make `Base.@assume_effects :nothrow` mark throwing expressions as `unreachable` about julia

[mikmoore]

This is a clever way to implement unreachable, but I don't like it. It's too hard to use properly. An unreachable() function or macro seems to be a much safer and more ergonomic implementation.

Base.@assume_effects is annotated at the function level. This makes it too coarse-grained to use responsibly for this purpose. EDIT: it can be applied to blocks in upcoming versions (thanks for the correction below). A :nothrow annotation like this would also seem to risk applying recursively to any functions called (or at least to inlined functions), which makes this dangerous in the first place and a maintenance nightmare if any of those called functions ever change.

I think we shouldn't have this for many of the same reasons we don't have global @fastmath (any more).

[nsajko]

Not true, since v1.11 @assume_effects may be appllied to a code block: https://docs.julialang.org/en/v1.12-dev/base/base/#Base.@assume_effects. For example the nothrow below should only apply to the 7::Union{} AFAIK:

Base.@assume_effects :nothrow let
    7::Union{}
end

Obviously unreachable() is intrinsically unsafe, however, given that we can apply nothrow to a code block, I think my proposal is as safe as possible?

[mikmoore]

Not true, since v1.11

Thanks for the correction. I vaguely recalled seeing something like that but it's hard to keep up with all the upcoming changes.

Really, you're right that (in at least some senses) @assume_effects should have this power. Otherwise it's not really assuming the stated effects. Although taking that line of thinking further, we should allow any provable violation of an effects flag to emit an unreachable.

That said, as far as I understand (though my understanding is poor), @assume_effects has virtually no impact on the internal operation of the function it's defined on. Its impact is actually on call sites of that function. Given that understanding, this proposal would dramatically expand the scope of @assume_effects's operation. So it seems like the wrong place for this functionality.

I also don't relish the idea of encouraging people to impose blatantly false effects to twist the compiler into emitting unreachable. There should be a more direct option (although it could co-exist with the proposed feature).

[nsajko]

xref #54436