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

[DAG] SimplifyMultipleUseDemandedBits - drop unnecessary *_EXTEND_VECTOR_INREG cases 

For little endian targets, if we only need the lowest element and none of the extended bits then we can just use the (bitcasted) source vector directly.

We already do this in SimplifyDemandedBits, this adds the SimplifyMultipleUseDemandedBits equivalent.
This commit is contained in:
Simon Pilgrim 2020-06-22 12:11:11 +01:00
parent b5b46601c0
commit ecc5d7ee0d
2 changed files with 15 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
View File

@ -773,6 +773,21 @@ SDValue TargetLowering::SimplifyMultipleUseDemandedBits(
return Op0;
break;
}
case ISD::ANY_EXTEND_VECTOR_INREG:
case ISD::SIGN_EXTEND_VECTOR_INREG:
case ISD::ZERO_EXTEND_VECTOR_INREG: {
// If we only want the lowest element and none of extended bits, then we can
// return the bitcasted source vector.
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
EVT DstVT = Op.getValueType();
if (DemandedElts == 1 && DstVT.getSizeInBits() == SrcVT.getSizeInBits() &&
DAG.getDataLayout().isLittleEndian() &&
DemandedBits.getActiveBits() <= SrcVT.getScalarSizeInBits()) {
return DAG.getBitcast(DstVT, Src);
}
break;
}
case ISD::INSERT_VECTOR_ELT: {
// If we don't demand the inserted element, return the base vector.
SDValue Vec = Op.getOperand(0);

3
llvm/test/CodeGen/X86/known-signbits-vector.ll
View File

@ -389,7 +389,6 @@ define float @signbits_ashr_sextvecinreg_bitops_extract_sitofp(<2 x i64> %a0, <4
; X86-NEXT: vmovdqa {{.*#+}} xmm2 = [4,0,8,0]
; X86-NEXT: vpxor %xmm2, %xmm0, %xmm0
; X86-NEXT: vpsubq %xmm2, %xmm0, %xmm0
; X86-NEXT: vpmovsxdq %xmm1, %xmm1
; X86-NEXT: vpand %xmm1, %xmm0, %xmm2
; X86-NEXT: vpor %xmm1, %xmm2, %xmm1
; X86-NEXT: vpxor %xmm0, %xmm1, %xmm0
@ -407,7 +406,6 @@ define float @signbits_ashr_sextvecinreg_bitops_extract_sitofp(<2 x i64> %a0, <4
; X64-AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [4,8]
; X64-AVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
; X64-AVX1-NEXT: vpsubq %xmm2, %xmm0, %xmm0
; X64-AVX1-NEXT: vpmovsxdq %xmm1, %xmm1
; X64-AVX1-NEXT: vpand %xmm1, %xmm0, %xmm2
; X64-AVX1-NEXT: vpor %xmm1, %xmm2, %xmm1
; X64-AVX1-NEXT: vpxor %xmm0, %xmm1, %xmm0
@ -420,7 +418,6 @@ define float @signbits_ashr_sextvecinreg_bitops_extract_sitofp(<2 x i64> %a0, <4
; X64-AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [4,8]
; X64-AVX2-NEXT: vpxor %xmm2, %xmm0, %xmm0
; X64-AVX2-NEXT: vpsubq %xmm2, %xmm0, %xmm0
; X64-AVX2-NEXT: vpmovsxdq %xmm1, %xmm1
; X64-AVX2-NEXT: vpand %xmm1, %xmm0, %xmm2
; X64-AVX2-NEXT: vpor %xmm1, %xmm2, %xmm1
; X64-AVX2-NEXT: vpxor %xmm0, %xmm1, %xmm0