[llvm] [InstCombine] Eliminate fptrunc/fpext if fast math flags allow it (PR #115027)

[Yingwei Zheng via llvm-commits]

================
@@ -1940,6 +1940,31 @@ Instruction *InstCombinerImpl::visitFPExt(CastInst &FPExt) {
       return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
   }
 
+  // fpext (fptrunc(x)) -> x, if the fast math flags allow it
+  if (auto *Trunc = dyn_cast<FPTruncInst>(Src)) {
+    // Whether this transformation is possible depends on the fast math flags of
+    // both the fpext and fptrunc.
+    FastMathFlags SrcFlags = Trunc->getFastMathFlags();
+    FastMathFlags DstFlags = FPExt.getFastMathFlags();
+    // Trunc can introduce inf and change the encoding of a nan, so the
+    // destination must have the nnan and ninf flags to indicate that we don't
+    // need to care about that. We are also removing a rounding step, and that
+    // requires both the source and destination to allow contraction.
+    if (DstFlags.noNaNs() && DstFlags.noInfs() && SrcFlags.allowContract() &&
----------------
dtcxzyw wrote:

I don't think `contract` is suitable here. Previously, `contract` means the optimizer is allowed to fold fmul+fadd into a fma. But the transformation here is too aggressive since clang may be the first compiler to fold this pattern: https://godbolt.org/z/bYfz334Pr


https://github.com/llvm/llvm-project/pull/115027