public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug rtl-optimization/56175] New: Issue with combine phase on x86.
@ 2013-02-01 15:52 ysrumyan at gmail dot com
  2013-02-01 15:59 ` [Bug rtl-optimization/56175] " ysrumyan at gmail dot com
                   ` (17 more replies)
  0 siblings, 18 replies; 19+ messages in thread
From: ysrumyan at gmail dot com @ 2013-02-01 15:52 UTC (permalink / raw)
  To: gcc-bugs


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56175

             Bug #: 56175
           Summary: Issue with combine phase on x86.
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: ysrumyan@gmail.com


Analyzing performance of important benchmark on x86 Atom in 32bit mode we found
out that the code produced for attached testcase is not optimal - the inner
loop contains 18 instructions instead of 12.
The problem is that 'combine' does not perform desired substitution for the
following stmt:

    t = (u8)((x & 1) ^ ((u8)y & 1));
It is not able to convert it to more optimal form like:
    t = (u8)((x ^ (u8)y ) & 1);

This issue can be explained using the following testcase:

int foo( unsigned char x, unsigned short y)
{
  unsigned char z;
 if (x ==0 || y == 0)
  return 0;
 x>>=1;
 y>>=1;
  z = (unsigned char)((x & 1) ^ ((unsigned char)y & 1));
  if (z == 1)
    return 1;
  return 0;
}

For this case combine performs needed transformation and we get optimal
assembly:
...
    xorl    %edx, %eax
    andl    $1, %eax
    ret

For this case combine tries to perform the following substitution:

Trying 22, 20 -> 23:
Failed to match this instruction:
(parallel [
        (set (reg:QI 83 [ D.1758 ])
            (and:QI (xor:QI (reg:QI 79 [ x ])
                    (subreg:QI (reg:HI 81 [ y ]) 0))
                (const_int 1 [0x1])))
        (clobber (reg:CC 17 flags))
    ])
Failed to match this instruction:
(set (reg:QI 83 [ D.1758 ])
    (and:QI (xor:QI (reg:QI 79 [ x ])
            (subreg:QI (reg:HI 81 [ y ]) 0))
        (const_int 1 [0x1])))
Successfully matched this instruction:
(set (reg:QI 82 [ D.1760 ])
    (xor:QI (reg:QI 79 [ x ])
        (subreg:QI (reg:HI 81 [ y ]) 0)))
Successfully matched this instruction:
(set (reg:QI 83 [ D.1758 ])
    (and:QI (reg:QI 82 [ D.1760 ])
        (const_int 1 [0x1])))
where
(insn 20 19 21 4 (parallel [
            (set (reg:QI 80 [ D.1759 ])
                (and:QI (reg:QI 79 [ x ])
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) t.c:8 405 {*andqi_1}
     (expr_list:REG_DEAD (reg:QI 79 [ x ])
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))
(insn 22 21 23 4 (parallel [
            (set (reg:HI 82 [ D.1760 ])
                (and:HI (reg:HI 81 [ y ])
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) t.c:8 404 {*andhi_1}
     (expr_list:REG_DEAD (reg:HI 81 [ y ])
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))
(insn 23 22 24 4 (parallel [
            (set (reg:QI 83 [ D.1758 ])
                (xor:QI (reg:QI 80 [ D.1759 ])
                    (subreg:QI (reg:HI 82 [ D.1760 ]) 0)))
            (clobber (reg:CC 17 flags))
        ]) t.c:8 426 {*xorqi_1}
     (expr_list:REG_DEAD (reg:HI 82 [ D.1760 ])
        (expr_list:REG_DEAD (reg:QI 80 [ D.1759 ])
            (expr_list:REG_UNUSED (reg:CC 17 flags)
                (nil)))))
but for more compicated test that is attached combine tries to do the same
substitution in reverse order of operands and it is failed:

