Re: [PATCH] arm64: lse: fix LSE atomics with LLVM's integrated assembler

[Nick Desaulniers]

On Mon, Oct 7, 2019 at 2:24 PM Sami Tolvanen <samitolvanen@xxxxxxxxxx> wrote:
>
> On Mon, Oct 7, 2019 at 1:28 PM 'Nick Desaulniers' via Clang Built
> Linux <clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
> > I tried adding `.arch armv8-a+lse` directives to all of the inline asm:
> > https://github.com/ClangBuiltLinux/linux/issues/573#issuecomment-535098996
>
> Yes, I had a similar patch earlier. I feel like using a command line
> parameter here is cleaner, but I'm fine with either solution.
>
> > One thing to be careful about is that blankets the entire kernel in
> > `+lse`, allowing LSE atomics to be selected at any point.
>
> Is that a problem? The current code allows LSE instructions with gcc
> in any file that includes <asm/lse.h>, which turns out to be quite a
> few places.

I may be mistaken, but I don't think inline asm directives allow the C
compiler to change what instructions it selects for C code, but
command line arguments to the C compiler do.

Grepping the kernel for some of the functions and memory orderings
turns up a few hits:
https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html

I'm worried that one of these might lower to LSE atomics without
ALTERNATIVE guards by blanketing all C code with `-march=armv8-a+lse`.
But I did just boot test this patch but using GAS in QEMU (on a -cpu
cortex-a72 which I suspect should not have lse instructions by default
IIUC), FWIW.
Tested-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
Maybe the maintainers have more thoughts?
-- 
Thanks,
~Nick Desaulniers