[llvm] r327239 - [X86][AVX] createVariablePermute - use 2xVPERMIL+PCMPGT+SELECT for v8i32/v8f32 and v4i64/v4f64 variable permutes
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Sun Mar 11 04:52:26 PDT 2018
Author: rksimon
Date: Sun Mar 11 04:52:26 2018
New Revision: 327239
URL: http://llvm.org/viewvc/llvm-project?rev=327239&view=rev
Log:
[X86][AVX] createVariablePermute - use 2xVPERMIL+PCMPGT+SELECT for v8i32/v8f32 and v4i64/v4f64 variable permutes
As VPERMILPS/VPERMILPD only selects elements based on the bits[1:0]/bit[1] then we can permute both the (repeated) lo/hi 128-bit vectors in each case and then select between these results based on whether the index was for for lo/hi.
For v4i64/v4f64 this avoids some rather nasty v4i64 multiples on the AVX2 implementation, which seems to be worse than the extra port5 pressure from the additional shuffles/blends.
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/var-permute-256.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=327239&r1=327238&r2=327239&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Mar 11 04:52:26 2018
@@ -8035,15 +8035,24 @@ SDValue createVariablePermute(MVT VT, SD
case MVT::v8i32:
if (Subtarget.hasAVX2())
Opcode = X86ISD::VPERMV;
- else if (Subtarget.hasXOP()) {
+ else if (Subtarget.hasAVX()) {
SrcVec = DAG.getBitcast(MVT::v8f32, SrcVec);
SDValue LoLo = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec,
{0, 1, 2, 3, 0, 1, 2, 3});
SDValue HiHi = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec,
{4, 5, 6, 7, 4, 5, 6, 7});
- return DAG.getBitcast(VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v8f32,
- LoLo, HiHi, IndicesVec,
- DAG.getConstant(0, DL, MVT::i8)));
+ if (Subtarget.hasXOP())
+ return DAG.getBitcast(VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v8f32,
+ LoLo, HiHi, IndicesVec,
+ DAG.getConstant(0, DL, MVT::i8)));
+ // Permute Lo and Hi and then select based on index range.
+ // This works as VPERMILPS only uses index bits[0:1] to permute elements.
+ SDValue Res = DAG.getSelectCC(
+ DL, IndicesVec, DAG.getConstant(3, DL, MVT::v8i32),
+ DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, HiHi, IndicesVec),
+ DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, LoLo, IndicesVec),
+ ISD::CondCode::SETGT);
+ return DAG.getBitcast(VT, Res);
}
break;
case MVT::v4i64:
@@ -8060,7 +8069,7 @@ SDValue createVariablePermute(MVT VT, SD
return extract256BitVector(Res, 0, DAG, DL);
}
Opcode = X86ISD::VPERMV;
- } else if (Subtarget.hasXOP()) {
+ } else if (Subtarget.hasAVX()) {
SrcVec = DAG.getBitcast(MVT::v4f64, SrcVec);
SDValue LoLo =
DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {0, 1, 0, 1});
@@ -8068,12 +8077,18 @@ SDValue createVariablePermute(MVT VT, SD
DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {2, 3, 2, 3});
// VPERMIL2PD selects with bit#1 of the index vector, so scale IndicesVec.
IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec);
- return DAG.getBitcast(VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v4f64,
- LoLo, HiHi, IndicesVec,
- DAG.getConstant(0, DL, MVT::i8)));
- } else if (Subtarget.hasAVX2()) {
- Opcode = X86ISD::VPERMV;
- ShuffleVT = MVT::v8f32;
+ if (Subtarget.hasXOP())
+ return DAG.getBitcast(VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v4f64,
+ LoLo, HiHi, IndicesVec,
+ DAG.getConstant(0, DL, MVT::i8)));
+ // Permute Lo and Hi and then select based on index range.
+ // This works as VPERMILPD only uses index bit[1] to permute elements.
