[Mlir-commits] [mlir] [mlir][amdgpu] Shared memory access optimization pass (PR #75627)

[Mehdi Amini]

================
@@ -43,6 +43,21 @@ void AMDGPUDialect::initialize() {
       >();
 }
 
+bool amdgpu::AMDGPUDialect::isSharedMemoryAddressSpace(Attribute memorySpace) {
+  if (!memorySpace)
+    return false;
+  if (auto intAttr = llvm::dyn_cast<IntegerAttr>(memorySpace))
+    return intAttr.getInt() == AMDGPUDialect::kSharedMemoryAddressSpace;
----------------
joker-eph wrote:

Other than at the LLVM level, I wouldn't think we'll have integer based memory spaces.

https://github.com/llvm/llvm-project/pull/75627