Improved optimization pass scheduling

[vouillon]

No description provided.

[hhugo]

I was playing with the same idea and got the same diffs in the tests, do you understand where they come from ?

[vouillon]

I was playing with the same idea and got the same diffs in the tests, do you understand where they come from ?

It's because we add a phi pass before the first deadcode pass. So the name of some variables which are actually unused get propagated to previous variables.

[hhugo]

@vouillon, I've rebased your branch and made some changes to accommodate the new check

[hhugo]

It's a bit difficult to analyze the benchmarks for this PR because the number of optimization passes is now different.

The only noticeable regression is with Fiat-Crypto, phi + flow went from 12s to 17s. I don't know if we can do anything about it.

[hhugo]

Can you also add a Changelog entry

[vouillon]

The only noticeable regression is with Fiat-Crypto, phi + flow went from 12s to 17s. I don't know if we can do anything about it.

I have pushed a commit to skip the phi simplification pass in the first round. It is expensive (since blocks still have a lot of parameters) and not very helpful.