update to rustc 1.98

[Firestar99]

Requires #610

Note

This is a nightly as close to rustc 1.98 as we can get

Notable changes:

• This commit does a few cleanups, assuming ScalarPairs are passed by PassMode::ScalarPair exclusively. But rust-gpu readjusts the fn abi here to pass any non-slice ScalarPair as PassMode::Direct. Removing this ABI change breaks a bunch of code in all kinds of weird ways, primarily ByteAddressableBuffer intrinsics and panic!() format args decompiler complaining about function pointers and pointer casts. It's just that rust-gpu can't properly handle PassMode::ScalarPair in so many places... So we'll keep our abi adjustments and I just undo that upstream cleanup.

close #605

[Firestar99]

Going from working nightly-2026-06-12 (rust-lang/rust@b30f3df) to nighty-2026-06-13 (rust-lang/rust@6540795) causes our compiletests to break due to some upstream change I can't identify.

The only two PRs that change anything in rustc_codegen_ssa in that range are:

• Move create_scope_map to rustc_codegen_ssa. rust-lang/rust#157722 which only touches debug info (and requires us to remove some todo!s that now get called) it's this PR
• codegen_ssa: peel trans. wrappers on scalable vecs rust-lang/rust#157768 which should only affect scaleable vecs (aarch64 SIMD stuff) which we're not using

UPDATE: We can fix the compiletests by just removing this flag:

-        "-Cdebuginfo=2",

This flag is responsible for controlling debug symbol emission, 2 means all the debug symbols and it defaults to 0 (I think) meaning none.

How the hell does turning on debug symbol emission cause fundamental compilation errors?

• Move create_scope_map to rustc_codegen_ssa. rust-lang/rust#157722 moved debug symbol emission from cg_lllvm to cg_ssa so cg_gcc could take advantage of the shared code path. This included refactoring the cg_ssa API to account for the new emission format.
• They removed this function in particular:

    fn create_function_debug_context(
        &self,
        _instance: Instance<'tcx>,
        _fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
        _llfn: Self::Function,
        _mir: &mir::Body<'tcx>,
    ) -> Option<FunctionDebugContext<'tcx, Self::DIScope, Self::DILocation>> {
        // TODO: This is ignored. Do we want to implement this at some point?
        None
    }

• Notice how it returns an Option and we always return None? With the debug symbol refactor, codegen backends can't choose to just return None and disable debug info emission. Whether to emit or not is exclusively determined by the debuginfo flag. Which is why I had to remove a bunch of todo!()s in all our dbg fn impls.
• To correctly emit debug information, the new cg_ssa impl needs to monomorphise inlined functions, which causes them to be emitted to our codegen backend. Specifically, I see a lot of <usize>::unchecked_add calling <usize>::unchecked_add::precondition_check with a &str as an argument.
• That &str value is a [u8] constant, which are represented as a RuntimeArray in rust-gpu's codegen and not well supported. Eventually it hits this zombification, erroring with error: unsupported unsized [u8] constant.

I beliebe the fix I made here is only a temporary fix, I suspect any debug build of sufficient complexity will fail. Don't yet know what to do about that.

[Firestar99]

Instead of opting out compiletests from debug info, I decided to just patch out the newest changes to debuginfo emission

[nazar-pc]

nightly-2026-07-16 just fixed a lot of issues with const generics under generic_const_args.

I really with rust-gpu switches to it soon, I could remove so many temporary workarounds then 🙏

[Firestar99]

@nazar-pc this PR only goes until nightly-2026-07-03, cause that's the commit rustc 1.98 branched. You'll find nightly-2026-07-16 in PR #630

I'm considering holding off a bit before merging this, would like alpha.2 out before that.

[FractalFir]

Made a quick pass, found some suspicious things here and there.

[FractalFir]

Seems problematic to ignore Align like this - how about checking that default_align is equal to this Align, or, better yet - use this align instead of b.primitive().default_align(self).abi?

[Firestar99]

To be completely honest, I have no idea what this function is even supposed to do. Sure you can probably find some unsafe rust code that reaches it, but like what it's supposed to do from the shader codegen perspective.

This is sadly the case with quite a bit of code in rustc_codegen_spirv, where I highly suspect it's effectively unreachable for any sane shader, there is no compiletest that reaches that code, so I really don't know what it's really supposed to do.

[FractalFir]

Align::ONE? The argument is seemingly ignored for now, but this is an issue for the future, when it will not. You already got the layout - why not just get the align from there?

[FractalFir]

Unrelated to the main changes, but this is a kind of silly way to do patches. It can silently fail when things change around: replace does not assert the OG source contains the lines we want to replace, and fails silently if it does not. It might be worth it to add a helper like this:

fn apply_patch(src:&str, og:&str, new:&str)->String{
    assert!(src.contains(og), "ppg_cg_ssa patch does not apply!");
    src.replace(og,new)
}

Would make the failure mode louder, and easier to debug.

[Firestar99]

I've thought about replacing it with a directory of *.patch files before. In fact I tripped up on a replace failing due to upstream changes and it not causing a compilation error but simply a behavior change. Luckily, I added that regex just a month ago so I knew what failed.