[llvm] [InstCombine] Fold fcmp ogt (x - y), 0 into fcmp ogt x, y #85245 (PR #85506)

[Matt Arsenault via llvm-commits]

================
@@ -7972,6 +7972,11 @@ Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
   Constant *RHSC;
   if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
     switch (LHSI->getOpcode()) {
+    case Instruction::FSub:
+      if (Pred == FCmpInst::FCMP_OGT && match(RHSC, m_PosZeroFP()) &&
----------------
arsenm wrote:

For the tests you have written here, yes it will always fail. You will need to have additional tests covering the no-infinity knowledge.

Also, you're not taking advantage of the fast math flags on the fsub itself by directly looking at the operands. IIRC there's another overload of isKnownNeverInfinity with passed in flags 

https://github.com/llvm/llvm-project/pull/85506