Add a compiler intrinsic to back bigint_helper_methods

[scottmcm]

cc #85532

This adds a new carrying_mul_add intrinsic, to implement wide_mul and carrying_mul.

It has fallback MIR for all types -- including u128, which isn't currently supported on nightly -- so that it'll continue to work on all backends, including CTFE.

Then it's overridden in cg_llvm to use wider intermediate types, including i256 for u128::carrying_mul.

[rustbot]

r? @ibraheemdev

rustbot has assigned @ibraheemdev.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[rustbot]

Some changes occurred to the intrinsics. Make sure the CTFE / Miri interpreter
gets adapted for the changes, if necessary.

cc @rust-lang/miri, @rust-lang/wg-const-eval

[clarfonthey]

Will take a look at this later, but I like that you managed to make fallback intrinsics work here when I couldn't quite get them to work for #132195.

[bors]

☔ The latest upstream changes (presumably #133728) made this pull request unmergeable. Please resolve the merge conflicts.

[clarfonthey]

Also just to clarify my opinion on where this stands relative to #132195: I think that this PR just adding the intrinsic is okay, although I would like the intrinsic to also handle the signed case, even if it's only used at the moment for the unsigned case. That way, my PR can just update the library methods to use the intrinsic, without having to worry about the slow implementation.

Although the signed version of the library methods will likely return an unsigned word always for the least-significant / high-order part, it vastly simplifies the intrinsic to just have the same type for the arguments and return value, so, I would prefer that we do this specifically for the intrinsic. Since it doesn't actually affect the returned bits, I think that it's okay to just cast half of it unsigned on the library side. You can see how I did this here in a previous (closed) PR.

Otherwise, LGTM. Also, while the codegen test should suffice, if you want to copy over the library tests I added in my PR to verify that the intrinsic works correctly, feel free. I specifically added a bunch of edge cases with MAX and MIN as operands to ensure that things work correctly.

[scottmcm]

Success! No more "oh my this is hacky", just a couple of impl consts with fairly normal code.

it vastly simplifies the intrinsic to just have the same type for the arguments and return value

It didn't, actually. Once the fallback was working, the implementation in cg_llvm was super-easy -- after all, at that level the signed and unsigned types are the same anyway.

[clarfonthey]

Just comparing to my version, this doesn't really matter that much, but const_undef feels slightly more accurate here just semantically. Not actually sure if this affects codegen though.

[scottmcm]

poison is always preferred where possible, because it doesn't have the same complications as undef. (undef << 1) & 1 is guaranteed to be 0, but (poison << 1) & 1 is still poison. You'll also see that both what we emit for tuples and what the optimizer does for tuples both use poison for this: https://rust.godbolt.org/z/WTnW4GPE6

[scottmcm]

Pop Quiz: after a double-width isize::MIN * isize::MIN (which is the largest possible positive result from a multiplication), how many more isize::MAXs can you add to the result before it would need to carry to triple-width?

Answer: isize::MAX + 2 of them!

(The carrying_mul_add really feels so much nicer for unsigned types, since uN::MAX * uN::MAX + uN::MAX + uN::MAX == u2N::MAX -- no more space available after that.)

[ibraheemdev]

This is outside my territory so I'm going to pass this along. r? libs

[bors]

☔ The latest upstream changes (presumably #134516) made this pull request unmergeable. Please resolve the merge conflicts.

[bors]

💔 Test failed - checks-actions

[scottmcm]

Fix was just changing some i64s to {{i32|i64}}s in GEPs.

@bors r=Amanieu

[bors]

📌 Commit 4669c0d has been approved by Amanieu

It is now in the queue for this repository.