+ SDValue Res = DAG.getSelectCC(
+ DL, IndicesVec, DAG.getConstant(2, DL, MVT::v4i64),
+ DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, HiHi, IndicesVec),
+ DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, LoLo, IndicesVec),
+ ISD::CondCode::SETGT);
+ return DAG.getBitcast(VT, Res);
}
break;
case MVT::v64i8:
Modified: llvm/trunk/test/CodeGen/X86/var-permute-256.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/var-permute-256.ll?rev=327239&r1=327238&r2=327239&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/var-permute-256.ll (original)
+++ llvm/trunk/test/CodeGen/X86/var-permute-256.ll Sun Mar 11 04:52:26 2018
@@ -23,45 +23,32 @@ define <4 x i64> @var_shuffle_v4i64(<4 x
;
; AVX1-LABEL: var_shuffle_v4i64:
; AVX1: # %bb.0:
-; AVX1-NEXT: pushq %rbp
-; AVX1-NEXT: movq %rsp, %rbp
-; AVX1-NEXT: andq $-32, %rsp
-; AVX1-NEXT: subq $64, %rsp
-; AVX1-NEXT: vmovq %xmm1, %rax
-; AVX1-NEXT: andl $3, %eax
-; AVX1-NEXT: vpextrq $1, %xmm1, %rcx
-; AVX1-NEXT: andl $3, %ecx
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovq %xmm1, %rdx
-; AVX1-NEXT: andl $3, %edx
-; AVX1-NEXT: vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT: andl $3, %esi
-; AVX1-NEXT: vmovaps %ymm0, (%rsp)
-; AVX1-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
-; AVX1-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: movq %rbp, %rsp
-; AVX1-NEXT: popq %rbp
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [2,2,2,2]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpaddq %xmm4, %xmm4, %xmm4
+; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpaddq %xmm1, %xmm1, %xmm1
+; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4i64:
; AVX2: # %bb.0:
+; AVX2-NEXT: vpaddq %ymm1, %ymm1, %ymm1
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [2,2,2,2]
-; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [8589934594,8589934594,8589934594,8589934594]
-; AVX2-NEXT: vpmuludq %ymm4, %ymm3, %ymm3
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm3 = [4294967296,4294967296,4294967296,4294967296]
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
+; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX2-NEXT: retq
;
; AVX512F-LABEL: var_shuffle_v4i64:
@@ -114,39 +101,17 @@ define <8 x i32> @var_shuffle_v8i32(<8 x
;
; AVX1-LABEL: var_shuffle_v8i32:
; AVX1: # %bb.0:
-; AVX1-NEXT: pushq %rbp
-; AVX1-NEXT: movq %rsp, %rbp
-; AVX1-NEXT: andq $-32, %rsp
-; AVX1-NEXT: subq $64, %rsp
-; AVX1-NEXT: vmovd %xmm1, %r8d
-; AVX1-NEXT: vpextrd $1, %xmm1, %r9d
-; AVX1-NEXT: vpextrd $2, %xmm1, %r10d
-; AVX1-NEXT: vpextrd $3, %xmm1, %esi
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovd %xmm1, %edi
-; AVX1-NEXT: vpextrd $1, %xmm1, %eax
-; AVX1-NEXT: vpextrd $2, %xmm1, %ecx
-; AVX1-NEXT: vpextrd $3, %xmm1, %edx
-; AVX1-NEXT: vmovaps %ymm0, (%rsp)
-; AVX1-NEXT: andl $7, %r8d
-; AVX1-NEXT: andl $7, %r9d
-; AVX1-NEXT: andl $7, %r10d
-; AVX1-NEXT: andl $7, %esi
-; AVX1-NEXT: andl $7, %edi
-; AVX1-NEXT: andl $7, %eax
-; AVX1-NEXT: andl $7, %ecx
-; AVX1-NEXT: andl $7, %edx
-; AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT: vpinsrd $1, (%rsp,%rax,4), %xmm0, %xmm0
-; AVX1-NEXT: vpinsrd $2, (%rsp,%rcx,4), %xmm0, %xmm0
-; AVX1-NEXT: vpinsrd $3, (%rsp,%rdx,4), %xmm0, %xmm0
-; AVX1-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT: vpinsrd $1, (%rsp,%r9,4), %xmm1, %xmm1
-; AVX1-NEXT: vpinsrd $2, (%rsp,%r10,4), %xmm1, %xmm1
-; AVX1-NEXT: vpinsrd $3, (%rsp,%rsi,4), %xmm1, %xmm1
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: movq %rbp, %rsp
-; AVX1-NEXT: popq %rbp
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [3,3,3,3,3,3,3,3]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpcmpgtd %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpcmpgtd %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vpermilps %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvps %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; INT256-LABEL: var_shuffle_v8i32:
@@ -1326,43 +1291,32 @@ define <4 x double> @var_shuffle_v4f64(<
;
; AVX1-LABEL: var_shuffle_v4f64:
; AVX1: # %bb.0:
-; AVX1-NEXT: pushq %rbp
-; AVX1-NEXT: movq %rsp, %rbp
-; AVX1-NEXT: andq $-32, %rsp
-; AVX1-NEXT: subq $64, %rsp
-; AVX1-NEXT: vmovq %xmm1, %rax
-; AVX1-NEXT: andl $3, %eax
-; AVX1-NEXT: vpextrq $1, %xmm1, %rcx
-; AVX1-NEXT: andl $3, %ecx
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovq %xmm1, %rdx
-; AVX1-NEXT: andl $3, %edx
