Merge linear control-flow before reconstruction

[Nadrieril]

This PR does several cfg-related things:

• Makes some ullbc terminators be statements instead; this reduces terminators to be the bare minimum needed for control flow: goto, switch, return and abort;
• Adds a pass that merges linear goto chains into a single block. This happens particularly in bodies with lots of function calls; this lightens the work for the control-flow reconstruction pass;
• These two transformations made the llbc output slightly worse, so this PR also adds a pass that improves the cfg.

[sonmarcho]

What does this do?

[Nadrieril]

That's a drive-by change that keeps more debuginfo in release mode, so that we get useful backtraces and flamegraphs. Otherwise a lot of info is lost and the flamegraphs are useless

[Nadrieril]

There's too high a risk of breaking libcrux because eurydice relies on specific control-flow shapes to detect loops. I'll remove the last pass added in this PR and keep it for when libcrux is fixed.

[sonmarcho]

There's too high a risk of breaking libcrux because eurydice relies on specific control-flow shapes to detect loops. I'll remove the last pass added in this PR and keep it for when libcrux is fixed.

I guess we should open an issue for this, then?
For libcrux: you mean to generate for and while loops, right?

[Nadrieril]

I guess we should open an issue for this, then?

I mean, I've got the branch ready for whenever it's fixed, no need

For libcrux: you mean to generate for and while loops, right?

yes

[sonmarcho]

I guess we should open an issue for this, then?

I mean, I've got the branch ready for whenever it's fixed, no need

As long as we don't forget it, I'm fine with that.