[clang] [CIR] Add custom assembly format for alloca op to fix flag parsing (PR #198962)

[via cfe-commits]

================
@@ -653,13 +631,13 @@ def CIR_AllocaOp : CIR_Op<"alloca", [
   }];
 
   let assemblyFormat = [{
-    $allocaType `,` qualified(type($addr)) `,`
+    $name
+    `align` `(` $alignment `)`
+    oilist( `init`              $init
+          | `const`             $constant
+          | `cleanup_dest_slot` $cleanup_dest_slot)
     ($dynAllocSize^ `:` type($dynAllocSize) `,`)?
-    `[` $name
-       (`,` `init` $init^)?
-       (`,` `const` $constant^)?
-       (`,` `cleanup_dest_slot` $cleanup_dest_slot^)?
-    `]`
+    $allocaType `->` qualified(type($addr))
----------------
Andres-Salamanca wrote:

I explored this a bit more and found that `InferTypeOpInterface` could be used here. One advantage is that callers would no longer need to explicitly construct the pointer type. For example, instead of:

```cpp
mlir::Type itrTy = cgf.cgm.convertType(cgf.getContext().UnsignedLongLongTy);
auto itrPtrTy = cir::PointerType::get(itrTy);
cir::AllocaOp::create(builder, loc, itrPtrTy, itrTy, "itr", itrAlign);
```

we could just pass the allocated type and let the result type be inferred.

That said, it doesn't actually address the storage redundancy you mentioned. Even with `inferReturnTypes` building `cir::PointerType::get(ctx, adaptor.getAllocaType())`, both the allocated type and the pointer result type still end up stored in the operation.

I also noticed that `llvm.alloca` follows the same approach and stores both `elem_type` and the pointer result type separately.

Following your suggestion, the format would look something like:

`%0 = cir.alloca "n" align(4) init -> !cir.ptr<!s32i>`

and for dynamic allocations:

`%0 = cir.alloca "n" align(4) init %sz : !u64i -> !cir.ptr<!s32i>`

Does that look good to you?😊

https://github.com/llvm/llvm-project/pull/198962