[Bug tree-optimization/54245] New: [4.8 regression] incorrect optimisation

[Bug tree-optimization/54245] New: [4.8 regression] incorrect optimisation

[mans at mansr dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

             Bug #: 54245
           Summary: [4.8 regression] incorrect optimisation
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: mans@mansr.com


Created attachment 28007
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28007
Test case

Since r190220 the attached test is compiled incorrectly at -O1 and higher.

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |wschmidt at gcc dot gnu.org
   Target Milestone|---                         |4.8.0

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-08-13 17:19:40 UTC ---
Confirmed.  slsr replaces:
  D.2219_3 = *row_2(D);
  D.2220_4 = (int) D.2219_3;
  a1_5 = D.2220_4 * 22725;
  D.2222_6 = MEM[(short int *)row_2(D) + 4B];
  D.2223_7 = (int) D.2222_6;
  D.2224_8 = D.2223_7 * 21407;
  a0_9 = D.2224_8 + a1_5;
  D.2225_10 = D.2223_7 * 8867;
- a1_11 = a1_5 + D.2225_10;
+ slsr.4_25 = D.2222_6 * 12540;
+ slsr.5_26 = (int) slsr.4_25;
+ a1_11 = a0_9 - slsr.5_26;

The multiplication is newly performed in short int, supposedly that is the
problem here.  Anyway, while the number of multiplications in the end is the
same, with slsr the code sequence is also 3 insns/4 bytes longer on x86_64.

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2012-08-13
         AssignedTo|unassigned at gcc dot       |wschmidt at gcc dot gnu.org
                   |gnu.org                     |
     Ever Confirmed|0                           |1

--- Comment #2 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-13 19:29:06 UTC ---
I'll take a look.  Might be a day or two as my queue is kind of full.

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

--- Comment #3 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-14 12:34:24 UTC ---
I'm putting together a for-now patch that disables the optimization when a
widening cast produces the stride.  In the long run this can be re-enabled so
long as we can retain the original cast and base the new multiply on that value
rather than doing it in the smaller type.  This is a bit subtle to get right so
I plan to defer that work until I have more time to concentrate on it.

I.e., there are two problems in this example: we did the multiply in a smaller
type and we introduced an extra cast that wasn't necessary if we were smart
enough (causing the longer sequence as Jakub noted).  The replacement would be
ok with a little re-design.

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

--- Comment #4 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-15 13:27:38 UTC ---
Author: wschmidt
Date: Wed Aug 15 13:27:29 2012
New Revision: 190412

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=190412
Log:
gcc:

2012-08-15  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

    PR tree-optimization/54245
    * gimple-ssa-strength-reduction.c (legal_cast_p_1): New function.
    (legal_cast_p): Split out logic to legal_cast_p_1.
    (analyze_increments): Avoid introducing multiplies in smaller types.


gcc/testsuite:

2012-08-15  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

    PR tree-optimization/54245
    * gcc.dg/tree-ssa/pr54245.c: New test.


Added:
    trunk/gcc/testsuite/gcc.dg/tree-ssa/pr54245.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/gimple-ssa-strength-reduction.c
    trunk/gcc/testsuite/ChangeLog

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #5 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-15 13:28:49 UTC ---
Fixed.

[Bug tree-optimization/54245] [4.8 regression] incorrect optimisation

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54245

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #4 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-15 13:27:38 UTC ---
Author: wschmidt
Date: Wed Aug 15 13:27:29 2012
New Revision: 190412

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=190412
Log:
gcc:

2012-08-15  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

    PR tree-optimization/54245
    * gimple-ssa-strength-reduction.c (legal_cast_p_1): New function.
    (legal_cast_p): Split out logic to legal_cast_p_1.
    (analyze_increments): Avoid introducing multiplies in smaller types.


gcc/testsuite:

2012-08-15  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

    PR tree-optimization/54245
    * gcc.dg/tree-ssa/pr54245.c: New test.


Added:
    trunk/gcc/testsuite/gcc.dg/tree-ssa/pr54245.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/gimple-ssa-strength-reduction.c
    trunk/gcc/testsuite/ChangeLog

--- Comment #5 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-08-15 13:28:49 UTC ---
Fixed.