Trying 14, 13 -> 15:
Failed to match this instruction:
(parallel [
        (set (reg:QI 63 [ D.1770 ])
            (xor:QI (and:QI (reg/v:QI 72 [ x ])
                    (const_int 1 [0x1]))
                (and:QI (subreg:QI (reg/v:HI 74 [ y ]) 0)
                    (const_int 1 [0x1]))))
        (clobber (reg:CC 17 flags))
    ])
Failed to match this instruction:
(set (reg:QI 63 [ D.1770 ])
    (xor:QI (and:QI (reg/v:QI 72 [ x ])
            (const_int 1 [0x1]))
        (and:QI (subreg:QI (reg/v:HI 74 [ y ]) 0)
            (const_int 1 [0x1]))))
Successfully matched this instruction:
(set (reg:QI 77 [ D.1771 ])
    (and:QI (subreg:QI (reg/v:HI 74 [ y ]) 0)
        (const_int 1 [0x1])))
Failed to match this instruction:
(set (reg:QI 63 [ D.1770 ])
    (xor:QI (and:QI (reg/v:QI 72 [ x ])
            (const_int 1 [0x1]))
        (reg:QI 77 [ D.1771 ])))
where
(insn 13 12 14 3 (parallel [
            (set (reg:HI 76 [ D.1772 ])
                (and:HI (reg/v:HI 74 [ y ])
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) t1.c:9 404 {*andhi_1}
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))
(insn 14 13 15 3 (parallel [
            (set (reg:QI 77 [ D.1771 ])
                (and:QI (reg/v:QI 72 [ x ])
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) t1.c:9 405 {*andqi_1}
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))
(insn 15 14 16 3 (parallel [
            (set (reg:QI 63 [ D.1770 ])
                (xor:QI (reg:QI 77 [ D.1771 ])
                    (subreg:QI (reg:HI 76 [ D.1772 ]) 0)))
            (clobber (reg:CC 17 flags))
        ]) t1.c:9 426 {*xorqi_1}
     (expr_list:REG_DEAD (reg:QI 77 [ D.1771 ])
        (expr_list:REG_DEAD (reg:HI 76 [ D.1772 ])
            (expr_list:REG_UNUSED (reg:CC 17 flags)
                (nil)))))
It seems that if we tried to combine 13, 14 -> 15 we will be successful.
Note also that an order of instructions is different after expand.


^ permalink raw reply	[flat|nested] 19+ messages in thread
end of thread, other threads:[~2014-06-12 13:23 UTC | newest]

Thread overview: 19+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2013-02-01 15:52 [Bug rtl-optimization/56175] New: Issue with combine phase on x86 ysrumyan at gmail dot com
2013-02-01 15:59 ` [Bug rtl-optimization/56175] " ysrumyan at gmail dot com
2013-02-02 12:02 ` ubizjak at gmail dot com
2013-02-04  8:55 ` ysrumyan at gmail dot com
2013-02-04 10:11 ` [Bug tree-optimization/56175] " rguenth at gcc dot gnu.org
2013-02-11 13:43 ` ysrumyan at gmail dot com
2013-02-11 14:38 ` rguenth at gcc dot gnu.org
2013-02-12 13:05 ` ysrumyan at gmail dot com
2013-02-12 13:26 ` rguenth at gcc dot gnu.org
2013-02-12 14:44 ` ysrumyan at gmail dot com
2013-02-12 14:47 ` jakub at gcc dot gnu.org
2013-02-14 12:04 ` ysrumyan at gmail dot com
2013-02-21 13:44 ` rguenth at gcc dot gnu.org
2013-02-21 13:48 ` [Bug tree-optimization/56175] [4.7/4.8 Regression] " rguenth at gcc dot gnu.org
2013-02-25 15:32 ` rguenth at gcc dot gnu.org
2013-02-25 16:00 ` [Bug tree-optimization/56175] [4.7 " rguenth at gcc dot gnu.org
2013-04-03  9:51 ` rguenth at gcc dot gnu.org
2013-04-11  8:00 ` rguenth at gcc dot gnu.org
2014-06-12 13:23 ` rguenth at gcc dot gnu.org

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).