[Bug rtl-optimization/56151] [4.8 Regression] Performance degradation after r194054 on x86 Atom.

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56151

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-02-03
                 CC|steven at gcc dot gnu.org   |
     Ever Confirmed|0                           |1

--- Comment #3 from Steven Bosscher <steven at gcc dot gnu.org> 2013-02-03 00:45:37 UTC ---
Confirmed.

My patch only avoids situations that ought not to happen in the first
place: self-referencing REG_EQUAL notes.  If this somehow prevents an
interesting code transformation I don't consider it my responsibility
to find out what that might be. Therefore I'm not taking this bug.

I did look at the difference in generated code. The regression is due
to the optabs.c change, reverting it results in the pre-r194054 code
again. The difference in code from the 'expand' pass is:

    86: NOTE_INSN_BASIC_BLOCK 13
    87: {r130:SI=r60:SI>>0x5;clobber flags:CC;}
    88: {r131:SI=r130:SI<<0x2;clobber flags:CC;}
    89: {r86:SI=r103:SI+r131:SI;clobber flags:CC;}
    90: {r132:SI=r60:SI&0x1f;clobber flags:CC;}
    91: r133:SI=`setmask'
-   92: r134:SI=[r132:SI*0x4+r133:SI]
-   93: {[r86:SI]=[r86:SI]|r134:SI;clobber flags:CC;}
+   94: r136:SI=[r132:SI*0x4+r133:SI]
+   95: r137:SI=[r86:SI]
+   96: {r135:SI=r137:SI|r136:SI;clobber flags:CC;}
+   97: [r86:SI]=r135:SI
       REG_EQUAL [r86:SI]|[r132:SI*0x4+r133:SI]
    // next basic block

The '+' code is the post-patch code, '-' is pre-patch (the continuous
numbering is due to the '-' code being rejected post-past, so those two
insn UIDs are lost).

Post-patch the extra moves of insns 95 and 97 are never cleaned up. The
extra move is introduced to make room for the REG_EQUAL note that is 
self-referencing in the pre-patch situation (SET of r86 with REG_EQUAL
note using r86). So the pre-patch RTL was indeed invalid.

Combine fails to restore the code to its pre-patch state because it
combines:

   94: r136:SI=[r132:SI*0x4+`setmask']
   95: r137:SI=[r86:SI]
   96: {r135:SI=r137:SI|r136:SI;clobber flags:CC;}

to:

   95: r137:SI=[r86:SI]
   96: {r135:SI=r137:SI|[r132:SI*0x4+`setmask'];clobber flags:CC;}

and the 'setmask' reference make merging the SET of r135 and the store
impossible:

Trying 96 -> 97:
Failed to match this instruction:
(parallel [
        (set (mem:SI (reg/f:SI 86 [ D.1803 ]) [2 *_41+0 S4 A32])
            (ior:SI (reg:SI 137 [ *_41 ])
                (mem:SI (plus:SI (mult:SI (reg:SI 132 [ D.1801 ])
                            (const_int 4 [0x4]))
                        (symbol_ref:SI ("setmask") <var_decl 0x7ffff63fb390
setmask>)) [2 setmaskD.1732 S4 A32])))
        (clobber (reg:CC 17 flags))
    ])
Failed to match this instruction:
(set (mem:SI (reg/f:SI 86 [ D.1803 ]) [2 *_41+0 S4 A32])
    (ior:SI (reg:SI 137 [ *_41 ])
        (mem:SI (plus:SI (mult:SI (reg:SI 132 [ D.1801 ])
                    (const_int 4 [0x4]))
                (symbol_ref:SI ("setmask") <var_decl 0x7ffff63fb390 setmask>))
[2 setmaskD.1732 S4 A32])))