V2 [PATCH] bench-strcmp.c: Add workloads on page boundary

[H.J. Lu]

On Wed, Sep 23, 2020 at 5:52 PM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 6/12/20 4:10 PM, H.J. Lu via Libc-alpha wrote:
> > Add strcmp workloads on page boundary.
> > ---
> >  benchtests/bench-strcmp.c | 48 +++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 48 insertions(+)
> >
>
> These additions to the benchmark exercise the page boundary conditions
> in the algorithms implemented by all the architectures. The test is
> designed from the experience with bug 25933 in mind so we should mention
> that in a comment. It could be made a bit more generic by parametrizing
> vector sizes and number of vector registers that might be grouped e.g.
> 32 * 4, but it's fine as it is for now to test the page boundary.
>
> OK with:
> - Function rename
> - Added comments.
> - s1a iterate over [30,21)
>
> Reviewed-by: Carlos O'Donell <carlos@redhat.com>
>
> > diff --git a/benchtests/bench-strcmp.c b/benchtests/bench-strcmp.c
> > index 47d0a35299..3ba1399a4d 100644
> > --- a/benchtests/bench-strcmp.c
> > +++ b/benchtests/bench-strcmp.c
> > @@ -144,6 +144,52 @@ do_test (json_ctx_t *json_ctx, size_t align1, size_t align2, size_t len, int
> >    json_element_object_end (json_ctx);
> >  }
> >
> > +static void
> > +do_test_page_boundary_1 (json_ctx_t *json_ctx, CHAR *s1, CHAR *s2,
>
> Rename to do_one_test_page_boundary ()

Done.

> > +                      size_t align1, size_t align2, size_t len,
> > +                      int exp_result)
> > +{
> > +  json_element_object_begin (json_ctx);
> > +  json_attr_uint (json_ctx, "length", (double) len);
> > +  json_attr_uint (json_ctx, "align1", (double) align1);
> > +  json_attr_uint (json_ctx, "align2", (double) align2);
> > +  json_array_begin (json_ctx, "timings");
> > +  FOR_EACH_IMPL (impl, 0)
> > +    do_one_test (json_ctx, impl, s1, s2, exp_result);
>
> OK.
>
> > +  json_array_end (json_ctx);
> > +  json_element_object_end (json_ctx);
> > +}
> > +
>
> Add a comment:
>
> /* To trigger bug 25933 we need a size that is equal to the
>    vector length times 4. In the case of AVX2 for Intel we
>    need 32 * 4. We make this test generic and run it for all
>    architectures as additional boundary testing for such
>    related algorithms.  */

Done.

> > +static void
> > +do_test_page_boundary (json_ctx_t *json_ctx)
> > +{
> > +  size_t size = 32 * 4;
> > +  size_t len;
> > +  CHAR *s1 = (CHAR *) (buf1 + (BUF1PAGES - 1) * page_size);
> > +  CHAR *s2 = (CHAR *) (buf2 + (BUF1PAGES - 1) * page_size);
> > +  int exp_result;
> > +
> > +  memset (s1, 'a', page_size);
> > +  memset (s2, 'a', page_size);
> > +
> > +  s1[(page_size / CHARBYTES) - 1] = (CHAR) 0;
> > +  s2[(page_size / CHARBYTES) - 1] = (CHAR) 0;
> > +
>
>
> Add comment:
>
> /* Iterate over a size that is just below where we expect
>    the bug to trigger up to the size we expect will trigger
>    the bug e.g. [99-128].  Likewise iterate the start of
>    two strings between 30 and 31 bytes away from the
>    boundary to simulate alignment changes.  */

Done.

> > +  for (size_t s = 99; s <= size; s++)
> > +    for (size_t s1a = 31; s1a < 32; s1a++)
>
> Make s1a iterate over [30,32) like s2a.

Done.

> > +      for (size_t s2a = 30; s2a < 32; s2a++)
> > +     {
> > +       size_t align1 = (page_size / CHARBYTES - s) - s1a;
> > +       size_t align2 = (page_size / CHARBYTES - s) - s2a;
> > +       CHAR *s1p = s1 + align1;
> > +       CHAR *s2p = s2 + align2;
> > +       len = (page_size / CHARBYTES) - 1 - align1;
> > +       exp_result = SIMPLE_STRCMP (s1p, s2p);
> > +       do_test_page_boundary_1 (json_ctx, s1p, s2p, align1, align2,
>
> Call do_one_test_page_boundary ()

Done.

> > +                                len, exp_result);
> > +     }
> > +}
> > +
> >  int
> >  test_main (void)
> >  {
> > @@ -197,6 +243,8 @@ test_main (void)
> >        do_test (&json_ctx, 2 * CHARBYTES * i, CHARBYTES * i, 8 << i, LARGECHAR, -1);
> >      }
> >
> > +  do_test_page_boundary (&json_ctx);
>
> OK.
>
> > +
> >    json_array_end (&json_ctx);
> >    json_attr_object_end (&json_ctx);
> >    json_attr_object_end (&json_ctx);
> >
>
>

Here is the updated patch.  OK for master?

Thanks.

-- 
H.J.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-bench-strcmp.c-Add-workloads-on-page-boundary.patch
Type: text/x-patch
Size: 3013 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20200924/13c96fed/attachment.bin>