[Wasm EH] Optimize throws caught by TryTable into breaks

[kripken]

E.g.

(try_table (catch_all $catch)
  (throw $e)
)

=>

(try_table (catch_all $catch)
  (br $catch)
)

This can then allow other passes to remove the TryTable, if no throwing
things remain.

[aheejin]

LGTM! By the way why would this routine be in RemoveUnusedBr, given that it rather creates branches? Not that I have another place in mind.. (Other than let's say Vacuum which has become a kitchensync of sorts...)

[aheejin]

What does this TODO mean?

[kripken]

Just that we can optimize exception handling here, both Try and TryTable. I'll do that soon.

[kripken]

(the handling will be similar to Loop)

[kripken]

By the way why would this routine be in RemoveUnusedBr, given that it rather creates branches?

My reasoning is that by lowering a throw to a break we can then optimize that break right here in the pass, without waiting to reach another pass (e.g. a test here shows a throw turning into a br that turn turns into a br_if).

But, yeah, this could have fit elsewhere. I see Vacuum more as "remove obviously unneeded (but not dead) code", so if I had to pick another place I'd use OptimizeInstructions, though this is a little more complex than the typical peephole optimization there...

[aheejin]

If a branch is created here, can it be optimized away in the same run of the pass? Apparently it can, as you've shown with the br_if example... Does that mean it traverses the program more than once?

[kripken]

Yes, this does loop over several simplifications, and keeps going until things stop changing. That main loop starts here:

binaryen/src/passes/RemoveUnusedBrs.cpp

Lines 920 to 923 in 7e14139

	void doWalkFunction(Function* func) {
	// multiple cycles may be needed
	do {
	anotherCycle = false;

This could probably be more efficient, but each traversal is practically linear. This pass didn't show up as slow in any of my recent performance measurements.

[aheejin]

I see, then it makes sense to put it here. Thanks!