comparison of nvcc vs clang nvptx backend, part 2

Part 1

First of all, I have to say that digging through LLVM source code is simply disgusting - its code tree is monstrously huge and parts from nvptx scattered across the most unexpected places. So to extract needed tables I just stuck libLLVM.so.21.1 (btw size 135Mb) into IDA Pro and reused the same IDC scripts for cicc, just removed strings decryption. Typical open-source, yeah

Mlir

cicc does not have MLIR at all. This fact probably deserves a larger font:

cicc does not have MLIR

The proof is left as an exercise for the reader very simple - just grep strings like mlir/linalg/affine etc in cicc binary. Let's assume that we have some way to load LLVM bc into cicc - even in this case, we wouldn't be able to force it to generate more efficient PTX for code containing MLIR

Btw this is not first case when serious amount of real code generation was done without mlir - for example vectorization for RISC-V too. And somewhere on the side, you’re bound to hear that argument from clueless managers: "if even Oprah Microsoft uses a monorepo, you must be happy using one too" 

Builtins

• from cicc - 870
• from clang -  715

Notably, neither Nvidia nor LLVM documented the list of supported builtins. "doc is also quite good in my opinion", of course

Intrinsics

• from cicc - 3183
• from clang - 1998
  

I think this is one of the secrets behind why Nvidia's code is faster - they just capable of generating 1.5 times as many PTX instructions. What happens when NVPTX backend can't compress group of operations into single PTX? Sure it generates several - like ld/cvt/something or something/cvt/st. And ptxas isn't enough smart to detect such patterns and apply more optimal instructions

 

Conclusion

Category	cicc	Clang (NVPTX)
Builtins	870	715
Intrinsics	3183	1998

If you use MLIR - then you cursed forever must use NVPTX backend only and your code will always be 15-20% slower