[PATCH v3 1/3] x86/string: Factor out large memmove implemention to seperate file

Sunil Pandey skpgkp2@gmail.com
Fri Nov 28 15:23:15 GMT 2025


On Thu, Nov 27, 2025 at 10:56 PM Noah Goldstein <goldstein.w.n@gmail.com>
wrote:

>
>
> On Fri, Nov 28, 2025, 08:56 Sunil Pandey <skpgkp2@gmail.com> wrote:
>
>>
>>
>>
>> On Thu, Nov 27, 2025 at 12:47 AM Noah Goldstein <goldstein.w.n@gmail.com>
>> wrote:
>>
>>> On Wed, Nov 26, 2025 at 8:53 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>> >
>>> > On Thu, Nov 27, 2025 at 9:19 AM Sunil Pandey <skpgkp2@gmail.com>
>>> wrote:
>>> > >
>>> > > Hi Noah,
>>> > >
>>> > > Looking into bugzilla
>>> https://sourceware.org/bugzilla/show_bug.cgi?id=32475
>>> > >
>>> > > Can you please modify patch to just address the issue listed in the
>>> bug, without any change on Intel
>>> > > processors. As far as I know, we don't have any perf issue on Intel
>>> side and patch is regressing on
>>> > > newer processors. On SPR this patch has no performance impact.
>>>
>>> Okay, added a 4th patch to the series that makes this default for all
>>> intel bigcore.
>>>
>>
>> How about not making any changes for Intel processors?
>>
>
> Do we have any evidence the page unrolled impl is good for atom?
>
>
So far, I have not found any perf benefit from this patch set on current
intel processors.


>
>> I still advise that you robustly test the page-unrolled vs standard
>>> implementation
>>> as it only seems like a matter of time until some hardware
>>> optimization makes the
>>> standard approach optimal.
>>>
>>> Thank you for benchmarking, glad to see that after SPR it still is
>>> yielding benefits :)
>>>
>>> >
>>> > SPR is old.  Please try it on the current Intel processors.
>>>
>>> I think he ran SPEC on some newer processors.
>>
>>
>>> >
>>> > > --Sunil
>>> > >
>>> > > On Tue, Nov 25, 2025 at 6:42 PM Noah Goldstein <
>>> goldstein.w.n@gmail.com> wrote:
>>> > >>
>>> > >> On Tue, Nov 25, 2025 at 8:44 PM Noah Goldstein <
>>> goldstein.w.n@gmail.com> wrote:
>>> > >> >
>>> > >> > On Tue, Nov 25, 2025 at 10:37 AM Sunil Pandey <skpgkp2@gmail.com>
>>> wrote:
>>> > >> > >
>>> > >> > >
>>> > >> > >
>>> > >> > > On Thu, Nov 20, 2025 at 3:14 PM H.J. Lu <hjl.tools@gmail.com>
>>> wrote:
>>> > >> > >>
>>> > >> > >> On Fri, Nov 21, 2025 at 1:41 AM Sunil Pandey <
>>> skpgkp2@gmail.com> wrote:
>>> > >> > >> >
>>> > >> > >> >
>>> > >> > >> >
>>> > >> > >> > On Sat, Nov 15, 2025 at 1:45 AM Noah Goldstein <
>>> goldstein.w.n@gmail.com> wrote:
>>> > >> > >> >>
>>> > >> > >> >> On Sat, Nov 15, 2025 at 4:41 AM Noah Goldstein <
>>> goldstein.w.n@gmail.com> wrote:
>>> > >> > >> >> >
>>> > >> > >> >> > On Sat, Nov 15, 2025 at 4:33 AM Noah Goldstein <
>>> goldstein.w.n@gmail.com> wrote:
>>> > >> > >> >> > >
>>> > >> > >> >> > > This is to enable us to support multiple large (size
>>> greater than
>>> > >> > >> >> > > non-temporal threshold) implementations.
>>> > >> > >> >> > >
>>> > >> > >> >> > > This patch has no affect on the resulting libc.so
>>> library.
>>> > >> > >> >> > > ---
>>> > >> > >> >> > >  .../memmove-vec-large-page-unrolled.S         | 272
>>> ++++++++++++++++++
>>> > >> > >> >> > >  .../multiarch/memmove-vec-unaligned-erms.S    | 272
>>> +-----------------
>>> > >> > >> >> > >  2 files changed, 279 insertions(+), 265 deletions(-)
>>> > >> > >> >> > >  create mode 100644
>>> sysdeps/x86_64/multiarch/memmove-vec-large-page-unrolled.S
>>> > >> > >> >> > >
>>> > >> > >> >> > > diff --git
>>> a/sysdeps/x86_64/multiarch/memmove-vec-large-page-unrolled.S
>>> b/sysdeps/x86_64/multiarch/memmove-vec-large-page-unrolled.S
>>> > >> > >> >> > > new file mode 100644
>>> > >> > >> >> > > index 0000000000..b5b4593279
>>> > >> > >> >> > > --- /dev/null
>>> > >> > >> >> > > +++
>>> b/sysdeps/x86_64/multiarch/memmove-vec-large-page-unrolled.S
>>> > >> > >> >> > > @@ -0,0 +1,272 @@
>>> > >> > >> >> > > +#ifdef MEMMOVE_LARGE_IMPL
>>> > >> > >> >> > > +# error "Multiple large memmove impls included!"
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +#define MEMMOVE_LARGE_IMPL     1
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +/* Copies large regions by copying multiple pages at
>>> once.  This is
>>> > >> > >> >> > > +       beneficial on some older Intel hardware
>>> (Broadwell, Skylake, and
>>> > >> > >> >> > > +       Icelake).
>>> > >> > >> >> > > +   1. If size < 16 *
>>> __x86_shared_non_temporal_threshold and
>>> > >> > >> >> > > +      source and destination do not page alias, copy
>>> from 2 pages
>>> > >> > >> >> > > +      at once using non-temporal stores. Page aliasing
>>> in this case is
>>> > >> > >> >> > > +      considered true if destination's page alignment
>>> - sources' page
>>> > >> > >> >> > > +      alignment is less than 8 * VEC_SIZE.
