From 340ad0a0c98fbef61086640251199cd2e59ec676 Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Thu, 21 Apr 2016 06:39:34 +0000 Subject: [PATCH] [AVX512] Add support for lowering CTTZ v64i8 and v32i16 with BWI instructions. llvm-svn: 266963 --- llvm/lib/Target/X86/X86ISelLowering.cpp | 1 + llvm/test/CodeGen/X86/vector-tzcnt-512.ll | 854 +++++------------------------- 2 files changed, 126 insertions(+), 729 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 6182504..cff4936 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -1486,6 +1486,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM, setOperationAction(ISD::MLOAD, VT, Legal); setOperationAction(ISD::MSTORE, VT, Legal); setOperationAction(ISD::CTPOP, VT, Custom); + setOperationAction(ISD::CTTZ, VT, Custom); setOperationPromotedToType(ISD::AND, VT, MVT::v8i64); setOperationPromotedToType(ISD::OR, VT, MVT::v8i64); diff --git a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll index 5d7bcd8..262e94e 100644 --- a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll +++ b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll @@ -1,6 +1,7 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; 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 +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512BW define <8 x i64> @testv8i64(<8 x i64> %in) nounwind { ; ALL-LABEL: testv8i64: @@ -169,109 +170,41 @@ define <32 x i16> @testv32i16(<32 x i16> %in) nounwind { ; ; 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: vpxord %zmm1, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpsubw %zmm0, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0 +; AVX512CDBW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512CDBW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CDBW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512CDBW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512CDBW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsllw $8, %zmm0, %zmm1 +; AVX512CDBW-NEXT: vpaddb %zmm0, %zmm1, %zmm0 +; AVX512CDBW-NEXT: vpsrlw $8, %zmm0, %zmm0 ; AVX512CDBW-NEXT: retq +; +; AVX512BW-LABEL: testv32i16: +; AVX512BW: ## BB#0: +; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1 +; AVX512BW-NEXT: vpsubw %zmm0, %zmm1, %zmm1 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0 +; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsllw $8, %zmm0, %zmm1 +; AVX512BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0 +; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0 +; AVX512BW-NEXT: retq %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 0) ret <32 x i16> %out } @@ -311,109 +244,41 @@ define <32 x i16> @testv32i16u(<32 x i16> %in) nounwind { ; ; 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: vpxord %zmm1, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpsubw %zmm0, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0 +; AVX512CDBW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512CDBW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CDBW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512CDBW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512CDBW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsllw $8, %zmm0, %zmm1 +; AVX512CDBW-NEXT: vpaddb %zmm0, %zmm1, %zmm0 +; AVX512CDBW-NEXT: vpsrlw $8, %zmm0, %zmm0 ; AVX512CDBW-NEXT: retq +; +; AVX512BW-LABEL: testv32i16u: +; AVX512BW: ## BB#0: +; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1 +; AVX512BW-NEXT: vpsubw %zmm0, %zmm1, %zmm1 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0 +; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsllw $8, %zmm0, %zmm1 +; AVX512BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0 +; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0 +; AVX512BW-NEXT: retq %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 -1) ret <32 x i16> %out } @@ -447,334 +312,35 @@ define <64 x i8> @testv64i8(<64 x i8> %in) nounwind { ; ; 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: vpxord %zmm1, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpsubb %zmm0, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0 +; AVX512CDBW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512CDBW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CDBW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512CDBW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512CDBW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 ; AVX512CDBW-NEXT: retq +; +; AVX512BW-LABEL: testv64i8: +; AVX512BW: ## BB#0: +; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1 +; AVX512BW-NEXT: vpsubb %zmm0, %zmm1, %zmm1 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0 +; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512BW-NEXT: retq %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 0) ret <64 x i8> %out } @@ -808,205 +374,35 @@ define <64 x i8> @testv64i8u(<64 x i8> %in) nounwind { ; ; 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: vpxord %zmm1, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpsubb %zmm0, %zmm1, %zmm1 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0 +; AVX512CDBW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512CDBW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512CDBW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512CDBW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512CDBW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512CDBW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 ; AVX512CDBW-NEXT: retq +; +; AVX512BW-LABEL: testv64i8u: +; AVX512BW: ## BB#0: +; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1 +; AVX512BW-NEXT: vpsubb %zmm0, %zmm1, %zmm1 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0 +; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15] +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2 +; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4] +; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2 +; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0 +; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 +; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 +; AVX512BW-NEXT: retq %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 -1) ret <64 x i8> %out } -- 2.7.4