[Mlir-commits] [mlir] [MLIR][NVVM] Fixed assertion failure for insufficient parsing validation of nvvm dialect PureSpecialRangeableRegisterOp (PR #163434)

[Stefan Mada]

================
@@ -279,6 +280,23 @@ class NVVM_PureSpecialRangeableRegisterOp<string mnemonic, list<Trait> traits =
         SetIntRangeFn setResultRanges) {
         nvvmInferResultRanges(getOperation(), getResult(), argRanges, setResultRanges);
     }
+
+    // Verify the range attribute satisfies LLVM ConstantRange constructor requirements.
+    ::llvm::LogicalResult $cppClass::verify() {
----------------
smada3 wrote:

There are 34 registers that are part of this class, adding even a 1 line check for each of the specializations would add over 100 lines to NVVMDialect.cpp with code that is identical between them all. Is there a way to template the C++ code to avoid this? I agree with making a function for the verification logic that resides in C++, but I think doing this in tablegen would have the least amount of new code added. I think it would be easy to debug if we had something like:

```
::llvm::LogicalResult $cppClass::verify() {
  return verifyConstantRangeAttr(getOperation(), getRange());
}
```

And then define `verifyConstantRangeAttr` in `NVVMDialect.cpp`. How does that seem?

https://github.com/llvm/llvm-project/pull/163434