>>> > >> > >> >> > > +   2. If size >= 16 *
>>> __x86_shared_non_temporal_threshold or source
>>> > >> > >> >> > > +      and destination do page alias copy from 4 pages
>>> at once using
>>> > >> > >> >> > > +      non-temporal stores.  */
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +#ifndef LOG_PAGE_SIZE
>>> > >> > >> >> > > +# define LOG_PAGE_SIZE 12
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +#if PAGE_SIZE != (1 << LOG_PAGE_SIZE)
>>> > >> > >> >> > > +# error Invalid LOG_PAGE_SIZE
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +/* Byte per page for large_memcpy inner loop.  */
>>> > >> > >> >> > > +#if VEC_SIZE == 64
>>> > >> > >> >> > > +# define LARGE_LOAD_SIZE       (VEC_SIZE * 2)
>>> > >> > >> >> > > +#else
>>> > >> > >> >> > > +# define LARGE_LOAD_SIZE       (VEC_SIZE * 4)
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +/* Amount to shift __x86_shared_non_temporal_threshold
>>> by for
>>> > >> > >> >> > > +   bound for memcpy_large_4x. This is essentially use
>>> to to
>>> > >> > >> >> > > +   indicate that the copy is far beyond the scope of L3
>>> > >> > >> >> > > +   (assuming no user config
>>> x86_non_temporal_threshold) and to
>>> > >> > >> >> > > +   use a more aggressively unrolled loop.  NB: before
>>> > >> > >> >> > > +   increasing the value also update initialization of
>>> > >> > >> >> > > +   x86_non_temporal_threshold.  */
>>> > >> > >> >> > > +#ifndef LOG_4X_MEMCPY_THRESH
>>> > >> > >> >> > > +# define LOG_4X_MEMCPY_THRESH  4
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +#if LARGE_LOAD_SIZE == (VEC_SIZE * 2)
>>> > >> > >> >> > > +# define LOAD_ONE_SET(base, offset, vec0, vec1, ...)
>>>  \
>>> > >> > >> >> > > +       VMOVU   (offset)base, vec0;     \
>>> > >> > >> >> > > +       VMOVU   ((offset) + VEC_SIZE)base, vec1;
>>> > >> > >> >> > > +# define STORE_ONE_SET(base, offset, vec0, vec1, ...)
>>> \
>>> > >> > >> >> > > +       VMOVNT  vec0, (offset)base;     \
>>> > >> > >> >> > > +       VMOVNT  vec1, ((offset) + VEC_SIZE)base;
>>> > >> > >> >> > > +#elif LARGE_LOAD_SIZE == (VEC_SIZE * 4)
>>> > >> > >> >> > > +# define LOAD_ONE_SET(base, offset, vec0, vec1, vec2,
>>> vec3)    \
>>> > >> > >> >> > > +       VMOVU   (offset)base, vec0;     \
>>> > >> > >> >> > > +       VMOVU   ((offset) + VEC_SIZE)base, vec1;
>>> \
>>> > >> > >> >> > > +       VMOVU   ((offset) + VEC_SIZE * 2)base, vec2;
>>> \
>>> > >> > >> >> > > +       VMOVU   ((offset) + VEC_SIZE * 3)base, vec3;
>>> > >> > >> >> > > +# define STORE_ONE_SET(base, offset, vec0, vec1, vec2,
>>> vec3)   \
>>> > >> > >> >> > > +       VMOVNT  vec0, (offset)base;     \
>>> > >> > >> >> > > +       VMOVNT  vec1, ((offset) + VEC_SIZE)base;
>>> \
>>> > >> > >> >> > > +       VMOVNT  vec2, ((offset) + VEC_SIZE * 2)base;
>>> \
>>> > >> > >> >> > > +       VMOVNT  vec3, ((offset) + VEC_SIZE * 3)base;
>>> > >> > >> >> > > +#else
>>> > >> > >> >> > > +# error Invalid LARGE_LOAD_SIZE
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       .p2align 4,, 10
>>> > >> > >> >> > > +#if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN
>>> (libc)
>>> > >> > >> >> > > +L(large_memcpy_check):
>>> > >> > >> >> > > +       /* Entry from L(large_memcpy_2x) has a
>>> redundant load of
>>> > >> > >> >> > > +          __x86_shared_non_temporal_threshold(%rip).
