[llvm] [InstCombine] Fold +0.0 - x to -x if we known x is non positive zero (PR #215871)

[Yingwei Zheng via llvm-commits]

================
@@ -3290,6 +3290,18 @@ Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
   if (match(&I, m_FNeg(m_Value(Op))))
     return UnaryOperator::CreateFNegFMF(Op, &I);
 
+  // fsub +0.0, X ==> fneg X if known X is never +0.0
+  if (match(I.getOperand(0), m_PosZeroFP())) {
+    Value *X = I.getOperand(1);
+    DenormalMode Mode = I.getFunction()->getDenormalMode(
+        X->getType()->getScalarType()->getFltSemantics());
+    KnownFPClass Known = computeKnownFPClass(
+        X, fcPosZero | fcSubnormal, getSimplifyQuery().getWithInstruction(&I));
+    if (Known.isKnownNeverLogicalPosZero(Mode) &&
+        (Mode.Input == DenormalMode::IEEE || Known.isKnownNeverSubnormal()))
----------------
dtcxzyw wrote:

I think most of the `fsub->fneg` transformations in visitFSub don't respect FTZ/DTZ...

Perhaps we can replace all the `UnaryOperator::CreateFNegFMF` below with `fsub -0.0, x`, then only handle denormal modes in the canonicalization of `fsub -0.0, X ==> fneg X`.


https://github.com/llvm/llvm-project/pull/215871