[Bug tree-optimization/98961] New: Failure to optimize successive comparisons with 0 into clz

[Bug tree-optimization/98961] New: Failure to optimize successive comparisons with 0 into clz

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

            Bug ID: 98961
           Summary: Failure to optimize successive comparisons with 0 into
                    clz
           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: ---

int f(int i, int j)
{
    return i == 0 || j == 0;
}

This can be optimized to `return (__builtin_clz(i) | __builtin_clz(j)) >> 5;`
(if `clz(0)` returns 0). LLVM does this transformation, but GCC does not.

On x86, for example, with lzcnt, while this does not seem to be a net win in
terms of performance (at least, not for this code alone), it also simply not a
loss. As it is a win in terms of code size (which should make it a net win in
most situations), I think that should make it a net win overall in actual code.
See also https://gcc.gnu.org/bugzilla/show_bug.cgi?id=10588

[Bug rtl-optimization/98961] Failure to optimize successive comparisons with 0 into clz

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

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

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

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
IMHO this is more for RTL.

[Bug rtl-optimization/98961] Failure to optimize successive comparisons with 0 into clz

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

Uroš Bizjak <ubizjak at gmail dot com> changed:

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

--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 50124
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50124&action=edit
Experimental x86 patch

Experimental patch for x86 target.

[Bug rtl-optimization/98961] Failure to optimize successive comparisons with 0 into clz

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

--- Comment #3 from Uroš Bizjak <ubizjak at gmail dot com> ---
Please note that LZCNT insn has it own set of problems (e.g.
TARGET_AVOID_FALSE_DEP_FOR_BMI), so I'm not convinced that even:

int z (int i)
{
  return i == 0;
}

benefits from using LZCNT:

   0:   31 c0                   xor    %eax,%eax
   2:   f3 0f bd c7             lzcnt  %edi,%eax
   6:   c1 e8 05                shr    $0x5,%eax
   9:   c3                      retq   

vs:

   0:   31 c0                   xor    %eax,%eax
   2:   85 ff                   test   %edi,%edi
   4:   0f 94 c0                sete   %al
   7:   c3                      retq

[Bug rtl-optimization/98961] Failure to optimize successive comparisons with 0 into clz

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug middle-end/98961] Failure to optimize successive comparisons with 0 into clz

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
          Component|rtl-optimization            |middle-end
   Last reconfirmed|                            |2023-05-20

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed, I think this should happen at expand time and only if the target
does not have conditional compares (e.g. like aarch64).

[Bug middle-end/98961] Failure to optimize successive comparisons with 0 into clz

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98961

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
or could be a cost thing ...