[Bug rtl-optimization/31799] New: Failed to optimize out test instruction

[Bug rtl-optimization/31799] New: Failed to optimize out test instruction

[hjl at lucon dot org]

For

void
foo (int x, int *y, int *z)
{
  *z = ++x;
  if (x != 0)
  *y = 1;
}

gcc 4.3 generates

foo:
.LFB2:
        addl    $1, %edi
        testl   %edi, %edi
        movl    %edi, (%rdx)
        je      .L3
        movl    $1, (%rsi)
.L3:
        rep ; ret

But "testl   %edi, %edi" isn't needed since "addl    $1, %edi" sets proper
FLAGS_REG.


-- 
           Summary: Failed to optimize out test instruction
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hjl at lucon dot org
GCC target triplet: x86_64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31799

[Bug rtl-optimization/31799] Failed to optimize out test instruction

[ubizjak at gmail dot com]

------- Comment #1 from ubizjak at gmail dot com  2007-05-03 17:08 -------
(In reply to comment #0)

> foo:
>         addl    $1, %edi
>         testl   %edi, %edi
>         movl    %edi, (%rdx)
>         je      .L3
>         movl    $1, (%rsi)
> .L3:
>         rep ; ret

This is just an unfortunate corner case. During combine phase we get:

(insn 10 6 11 2 (parallel [
            (set (reg/v:SI 58 [ x.27 ])
                (plus:SI (reg/v:SI 59 [ x ])
                    (const_int 1 [0x1])))
            (clobber (reg:CC 17 flags))
        ]) 210 {*addsi_1} (insn_list:REG_DEP_TRUE 3 (nil))
    (expr_list:REG_DEAD (reg/v:SI 59 [ x ])
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))

(insn 11 10 13 2 (set (mem:SI (reg/v/f:DI 61 [ z ]) [2 S4 A32])
        (reg/v:SI 58 [ x.27 ])) 40 {*movsi_1} (insn_list:REG_DEP_TRUE 5
(insn_list:REG_DEP_TRUE 10 (nil)))
    (expr_list:REG_DEAD (reg/v/f:DI 61 [ z ])
        (nil)))

(insn 13 11 14 2 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg/v:SI 58 [ x.27 ])
            (const_int -1 [0xffffffffffffffff]))) 5 {*cmpsi_1_insn} (nil)
    (expr_list:REG_DEAD (reg/v:SI 58 [ x.27 ])
        (nil)))


It looks that insn 10 and insn 13 don't combine into flags setting insn due to
insn 11 interfering in some way. If the code is changed to:

void
foo (int x, int *y, int *z)
{
  if (++x != 0)
    *y = 1;
  *z = x;
}

then gcc generates expected asm without test insn.


-- 

ubizjak at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-05-03 17:08:39
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31799

[Bug rtl-optimization/31799] Failed to optimize out test instruction

[ubizjak at gmail dot com]

------- Comment #2 from ubizjak at gmail dot com  2007-05-03 17:22 -------
Ah, for some reason insn 13 does not have the insn link pointing to insn 10.
Probably dataflow should/could fix this?


-- 

ubizjak at gmail dot com changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31799

[Bug rtl-optimization/31799] Failed to optimize out test instruction

[hjl at lucon dot org]

------- Comment #3 from hjl at lucon dot org  2007-05-05 02:31 -------
The problem may be patterns like

 [(set (match_operand:SI 0 "nonimmediate_operand" "=r,rm,r")
        (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0,r")
                 (match_operand:SI 2 "general_operand" "rmni,rni,lni")))
   (clobber (reg:CC FLAGS_REG))]

What if we replace (clobber (reg:CC FLAGS_REG)) with something more
reasonable like

(set (reg FLAGS_REG)
        (compare
          (plus:SI (match_dup 1) (match_dup 2))
          (const_int 0)))


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31799

[Bug rtl-optimization/31799] Failed to optimize out test instruction

[ubizjak at gmail dot com]

------- Comment #4 from ubizjak at gmail dot com  2007-05-05 18:49 -------
(In reply to comment #3)
> The problem may be patterns like

No, the patterns are OK. Combine pass merges these patterns without problems.
But without links, combine doesn't even try to merge them.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31799