Re: [PATCH v8 02/10] x86/bhi: Make clear_bhb_loop() effective on newer CPUs

[Pawan Gupta]

On Wed, Apr 01, 2026 at 10:02:00AM +0100, David Laight wrote:
> > > As well as swapping %al <-> %ah try changing the outer loop decrement to
> > > 	sub $0x100, %ax
> > > since %al is zero that will set the z flag the same.  
> > 
> > Unfortunately, using "sub $0x100, %ax"(with %al as inner loop) isn't better
> > than just using "sub $1, %ah" in the outer loop:
> > 
> >   Event                     %al inner      + sub %ax       Delta
> >   ----------------------  -------------  -------------  ----------
> >   cycles                    776,775,020    813,372,036     +4.7%
> >   instructions/cycle               1.23           1.17     -4.5%
> >   branch-misses               4,792,502      7,610,323    +58.8%
> >   uops_issued.any           768,019,010    827,465,137     +7.7%
> >   time elapsed                 0.1627s        0.1707s      +4.9%
> 
> That is even more interesting.
> The 'sub %ax' version has more uops and more branch-misses.
> Looks like the extra cost of the %ah access is less than the cost
> of the extra mis-predicted branches.
> 
> Makes me wonder where a version that uses %cl fits?
> (Or use a zero-extending read and %eax/%ecx - likely to be the same.)
> I'll bet 'one beer' that is nearest the 'sub %ax' version.

%cl didn't make a noticeable difference, but ...

    Event                      %al/%ah        %al/%cl        Delta
                             (inner/outer)  (inner/outer)
    ----------------------  -------------  -------------  ----------
    cycles                    776,380,149    778,294,183     +0.2%
    instructions/cycle               1.23           1.22     -0.4%
    branch-misses               4,986,437      5,679,599    +13.9%
    uops_issued.any           773,223,387    765,724,878     -1.0%
    time elapsed                 0.1631s        0.1637s      +0.4%

... there are meaningful gains with 32-bit registers:

    Event                      %al/%ah        %eax/%ecx      Delta
                             (inner/outer)  (inner/outer)
    ----------------------  -------------  -------------  ----------
    cycles                    776,380,149    706,331,177     -9.0%
    instructions/cycle               1.23           1.35     +9.9%
    branch-misses               4,986,437      6,089,306    +22.1%
    uops_issued.any           773,223,387    774,539,522     +0.2%
    time elapsed                 0.1631s        0.1482s      -9.1%

These values are for userspace tests with immediates. Next, I will test how
they perform with memory loads in kernel. Before we finalize these uarch
nuances needs to be tested on a variety of CPUs.