[Bug target/11877] gcc should use xor trick with -Os

[Bug target/11877] gcc should use xor trick with -Os

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Roger Sayle <sayle@gcc.gnu.org>:

https://gcc.gnu.org/g:9cedbaab8e048b90ceb9ceef0d851385fae67cde

commit r12-1668-g9cedbaab8e048b90ceb9ceef0d851385fae67cde
Author: Roger Sayle <roger@nextmovesoftware.com>
Date:   Mon Jun 21 08:54:50 2021 +0100

    PR target/11877: Use xor to write zero to memory with -Os

    The following patch attempts to resolve PR target/11877 (without
    triggering PR/23102).  On x86_64, writing an SImode or DImode zero
    to memory uses an instruction encoding that is larger than first
    clearing a register (using xor) then writing that to memory.  Hence,
    after reload, the peephole2 pass can determine if there's a suitable
    free register, and if so, use that to shrink the code size with -Os.

    To improve code size, and avoid inserting a large number of xor
    instructions (PR target/23102), this patch makes use of peephole2's
    efficient pattern matching to use a single temporary for a run of
    consecutive writes.  In theory, one could do better still with a
    new target-specific pass, gated on -Os, to shrink these instructions
    (like stv), but that's probably overkill for the little remaining
    space savings.

    Evaluating this patch on the CSiBE benchmark (v2.1.1) results in a
    0.26% code size improvement (3715273 bytes down to 3705477) on x86_64
    with -Os [saving 1 byte every 400].  549 of 894 tests improve, two
    tests grow larger.  Analysis of these 2 pathological cases reveals
    that although peephole2's match_scratch prefers to use a call-clobbered
    register (to avoid requiring a new stack frame), very rarely this
    interacts with GCC's shrink wrapping optimization, which may previously
    have avoided saving/restoring a call clobbered register, such as %eax,
    in the calling function.

    2021-06-21  Roger Sayle  <roger@nextmovesoftware.com>

    gcc/ChangeLog
            PR target/11877
            * config/i386/i386.md: New define_peephole2s to shrink writing
            1, 2 or 4 consecutive zeros to memory when optimizing for size.

    gcc/testsuite/ChangeLog
            PR target/11877
            * gcc.target/i386/pr11877.c: New test case.

[Bug target/11877] gcc should use xor trick with -Os

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:d58a66aa0faa64bfbd85e528be5104293dd41d0e

commit r12-1712-gd58a66aa0faa64bfbd85e528be5104293dd41d0e
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Jun 22 10:16:18 2021 +0200

    i386: Use xor to write zero to memory with -Os even for more than 4 stores
[PR11877]

    > > 2021-06-20  Roger Sayle  <roger@nextmovesoftware.com>
    > >
    > > gcc/ChangeLog
    > >         PR target/11877
    > >         * config/i386/i386.md: New define_peephole2s to shrink writing
    > >         1, 2 or 4 consecutive zeros to memory when optimizing for size.

    It unfortunately doesn't extend well to larger memory clearing.
    Consider e.g.
    void
    foo (int *p)
    {
      p[0] = 0;
      p[7] = 0;
      p[23] = 0;
      p[41] = 0;
      p[48] = 0;
      p[59] = 0;
      p[69] = 0;
      p[78] = 0;
      p[83] = 0;
      p[89] = 0;
      p[98] = 0;
      p[121] = 0;
      p[132] = 0;
      p[143] = 0;
      p[154] = 0;
    }
    where with the patch we emit:
            xorl    %eax, %eax
            xorl    %edx, %edx
            xorl    %ecx, %ecx
            xorl    %esi, %esi
            xorl    %r8d, %r8d
            movl    %eax, (%rdi)
            movl    %eax, 28(%rdi)
            movl    %eax, 92(%rdi)
            movl    %eax, 164(%rdi)
            movl    %edx, 192(%rdi)
            movl    %edx, 236(%rdi)
            movl    %edx, 276(%rdi)
            movl    %edx, 312(%rdi)
            movl    %ecx, 332(%rdi)
            movl    %ecx, 356(%rdi)
            movl    %ecx, 392(%rdi)
            movl    %ecx, 484(%rdi)
            movl    %esi, 528(%rdi)
            movl    %esi, 572(%rdi)
            movl    %r8d, 616(%rdi)
    Here is an incremental patch that emits:
            xorl    %eax, %eax
            movl    %eax, (%rdi)
            movl    %eax, 28(%rdi)
            movl    %eax, 92(%rdi)
            movl    %eax, 164(%rdi)
            movl    %eax, 192(%rdi)
            movl    %eax, 236(%rdi)
            movl    %eax, 276(%rdi)
            movl    %eax, 312(%rdi)
            movl    %eax, 332(%rdi)
            movl    %eax, 356(%rdi)
            movl    %eax, 392(%rdi)
            movl    %eax, 484(%rdi)
            movl    %eax, 528(%rdi)
            movl    %eax, 572(%rdi)
            movl    %eax, 616(%rdi)
    instead.

