sparc-unknown-none-elf target broken

[thejpster]

If I try and build https://github.com/ferrous-systems/sparc-experiments with today's nightly, I get:

/home/jonathan/Downloads/gaisler/sparc-bcc-2.3.0-llvm/bin/sparc-gaisler-elf-ld: unknown architecture of input file `/tmp/rustc4jBo9z/symbols.o' is incompatible with sparc output

Using -Csave-temps, I can see:

$ file /tmp/rustc4jBo9z/symbols.o
/tmp/rustc4jBo9z/symbols.o: ELF 32-bit MSB relocatable, SPARC32PLUS, total store ordering, version 1 (SYSV), not stripped
$ file ./target/sparc-unknown-none-elf/debug/deps/sparc_demo_rust-8c8aedb85be3cab2.0unncljkfm3eoay971moazawq.rcgu.o 
./target/sparc-unknown-none-elf/debug/deps/sparc_demo_rust-8c8aedb85be3cab2.0unncljkfm3eoay971moazawq.rcgu.o: ELF 32-bit MSB relocatable, SPARC, version 1 (SYSV), with debug_info, not stripped

They are not the same.

Doing a bisect, it falls over on 2024-08-08. I suspect it might be d1d21ed, which probably should have picked Architecture::Sparc instead.

diff --git a/compiler/rustc_codegen_ssa/src/back/metadata.rs b/compiler/rustc_codegen_ssa/src/back/metadata.rs
index 0fd9d7fffe8..ea9327dc0f1 100644
--- a/compiler/rustc_codegen_ssa/src/back/metadata.rs
+++ b/compiler/rustc_codegen_ssa/src/back/metadata.rs
@@ -208,7 +208,7 @@ pub(crate) fn create_object_file(sess: &Session) -> Option<write::Object<'static
         "powerpc64" => (Architecture::PowerPc64, None),
         "riscv32" => (Architecture::Riscv32, None),
         "riscv64" => (Architecture::Riscv64, None),
-        "sparc" => (Architecture::Sparc32Plus, None),
+        "sparc" => (Architecture::Sparc, None),
         "sparc64" => (Architecture::Sparc64, None),
         "avr" => (Architecture::Avr, None),
         "msp430" => (Architecture::Msp430, None),

I did a build and my repo now builds and runs.

[workingjubilee]

@glaubitz Can you explain why Sparc32Plus is the "proper" architecture?

[glaubitz]

@glaubitz Can you explain why Sparc32Plus is the "proper" architecture?

It's been the default baseline for 32-bit SPARC on Linux for a long time already as it was the only 32-bit baseline for a long time to fully support atomic operations until Leon came around.

If you need to support a 32-bit SPARC baseline with EM_SPARC, we should find a way to differentiate between 32-bit SPARC on Linux and the sparc-unknown-none-elf target.

[workingjubilee]

I think we should probably conditionalize this on Linux, then (or on "none" as an OS, whatevr)

[bjorn3]

Do the linux and none targets set a different target cpu? If so maybe switching based on the target cpu would work?

[workingjubilee]

Switching based on the target cpu seems incorrect?

[bjorn3]

If you have a target cpu which doesn't support the sparc32plus isa, then using regular sparc32 as elf machine is correct and if the target cpu does support it, LLVM will likely emit instructions that require it and thus sparc32plus is the correct elf machine. Or am I misunderstanding the difference between the two elf machine ids?

[workingjubilee]

...wtf, they made their ISA extension level part of the ABI

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

Also referencing the Zulip discussion.

@rustbot label -I-prioritize +P-low