[CombToArith] Fix coarsening of division by zero UB

[maerhart]

Fixes #6938

[fzi-hielscher]

Thanks! This will probably do for most practical cases. But I am unsure if we may not still be in trouble if we try to do 0 / 0. Going by the specification of the arith dialect, any division by zero is (immediate) UB. Alive2 also confirms this if we compare a clang compilation of the b == 0 ? 0 : a/b reference against a slightly massaged output of the integration test:

Transformation doesn't verify!

ERROR: Source is more defined than target

Example:
i8 %#0 = #x00 (0)
i8 %#1 = #x00 (0)

Source:
i1 %#5 = #x1 (1)
  >> Jump to %#4
i8 %#7 = #x00 (0)

Target:
i1 %#3 = #x1 (1)
i8 %#4 = #x00 (0)
i8 %#5 = UB triggered!

For practical purposes, I tested it on both x86_64 and ARM64 and the test produced 0 even for zero divded by zero. This is not so surprising for ARM64. I would have expected a trap on x86_64 though, see #6726 . This could be a subtlety of the JIT engine and the way it handles UB. But maybe I've just messed something up, which wouldn't be surprising either. 😖

On a more general note, I do think we need to tinker with the HW/Comb specification to get a clean separation between undefined behavior, undefined values (undef, poison, frozen poison, X, you name it...) and the conditions which must/may/mustn't terminate a simulation.

[fabianschuiki]

LGTM!

[darthscsi]

LGTM

[maerhart]

@fzi-hielscher Thanks for the detailed review! I've changed it to divide by 1 if the divisor is 0 originally.

[fzi-hielscher]

Nice, thank you.