[llvm] Check for side effects when lowering target intrinsics, update NVVM ldu/ldg intrinsics with IntrWillReturn and test for DCE (PR #98968)

[Kevin McAfee via llvm-commits]

================
@@ -5227,7 +5227,7 @@ void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
   // definition.
   const Function *F = I.getCalledFunction();
   bool HasChain = !F->doesNotAccessMemory();
-  bool OnlyLoad = HasChain && F->onlyReadsMemory();
+  bool OnlyLoad = HasChain && F->onlyReadsMemory() && !I.mayHaveSideEffects();
----------------
kalxr wrote:

Thanks for explaining!
> The code previously only considers the attributes of the underlying function definition, which is what all other intrinsic code expects (and changes the opcode between INTRINSIC_WO_CHAIN/INTRINSIC_W_CHAIN/INTRINSIC_VOID). This now will swap the opcode and operand structure depending on whether the callsite has a memory(none) attribute.

I understand that it will affect the operands, but I don't see where `OnlyLoad` affects the opcode. That seems like something determined by `HasChain`.

> ... you can only consider the underlying function declaration and not the instruction / call site instance

If the issue is with using the instruction, would it then be reasonable to use `F->willReturn()` instead of `!I.mayHaveSideEffects()`? It still appears to me incorrect to treat an intrinsic as only a load when it may not return, as the previous code is doing.

https://github.com/llvm/llvm-project/pull/98968