[Bug tree-optimization/96271] New: Failure to optimize memcmp of doubles to avoid going through memory

[Bug tree-optimization/96271] New: Failure to optimize memcmp of doubles to avoid going through memory

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

            Bug ID: 96271
           Summary: Failure to optimize memcmp of doubles to avoid going
                    through memory
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gabravier at gmail dot com
  Target Milestone: ---

bool f(double a, double b)
{
   return memcmp(&a, &b, sizeof(double)) == 0;
}

With -O3, LLVM outputs this :

f(double, double):
  movq rax, xmm0
  movq rcx, xmm1
  cmp rax, rcx
  sete al
  ret

GCC outputs this:

f(double, double):
  movsd QWORD PTR [rsp-8], xmm0
  movsd QWORD PTR [rsp-16], xmm1
  mov rax, QWORD PTR [rsp-16]
  cmp QWORD PTR [rsp-8], rax
  sete al
  ret

[Bug target/96271] Failure to optimize memcmp of doubles to avoid going through memory

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |target

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Did you try -march=intel ?

[Bug target/96271] Failure to optimize memcmp of doubles to avoid going through memory

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

--- Comment #2 from Gabriel Ravier <gabravier at gmail dot com> ---
What do you mean ? -march=intel doesn't exist and I have tried -mtune=intel, I
see no difference in the handling of this regardless of any -march/tune flags
(I've tried : none, -march=nocona, -march=tigerlake, -mtune=intel,
-march=znver2), unless there is somehow a big problem with `movq r64, xmm` on
all x86-64 processors, this should be considered a bug.

[Bug target/96271] Failure to optimize memcmp of doubles to avoid going through memory

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

Hongtao.liu <crazylht at gmail dot com> changed:

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

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
in for testcase.c.267r.dse1

pass_dse1 fail to delete

---
trying to replace DImode load in insn 7 from DFmode store in insn 2
-- could not extract bits of stored value

trying to replace DImode load in insn 8 from DFmode store in insn 3
-- could not extract bits of stored value

...


(insn 2 5 3 2 (set (mem/c:DF (plus:DI (reg/f:DI 19 frame)
                (const_int -8 [0xfffffffffffffff8])) [1 a+0 S8 A64])
        (reg:DF 20 xmm0 [ a ])) "pr96271_double.c":4:1 135 {*movdf_internal}
     (expr_list:REG_DEAD (reg:DF 20 xmm0 [ a ])

...


(insn 7 4 8 2 (set (reg:DI 87 [ MEM <unsigned long> [(char * {ref-all})&a] ])
        (mem/c:DI (plus:DI (reg/f:DI 19 frame)
                (const_int -8 [0xfffffffffffffff8])) [0 MEM <unsigned long>
[(char * {ref-all})&a]+0 S8 A64])) "pr96271_double.c":5:37 74 {*movdi_internal}
     (nil))

...

---

Shouldn't DImode load behave the same with DFmode load?

[Bug tree-optimization/96271] Failure to optimize memcmp of doubles to avoid going through memory

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
          Component|target                      |tree-optimization
   Last reconfirmed|                            |2020-07-22
     Ever confirmed|0                           |1

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note this could be folded on the GIMPLE level already similar to how we
inline string comparisons of size 1.  We should already fold this to
MEMCMP_EQ builtins and those can be inlined for bigger sizes, profitable
when we can emit a single load to a register for both values (thus
power-of-two size up to word_mode at least).  Desired gimple:

  _1 = VIEW_CONVERT <unsigned long> (a);
  _2 = VIEW_CONVERT <unsigned long> (b);
  _3 = _1 != _2;
  return _3;

and memcmp of size 1 can be inlined as two char loads and subtraction.

[Bug tree-optimization/96271] Failure to optimize memcmp of doubles to avoid going through memory

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 49974
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49974&action=edit
gcc11-pr96271.patch

We actually do that at the GIMPLE level, just quite late and don't set
TODO_update_address_taken if we do.

[Bug tree-optimization/96271] Failure to optimize memcmp of doubles to avoid going through memory

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

--- Comment #6 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:b673e7547fb95d1f0d5cd17ae9e3874742cade66

commit r11-6747-gb673e7547fb95d1f0d5cd17ae9e3874742cade66
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Sat Jan 16 09:17:38 2021 +0100

    cd_dce: Return TODO_update_address_taken from last cd_dce [PR96271]

    On the following testcase, handle_builtin_memcmp in the strlen pass folds
    the memcmp into comparison of two MEM_REFs.  But nothing triggers updating
    of addressable vars afterwards, so even when the parameters are no longer
    address taken, we force the parameters to stack and back anyway.

    This patch causes TODO_update_address_taken to happen right before last
forwprop
    pass (at the end of last cd_dce), so after strlen1 too.

    2021-01-16  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/96271
            * passes.def: Pass false argument to first two pass_cd_dce
            instances and true to last instance.  Add comment that
            last instance rewrites no longer addressed locals.
            * tree-ssa-dce.c (pass_cd_dce): Add update_address_taken_p member
and
            initialize it.
            (pass_cd_dce::set_pass_param): New method.
            (pass_cd_dce::execute): Return TODO_update_address_taken from
            last cd_dce instance.

            * gcc.target/i386/pr96271.c: New test.

[Bug tree-optimization/96271] Failure to optimize memcmp of doubles to avoid going through memory

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed.  Probably the last of these missed-optimization bugfixes from me for GCC
11, tomorrow starts stage4 and only regression bugfixes will be allowed.

[Bug tree-optimization/96271] Failure to optimize memcmp of doubles to avoid going through memory

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96271

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
.