maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[X86] getShuffleScalarElt - peek through insert/extract subvector nodes. 

llvm-svn: 355288
This commit is contained in:
Simon Pilgrim 2019-03-03 14:11:05 +00:00
parent 11149ea433
commit d8e91a54c0
2 changed files with 23 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -6884,6 +6884,29 @@ static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
Depth+1);
}
// Recurse into insert_subvector base/sub vector to find scalars.
if (Opcode == ISD::INSERT_SUBVECTOR &&
isa<ConstantSDNode>(N->getOperand(2))) {
SDValue Vec = N->getOperand(0);
SDValue Sub = N->getOperand(1);
EVT SubVT = Sub.getValueType();
unsigned NumSubElts = SubVT.getVectorNumElements();
uint64_t SubIdx = N->getConstantOperandVal(2);
if (SubIdx <= Index && Index < (SubIdx + NumSubElts))
return getShuffleScalarElt(Sub.getNode(), Index - SubIdx, DAG, Depth + 1);
return getShuffleScalarElt(Vec.getNode(), Index, DAG, Depth + 1);
}
// Recurse into extract_subvector src vector to find scalars.
if (Opcode == ISD::EXTRACT_SUBVECTOR &&
isa<ConstantSDNode>(N->getOperand(1))) {
SDValue Src = N->getOperand(0);
EVT SrcVT = Src.getValueType();
uint64_t SrcIdx = N->getConstantOperandVal(1);
return getShuffleScalarElt(Src.getNode(), Index + SrcIdx, DAG, Depth + 1);
}
// Actual nodes that may contain scalar elements
if (Opcode == ISD::BITCAST) {
V = V.getOperand(0);

10
llvm/test/CodeGen/X86/avx-load-store.ll
View File

@ -85,11 +85,6 @@ define <8 x float> @mov00(<8 x float> %v, float * %ptr) nounwind {
; CHECK_O0-LABEL: mov00:
; CHECK_O0: # %bb.0:
; CHECK_O0-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; CHECK_O0-NEXT: # implicit-def: $ymm1
; CHECK_O0-NEXT: vmovaps %xmm0, %xmm1
; CHECK_O0-NEXT: vmovaps %xmm1, %xmm0
; CHECK_O0-NEXT: vxorps %xmm2, %xmm2, %xmm2
; CHECK_O0-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
; CHECK_O0-NEXT: # kill: def $ymm0 killed $xmm0
; CHECK_O0-NEXT: retq
%val = load float, float* %ptr
@ -106,11 +101,6 @@ define <4 x double> @mov01(<4 x double> %v, double * %ptr) nounwind {
; CHECK_O0-LABEL: mov01:
; CHECK_O0: # %bb.0:
; CHECK_O0-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; CHECK_O0-NEXT: # implicit-def: $ymm1
; CHECK_O0-NEXT: vmovaps %xmm0, %xmm1
; CHECK_O0-NEXT: vmovaps %xmm1, %xmm0
; CHECK_O0-NEXT: vxorps %xmm2, %xmm2, %xmm2
; CHECK_O0-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm2[1]
; CHECK_O0-NEXT: # kill: def $ymm0 killed $xmm0
; CHECK_O0-NEXT: retq
%val = load double, double* %ptr