[Mlir-commits] [mlir] [mlir] Do not merge blocks during canonicalization by default (PR #95057)

[Mehdi Amini]

joker-eph wrote:

> I had one ready to introduce a numBlocksArgs threshold. I am not sure which one is better, but my idea was that reducing the number of blocks, up to a point, was something useful (especially on GPUs where branches are so bad)

What is the problem you're seeing with the number of block arguments? (in terms of code generation they are lowered to "nothing", basically just a register constraint, worst case a register copy)

https://github.com/llvm/llvm-project/pull/95057