[ConstraintElim] Add facts about non-poison intrinsics on demand

[dtcxzyw]

Address comment #80121 (comment).
TODO: caching added non-poison values

[dtcxzyw]

Compile-time overhead is high: https://llvm-compile-time-tracker.com/compare.php?from=2ba455ff3dd192ff814514647d485491510a6342&to=7338200a504c6f16f4394fbcbbe632bcfcc54b6e&stat=instructions%3Au
@fhahn @nikic Do you think this patch is in the right direction? If so, I can further reduce the overhead by maintaining a global cache for visited non-poison values.

[nikic]

I don't really understand the approach you are using here. What I expected to see is something along the lines of how we handle IsKnownNonNegative, where we add additional constraints to ExtraInfo, which will be used while checking the condition only.

Basically, would it be possible to extend ConstraintInfo::getConstraint to go through the variables, and if they are umin/etc add the additional constraint to ExtraInfo? Or is that insufficient?

[dtcxzyw]

I don't really understand the approach you are using here.

If an instruction is guaranteed not to be a poison and not to produce a poison, its operands are also non-poison (except for select). We collect a set of non-poison values used as conditions first, then find the closure of the set with the rule above. Finally we add constraints for non-poison intrinsics.