[llvm] [AMDGPU] Skip fold_pow constant-exponent shortcuts for powr on possibly-negative base (PR #200579)

[Jay Foad via llvm-commits]

================
@@ -898,66 +898,76 @@ bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B,
   // 0x1111111 means that we don't do anything for this call.
   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
 
-  if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) {
-    //  pow/powr/pown(x, 0) == 1
-    LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n");
-    Constant *cnval = ConstantFP::get(eltType, 1.0);
-    if (getVecSize(FInfo) > 1) {
-      cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
+  // OpenCL powr(x<0, y) = NaN, but the folds below would turn it into a
+  // finite number. Skip them unless NaNs are ignored or the base is known
+  // non-negative.
+  bool IsPowr = FInfo.getId() == AMDGPULibFunc::EI_POWR ||
+                FInfo.getId() == AMDGPULibFunc::EI_POWR_FAST;
+  bool SkipConstantFolds =
+      (CF || CINT) && IsPowr && !FPOp->hasNoNaNs() &&
----------------
jayfoad wrote:

Could simplify the condition:
```suggestion
      IsPowr && !FPOp->hasNoNaNs() &&
```

https://github.com/llvm/llvm-project/pull/200579