On Tue. 2 janv. 2024 at 19:28, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
Hi Vincent,
Thanks for your patch!
Thanks for the review and for running the benchmark.
On Sun, Dec 17, 2023 at 8:13 AM Vincent Mailhol
<mailhol.vincent@xxxxxxxxxx> wrote:
The inline keyword actually does not guarantee that the compiler will
inline a functions. Whenever the goal is to actually inline a
function, __always_inline should always be preferred instead.
On an allyesconfig, with GCC 13.2.1, it saves roughly 5 KB.
$ size --format=GNU vmlinux.before vmlinux.after
text data bss total filename
60449738 70975612 2288988 133714338 vmlinux.before
60446534 70972412 2289596 133708542 vmlinux.after
With gcc 9.5.0-1ubuntu1~22.04, the figures are completely different
(i.e. a size increase):
Those results are not normal, there should not be such a big
discrepancy between two versions of the same compiler. I double
checked everything and found out that I made a mistake when computing
the figures: not sure what exactly, but at some point, the ASLR seeds
(or other similar randomization feature) got reset and so, the
decrease I witnessed was just a "lucky roll".
After rerunning the benchmark (making sure to keep every seeds), I got
similar results as you:
text data bss total filename
60449738 70975356 2288988 133714082
vmlinux_allyesconfig.before_this_series
60446534 70979068 2289596 133715198
vmlinux_allyesconfig.after_first_patch
60429746 70979132 2291676 133700554
vmlinux_allyesconfig.final_second_patch
Note that there are still some kind of randomness on the data segment
as shown in those other benchmarks I run:
text data bss total filename
60449738 70976124 2288988 133714850
vmlinux_allyesconfig.before_this_series
60446534 70980092 2289596 133716222
vmlinux_allyesconfig.after_first_patch
60429746 70979388 2291676 133700810
vmlinux_allyesconfig.after_second_patch
text data bss total filename
60449738 70975612 2288988 133714338
vmlinux_allyesconfig.before_this_series
60446534 70980348 2289596 133716478
vmlinux_allyesconfig.after_first_patch
60429746 70979900 2291676 133701322
vmlinux_allyesconfig.after_second_patch
But the error margin is within 1K.
So, in short, I inlined some functions which I shouldn't have. I am
preparing a v4 in which I will only inline the bit-find functions
(namely: __ffs(), ffs(), ffz(), __fls(), fls() and fls64()). Here are
the new figures:
text data bss total filename
60453552 70955485 2288620 133697657
vmlinux_allyesconfig.before_this_series
60450304 70953085 2289260 133692649
vmlinux_allyesconfig.after_first_patch
60433536 70952637 2291340 133677513
vmlinux_allyesconfig.after_second_patch
N.B. The new figures were after a rebase, so do not try to compare
with the previous benchmarks. I will send the v4 soon, after I finish
to update the patch comments and double check things.
Concerning the other functions in bitops.h, there may be some other
ones worth a __always_inline. But I will narrow the scope of this
series only to the bit-find function. If a good samaritan wants to
investigate the other functions, go ahead!
Yours sincerely,
Vincent Mailhol
allyesconfig:
text data bss total filename
58878600 72415994 2283652 133578246 vmlinux.before
58882250 72419706 2284004 133585960 vmlinux.after
atari_defconfig:
text data bss total filename
4112060 1579862 151680 5843602 vmlinux-v6.7-rc8
4117008 1579350 151680 5848038
vmlinux-v6.7-rc8-1-m68k-bitops-force-inlining
The next patch offsets that for allyesconfig, but not for atari_defconfig.
Reference: commit 8dd5032d9c54 ("x86/asm/bitops: Force inlining of
test_and_set_bit and friends")
Please don't split lines containing tags.
Link: https://git.kernel.org/torvalds/c/8dd5032d9c54
Signed-off-by: Vincent Mailhol <mailhol.vincent@xxxxxxxxxx>
Reviewed-by: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
