Use of RiscV "+forced-atomics" target feature triggers hard-float ABI for symbols.o

[paulmenage]

When building for RiscV-32, adding the "+forced-atomics" target feature (in this case, adding it to a custom target .json file) causes the symbols.o file to be built with the single-float ABI rather than the soft-float ABI used for the all .rcgu.o files, which causes the link to fail due to incompatible float ABIs.

It appears to be due to this code in compiler/rustc_codegen_ssa/src/back/metadata.rs:

        Architecture::Riscv32 | Architecture::Riscv64 => {
            // Source: https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/079772828bd10933d34121117a222b4cc0ee2200/riscv-elf.adoc
            let mut e_flags: u32 = 0x0;
            let features = &sess.target.options.features;
            // Check if compressed is enabled
            if features.contains("+c") {
                e_flags |= elf::EF_RISCV_RVC;
            }

            // Select the appropriate floating-point ABI
            if features.contains("+d") {
                e_flags |= elf::EF_RISCV_FLOAT_ABI_DOUBLE;
            } else if features.contains("+f") {
                e_flags |= elf::EF_RISCV_FLOAT_ABI_SINGLE;
            } else {
                e_flags |= elf::EF_RISCV_FLOAT_ABI_SOFT;
            }
            e_flags
        }

which assumes that any feature beginning with "+f" is in fact exactly "+f", so the addition of "+forced-atomics" triggers the single-float ABI.

Should this check (and most other feature checks) actually be something like: features.split(",").any(|f| f == "+f") (probably useful as a TargetOptions::has_feature() helper method)?

[paulmenage]

Cc: @nikic

[Noratrieb]

cc @nbdd0121 , I think you added this. I looked a bit through uses of target features and found

rust/compiler/rustc_codegen_ssa/src/back/link.rs

Lines 2317 to 2323 in 0441150

	if let Some(feat) = [sess.opts.cg.target_feature.as_str(), &sess.target.options.features]
	.into_iter()
	.find(|feat| !feat.is_empty())
	{
	cmd.arg("--cpu-features");
	cmd.arg(feat);
	}

which looks pretty suspicious too. Fixing this bug is indeed simple but making sure this doesn't happen again is less so (although these two things of course don't have to happen in the same PR).

[nbdd0121]

I didn't recall changing this bit. This function has existed for quite a while for creating metadata files, I just also call it to create symbols.o.

That said, I'm happy to work on a fix.

@rustbot claim