[Bug other/57668] New: Regression in vectorizing memcpy pattern.

[Bug other/57668] New: Regression in vectorizing memcpy pattern.

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57668

            Bug ID: 57668
           Summary: Regression in vectorizing memcpy pattern.
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: other
          Assignee: unassigned at gcc dot gnu.org
          Reporter: neleai at seznam dot cz

Hi,
When I ran atached benchmark that test how gcc can optimize byte by byte
memcpy(attached memcpy_byte.c) I got a regression on nehalem and ivy_bridge
architectures.
I ran it by commands ./run machine 2> machine_result

For ivy bridge results between 4.7 and 4.9 are:

memcpyO3-4.7.s
0.66
0.65
0.64
0.64
0.64
memcpyO3-4.9.s
0.74
0.74
0.73
0.74
0.74

Also when I look at assemblies and 4.9 version is excessively large compared to
4.7 one.

[Bug other/57668] Regression in vectorizing memcpy pattern.

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57668

--- Comment #1 from Ondrej Bilka <neleai at seznam dot cz> ---
Created attachment 30333
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30333&action=edit
benchmark for memcpy

[Bug other/57668] Regression in vectorizing memcpy pattern.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57668

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-06-24
                 CC|                            |hubicka at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.

memcpy_byte.c:2: note: Completely unroll loop 14 times

the new unrolling code unrolls the epilogue loop completely:

Estimating sizes for loop 3
 BB: 8, after_exit: 0
  size:   0 i.0_40 = (sizetype) i_38;
  size:   1 _41 = s_7(D) + i.0_40;
  size:   1 _42 = d_9(D) + i.0_40;
  size:   1 _43 = *_42;
  size:   1 *_41 = _43;
  size:   1 i_45 = i_38 + 1;
  size:   2 if (i_45 < _16)
   Exit condition will be eliminated in last copy.
 BB: 9, after_exit: 1
size: 7-0, last_iteration: 7-2
  Loop size: 7
  Estimated size after unrolling: 68

And loop distribution fails to detect the memcpy clone, too.

[Bug tree-optimization/57668] Regression in vectorizing memcpy pattern.

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57668

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|2013-06-24 00:00:00         |2021-7-19
         Resolution|---                         |FIXED
             Status|NEW                         |RESOLVED
   Target Milestone|---                         |8.3

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed in 8.3 at least for skylake-512 (Intel(R) Xeon(R) D-2166NT CPU @
2.00GHz):
memcpyO3-831.s
0.68
memcpyO3-4.7.s
0.62
memcpyO3-4.9.s
0.78
memcpyO3-11.s
0.63