[Bug tree-optimization/45256] New: Missed arithmetic simplification at tree level

[Bug tree-optimization/45256] New: Missed arithmetic simplification at tree level

[bernds at gcc dot gnu dot org]

I'll attach a testcase, which shows a missed simplification at tree level:

  D.2276_42 = i_53 + 1;
  D.2277_43 = D.2276_42 * 32;
  iftmp.3_55 = __fswab32 (xb_54);
  __asm__("clz  %0, %1" : "=r" ret_56 : "r" iftmp.3_55 : "cc");
  ret_58 = 32 - ret_56;
  ret_59 = D.2277_43 - ret_58;

In effect, the constant 32 is both added and subtracted from the result.  With
a four-insn combiner, this is caught at the RTL stage (compiling for Thumb-1):

-       add     r2, r2, #1
        lsl     r2, r2, #5
-       add     r3, r3, r2
-       sub     r3, r3, #32
+       add     r3, r2, r3


-- 
           Summary: Missed arithmetic simplification at tree level
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: bernds at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: arm-none-linux-gnueabi


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45256

[Bug tree-optimization/45256] Missed arithmetic simplification at tree level

[bernds at gcc dot gnu dot org]

------- Comment #1 from bernds at gcc dot gnu dot org  2010-08-11 15:19 -------
Created an attachment (id=21454)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=21454&action=view)
Testcase


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45256

[Bug tree-optimization/45256] Missed arithmetic simplification at tree level

[ibolton at gcc dot gnu dot org]

------- Comment #2 from ibolton at gcc dot gnu dot org  2010-09-07 13:05 -------
(In reply to comment #1)
> Created an attachment (id=21454)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=21454&action=view) [edit]
> Testcase
> 

OK.  I can see this happening on trunk, when compiled as follows:

arm-none-linux-gnueabi/obj/gcc3/gcc/cc1 -O2 pr45256.i -fdump-tree-all
-fdump-rtl-all.

The 148t.optimized dump shows exactly what has been posted above, where
effectively 32 is both added and then subtracted.

The 183r.dce dump still shows this extra work happening and the 184r.combine
dump shows that it has been optimised away.  It should be possible to do this
optimisation at the tree-level, so this bug is confirmed.

Should I check whether this happens for 4.4 and 4.5 too?


-- 

ibolton at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
           Keywords|                            |missed-optimization
      Known to fail|                            |4.6.0
   Last reconfirmed|0000-00-00 00:00:00         |2010-09-07 13:05:49
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45256

[Bug tree-optimization/45256] Missed arithmetic simplification at tree level

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2010-09-07 18:41 -------
ret_59 = (i_53 + 1) * 32 - (32 - ret_56)

So this looks like a re-association issue.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45256

[Bug tree-optimization/45256] Missed arithmetic simplification at tree level

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=45256

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
With IVOPTs we get:
  # ivtmp.13_97 = PHI <ivtmp.13_96(12), 1(7)>
  _34 = MEM[(const __be32 *)_28 + -4B + ivtmp.13_97 * 4];
  _36 = MEM[(const __be32 *)_27 + -4B + ivtmp.13_97 * 4];
  if (_34 != _36)
    goto <bb 9>; [5.50%]
  else
    goto <bb 12>; [94.50%]

  <bb 9> [local count: 217424209]:
  _22 = (int) ivtmp.13_97;
  _40 = _22 * 32;
  xb_37 = _34 ^ _36;
  iftmp.1_53 = __fswab32 (xb_37);
  __asm__("clz  %0, %1" : "=r" ret_54 : "r" iftmp.1_53 : "cc");
  ret_56 = 32 - ret_54;
  _57 = _40 - ret_56;
  MEM[(struct ipv6_saddr_score *)&scores + 32B].matchlen = _57;
  if (_57 != 0)
    goto <bb 13>; [0.00%]
  else
    goto <bb 10>; [100.00%]

So there is no extra add one any more.
BUT with -fno-ivopts I still see the add one.