[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-06-06
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-06-06 10:05:57 UTC ---
Confirmed with -Os on trunk (4.8).  With -O2 we unroll completely to

f:
.LFB0:
        .cfi_startproc
        movq    $0, (%rdi)
        movq    $0, 8(%rdi)
        movq    $0, 16(%rdi)
        movq    $0, 24(%rdi)
        movq    $0, 32(%rdi)
        movq    $0, 40(%rdi)
        movq    $0, 48(%rdi)
        movq    $0, 56(%rdi)
        movq    $0, 64(%rdi)
        movq    $0, 72(%rdi)
        ret

which lacks the size optimization to use a zeroed %rax.  Likewise
for -Os which now looks like

   0:   48 8d 57 30             lea    0x30(%rdi),%rdx
   4:   48 c7 07 00 00 00 00    movq   $0x0,(%rdi)
   b:   48 c7 47 08 00 00 00    movq   $0x0,0x8(%rdi)
  12:   00 
  13:   48 c7 47 10 00 00 00    movq   $0x0,0x10(%rdi)
  1a:   00 
  1b:   48 c7 47 18 00 00 00    movq   $0x0,0x18(%rdi)
  22:   00 
  23:   b9 08 00 00 00          mov    $0x8,%ecx
  28:   48 c7 47 20 00 00 00    movq   $0x0,0x20(%rdi)
  2f:   00 
  30:   48 c7 47 28 00 00 00    movq   $0x0,0x28(%rdi)
  37:   00 
  38:   31 c0                   xor    %eax,%eax
  3a:   48 89 d7                mov    %rdx,%rdi
  3d:   f3 ab                   rep stos %eax,%es:(%rdi)
  3f:   c3                      retq   

I suppose with -Os we use rep stosl because that's one byte smaller ...(?)

I suppose doing the $0x0 optimization should be done post-reload.

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

--- Comment #2 from Jan Hubicka <hubicka at ucw dot cz> 2012-06-09 22:17:07 UTC ---
> I suppose doing the $0x0 optimization should be done post-reload.
I was wondering how to implement this nice for some years already.... I don't
see how this can be done without specialized pass, really, and the interface is
probably going to be bit weird, since it is very weird property of x86
instruction set that there are no stores with short immediate...

Honza

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> 2012-06-11 08:39:45 UTC ---
On Sat, 9 Jun 2012, hubicka at ucw dot cz wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629
> 
> --- Comment #2 from Jan Hubicka <hubicka at ucw dot cz> 2012-06-09 22:17:07 UTC ---
> > I suppose doing the $0x0 optimization should be done post-reload.
> I was wondering how to implement this nice for some years already.... I don't
> see how this can be done without specialized pass, really, and the interface is
> probably going to be bit weird, since it is very weird property of x86
> instruction set that there are no stores with short immediate...

I wonder if we can use a peephole2 and DF information (and update it
on-the-fly).  Thus, when seeing

  mov $0, ...
  mov $0, ...

transform it incrementally to

  xor %eax, %eax
  mov %eax, ...
  mov $0, ...

and then have a 2nd peephole2 with higher priority that looks for
a register with zero content (eh ... that's the interesting part ;))
and do

  xor %eax, %eax
  mov %eax, ...
  mov %eax, ...

I suppose we have the first peephole already - though we'd end up
with

  xor %eax, %eax
  mov %eax, ...
  xor %eax, %eax
  mov %eax, ...

and rely on postreload-cse to clean that up (which isn't run after
the peephole2 in postreload queue ...)

Richard.

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

--- Comment #4 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Yep, the problem with dealing arbitrarily long sequences is something we need
to solve.  Also memcpy/memset ought to use vector moves by itself in these
cases..

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
We could also make $0 a not legitimate constant on x86... (and undo that late
with a peephole2)

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[hubicka at ucw dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

--- Comment #6 from Jan Hubicka <hubicka at ucw dot cz> ---
> We could also make $0 a not legitimate constant on x86... (and undo that late
> with a peephole2)

I tried that in 90's. At that time it increased register pressure and was not
win...

Honza

[Bug rtl-optimization/32629] missing CSE for constant in registers / inefficient memset

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=32629

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |DUPLICATE
             Status|NEW                         |RESOLVED

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
On the trunk at -O2 we have:
        pxor    %xmm0, %xmm0
        movups  %xmm0, (%rdi)
        movups  %xmm0, 16(%rdi)
        movups  %xmm0, 32(%rdi)
        pxor    %xmm0, %xmm0
        movups  %xmm0, 48(%rdi)
        movups  %xmm0, 64(%rdi)
        ret

With -mno-sse we get:
        movq    $0, (%rdi)
        movq    $0, 8(%rdi)
        movq    $0, 16(%rdi)
        movq    $0, 24(%rdi)
        movq    $0, 32(%rdi)
        movq    $0, 40(%rdi)
        movq    $0, 48(%rdi)
        movq    $0, 56(%rdi)
        movq    $0, 64(%rdi)
        movq    $0, 72(%rdi)
        ret

at -Os we get:
        xorl    %eax, %eax
        leaq    48(%rdi), %rdx
        movl    $8, %ecx
        movq    %rax, (%rdi)
        movq    %rax, 8(%rdi)
        movq    %rax, 16(%rdi)
        movq    %rax, 24(%rdi)
        movq    %rax, 32(%rdi)
        movq    %rax, 40(%rdi)
        xorl    %eax, %eax
        movq    %rdx, %rdi

Which was implemented by PR 11877.
So I am going to close this as a dup of bug 11877.

*** This bug has been marked as a duplicate of bug 11877 ***