gccrs: 128bit integer support is limited on 32 bit platforms

[philberty]

This is a wider gcc issue we will need to address at some point but this test runs fine under 32 bit more without the 128bit lang item for rotate here.

gcc/testsuite/ChangeLog:

* rust/execute/torture/sip-hasher.rs: remove 128bit lang item

[thesamesam]

Maybe as @tschwinge suggested it'd be better to dg-ice for 32-bit?

[thesamesam]

Should link to #4003 in the commit message anyway

[philberty]

yeah i was testing quite a bit here and normal things like the arithmetic operators like +, - etc wok on 128 bit integers but rotate doesnt the C front-end explicitly says 128bit doesnt work on 32 bit.

From testing it seems as though rotates and shifts dont work.

The RTL passes do try to expand the rotate into two parts like two u64's as far as i can tell in gdb but this a bit beyond my ability right now its hitting stuff in the code relating to the target option support so i think its a wider thing to ask GCC folks about at this point.

[philberty]

@tschwinge @dkm @CohenArthur what do you think i should do? I think i can tweak this test case to pass for this alpine 32 bit CI but not sure who to poke in GCC about this maybe i should just ask on the gcc mailing list i guess?

[tschwinge]

As far as I know, GCC supports 128-bit integers only for certain (64-bit?) targets, but not for 32-bit targets. As far as I remember, the new GCC/Cobol support is also running into that (because they need 128-bit integers) -- so maybe that'll get resolved sooner or later.

I see C, C++ test cases conditionalize __int128 usage on #ifdef __SIZEOF_INT128__, or effective-target int128, for example. The latter works for GCC/Rust, too, so I suppose we could do something like dg-additional-options -frust-cfg=int128 { target int128 }, and conditionalize u128, i128 on #[cfg(int128)] (untested)?

[tschwinge]

A follow-on question then is whether Rust mandates u128, i128 support unconditionally, so we'll have to make that work also for 32-bit configurations? (Like Cobol, apparently.)

[tschwinge]

By the way, I see u128, i128 appear in some of our test cases, and commented out in others. All that could -- at least for now -- use effective-target int128, I suppose?

[powerboat9]

https://blog.rust-lang.org/2024/03/30/i128-layout-update/
rust-lang/lang-team#255
https://godbolt.org/z/EYMbdx63K

Looks like rustc didn't have a stable ABI for i128 in 1.49. However, I'm fairly certain rustc supports i128 on ~all targets.

[powerboat9]

rust-lang/rustc_codegen_gcc#71
rust-lang/rustc_codegen_gcc#103

[dkm]

@tschwinge @dkm @CohenArthur what do you think i should do? I think i can tweak this test case to pass for this alpine 32 bit CI but not sure who to poke in GCC about this maybe i should just ask on the gcc mailing list i guess?

if you don't feel like sending a mail, IRC is also an option.