-; AVX1-NEXT: vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT: andl $3, %esi
-; AVX1-NEXT: vmovaps %ymm0, (%rsp)
-; AVX1-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: movq %rbp, %rsp
-; AVX1-NEXT: popq %rbp
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [2,2,2,2]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpaddq %xmm4, %xmm4, %xmm4
+; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpaddq %xmm1, %xmm1, %xmm1
+; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4f64:
; AVX2: # %bb.0:
+; AVX2-NEXT: vpaddq %ymm1, %ymm1, %ymm1
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [2,2,2,2]
-; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [8589934594,8589934594,8589934594,8589934594]
-; AVX2-NEXT: vpmuludq %ymm4, %ymm3, %ymm3
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm3 = [4294967296,4294967296,4294967296,4294967296]
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
+; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX2-NEXT: retq
;
; AVX512F-LABEL: var_shuffle_v4f64:
@@ -1415,39 +1369,17 @@ define <8 x float> @var_shuffle_v8f32(<8
;
; AVX1-LABEL: var_shuffle_v8f32:
; AVX1: # %bb.0:
-; AVX1-NEXT: pushq %rbp
-; AVX1-NEXT: movq %rsp, %rbp
-; AVX1-NEXT: andq $-32, %rsp
-; AVX1-NEXT: subq $64, %rsp
-; AVX1-NEXT: vmovd %xmm1, %esi
-; AVX1-NEXT: vpextrd $1, %xmm1, %r8d
-; AVX1-NEXT: vpextrd $2, %xmm1, %r9d
-; AVX1-NEXT: vpextrd $3, %xmm1, %r10d
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovd %xmm1, %edx
-; AVX1-NEXT: vpextrd $1, %xmm1, %edi
-; AVX1-NEXT: vpextrd $2, %xmm1, %eax
-; AVX1-NEXT: vpextrd $3, %xmm1, %ecx
-; AVX1-NEXT: vmovaps %ymm0, (%rsp)
-; AVX1-NEXT: andl $7, %esi
-; AVX1-NEXT: andl $7, %r8d
-; AVX1-NEXT: andl $7, %r9d
-; AVX1-NEXT: andl $7, %r10d
-; AVX1-NEXT: andl $7, %edx
-; AVX1-NEXT: andl $7, %edi
-; AVX1-NEXT: andl $7, %eax
-; AVX1-NEXT: andl $7, %ecx
-; AVX1-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
-; AVX1-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT: movq %rbp, %rsp
-; AVX1-NEXT: popq %rbp
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [3,3,3,3,3,3,3,3]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpcmpgtd %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpcmpgtd %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vpermilps %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvps %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; INT256-LABEL: var_shuffle_v8f32:
@@ -1500,40 +1432,34 @@ define <4 x i64> @var_shuffle_v4i64_from
;
; AVX1-LABEL: var_shuffle_v4i64_from_v2i64:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovq %xmm1, %rax
-; AVX1-NEXT: andl $1, %eax
-; AVX1-NEXT: vpextrq $1, %xmm1, %rcx
-; AVX1-NEXT: andl $1, %ecx
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovq %xmm1, %rdx
-; AVX1-NEXT: andl $1, %edx
-; AVX1-NEXT: vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT: andl $1, %esi
-; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
-; AVX1-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
-; AVX1-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [2,2,2,2]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpaddq %xmm4, %xmm4, %xmm4
+; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpaddq %xmm1, %xmm1, %xmm1
+; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4i64_from_v2i64:
; AVX2: # %bb.0:
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vpaddq %ymm1, %ymm1, %ymm1
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [2,2,2,2]
-; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [8589934594,8589934594,8589934594,8589934594]
-; AVX2-NEXT: vpmuludq %ymm4, %ymm3, %ymm3
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm3 = [4294967296,4294967296,4294967296,4294967296]
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
+; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX2-NEXT: retq
;
; AVX512F-LABEL: var_shuffle_v4i64_from_v2i64:
@@ -1589,33 +1515,18 @@ define <8 x i32> @var_shuffle_v8i32_from
;
; AVX1-LABEL: var_shuffle_v8i32_from_v4i32:
; AVX1: # %bb.0: # %entry
-; AVX1-NEXT: vmovd %xmm1, %r8d
-; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
-; AVX1-NEXT: andl $3, %r8d
-; AVX1-NEXT: vpextrd $1, %xmm1, %r9d
-; AVX1-NEXT: andl $3, %r9d
-; AVX1-NEXT: vpextrd $2, %xmm1, %r10d
-; AVX1-NEXT: andl $3, %r10d
-; AVX1-NEXT: vpextrd $3, %xmm1, %esi
-; AVX1-NEXT: andl $3, %esi
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %edi
-; AVX1-NEXT: andl $3, %edi
