[clang] [Hlsl][Spirv] Don't emit global variables for structs containing resources (PR #234797)

via cfe-commits cfe-commits at lists.llvm.org
Thu Nov 5 23:35:54 PST 2026


rj123-nv wrote:

> Thanks for fixing this.
> 
> Could you make sure the offload-test-suite `ResourcesInStructs` tests still behave as expected with this change?
> 
> OK with the above changes.

with current PR i get:

array-of-structs-with-res.test         : pass
res-in-struct-local-arg.test             : fail (compiler error)
res-in-struct-simple-array.test       : pass
res-in-struct-this-access.test         : fail (compiler error)
res-array-of-matrix-in-struct.test   : fail (runtime diff)
res-in-struct-mix.test                      : pass
res-in-struct-simple-one.test         : pass
res-of-matrix-in-struct.test            : pass

all the ones passing here fail on main in spirv validation.

res-in-struct-this-access seems out of scope here.
res-in-struct-local-arg is something I realize I should address here. Basically the issue is that when you have 

```
  struct S {
    RWBuffer<float> a;
    RWBuffer<float> b;
  };

  S s;

  [numthreads(1, 1, 1)]
  void main() {
    S tmp = s;
    tmp.a[0] = tmp.b[0] + 1.0f;
  }
```

We effectively split 's' into the actual globals 's.a' and 's.b' (assume '.' can be used in names)
according to https://github.com/llvm/wg-hlsl/blob/main/proposals/0038-resources-in-structs.md

When the whole struct 's' is used by value, like assignment or as a function arg, we need to reconstruct a struct type from the 2 globals s.a and s.b.

Effectively doing

```
struct S {
  RWBuffer<float> a;
  RWBuffer<float> b;
};

RWBuffer<float> s_a;
RWBuffer<float> s_b;

...

S tmp = { s_a, s_b };  // SROA should remove this eventually
```

There is logic that exists for mixed structs to do this (globals structs with both numeric members and resources, which are given hlsl_constant addr space) that I think can be adapted here.
https://github.com/llvm/llvm-project/blob/0dfd901f3757/clang/lib/CodeGen/CGHLSLRuntime.cpp#L538



https://github.com/llvm/llvm-project/pull/234797


More information about the cfe-commits mailing list