[Bug c/101705] New: Missed optimization opportunity when copying lots of bitfields

[Bug c/101705] New: Missed optimization opportunity when copying lots of bitfields

[jengelh at inai dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101705

            Bug ID: 101705
           Summary: Missed optimization opportunity when copying lots of
                    bitfields
           Product: gcc
           Version: 11.1.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jengelh at inai dot de
  Target Milestone: ---

Using gcc-11.1.1 [revision 62bbb113ae68a7e724255e17143520735bcb9ec9], I observe
that gcc is able to recognize and combine lots of "structb->member =
structa->member" assignments into SIMD instructions. However, this only works
as long as "member" has exactly 8*n bits. It would appear gcc is not smart
enough to consider smaller entities (if and when they add up to multiples of 8
or 64).

Observed
========
[On x86_64]

»gcc -O3 -c 1.c; gcc -O3 -c 2.c
»objdump -d 1.o
   0:   f3 0f 6f 3e             movdqu (%rsi),%xmm7
...
»objdump -d 2.o
   0:   0f b6 16                movzbl (%rsi),%edx
   3:   0f b6 07                movzbl (%rdi),%eax
   6:   83 e2 01                and    $0x1,%edx
   9:   83 e0 fe                and    $0xfffffffe,%eax
   c:   09 d0                   or     %edx,%eax
   e:   88 07                   mov    %al,(%rdi)
  10:   0f b6 16                movzbl (%rsi),%edx
  13:   83 e0 fd                and    $0xfffffffd,%eax
  16:   83 e2 02                and    $0x2,%edx
  19:   09 d0                   or     %edx,%eax
  1b:   88 07                   mov    %al,(%rdi)
...

Expected
========
Emit some movdqu/movups even for 2.c.

Other info
==========
gcc version 11.1.1 20210625 [revision 62bbb113ae68a7e724255e17143520735bcb9ec9]
(SUSE Linux) &
gcc version 7.5.0 (SUSE Linux)

[Bug c/101705] Missed optimization opportunity when copying lots of bitfields

[jengelh at inai dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101705

--- Comment #1 from Jan Engelhardt <jengelh at inai dot de> ---
Created attachment 51228
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51228&action=edit
testcase, 1.c

[Bug c/101705] Missed optimization opportunity when copying lots of bitfields

[jengelh at inai dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101705

--- Comment #2 from Jan Engelhardt <jengelh at inai dot de> ---
Created attachment 51229
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51229&action=edit
testcase, 2.c

[Bug tree-optimization/101705] Missed optimization opportunity when copying lots of bitfields

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101705

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-07-31
             Status|UNCONFIRMED                 |ASSIGNED
           Keywords|                            |missed-optimization
     Ever confirmed|0                           |1
          Component|c                           |tree-optimization
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Mine but for GCC 13.  There are other bugs which say the same thing too.

[Bug tree-optimization/101705] Missed optimization opportunity when copying lots of bitfields

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101705

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Blocks|                            |53947

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
SLP vectorization handles this and it even splits the groups and would use
smaller groups for the tail but then it fails:

t.i:244:10: note:   === vect_slp_analyze_instance_alignment ===
t.i:244:10: missed:   desired vector type conflicts with earlier one for _113 =
s_129(D)->b113;
t.i:244:10: note:  removing SLP instance operations starting from:
d_130(D)->b113 = _113;
t.i:252:10: note:   === vect_slp_analyze_instance_alignment ===
t.i:252:10: missed:   desired vector type conflicts with earlier one for _121 =
s_129(D)->b121;
t.i:252:10: note:  removing SLP instance operations starting from:
d_130(D)->b121 = _121;
t.i:256:10: note:   === vect_slp_analyze_instance_alignment ===
t.i:256:10: missed:   desired vector type conflicts with earlier one for _125 =
s_129(D)->b125;
t.i:256:10: note:  removing SLP instance operations starting from:
d_130(D)->b125 = _125;

which is a pre-existing issue (and I think there's a duplicate for the
vectorizer issue).  Not sure what part Andrew wants to look at.  Without
vectorization we don't merge anything.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations