On Sun, Oct 6, 2024 at 1:37 AM H. Peter Anvin <hpa@xxxxxxxxx> wrote:
>
> On 10/5/24 01:31, Uros Bizjak wrote:
> >>
> >> movq $sym to leaq sym(%rip) which you said ought to be smaller (and in
> >> reality appears to be the same size, 7 bytes) seems like a no-brainer
> >> and can be treated as a code quality issue -- in other words, file bug
> >> reports against gcc and clang.
> >
> > It is the kernel assembly source that should be converted to
> > rip-relative form, gcc (and probably clang) have nothing with it.
> >
>
> Sadly, that is not correct; neither gcc nor clang uses lea:
>
> -hpa
>
>
> gcc version 14.2.1 20240912 (Red Hat 14.2.1-3) (GCC)
>
> hpa@tazenda:/tmp$ cat foo.c
> int foobar;
>
> int *where_is_foobar(void)
> {
> return &foobar;
> }
>
> hpa@tazenda:/tmp$ gcc -mcmodel=kernel -O2 -c -o foo.o foo.c
Indeed, but my reply was in the context of -fpie, which guarantees RIP
relative access. IOW, the compiler will always generate sym(%rip) with
-fpie, but (obviously) can't change assembly code in the kernel when
the PIE is requested.
Otherwise, MOV $immediate, %reg is faster when PIE is not required,
which is the case with -mcmodel=kernel. IIRC, LEA with %rip had some
performance issues, which may not be the case anymore with newer
processors.
Due to the non-negligible impact of PIE, perhaps some kind of
CONFIG_PIE config definition should be introduced, so the assembly
code would be able to choose optimal asm sequence when PIE and non-PIE
is requested?
Uros.
