diff mbox series

[5/5] X86-64: Optimize memmove-vec-unaligned-erms.S

Message ID 20210824082753.3356637-5-goldstein.w.n@gmail.com
State Superseded
Headers show
Series [1/5] string: Make tests birdirectional test-memcpy.c | expand

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent
dj/TryBot-32bit success Build for i686

Commit Message

Noah Goldstein Aug. 24, 2021, 8:27 a.m. UTC
No bug. This commit optimizes memmove-vec-unaligned.S.

The optimizations are in descending order of importance to the
L(less_vec), L(movsb), the 8x forward/backward loops and various
target alignments that have minimal code size impact.

The L(less_vec) optimizations are to:

    1. Readjust the branch order to either given hotter paths a fall
    through case or have less branches in there way.
    2. Moderately change the size classes to make hot branches hotter
    and thus increase predictability.
    3. Try and minimize branch aliasing to avoid BPU thrashing based
    misses.
    4. 64 byte the prior function entry. This is to avoid cases where
    seemingly unrelated changes end up have severe negative
    performance impacts.

The L(movsb) optimizations are to:

    1. Reduce the number of taken branches needed to determine if
    movsb should be used.
    2. 64 byte align either dst if the CPU has fsrm or if dst and src
    do not 4k alias.
    3. 64 byte align src if the CPU does not have fsrm and dst and src
    do 4k alias.

The 8x forward/backward loop optimizations are to:

    1. Reduce instructions needed for aligning to VEC_SIZE.
    2. Reduce uops and code size of the loops.

All tests in string/ passing.
---
See performance data attached.
Included benchmarks: memcpy-random, memcpy, memmove, memcpy-walk, memmove-walk, memcpy-large  

The first page is a summary with the ifunc selection version for
erms/non-erms for each computers. Then in the following 4 sheets are
all the numbers for sse2, avx for Skylake and sse2, avx2, evex, and
avx512 for Tigerlake.

Benchmark CPUS: Skylake:
https://ark.intel.com/content/www/us/en/ark/products/149091/intel-core-i7-8565u-processor-8m-cache-up-to-4-60-ghz.html

Tigerlake:
https://ark.intel.com/content/www/us/en/ark/products/208921/intel-core-i7-1165g7-processor-12m-cache-up-to-4-70-ghz-with-ipu.html

All times are geometric mean of N=30.

"Cur" refers to the current implementation "New" refers to this
patches implementation

Score refers to new/cur (low means improvement, high means
degragation). Scores are color coded. The more green the better, the
more red the worse.


Some notes on the numbers:

In my opinion most of the benchmarks where src/dst align are in [0,
64] have some unpredictable and unfortunate noise from non-obvious
false dependencies between stores to dst and next iterations loads
from src. For example in the 8x forward case, the store of VEC(4) will
end up stalling next iterations load queue, so if size was large
enough that the begining of dst was flushed from L1 this can have a
seemingly random but significant impact on the benchmark result.

There are significant performance improvements/degregations in the [0,
VEC_SIZE]. I didn't treat these as imporant as I think in this size
range the branch pattern indicated by the random tests is more
important. On the random tests the new implementation performance
significantly better.

I also added logic to align before L(movsb). If you see the new random
benchmarks with fixed size this leads to roughly a 10-20% performance
improvement for some hot sizes. I am not 100% convinced this is needed
as generally for larger copies that would go to movsb they are already
aligned but even in the fixed loop cases, especially on Skylake w.o
FSRM it seems aligning before movsb pays off. Let me know if you think
this is unnecessary.

There are occasional performance degregations at odd splots throughout
the medium range sizes in the fixed memcpy benchmarks. I think
generally there is more good than harm here and at the moment I don't
have an explination for why these certain configurations seem to
perform worse. On the plus side, however, it also seems that there are
unexplained improvements of the same magnitude patterened with the
degregations (and both are sparse) so I ultimately believe it should
be acceptable. if this is not the case let me know.

The memmove benchmarks look a bit worse, especially for the erms
case. Part of this is from the nop cases which I didn't treat as
important. But part of it is also because to optimize for what I
expect to be the common case of no overlap the overlap case has extra
branches and overhead. I think this is inevitable when implementing
memmove and memcpy in the same file, but if this is unacceptable let
me know.


Note: I benchmarks before two changes that made it into the final version:

-#if !defined USE_MULTIARCH || !IS_IN (libc)
-L(nop):
-       ret
-#else
+       VMOVU   %VEC(1), -VEC_SIZE(%rdi, %rdx)
        VZEROUPPER_RETURN
-#endif



And

+       testl   $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB, __x86_string_control(%rip)
-       andl    $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB, __x86_string_control(%rip)


I don't think either of these should have any impact.

I made the former change because I think it was a bug that could cause
use of avx2 w.o vzeroupper and the latter because I think it could
cause issues on multicore platforms.

    
 sysdeps/x86/sysdep.h                          |  13 +-
 .../multiarch/memmove-vec-unaligned-erms.S    | 484 +++++++++++-------
 2 files changed, 317 insertions(+), 180 deletions(-)

Comments

Noah Goldstein Aug. 24, 2021, 9:12 a.m. UTC | #1
On Tue, Aug 24, 2021 at 4:29 AM Noah Goldstein <goldstein.w.n@gmail.com>
wrote:

