[llvm] [InstCombine] Support reassoc for foldLogicOfFCmps (PR #116065)

[Matt Arsenault via llvm-commits]

================
@@ -1465,11 +1465,16 @@ Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS,
 
     // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
     // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
-    if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP()))
+    if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP())) {
       // Ignore the constants because they are obviously not NANs:
       // (fcmp ord x, 0.0) & (fcmp ord y, 0.0)  -> (fcmp ord x, y)
       // (fcmp uno x, 0.0) | (fcmp uno y, 0.0)  -> (fcmp uno x, y)
+      IRBuilder<>::FastMathFlagGuard FMFG(Builder);
+      FastMathFlags FMF = LHS->getFastMathFlags();
+      FMF &= RHS->getFastMathFlags();
+      Builder.setFastMathFlags(FMF);
----------------
arsenm wrote:

You can do the and inline now
```suggestion
      Builder.setFastMathFlags(LHS->getFastMathFlags() & RHS->getFastMathFlags());
```

https://github.com/llvm/llvm-project/pull/116065