[X86] Use zero vector to extend to 512-bits for strict_fp_to_uint v2i1->v2f64 on targets with AVX512F, but not AVX512VL.

In the worst case, this requires a 128-bit move instruction to
implicitly zero the upper bits. In the common case, we should
recognize the producing instruction already zeroed the upper bits.
This commit is contained in:
Craig Topper 2019-12-25 10:40:40 -08:00
parent 4af5b23db3
commit c5b4a2386b
2 changed files with 8 additions and 5 deletions

View File

@ -19776,9 +19776,13 @@ SDValue X86TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
Opc = IsSigned ? ISD::STRICT_FP_TO_SINT : ISD::STRICT_FP_TO_UINT;
else
Opc = IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
Src = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8f64,
DAG.getUNDEF(MVT::v8f64),
Src, DAG.getIntPtrConstant(0, dl));
// Need to concat with zero vector for strict fp to avoid spurious
// exceptions.
// TODO: Should we just do this for non-strict as well?
SDValue Tmp = IsStrict ? DAG.getConstantFP(0.0, dl, MVT::v8f64)
: DAG.getUNDEF(MVT::v8f64);
Src = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8f64, Tmp, Src,
DAG.getIntPtrConstant(0, dl));
}
SDValue Res, Chain;
if (IsStrict) {

View File

@ -1754,10 +1754,9 @@ define <2 x i1> @strict_vector_fptoui_v2f64_to_v2i1(<2 x double> %a) #0 {
; AVX512VL-NEXT: vmovdqa64 %xmm0, %xmm0 {%k1} {z}
; AVX512VL-NEXT: ret{{[l|q]}}
;
; FIXME: This is an unsafe behavior for strict FP
; AVX512DQ-LABEL: strict_vector_fptoui_v2f64_to_v2i1:
; AVX512DQ: # %bb.0:
; AVX512DQ-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
; AVX512DQ-NEXT: vmovaps %xmm0, %xmm0
; AVX512DQ-NEXT: vcvttpd2udq %zmm0, %ymm0
; AVX512DQ-NEXT: vpslld $31, %ymm0, %ymm0
; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0