[llvm] [NVPTX] Fix NaN + overflow semantics of f2ll/d2i (PR #159530)

[Durgadoss R via llvm-commits]

================
@@ -281,21 +281,12 @@ static Instruction *convertNvvmIntrinsicToLlvm(InstCombiner &IC,
       return {Intrinsic::trunc, FTZ_MustBeOn};
 
     // NVVM intrinsics that map to LLVM cast operations.
-    //
-    // Note that llvm's target-generic conversion operators correspond to the rz
-    // (round to zero) versions of the nvvm conversion intrinsics, even though
-    // most everything else here uses the rn (round to nearest even) nvvm ops.
-    case Intrinsic::nvvm_d2i_rz:
-    case Intrinsic::nvvm_f2i_rz:
-    case Intrinsic::nvvm_d2ll_rz:
-    case Intrinsic::nvvm_f2ll_rz:
-      return {Instruction::FPToSI};
-    case Intrinsic::nvvm_d2ui_rz:
-    case Intrinsic::nvvm_f2ui_rz:
-    case Intrinsic::nvvm_d2ull_rz:
-    case Intrinsic::nvvm_f2ull_rz:
-      return {Instruction::FPToUI};
-    // Integer to floating-point uses RN rounding, not RZ
+    // Note - we cannot map intrinsics like nvvm_d2ll_rz to LLVM's
+    // FPToSI, as NaN to int conversion with FPToSI is considered UB and is
+    // eliminated. NVVM conversion intrinsics are translated to PTX cvt
+    // instructions which define the outcome for NaN rather than leaving as UB.
+    // Therefore, translate NVVM intrinsics to sitofp/uitofp, but not to
+    // fptosi/fptoui.
----------------
durga4github wrote:

Thanks for this comment!

https://github.com/llvm/llvm-project/pull/159530