[Mlir-commits] [mlir] [mlir] Make symbol-user type verification proportional to participating IR (PR #212354)

[Will Dietz]

dtzSiFive wrote:

> @dtzSiFive are you able to easily share your benchmark so that I can benchmark these different options on a real workload?

The real workload, no, it's a private design.

I've also been using a public RocketChip core and can direct you how to generate that or share the MLIR with you if that's useful.  However it's primarily using dialects specific to the CIRCT project, so I'm not sure if that's how you'd like to proceed. 

> The other we tested against the current simpler implementation and found it was slightly worse (+3.5%) on our synthetic tests.

Good catch, apologies.  This was the "minimal" suggestion I made, right? Yeah that was a mistake on my end.
The full version drops the walker entirely and combines its memo table with a local cache / state bits for the walk, essentially.  That was the improvement.  Thanks for the careful benchmarking on your end.

FWIW I don't see in your table entries for "before this interface was added", which might help re:apples-to-apples as we weigh the cost of this verification.  If you're able to see 2-3x, that's about what we're seeing, then your synthetic workloads are likely enough to continue development against.

https://github.com/llvm/llvm-project/pull/212354