>>> L(large_memcpy_2x)
>>> > >> > >> >> > > +          is only use for the non-erms memmove which
>>> is generally less
>>> > >> > >> >> > > +          common.  */
>>> > >> > >> >> > > +L(large_memcpy):
>>> > >> > >> >> > > +       mov
>>>  __x86_shared_non_temporal_threshold(%rip), %R11_LP
>>> > >> > >> >> > > +       cmp     %R11_LP, %RDX_LP
>>> > >> > >> >> > > +       jb      L(more_8x_vec_check)
>>> > >> > >> >> > > +       /* To reach this point it is impossible for dst
>>> > src and
>>> > >> > >> >> > > +          overlap. Remaining to check is src > dst and
>>> overlap. rcx
>>> > >> > >> >> > > +          already contains dst - src. Negate rcx to
>>> get src - dst. If
>>> > >> > >> >> > > +          length > rcx then there is overlap and
>>> forward copy is best.  */
>>> > >> > >> >> > > +       negq    %rcx
>>> > >> > >> >> > > +       cmpq    %rcx, %rdx
>>> > >> > >> >> > > +       ja      L(more_8x_vec_forward)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Cache align destination. First store the
>>> first 64 bytes then
>>> > >> > >> >> > > +          adjust alignments.  */
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* First vec was also loaded into VEC(0).  */
>>> > >> > >> >> > > +# if VEC_SIZE < 64
>>> > >> > >> >> > > +       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > +#  if VEC_SIZE < 32
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > +#  endif
>>> > >> > >> >> > > +# endif
>>> > >> > >> >> > > +       VMOVU   %VMM(0), (%rdi)
>>> > >> > >> >> > > +# if VEC_SIZE < 64
>>> > >> > >> >> > > +       VMOVU   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > +#  if VEC_SIZE < 32
>>> > >> > >> >> > > +       VMOVU   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > +       VMOVU   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > +#  endif
>>> > >> > >> >> > > +# endif
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Adjust source, destination, and size.  */
>>> > >> > >> >> > > +       movq    %rdi, %r8
>>> > >> > >> >> > > +       andq    $63, %r8
>>> > >> > >> >> > > +       /* Get the negative of offset for alignment.  */
>>> > >> > >> >> > > +       subq    $64, %r8
>>> > >> > >> >> > > +       /* Adjust source.  */
>>> > >> > >> >> > > +       subq    %r8, %rsi
>>> > >> > >> >> > > +       /* Adjust destination which should be aligned
>>> now.  */
>>> > >> > >> >> > > +       subq    %r8, %rdi
>>> > >> > >> >> > > +       /* Adjust length.  */
>>> > >> > >> >> > > +       addq    %r8, %rdx
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Test if source and destination addresses
>>> will alias. If they
>>> > >> > >> >> > > +          do the larger pipeline in large_memcpy_4x
>>> alleviated the
>>> > >> > >> >> > > +          performance drop.  */
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* ecx contains -(dst - src). not ecx will
>>> return dst - src - 1
>>> > >> > >> >> > > +          which works for testing aliasing.  */
>>> > >> > >> >> > > +       notl    %ecx
>>> > >> > >> >> > > +       movq    %rdx, %r10
>>> > >> > >> >> > > +       testl   $(PAGE_SIZE - VEC_SIZE * 8), %ecx
>>> > >> > >> >> > > +       jz      L(large_memcpy_4x)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* r11 has
>>> __x86_shared_non_temporal_threshold.  Shift it left
>>> > >> > >> >> > > +          by LOG_4X_MEMCPY_THRESH to get
>>> L(large_memcpy_4x) threshold.  */
>>> > >> > >> >> > > +       shlq    $LOG_4X_MEMCPY_THRESH, %r11
>>> > >> > >> >> > > +       cmp     %r11, %rdx
>>> > >> > >> >> > > +       jae     L(large_memcpy_4x)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* edx will store remainder size for copying
>>> tail.  */
>>> > >> > >> >> > > +       andl    $(PAGE_SIZE * 2 - 1), %edx
>>> > >> > >> >> > > +       /* r10 stores outer loop counter.  */
>>> > >> > >> >> > > +       shrq    $(LOG_PAGE_SIZE + 1), %r10
>>> > >> > >> >> > > +       /* Copy 4x VEC at a time from 2 pages.  */
>>> > >> > >> >> > > +       .p2align 4
>>> > >> > >> >> > > +L(loop_large_memcpy_2x_outer):
>>> > >> > >> >> > > +       /* ecx stores inner loop counter.  */
>>> > >> > >> >> > > +       movl    $(PAGE_SIZE / LARGE_LOAD_SIZE), %ecx
>>> > >> > >> >> > > +L(loop_large_memcpy_2x_inner):
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE * 2)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE * 2)
>>> > >> > >> >> > > +       /* Load vectors from rsi.  */
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > +       subq    $-LARGE_LOAD_SIZE, %rsi
>>> > >> > >> >> > > +       /* Non-temporal store vectors to rdi.  */
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > +       subq    $-LARGE_LOAD_SIZE, %rdi
>>> > >> > >> >> > > +       decl    %ecx
>>> > >> > >> >> > > +       jnz     L(loop_large_memcpy_2x_inner)
>>> > >> > >> >> > > +       addq    $PAGE_SIZE, %rdi
>>> > >> > >> >> > > +       addq    $PAGE_SIZE, %rsi
>>> > >> > >> >> > > +       decq    %r10
>>> > >> > >> >> > > +       jne     L(loop_large_memcpy_2x_outer)
>>> > >> > >> >> > > +       sfence
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Check if only last 4 loads are needed.  */
>>> > >> > >> >> > > +       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       jbe     L(large_memcpy_2x_end)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Handle the last 2 * PAGE_SIZE bytes.  */
>>> > >> > >> >> > > +L(loop_large_memcpy_2x_tail):
>>> > >> > >> >> > > +       /* Copy 4 * VEC a time forward with
>>> non-temporal stores.  */
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rdi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       VMOVU   (%rsi), %VMM(0)
>>> > >> > >> >> > > +       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > +       subq    $-(VEC_SIZE * 4), %rsi
>>> > >> > >> >> > > +       addl    $-(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       VMOVA   %VMM(0), (%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > +       subq    $-(VEC_SIZE * 4), %rdi
>>> > >> > >> >> > > +       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       ja      L(loop_large_memcpy_2x_tail)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +L(large_memcpy_2x_end):
>>> > >> > >> >> > > +       /* Store the last 4 * VEC.  */
