[llvm] [NVPTX] Constant fold NVVM fmin and fmax (PR #121966)

[Artem Belevich via llvm-commits]

================
@@ -1689,6 +1689,28 @@ bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
   case Intrinsic::x86_avx512_cvttsd2usi64:
     return !Call->isStrictFP();
 
+  // NVVM FMax intrinsics
+  case Intrinsic::nvvm_fmax_d:
+  case Intrinsic::nvvm_fmax_f:
+  case Intrinsic::nvvm_fmax_ftz_f:
+  case Intrinsic::nvvm_fmax_ftz_nan_f:
+  case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
+  case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
+  case Intrinsic::nvvm_fmax_nan_f:
+  case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
+  case Intrinsic::nvvm_fmax_xorsign_abs_f:
+
+  // NVVM FMin intrinsics
+  case Intrinsic::nvvm_fmin_d:
+  case Intrinsic::nvvm_fmin_f:
+  case Intrinsic::nvvm_fmin_ftz_f:
+  case Intrinsic::nvvm_fmin_ftz_nan_f:
+  case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
+  case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
+  case Intrinsic::nvvm_fmin_nan_f:
+  case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
+  case Intrinsic::nvvm_fmin_xorsign_abs_f:
----------------
Artem-B wrote:

Should we follow the folding behavior of the builtin min/max intrinsics based on `isStrictFP()` ? Granted, "FP environment" on the GPU is effectively fixed, but it may be prudent to make the folding decisions consistently across host and GPU compilations.


https://github.com/llvm/llvm-project/pull/121966