[Bug rtl-optimization/54919] New: [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[Bug rtl-optimization/54919] New: [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

             Bug #: 54919
           Summary: [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c
                    FAILs with -fvariable-expansion-in-unroller
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zsojka@seznam.cz


Created attachment 28439
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28439
preprocessed source

Output:
$ gcc -O2 -ffast-math -funroll-loops -fvariable-expansion-in-unroller
testcase.c
$ ./a.out 
Aborted

What happens is that foo() returns -63 instead of -64.

.L15:
    movapd    xmm0, xmm2    # d, tmp71
    sub    eax, 8    # D.1761,
    subsd    xmm0, xmm1    # d, tmp69
    subsd    xmm0, xmm1    # d, tmp69
    subsd    xmm0, xmm1    # d, tmp69
1==>    movapd    xmm2, xmm0    # tmp71, d
2==>    subsd    xmm2, xmm1    # tmp71, tmp69
    jne    .L15    #,
3==>    addsd    xmm0, xmm2    # d, tmp71
    cvttsd2si    edx, xmm0    # D.1762, d

In each loop, 4 is subtracted, and finally, the sum is multiplied by 2 at (3).

The problem is that when (3) is executed, xmm2==-32 and xmm0==-31 ; so the
final sum is -63.

If (1) and (2) are swapped (and (2) uses xmm0 instead of xmm2), the code works
fine:

.L15:
    movapd    xmm0, xmm2    # d, tmp71
    sub    eax, 8    # D.1761,
    subsd    xmm0, xmm1    # d, tmp69
    subsd    xmm0, xmm1    # d, tmp69
    subsd    xmm0, xmm1    # d, tmp69
2==>    subsd    xmm0, xmm1    # tmp71, tmp69
1==>    movapd    xmm2, xmm0    # tmp71, d
    jne    .L15    #,
3==>    addsd    xmm0, xmm2    # d, tmp71
    cvttsd2si    edx, xmm0    # D.1762, d

Note foo() is inlined into main(), so the change has to be done in main() (or
the noinline attribute has to be added).

Tested revisions:
r192420 - fail
4.7 r191640 - fail
4.6 r191640 - fail
4.5 r191640 - fail
4.4 r191640 - OK

[Bug rtl-optimization/54919] [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-10-13
         AssignedTo|unassigned at gcc dot       |steven at gcc dot gnu.org
                   |gnu.org                     |
     Ever Confirmed|0                           |1

--- Comment #1 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-13 10:38:51 UTC ---
I'll have a look.

[Bug rtl-optimization/54919] [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

--- Comment #2 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-13 16:38:14 UTC ---
At the ".178r.loop2_invariant" dump:

    5 r61:SI=0x40 # 64
    6 r62:DF=[`*.LC0'] # 0.0
   34 r68:DF=[`*.LC1'] # 1.0e+0
L16:
   11 r62:DF=r62:DF-r68:DF
   12 {r61:SI=r61:SI-0x1;clobber flags:CC;}
   13 flags:CCZ=cmp(r61:SI,0)
   14 pc={(flags:CCZ==0)?L19:pc}
   # {implicit: pc=L16} 
L19:
   21 r67:SI=fix(r62:DF)
   26 ax:SI=r67:SI
   29 use ax:SI


Clearly the above loop iterates 64 times. But in the next pass
(".180r.loop2_unroll") the iterations estimate is not 64:

;; *** Considering loop 1 for complete peeling ***

;; Considering peeling once rolling loop
...
Loop 1 is simple:
  simple exit 3 -> 5
  number of iterations: (const_int 63 [0x3f])
  upper bound: 63
  realistic bound: 63
;; Unable to prove that the loop rolls exactly once

;; Considering peeling completely
;; Not peeling loop completely, rolls too much (63 iterations > 16 [maximum
peelings])

;; *** Considering loop 1 ***

;; Considering unrolling loop with constant number of iterations
;; max_unroll 8 (8 copies, initial 8).
;; Decided to unroll the constant times rolling loop, 7 times.

But with the wrong iteration estimate, the unrolled loop size is wrong.

[Bug rtl-optimization/54919] [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org

--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-10-13 18:06:29 UTC ---
I think the dump is right. Number of iterations as defined by loop
infrastructure is number of execution of loopback edges. For loop after loop
header copying has one iteration fewer than one would naturaly assume.
The dump seems right - the constant unrolling unrolls 8x that it reports as 7
peelings (that means "produce 7 extra copies of body")

[Bug rtl-optimization/54919] [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

--- Comment #4 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-13 21:51:15 UTC ---
Created attachment 28442
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28442
Perform replacement in REG_EQUAL notes too

[Bug rtl-optimization/54919] [4.6/4.7/4.8 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

--- Comment #5 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-14 19:29:22 UTC ---
Author: steven
Date: Sun Oct 14 19:29:18 2012
New Revision: 192439

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192439
Log:
gcc/
    PR rtl-optimization/54919
    * loop-unroll.c (struct var_to_expand): Remove accum_pos field.
    (analyze_insn_to_expand_var): Do not record accum_pos.
    (expand_var_during_unrolling): Use validate_replace_rtx_group to
    perform replacement of all references to SET_DEST (set) with the
    new register, including references in REG_EQUAL notes.
    (insert_var_expansion_initialization): Insert initializatio insns
    at the bottom of the pre-header of the loop.

testsuite/
    PR rtl-optimization/54919
    * gcc.dg/pr54919.c: New testcase.


Added:
    trunk/gcc/testsuite/gcc.dg/pr54919.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/loop-unroll.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/54919] [4.6/4.7 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.8.0
         AssignedTo|steven at gcc dot gnu.org   |unassigned at gcc dot
                   |                            |gnu.org
            Summary|[4.6/4.7/4.8 Regression]    |[4.6/4.7 Regression]
                   |gcc.dg/torture/pr54877.c    |gcc.dg/torture/pr54877.c
                   |FAILs with                  |FAILs with
                   |-fvariable-expansion-in-unr |-fvariable-expansion-in-unr
                   |oller                       |oller
      Known to fail|4.8.0                       |

--- Comment #6 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-14 20:30:42 UTC ---
Fixed on the trunk, see r192439.

I have no plans to work on back-ports of my patch for the release
branches, so unassigning...

[Bug rtl-optimization/54919] [4.6/4.7 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch
   Target Milestone|---                         |4.6.4

[Bug rtl-optimization/54919] [4.6/4.7 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2

[Bug rtl-optimization/54919] [4.7 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.6.4                       |4.7.4

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-04-12 15:17:08 UTC ---
GCC 4.6.4 has been released and the branch has been closed.

[Bug rtl-optimization/54919] [4.7 Regression] gcc.dg/torture/pr54877.c FAILs with -fvariable-expansion-in-unroller

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54919

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED
   Target Milestone|4.7.4                       |4.8.0
      Known to fail|4.7.3                       |4.7.4

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed for 4.8.0.