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VMM(0)
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VMM(1)
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VMM(2)
>>> > >> > >> >> > > +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VMM(3)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       VMOVU   %VMM(0), -(VEC_SIZE * 4)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(1), -(VEC_SIZE * 3)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(2), -(VEC_SIZE * 2)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(3), -VEC_SIZE(%rdi, %rdx)
>>> > >> > >> >> > > +       VZEROUPPER_RETURN
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       .p2align 4
>>> > >> > >> >> > > +L(large_memcpy_4x):
>>> > >> > >> >> > > +       /* edx will store remainder size for copying
>>> tail.  */
>>> > >> > >> >> > > +       andl    $(PAGE_SIZE * 4 - 1), %edx
>>> > >> > >> >> > > +       /* r10 stores outer loop counter.  */
>>> > >> > >> >> > > +       shrq    $(LOG_PAGE_SIZE + 2), %r10
>>> > >> > >> >> > > +       /* Copy 4x VEC at a time from 4 pages.  */
>>> > >> > >> >> > > +       .p2align 4
>>> > >> > >> >> > > +L(loop_large_memcpy_4x_outer):
>>> > >> > >> >> > > +       /* ecx stores inner loop counter.  */
>>> > >> > >> >> > > +       movl    $(PAGE_SIZE / LARGE_LOAD_SIZE), %ecx
>>> > >> > >> >> > > +L(loop_large_memcpy_4x_inner):
>>> > >> > >> >> > > +       /* Only one prefetch set per page as doing 4
>>> pages give more
>>> > >> > >> >> > > +          time for prefetcher to keep up.  */
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi), PAGE_SIZE * 2 +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi), PAGE_SIZE * 3 +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       /* Load vectors from rsi.  */
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), PAGE_SIZE * 2, %VMM(8),
>>> %VMM(9), %VMM(10), %VMM(11))
>>> > >> > >> >> > > +       LOAD_ONE_SET ((%rsi), PAGE_SIZE * 3, %VMM(12),
>>> %VMM(13), %VMM(14), %VMM(15))
>>> > >> > >> >> > > +       subq    $-LARGE_LOAD_SIZE, %rsi
>>> > >> > >> >> > > +       /* Non-temporal store vectors to rdi.  */
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), PAGE_SIZE * 2, %VMM(8),
>>> %VMM(9), %VMM(10), %VMM(11))
>>> > >> > >> >> > > +       STORE_ONE_SET ((%rdi), PAGE_SIZE * 3, %VMM(12),
>>> %VMM(13), %VMM(14), %VMM(15))
>>> > >> > >> >> > > +       subq    $-LARGE_LOAD_SIZE, %rdi
>>> > >> > >> >> > > +       decl    %ecx
>>> > >> > >> >> > > +       jnz     L(loop_large_memcpy_4x_inner)
>>> > >> > >> >> > > +       addq    $(PAGE_SIZE * 3), %rdi
>>> > >> > >> >> > > +       addq    $(PAGE_SIZE * 3), %rsi
>>> > >> > >> >> > > +       decq    %r10
>>> > >> > >> >> > > +       jne     L(loop_large_memcpy_4x_outer)
>>> > >> > >> >> > > +       sfence
>>> > >> > >> >> > > +       /* Check if only last 4 loads are needed.  */
>>> > >> > >> >> > > +       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       jbe     L(large_memcpy_4x_end)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       /* Handle the last 4  * PAGE_SIZE bytes.  */
>>> > >> > >> >> > > +L(loop_large_memcpy_4x_tail):
>>> > >> > >> >> > > +       /* Copy 4 * VEC a time forward with
>>> non-temporal stores.  */
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       PREFETCH_ONE_SET (1, (%rdi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > +       VMOVU   (%rsi), %VMM(0)
>>> > >> > >> >> > > +       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > +       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > +       subq    $-(VEC_SIZE * 4), %rsi
>>> > >> > >> >> > > +       addl    $-(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       VMOVA   %VMM(0), (%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > +       VMOVA   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > +       subq    $-(VEC_SIZE * 4), %rdi
>>> > >> > >> >> > > +       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > +       ja      L(loop_large_memcpy_4x_tail)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +L(large_memcpy_4x_end):
>>> > >> > >> >> > > +       /* Store the last 4 * VEC.  */
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VMM(0)
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VMM(1)
>>> > >> > >> >> > > +       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VMM(2)
>>> > >> > >> >> > > +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VMM(3)
>>> > >> > >> >> > > +
>>> > >> > >> >> > > +       VMOVU   %VMM(0), -(VEC_SIZE * 4)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(1), -(VEC_SIZE * 3)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(2), -(VEC_SIZE * 2)(%rdi, %rdx)
>>> > >> > >> >> > > +       VMOVU   %VMM(3), -VEC_SIZE(%rdi, %rdx)
>>> > >> > >> >> > > +       VZEROUPPER_RETURN
>>> > >> > >> >> > > +#endif
>>> > >> > >> >> > > diff --git
>>> a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
>>> b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
>>> > >> > >> >> > > index 5cd8a6286e..70d303687c 100644
>>> > >> > >> >> > > ---
>>> a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
>>> > >> > >> >> > > +++
>>> b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
>>> > >> > >> >> > > @@ -34,17 +34,8 @@
>>> > >> > >> >> > >        __x86_rep_movsb_threshold and less than
>>> > >> > >> >> > >        __x86_rep_movsb_stop_threshold, then REP MOVSB
>>> will be used.
>>> > >> > >> >> > >     7. If size >= __x86_shared_non_temporal_threshold
>>> and there is no
>>> > >> > >> >> > > -      overlap between destination and source, use
>>> non-temporal store
>>> > >> > >> >> > > -      instead of aligned store copying from either 2
>>> or 4 pages at
>>> > >> > >> >> > > -      once.
>>> > >> > >> >> > > -   8. For point 7) if size < 16 *
>>> __x86_shared_non_temporal_threshold
>>> > >> > >> >> > > -      and source and destination do not page alias,
>>> copy from 2 pages
>>> > >> > >> >> > > -      at once using non-temporal stores. Page aliasing
>>> in this case is
>>> > >> > >> >> > > -      considered true if destination's page alignment
>>> - sources' page
>>> > >> > >> >> > > -      alignment is less than 8 * VEC_SIZE.