> No bug. This commit optimizes memmove-vec-unaligned.S.
>
> The optimizations are in descending order of importance to the
> L(less_vec), L(movsb), the 8x forward/backward loops and various
> target alignments that have minimal code size impact.
>
> The L(less_vec) optimizations are to:
>
>     1. Readjust the branch order to either given hotter paths a fall
>     through case or have less branches in there way.
>     2. Moderately change the size classes to make hot branches hotter
>     and thus increase predictability.
>     3. Try and minimize branch aliasing to avoid BPU thrashing based
>     misses.
>     4. 64 byte the prior function entry. This is to avoid cases where
>     seemingly unrelated changes end up have severe negative
>     performance impacts.
>
> The L(movsb) optimizations are to:
>
>     1. Reduce the number of taken branches needed to determine if
>     movsb should be used.
>     2. 64 byte align either dst if the CPU has fsrm or if dst and src
>     do not 4k alias.
>     3. 64 byte align src if the CPU does not have fsrm and dst and src
>     do 4k alias.
>
> The 8x forward/backward loop optimizations are to:
>
>     1. Reduce instructions needed for aligning to VEC_SIZE.
>     2. Reduce uops and code size of the loops.
>
> All tests in string/ passing.
> ---
> See performance data attached.
> Included benchmarks: memcpy-random, memcpy, memmove, memcpy-walk,
> memmove-walk, memcpy-large
>
> The first page is a summary with the ifunc selection version for
> erms/non-erms for each computers. Then in the following 4 sheets are
> all the numbers for sse2, avx for Skylake and sse2, avx2, evex, and
> avx512 for Tigerlake.
>
> Benchmark CPUS: Skylake:
>
> https://ark.intel.com/content/www/us/en/ark/products/149091/intel-core-i7-8565u-processor-8m-cache-up-to-4-60-ghz.html
>
> Tigerlake:
>
> https://ark.intel.com/content/www/us/en/ark/products/208921/intel-core-i7-1165g7-processor-12m-cache-up-to-4-70-ghz-with-ipu.html
>
> All times are geometric mean of N=30.
>
> "Cur" refers to the current implementation "New" refers to this
> patches implementation
>
> Score refers to new/cur (low means improvement, high means
> degragation). Scores are color coded. The more green the better, the
> more red the worse.
>
>
> Some notes on the numbers:
>
> In my opinion most of the benchmarks where src/dst align are in [0,
> 64] have some unpredictable and unfortunate noise from non-obvious
> false dependencies between stores to dst and next iterations loads
> from src. For example in the 8x forward case, the store of VEC(4) will
> end up stalling next iterations load queue, so if size was large
> enough that the begining of dst was flushed from L1 this can have a
> seemingly random but significant impact on the benchmark result.
>
> There are significant performance improvements/degregations in the [0,
> VEC_SIZE]. I didn't treat these as imporant as I think in this size
> range the branch pattern indicated by the random tests is more
> important. On the random tests the new implementation performance
> significantly better.
>
> I also added logic to align before L(movsb). If you see the new random
> benchmarks with fixed size this leads to roughly a 10-20% performance
> improvement for some hot sizes. I am not 100% convinced this is needed
> as generally for larger copies that would go to movsb they are already
> aligned but even in the fixed loop cases, especially on Skylake w.o
> FSRM it seems aligning before movsb pays off. Let me know if you think
> this is unnecessary.
>
> There are occasional performance degregations at odd splots throughout
> the medium range sizes in the fixed memcpy benchmarks. I think
> generally there is more good than harm here and at the moment I don't
> have an explination for why these certain configurations seem to
> perform worse. On the plus side, however, it also seems that there are
> unexplained improvements of the same magnitude patterened with the
> degregations (and both are sparse) so I ultimately believe it should
> be acceptable. if this is not the case let me know.
>
> The memmove benchmarks look a bit worse, especially for the erms
> case. Part of this is from the nop cases which I didn't treat as
> important. But part of it is also because to optimize for what I
> expect to be the common case of no overlap the overlap case has extra
> branches and overhead. I think this is inevitable when implementing
> memmove and memcpy in the same file, but if this is unacceptable let
> me know.
>
>
> Note: I benchmarks before two changes that made it into the final version:
>
> -#if !defined USE_MULTIARCH || !IS_IN (libc)
> -L(nop):
> -       ret
> -#else
> +       VMOVU   %VEC(1), -VEC_SIZE(%rdi, %rdx)
>         VZEROUPPER_RETURN
> -#endif
>
>
>
> And
>
> +       testl   $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB,
> __x86_string_control(%rip)
> -       andl    $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB,
> __x86_string_control(%rip)
>
>
> I don't think either of these should have any impact.
>
> I made the former change because I think it was a bug that could cause
> use of avx2 w.o vzeroupper and the latter because I think it could
> cause issues on multicore platforms.
>
>
>  sysdeps/x86/sysdep.h                          |  13 +-
>  .../multiarch/memmove-vec-unaligned-erms.S    | 484 +++++++++++-------
>  2 files changed, 317 insertions(+), 180 deletions(-)
>
> diff --git a/sysdeps/x86/sysdep.h b/sysdeps/x86/sysdep.h
> index cac1d762fb..9226d2c6c9 100644
> --- a/sysdeps/x86/sysdep.h
> +++ b/sysdeps/x86/sysdep.h
> @@ -78,15 +78,18 @@ enum cf_protection_level
>  #define ASM_SIZE_DIRECTIVE(name) .size name,.-name;
>
>  /* Define an entry point visible from C.  */
> -#define        ENTRY(name)
>            \
> -  .globl C_SYMBOL_NAME(name);
>     \
> -  .type C_SYMBOL_NAME(name),@function;
>    \
> -  .align ALIGNARG(4);
>     \
> +#define        P2ALIGN_ENTRY(name, alignment)
>                             \
> +  .globl C_SYMBOL_NAME(name);
>             \
> +  .type C_SYMBOL_NAME(name),@function;
>                    \
> +  .align ALIGNARG(alignment);
>             \
>    C_LABEL(name)
>             \
>    cfi_startproc;
>    \
> -  _CET_ENDBR;
>     \
> +  _CET_ENDBR;                                                        \
>    CALL_MCOUNT
>
> +#define        ENTRY(name) P2ALIGN_ENTRY(name, 4)
> +
> +
>  #undef END
>  #define END(name)
>     \
>    cfi_endproc;
>    \
> diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> index 9f02624375..75b6efe969 100644
> --- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> @@ -165,6 +165,32 @@
>  # error Invalid LARGE_LOAD_SIZE
>  #endif
>
> +/* Whether to align before movsb. Ultimately we want 64 byte align
> +   and not worth it to load 4x VEC for VEC_SIZE == 16.  */
> +#define ALIGN_MOVSB    (VEC_SIZE       >       16)
> +
> +/* Number of VECs to align movsb to.  */
> +#if VEC_SIZE == 64
> +# define MOVSB_ALIGN_TO        (VEC_SIZE)
> +#else
> +# define MOVSB_ALIGN_TO        (VEC_SIZE       *       2)
> +#endif
> +
> +/* Macro for copying inclusive power of 2 range with two register
> +   loads.  */
> +#define COPY_BLOCK(mov_inst, src_reg, dst_reg, size_reg, len, tmp_reg0,
> tmp_reg1)      \
> +       mov_inst (%src_reg), %tmp_reg0; \
> +       mov_inst -(len)(%src_reg, %size_reg), %tmp_reg1; \
> +       mov_inst %tmp_reg0, (%dst_reg); \
> +       mov_inst %tmp_reg1, -(len)(%dst_reg, %size_reg);
> +
> +/* Define all copies used by L(less_vec) for VEC_SIZE of 16, 32, or
> +   64.  */
> +#define COPY_4_8       COPY_BLOCK(movl, rsi, rdi, rdx, 4, ecx, esi)
