[llvm] [NVPTX] Enable .ptr .global .align attributes for kernel attributes for CUDA (PR #79646)

[Artem Belevich via llvm-commits]

================
@@ -1610,6 +1610,10 @@ void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
             }
             Align ParamAlign = I->getParamAlign().valueOrOne();
             O << ".align " << ParamAlign.value() << " ";
+          } else if (I->getParamAlign().valueOrOne() != 1) {
+            O << ".ptr .global ";
----------------
Artem-B wrote:

I think the code would be more concise if we would rearrange the code a bit :

```
if (PTy) {
  print ".param"
  if (CUDA) {
    assert(AS == generic)
    print ".ptr.global"
  } else {
    switch(AS){
     ...
    }
  }
  print(".align") // see note regarding alignment below.
}
```

I think we may want to annotate pointers & AS regardless of whether alignment is specified. E.g. the second test case should've generated '.praram .ptr .global func_param_0'


Another interesting quirk is that the default alignement assumed by ptxas is 4. I do not think LLVM has the same idea when alignment is not specified. So, I think the non-CUDA branch does the right thing always specifying it explicitly.



https://github.com/llvm/llvm-project/pull/79646