>>> > >> > >> >> > > -   9. If size >= 16 *
>>> __x86_shared_non_temporal_threshold or source
>>> > >> > >> >> > > -      and destination do page alias copy from 4 pages
>>> at once using
>>> > >> > >> >> > > -      non-temporal stores.  */
>>> > >> > >> >> > > +      overlap between destination and source, the
>>> exact method varies
>>> > >> > >> >> > > +      and is set with MEMMOVE_VEC_LARGE_IMPL".  */
>>> > >> > >> >> > >
>>> > >> > >> >> > >  #include <sysdep.h>
>>> > >> > >> >> > >
>>> > >> > >> >> > > @@ -95,31 +86,6 @@
>>> > >> > >> >> > >  # error Unsupported PAGE_SIZE
>>> > >> > >> >> > >  #endif
>>> > >> > >> >> > >
>>> > >> > >> >> > > -#ifndef LOG_PAGE_SIZE
>>> > >> > >> >> > > -# define LOG_PAGE_SIZE 12
>>> > >> > >> >> > > -#endif
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -#if PAGE_SIZE != (1 << LOG_PAGE_SIZE)
>>> > >> > >> >> > > -# error Invalid LOG_PAGE_SIZE
>>> > >> > >> >> > > -#endif
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -/* Byte per page for large_memcpy inner loop.  */
>>> > >> > >> >> > > -#if VEC_SIZE == 64
>>> > >> > >> >> > > -# define LARGE_LOAD_SIZE (VEC_SIZE * 2)
>>> > >> > >> >> > > -#else
>>> > >> > >> >> > > -# define LARGE_LOAD_SIZE (VEC_SIZE * 4)
>>> > >> > >> >> > > -#endif
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -/* Amount to shift __x86_shared_non_temporal_threshold
>>> by for
>>> > >> > >> >> > > -   bound for memcpy_large_4x. This is essentially use
>>> to to
>>> > >> > >> >> > > -   indicate that the copy is far beyond the scope of L3
>>> > >> > >> >> > > -   (assuming no user config
>>> x86_non_temporal_threshold) and to
>>> > >> > >> >> > > -   use a more aggressively unrolled loop.  NB: before
>>> > >> > >> >> > > -   increasing the value also update initialization of
>>> > >> > >> >> > > -   x86_non_temporal_threshold.  */
>>> > >> > >> >> > > -#ifndef LOG_4X_MEMCPY_THRESH
>>> > >> > >> >> > > -# define LOG_4X_MEMCPY_THRESH 4
>>> > >> > >> >> > > -#endif
>>> > >> > >> >> > >
>>> > >> > >> >> > >  /* Avoid short distance rep movsb only with non-SSE
>>> vector.  */
>>> > >> > >> >> > >  #ifndef AVOID_SHORT_DISTANCE_REP_MOVSB
>>> > >> > >> >> > > @@ -160,26 +126,8 @@
>>> > >> > >> >> > >  # error Unsupported PREFETCH_SIZE!
>>> > >> > >> >> > >  #endif
>>> > >> > >> >> > >
>>> > >> > >> >> > > -#if LARGE_LOAD_SIZE == (VEC_SIZE * 2)
>>> > >> > >> >> > > -# define LOAD_ONE_SET(base, offset, vec0, vec1, ...) \
>>> > >> > >> >> > > -       VMOVU   (offset)base, vec0; \
>>> > >> > >> >> > > -       VMOVU   ((offset) + VEC_SIZE)base, vec1;
>>> > >> > >> >> > > -# define STORE_ONE_SET(base, offset, vec0, vec1, ...) \
>>> > >> > >> >> > > -       VMOVNT  vec0, (offset)base; \
>>> > >> > >> >> > > -       VMOVNT  vec1, ((offset) + VEC_SIZE)base;
>>> > >> > >> >> > > -#elif LARGE_LOAD_SIZE == (VEC_SIZE * 4)
>>> > >> > >> >> > > -# define LOAD_ONE_SET(base, offset, vec0, vec1, vec2,
>>> vec3) \
>>> > >> > >> >> > > -       VMOVU   (offset)base, vec0; \
>>> > >> > >> >> > > -       VMOVU   ((offset) + VEC_SIZE)base, vec1; \
>>> > >> > >> >> > > -       VMOVU   ((offset) + VEC_SIZE * 2)base, vec2; \
>>> > >> > >> >> > > -       VMOVU   ((offset) + VEC_SIZE * 3)base, vec3;
>>> > >> > >> >> > > -# define STORE_ONE_SET(base, offset, vec0, vec1, vec2,
>>> vec3) \
>>> > >> > >> >> > > -       VMOVNT  vec0, (offset)base; \
>>> > >> > >> >> > > -       VMOVNT  vec1, ((offset) + VEC_SIZE)base; \
>>> > >> > >> >> > > -       VMOVNT  vec2, ((offset) + VEC_SIZE * 2)base; \
>>> > >> > >> >> > > -       VMOVNT  vec3, ((offset) + VEC_SIZE * 3)base;
>>> > >> > >> >> > > -#else
>>> > >> > >> >> > > -# error Invalid LARGE_LOAD_SIZE
>>> > >> > >> >> > > +#ifndef MEMMOVE_VEC_LARGE_IMPL
>>> > >> > >> >> > > +# define MEMMOVE_VEC_LARGE_IMPL
>>> "memmove-vec-large-page-unrolled.S"
>>> > >> > >> >> > >  #endif
>>> > >> > >> >> > >
>>> > >> > >> >> > >  #ifndef SECTION
>>> > >> > >> >> > > @@ -426,7 +374,7 @@ L(more_8x_vec):
>>> > >> > >> >> > >  #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN
>>> (libc)
>>> > >> > >> >> > >         /* Check non-temporal store threshold.  */
>>> > >> > >> >> > >         cmp
>>>  __x86_shared_non_temporal_threshold(%rip), %RDX_LP
>>> > >> > >> >> > > -       ja      L(large_memcpy_2x)
>>> > >> > >> >> > > +       ja      L(large_memcpy)
>>> > >> > >> >> > >  #endif
>>> > >> > >> >> > >         /* To reach this point there cannot be overlap
>>> and dst > src. So
>>> > >> > >> >> > >            check for overlap and src > dst in which
>>> case correctness
>>> > >> > >> >> > > @@ -613,7 +561,7 @@ L(movsb):
>>> > >> > >> >> > >         /* If above __x86_rep_movsb_stop_threshold most
>>> likely is
>>> > >> > >> >> > >            candidate for NT moves as well.  */
>>> > >> > >> >> > >         cmp     __x86_rep_movsb_stop_threshold(%rip),
>>> %RDX_LP
>>> > >> > >> >> > > -       jae     L(large_memcpy_2x_check)
>>> > >> > >> >> > > +       jae     L(large_memcpy_check)
>>> > >> > >> >> > >  # if AVOID_SHORT_DISTANCE_REP_MOVSB || ALIGN_MOVSB
>>> > >> > >> >> > >         /* Only avoid short movsb if CPU has FSRM.  */
>>> > >> > >> >> > >  #  if
>>> X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB < 256
>>> > >> > >> >> > > @@ -673,214 +621,8 @@ L(skip_short_movsb_check):
>>> > >> > >> >> > >  # endif
>>> > >> > >> >> > >  #endif
>>> > >> > >> >> > >
>>> > >> > >> >> > > -       .p2align 4,, 10
>>> > >> > >> >> > > -#if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN
>>> (libc)
>>> > >> > >> >> > > -L(large_memcpy_2x_check):
>>> > >> > >> >> > > -       /* Entry from L(large_memcpy_2x) has a
>>> redundant load of
>>> > >> > >> >> > > -          __x86_shared_non_temporal_threshold(%rip).