> +#define COPY_8_16      COPY_BLOCK(movq, rsi, rdi, rdx, 8, rcx, rsi)
> +#define COPY_16_32     COPY_BLOCK(vmovdqu, rsi, rdi, rdx, 16, xmm0, xmm1)
> +#define COPY_32_64     COPY_BLOCK(vmovdqu64, rsi, rdi, rdx, 32, ymm16,
> ymm17)
> +
>  #ifndef SECTION
>  # error SECTION is not defined!
>  #endif
> @@ -198,7 +224,13 @@ L(start):
>         movl    %edx, %edx
>  # endif
>         cmp     $VEC_SIZE, %RDX_LP
> +       /* Based on SPEC2017 distribution both 16 and 32 memcpy calls are
> +          really hot so we want them to take the same branch path.  */
> +#if VEC_SIZE > 16
> +       jbe     L(less_vec)
> +#else
>         jb      L(less_vec)
> +#endif
>         cmp     $(VEC_SIZE * 2), %RDX_LP
>         ja      L(more_2x_vec)
>  #if !defined USE_MULTIARCH || !IS_IN (libc)
> @@ -206,15 +238,10 @@ L(last_2x_vec):
>  #endif
>         /* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE.  */
>         VMOVU   (%rsi), %VEC(0)
> -       VMOVU   -VEC_SIZE(%rsi,%rdx), %VEC(1)
> +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VEC(1)
>         VMOVU   %VEC(0), (%rdi)
> -       VMOVU   %VEC(1), -VEC_SIZE(%rdi,%rdx)
> -#if !defined USE_MULTIARCH || !IS_IN (libc)
> -L(nop):
> -       ret
> -#else
> +       VMOVU   %VEC(1), -VEC_SIZE(%rdi, %rdx)
>         VZEROUPPER_RETURN
> -#endif
>  #if defined USE_MULTIARCH && IS_IN (libc)
>  END (MEMMOVE_SYMBOL (__memmove, unaligned))
>
> @@ -289,7 +316,9 @@ ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk,
> unaligned_erms))
>  END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms))
>  # endif
>
> -ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms))
> +/* Cache align entry so that branch heavy L(less_vec) maintains good
> +   alignment.  */
> +P2ALIGN_ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms), 6)
>         movq    %rdi, %rax
>  L(start_erms):
>  # ifdef __ILP32__
> @@ -297,123 +326,217 @@ L(start_erms):
>         movl    %edx, %edx
>  # endif
>         cmp     $VEC_SIZE, %RDX_LP
> +       /* Based on SPEC2017 distribution both 16 and 32 memcpy calls are
> +          really hot so we want them to take the same branch path.  */
> +# if VEC_SIZE > 16
> +       jbe     L(less_vec)
> +# else
>         jb      L(less_vec)
> +# endif
>         cmp     $(VEC_SIZE * 2), %RDX_LP
>         ja      L(movsb_more_2x_vec)
>  L(last_2x_vec):
> -       /* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE. */
> +       /* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE.  */
>         VMOVU   (%rsi), %VEC(0)
> -       VMOVU   -VEC_SIZE(%rsi,%rdx), %VEC(1)
> +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VEC(1)
>         VMOVU   %VEC(0), (%rdi)
> -       VMOVU   %VEC(1), -VEC_SIZE(%rdi,%rdx)
> +       VMOVU   %VEC(1), -VEC_SIZE(%rdi, %rdx)
>  L(return):
> -#if VEC_SIZE > 16
> +# if VEC_SIZE > 16
>         ZERO_UPPER_VEC_REGISTERS_RETURN
> -#else
> +# else
>         ret
> +# endif
>  #endif
> +#if VEC_SIZE == 64
> +L(copy_8_15):
> +       COPY_8_16
> +       ret
>
> -L(movsb):
> -       cmp     __x86_rep_movsb_stop_threshold(%rip), %RDX_LP
> -       jae     L(more_8x_vec)
> -       cmpq    %rsi, %rdi
> -       jb      1f
> -       /* Source == destination is less common.  */
> -       je      L(nop)
> -       leaq    (%rsi,%rdx), %r9
> -       cmpq    %r9, %rdi
> -       /* Avoid slow backward REP MOVSB.  */
> -       jb      L(more_8x_vec_backward)
> -# if AVOID_SHORT_DISTANCE_REP_MOVSB
> -       andl    $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB,
> __x86_string_control(%rip)
> -       jz      3f
> -       movq    %rdi, %rcx
> -       subq    %rsi, %rcx
> -       jmp     2f
> -# endif
> -1:
> -# if AVOID_SHORT_DISTANCE_REP_MOVSB
> -       andl    $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB,
> __x86_string_control(%rip)
> -       jz      3f
> -       movq    %rsi, %rcx
> -       subq    %rdi, %rcx
> -2:
> -/* Avoid "rep movsb" if RCX, the distance between source and destination,
> -   is N*4GB + [1..63] with N >= 0.  */
> -       cmpl    $63, %ecx
> -       jbe     L(more_2x_vec)  /* Avoid "rep movsb" if ECX <= 63.  */
> -3:
> -# endif
> -       mov     %RDX_LP, %RCX_LP
> -       rep movsb
> -L(nop):
> +L(copy_33_63):
> +       COPY_32_64
>         ret
>  #endif
> -
> +       /* Only worth aligning if near end of 16 byte block and won't get
> +          first branch in first decode after jump.  */
> +       .p2align 4,, 6
>  L(less_vec):
> -       /* Less than 1 VEC.  */
>  #if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64
>  # error Unsupported VEC_SIZE!
>  #endif
> -#if VEC_SIZE > 32
> -       cmpb    $32, %dl
> -       jae     L(between_32_63)
> +       /* Second set of branches for smallest copies.  */
> +       cmpl    $(VEC_SIZE / 4), %edx
> +       jb      L(less_quarter_vec)
> +
> +       cmpl    $(VEC_SIZE / 2), %edx
> +#if VEC_SIZE == 64
> +       /* We branch to [33, 63] instead of [16, 32] to give [16, 32] fall
> +          through path as [16, 32] is hotter.  */
> +       ja      L(copy_33_63)
> +       COPY_16_32
> +#elif VEC_SIZE == 32
> +       /* Branch to [8, 15]. Fall through to [16, 32].  */
> +       jb      L(copy_8_15)
> +       COPY_16_32
> +#else
> +       /* Branch to [4, 7]. Fall through to [8, 15].  */
> +       jb      L(copy_4_7)
> +       COPY_8_16
>  #endif
> -#if VEC_SIZE > 16
> -       cmpb    $16, %dl
> -       jae     L(between_16_31)
> -#endif
> -       cmpb    $8, %dl
> -       jae     L(between_8_15)
> -       cmpb    $4, %dl
> -       jae     L(between_4_7)
> -       cmpb    $1, %dl
> -       ja      L(between_2_3)
> -       jb      1f
> +       ret
> +       /* Align if won't cost too many bytes.  */
> +       .p2align 4,, 6
> +L(copy_4_7):
> +       COPY_4_8
> +       ret
> +
> +       /* Cold target. No need to align.  */
> +L(copy_1):
>         movzbl  (%rsi), %ecx
>         movb    %cl, (%rdi)
> -1:
>         ret
> +
> +       /* Colder copy case for [0, VEC_SIZE / 4 - 1].  */
> +L(less_quarter_vec):
>  #if VEC_SIZE > 32
> -L(between_32_63):
> -       /* From 32 to 63.  No branch when size == 32.  */
> -       VMOVU   (%rsi), %YMM0
> -       VMOVU   -32(%rsi,%rdx), %YMM1
> -       VMOVU   %YMM0, (%rdi)
> -       VMOVU   %YMM1, -32(%rdi,%rdx)
> -       VZEROUPPER_RETURN
> +       cmpl    $8, %edx
> +       jae     L(copy_8_15)
>  #endif
>  #if VEC_SIZE > 16
> -       /* From 16 to 31.  No branch when size == 16.  */
> -L(between_16_31):
> -       VMOVU   (%rsi), %XMM0
> -       VMOVU   -16(%rsi,%rdx), %XMM1
> -       VMOVU   %XMM0, (%rdi)
> -       VMOVU   %XMM1, -16(%rdi,%rdx)
> -       VZEROUPPER_RETURN
> -#endif
> -L(between_8_15):
> -       /* From 8 to 15.  No branch when size == 8.  */
> -       movq    -8(%rsi,%rdx), %rcx
> -       movq    (%rsi), %rsi
> -       movq    %rcx, -8(%rdi,%rdx)
> -       movq    %rsi, (%rdi)
> -       ret
> -L(between_4_7):
> -       /* From 4 to 7.  No branch when size == 4.  */
> -       movl    -4(%rsi,%rdx), %ecx
> -       movl    (%rsi), %esi
> -       movl    %ecx, -4(%rdi,%rdx)
> -       movl    %esi, (%rdi)
> +       cmpl    $4, %edx
> +       jae     L(copy_4_7)
> +#endif
> +       cmpl    $1, %edx
> +       je      L(copy_1)
> +       jb      L(copy_0)
> +       /* Fall through into copy [2, 3] as it is more common than [0, 1].
> +        */
> +       movzwl  (%rsi), %ecx
> +       movzbl  -1(%rsi, %rdx), %esi
> +       movw    %cx, (%rdi)
> +       movb    %sil, -1(%rdi, %rdx)
> +L(copy_0):
>         ret
> -L(between_2_3):
> -       /* From 2 to 3.  No branch when size == 2.  */
> -       movzwl  -2(%rsi,%rdx), %ecx
> -       movzwl  (%rsi), %esi
