diff options
Diffstat (limited to 'llvm/test')
-rw-r--r-- | llvm/test/CodeGen/X86/vector-popcnt-128.ll | 28 | ||||
-rw-r--r-- | llvm/test/CodeGen/X86/vector-popcnt-256.ll | 26 | ||||
-rw-r--r-- | llvm/test/CodeGen/X86/vector-popcnt-512.ll | 11 | ||||
-rw-r--r-- | llvm/test/CodeGen/X86/vector-tzcnt-128.ll | 46 | ||||
-rw-r--r-- | llvm/test/CodeGen/X86/vector-tzcnt-256.ll | 44 | ||||
-rw-r--r-- | llvm/test/CodeGen/X86/vector-tzcnt-512.ll | 21 | ||||
-rw-r--r-- | llvm/test/MC/X86/avx512bitalg-encoding.s | 114 | ||||
-rw-r--r-- | llvm/test/MC/X86/avx512vl_bitalg-encoding.s | 226 |
8 files changed, 516 insertions, 0 deletions
diff --git a/llvm/test/CodeGen/X86/vector-popcnt-128.ll b/llvm/test/CodeGen/X86/vector-popcnt-128.ll index d2f33785530..9a0f85bad77 100644 --- a/llvm/test/CodeGen/X86/vector-popcnt-128.ll +++ b/llvm/test/CodeGen/X86/vector-popcnt-128.ll @@ -6,6 +6,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512VPOPCNTDQ +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=BITALG_NOVLX +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=BITALG define <2 x i64> @testv2i64(<2 x i64> %in) nounwind { ; SSE2-LABEL: testv2i64: @@ -381,6 +383,19 @@ define <8 x i16> @testv8i16(<8 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpmovqw %zmm0, %xmm0 ; AVX512VPOPCNTDQ-NEXT: vzeroupper ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG_NOVLX-LABEL: testv8i16: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<def> +; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: vzeroupper +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv8i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntw %xmm0, %xmm0 +; BITALG-NEXT: retq %out = call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %in) ret <8 x i16> %out } @@ -485,6 +500,19 @@ define <16 x i8> @testv16i8(<16 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpmovdb %zmm0, %xmm0 ; AVX512VPOPCNTDQ-NEXT: vzeroupper ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG_NOVLX-LABEL: testv16i8: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<def> +; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: vzeroupper +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv16i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntb %xmm0, %xmm0 +; BITALG-NEXT: retq %out = call <16 x i8> @llvm.ctpop.v16i8(<16 x i8> %in) ret <16 x i8> %out } diff --git a/llvm/test/CodeGen/X86/vector-popcnt-256.ll b/llvm/test/CodeGen/X86/vector-popcnt-256.ll index b2743f1efdd..7f7fe1f8178 100644 --- a/llvm/test/CodeGen/X86/vector-popcnt-256.ll +++ b/llvm/test/CodeGen/X86/vector-popcnt-256.ll @@ -2,6 +2,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512VPOPCNTDQ +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=BITALG_NOVLX +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=BITALG define <4 x i64> @testv4i64(<4 x i64> %in) nounwind { ; AVX1-LABEL: testv4i64: @@ -159,6 +161,18 @@ define <16 x i16> @testv16i16(<16 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0 ; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm0, %ymm0 ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG_NOVLX-LABEL: testv16i16: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> +; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv16i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntw %ymm0, %ymm0 +; BITALG-NEXT: retq %out = call <16 x i16> @llvm.ctpop.v16i16(<16 x i16> %in) ret <16 x i16> %out } @@ -207,6 +221,18 @@ define <32 x i8> @testv32i8(<32 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0 ; AVX512VPOPCNTDQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0 ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG_NOVLX-LABEL: testv32i8: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> +; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv32i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntb %ymm0, %ymm0 +; BITALG-NEXT: retq %out = call <32 x i8> @llvm.ctpop.v32i8(<32 x i8> %in) ret <32 x i8> %out } diff --git a/llvm/test/CodeGen/X86/vector-popcnt-512.ll b/llvm/test/CodeGen/X86/vector-popcnt-512.ll index 4421f32c56c..6bd64ccdd99 100644 --- a/llvm/test/CodeGen/X86/vector-popcnt-512.ll +++ b/llvm/test/CodeGen/X86/vector-popcnt-512.ll @@ -3,6 +3,7 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512BW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512VPOPCNTDQ --check-prefix=AVX512VPOPCNTDQ-NOBW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq,+avx512bw | FileCheck %s --check-prefix=AVX512 --check-prefix=AVX512VPOPCNTDQ --check-prefix=AVX512VPOPCNTDQ-BW +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=AVX512 --check-prefix=BITALG define <8 x i64> @testv8i64(<8 x i64> %in) nounwind { ; AVX512F-LABEL: testv8i64: @@ -172,6 +173,11 @@ define <32 x i16> @testv32i16(<32 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0 ; AVX512VPOPCNTDQ-BW-NEXT: vpsrlw $8, %zmm0, %zmm0 ; AVX512VPOPCNTDQ-BW-NEXT: retq +; +; BITALG-LABEL: testv32i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG-NEXT: retq %out = call <32 x i16> @llvm.ctpop.v32i16(<32 x i16> %in) ret <32 x i16> %out } @@ -236,6 +242,11 @@ define <64 x i8> @testv64i8(<64 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0 ; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0 ; AVX512VPOPCNTDQ-BW-NEXT: retq +; +; BITALG-LABEL: testv64i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG-NEXT: retq %out = call <64 x i8> @llvm.ctpop.v64i8(<64 x i8> %in) ret <64 x i8> %out } diff --git a/llvm/test/CodeGen/X86/vector-tzcnt-128.ll b/llvm/test/CodeGen/X86/vector-tzcnt-128.ll index 5f00e55e225..bf604f7df95 100644 --- a/llvm/test/CodeGen/X86/vector-tzcnt-128.ll +++ b/llvm/test/CodeGen/X86/vector-tzcnt-128.ll @@ -8,6 +8,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CDVL ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,-avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CD ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX512VPOPCNTDQ +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=BITALG_NOVLX +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=BITALG ; ; Just one 32-bit run to make sure we do reasonable things for i64 tzcnt. ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=X32-SSE --check-prefix=X32-SSE41 @@ -934,6 +936,28 @@ define <8 x i16> @testv8i16(<8 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vzeroupper ; AVX512VPOPCNTDQ-NEXT: retq ; +; BITALG_NOVLX-LABEL: testv8i16: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpsubw %xmm0, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 +; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpaddw %xmm1, %xmm0, %xmm0 +; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: vzeroupper +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv8i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubw %xmm0, %xmm1, %xmm1 +; BITALG-NEXT: vpand %xmm1, %xmm0, %xmm0 +; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpaddw %xmm1, %xmm0, %xmm0 +; BITALG-NEXT: vpopcntw %xmm0, %xmm0 +; BITALG-NEXT: retq +; ; X32-SSE-LABEL: testv8i16: ; X32-SSE: # BB#0: ; X32-SSE-NEXT: pxor %xmm1, %xmm1 @@ -1235,6 +1259,28 @@ define <16 x i8> @testv16i8(<16 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vzeroupper ; AVX512VPOPCNTDQ-NEXT: retq ; +; BITALG_NOVLX-LABEL: testv16i8: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpsubb %xmm0, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 +; BITALG_NOVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpaddb %xmm1, %xmm0, %xmm0 +; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %XMM0<def> %XMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: vzeroupper +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv16i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubb %xmm0, %xmm1, %xmm1 +; BITALG-NEXT: vpand %xmm1, %xmm0, %xmm0 +; BITALG-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpaddb %xmm1, %xmm0, %xmm0 +; BITALG-NEXT: vpopcntb %xmm0, %xmm0 +; BITALG-NEXT: retq +; ; X32-SSE-LABEL: testv16i8: ; X32-SSE: # BB#0: ; X32-SSE-NEXT: pxor %xmm1, %xmm1 diff --git a/llvm/test/CodeGen/X86/vector-tzcnt-256.ll b/llvm/test/CodeGen/X86/vector-tzcnt-256.ll index 39d6d4c35ed..7b6fe21ca4d 100644 --- a/llvm/test/CodeGen/X86/vector-tzcnt-256.ll +++ b/llvm/test/CodeGen/X86/vector-tzcnt-256.ll @@ -4,6 +4,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CDVL ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,-avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512CD ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512VPOPCNTDQ +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=BITALG_NOVLX +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=BITALG ; ; Just one 32-bit run to make sure we do reasonable things for i64 tzcnt. ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=X32-AVX --check-prefix=X32-AVX2 @@ -589,6 +591,27 @@ define <16 x i16> @testv16i16(<16 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm0, %ymm0 ; AVX512VPOPCNTDQ-NEXT: retq ; +; BITALG_NOVLX-LABEL: testv16i16: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpsubw %ymm0, %ymm1, %ymm1 +; BITALG_NOVLX-NEXT: vpand %ymm1, %ymm0, %ymm0 +; BITALG_NOVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 +; BITALG_NOVLX-NEXT: vpaddw %ymm1, %ymm0, %ymm0 +; BITALG_NOVLX-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv16i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubw %ymm0, %ymm1, %ymm1 +; BITALG-NEXT: vpand %ymm1, %ymm0, %ymm0 +; BITALG-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 +; BITALG-NEXT: vpaddw %ymm1, %ymm0, %ymm0 +; BITALG-NEXT: vpopcntw %ymm0, %ymm0 +; BITALG-NEXT: retq +; ; X32-AVX-LABEL: testv16i16: ; X32-AVX: # BB#0: ; X32-AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1 @@ -839,6 +862,27 @@ define <32 x i8> @testv32i8(<32 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0 ; AVX512VPOPCNTDQ-NEXT: retq ; +; BITALG_NOVLX-LABEL: testv32i8: +; BITALG_NOVLX: # BB#0: +; BITALG_NOVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG_NOVLX-NEXT: vpsubb %ymm0, %ymm1, %ymm1 +; BITALG_NOVLX-NEXT: vpand %ymm1, %ymm0, %ymm0 +; BITALG_NOVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 +; BITALG_NOVLX-NEXT: vpaddb %ymm1, %ymm0, %ymm0 +; BITALG_NOVLX-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG_NOVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill> +; BITALG_NOVLX-NEXT: retq +; +; BITALG-LABEL: testv32i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubb %ymm0, %ymm1, %ymm1 +; BITALG-NEXT: vpand %ymm1, %ymm0, %ymm0 +; BITALG-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 +; BITALG-NEXT: vpaddb %ymm1, %ymm0, %ymm0 +; BITALG-NEXT: vpopcntb %ymm0, %ymm0 +; BITALG-NEXT: retq +; ; X32-AVX-LABEL: testv32i8: ; X32-AVX: # BB#0: ; X32-AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1 diff --git a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll index 4d3858863e3..0890ae5012c 100644 --- a/llvm/test/CodeGen/X86/vector-tzcnt-512.ll +++ b/llvm/test/CodeGen/X86/vector-tzcnt-512.ll @@ -3,6 +3,7 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512cd,+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512CDBW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512BW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefix=ALL --check-prefix=AVX512VPOPCNTDQ +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=ALL --check-prefix=BITALG define <8 x i64> @testv8i64(<8 x i64> %in) nounwind { ; AVX512CD-LABEL: testv8i64: @@ -374,6 +375,16 @@ define <32 x i16> @testv32i16(<32 x i16> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm1, %zmm1 ; AVX512VPOPCNTDQ-NEXT: vpmovdw %zmm1, %ymm1 ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG-LABEL: testv32i16: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubw %zmm0, %zmm1, %zmm1 +; BITALG-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; BITALG-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 +; BITALG-NEXT: vpaddw %zmm1, %zmm0, %zmm0 +; BITALG-NEXT: