[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.

2 files changed, 8 insertions, 5 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 27983a1..c383d92 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 662afe2..8326069 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