> -       movw    %cx, -2(%rdi,%rdx)
> -       movw    %si, (%rdi)
> +
> +       .p2align 4
> +#if VEC_SIZE == 32
> +L(copy_8_15):
> +       COPY_8_16
>         ret
> +       /* COPY_8_16 is exactly 17 bytes so don't want to p2align after as
> +          it wastes 15 bytes of code and 1 byte off is fine.  */
> +#endif
> +
> +#if defined USE_MULTIARCH && IS_IN (libc)
> +L(movsb):
> +       movq    %rdi, %rcx
> +       subq    %rsi, %rcx
> +       /* Go to backwards temporal copy if overlap no matter what as
> +          backward movsb is slow.  */
> +       cmpq    %rdx, %rcx
> +       /* L(more_8x_vec_backward_check_nop) checks for src == dst.  */
> +       jb      L(more_8x_vec_backward_check_nop)
> +       /* If above __x86_rep_movsb_stop_threshold most likely is candidate
> +          for NT moves aswell.  */
> +       cmp     __x86_rep_movsb_stop_threshold(%rip), %RDX_LP
> +       jae     L(large_memcpy_2x_check)
> +# if ALIGN_MOVSB
> +       VMOVU   (%rsi), %VEC(0)
> +#  if MOVSB_ALIGN_TO > VEC_SIZE
> +       VMOVU   VEC_SIZE(%rsi), %VEC(1)
> +#  endif
> +#  if MOVSB_ALIGN_TO > (VEC_SIZE * 2)
> +#   error Unsupported MOVSB_ALIGN_TO
> +#  endif
> +       /* Store dst for use after rep movsb.  */
> +       movq    %rdi, %r8
> +# endif
> +# if AVOID_SHORT_DISTANCE_REP_MOVSB
> +       /* Only avoid short movsb if CPU has FSRM.  */
> +       testl   $X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB,
> __x86_string_control(%rip)
> +       jz      L(skip_short_movsb_check)
> +       /* Avoid "rep movsb" if RCX, the distance between source and
> +          destination, is N*4GB + [1..63] with N >= 0.  */
> +
> +       /* ecx contains dst - src. Early check for backward copy conditions
> +          means only case of slow movsb with src = dst + [0, 63] is ecx in
> +          [-63, 0]. Use unsigned comparison with -64 check for that
> case.  */
> +       cmpl    $-64, %ecx
> +       ja      L(more_8x_vec_forward)
> +# endif
> +# if ALIGN_MOVSB
> +       /* Fall through means cpu has FSRM. In that case exclusively align
> +          destination.  */
> +
> +       /* Subtract dst from src. Add back after dst aligned.  */
> +       subq    %rdi, %rsi
> +       /* Add dst to len. Subtract back after dst aligned.  */
> +       leaq    (%rdi, %rdx), %rcx
> +       /* Exclusively align dst to MOVSB_ALIGN_TO (64).  */
> +       addq    $(MOVSB_ALIGN_TO - 1), %rdi
> +       andq    $-(MOVSB_ALIGN_TO), %rdi
> +       /* Restore src and len adjusted with new values for aligned dst.
> */
> +       addq    %rdi, %rsi
> +       subq    %rdi, %rcx
>
> +       rep     movsb
> +       VMOVU   %VEC(0), (%r8)
> +#  if MOVSB_ALIGN_TO > VEC_SIZE
> +       VMOVU   %VEC(1), VEC_SIZE(%r8)
> +#  endif
> +       VZEROUPPER_RETURN
> +L(movsb_align_dst):
> +       /* Subtract dst from src. Add back after dst aligned.  */
> +       subq    %rdi, %rsi
> +       /* Add dst to len. Subtract back after dst aligned. -1 because dst
> +          is initially aligned to MOVSB_ALIGN_TO - 1.  */
> +       leaq    -(1)(%rdi, %rdx), %rcx
> +       /* Inclusively align dst to MOVSB_ALIGN_TO - 1.  */
> +       orq     $(MOVSB_ALIGN_TO - 1), %rdi
> +       leaq    1(%rdi, %rsi), %rsi
> +       /* Restore src and len adjusted with new values for aligned dst.
> */
> +       subq    %rdi, %rcx
> +       /* Finish aligning dst.  */
> +       incq    %rdi
> +       rep     movsb
> +       VMOVU   %VEC(0), (%r8)
> +#  if MOVSB_ALIGN_TO > VEC_SIZE
> +       VMOVU   %VEC(1), VEC_SIZE(%r8)
> +#  endif
> +       VZEROUPPER_RETURN
> +
> +L(skip_short_movsb_check):
> +       /* If CPU does not have FSRM two options for aligning. Align src if
> +          dst and src 4k alias. Otherwise align dst.  */
> +       testl   $(PAGE_SIZE - 512), %ecx
> +       jnz     L(movsb_align_dst)
> +       /* rcx already has dst - src.  */
> +       movq    %rcx, %r9
> +       /* Add src to len. Subtract back after src aligned. -1 because src
> +          is initially aligned to MOVSB_ALIGN_TO - 1.  */
> +       leaq    -(1)(%rsi, %rdx), %rcx
> +       /* Inclusively align src to MOVSB_ALIGN_TO - 1.  */
> +       orq     $(MOVSB_ALIGN_TO - 1), %rsi
> +       /* Restore dst and len adjusted with new values for aligned dst.
> */
> +       leaq    1(%rsi, %r9), %rdi
> +       subq    %rsi, %rcx
> +       /* Finish aligning src.  */
> +       incq    %rsi
> +       rep     movsb
> +       VMOVU   %VEC(0), (%r8)
> +#  if MOVSB_ALIGN_TO > VEC_SIZE
> +       VMOVU   %VEC(1), VEC_SIZE(%r8)
> +#  endif
> +       VZEROUPPER_RETURN
> +# else
> +       /* Not alignined rep movsb so just copy.  */
> +       mov     %RDX_LP, %RCX_LP
> +       rep     movsb
> +       ret
> +# endif
> +#endif
> +       /* Align if doesn't cost too many bytes.  */
> +       .p2align 4,, 6
>  #if defined USE_MULTIARCH && IS_IN (libc)
>  L(movsb_more_2x_vec):
>         cmp     __x86_rep_movsb_threshold(%rip), %RDX_LP
> @@ -426,50 +549,60 @@ L(more_2x_vec):
>         ja      L(more_8x_vec)
>         cmpq    $(VEC_SIZE * 4), %rdx
>         jbe     L(last_4x_vec)
> -       /* Copy from 4 * VEC + 1 to 8 * VEC, inclusively. */
> +       /* Copy from 4 * VEC + 1 to 8 * VEC, inclusively.  */
>         VMOVU   (%rsi), %VEC(0)
>         VMOVU   VEC_SIZE(%rsi), %VEC(1)
>         VMOVU   (VEC_SIZE * 2)(%rsi), %VEC(2)
>         VMOVU   (VEC_SIZE * 3)(%rsi), %VEC(3)
> -       VMOVU   -VEC_SIZE(%rsi,%rdx), %VEC(4)
> -       VMOVU   -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(5)
> -       VMOVU   -(VEC_SIZE * 3)(%rsi,%rdx), %VEC(6)
> -       VMOVU   -(VEC_SIZE * 4)(%rsi,%rdx), %VEC(7)
> +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VEC(4)
> +       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VEC(5)
> +       VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VEC(6)
> +       VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VEC(7)
>         VMOVU   %VEC(0), (%rdi)
>         VMOVU   %VEC(1), VEC_SIZE(%rdi)
>         VMOVU   %VEC(2), (VEC_SIZE * 2)(%rdi)
>         VMOVU   %VEC(3), (VEC_SIZE * 3)(%rdi)
> -       VMOVU   %VEC(4), -VEC_SIZE(%rdi,%rdx)
> -       VMOVU   %VEC(5), -(VEC_SIZE * 2)(%rdi,%rdx)
> -       VMOVU   %VEC(6), -(VEC_SIZE * 3)(%rdi,%rdx)
> -       VMOVU   %VEC(7), -(VEC_SIZE * 4)(%rdi,%rdx)
> +       VMOVU   %VEC(4), -VEC_SIZE(%rdi, %rdx)
> +       VMOVU   %VEC(5), -(VEC_SIZE * 2)(%rdi, %rdx)
> +       VMOVU   %VEC(6), -(VEC_SIZE * 3)(%rdi, %rdx)
> +       VMOVU   %VEC(7), -(VEC_SIZE * 4)(%rdi, %rdx)
>         VZEROUPPER_RETURN
> +       /* Align if doesn't cost too much code size. 6 bytes so that after
> +          jump to target a full mov instruction will always be able to be
> +          fetched.  */
> +       .p2align 4,, 6
>  L(last_4x_vec):
> -       /* Copy from 2 * VEC + 1 to 4 * VEC, inclusively. */
> +       /* Copy from 2 * VEC + 1 to 4 * VEC, inclusively.  */
>         VMOVU   (%rsi), %VEC(0)
>         VMOVU   VEC_SIZE(%rsi), %VEC(1)
> -       VMOVU   -VEC_SIZE(%rsi,%rdx), %VEC(2)
> -       VMOVU   -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(3)
> +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VEC(2)
> +       VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VEC(3)
>         VMOVU   %VEC(0), (%rdi)
>         VMOVU   %VEC(1), VEC_SIZE(%rdi)
> -       VMOVU   %VEC(2), -VEC_SIZE(%rdi,%rdx)
> -       VMOVU   %VEC(3), -(VEC_SIZE * 2)(%rdi,%rdx)
> +       VMOVU   %VEC(2), -VEC_SIZE(%rdi, %rdx)
> +       VMOVU   %VEC(3), -(VEC_SIZE * 2)(%rdi, %rdx)
> +       /* Keep nop target close to jmp for 2-byte encoding.  */
> +L(nop):
>         VZEROUPPER_RETURN
> -
> +       /* Align if doesn't cost too much code size.  */
