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

[Arseniy Obolenskiy via llvm-commits]

================
@@ -898,6 +898,16 @@ 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);
 
+  // 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.
+  if ((CF || CINT) && !FPOp->hasNoNaNs() &&
+      (FInfo.getId() == AMDGPULibFunc::EI_POWR ||
+       FInfo.getId() == AMDGPULibFunc::EI_POWR_FAST) &&
+      !cannotBeOrderedLessThanZero(
+          opr0, SQ.getWithInstruction(cast<Instruction>(FPOp))))
+    return false;
----------------
aobolensk wrote:

Actually, both are valid and produce the same result, but the diff is much more verbose. Applied.

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