[llvm] [InstCombine] Enable more fabs fold when the user ignores sign bit of zero/NaN (PR #139861)

[Matt Arsenault via llvm-commits]

================
@@ -2773,6 +2773,47 @@ Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op,
   return nullptr;
 }
 
+/// Return true if the sign bit of result can be ignored when the result is
+/// zero.
+static bool ignoreSignBitOfZero(Instruction &I) {
+  if (I.hasNoSignedZeros())
+    return true;
+
+  // Check if the sign bit is ignored by the only user.
+  if (!I.hasOneUse())
+    return false;
+  Instruction *User = I.user_back();
+
+  // fcmp treats both positive and negative zero as equal.
+  if (User->getOpcode() == Instruction::FCmp)
+    return true;
+
+  if (auto *FPOp = dyn_cast<FPMathOperator>(User))
+    return FPOp->hasNoSignedZeros();
+
+  return false;
+}
+
+/// Return true if the sign bit of result can be ignored when the result is NaN.
+static bool ignoreSignBitOfNaN(Instruction &I) {
+  if (I.hasNoNaNs())
+    return true;
+
+  // Check if the sign bit is ignored by the only user.
+  if (!I.hasOneUse())
+    return false;
+  Instruction *User = I.user_back();
+
+  // fcmp ignores the sign bit of NaN.
----------------
arsenm wrote:

All proper floating point instructions kind of ignore the sign bit of a nan, this is just one particular instance. Eventually we should have a utility function to identify all potentially canonicalizing instructions which we can ignore the nan sign bit from 

https://github.com/llvm/llvm-project/pull/139861