> +       .p2align 4,, 10
>  L(more_8x_vec):
>         /* Check if non-temporal move candidate.  */
>  #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc)
>         /* Check non-temporal store threshold.  */
> -       cmp __x86_shared_non_temporal_threshold(%rip), %RDX_LP
> +       cmp     __x86_shared_non_temporal_threshold(%rip), %RDX_LP
>         ja      L(large_memcpy_2x)
>  #endif
> -       /* Entry if rdx is greater than non-temporal threshold but there
> -       is overlap.  */
> +       /* Entry if rdx is greater than non-temporal threshold but there is
> +          overlap.  */
>  L(more_8x_vec_check):
>         cmpq    %rsi, %rdi
>         ja      L(more_8x_vec_backward)
>         /* Source == destination is less common.  */
>         je      L(nop)
> +       /* Entry if rdx is greater than movsb or stop movsb threshold but
> +          there is overlap with dst > src.  */
> +L(more_8x_vec_forward):
>         /* Load the first VEC and last 4 * VEC to support overlapping
>            addresses.  */
>         VMOVU   (%rsi), %VEC(4)
> @@ -477,22 +610,18 @@ L(more_8x_vec_check):
>         VMOVU   -(VEC_SIZE * 2)(%rsi, %rdx), %VEC(6)
>         VMOVU   -(VEC_SIZE * 3)(%rsi, %rdx), %VEC(7)
>         VMOVU   -(VEC_SIZE * 4)(%rsi, %rdx), %VEC(8)
> -       /* Save start and stop of the destination buffer.  */
> -       movq    %rdi, %r11
> -       leaq    -VEC_SIZE(%rdi, %rdx), %rcx
> -       /* Align destination for aligned stores in the loop.  Compute
> -          how much destination is misaligned.  */
> -       movq    %rdi, %r8
> -       andq    $(VEC_SIZE - 1), %r8
> -       /* Get the negative of offset for alignment.  */
> -       subq    $VEC_SIZE, %r8
> -       /* Adjust source.  */
> -       subq    %r8, %rsi
> -       /* Adjust destination which should be aligned now.  */
> -       subq    %r8, %rdi
> -       /* Adjust length.  */
> -       addq    %r8, %rdx
> -
> +       /* Subtract dst from src. Add back after dst aligned.  */
> +       subq    %rdi, %rsi
> +       /* Store end of buffer minus tail in rdx.  */
> +       leaq    (VEC_SIZE * -4)(%rdi, %rdx), %rdx
> +       /* Save begining of dst.  */
> +       movq    %rdi, %rcx
> +       /* Align dst to VEC_SIZE - 1.  */
> +       orq     $(VEC_SIZE - 1), %rdi
> +       /* Restore src adjusted with new value for aligned dst.  */
> +       leaq    1(%rdi, %rsi), %rsi
> +       /* Finish aligning dst.  */
> +       incq    %rdi
>         .p2align 4
>  L(loop_4x_vec_forward):
>         /* Copy 4 * VEC a time forward.  */
> @@ -501,23 +630,27 @@ L(loop_4x_vec_forward):
>         VMOVU   (VEC_SIZE * 2)(%rsi), %VEC(2)
>         VMOVU   (VEC_SIZE * 3)(%rsi), %VEC(3)
>         subq    $-(VEC_SIZE * 4), %rsi
> -       addq    $-(VEC_SIZE * 4), %rdx
>         VMOVA   %VEC(0), (%rdi)
>         VMOVA   %VEC(1), VEC_SIZE(%rdi)
>         VMOVA   %VEC(2), (VEC_SIZE * 2)(%rdi)
>         VMOVA   %VEC(3), (VEC_SIZE * 3)(%rdi)
>         subq    $-(VEC_SIZE * 4), %rdi
> -       cmpq    $(VEC_SIZE * 4), %rdx
> +       cmpq    %rdi, %rdx
>         ja      L(loop_4x_vec_forward)
>         /* Store the last 4 * VEC.  */
> -       VMOVU   %VEC(5), (%rcx)
> -       VMOVU   %VEC(6), -VEC_SIZE(%rcx)
> -       VMOVU   %VEC(7), -(VEC_SIZE * 2)(%rcx)
> -       VMOVU   %VEC(8), -(VEC_SIZE * 3)(%rcx)
> +       VMOVU   %VEC(5), (VEC_SIZE * 3)(%rdx)
> +       VMOVU   %VEC(6), (VEC_SIZE * 2)(%rdx)
> +       VMOVU   %VEC(7), VEC_SIZE(%rdx)
> +       VMOVU   %VEC(8), (%rdx)
>         /* Store the first VEC.  */
> -       VMOVU   %VEC(4), (%r11)
> +       VMOVU   %VEC(4), (%rcx)
> +       /* Keep nop target close to jmp for 2-byte encoding.  */
> +L(nop2):
>         VZEROUPPER_RETURN
> -
> +       /* Entry from fail movsb. Need to test if dst - src == 0 still.  */
> +L(more_8x_vec_backward_check_nop):
> +       testq   %rcx, %rcx
> +       jz      L(nop2)
>  L(more_8x_vec_backward):
>         /* Load the first 4 * VEC and last VEC to support overlapping
>            addresses.  */
> @@ -525,49 +658,50 @@ L(more_8x_vec_backward):
>         VMOVU   VEC_SIZE(%rsi), %VEC(5)
>         VMOVU   (VEC_SIZE * 2)(%rsi), %VEC(6)
>         VMOVU   (VEC_SIZE * 3)(%rsi), %VEC(7)
> -       VMOVU   -VEC_SIZE(%rsi,%rdx), %VEC(8)
> -       /* Save stop of the destination buffer.  */
> -       leaq    -VEC_SIZE(%rdi, %rdx), %r11
> -       /* Align destination end for aligned stores in the loop.  Compute
> -          how much destination end is misaligned.  */
> -       leaq    -VEC_SIZE(%rsi, %rdx), %rcx
> -       movq    %r11, %r9
> -       movq    %r11, %r8
> -       andq    $(VEC_SIZE - 1), %r8
> -       /* Adjust source.  */
> -       subq    %r8, %rcx
> -       /* Adjust the end of destination which should be aligned now.  */
> -       subq    %r8, %r9
> -       /* Adjust length.  */
> -       subq    %r8, %rdx
> -
> -       .p2align 4
> +       VMOVU   -VEC_SIZE(%rsi, %rdx), %VEC(8)
> +       /* Subtract dst from src. Add back after dst aligned.  */
> +       subq    %rdi, %rsi
> +       /* Save begining of buffer.  */
> +       movq    %rdi, %rcx
> +       /* Set dst to begining of region to copy. -1 for inclusive
> +          alignment.  */
> +       leaq    (VEC_SIZE * -4 + -1)(%rdi, %rdx), %rdi
> +       /* Align dst.  */
> +       andq    $-(VEC_SIZE), %rdi
> +       /* Restore src.  */
> +       addq    %rdi, %rsi
> +       /* Don't use multi-byte nop to align.  */
> +       .p2align 4,, 11
>  L(loop_4x_vec_backward):
>         /* Copy 4 * VEC a time backward.  */
> -       VMOVU   (%rcx), %VEC(0)
> -       VMOVU   -VEC_SIZE(%rcx), %VEC(1)
> -       VMOVU   -(VEC_SIZE * 2)(%rcx), %VEC(2)
> -       VMOVU   -(VEC_SIZE * 3)(%rcx), %VEC(3)
> -       addq    $-(VEC_SIZE * 4), %rcx
> -       addq    $-(VEC_SIZE * 4), %rdx
> -       VMOVA   %VEC(0), (%r9)
> -       VMOVA   %VEC(1), -VEC_SIZE(%r9)
> -       VMOVA   %VEC(2), -(VEC_SIZE * 2)(%r9)
> -       VMOVA   %VEC(3), -(VEC_SIZE * 3)(%r9)
> -       addq    $-(VEC_SIZE * 4), %r9
> -       cmpq    $(VEC_SIZE * 4), %rdx
> -       ja      L(loop_4x_vec_backward)
> +       VMOVU   (VEC_SIZE * 3)(%rsi), %VEC(0)
> +       VMOVU   (VEC_SIZE * 2)(%rsi), %VEC(1)
> +       VMOVU   (VEC_SIZE * 1)(%rsi), %VEC(2)
> +       VMOVU   (VEC_SIZE * 0)(%rsi), %VEC(3)
> +       addq    $(VEC_SIZE * -4), %rsi
> +       VMOVA   %VEC(0), (VEC_SIZE * 3)(%rdi)
> +       VMOVA   %VEC(1), (VEC_SIZE * 2)(%rdi)
> +       VMOVA   %VEC(2), (VEC_SIZE * 1)(%rdi)
> +       VMOVA   %VEC(3), (VEC_SIZE * 0)(%rdi)
> +       addq    $(VEC_SIZE * -4), %rdi
> +       cmpq    %rdi, %rcx
> +       jb      L(loop_4x_vec_backward)
>         /* Store the first 4 * VEC.  */
> -       VMOVU   %VEC(4), (%rdi)
> -       VMOVU   %VEC(5), VEC_SIZE(%rdi)
> -       VMOVU   %VEC(6), (VEC_SIZE * 2)(%rdi)
> -       VMOVU   %VEC(7), (VEC_SIZE * 3)(%rdi)
> +       VMOVU   %VEC(4), (%rcx)
> +       VMOVU   %VEC(5), VEC_SIZE(%rcx)
> +       VMOVU   %VEC(6), (VEC_SIZE * 2)(%rcx)
> +       VMOVU   %VEC(7), (VEC_SIZE * 3)(%rcx)
>         /* Store the last VEC.  */
> -       VMOVU   %VEC(8), (%r11)
> +       VMOVU   %VEC(8), -VEC_SIZE(%rdx, %rcx)
>         VZEROUPPER_RETURN
>
>  #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc)
>         .p2align 4
> +       /* Entry if dst > stop movsb threshold (usually set to non-temporal
> +          threshold).  */
> +L(large_memcpy_2x_check):
> +       cmp     __x86_shared_non_temporal_threshold(%rip), %RDX_LP
> +       jb      L(more_8x_vec_forward)
>  L(large_memcpy_2x):
>         /* Compute absolute value of difference between source and
>            destination.  */
> --
> 2.25.1
>
>
diff mbox series