-; AVX1-NEXT: vpextrd $1, %xmm0, %eax
-; AVX1-NEXT: andl $3, %eax
-; AVX1-NEXT: vpextrd $2, %xmm0, %ecx
-; AVX1-NEXT: andl $3, %ecx
-; AVX1-NEXT: vpextrd $3, %xmm0, %edx
-; AVX1-NEXT: andl $3, %edx
-; AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT: vpinsrd $1, -24(%rsp,%rax,4), %xmm0, %xmm0
-; AVX1-NEXT: vpinsrd $2, -24(%rsp,%rcx,4), %xmm0, %xmm0
-; AVX1-NEXT: vpinsrd $3, -24(%rsp,%rdx,4), %xmm0, %xmm0
-; AVX1-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT: vpinsrd $1, -24(%rsp,%r9,4), %xmm1, %xmm1
-; AVX1-NEXT: vpinsrd $2, -24(%rsp,%r10,4), %xmm1, %xmm1
-; AVX1-NEXT: vpinsrd $3, -24(%rsp,%rsi,4), %xmm1, %xmm1
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [3,3,3,3,3,3,3,3]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpcmpgtd %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpcmpgtd %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vpermilps %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvps %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; INT256-LABEL: var_shuffle_v8i32_from_v4i32:
@@ -2738,38 +2649,34 @@ define <4 x double> @var_shuffle_v4f64_f
;
; AVX1-LABEL: var_shuffle_v4f64_from_v2f64:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovq %xmm1, %rax
-; AVX1-NEXT: andl $1, %eax
-; AVX1-NEXT: vpextrq $1, %xmm1, %rcx
-; AVX1-NEXT: andl $1, %ecx
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vmovq %xmm1, %rdx
-; AVX1-NEXT: andl $1, %edx
-; AVX1-NEXT: vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT: andl $1, %esi
-; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
-; AVX1-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
-; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX1-NEXT: vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [2,2,2,2]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpaddq %xmm4, %xmm4, %xmm4
+; AVX1-NEXT: vpcmpgtq %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpaddq %xmm1, %xmm1, %xmm1
+; AVX1-NEXT: vpcmpgtq %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: var_shuffle_v4f64_from_v2f64:
; AVX2: # %bb.0:
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vpaddq %ymm1, %ymm1, %ymm1
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [2,2,2,2]
-; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [8589934594,8589934594,8589934594,8589934594]
-; AVX2-NEXT: vpmuludq %ymm4, %ymm3, %ymm3
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm3 = [4294967296,4294967296,4294967296,4294967296]
-; AVX2-NEXT: vpaddq %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpcmpgtq %ymm2, %ymm1, %ymm2
+; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm3, %ymm3
+; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX2-NEXT: vpermilpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vblendvpd %ymm2, %ymm3, %ymm0, %ymm0
; AVX2-NEXT: retq
;
; AVX512F-LABEL: var_shuffle_v4f64_from_v2f64:
@@ -2825,33 +2732,18 @@ define <8 x float> @var_shuffle_v8f32_fr
;
; AVX1-LABEL: var_shuffle_v8f32_from_v4f32:
; AVX1: # %bb.0: # %entry
-; AVX1-NEXT: vmovd %xmm1, %r8d
-; AVX1-NEXT: vmovaps %xmm0, -{{[0-9]+}}(%rsp)
-; AVX1-NEXT: andl $3, %r8d
-; AVX1-NEXT: vpextrd $1, %xmm1, %r9d
-; AVX1-NEXT: andl $3, %r9d
-; AVX1-NEXT: vpextrd $2, %xmm1, %r10d
-; AVX1-NEXT: andl $3, %r10d
-; AVX1-NEXT: vpextrd $3, %xmm1, %esi
-; AVX1-NEXT: andl $3, %esi
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %edi
-; AVX1-NEXT: andl $3, %edi
-; AVX1-NEXT: vpextrd $1, %xmm0, %eax
-; AVX1-NEXT: andl $3, %eax
-; AVX1-NEXT: vpextrd $2, %xmm0, %ecx
-; AVX1-NEXT: andl $3, %ecx
-; AVX1-NEXT: vpextrd $3, %xmm0, %edx
-; AVX1-NEXT: andl $3, %edx
-; AVX1-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
-; AVX1-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
-; AVX1-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
-; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vmovdqa {{.*#+}} ymm2 = [3,3,3,3,3,3,3,3]
+; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3
+; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT: vpcmpgtd %xmm3, %xmm4, %xmm3
+; AVX1-NEXT: vpcmpgtd %xmm2, %xmm1, %xmm2
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX1-NEXT: vpermilps %ymm1, %ymm3, %ymm3
+; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT: vpermilps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vblendvps %ymm2, %ymm3, %ymm0, %ymm0
; AVX1-NEXT: retq
;
; INT256-LABEL: var_shuffle_v8f32_from_v4f32:
More information about the llvm-commits
mailing list