[Mlir-commits] [mlir] [MLIR][GPU] Support grid constant, byval, byref on gpu.func (PR #172037)

[Mehdi Amini]

================
@@ -369,7 +370,10 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
     }
 
     if (lowersToPointer) {
+      copyPointerAttribute(mlir::NVVM::NVVMDialect::getGridConstantAttrName());
----------------
joker-eph wrote:

> [...] because it's most likely what people want.

I'm skeptical of this kind of arguments: a compiler first principle is to not miscompile, and this line of thinking around "the majority of what people want" does not mesh well with it, we've refrained from designing the system that way for good reasons. This is a recurring issue with dropping discardable attributes.

Note I'm arguing on the reasoning you're presenting, not necessarily the particular conclusion here (in case you missed [my comment yesterday](https://github.com/llvm/llvm-project/pull/172037#discussion_r2635424380)).




https://github.com/llvm/llvm-project/pull/172037