Patch

diff --git a/sysdeps/x86/sysdep.h b/sysdeps/x86/sysdep.h
index cac1d762fb..9226d2c6c9 100644
--- a/sysdeps/x86/sysdep.h
+++ b/sysdeps/x86/sysdep.h
@@ -78,15 +78,18 @@  enum cf_protection_level
 #define ASM_SIZE_DIRECTIVE(name) .size name,.-name;
 
 /* Define an entry point visible from C.  */
-#define	ENTRY(name)							      \
-  .globl C_SYMBOL_NAME(name);						      \
-  .type C_SYMBOL_NAME(name),@function;					      \
-  .align ALIGNARG(4);							      \
+#define	P2ALIGN_ENTRY(name, alignment)							      \
+  .globl C_SYMBOL_NAME(name);							      \
+  .type C_SYMBOL_NAME(name),@function;							      \
+  .align ALIGNARG(alignment);							      \
   C_LABEL(name)								      \
   cfi_startproc;							      \
-  _CET_ENDBR;								      \
+  _CET_ENDBR;							      \
   CALL_MCOUNT
 
+#define	ENTRY(name) P2ALIGN_ENTRY(name, 4)
+
+
 #undef	END
 #define END(name)							      \
   cfi_endproc;								      \
diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
index 9f02624375..75b6efe969 100644
--- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
@@ -165,6 +165,32 @@ 
 # error Invalid LARGE_LOAD_SIZE
 #endif
 
+/* Whether to align before movsb. Ultimately we want 64 byte align
+   and not worth it to load 4x VEC for VEC_SIZE == 16.  */
+#define ALIGN_MOVSB	(VEC_SIZE	>	16)
+
+/* Number of VECs to align movsb to.  */
+#if VEC_SIZE == 64
+# define MOVSB_ALIGN_TO	(VEC_SIZE)
+#else
+# define MOVSB_ALIGN_TO	(VEC_SIZE	*	2)
+#endif
+
+/* Macro for copying inclusive power of 2 range with two register
+   loads.  */
+#define COPY_BLOCK(mov_inst, src_reg, dst_reg, size_reg, len, tmp_reg0, tmp_reg1)	\
+	mov_inst (%src_reg), %tmp_reg0; \
+	mov_inst -(len)(%src_reg, %size_reg), %tmp_reg1; \
+	mov_inst %tmp_reg0, (%dst_reg); \
+	mov_inst %tmp_reg1, -(len)(%dst_reg, %size_reg);
+
+/* Define all copies used by L(less_vec) for VEC_SIZE of 16, 32, or
+   64.  */
+#define COPY_4_8	COPY_BLOCK(movl, rsi, rdi, rdx, 4, ecx, esi)
+#define COPY_8_16	COPY_BLOCK(movq, rsi, rdi, rdx, 8, rcx, rsi)
+#define COPY_16_32	COPY_BLOCK(vmovdqu, rsi, rdi, rdx, 16, xmm0, xmm1)
+#define COPY_32_64	COPY_BLOCK(vmovdqu64, rsi, rdi, rdx, 32, ymm16, ymm17)
+
 #ifndef SECTION
 # error SECTION is not defined!
 #endif
@@ -198,7 +224,13 @@  L(start):
 	movl	%edx, %edx
 # endif
 	cmp	$VEC_SIZE, %RDX_LP
+	/* Based on SPEC2017 distribution both 16 and 32 memcpy calls are
+	   really hot so we want them to take the same branch path.  */
+#if VEC_SIZE > 16
+	jbe	L(less_vec)
+#else
 	jb	L(less_vec)
+#endif
 	cmp	$(VEC_SIZE * 2), %RDX_LP
 	ja	L(more_2x_vec)
 #if !defined USE_MULTIARCH || !IS_IN (libc)
@@ -206,15 +238,10 @@  L(last_2x_vec):
 #endif
 	/* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE.  */
 	VMOVU	(%rsi), %VEC(0)
-	VMOVU	-VEC_SIZE(%rsi,%rdx), %VEC(1)
+	VMOVU	-VEC_SIZE(%rsi, %rdx), %VEC(1)
 	VMOVU	%VEC(0), (%rdi)
-	VMOVU	%VEC(1), -VEC_SIZE(%rdi,%rdx)
-#if !defined USE_MULTIARCH || !IS_IN (libc)
-L(nop):
-	ret
-#else
+	VMOVU	%VEC(1), -VEC_SIZE(%rdi, %rdx)
 	VZEROUPPER_RETURN
-#endif
 #if defined USE_MULTIARCH && IS_IN (libc)
 END (MEMMOVE_SYMBOL (__memmove, unaligned))
 
