From 77ab058bf80f0f04a2c8d22fd375795d6ddd7483 Mon Sep 17 00:00:00 2001 From: Elena Demikhovsky Date: Sun, 1 Nov 2015 11:45:47 +0000 Subject: [PATCH] AVX-512: Optimized SIMD truncate operations for AVX512F set. Optimized <8 x i32> to <8 x i16> <4 x i64> to < 4 x i32> <16 x i16> to <16 x i8> All these oprtrations use now AVX512F set (KNL). Before this change it was implemented with AVX2 set. Differential Revision: http://reviews.llvm.org/D14108 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251764 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/X86/X86ISelLowering.cpp | 9 +- lib/Target/X86/X86InstrAVX512.td | 17 ++ test/CodeGen/X86/avx512-trunc.ll | 275 +++++++++++++++++++-------- test/CodeGen/X86/vector-lzcnt-128.ll | 6 +- 4 files changed, 224 insertions(+), 83 deletions(-) diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index 74299e92d4a..3a013a24111 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -13102,10 +13102,13 @@ SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { } // vpmovqb/w/d, vpmovdb/w, vpmovwb - if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) && - (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI())) + if (Subtarget->hasAVX512()) { + // word to byte only under BWI + if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8 + return DAG.getNode(X86ISD::VTRUNC, DL, VT, + DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In)); return DAG.getNode(X86ISD::VTRUNC, DL, VT, In); - + } if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) { // On AVX2, v4i64 -> v4i32 becomes VPERMD. if (Subtarget->hasInt256()) { diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td index ca124ed253c..60ba687da62 100644 --- a/lib/Target/X86/X86InstrAVX512.td +++ b/lib/Target/X86/X86InstrAVX512.td @@ -6036,6 +6036,23 @@ defm VPMOVWB : avx512_trunc_wb<0x30, "vpmovwb", X86vtrunc>; defm VPMOVSWB : avx512_trunc_sat_wb<0x20, "s", X86vtruncs>; defm VPMOVUSWB : avx512_trunc_sat_wb<0x10, "us", X86vtruncus>; +let Predicates = [HasAVX512, NoVLX] in { +def: Pat<(v8i16 (X86vtrunc (v8i32 VR256X:$src))), + (v8i16 (EXTRACT_SUBREG + (v16i16 (VPMOVDWZrr (v16i32 (SUBREG_TO_REG (i32 0), + VR256X:$src, sub_ymm)))), sub_xmm))>; +def: Pat<(v4i32 (X86vtrunc (v4i64 VR256X:$src))), + (v4i32 (EXTRACT_SUBREG + (v8i32 (VPMOVQDZrr (v8i64 (SUBREG_TO_REG (i32 0), + VR256X:$src, sub_ymm)))), sub_xmm))>; +} + +let Predicates = [HasBWI, NoVLX] in { +def: Pat<(v16i8 (X86vtrunc (v16i16 VR256X:$src))), + (v16i8 (EXTRACT_SUBREG (VPMOVWBZrr (v32i16 (SUBREG_TO_REG (i32 0), + VR256X:$src, sub_ymm))), sub_xmm))>; +} + multiclass avx512_extend_common opc, string OpcodeStr, X86VectorVTInfo DestInfo, X86VectorVTInfo SrcInfo, X86MemOperand x86memop, PatFrag LdFrag, SDNode OpNode>{ diff --git a/test/CodeGen/X86/avx512-trunc.ll b/test/CodeGen/X86/avx512-trunc.ll index 9205feda7eb..833a057b9a3 100644 --- a/test/CodeGen/X86/avx512-trunc.ll +++ b/test/CodeGen/X86/avx512-trunc.ll @@ -1,52 +1,60 @@ -; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck %s --check-prefix=KNL -; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx | FileCheck %s --check-prefix=SKX - +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f | FileCheck %s --check-prefix=ALL --check-prefix=KNL +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f -mattr=+avx512vl -mattr=+avx512bw -mattr=+avx512dq | FileCheck %s --check-prefix=ALL --check-prefix=SKX + attributes #0 = { nounwind } - -; KNL-LABEL: trunc_16x32_to_16x8 -; KNL: vpmovdb -; KNL: ret + define <16 x i8> @trunc_16x32_to_16x8(<16 x i32> %i) #0 { +; ALL-LABEL: trunc_16x32_to_16x8: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdb %zmm0, %xmm0 +; ALL-NEXT: retq %x = trunc <16 x i32> %i to <16 x i8> ret <16 x i8> %x } -; KNL-LABEL: trunc_8x64_to_8x16 -; KNL: vpmovqw -; KNL: ret define <8 x i16> @trunc_8x64_to_8x16(<8 x i64> %i) #0 { +; ALL-LABEL: trunc_8x64_to_8x16: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqw %zmm0, %xmm0 +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i16> ret <8 x i16> %x } -; KNL-LABEL: trunc_v16i32_to_v16i16 -; KNL: vpmovdw -; KNL: ret define <16 x i16> @trunc_v16i32_to_v16i16(<16 x i32> %x) #0 { +; ALL-LABEL: trunc_v16i32_to_v16i16: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdw %zmm0, %ymm0 +; ALL-NEXT: retq %1 = trunc <16 x i32> %x to <16 x i16> ret <16 x i16> %1 } define <8 x i8> @trunc_qb_512(<8 x i64> %i) #0 { -; SKX-LABEL: trunc_qb_512: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqw %zmm0, %xmm0 -; SKX-NEXT: retq +; ALL-LABEL: trunc_qb_512: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqw %zmm0, %xmm0 +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i8> ret <8 x i8> %x } define void @trunc_qb_512_mem(<8 x i64> %i, <8 x i8>* %res) #0 { -; SKX-LABEL: trunc_qb_512_mem: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqb %zmm0, (%rdi) -; SKX-NEXT: retq +; ALL-LABEL: trunc_qb_512_mem: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqb %zmm0, (%rdi) +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i8> store <8 x i8> %x, <8 x i8>* %res ret void } define <4 x i8> @trunc_qb_256(<4 x i64> %i) #0 { +; KNL-LABEL: trunc_qb_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qb_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqd %ymm0, %xmm0 @@ -56,6 +64,13 @@ define <4 x i8> @trunc_qb_256(<4 x i64> %i) #0 { } define void @trunc_qb_256_mem(<4 x i64> %i, <4 x i8>* %res) #0 { +; KNL-LABEL: trunc_qb_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u] +; KNL-NEXT: vmovd %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qb_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqb %ymm0, (%rdi) @@ -66,14 +81,21 @@ define void @trunc_qb_256_mem(<4 x i64> %i, <4 x i8>* %res) #0 { } define <2 x i8> @trunc_qb_128(<2 x i64> %i) #0 { -; SKX-LABEL: trunc_qb_128: -; SKX: ## BB#0: -; SKX-NEXT: retq +; ALL-LABEL: trunc_qb_128: +; ALL: ## BB#0: +; ALL-NEXT: retq %x = trunc <2 x i64> %i to <2 x i8> ret <2 x i8> %x } define void @trunc_qb_128_mem(<2 x i64> %i, <2 x i8>* %res) #0 { +; KNL-LABEL: trunc_qb_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u] +; KNL-NEXT: vmovd %xmm0, %eax +; KNL-NEXT: movw %ax, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qb_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqb %xmm0, (%rdi) @@ -84,25 +106,30 @@ define void @trunc_qb_128_mem(<2 x i64> %i, <2 x i8>* %res) #0 { } define <8 x i16> @trunc_qw_512(<8 x i64> %i) #0 { -; SKX-LABEL: trunc_qw_512: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqw %zmm0, %xmm0 -; SKX-NEXT: retq +; ALL-LABEL: trunc_qw_512: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqw %zmm0, %xmm0 +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i16> ret <8 x i16> %x } define void @trunc_qw_512_mem(<8 x i64> %i, <8 x i16>* %res) #0 { -; SKX-LABEL: trunc_qw_512_mem: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqw %zmm0, (%rdi) -; SKX-NEXT: retq +; ALL-LABEL: trunc_qw_512_mem: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqw %zmm0, (%rdi) +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i16> store <8 x i16> %x, <8 x i16>* %res ret void } define <4 x i16> @trunc_qw_256(<4 x i64> %i) #0 { +; KNL-LABEL: trunc_qw_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qw_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqd %ymm0, %xmm0 @@ -112,6 +139,13 @@ define <4 x i16> @trunc_qw_256(<4 x i64> %i) #0 { } define void @trunc_qw_256_mem(<4 x i64> %i, <4 x i16>* %res) #0 { +; KNL-LABEL: trunc_qw_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] +; KNL-NEXT: vmovq %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qw_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqw %ymm0, (%rdi) @@ -122,14 +156,21 @@ define void @trunc_qw_256_mem(<4 x i64> %i, <4 x i16>* %res) #0 { } define <2 x i16> @trunc_qw_128(<2 x i64> %i) #0 { -; SKX-LABEL: trunc_qw_128: -; SKX: ## BB#0: -; SKX-NEXT: retq +; ALL-LABEL: trunc_qw_128: +; ALL: ## BB#0: +; ALL-NEXT: retq %x = trunc <2 x i64> %i to <2 x i16> ret <2 x i16> %x } define void @trunc_qw_128_mem(<2 x i64> %i, <2 x i16>* %res) #0 { +; KNL-LABEL: trunc_qw_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] +; KNL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] +; KNL-NEXT: vmovd %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qw_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqw %xmm0, (%rdi) @@ -140,25 +181,30 @@ define void @trunc_qw_128_mem(<2 x i64> %i, <2 x i16>* %res) #0 { } define <8 x i32> @trunc_qd_512(<8 x i64> %i) #0 { -; SKX-LABEL: trunc_qd_512: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqd %zmm0, %ymm0 -; SKX-NEXT: retq +; ALL-LABEL: trunc_qd_512: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqd %zmm0, %ymm0 +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i32> ret <8 x i32> %x } define void @trunc_qd_512_mem(<8 x i64> %i, <8 x i32>* %res) #0 { -; SKX-LABEL: trunc_qd_512_mem: -; SKX: ## BB#0: -; SKX-NEXT: vpmovqd %zmm0, (%rdi) -; SKX-NEXT: retq +; ALL-LABEL: trunc_qd_512_mem: +; ALL: ## BB#0: +; ALL-NEXT: vpmovqd %zmm0, (%rdi) +; ALL-NEXT: retq %x = trunc <8 x i64> %i to <8 x i32> store <8 x i32> %x, <8 x i32>* %res ret void } define <4 x i32> @trunc_qd_256(<4 x i64> %i) #0 { +; KNL-LABEL: trunc_qd_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qd_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqd %ymm0, %xmm0 @@ -168,6 +214,12 @@ define <4 x i32> @trunc_qd_256(<4 x i64> %i) #0 { } define void @trunc_qd_256_mem(<4 x i64> %i, <4 x i32>* %res) #0 { +; KNL-LABEL: trunc_qd_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovqd %zmm0, %ymm0 +; KNL-NEXT: vmovaps %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qd_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqd %ymm0, (%rdi) @@ -178,14 +230,20 @@ define void @trunc_qd_256_mem(<4 x i64> %i, <4 x i32>* %res) #0 { } define <2 x i32> @trunc_qd_128(<2 x i64> %i) #0 { -; SKX-LABEL: trunc_qd_128: -; SKX: ## BB#0: -; SKX-NEXT: retq +; ALL-LABEL: trunc_qd_128: +; ALL: ## BB#0: +; ALL-NEXT: retq %x = trunc <2 x i64> %i to <2 x i32> ret <2 x i32> %x } define void @trunc_qd_128_mem(<2 x i64> %i, <2 x i32>* %res) #0 { +; KNL-LABEL: trunc_qd_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] +; KNL-NEXT: vmovq %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_qd_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovqd %xmm0, (%rdi) @@ -196,25 +254,30 @@ define void @trunc_qd_128_mem(<2 x i64> %i, <2 x i32>* %res) #0 { } define <16 x i8> @trunc_db_512(<16 x i32> %i) #0 { -; SKX-LABEL: trunc_db_512: -; SKX: ## BB#0: -; SKX-NEXT: vpmovdb %zmm0, %xmm0 -; SKX-NEXT: retq +; ALL-LABEL: trunc_db_512: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdb %zmm0, %xmm0 +; ALL-NEXT: retq %x = trunc <16 x i32> %i to <16 x i8> ret <16 x i8> %x } define void @trunc_db_512_mem(<16 x i32> %i, <16 x i8>* %res) #0 { -; SKX-LABEL: trunc_db_512_mem: -; SKX: ## BB#0: -; SKX-NEXT: vpmovdb %zmm0, (%rdi) -; SKX-NEXT: retq +; ALL-LABEL: trunc_db_512_mem: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdb %zmm0, (%rdi) +; ALL-NEXT: retq %x = trunc <16 x i32> %i to <16 x i8> store <16 x i8> %x, <16 x i8>* %res ret void } define <8 x i8> @trunc_db_256(<8 x i32> %i) #0 { +; KNL-LABEL: trunc_db_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovdw %zmm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_db_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdw %ymm0, %xmm0 @@ -224,6 +287,13 @@ define <8 x i8> @trunc_db_256(<8 x i32> %i) #0 { } define void @trunc_db_256_mem(<8 x i32> %i, <8 x i8>* %res) #0 { +; KNL-LABEL: trunc_db_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovdw %zmm0, %ymm0 +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] +; KNL-NEXT: vmovq %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_db_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdb %ymm0, (%rdi) @@ -234,14 +304,20 @@ define void @trunc_db_256_mem(<8 x i32> %i, <8 x i8>* %res) #0 { } define <4 x i8> @trunc_db_128(<4 x i32> %i) #0 { -; SKX-LABEL: trunc_db_128: -; SKX: ## BB#0: -; SKX-NEXT: retq +; ALL-LABEL: trunc_db_128: +; ALL: ## BB#0: +; ALL-NEXT: retq %x = trunc <4 x i32> %i to <4 x i8> ret <4 x i8> %x } define void @trunc_db_128_mem(<4 x i32> %i, <4 x i8>* %res) #0 { +; KNL-LABEL: trunc_db_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u] +; KNL-NEXT: vmovd %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_db_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdb %xmm0, (%rdi) @@ -252,25 +328,30 @@ define void @trunc_db_128_mem(<4 x i32> %i, <4 x i8>* %res) #0 { } define <16 x i16> @trunc_dw_512(<16 x i32> %i) #0 { -; SKX-LABEL: trunc_dw_512: -; SKX: ## BB#0: -; SKX-NEXT: vpmovdw %zmm0, %ymm0 -; SKX-NEXT: retq +; ALL-LABEL: trunc_dw_512: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdw %zmm0, %ymm0 +; ALL-NEXT: retq %x = trunc <16 x i32> %i to <16 x i16> ret <16 x i16> %x } define void @trunc_dw_512_mem(<16 x i32> %i, <16 x i16>* %res) #0 { -; SKX-LABEL: trunc_dw_512_mem: -; SKX: ## BB#0: -; SKX-NEXT: vpmovdw %zmm0, (%rdi) -; SKX-NEXT: retq +; ALL-LABEL: trunc_dw_512_mem: +; ALL: ## BB#0: +; ALL-NEXT: vpmovdw %zmm0, (%rdi) +; ALL-NEXT: retq %x = trunc <16 x i32> %i to <16 x i16> store <16 x i16> %x, <16 x i16>* %res ret void } define <8 x i16> @trunc_dw_256(<8 x i32> %i) #0 { +; KNL-LABEL: trunc_dw_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovdw %zmm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_dw_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdw %ymm0, %xmm0 @@ -280,6 +361,12 @@ define <8 x i16> @trunc_dw_256(<8 x i32> %i) #0 { } define void @trunc_dw_256_mem(<8 x i32> %i, <8 x i16>* %res) #0 { +; KNL-LABEL: trunc_dw_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovdw %zmm0, %ymm0 +; KNL-NEXT: vmovaps %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_dw_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdw %ymm0, (%rdi) @@ -289,15 +376,13 @@ define void @trunc_dw_256_mem(<8 x i32> %i, <8 x i16>* %res) #0 { ret void } -define <4 x i16> @trunc_dw_128(<4 x i32> %i) #0 { -; SKX-LABEL: trunc_dw_128: -; SKX: ## BB#0: -; SKX-NEXT: retq - %x = trunc <4 x i32> %i to <4 x i16> - ret <4 x i16> %x -} - define void @trunc_dw_128_mem(<4 x i32> %i, <4 x i16>* %res) #0 { +; KNL-LABEL: trunc_dw_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] +; KNL-NEXT: vmovq %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_dw_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovdw %xmm0, (%rdi) @@ -308,6 +393,15 @@ define void @trunc_dw_128_mem(<4 x i32> %i, <4 x i16>* %res) #0 { } define <32 x i8> @trunc_wb_512(<32 x i16> %i) #0 { +; KNL-LABEL: trunc_wb_512: +; KNL: ## BB#0: +; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 +; KNL-NEXT: vpmovdb %zmm0, %xmm0 +; KNL-NEXT: vpmovsxwd %ymm1, %zmm1 +; KNL-NEXT: vpmovdb %zmm1, %xmm1 +; KNL-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_wb_512: ; SKX: ## BB#0: ; SKX-NEXT: vpmovwb %zmm0, %ymm0 @@ -317,6 +411,16 @@ define <32 x i8> @trunc_wb_512(<32 x i16> %i) #0 { } define void @trunc_wb_512_mem(<32 x i16> %i, <32 x i8>* %res) #0 { +; KNL-LABEL: trunc_wb_512_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 +; KNL-NEXT: vpmovdb %zmm0, %xmm0 +; KNL-NEXT: vpmovsxwd %ymm1, %zmm1 +; KNL-NEXT: vpmovdb %zmm1, %xmm1 +; KNL-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 +; KNL-NEXT: vmovaps %ymm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_wb_512_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovwb %zmm0, (%rdi) @@ -327,6 +431,12 @@ define void @trunc_wb_512_mem(<32 x i16> %i, <32 x i8>* %res) #0 { } define <16 x i8> @trunc_wb_256(<16 x i16> %i) #0 { +; KNL-LABEL: trunc_wb_256: +; KNL: ## BB#0: +; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 +; KNL-NEXT: vpmovdb %zmm0, %xmm0 +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_wb_256: ; SKX: ## BB#0: ; SKX-NEXT: vpmovwb %ymm0, %xmm0 @@ -336,6 +446,13 @@ define <16 x i8> @trunc_wb_256(<16 x i16> %i) #0 { } define void @trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) #0 { +; KNL-LABEL: trunc_wb_256_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 +; KNL-NEXT: vpmovdb %zmm0, %xmm0 +; KNL-NEXT: vmovaps %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_wb_256_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovwb %ymm0, (%rdi) @@ -346,14 +463,20 @@ define void @trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) #0 { } define <8 x i8> @trunc_wb_128(<8 x i16> %i) #0 { -; SKX-LABEL: trunc_wb_128: -; SKX: ## BB#0: -; SKX-NEXT: retq +; ALL-LABEL: trunc_wb_128: +; ALL: ## BB#0: +; ALL-NEXT: retq %x = trunc <8 x i16> %i to <8 x i8> ret <8 x i8> %x } define void @trunc_wb_128_mem(<8 x i16> %i, <8 x i8>* %res) #0 { +; KNL-LABEL: trunc_wb_128_mem: +; KNL: ## BB#0: +; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] +; KNL-NEXT: vmovq %xmm0, (%rdi) +; KNL-NEXT: retq +; ; SKX-LABEL: trunc_wb_128_mem: ; SKX: ## BB#0: ; SKX-NEXT: vpmovwb %xmm0, (%rdi) diff --git a/test/CodeGen/X86/vector-lzcnt-128.ll b/test/CodeGen/X86/vector-lzcnt-128.ll index 2ab847ccc3d..8bf0af68e6d 100644 --- a/test/CodeGen/X86/vector-lzcnt-128.ll +++ b/test/CodeGen/X86/vector-lzcnt-128.ll @@ -753,8 +753,7 @@ define <8 x i16> @testv8i16(<8 x i16> %in) nounwind { ; AVX512CD: ## BB#0: ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0 -; AVX512CD-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,20,21,24,25,28,29],zero,zero,zero,zero,zero,zero,zero,zero -; AVX512CD-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] +; AVX512CD-NEXT: vpmovdw %zmm0, %ymm0 ; AVX512CD-NEXT: vpsubw {{.*}}(%rip), %xmm0, %xmm0 ; AVX512CD-NEXT: retq %out = call <8 x i16> @llvm.ctlz.v8i16(<8 x i16> %in, i1 0) @@ -976,8 +975,7 @@ define <8 x i16> @testv8i16u(<8 x i16> %in) nounwind { ; AVX512CD: ## BB#0: ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0 -; AVX512CD-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,20,21,24,25,28,29],zero,zero,zero,zero,zero,zero,zero,zero -; AVX512CD-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3] +; AVX512CD-NEXT: vpmovdw %zmm0, %ymm0 ; AVX512CD-NEXT: vpsubw {{.*}}(%rip), %xmm0, %xmm0 ; AVX512CD-NEXT: retq %out = call <8 x i16> @llvm.ctlz.v8i16(<8 x i16> %in, i1 -1) -- 2.34.1