    2021-06-22  Jakub Jelinek  <jakub@redhat.com>

            PR target/11877
            * config/i386/i386-protos.h (ix86_last_zero_store_uid): Declare.
            * config/i386/i386-expand.c (ix86_last_zero_store_uid): New
variable.
            * config/i386/i386.c (ix86_expand_prologue): Clear it.
            * config/i386/i386.md (peephole2s for 1/2/4 stores of const0_rtx):
            Remove "" from match_operand.  Emit new insns using emit_move_insn
and
            set ix86_last_zero_store_uid to INSN_UID of the last store.
            Add peephole2s for 1/2/4 stores of const0_rtx following previous
            successful peep2s.

            * gcc.target/i386/pr11877-2.c: New test.

[Bug target/11877] gcc should use xor trick with -Os

[roger at nextmovesoftware dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
                 CC|                            |roger at nextmovesoftware dot com
             Status|ASSIGNED                    |RESOLVED
   Target Milestone|---                         |12.0

--- Comment #12 from Roger Sayle <roger at nextmovesoftware dot com> ---
Fixed on mainline thanks to Jakub's patch.

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |msharov at users dot sourceforge.n
                   |                            |et

--- Comment #13 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 49127 has been marked as a duplicate of this bug. ***

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jeffreyalaw at gmail dot com

--- Comment #14 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 41505 has been marked as a duplicate of this bug. ***

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |andi-gcc at firstfloor dot org

--- Comment #15 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 32629 has been marked as a duplicate of this bug. ***

[Bug target/11877] gcc should use xor trick with -Os

[dann at godzilla dot ics dot uci dot edu]

------- Comment #9 from dann at godzilla dot ics dot uci dot edu  2006-01-05 20:22 -------
(In reply to comment #7)
> *** Bug 23338 has been marked as a duplicate of this bug. ***
> 

Bug 23338 contained a patch that might fixed this issue. Here it is, so
that it can be evaluated.


*** i386.md     08 Aug 2005 16:38:37 -0700      1.652
--- i386.md     11 Aug 2005 11:27:11 -0700      
***************
*** 18874,18881 ****
    [(match_scratch:SI 1 "r")
     (set (match_operand:SI 0 "memory_operand" "")
          (const_int 0))]
!   "! optimize_size
!    && ! TARGET_USE_MOV0
     && TARGET_SPLIT_LONG_MOVES
     && get_attr_length (insn) >= ix86_cost->large_insn
     && peep2_regno_dead_p (0, FLAGS_REG)"
--- 18874,18880 ----
    [(match_scratch:SI 1 "r")
     (set (match_operand:SI 0 "memory_operand" "")
          (const_int 0))]
!   "! TARGET_USE_MOV0
     && TARGET_SPLIT_LONG_MOVES
     && get_attr_length (insn) >= ix86_cost->large_insn
     && peep2_regno_dead_p (0, FLAGS_REG)"


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-08-12 05:27 -------
*** Bug 23338 has been marked as a duplicate of this bug. ***

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dann at godzilla dot ics dot
                   |                            |uci dot edu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-08-12 05:28 -------
PR 23102 is the bug for multiple xors.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  BugsThisDependsOn|                            |23102


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[kazu at cs dot umass dot edu]

------- Additional Comments From kazu at cs dot umass dot edu  2004-03-25 19:16 -------
Even if you split the long move in ix86_expand_move,
the constant 0 is propagated into the two moves.
I guess the right way may be uncse sometime after register allocation.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[kazu at cs dot umass dot edu]

------- Additional Comments From kazu at cs dot umass dot edu  2004-02-01 17:24 -------
Then we need something like an un-cse pass.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-01-04 07:50 -------
What about expanding (set (mem:DI ...) (const_int 0)) at expand time, this will cause more 
opportunities to happen and then the discusion is up to other parts of the compiler.
It looks like an easy change to ix86_expand_move.
Also interesting is that this testcase:
void
foo (long *p)
{
  *p = 0;
  p[1] = 0;
}
Does not use the xor trick either.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[kazu at cs dot umass dot edu]

------- Additional Comments From kazu at cs dot umass dot edu  2004-01-04 06:55 -------
Patch posted:

http://gcc.gnu.org/ml/gcc-patches/2004-01/msg00153.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[kazu at cs dot umass dot edu]

------- Additional Comments From kazu at cs dot umass dot edu  2003-12-31 21:19 -------
(set (mem:DI ...) (const_int 0)) is split into two moves in SImode after reload.
We could delay the split until after peephole2.
In peephole2, if a scratch reg is available,
load 0 into it with XOR and then copy that reg to two mem:SI locations.

Reduced to:

void
foo (long long *p)
{
  *p = 0;
}

The reduction from 13 bytes down to 7 bytes sounds impressive.
My proposed solution would still leave two XORs, though.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kazu at cs dot umass dot edu
         AssignedTo|unassigned at gcc dot gnu   |kazu at cs dot umass dot edu
                   |dot org                     |
             Status|NEW                         |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877

[Bug target/11877] gcc should use xor trick with -Os

[dhazeghi at yahoo dot com]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877


dhazeghi at yahoo dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  GCC build triplet|386-linux                   |i386-linux
 GCC target triplet|386-linux                   |i386-linux
   Target Milestone|3.4                         |---

[Bug target/11877] gcc should use xor trick with -Os

[pinskia at gcc dot gnu dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11877


pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2003-08-10 16:03:50
               date|                            |


------- Additional Comments From pinskia at gcc dot gnu dot org  2003-08-10 16:03 -------
I can confirm that the xor trick will be a size win.

The movl still happens with the mainline (20030810).