[llvm] c5b4a23 - [X86] Use zero vector to extend to 512-bits for strict_fp_to_uint v2i1->v2f64 on targets with AVX512F, but not AVX512VL.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Wed Dec 25 10:46:45 PST 2019
Author: Craig Topper
Date: 2019-12-25T10:46:00-08:00
New Revision: c5b4a2386b51a18daad7e42040c685c2e9708c47
URL: https://github.com/llvm/llvm-project/commit/c5b4a2386b51a18daad7e42040c685c2e9708c47
DIFF: https://github.com/llvm/llvm-project/commit/c5b4a2386b51a18daad7e42040c685c2e9708c47.diff
LOG: [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.
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/vec-strict-fptoint-128.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 27983a11e514..c383d9252b39 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -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) {
diff --git a/llvm/test/CodeGen/X86/vec-strict-fptoint-128.ll b/llvm/test/CodeGen/X86/vec-strict-fptoint-128.ll
index 662afe22cfc9..832606947576 100644
--- a/llvm/test/CodeGen/X86/vec-strict-fptoint-128.ll
+++ b/llvm/test/CodeGen/X86/vec-strict-fptoint-128.ll
@@ -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
More information about the llvm-commits
mailing list