CD-NEXT: vpsrlw $8, %ymm1, %ymm1 +; AVX512CD-NEXT: retq +; +; AVX512CDBW-LABEL: testv32i16: +; AVX512CDBW: ## BB#0: +; AVX512CDBW-NEXT: vextracti32x4 $3, %zmm0, %xmm1 +; AVX512CDBW-NEXT: vpextrw $1, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm2 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $2, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $3, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $4, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $5, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $6, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $7, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm2, %xmm1 +; AVX512CDBW-NEXT: vextracti32x4 $2, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1 +; AVX512CDBW-NEXT: vextracti32x4 $1, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm0 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 +; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: retq %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 0) ret <32 x i16> %out } define <32 x i16> @testv32i16u(<32 x i16> %in) nounwind { -; ALL-LABEL: testv32i16u: -; ALL: ## BB#0: -; ALL-NEXT: vpxor %ymm2, %ymm2, %ymm2 -; ALL-NEXT: vpsubw %ymm0, %ymm2, %ymm3 -; ALL-NEXT: vpand %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] -; ALL-NEXT: vpsubw %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm5 -; ALL-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] -; ALL-NEXT: vpshufb %ymm5, %ymm6, %ymm5 -; ALL-NEXT: vpsrlw $4, %ymm0, %ymm0 -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm0 -; ALL-NEXT: vpshufb %ymm0, %ymm6, %ymm0 -; ALL-NEXT: vpaddb %ymm5, %ymm0, %ymm0 -; ALL-NEXT: vpsllw $8, %ymm0, %ymm5 -; ALL-NEXT: vpaddb %ymm0, %ymm5, %ymm0 -; ALL-NEXT: vpsrlw $8, %ymm0, %ymm0 -; ALL-NEXT: vpsubw %ymm1, %ymm2, %ymm2 -; ALL-NEXT: vpand %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsubw %ymm3, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm2 -; ALL-NEXT: vpshufb %ymm2, %ymm6, %ymm2 -; ALL-NEXT: vpsrlw $4, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm1 -; ALL-NEXT: vpshufb %ymm1, %ymm6, %ymm1 -; ALL-NEXT: vpaddb %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsllw $8, %ymm1, %ymm2 -; ALL-NEXT: vpaddb %ymm1, %ymm2, %ymm1 -; ALL-NEXT: vpsrlw $8, %ymm1, %ymm1 -; ALL-NEXT: retq +; AVX512CD-LABEL: testv32i16u: +; AVX512CD: ## BB#0: +; AVX512CD-NEXT: vpxor %ymm2, %ymm2, %ymm2 +; AVX512CD-NEXT: vpsubw %ymm0, %ymm2, %ymm3 +; AVX512CD-NEXT: vpand %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] +; AVX512CD-NEXT: vpsubw %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm5 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CD-NEXT: vpshufb %ymm5, %ymm6, %ymm5 +; AVX512CD-NEXT: vpsrlw $4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpshufb %ymm0, %ymm6, %ymm0 +; AVX512CD-NEXT: vpaddb %ymm5, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsllw $8, %ymm0, %ymm5 +; AVX512CD-NEXT: vpaddb %ymm0, %ymm5, %ymm0 +; AVX512CD-NEXT: vpsrlw $8, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsubw %ymm1, %ymm2, %ymm2 +; AVX512CD-NEXT: vpand %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsubw %ymm3, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm2 +; AVX512CD-NEXT: vpshufb %ymm2, %ymm6, %ymm2 +; AVX512CD-NEXT: vpsrlw $4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpshufb %ymm1, %ymm6, %ymm1 +; AVX512CD-NEXT: vpaddb %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsllw $8, %ymm1, %ymm2 +; AVX512CD-NEXT: vpaddb %ymm1, %ymm2, %ymm1 +; AVX512CD-NEXT: vpsrlw $8, %ymm1, %ymm1 +; AVX512CD-NEXT: retq +; +; AVX512CDBW-LABEL: testv32i16u: +; AVX512CDBW: ## BB#0: +; AVX512CDBW-NEXT: vextracti32x4 $3, %zmm0, %xmm1 +; AVX512CDBW-NEXT: vpextrw $1, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm2 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $2, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $3, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $4, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $5, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $6, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrw $7, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm2, %xmm1 +; AVX512CDBW-NEXT: vextracti32x4 $2, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1 +; AVX512CDBW-NEXT: vextracti32x4 $1, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vpextrw $1, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vmovd %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntw %cx, %cx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $2, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $3, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $4, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $5, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $6, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrw $7, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntw %ax, %ax +; AVX512CDBW-NEXT: vpinsrw $7, %eax, %xmm3, %xmm0 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 +; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: retq %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 -1) ret <32 x i16> %out } define <64 x i8> @testv64i8(<64 x i8> %in) nounwind { -; ALL-LABEL: testv64i8: -; ALL: ## BB#0: -; ALL-NEXT: vpxor %ymm2, %ymm2, %ymm2 -; ALL-NEXT: vpsubb %ymm0, %ymm2, %ymm3 -; ALL-NEXT: vpand %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] -; ALL-NEXT: vpsubb %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm5 -; ALL-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] -; ALL-NEXT: vpshufb %ymm5, %ymm6, %ymm5 -; ALL-NEXT: vpsrlw $4, %ymm0, %ymm0 -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm0 -; ALL-NEXT: vpshufb %ymm0, %ymm6, %ymm0 -; ALL-NEXT: vpaddb %ymm5, %ymm0, %ymm0 -; ALL-NEXT: vpsubb %ymm1, %ymm2, %ymm2 -; ALL-NEXT: vpand %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsubb %ymm3, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm2 -; ALL-NEXT: vpshufb %ymm2, %ymm6, %ymm2 -; ALL-NEXT: vpsrlw $4, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm1 -; ALL-NEXT: vpshufb %ymm1, %ymm6, %ymm1 -; ALL-NEXT: vpaddb %ymm2, %ymm1, %ymm1 -; ALL-NEXT: retq +; AVX512CD-LABEL: testv64i8: +; AVX512CD: ## BB#0: +; AVX512CD-NEXT: vpxor %ymm2, %ymm2, %ymm2 +; AVX512CD-NEXT: vpsubb %ymm0, %ymm2, %ymm3 +; AVX512CD-NEXT: vpand %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] +; AVX512CD-NEXT: vpsubb %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm5 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CD-NEXT: vpshufb %ymm5, %ymm6, %ymm5 +; AVX512CD-NEXT: vpsrlw $4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpshufb %ymm0, %ymm6, %ymm0 +; AVX512CD-NEXT: vpaddb %ymm5, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsubb %ymm1, %ymm2, %ymm2 +; AVX512CD-NEXT: vpand %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsubb %ymm3, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm2 +; AVX512CD-NEXT: vpshufb %ymm2, %ymm6, %ymm2 +; AVX512CD-NEXT: vpsrlw $4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpshufb %ymm1, %ymm6, %ymm1 +; AVX512CD-NEXT: vpaddb %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: retq +; +; AVX512CDBW-LABEL: testv64i8: +; AVX512CDBW: ## BB#0: +; AVX512CDBW-NEXT: vextracti32x4 $3, %zmm0, %xmm1 +; AVX512CDBW-NEXT: vpextrb $1, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: movl $8, %eax +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpextrb $0, %xmm1, %edx +; AVX512CDBW-NEXT: tzcntl %edx, %edx +; AVX512CDBW-NEXT: cmpl $32, %edx +; AVX512CDBW-NEXT: cmovel %eax, %edx +; AVX512CDBW-NEXT: vmovd %edx, %xmm2 +; AVX512CDBW-NEXT: vpinsrb $1, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $2, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $2, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $3, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $3, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $4, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $4, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $5, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $5, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $6, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $6, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $7, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $7, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $8, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $8, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $9, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $9, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $10, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $10, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $11, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $11, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $12, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $12, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $13, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $13, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $14, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $14, %ecx, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $15, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $15, %ecx, %xmm2, %xmm1 +; AVX512CDBW-NEXT: vextracti32x4 $2, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpextrb $0, %xmm2, %edx +; AVX512CDBW-NEXT: tzcntl %edx, %edx +; AVX512CDBW-NEXT: cmpl $32, %edx +; AVX512CDBW-NEXT: cmovel %eax, %edx +; AVX512CDBW-NEXT: vmovd %edx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $2, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $3, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $4, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $5, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $6, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $7, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $8, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $9, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $10, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $11, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $12, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $13, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $14, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $15, %ecx, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1 +; AVX512CDBW-NEXT: vextracti32x4 $1, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpextrb $0, %xmm2, %edx +; AVX512CDBW-NEXT: tzcntl %edx, %edx +; AVX512CDBW-NEXT: cmpl $32, %edx +; AVX512CDBW-NEXT: cmovel %eax, %edx +; AVX512CDBW-NEXT: vmovd %edx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $2, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $3, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $4, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $5, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $6, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $7, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $8, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $9, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $10, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $11, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $12, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $13, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $14, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $15, %ecx, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpextrb $0, %xmm0, %edx +; AVX512CDBW-NEXT: tzcntl %edx, %edx +; AVX512CDBW-NEXT: cmpl $32, %edx +; AVX512CDBW-NEXT: cmovel %eax, %edx +; AVX512CDBW-NEXT: vmovd %edx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $2, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $3, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $4, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $5, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $6, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $7, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $8, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $9, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $10, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $11, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $12, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $13, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $14, %ecx, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: cmpl $32, %ecx +; AVX512CDBW-NEXT: cmovel %eax, %ecx +; AVX512CDBW-NEXT: vpinsrb $15, %ecx, %xmm3, %xmm0 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 +; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: retq %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 0) ret <64 x i8> %out } define <64 x i8> @testv64i8u(<64 x i8> %in) nounwind { -; ALL-LABEL: testv64i8u: -; ALL: ## BB#0: -; ALL-NEXT: vpxor %ymm2, %ymm2, %ymm2 -; ALL-NEXT: vpsubb %ymm0, %ymm2, %ymm3 -; ALL-NEXT: vpand %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] -; ALL-NEXT: vpsubb %ymm3, %ymm0, %ymm0 -; ALL-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm5 -; ALL-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] -; ALL-NEXT: vpshufb %ymm5, %ymm6, %ymm5 -; ALL-NEXT: vpsrlw $4, %ymm0, %ymm0 -; ALL-NEXT: vpand %ymm4, %ymm0, %ymm0 -; ALL-NEXT: vpshufb %ymm0, %ymm6, %ymm0 -; ALL-NEXT: vpaddb %ymm5, %ymm0, %ymm0 -; ALL-NEXT: vpsubb %ymm1, %ymm2, %ymm2 -; ALL-NEXT: vpand %ymm2, %ymm1, %ymm1 -; ALL-NEXT: vpsubb %ymm3, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm2 -; ALL-NEXT: vpshufb %ymm2, %ymm6, %ymm2 -; ALL-NEXT: vpsrlw $4, %ymm1, %ymm1 -; ALL-NEXT: vpand %ymm4, %ymm1, %ymm1 -; ALL-NEXT: vpshufb %ymm1, %ymm6, %ymm1 -; ALL-NEXT: vpaddb %ymm2, %ymm1, %ymm1 -; ALL-NEXT: retq +; AVX512CD-LABEL: testv64i8u: +; AVX512CD: ## BB#0: +; AVX512CD-NEXT: vpxor %ymm2, %ymm2, %ymm2 +; AVX512CD-NEXT: vpsubb %ymm0, %ymm2, %ymm3 +; AVX512CD-NEXT: vpand %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] +; AVX512CD-NEXT: vpsubb %ymm3, %ymm0, %ymm0 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm5 +; AVX512CD-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CD-NEXT: vpshufb %ymm5, %ymm6, %ymm5 +; AVX512CD-NEXT: vpsrlw $4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpand %ymm4, %ymm0, %ymm0 +; AVX512CD-NEXT: vpshufb %ymm0, %ymm6, %ymm0 +; AVX512CD-NEXT: vpaddb %ymm5, %ymm0, %ymm0 +; AVX512CD-NEXT: vpsubb %ymm1, %ymm2, %ymm2 +; AVX512CD-NEXT: vpand %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: vpsubb %ymm3, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm2 +; AVX512CD-NEXT: vpshufb %ymm2, %ymm6, %ymm2 +; AVX512CD-NEXT: vpsrlw $4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpand %ymm4, %ymm1, %ymm1 +; AVX512CD-NEXT: vpshufb %ymm1, %ymm6, %ymm1 +; AVX512CD-NEXT: vpaddb %ymm2, %ymm1, %ymm1 +; AVX512CD-NEXT: retq +; +; AVX512CDBW-LABEL: testv64i8u: +; AVX512CDBW: ## BB#0: +; AVX512CDBW-NEXT: vextracti32x4 $3, %zmm0, %xmm1 +; AVX512CDBW-NEXT: vpextrb $1, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpextrb $0, %xmm1, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm2 +; AVX512CDBW-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $2, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $3, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $4, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $5, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $6, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $7, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $8, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $9, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $10, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $11, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $12, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $13, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $14, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 +; AVX512CDBW-NEXT: vpextrb $15, %xmm1, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $15, %eax, %xmm2, %xmm1 +; AVX512CDBW-NEXT: vextracti32x4 $2, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpextrb $0, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $15, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1 +; AVX512CDBW-NEXT: vextracti32x4 $1, %zmm0, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpextrb $0, %xmm2, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm2, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $15, %eax, %xmm3, %xmm2 +; AVX512CDBW-NEXT: vpextrb $1, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpextrb $0, %xmm0, %ecx +; AVX512CDBW-NEXT: tzcntl %ecx, %ecx +; AVX512CDBW-NEXT: vmovd %ecx, %xmm3 +; AVX512CDBW-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $2, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $3, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $4, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $5, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $6, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $7, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $8, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $9, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $10, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $11, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $12, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $13, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $14, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 +; AVX512CDBW-NEXT: vpextrb $15, %xmm0, %eax +; AVX512CDBW-NEXT: tzcntl %eax, %eax +; AVX512CDBW-NEXT: vpinsrb $15, %eax, %xmm3, %xmm0 +; AVX512CDBW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 +; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: retq %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 -1) ret <64 x i8> %out } -- 2.7.4