>>> L(large_memcpy_2x)
>>> > >> > >> >> > > -          is only use for the non-erms memmove which
>>> is generally less
>>> > >> > >> >> > > -          common.  */
>>> > >> > >> >> > > -L(large_memcpy_2x):
>>> > >> > >> >> > > -       mov
>>>  __x86_shared_non_temporal_threshold(%rip), %R11_LP
>>> > >> > >> >> > > -       cmp     %R11_LP, %RDX_LP
>>> > >> > >> >> > > -       jb      L(more_8x_vec_check)
>>> > >> > >> >> > > -       /* To reach this point it is impossible for dst
>>> > src and
>>> > >> > >> >> > > -          overlap. Remaining to check is src > dst and
>>> overlap. rcx
>>> > >> > >> >> > > -          already contains dst - src. Negate rcx to
>>> get src - dst. If
>>> > >> > >> >> > > -          length > rcx then there is overlap and
>>> forward copy is best.  */
>>> > >> > >> >> > > -       negq    %rcx
>>> > >> > >> >> > > -       cmpq    %rcx, %rdx
>>> > >> > >> >> > > -       ja      L(more_8x_vec_forward)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* Cache align destination. First store the
>>> first 64 bytes then
>>> > >> > >> >> > > -          adjust alignments.  */
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* First vec was also loaded into VEC(0).  */
>>> > >> > >> >> > > -# if VEC_SIZE < 64
>>> > >> > >> >> > > -       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > -#  if VEC_SIZE < 32
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > -#  endif
>>> > >> > >> >> > > -# endif
>>> > >> > >> >> > > -       VMOVU   %VMM(0), (%rdi)
>>> > >> > >> >> > > -# if VEC_SIZE < 64
>>> > >> > >> >> > > -       VMOVU   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > -#  if VEC_SIZE < 32
>>> > >> > >> >> > > -       VMOVU   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > -       VMOVU   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > -#  endif
>>> > >> > >> >> > > -# endif
>>> > >> > >> >> > > +#include MEMMOVE_VEC_LARGE_IMPL
>>> > >> > >> >> > >
>>> > >> > >> >> > > -       /* Adjust source, destination, and size.  */
>>> > >> > >> >> > > -       movq    %rdi, %r8
>>> > >> > >> >> > > -       andq    $63, %r8
>>> > >> > >> >> > > -       /* Get the negative of offset for alignment.  */
>>> > >> > >> >> > > -       subq    $64, %r8
>>> > >> > >> >> > > -       /* Adjust source.  */
>>> > >> > >> >> > > -       subq    %r8, %rsi
>>> > >> > >> >> > > -       /* Adjust destination which should be aligned
>>> now.  */
>>> > >> > >> >> > > -       subq    %r8, %rdi
>>> > >> > >> >> > > -       /* Adjust length.  */
>>> > >> > >> >> > > -       addq    %r8, %rdx
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* Test if source and destination addresses
>>> will alias. If they
>>> > >> > >> >> > > -          do the larger pipeline in large_memcpy_4x
>>> alleviated the
>>> > >> > >> >> > > -          performance drop.  */
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* ecx contains -(dst - src). not ecx will
>>> return dst - src - 1
>>> > >> > >> >> > > -          which works for testing aliasing.  */
>>> > >> > >> >> > > -       notl    %ecx
>>> > >> > >> >> > > -       movq    %rdx, %r10
>>> > >> > >> >> > > -       testl   $(PAGE_SIZE - VEC_SIZE * 8), %ecx
>>> > >> > >> >> > > -       jz      L(large_memcpy_4x)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* r11 has
>>> __x86_shared_non_temporal_threshold.  Shift it left
>>> > >> > >> >> > > -          by LOG_4X_MEMCPY_THRESH to get
>>> L(large_memcpy_4x) threshold.
