sparc: make ABI consistent with clang by folkertdev · Pull Request #160401 · rust-lang/rust

folkertdev · 2026-08-02T20:16:49Z

I've hacked together a version of abi-cafe that can cross-compile. It finds many GCC vs. Clang inconsistencies, but for this target it found 3 ABI bugs that are specific to the rust implementation.

For reference

Pass ZST arguments

https://godbolt.org/z/P7KoqTWvr

The sparc ABI does not skip zero-sized arguments.

Pass and return `f128` (i.e. `long double`) indirectly

The long double type, which corresponds to rust f128, is special-cased, it uses getNaturalAlignIndirect for both arguments and returns.

pass aggregate arguments by reference

https://godbolt.org/z/hM8Gh1bfY

The implementation defers to DefaultABIInfo::{classifyReturnType, classifyArgumentType}, which use byval and sret for aggregate types.

long double (i.e. f128) is special-case in the sparc abi

folkertdev · 2026-08-02T20:23:11Z

r? workingjubilee (who else lol)
cc @jonathanpallant as target maintainer

This is totally tier 3, but if we have it, it should be correct.

rustbot · 2026-08-02T20:23:15Z

workingjubilee is currently at their maximum review capacity.
They may take a while to respond.

jonathanpallant · 2026-08-03T09:28:08Z

I will dig out my TSIM3 licence and double check this on the LEON3 simulator.

I'm pinging @glaubitz who knows more about the 64-bit SPARC Linux side than I do.

(Also, sparc64-unknown-linux-gnu is Tier 2 with no platform docs, and hence no recorded maintainer ... so that's not ideal)

folkertdev · 2026-08-03T10:21:23Z

Neat, to be clear I validated the changes with qemu, clang and gcc 16.

I'm pinging glaubitz who knows more about the 64-bit SPARC Linux side than I do.

The changes here only touch 32-bit SPARC, and the ABI is, from what I can tell, quite different to SPARC64.

beetrees

Since this PR is to fix the SPARC struct ABI could you add "Fixes: #43894"?

View changes since this review

beetrees · 2026-08-03T12:51:59Z

-        let pad_i32 = !offset.is_aligned(align);
-        arg.cast_to_and_pad_i32(Uniform::new(Reg::i32(), size), pad_i32);
+    if is_long_double(arg.layout.backend_repr) || arg.layout.is_aggregate() {
+        arg.pass_by_stack_offset(None);


Suggested change

arg.pass_by_stack_offset(None);

arg.make_indirect();

According to the 32-bit SPARC ABI, this should be passed indirectly, not by stack offset ("[...] structure and union objects are not passed directly in the argument list."). pass_by_stack_offset appears to work with LLVM as the LLVM SPARC backend appears to ignore the byval attribute (compiler explorer), but we should use make_indirect as other codegen backends might not ignore it.

folkertdev

Since this PR is to fix the SPARC struct ABI could you add "Fixes: #43894"?

added, but also wow from 2017 huh

View changes since this review

folkertdev · 2026-08-03T13:14:14Z

-        let pad_i32 = !offset.is_aligned(align);
-        arg.cast_to_and_pad_i32(Uniform::new(Reg::i32(), size), pad_i32);
+    if is_long_double(arg.layout.backend_repr) || arg.layout.is_aggregate() {
+        arg.pass_by_stack_offset(None);


this matches GCC and Clang.

jonathanpallant · 2026-08-03T13:37:26Z

Do our SPARC Linux and SPARC bare-metal targets have the same ABI? Bare metal SPARC is aimed at the LEON3 but I think SPARC Linux is for 32 bit applications for 64 bit SPARC machines?

Which QEMU machine did you test with?

folkertdev · 2026-08-03T13:49:25Z

LLVM does not take the OS into account at all

https://github.com/llvm/llvm-project/blob/6849789a8fa1fc4a09f4f81dc655235e648eb15a/clang/lib/CodeGen/Targets/Sparc.cpp#L34-L59

My config is

[target.sparc-unknown-linux-gnu]
runner = "qemu-sparc32plus -L /usr/sparc-linux-gnu"

So that runs the v9-flavored v8, at least based on https://stackoverflow.com/questions/23506538/what-is-em-sparc32plus-for. That is needed because we configure the +v8plus feature for the linux target

rust/compiler/rustc_target/src/spec/targets/sparc_unknown_linux_gnu.rs

Lines 17 to 24 in 7c329d6

    
           options: TargetOptions { 
        
               features: "+v8plus".into(), 
        
               cpu: "v9".into(), 
        
               endian: Endian::Big, 
        
               late_link_args: TargetOptions::link_args( 
        
                   LinkerFlavor::Gnu(Cc::Yes, Lld::No), 
        
                   &["-mcpu=v9", "-m32"], 
        
               ),

To clarify: I believe this uses V9 instructions (or something like it), but still uses the V8 abi.

taiki-e · 2026-08-04T11:37:11Z

To clarify: I believe this uses V9 instructions (or something like it), but still uses the V8 abi.

+v8plus enables the use of the v8+ ABI. See #132552 for details. (I think it would actually be more appropriate to use the rustc_abi field here in addition to features, but I don't think that field was implemented at the time.)

RalfJung · 2026-08-05T13:41:47Z

Do our SPARC Linux and SPARC bare-metal targets have the same ABI?

I don't think so, no. At least, sparc-unknown-none-elf does not set v8plus which I think means it uses a different ABI.

RalfJung · 2026-08-05T13:49:49Z

#160562 adds a rustc_abi toggle for v8plus.

folkertdev added 2 commits August 2, 2026 21:31

sparc: pass ZST arguments

76d32d5

sparc: pass and return f128 indirectly

78fb9e1

long double (i.e. f128) is special-case in the sparc abi

rustbot added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. labels Aug 2, 2026

rustbot assigned workingjubilee Aug 2, 2026

folkertdev marked this pull request as ready for review August 2, 2026 22:55

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Aug 2, 2026

beetrees suggested changes Aug 3, 2026

View reviewed changes

folkertdev commented Aug 3, 2026

View reviewed changes

folkertdev added the O-SPARC Target: SPARC processors label Aug 3, 2026

sparc: pass aggregate arguments by reference

4873fa1

this matches GCC and Clang.

folkertdev force-pushed the sparc-abi-fixes branch from ccdb6c7 to 4873fa1 Compare August 3, 2026 13:29

taiki-e mentioned this pull request Aug 5, 2026

add target feature ABI checks for SPARC #160562

Open

Uh oh!

Uh oh!

Conversation

folkertdev commented Aug 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pass ZST arguments

Pass and return f128 (i.e. long double) indirectly

pass aggregate arguments by reference

Uh oh!

folkertdev commented Aug 2, 2026

Uh oh!

rustbot commented Aug 2, 2026

Uh oh!

jonathanpallant commented Aug 3, 2026

Uh oh!

folkertdev commented Aug 3, 2026

Uh oh!

beetrees left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

beetrees Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

folkertdev Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

folkertdev left a comment • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

folkertdev Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

jonathanpallant commented Aug 3, 2026

Uh oh!

folkertdev commented Aug 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

taiki-e commented Aug 4, 2026

Uh oh!

RalfJung commented Aug 5, 2026

Uh oh!

RalfJung commented Aug 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

folkertdev commented Aug 2, 2026 •

edited

Loading

Pass and return `f128` (i.e. `long double`) indirectly

beetrees left a comment •

edited

Loading

folkertdev left a comment •

edited by rustbot

Loading

folkertdev commented Aug 3, 2026 •

edited

Loading