[codegen] assume the tag, not the relative discriminant

[scottmcm]

Address the issue mentioned in llvm/llvm-project#134024 (comment) by changing discriminant calculation to assume on the originally-loaded tag, rather than on cast(tag)-OFFSET.

The previous way does make the purpose of the assume clearer, IMHO, since you see assume(x != 4); if p { x } else { 4 }, but doing it this way instead means that the adds optimize away in LLVM21, which is more important. And this new way is still easily thought of as being like metadata on the load saying specifically which value is impossible.

Demo of the LLVM20 vs LLVM21 difference: https://llvm.godbolt.org/z/n54x5Mq1T

r? @nikic

[rustbot]

Some changes occurred in compiler/rustc_codegen_ssa

cc @WaffleLapkin

[nikic]

Wouldn't this select be at the bottom for LLVM 21?

[scottmcm]

Oops, I completely missed that! Thanks.

[nikic]

This looks like a reasonable thing to do to me, but I'm probably the wrong person to check whether the logic is correct. The surrounding niche/discriminant logic looks kinda tricky.

[scottmcm]

r? codegen

[scottmcm]

In case this helps review, here's how you can see this just from the diff, rather than from https://doc.rust-lang.org/nightly/nightly-rustc/rustc_abi/enum.TagEncoding.html#variant.Niche:

Previously we were doing tagged_discr != untagged_variant here.

But

relative_discr = tag - niche_start
delta = niche_variants.start()
tagged_discr  = relative_discr + delta

so

   tagged_discr != untagged_variant
=> relative_discr + delta != untagged_variant
=> (tag - niche_start) + niche_variants.start() != untagged_variant
=> tag != niche_start  + untagged_variant - niche_variants.start() 

which is the calculation on line 522.

[WaffleLapkin]

@bors r+

[bors]

📌 Commit b90e0fb has been approved by WaffleLapkin

It is now in the queue for this repository.