>>> > >> > >> >> > > -        */
>>> > >> > >> >> > > -       shlq    $LOG_4X_MEMCPY_THRESH, %r11
>>> > >> > >> >> > > -       cmp     %r11, %rdx
>>> > >> > >> >> > > -       jae     L(large_memcpy_4x)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* edx will store remainder size for copying
>>> tail.  */
>>> > >> > >> >> > > -       andl    $(PAGE_SIZE * 2 - 1), %edx
>>> > >> > >> >> > > -       /* r10 stores outer loop counter.  */
>>> > >> > >> >> > > -       shrq    $(LOG_PAGE_SIZE + 1), %r10
>>> > >> > >> >> > > -       /* Copy 4x VEC at a time from 2 pages.  */
>>> > >> > >> >> > > -       .p2align 4
>>> > >> > >> >> > > -L(loop_large_memcpy_2x_outer):
>>> > >> > >> >> > > -       /* ecx stores inner loop counter.  */
>>> > >> > >> >> > > -       movl    $(PAGE_SIZE / LARGE_LOAD_SIZE), %ecx
>>> > >> > >> >> > > -L(loop_large_memcpy_2x_inner):
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi),
>>> PREFETCHED_LOAD_SIZE * 2)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE * 2)
>>> > >> > >> >> > > -       /* Load vectors from rsi.  */
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > -       subq    $-LARGE_LOAD_SIZE, %rsi
>>> > >> > >> >> > > -       /* Non-temporal store vectors to rdi.  */
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > -       subq    $-LARGE_LOAD_SIZE, %rdi
>>> > >> > >> >> > > -       decl    %ecx
>>> > >> > >> >> > > -       jnz     L(loop_large_memcpy_2x_inner)
>>> > >> > >> >> > > -       addq    $PAGE_SIZE, %rdi
>>> > >> > >> >> > > -       addq    $PAGE_SIZE, %rsi
>>> > >> > >> >> > > -       decq    %r10
>>> > >> > >> >> > > -       jne     L(loop_large_memcpy_2x_outer)
>>> > >> > >> >> > > -       sfence
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* Check if only last 4 loads are needed.  */
>>> > >> > >> >> > > -       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       jbe     L(large_memcpy_2x_end)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* Handle the last 2 * PAGE_SIZE bytes.  */
>>> > >> > >> >> > > -L(loop_large_memcpy_2x_tail):
>>> > >> > >> >> > > -       /* Copy 4 * VEC a time forward with
>>> non-temporal stores.  */
>>> > >> > >> >> > > -       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET (1, (%rdi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       VMOVU   (%rsi), %VMM(0)
>>> > >> > >> >> > > -       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > -       subq    $-(VEC_SIZE * 4), %rsi
>>> > >> > >> >> > > -       addl    $-(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       VMOVA   %VMM(0), (%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > -       subq    $-(VEC_SIZE * 4), %rdi
>>> > >> > >> >> > > -       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       ja      L(loop_large_memcpy_2x_tail)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -L(large_memcpy_2x_end):
>>> > >> > >> >> > > -       /* Store the last 4 * VEC.  */
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VMM(0)
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VMM(1)
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VMM(2)
>>> > >> > >> >> > > -       VMOVU   -VEC_SIZE(%rsi, %rdx), %VMM(3)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       VMOVU   %VMM(0), -(VEC_SIZE * 4)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(1), -(VEC_SIZE * 3)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(2), -(VEC_SIZE * 2)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(3), -VEC_SIZE(%rdi, %rdx)
>>> > >> > >> >> > > -       VZEROUPPER_RETURN
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       .p2align 4
>>> > >> > >> >> > > -L(large_memcpy_4x):
>>> > >> > >> >> > > -       /* edx will store remainder size for copying
>>> tail.  */
>>> > >> > >> >> > > -       andl    $(PAGE_SIZE * 4 - 1), %edx
>>> > >> > >> >> > > -       /* r10 stores outer loop counter.  */
>>> > >> > >> >> > > -       shrq    $(LOG_PAGE_SIZE + 2), %r10
>>> > >> > >> >> > > -       /* Copy 4x VEC at a time from 4 pages.  */
>>> > >> > >> >> > > -       .p2align 4
>>> > >> > >> >> > > -L(loop_large_memcpy_4x_outer):
>>> > >> > >> >> > > -       /* ecx stores inner loop counter.  */
>>> > >> > >> >> > > -       movl    $(PAGE_SIZE / LARGE_LOAD_SIZE), %ecx
>>> > >> > >> >> > > -L(loop_large_memcpy_4x_inner):
>>> > >> > >> >> > > -       /* Only one prefetch set per page as doing 4
>>> pages give more
>>> > >> > >> >> > > -          time for prefetcher to keep up.  */
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi), PAGE_SIZE +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi), PAGE_SIZE * 2 +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET(1, (%rsi), PAGE_SIZE * 3 +
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       /* Load vectors from rsi.  */
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), PAGE_SIZE * 2, %VMM(8),
>>> %VMM(9), %VMM(10), %VMM(11))
>>> > >> > >> >> > > -       LOAD_ONE_SET((%rsi), PAGE_SIZE * 3, %VMM(12),
>>> %VMM(13), %VMM(14), %VMM(15))
>>> > >> > >> >> > > -       subq    $-LARGE_LOAD_SIZE, %rsi
>>> > >> > >> >> > > -       /* Non-temporal store vectors to rdi.  */
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), 0, %VMM(0), %VMM(1),
>>> %VMM(2), %VMM(3))
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), PAGE_SIZE, %VMM(4),
>>> %VMM(5), %VMM(6), %VMM(7))
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), PAGE_SIZE * 2, %VMM(8),
>>> %VMM(9), %VMM(10), %VMM(11))
>>> > >> > >> >> > > -       STORE_ONE_SET((%rdi), PAGE_SIZE * 3, %VMM(12),
>>> %VMM(13), %VMM(14), %VMM(15))
>>> > >> > >> >> > > -       subq    $-LARGE_LOAD_SIZE, %rdi
>>> > >> > >> >> > > -       decl    %ecx
>>> > >> > >> >> > > -       jnz     L(loop_large_memcpy_4x_inner)