vpopcntw %zmm0, %zmm0 +; BITALG-NEXT: retq %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 0) ret <32 x i16> %out } @@ -558,6 +569,16 @@ define <64 x i8> @testv64i8(<64 x i8> %in) nounwind { ; AVX512VPOPCNTDQ-NEXT: vpshufb %ymm1, %ymm6, %ymm1 ; AVX512VPOPCNTDQ-NEXT: vpaddb %ymm2, %ymm1, %ymm1 ; AVX512VPOPCNTDQ-NEXT: retq +; +; BITALG-LABEL: testv64i8: +; BITALG: # BB#0: +; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1 +; BITALG-NEXT: vpsubb %zmm0, %zmm1, %zmm1 +; BITALG-NEXT: vpandq %zmm1, %zmm0, %zmm0 +; BITALG-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 +; BITALG-NEXT: vpaddb %zmm1, %zmm0, %zmm0 +; BITALG-NEXT: vpopcntb %zmm0, %zmm0 +; BITALG-NEXT: retq %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 0) ret <64 x i8> %out } diff --git a/llvm/test/MC/X86/avx512bitalg-encoding.s b/llvm/test/MC/X86/avx512bitalg-encoding.s new file mode 100644 index 00000000000..bd128bd129a --- /dev/null +++ b/llvm/test/MC/X86/avx512bitalg-encoding.s @@ -0,0 +1,114 @@ +// RUN: llvm-mc -triple x86_64-unknown-unknown -mcpu=knl -mattr=+avx512bitalg --show-encoding < %s | FileCheck %s + +// CHECK: vpopcntb %zmm23, %zmm21 +// CHECK: encoding: [0x62,0xa2,0x7d,0x48,0x54,0xef] + vpopcntb %zmm23, %zmm21 + +// CHECK: vpopcntw %zmm23, %zmm21 +// CHECK: encoding: [0x62,0xa2,0xfd,0x48,0x54,0xef] + vpopcntw %zmm23, %zmm21 + +// CHECK: vpopcntb %zmm3, %zmm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0x7d,0x4a,0x54,0xcb] + vpopcntb %zmm3, %zmm1 {%k2} + +// CHECK: vpopcntw %zmm3, %zmm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0xfd,0x4a,0x54,0xcb] + vpopcntw %zmm3, %zmm1 {%k2} + +// CHECK: vpopcntb (%rcx), %zmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x48,0x54,0x09] + vpopcntb (%rcx), %zmm1 + +// CHECK: vpopcntb -256(%rsp), %zmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x48,0x54,0x4c,0x24,0xfc] + vpopcntb -256(%rsp), %zmm1 + +// CHECK: vpopcntb 256(%rsp), %zmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x48,0x54,0x4c,0x24,0x04] + vpopcntb 256(%rsp), %zmm1 + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x48,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x48,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x48,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntw (%rcx), %zmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x48,0x54,0x09] + vpopcntw (%rcx), %zmm1 + +// CHECK: vpopcntw -256(%rsp), %zmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x48,0x54,0x4c,0x24,0xfc] + vpopcntw -256(%rsp), %zmm1 + +// CHECK: vpopcntw 256(%rsp), %zmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x48,0x54,0x4c,0x24,0x04] + vpopcntw 256(%rsp), %zmm1 + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x48,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x48,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %zmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x48,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %zmm1 + +// CHECK: vpopcntb (%rcx), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x4a,0x54,0x29] + vpopcntb (%rcx), %zmm21 {%k2} + +// CHECK: vpopcntb -256(%rsp), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x4a,0x54,0x6c,0x24,0xfc] + vpopcntb -256(%rsp), %zmm21 {%k2} + +// CHECK: vpopcntb 256(%rsp), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x4a,0x54,0x6c,0x24,0x04] + vpopcntb 256(%rsp), %zmm21 {%k2} + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x4a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %zmm21 {%k2} + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x4a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %zmm21 {%k2} + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x4a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %zmm21 {%k2} + +// CHECK: vpopcntw (%rcx), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x4a,0x54,0x29] + vpopcntw (%rcx), %zmm21 {%k2} + +// CHECK: vpopcntw -256(%rsp), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x4a,0x54,0x6c,0x24,0xfc] + vpopcntw -256(%rsp), %zmm21 {%k2} + +// CHECK: vpopcntw 256(%rsp), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x4a,0x54,0x6c,0x24,0x04] + vpopcntw 256(%rsp), %zmm21 {%k2} + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x4a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %zmm21 {%k2} + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x4a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %zmm21 {%k2} + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %zmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x4a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %zmm21 {%k2} + diff --git a/llvm/test/MC/X86/avx512vl_bitalg-encoding.s b/llvm/test/MC/X86/avx512vl_bitalg-encoding.s new file mode 100644 index 00000000000..53898664b5e --- /dev/null +++ b/llvm/test/MC/X86/avx512vl_bitalg-encoding.s @@ -0,0 +1,226 @@ +// RUN: llvm-mc -triple x86_64-unknown-unknown -mcpu=knl -mattr=+avx512vl,+avx512bitalg --show-encoding < %s | FileCheck %s + +// CHECK: vpopcntb %xmm23, %xmm21 +// CHECK: encoding: [0x62,0xa2,0x7d,0x08,0x54,0xef] + vpopcntb %xmm23, %xmm21 + +// CHECK: vpopcntw %xmm23, %xmm21 +// CHECK: encoding: [0x62,0xa2,0xfd,0x08,0x54,0xef] + vpopcntw %xmm23, %xmm21 + +// CHECK: vpopcntb %xmm3, %xmm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0x7d,0x0a,0x54,0xcb] + vpopcntb %xmm3, %xmm1 {%k2} + +// CHECK: vpopcntw %xmm3, %xmm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0xfd,0x0a,0x54,0xcb] + vpopcntw %xmm3, %xmm1 {%k2} + +// CHECK: vpopcntb (%rcx), %xmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x08,0x54,0x09] + vpopcntb (%rcx), %xmm1 + +// CHECK: vpopcntb -64(%rsp), %xmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x08,0x54,0x4c,0x24,0xfc] + vpopcntb -64(%rsp), %xmm1 + +// CHECK: vpopcntb 64(%rsp), %xmm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x08,0x54,0x4c,0x24,0x04] + vpopcntb 64(%rsp), %xmm1 + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x08,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x08,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x08,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntw (%rcx), %xmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x08,0x54,0x09] + vpopcntw (%rcx), %xmm1 + +// CHECK: vpopcntw -64(%rsp), %xmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x08,0x54,0x4c,0x24,0xfc] + vpopcntw -64(%rsp), %xmm1 + +// CHECK: vpopcntw 64(%rsp), %xmm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x08,0x54,0x4c,0x24,0x04] + vpopcntw 64(%rsp), %xmm1 + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x08,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x08,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %xmm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x08,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %xmm1 + +// CHECK: vpopcntb (%rcx), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x0a,0x54,0x29] + vpopcntb (%rcx), %xmm21 {%k2} + +// CHECK: vpopcntb -64(%rsp), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x0a,0x54,0x6c,0x24,0xfc] + vpopcntb -64(%rsp), %xmm21 {%k2} + +// CHECK: vpopcntb 64(%rsp), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x0a,0x54,0x6c,0x24,0x04] + vpopcntb 64(%rsp), %xmm21 {%k2} + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x0a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x0a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x0a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntw (%rcx), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x0a,0x54,0x29] + vpopcntw (%rcx), %xmm21 {%k2} + +// CHECK: vpopcntw -64(%rsp), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x0a,0x54,0x6c,0x24,0xfc] + vpopcntw -64(%rsp), %xmm21 {%k2} + +// CHECK: vpopcntw 64(%rsp), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x0a,0x54,0x6c,0x24,0x04] + vpopcntw 64(%rsp), %xmm21 {%k2} + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x0a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x0a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %xmm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x0a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %xmm21 {%k2} + +// CHECK: vpopcntb %ymm23, %ymm21 +// CHECK: encoding: [0x62,0xa2,0x7d,0x28,0x54,0xef] + vpopcntb %ymm23, %ymm21 + +// CHECK: vpopcntw %ymm23, %ymm21 +// CHECK: encoding: [0x62,0xa2,0xfd,0x28,0x54,0xef] + vpopcntw %ymm23, %ymm21 + +// CHECK: vpopcntb %ymm3, %ymm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0x7d,0x2a,0x54,0xcb] + vpopcntb %ymm3, %ymm1 {%k2} + +// CHECK: vpopcntw %ymm3, %ymm1 {%k2} +// CHECK: encoding: [0x62,0xf2,0xfd,0x2a,0x54,0xcb] + vpopcntw %ymm3, %ymm1 {%k2} + +// CHECK: vpopcntb (%rcx), %ymm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x28,0x54,0x09] + vpopcntb (%rcx), %ymm1 + +// CHECK: vpopcntb -128(%rsp), %ymm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x28,0x54,0x4c,0x24,0xfc] + vpopcntb -128(%rsp), %ymm1 + +// CHECK: vpopcntb 128(%rsp), %ymm1 +// CHECK: encoding: [0x62,0xf2,0x7d,0x28,0x54,0x4c,0x24,0x04] + vpopcntb 128(%rsp), %ymm1 + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x28,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x28,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0x7d,0x28,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntw (%rcx), %ymm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x28,0x54,0x09] + vpopcntw (%rcx), %ymm1 + +// CHECK: vpopcntw -128(%rsp), %ymm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x28,0x54,0x4c,0x24,0xfc] + vpopcntw -128(%rsp), %ymm1 + +// CHECK: vpopcntw 128(%rsp), %ymm1 +// CHECK: encoding: [0x62,0xf2,0xfd,0x28,0x54,0x4c,0x24,0x04] + vpopcntw 128(%rsp), %ymm1 + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x28,0x54,0x8c,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x28,0x54,0x8c,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %ymm1 +// CHECK: encoding: [0x62,0xb2,0xfd,0x28,0x54,0x8c,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %ymm1 + +// CHECK: vpopcntb (%rcx), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x2a,0x54,0x29] + vpopcntb (%rcx), %ymm21 {%k2} + +// CHECK: vpopcntb -128(%rsp), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x2a,0x54,0x6c,0x24,0xfc] + vpopcntb -128(%rsp), %ymm21 {%k2} + +// CHECK: vpopcntb 128(%rsp), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0x7d,0x2a,0x54,0x6c,0x24,0x04] + vpopcntb 128(%rsp), %ymm21 {%k2} + +// CHECK: vpopcntb 268435456(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x2a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntb 268435456(%rcx,%r14,8), %ymm21 {%k2} + +// CHECK: vpopcntb -536870912(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x2a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntb -536870912(%rcx,%r14,8), %ymm21 {%k2} + +// CHECK: vpopcntb -536870910(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0x7d,0x2a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntb -536870910(%rcx,%r14,8), %ymm21 {%k2} + +// CHECK: vpopcntw (%rcx), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x2a,0x54,0x29] + vpopcntw (%rcx), %ymm21 {%k2} + +// CHECK: vpopcntw -128(%rsp), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x2a,0x54,0x6c,0x24,0xfc] + vpopcntw -128(%rsp), %ymm21 {%k2} + +// CHECK: vpopcntw 128(%rsp), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xe2,0xfd,0x2a,0x54,0x6c,0x24,0x04] + vpopcntw 128(%rsp), %ymm21 {%k2} + +// CHECK: vpopcntw 268435456(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x2a,0x54,0xac,0xf1,0x00,0x00,0x00,0x10] + vpopcntw 268435456(%rcx,%r14,8), %ymm21 {%k2} + +// CHECK: vpopcntw -536870912(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x2a,0x54,0xac,0xf1,0x00,0x00,0x00,0xe0] + vpopcntw -536870912(%rcx,%r14,8), %ymm21 {%k2} + +// CHECK: vpopcntw -536870910(%rcx,%r14,8), %ymm21 {%k2} +// CHECK: encoding: [0x62,0xa2,0xfd,0x2a,0x54,0xac,0xf1,0x02,0x00,0x00,0xe0] + vpopcntw -536870910(%rcx,%r14,8), %ymm21 {%k2} + |