[Mlir-commits] [mlir] [mlir][gpu] GPUToROCDL/NVVM: use generic llvm conversion interface instead of hardcoded conversions. (PR #124439)

[Jakub Kuderski]

================
@@ -376,17 +376,44 @@ struct LowerGpuOpsToNVVMOpsPass
     LLVMTypeConverter converter(m.getContext(), options);
     configureGpuToNVVMTypeConverter(converter);
     RewritePatternSet llvmPatterns(m.getContext());
+    LLVMConversionTarget target(getContext());
+
+    if (!filterDialects.empty()) {
+      for (StringRef dialectName : filterDialects) {
+        Dialect *dialect = getContext().getLoadedDialect(dialectName);
+        // Dialect may not be loaded if it wasn't used in source module, ignore.
+        if (!dialect)
+          continue;
+
+        auto *iface = dyn_cast<ConvertToLLVMPatternInterface>(dialect);
+        if (!iface) {
+          m.emitError()
+              << "dialect does not implement ConvertToLLVMPatternInterface: "
+              << dialectName << "\n";
+          return signalPassFailure();
+        }
+
+        iface->populateConvertToLLVMConversionPatterns(target, converter,
+                                                       llvmPatterns);
+      }
+    } else {
+      for (Dialect *dialect : getContext().getLoadedDialects()) {
+        if (isa<math::MathDialect>(dialect)) // Need custom math lowering
----------------
kuhar wrote:

```suggestion
        if (isa<math::MathDialect>(dialect)) // Need custom math lowering.
```

https://github.com/llvm/llvm-project/pull/124439