@@ -289,7 +316,9 @@  ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms))
 END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms))
 # endif
 
-ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms))
+/* Cache align entry so that branch heavy L(less_vec) maintains good
+   alignment.  */
+P2ALIGN_ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms), 6)
 	movq	%rdi, %rax
 L(start_erms):
 # ifdef __ILP32__
@@ -297,123 +326,217 @@  L(start_erms):
 	movl	%edx, %edx
 # endif
 	cmp	$VEC_SIZE, %RDX_LP
+	/* Based on SPEC2017 distribution both 16 and 32 memcpy calls are
+	   really hot so we want them to take the same branch path.  */
+# if VEC_SIZE > 16
+	jbe	L(less_vec)
+# else
 	jb	L(less_vec)
+# endif
 	cmp	$(VEC_SIZE * 2), %RDX_LP
 	ja	L(movsb_more_2x_vec)
 L(last_2x_vec):
-	/* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE. */
+	/* From VEC and to 2 * VEC.  No branch when size == VEC_SIZE.  */
 	VMOVU	(%rsi), %VEC(0)
-	VMOVU	-VEC_SIZE(%rsi,%rdx), %VEC(1)
+	VMOVU	-VEC_SIZE(%rsi, %rdx), %VEC(1)
 	VMOVU	%VEC(0), (%rdi)
-	VMOVU	%VEC(1), -VEC_SIZE(%rdi,%rdx)
+	VMOVU	%VEC(1), -VEC_SIZE(%rdi, %rdx)
 L(return):
-#if VEC_SIZE > 16
+# if VEC_SIZE > 16
 	ZERO_UPPER_VEC_REGISTERS_RETURN
-#else
+# else
 	ret
+# endif
 #endif
+#if VEC_SIZE == 64
+L(copy_8_15):
+	COPY_8_16
+	ret
 
-L(movsb):
-	cmp     __x86_rep_movsb_stop_threshold(%rip), %RDX_LP
-	jae	L(more_8x_vec)
-	cmpq	%rsi, %rdi
-	jb	1f
-	/* Source == destination is less common.  */
-	je	L(nop)
-	leaq	(%rsi,%rdx), %r9
-	cmpq	%r9, %rdi
-	/* Avoid slow backward REP MOVSB.  */
-	jb	L(more_8x_vec_backward)
-# if AVOID_SHORT_DISTANCE_REP_MOVSB
-	andl	$X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB, __x86_string_control(%rip)
-	jz	3f
-	movq	%rdi, %rcx
-	subq	%rsi, %rcx
-	jmp	2f
-# endif
-1:
-# if AVOID_SHORT_DISTANCE_REP_MOVSB
-	andl	$X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB, __x86_string_control(%rip)
-	jz	3f
-	movq	%rsi, %rcx
-	subq	%rdi, %rcx
-2:
-/* Avoid "rep movsb" if RCX, the distance between source and destination,
-   is N*4GB + [1..63] with N >= 0.  */
-	cmpl	$63, %ecx
-	jbe	L(more_2x_vec)	/* Avoid "rep movsb" if ECX <= 63.  */
-3:
-# endif
-	mov	%RDX_LP, %RCX_LP
-	rep movsb
-L(nop):
+L(copy_33_63):
+	COPY_32_64
 	ret
 #endif
-
+	/* Only worth aligning if near end of 16 byte block and won't get
+	   first branch in first decode after jump.  */
+	.p2align 4,, 6
 L(less_vec):
-	/* Less than 1 VEC.  */
 #if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64
 # error Unsupported VEC_SIZE!
 #endif
-#if VEC_SIZE > 32
-	cmpb	$32, %dl
-	jae	L(between_32_63)
+	/* Second set of branches for smallest copies.  */
+	cmpl	$(VEC_SIZE / 4), %edx
+	jb	L(less_quarter_vec)
+
+	cmpl	$(VEC_SIZE / 2), %edx
+#if VEC_SIZE == 64
+	/* We branch to [33, 63] instead of [16, 32] to give [16, 32] fall
+	   through path as [16, 32] is hotter.  */
+	ja	L(copy_33_63)
+	COPY_16_32
+#elif VEC_SIZE == 32
+	/* Branch to [8, 15]. Fall through to [16, 32].  */
+	jb	L(copy_8_15)
+	COPY_16_32
+#else
+	/* Branch to [4, 7]. Fall through to [8, 15].  */
+	jb	L(copy_4_7)
+	COPY_8_16
 #endif
-#if VEC_SIZE > 16
-	cmpb	$16, %dl
-	jae	L(between_16_31)
-#endif
-	cmpb	$8, %dl
-	jae	L(between_8_15)
-	cmpb	$4, %dl
-	jae	L(between_4_7)
-	cmpb	$1, %dl
-	ja	L(between_2_3)
-	jb	1f
+	ret
+	/* Align if won't cost too many bytes.  */
+	.p2align 4,, 6
+L(copy_4_7):
+	COPY_4_8
+	ret
+
+	/* Cold target. No need to align.  */
+L(copy_1):
 	movzbl	(%rsi), %ecx
 	movb	%cl, (%rdi)
-1:
 	ret
+
+	/* Colder copy case for [0, VEC_SIZE / 4 - 1].  */
+L(less_quarter_vec):
 #if VEC_SIZE > 32
-L(between_32_63):
-	/* From 32 to 63.  No branch when size == 32.  */
-	VMOVU	(%rsi), %YMM0
-	VMOVU	-32(%rsi,%rdx), %YMM1
-	VMOVU	%YMM0, (%rdi)
-	VMOVU	%YMM1, -32(%rdi,%rdx)
-	VZEROUPPER_RETURN
+	cmpl	$8, %edx
+	jae	L(copy_8_15)
 #endif
 #if VEC_SIZE > 16
-	/* From 16 to 31.  No branch when size == 16.  */
-L(between_16_31):
-	VMOVU	(%rsi), %XMM0
-	VMOVU	-16(%rsi,%rdx), %XMM1
-	VMOVU	%XMM0, (%rdi)
-	VMOVU	%XMM1, -16(%rdi,%rdx)
-	VZEROUPPER_RETURN
-#endif
-L(between_8_15):
-	/* From 8 to 15.  No branch when size == 8.  */
-	movq	-8(%rsi,%rdx), %rcx
-	movq	(%rsi), %rsi
-	movq	%rcx, -8(%rdi,%rdx)
-	movq	%rsi, (%rdi)
-	ret
-L(between_4_7):
-	/* From 4 to 7.  No branch when size == 4.  */
-	movl	-4(%rsi,%rdx), %ecx
-	movl	(%rsi), %esi
-	movl	%ecx, -4(%rdi,%rdx)
-	movl	%esi, (%rdi)
+	cmpl	$4, %edx
+	jae	L(copy_4_7)
+#endif
+	cmpl	$1, %edx
+	je	L(copy_1)
+	jb	L(copy_0)
+	/* Fall through into copy [2, 3] as it is more common than [0, 1].
+	 */
+	movzwl	(%rsi), %ecx
+	movzbl	-1(%rsi, %rdx), %esi
+	movw	%cx, (%rdi)
+	movb	%sil, -1(%rdi, %rdx)
+L(copy_0):
 	ret
-L(between_2_3):
-	/* From 2 to 3.  No branch when size == 2.  */
-	movzwl	-2(%rsi,%rdx), %ecx
-	movzwl	(%rsi), %esi
-	movw	%cx, -2(%rdi,%rdx)
-	movw	%si, (%rdi)
+
+	.p2align 4
+#if VEC_SIZE == 32
+L(copy_8_15):
+	COPY_8_16
 	ret
+	/* COPY_8_16 is exactly 17 bytes so don't want to p2align after as
+	   it wastes 15 bytes of code and 1 byte off is fine.  */
+#endif
+
+#if defined USE_MULTIARCH && IS_IN (libc)
+L(movsb):
+	movq	%rdi, %rcx
+	subq	%rsi, %rcx
+	/* Go to backwards temporal copy if overlap no matter what as
+	   backward movsb is slow.  */
+	cmpq	%rdx, %rcx
+	/* L(more_8x_vec_backward_check_nop) checks for src == dst.  */
+	jb	L(more_8x_vec_backward_check_nop)
+	/* If above __x86_rep_movsb_stop_threshold most likely is candidate
+	   for NT moves aswell.  */
+	cmp	__x86_rep_movsb_stop_threshold(%rip), %RDX_LP
+	jae	L(large_memcpy_2x_check)
+# if ALIGN_MOVSB
+	VMOVU	(%rsi), %VEC(0)
+#  if MOVSB_ALIGN_TO > VEC_SIZE
+	VMOVU	VEC_SIZE(%rsi), %VEC(1)
+#  endif
+#  if MOVSB_ALIGN_TO > (VEC_SIZE * 2)
+#   error Unsupported MOVSB_ALIGN_TO
+#  endif
+	/* Store dst for use after rep movsb.  */
+	movq	%rdi, %r8
+# endif
+# if AVOID_SHORT_DISTANCE_REP_MOVSB
+	/* Only avoid short movsb if CPU has FSRM.  */
+	testl	$X86_STRING_CONTROL_AVOID_SHORT_DISTANCE_REP_MOVSB, __x86_string_control(%rip)
+	jz	L(skip_short_movsb_check)
+	/* Avoid "rep movsb" if RCX, the distance between source and
+	   destination, is N*4GB + [1..63] with N >= 0.  */
+
+	/* ecx contains dst - src. Early check for backward copy conditions
+	   means only case of slow movsb with src = dst + [0, 63] is ecx in
+	   [-63, 0]. Use unsigned comparison with -64 check for that case.  */
+	cmpl	$-64, %ecx
+	ja	L(more_8x_vec_forward)
+# endif
+# if ALIGN_MOVSB
+	/* Fall through means cpu has FSRM. In that case exclusively align
+	   destination.  */
+
+	/* Subtract dst from src. Add back after dst aligned.  */
+	subq	%rdi, %rsi
+	/* Add dst to len. Subtract back after dst aligned.  */
+	leaq	(%rdi, %rdx), %rcx
+	/* Exclusively align dst to MOVSB_ALIGN_TO (64).  */
+	addq	$(MOVSB_ALIGN_TO - 1), %rdi
+	andq	$-(MOVSB_ALIGN_TO), %rdi
+	/* Restore src and len adjusted with new values for aligned dst.  */
+	addq	%rdi, %rsi
+	subq	%rdi, %rcx
 
