On Sun, 3 Sep 2023 17:50:59 +0800
Kuan-Wei Chiu <[email protected]> wrote:
> Replace the conditional statements in the cmp_stripe() function with a
> branchless version to improve code readability and potentially enhance
> performance.
The new code will always do two comparisons and a subtraction (3
instructions in total), whereas the old version could return after just 1
comparison, or after 2 comparisons. So depending on the data values it is 3x
to 1.5x as much operations performed than before, there unlikely to be any
enhancement of performance.
Also IMO the previous version is more easily readable.
> The new code calculates the result using a subtraction of
> comparison results, making it more concise and avoiding conditional
> branches. This change does not alter the functionality of the code.
>
> Signed-off-by: Kuan-Wei Chiu <[email protected]>
> ---
> drivers/md/raid5.c | 6 +-----
> 1 file changed, 1 insertion(+), 5 deletions(-)
>
> diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
> index 4cb9c608ee19..b14d7ba38f0f 100644
> --- a/drivers/md/raid5.c
> +++ b/drivers/md/raid5.c
> @@ -1035,11 +1035,7 @@ static int cmp_stripe(void *priv, const struct list_head *a,
> struct r5pending_data, sibling);
> const struct r5pending_data *db = list_entry(b,
> struct r5pending_data, sibling);
> - if (da->sector > db->sector)
> - return 1;
> - if (da->sector < db->sector)
> - return -1;
> - return 0;
> + return (da->sector > db->sector) - (da->sector < db->sector);
> }
>
> static void dispatch_defer_bios(struct r5conf *conf, int target,
--
With respect,
Roman
On Sun, Sep 03, 2023 at 06:30:58PM +0500, Roman Mamedov wrote:
> On Sun, 3 Sep 2023 17:50:59 +0800
> Kuan-Wei Chiu <[email protected]> wrote:
>
> > Replace the conditional statements in the cmp_stripe() function with a
> > branchless version to improve code readability and potentially enhance
> > performance.
>
> The new code will always do two comparisons and a subtraction (3
> instructions in total), whereas the old version could return after just 1
> comparison, or after 2 comparisons. So depending on the data values it is 3x
> to 1.5x as much operations performed than before, there unlikely to be any
> enhancement of performance.
>
> Also IMO the previous version is more easily readable.
>
The reason behind my proposed changes was to eliminate conditional
branches in the code. While the original code could occasionally achieve
early returns, many compilers, such as x86-64 gcc 13.2 compiling with
-O2 flag, still generate branch instructions. Processors typically have
deep pipelines, and a branch prediction miss can result in a high
penalty. Therefore, even though early return might not be possible, the
new branchless version of code could still offer efficiency
improvements.
> > The new code calculates the result using a subtraction of
> > comparison results, making it more concise and avoiding conditional
> > branches. This change does not alter the functionality of the code.
> >
> > Signed-off-by: Kuan-Wei Chiu <[email protected]>
> > ---
> > drivers/md/raid5.c | 6 +-----
> > 1 file changed, 1 insertion(+), 5 deletions(-)
> >
> > diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
> > index 4cb9c608ee19..b14d7ba38f0f 100644
> > --- a/drivers/md/raid5.c
> > +++ b/drivers/md/raid5.c
> > @@ -1035,11 +1035,7 @@ static int cmp_stripe(void *priv, const struct list_head *a,
> > struct r5pending_data, sibling);
> > const struct r5pending_data *db = list_entry(b,
> > struct r5pending_data, sibling);
> > - if (da->sector > db->sector)
> > - return 1;
> > - if (da->sector < db->sector)
> > - return -1;
> > - return 0;
> > + return (da->sector > db->sector) - (da->sector < db->sector);
> > }
> >
> > static void dispatch_defer_bios(struct r5conf *conf, int target,
>
>
> --
> With respect,
> Roman
--
Best regards,
Kuan-Wei Chiu