Re: [PATCH 1/3] crypto: X25519 low-level primitives for ppc64le.

From: Andy Polyakov
Date: Thu May 16 2024 - 09:42:28 EST

Hi,

+.abiversion 2

I'd prefer that was left to the compiler flags.

Problem is that it's the compiler that is responsible for providing this
directive in the intermediate .s prior invoking the assembler. And there
is no assembler flag to pass through -Wa.

Hmm, right. But none of our existing .S files include .abiversion
directives.

We build .S files with gcc, passing -mabi=elfv2, but it seems to have no
effect.

So all the intermediate .o's generated from .S files are not ELFv2:

$ find .build/ -name '*.o' | xargs file | grep Unspecified
.build/arch/powerpc/kernel/vdso/note-64.o: ELF 64-bit LSB relocatable, 64-bit PowerPC or cisco 7500, Unspecified or Power ELF V1 ABI, version 1 (SYSV), not stripped

I would guess that contemporary linker is more forgiving than it was back then when the .abiversion directive was added. If it works now, then it of course can be omitted. I suppose my original remark should be viewed rather as "you can't replace it with a command line option" than "you can't make it work without it." :-)

But the actual code follows ELFv2, because we wrote it that way, and I
guess the linker doesn't look at the actual ABI version of the .o ?

So it currently works. But it's kind of gross that those .o files are
not ELFv2 for an ELFv2 build.

Well, as far as passing base types and pointers to/from assembly goes, there are no differences between the versions. Then it's a question of meaning assigned to r2 and r13, but as long as you don't touch them, you can freely reuse the code with either ABI. With this in mind the abiversion directive is effectively reduced to just a marker in the .o file. In other words the instruction sequences by themselves are customarily ABI-neutral, at least in "general calculation" modules such as the suggested one, so that if it works 100% without the .abiversion directive, then it can be safely omitted.

Cheers.