+	rep	movsb
+	VMOVU	%VEC(0), (%r8)
+#  if MOVSB_ALIGN_TO > VEC_SIZE
+	VMOVU	%VEC(1), VEC_SIZE(%r8)
+#  endif
+	VZEROUPPER_RETURN
+L(movsb_align_dst):
+	/* Subtract dst from src. Add back after dst aligned.  */
+	subq	%rdi, %rsi
+	/* Add dst to len. Subtract back after dst aligned. -1 because dst
+	   is initially aligned to MOVSB_ALIGN_TO - 1.  */
+	leaq	-(1)(%rdi, %rdx), %rcx
+	/* Inclusively align dst to MOVSB_ALIGN_TO - 1.  */
+	orq	$(MOVSB_ALIGN_TO - 1), %rdi
+	leaq	1(%rdi, %rsi), %rsi
+	/* Restore src and len adjusted with new values for aligned dst.  */
+	subq	%rdi, %rcx
+	/* Finish aligning dst.  */
+	incq	%rdi
+	rep	movsb
+	VMOVU	%VEC(0), (%r8)
+#  if MOVSB_ALIGN_TO > VEC_SIZE
+	VMOVU	%VEC(1), VEC_SIZE(%r8)
+#  endif
+	VZEROUPPER_RETURN
+
+L(skip_short_movsb_check):
+	/* If CPU does not have FSRM two options for aligning. Align src if
+	   dst and src 4k alias. Otherwise align dst.  */
+	testl	$(PAGE_SIZE - 512), %ecx
+	jnz	L(movsb_align_dst)
+	/* rcx already has dst - src.  */
+	movq	%rcx, %r9
+	/* Add src to len. Subtract back after src aligned. -1 because src
+	   is initially aligned to MOVSB_ALIGN_TO - 1.  */
+	leaq	-(1)(%rsi, %rdx), %rcx
+	/* Inclusively align src to MOVSB_ALIGN_TO - 1.  */
+	orq	$(MOVSB_ALIGN_TO - 1), %rsi
+	/* Restore dst and len adjusted with new values for aligned dst.  */
+	leaq	1(%rsi, %r9), %rdi
+	subq	%rsi, %rcx
+	/* Finish aligning src.  */
+	incq	%rsi
+	rep	movsb
+	VMOVU	%VEC(0), (%r8)
+#  if MOVSB_ALIGN_TO > VEC_SIZE
+	VMOVU	%VEC(1), VEC_SIZE(%r8)
+#  endif
+	VZEROUPPER_RETURN
+# else
+	/* Not alignined rep movsb so just copy.  */
+	mov	%RDX_LP, %RCX_LP
+	rep	movsb
+	ret
+# endif
+#endif
+	/* Align if doesn't cost too many bytes.  */
+	.p2align 4,, 6
 #if defined USE_MULTIARCH && IS_IN (libc)
 L(movsb_more_2x_vec):
 	cmp	__x86_rep_movsb_threshold(%rip), %RDX_LP
@@ -426,50 +549,60 @@  L(more_2x_vec):
 	ja	L(more_8x_vec)
 	cmpq	$(VEC_SIZE * 4), %rdx
 	jbe	L(last_4x_vec)
-	/* Copy from 4 * VEC + 1 to 8 * VEC, inclusively. */
+	/* Copy from 4 * VEC + 1 to 8 * VEC, inclusively.  */
 	VMOVU	(%rsi), %VEC(0)
 	VMOVU	VEC_SIZE(%rsi), %VEC(1)
 	VMOVU	(VEC_SIZE * 2)(%rsi), %VEC(2)
 	VMOVU	(VEC_SIZE * 3)(%rsi), %VEC(3)
-	VMOVU	-VEC_SIZE(%rsi,%rdx), %VEC(4)
-	VMOVU	-(VEC_SIZE * 2)(%rsi,%rdx), %VEC(5)
-	VMOVU	-(VEC_SIZE * 3)(%rsi,%rdx), %VEC(6)
-	VMOVU	-(VEC_SIZE * 4)(%rsi,%rdx), %VEC(7)
+	VMOVU	-VEC_SIZE(%rsi, %rdx), %VEC(4)
+	VMOVU	-(VEC_SIZE * 2)(%rsi, %rdx), %VEC(5)
+	VMOVU	-(VEC_SIZE * 3)(%rsi, %rdx), %VEC(6)
+	VMOVU	-(VEC_SIZE * 4)(%rsi, %rdx), %VEC(7)
 	VMOVU	%VEC(0), (%rdi)
 	VMOVU	%VEC(1), VEC_SIZE(%rdi)
 	VMOVU	%VEC(2), (VEC_SIZE * 2)(%rdi)
 	VMOVU	%VEC(3), (VEC_SIZE * 3)(%rdi)
-	VMOVU	%VEC(4), -VEC_SIZE(%rdi,%rdx)
-	VMOVU	%VEC(5), -(VEC_SIZE * 2)(%rdi,%rdx)
-	VMOVU	%VEC(6), -(VEC_SIZE * 3)(%rdi,%rdx)
-	VMOVU	%VEC(7), -(VEC_SIZE * 4)(%rdi,%rdx)
+	VMOVU	%VEC(4), -VEC_SIZE(%rdi, %rdx)
+	VMOVU	%VEC(5), -(VEC_SIZE * 2)(%rdi, %rdx)
+	VMOVU	%VEC(6), -(VEC_SIZE * 3)(%rdi, %rdx)
+	VMOVU	%VEC(7), -(VEC_SIZE * 4)(%rdi, %rdx)
 	VZEROUPPER_RETURN
+	/* Align if doesn't cost too much code size. 6 bytes so that after
+	   jump to target a full mov instruction will always be able to be
+	   fetched.  */
+	.p2align 4,, 6
 L(last_4x_vec):
-	/* Copy from 2 * VEC + 1 to 4 * VEC, inclusively. */
+	/* Copy from 2 * VEC + 1 to 4 * VEC, inclusively.  */
 	VMOVU	(%rsi), %VEC(0)
 	VMOVU	VEC_SIZE(%rsi), %VEC(1)
-	VMOVU	-VEC_SIZE(%rsi,%rdx), %VEC(2)
-	VMOVU	-(VEC_SIZE * 2)(%rsi,%rdx), %VEC(3)
+	VMOVU	-VEC_SIZE(%rsi, %rdx), %VEC(2)
+	VMOVU	-(VEC_SIZE * 2)(%rsi, %rdx), %VEC(3)
 	VMOVU	%VEC(0), (%rdi)
 	VMOVU	%VEC(1), VEC_SIZE(%rdi)
-	VMOVU	%VEC(2), -VEC_SIZE(%rdi,%rdx)
-	VMOVU	%VEC(3), -(VEC_SIZE * 2)(%rdi,%rdx)
+	VMOVU	%VEC(2), -VEC_SIZE(%rdi, %rdx)
+	VMOVU	%VEC(3), -(VEC_SIZE * 2)(%rdi, %rdx)
+	/* Keep nop target close to jmp for 2-byte encoding.  */
+L(nop):
 	VZEROUPPER_RETURN
-
+	/* Align if doesn't cost too much code size.  */
+	.p2align 4,, 10
 L(more_8x_vec):
 	/* Check if non-temporal move candidate.  */
 #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc)
 	/* Check non-temporal store threshold.  */
-	cmp __x86_shared_non_temporal_threshold(%rip), %RDX_LP
+	cmp	__x86_shared_non_temporal_threshold(%rip), %RDX_LP
 	ja	L(large_memcpy_2x)
 #endif
-	/* Entry if rdx is greater than non-temporal threshold but there
-       is overlap.  */
+	/* Entry if rdx is greater than non-temporal threshold but there is
+	   overlap.  */
 L(more_8x_vec_check):
 	cmpq	%rsi, %rdi
 	ja	L(more_8x_vec_backward)
 	/* Source == destination is less common.  */
 	je	L(nop)
+	/* Entry if rdx is greater than movsb or stop movsb threshold but
+	   there is overlap with dst > src.  */
+L(more_8x_vec_forward):
 	/* Load the first VEC and last 4 * VEC to support overlapping
 	   addresses.  */
 	VMOVU	(%rsi), %VEC(4)
@@ -477,22 +610,18 @@  L(more_8x_vec_check):
 	VMOVU	-(VEC_SIZE * 2)(%rsi, %rdx), %VEC(6)
 	VMOVU	-(VEC_SIZE * 3)(%rsi, %rdx), %VEC(7)
 	VMOVU	-(VEC_SIZE * 4)(%rsi, %rdx), %VEC(8)
-	/* Save start and stop of the destination buffer.  */
-	movq	%rdi, %r11
-	leaq	-VEC_SIZE(%rdi, %rdx), %rcx
-	/* Align destination for aligned stores in the loop.  Compute
-	   how much destination is misaligned.  */
-	movq	%rdi, %r8
-	andq	$(VEC_SIZE - 1), %r8
-	/* Get the negative of offset for alignment.  */
-	subq	$VEC_SIZE, %r8
-	/* Adjust source.  */
-	subq	%r8, %rsi
-	/* Adjust destination which should be aligned now.  */
-	subq	%r8, %rdi
-	/* Adjust length.  */
-	addq	%r8, %rdx
-
+	/* Subtract dst from src. Add back after dst aligned.  */
+	subq	%rdi, %rsi
+	/* Store end of buffer minus tail in rdx.  */
+	leaq	(VEC_SIZE * -4)(%rdi, %rdx), %rdx
+	/* Save begining of dst.  */
+	movq	%rdi, %rcx
+	/* Align dst to VEC_SIZE - 1.  */
+	orq	$(VEC_SIZE - 1), %rdi
+	/* Restore src adjusted with new value for aligned dst.  */
+	leaq	1(%rdi, %rsi), %rsi
+	/* Finish aligning dst.  */
+	incq	%rdi
 	.p2align 4
 L(loop_4x_vec_forward):
 	/* Copy 4 * VEC a time forward.  */
@@ -501,23 +630,27 @@  L(loop_4x_vec_forward):
 	VMOVU	(VEC_SIZE * 2)(%rsi), %VEC(2)
 	VMOVU	(VEC_SIZE * 3)(%rsi), %VEC(3)
 	subq	$-(VEC_SIZE * 4), %rsi
-	addq	$-(VEC_SIZE * 4), %rdx
 	VMOVA	%VEC(0), (%rdi)
 	VMOVA	%VEC(1), VEC_SIZE(%rdi)
 	VMOVA	%VEC(2), (VEC_SIZE * 2)(%rdi)
 	VMOVA	%VEC(3), (VEC_SIZE * 3)(%rdi)
 	subq	$-(VEC_SIZE * 4), %rdi
-	cmpq	$(VEC_SIZE * 4), %rdx
+	cmpq	%rdi, %rdx
 	ja	L(loop_4x_vec_forward)
 	/* Store the last 4 * VEC.  */
-	VMOVU	%VEC(5), (%rcx)
-	VMOVU	%VEC(6), -VEC_SIZE(%rcx)
-	VMOVU	%VEC(7), -(VEC_SIZE * 2)(%rcx)
-	VMOVU	%VEC(8), -(VEC_SIZE * 3)(%rcx)
+	VMOVU	%VEC(5), (VEC_SIZE * 3)(%rdx)
+	VMOVU	%VEC(6), (VEC_SIZE * 2)(%rdx)
+	VMOVU	%VEC(7), VEC_SIZE(%rdx)
+	VMOVU	%VEC(8), (%rdx)
 	/* Store the first VEC.  */
-	VMOVU	%VEC(4), (%r11)
+	VMOVU	%VEC(4), (%rcx)
+	/* Keep nop target close to jmp for 2-byte encoding.  */
+L(nop2):
 	VZEROUPPER_RETURN
-
+	/* Entry from fail movsb. Need to test if dst - src == 0 still.  */
+L(more_8x_vec_backward_check_nop):
+	testq	%rcx, %rcx
+	jz	L(nop2)
 L(more_8x_vec_backward):
 	/* Load the first 4 * VEC and last VEC to support overlapping
 	   addresses.  */
@@ -525,49 +658,50 @@  L(more_8x_vec_backward):
 	VMOVU	VEC_SIZE(%rsi), %VEC(5)
 	VMOVU	(VEC_SIZE * 2)(%rsi), %VEC(6)
 	VMOVU	(VEC_SIZE * 3)(%rsi), %VEC(7)
-	VMOVU	-VEC_SIZE(%rsi,%rdx), %VEC(8)
-	/* Save stop of the destination buffer.  */
-	leaq	-VEC_SIZE(%rdi, %rdx), %r11
-	/* Align destination end for aligned stores in the loop.  Compute
-	   how much destination end is misaligned.  */
-	leaq	-VEC_SIZE(%rsi, %rdx), %rcx
-	movq	%r11, %r9
-	movq	%r11, %r8
-	andq	$(VEC_SIZE - 1), %r8
-	/* Adjust source.  */
-	subq	%r8, %rcx
-	/* Adjust the end of destination which should be aligned now.  */
-	subq	%r8, %r9
-	/* Adjust length.  */
-	subq	%r8, %rdx
-
-	.p2align 4
+	VMOVU	-VEC_SIZE(%rsi, %rdx), %VEC(8)
+	/* Subtract dst from src. Add back after dst aligned.  */
+	subq	%rdi, %rsi
+	/* Save begining of buffer.  */
+	movq	%rdi, %rcx
+	/* Set dst to begining of region to copy. -1 for inclusive
+	   alignment.  */
+	leaq	(VEC_SIZE * -4 + -1)(%rdi, %rdx), %rdi
+	/* Align dst.  */
+	andq	$-(VEC_SIZE), %rdi
+	/* Restore src.  */
+	addq	%rdi, %rsi
+	/* Don't use multi-byte nop to align.  */
+	.p2align 4,, 11
 L(loop_4x_vec_backward):
 	/* Copy 4 * VEC a time backward.  */
-	VMOVU	(%rcx), %VEC(0)
-	VMOVU	-VEC_SIZE(%rcx), %VEC(1)
-	VMOVU	-(VEC_SIZE * 2)(%rcx), %VEC(2)
-	VMOVU	-(VEC_SIZE * 3)(%rcx), %VEC(3)
-	addq	$-(VEC_SIZE * 4), %rcx
-	addq	$-(VEC_SIZE * 4), %rdx
-	VMOVA	%VEC(0), (%r9)
-	VMOVA	%VEC(1), -VEC_SIZE(%r9)
-	VMOVA	%VEC(2), -(VEC_SIZE * 2)(%r9)
-	VMOVA	%VEC(3), -(VEC_SIZE * 3)(%r9)
-	addq	$-(VEC_SIZE * 4), %r9
-	cmpq	$(VEC_SIZE * 4), %rdx
-	ja	L(loop_4x_vec_backward)
+	VMOVU	(VEC_SIZE * 3)(%rsi), %VEC(0)
+	VMOVU	(VEC_SIZE * 2)(%rsi), %VEC(1)
+	VMOVU	(VEC_SIZE * 1)(%rsi), %VEC(2)
+	VMOVU	(VEC_SIZE * 0)(%rsi), %VEC(3)
+	addq	$(VEC_SIZE * -4), %rsi
+	VMOVA	%VEC(0), (VEC_SIZE * 3)(%rdi)
+	VMOVA	%VEC(1), (VEC_SIZE * 2)(%rdi)
+	VMOVA	%VEC(2), (VEC_SIZE * 1)(%rdi)
+	VMOVA	%VEC(3), (VEC_SIZE * 0)(%rdi)
+	addq	$(VEC_SIZE * -4), %rdi
+	cmpq	%rdi, %rcx
+	jb	L(loop_4x_vec_backward)
 	/* Store the first 4 * VEC.  */
-	VMOVU	%VEC(4), (%rdi)
-	VMOVU	%VEC(5), VEC_SIZE(%rdi)
-	VMOVU	%VEC(6), (VEC_SIZE * 2)(%rdi)
-	VMOVU	%VEC(7), (VEC_SIZE * 3)(%rdi)
+	VMOVU	%VEC(4), (%rcx)
+	VMOVU	%VEC(5), VEC_SIZE(%rcx)
+	VMOVU	%VEC(6), (VEC_SIZE * 2)(%rcx)
+	VMOVU	%VEC(7), (VEC_SIZE * 3)(%rcx)
 	/* Store the last VEC.  */
-	VMOVU	%VEC(8), (%r11)
+	VMOVU	%VEC(8), -VEC_SIZE(%rdx, %rcx)
 	VZEROUPPER_RETURN
 
 #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc)
 	.p2align 4
+	/* Entry if dst > stop movsb threshold (usually set to non-temporal
+	   threshold).  */
+L(large_memcpy_2x_check):
+	cmp	__x86_shared_non_temporal_threshold(%rip), %RDX_LP
+	jb	L(more_8x_vec_forward)
 L(large_memcpy_2x):
 	/* Compute absolute value of difference between source and
 	   destination.  */