[Mlir-commits] [mlir] [MLIR][GPU-LLVM] Add in-pass signature update option for opencl kernels (PR #105664)

[Petr Kurapov]

================
@@ -325,6 +369,20 @@ struct GPUToLLVMSPVConversionPass final
     LLVMTypeConverter converter(context, options);
     LLVMConversionTarget target(*context);
 
+    if (forceOpenclAddressSpaces) {
+      MemorySpaceToOpenCLMemorySpaceConverter converter;
+      AttrTypeReplacer replacer;
+      replacer.addReplacement([&converter](BaseMemRefType origType)
+                                  -> std::optional<BaseMemRefType> {
+        return converter.convertType<BaseMemRefType>(origType);
+      });
+
+      replacer.recursivelyReplaceElementsIn(getOperation(),
+                                            /*replaceAttrs=*/true,
+                                            /*replaceLocs=*/false,
+                                            /*replaceTypes=*/true);
+    }
+
----------------
kurapov-peter wrote:

I tried a natural way - putting additional conversion to the llvm converter but it didn't work out. First, it expects a legal value as the output, so I can't just add a memref. I tried to run `converter.convertType(<newMemRefTypeWithAttribute>)` for this. It does the job for memrefs in the function body but not for its signature. I could not access the signature type by adding a conversion (the type never reaches the hook). And even though the internal signature conversion calls type converter recursively the pattern is never called. I think it should work since I'm adding the conversion after all the default ones, so it should be called first. Yet, that doesn't happen somehow.

So instead, I added this. It doesn't have to think about legality and all since it's not an llvm converter.

https://github.com/llvm/llvm-project/pull/105664