[mlir] fix the types used during the generation of the kernel param array

[gysit]

Summary

The patch fixes the types used to access the elements of the kernel parameter structure from a pointer to the structure to a pointer to the actual parameter type.

[csigg]

Thanks!

mlir/lib/Conversion/GPUCommon/ConvertLaunchFuncToRuntimeCalls.cpp

auto arraySize = builder.create<LLVM::ConstantOp>(
       loc, llvmInt32Type, builder.getI32IntegerAttr(numArguments));
   auto arrayPtr = builder.create<LLVM::AllocaOp>(loc, llvmPointerPointerType,
                                                  arraySize, /*alignment=*/0);
   auto zero = builder.create<LLVM::ConstantOp>(loc, llvmInt32Type,
                                                builder.getI32IntegerAttr(0));
   for (auto en : llvm::enumerate(arguments)) {
     auto index = builder.create<LLVM::ConstantOp>(
         loc, llvmInt32Type, builder.getI32IntegerAttr(en.index()));
-    auto fieldPtr =
-        builder.create<LLVM::GEPOp>(loc, structType.getPointerTo(), structPtr,
-                                    ArrayRef<Value>{zero, index.getResult()});
+    auto fieldPtr = builder.create<LLVM::GEPOp>(
+        loc, argumentTypes[en.index()].getPointerTo(), structPtr,
     builder.create<LLVM::StoreOp>(loc, en.value(), fieldPtr);
     auto elementPtr = builder.create<LLVM::GEPOp>(loc, llvmPointerPointerType,
                                                   arrayPtr, index.getResult());
     auto casted =
         builder.create<LLVM::BitcastOp>(loc, llvmPointerType, fieldPtr);
     builder.create<LLVM::StoreOp>(loc, casted, elementPtr);
   }
   return arrayPtr;