>>> > >> > >> >> > > -       addq    $(PAGE_SIZE * 3), %rdi
>>> > >> > >> >> > > -       addq    $(PAGE_SIZE * 3), %rsi
>>> > >> > >> >> > > -       decq    %r10
>>> > >> > >> >> > > -       jne     L(loop_large_memcpy_4x_outer)
>>> > >> > >> >> > > -       sfence
>>> > >> > >> >> > > -       /* Check if only last 4 loads are needed.  */
>>> > >> > >> >> > > -       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       jbe     L(large_memcpy_4x_end)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       /* Handle the last 4  * PAGE_SIZE bytes.  */
>>> > >> > >> >> > > -L(loop_large_memcpy_4x_tail):
>>> > >> > >> >> > > -       /* Copy 4 * VEC a time forward with
>>> non-temporal stores.  */
>>> > >> > >> >> > > -       PREFETCH_ONE_SET (1, (%rsi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       PREFETCH_ONE_SET (1, (%rdi),
>>> PREFETCHED_LOAD_SIZE)
>>> > >> > >> >> > > -       VMOVU   (%rsi), %VMM(0)
>>> > >> > >> >> > > -       VMOVU   VEC_SIZE(%rsi), %VMM(1)
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 2)(%rsi), %VMM(2)
>>> > >> > >> >> > > -       VMOVU   (VEC_SIZE * 3)(%rsi), %VMM(3)
>>> > >> > >> >> > > -       subq    $-(VEC_SIZE * 4), %rsi
>>> > >> > >> >> > > -       addl    $-(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       VMOVA   %VMM(0), (%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(1), VEC_SIZE(%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(2), (VEC_SIZE * 2)(%rdi)
>>> > >> > >> >> > > -       VMOVA   %VMM(3), (VEC_SIZE * 3)(%rdi)
>>> > >> > >> >> > > -       subq    $-(VEC_SIZE * 4), %rdi
>>> > >> > >> >> > > -       cmpl    $(VEC_SIZE * 4), %edx
>>> > >> > >> >> > > -       ja      L(loop_large_memcpy_4x_tail)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -L(large_memcpy_4x_end):
>>> > >> > >> >> > > -       /* Store the last 4 * VEC.  */
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VMM(0)
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VMM(1)
>>> > >> > >> >> > > -       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VMM(2)
>>> > >> > >> >> > > -       VMOVU   -VEC_SIZE(%rsi, %rdx), %VMM(3)
>>> > >> > >> >> > > -
>>> > >> > >> >> > > -       VMOVU   %VMM(0), -(VEC_SIZE * 4)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(1), -(VEC_SIZE * 3)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(2), -(VEC_SIZE * 2)(%rdi, %rdx)
>>> > >> > >> >> > > -       VMOVU   %VMM(3), -VEC_SIZE(%rdi, %rdx)
>>> > >> > >> >> > > -       VZEROUPPER_RETURN
>>> > >> > >> >> > > -#endif
>>> > >> > >> >> > >  END (MEMMOVE_SYMBOL (__memmove, unaligned_erms))
>>> > >> > >> >> > >
>>> > >> > >> >> > >  #if IS_IN (libc)
>>> > >> > >> >> > > --
>>> > >> > >> >> > > 2.43.0
>>> > >> > >> >> > >
>>> > >> > >> >> >
>>> > >> > >> >> > For the entire series:
>>> > >> > >> >> > I tested this natively on x86_64 when I wrote this, but I
>>> no longer
>>> > >> > >> >> > have any x86 machines.
>>> > >> > >> >> > I re-tested w/ QEMU before re-posting, but I think it
>>> would be worthwhile if
>>> > >> > >> >> ... if someone could re-check natively before commit (after
>>> review has
>>> > >> > >> >> been completed).
>>> > >> > >> >
>>> > >> > >> >
>>> > >> > >> > Hi Noah,
>>> > >> > >> >
>>> > >> > >> > I tested this patch and got geomean regression upto 6% on
>>> large data size microbenchmark.
>>> > >> > >> >
>>> > >> > >> > Thanks,
>>> > >> > >> > Sunil
>>> > >> > >>
>>> > >> > >> Any impact on SPEC CPU?
>>> > >> > >>
>>> > >> > >
>>> > >> > > Spec2017 regression(502.gcc_r , 520.omnetpp_r) on newer AVX-256
>>> system after applying 3 patch sets.
>>> > >> > >
>>> > >> > > 503.bwaves_r -1%
>>> > >> > > 507.cactuBSSN_r -1%
>>> > >> > > 508.namd_r 0%
>>> > >> > > 510.parest_r 0%
>>> > >> > > 511.povray_r 0%
>>> > >> > > 519.lbm_r 0%
>>> > >> > > 521.wrf_r 0%
>>> > >> > > 526.blender_r 1%
>>> > >> > > 527.cam4_r 1%
>>> > >> > > 538.imagick_r 0%
>>> > >> > > 544.nab_r 0%
>>> > >> > > 549.fotonik3d_r 1%
>>> > >> > > 554.roms_r 0%
>>> > >> > > 500.perlbench_r 0%
>>> > >> > > 502.gcc_r -2%
>>> > >> > > 505.mcf_r 0%
>>> > >> > > 520.omnetpp_r -3%
>>> > >> > > 523.xalancbmk_r 1%
>>> > >> > > 525.x264_r -1%
>>> > >> > > 531.deepsjeng_r 0%
>>> > >> > > 541.leela_r 0%
>>> > >> > > 548.exchange2_r 0%
>>> > >> > > 557.xz_r -1%
>>> > >> >
>>> > >> > That's possibly within the noise, but as discussed, this should be
>>> > >> > handled in a follow up
>>> > >> > by adjusting the preferred implementation on whatever newer
>>> system you
>>> > >> > had in mind.
>>> > >>
>>> > >>
>>> > >> Alternatively, you can give me a list of CPUs you would prefer to
>>> keep the old
>>> > >> behavior on.
>>> > >> I do think it should be CPU by CPU w/ strong benchmarks to support
>>> it,
>>> > >> but if you
>>> > >> want more aggressive continuation of current behavior on Intel
>>> machines that is
>>> > >> understandable too.
>>> > >>
>>> > >> I will note, before leaving Intel I did do some SPR benchmarks of
>>> page unrolled
>>> > >> vs normal loop:
>>> > >>
>>> > >>
>>> https://docs.google.com/spreadsheets/d/1-8M-VNeERLxzsghX3CnYUS09pURckmAJLcUkAzTJaVE/edit?usp=sharing
>>> > >>
>>> > >> Comparing NT vs NT_4x. Lower numbers are better.
>>> > >> (CD is movdir64 just for reference).
>>> > >>
>>> > >> The benchmark code does slightly vary from what we have in GLIBC,
>>> > >> but I take it as evidence that it is not clear we should blindly be
>>> continuing
>>> > >> the page unrolled implementation.
>>> > >>
>>> > >>
>>> > >>
>>> > >> > >
>>> > >> > > --Sunil
>>> > >> > >
>>> > >> > >>
>>> > >> > >> --
>>> > >> > >> H.J.
>>> >
>>> >
>>> >
>>> > --
>>> > H.J.
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20251128/55f6f14a/attachment-0001.htm>


More information about the Libc-alpha mailing list