In the origin implementation in function bbr_update_ack_aggregation(),
we utilize a lambda expression to flip the bit value of
bbr->extra_acked_win_idx. Since the data type of
bbr->extra_acked_win_idx is simply a single bit, we are actually trying
to perform a bit flip operation, under the fact we can simply perform a
bitwise not operation on bbr->extra_acked_win_idx.
This way we can elimate the need of possible branches which generate by
the lambda function, they could result in branch misses sometimes.
Perform a bitwise not operation is more straightforward and wouldn't
generate branches.
Signed-off-by: I Hsin Cheng <[email protected]>
---
net/ipv4/tcp_bbr.c | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
index 146792cd2..75068ba25 100644
--- a/net/ipv4/tcp_bbr.c
+++ b/net/ipv4/tcp_bbr.c
@@ -829,8 +829,7 @@ static void bbr_update_ack_aggregation(struct sock *sk,
bbr->extra_acked_win_rtts + 1);
if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
bbr->extra_acked_win_rtts = 0;
- bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
- 0 : 1;
+ bbr->extra_acked_win_idx = ~(bbr->extra_acked_win_idx);
bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
}
}
--
2.34.1
On Fri, Apr 26, 2024 at 5:20 PM I Hsin Cheng <[email protected]> wrote:
>
> In the origin implementation in function bbr_update_ack_aggregation(),
> we utilize a lambda expression to flip the bit value of
> bbr->extra_acked_win_idx. Since the data type of
> bbr->extra_acked_win_idx is simply a single bit, we are actually trying
> to perform a bit flip operation, under the fact we can simply perform a
> bitwise not operation on bbr->extra_acked_win_idx.
>
> This way we can elimate the need of possible branches which generate by
> the lambda function, they could result in branch misses sometimes.
> Perform a bitwise not operation is more straightforward and wouldn't
> generate branches.
>
> Signed-off-by: I Hsin Cheng <[email protected]>
> ---
> net/ipv4/tcp_bbr.c | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
> index 146792cd2..75068ba25 100644
> --- a/net/ipv4/tcp_bbr.c
> +++ b/net/ipv4/tcp_bbr.c
> @@ -829,8 +829,7 @@ static void bbr_update_ack_aggregation(struct sock *sk,
> bbr->extra_acked_win_rtts + 1);
> if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
> bbr->extra_acked_win_rtts = 0;
> - bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
> - 0 : 1;
> + bbr->extra_acked_win_idx = ~(bbr->extra_acked_win_idx);
> bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
> }
> }
Or
bbr->extra_acked_win_idx ^= 1;
Note that C compilers generate the same code, for the 3 variants.
They do not generate branches for something simple like this.
On Fri, Apr 26, 2024 at 05:32:57PM +0200, Eric Dumazet wrote:
> On Fri, Apr 26, 2024 at 5:20 PM I Hsin Cheng <[email protected]> wrote:
> >
> > In the origin implementation in function bbr_update_ack_aggregation(),
> > we utilize a lambda expression to flip the bit value of
> > bbr->extra_acked_win_idx. Since the data type of
> > bbr->extra_acked_win_idx is simply a single bit, we are actually trying
> > to perform a bit flip operation, under the fact we can simply perform a
> > bitwise not operation on bbr->extra_acked_win_idx.
> >
> > This way we can elimate the need of possible branches which generate by
> > the lambda function, they could result in branch misses sometimes.
> > Perform a bitwise not operation is more straightforward and wouldn't
> > generate branches.
> >
> > Signed-off-by: I Hsin Cheng <[email protected]>
> > ---
> > net/ipv4/tcp_bbr.c | 3 +--
> > 1 file changed, 1 insertion(+), 2 deletions(-)
> >
> > diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
> > index 146792cd2..75068ba25 100644
> > --- a/net/ipv4/tcp_bbr.c
> > +++ b/net/ipv4/tcp_bbr.c
> > @@ -829,8 +829,7 @@ static void bbr_update_ack_aggregation(struct sock *sk,
> > bbr->extra_acked_win_rtts + 1);
> > if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
> > bbr->extra_acked_win_rtts = 0;
> > - bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
> > - 0 : 1;
> > + bbr->extra_acked_win_idx = ~(bbr->extra_acked_win_idx);
> > bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
> > }
> > }
>
> Or
>
> bbr->extra_acked_win_idx ^= 1;
>
> Note that C compilers generate the same code, for the 3 variants.
>
> They do not generate branches for something simple like this.
I see, thanks for your explanation.
I thought the compilers behavior might alters due to different
architecture or different compilers.
So would you recommend on the proposed changes or we should stick to
the original implementation?
Personally I think my version or your proposed change are both more
understandable and elegant than the lambda expression.
On Sat, Apr 27, 2024 at 01:01:21AM +0800, I Hsin Cheng wrote:
> I see, thanks for your explanation.
> I thought the compilers behavior might alters due to different
> architecture or different compilers.
> So would you recommend on the proposed changes or we should stick to
> the original implementation?
> Personally I think my version or your proposed change are both more
> understandable and elegant than the lambda expression.
Out of curiosity, where do you see any lambda expressions in the entire
thing?
On Fri, Apr 26, 2024 at 09:19:02PM +0100, Al Viro wrote:
> On Sat, Apr 27, 2024 at 01:01:21AM +0800, I Hsin Cheng wrote:
>
> > I see, thanks for your explanation.
> > I thought the compilers behavior might alters due to different
> > architecture or different compilers.
> > So would you recommend on the proposed changes or we should stick to
> > the original implementation?
> > Personally I think my version or your proposed change are both more
> > understandable and elegant than the lambda expression.
>
> Out of curiosity, where do you see any lambda expressions in the entire
> thing?
Sorry, it's my fault to address the expression as "lambda expression",
it should be called as "conditional" or "ternary" operator.
Thanks for your remind.