[llvm] [AMDGPU][LowerModuleLDS] Handle non-dynlds zero sized type (PR #84742)

[Pierre van Houtryve via llvm-commits]

================
@@ -1398,8 +1398,8 @@ class AMDGPULowerModuleLDS {
           LDSVarsToTransform.begin(), LDSVarsToTransform.end()));
 
       for (GlobalVariable *GV : Sorted) {
-        OptimizedStructLayoutField F(GV,
-                                     DL.getTypeAllocSize(GV->getValueType()),
+        unsigned Size = DL.getTypeAllocSize(GV->getValueType());
+        OptimizedStructLayoutField F(GV, Size ? Size : 1,
----------------
Pierre-vh wrote:

I think it'd be for a variable length array. It's why we lower as dynlds for external ones. I assume we just can't lower them then and we should probably just fatal_error. It's a bit ugly to fatal in the backend but better than asserting

https://github.com/llvm/llvm-project/pull/84742