[PATCH] fat: Refactor fat_tolower with branchless implementation

[PATCH] fat: Refactor fat_tolower with branchless implementation

[I Hsin Cheng]

Elimate the need of if-else branch within fat_tolower, replace it with a
branchless bitwise operation. This can reduce the number of branch ~130
regarding to the test script[1].

Test size can also be reduced:
$ ./scripts/bloat-o-meter vmlinux_old vmlinux_new
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-68 (-68)
Function                                     old     new   delta
fat_parse_short                             1901    1833     -68
Total: Before=22471023, After=22470955, chg -0.00%

Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
---
[1]:
static inline unsigned char old_tolower(unsigned char c)
{
    return ((c >= 'A') && (c <= 'Z')) ? c+32 : c;
}

static inline unsigned char new_tolower(unsigned char c)
{
    return c | 0x20;
}

int main(void) {
    for (unsigned char i = 0; i < 26; i++) {
        if (old_tolower('a' + i) != old_tolower('A' + i))
            return 1;
    }

    return 0;
}

Utilize perf to profile the difference when using old_tolower() and
new_tolower().

$ perf stat -e branches,branch-misses --repeat 100 ./old

 Performance counter stats for './old':

            2,6302      branches:u
              2334      branch-misses:u

       0.000754710 seconds time elapsed

       0.000000000 seconds user
       0.000804000 seconds sys

$ perf stat -e branches,branch-misses --repeat 100 ./new

 Performance counter stats for './main':

            2,6172      branches:u
              2338      branch-misses:u

       0.000782670 seconds time elapsed

       0.000853000 seconds user
       0.000000000 seconds sys

Best regards,
I Hsin Cheng
---
 fs/fat/dir.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/fat/dir.c b/fs/fat/dir.c
index acbec5bdd521..77d212b4d4db 100644
--- a/fs/fat/dir.c
+++ b/fs/fat/dir.c
@@ -35,7 +35,7 @@
 
 static inline unsigned char fat_tolower(unsigned char c)
 {
-	return ((c >= 'A') && (c <= 'Z')) ? c+32 : c;
+	return c | 0x20;
 }
 
 static inline loff_t fat_make_i_pos(struct super_block *sb,
-- 
2.43.0

Re: [PATCH] fat: Refactor fat_tolower with branchless implementation

[OGAWA Hirofumi]

I Hsin Cheng <richard120310@gmail.com> writes:

> Elimate the need of if-else branch within fat_tolower, replace it with a
> branchless bitwise operation. This can reduce the number of branch ~130
> regarding to the test script[1].
>
> Test size can also be reduced:
> $ ./scripts/bloat-o-meter vmlinux_old vmlinux_new
> add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-68 (-68)
> Function                                     old     new   delta
> fat_parse_short                             1901    1833     -68
> Total: Before=22471023, After=22470955, chg -0.00%
>
> Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
> ---
> [1]:
> static inline unsigned char old_tolower(unsigned char c)
> {
>     return ((c >= 'A') && (c <= 'Z')) ? c+32 : c;
> }
>
> static inline unsigned char new_tolower(unsigned char c)
> {
>     return c | 0x20;
> }

Looks like doesn't work correctly. For example, new one changes TAB to ')'.

Thanks.

> int main(void) {
>     for (unsigned char i = 0; i < 26; i++) {
>         if (old_tolower('a' + i) != old_tolower('A' + i))
>             return 1;
>     }
>
>     return 0;
> }
>
> Utilize perf to profile the difference when using old_tolower() and
> new_tolower().
>
> $ perf stat -e branches,branch-misses --repeat 100 ./old
>
>  Performance counter stats for './old':
>
>             2,6302      branches:u
>               2334      branch-misses:u
>
>        0.000754710 seconds time elapsed
>
>        0.000000000 seconds user
>        0.000804000 seconds sys
>
> $ perf stat -e branches,branch-misses --repeat 100 ./new
>
>  Performance counter stats for './main':
>
>             2,6172      branches:u
>               2338      branch-misses:u
>
>        0.000782670 seconds time elapsed
>
>        0.000853000 seconds user
>        0.000000000 seconds sys
>
> Best regards,
> I Hsin Cheng
> ---
>  fs/fat/dir.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/fs/fat/dir.c b/fs/fat/dir.c
> index acbec5bdd521..77d212b4d4db 100644
> --- a/fs/fat/dir.c
> +++ b/fs/fat/dir.c
> @@ -35,7 +35,7 @@
>  
>  static inline unsigned char fat_tolower(unsigned char c)
>  {
> -	return ((c >= 'A') && (c <= 'Z')) ? c+32 : c;
> +	return c | 0x20;
>  }
>  
>  static inline loff_t fat_make_i_pos(struct super_block *sb,

-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] fat: Refactor fat_tolower with branchless implementation

[Al Viro]

On Tue, Mar 18, 2025 at 11:43:09AM +0800, I Hsin Cheng wrote:
> Elimate the need of if-else branch within fat_tolower, replace it with a
> branchless bitwise operation. This can reduce the number of branch ~130
> regarding to the test script[1].

... and for values outside of 32..126 it would give wrong results.

Re: [PATCH] fat: Refactor fat_tolower with branchless implementation

[I Hsin Cheng]

On Tue, Mar 18, 2025 at 04:03:23AM +0000, Al Viro wrote:
> On Tue, Mar 18, 2025 at 11:43:09AM +0800, I Hsin Cheng wrote:
> > Elimate the need of if-else branch within fat_tolower, replace it with a
> > branchless bitwise operation. This can reduce the number of branch ~130
> > regarding to the test script[1].
> 
> ... and for values outside of 32..126 it would give wrong results.

Ahh, sorry I didn't take these into consideration, I guess branch is
necessary here then. Thanks for your review!

Best regards,
I Hsin Cheng