[llvm] goldsteinn/x86 nneg uitofp (PR #86694)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: None (goldsteinn)

<details>
<summary>Changes</summary>

- **[X86] Add tests for `uitofp nneg` -> `sitofp`; NFC**
- **[X86] Use `nneg` flag when trying to convert `uitofp` -> `sitofp`**


---
Full diff: https://github.com/llvm/llvm-project/pull/86694.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+3-2) 
- (modified) llvm/test/CodeGen/X86/uint_to_fp.ll (+54) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 1ce742a7c19a74..c5b391df20c837 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -48270,7 +48270,7 @@ static SDValue combineAndShuffleNot(SDNode *N, SelectionDAG &DAG,
 
   // We do not split for SSE at all, but we need to split vectors for AVX1 and
   // AVX2.
-  if (!Subtarget.useAVX512Regs() && VT.is512BitVector() && 
+  if (!Subtarget.useAVX512Regs() && VT.is512BitVector() &&
       TLI.isTypeLegal(VT.getHalfNumVectorElementsVT(*DAG.getContext()))) {
     SDValue LoX, HiX;
     std::tie(LoX, HiX) = splitVector(X, DAG, DL);
@@ -54065,7 +54065,8 @@ static SDValue combineUIntToFP(SDNode *N, SelectionDAG &DAG,
   // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
   // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
   // the optimization here.
-  if (DAG.SignBitIsZero(Op0)) {
+  SDNodeFlags Flags = N->getFlags();
+  if (Flags.hasNonNeg() || DAG.SignBitIsZero(Op0)) {
     if (IsStrict)
       return DAG.getNode(ISD::STRICT_SINT_TO_FP, SDLoc(N), {VT, MVT::Other},
                          {N->getOperand(0), Op0});
diff --git a/llvm/test/CodeGen/X86/uint_to_fp.ll b/llvm/test/CodeGen/X86/uint_to_fp.ll
index d8e0b61ed199fa..8c8cbb151974d6 100644
--- a/llvm/test/CodeGen/X86/uint_to_fp.ll
+++ b/llvm/test/CodeGen/X86/uint_to_fp.ll
@@ -25,3 +25,57 @@ entry:
   store float %1, ptr %y
   ret void
 }
+
+define float @test_without_nneg(i32 %x) nounwind {
+; X86-LABEL: test_without_nneg:
+; X86:       ## %bb.0:
+; X86-NEXT:    pushl %eax
+; X86-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-NEXT:    orpd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-NEXT:    subsd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-NEXT:    cvtsd2ss %xmm0, %xmm0
+; X86-NEXT:    movss %xmm0, (%esp)
+; X86-NEXT:    flds (%esp)
+; X86-NEXT:    popl %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_without_nneg:
+; X64:       ## %bb.0:
+; X64-NEXT:    movl %edi, %eax
+; X64-NEXT:    cvtsi2ss %rax, %xmm0
+; X64-NEXT:    retq
+  %r = uitofp i32 %x to float
+  ret float %r
+}
+
+define float @test_with_nneg(i32 %x) nounwind {
+; X86-LABEL: test_with_nneg:
+; X86:       ## %bb.0:
+; X86-NEXT:    pushl %eax
+; X86-NEXT:    cvtsi2ssl {{[0-9]+}}(%esp), %xmm0
+; X86-NEXT:    movss %xmm0, (%esp)
+; X86-NEXT:    flds (%esp)
+; X86-NEXT:    popl %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_with_nneg:
+; X64:       ## %bb.0:
+; X64-NEXT:    cvtsi2ss %edi, %xmm0
+; X64-NEXT:    retq
+  %r = uitofp nneg i32 %x to float
+  ret float %r
+}
+
+define <4 x float> @test_with_nneg_vec(<4 x i32> %x) nounwind {
+; X86-LABEL: test_with_nneg_vec:
+; X86:       ## %bb.0:
+; X86-NEXT:    cvtdq2ps %xmm0, %xmm0
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_with_nneg_vec:
+; X64:       ## %bb.0:
+; X64-NEXT:    cvtdq2ps %xmm0, %xmm0
+; X64-NEXT:    retq
+  %r = uitofp nneg <4 x i32> %x to <4 x float>
+  ret <4 x float> %r
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/86694