[Bug rtl-optimization/46804] New: [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[Bug rtl-optimization/46804] New: [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

           Summary: [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90
                    FAILs with -fregmove
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zsojka@seznam.cz
              Host: x86_64-pc-linux-gnu
            Target: x86_64-pc-linux-gnu


Created attachment 22632
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22632
auto-reduced testcase

Output:
$ /mnt/svn/gcc-trunk/binary-167453-lto-fortran-checking-yes-rtl-df/bin/gfortran
-O -fPIC -fexpensive-optimizations -fgcse -foptimize-register-move -fpeel-loops
-fno-tree-loop-optimize pr46804.f90
$ ./a.out 
Aborted

In the assembly, the problem is apparent (intel syntax):
...
    mov    r12d, DWORD PTR A.4.1573[rip+3]    # tmp310, A.4
    mov    DWORD PTR 16[rsp+r8*4], r12d    # shift3, tmp310
...
32bit (unaligned) value is loaded instead of sign-extend byte as is done
without -fregmove:
...
    movsx    r11d, BYTE PTR A.4.1573[rip+3]    # tmp310, A.4
    mov    DWORD PTR 16[rsp+r8*4], r11d    # shift3, tmp310
...

so shift3 contains invalid data

The problem seems to first appear in the 182r.regmove dump:
$ diff pr46804.f90.180r.ce2 pr46804.f90.182r.regmove
...
2009,2010c1409,1410
<         (sign_extend:SI (reg:QI 236))) pr46804.f90:17 132 {extendqisi2}
<      (expr_list:REG_DEAD (reg:QI 236)
---
>         (reg:SI 236)) pr46804.f90:17 64 {*movsi_internal}
>      (expr_list:REG_DEAD (reg:SI 236)
...

Output:
r167453 - fail
r158095 - fail
4.5 r166509 - fail
4.4 r166509 - OK

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

--- Comment #1 from Zdenek Sojka <zsojka at seznam dot cz> 2010-12-06 14:56:13 UTC ---
There are two likely related failures, but they both need -fno-tree-ter:
gfortran.dg/char_eoshift_[24].f90

$ gfortran -O -fPIC -fexpensive-optimizations -fgcse -foptimize-register-move
-fpeel-loops -fno-tree-loop-optimize -fno-tree-ter char_eoshift_2.f90 &&
./a.out
Aborted
$ gfortran -O -fPIC -fexpensive-optimizations -fgcse -foptimize-register-move
-fpeel-loops -fno-tree-loop-optimize -fno-tree-ter char_eoshift_4.f90 &&
./a.out 
Aborted

Neither of them fails in 4.4.

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.12.06 16:38:55
                 CC|                            |bernds at codesourcery dot
                   |                            |com
   Target Milestone|---                         |4.5.2
     Ever Confirmed|0                           |1

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> 2010-12-06 16:38:55 UTC ---
It is caused by revision 145357:

http://gcc.gnu.org/ml/gcc-cvs/2009-03/msg00864.html

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

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

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

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-08 10:49:42 UTC ---
Regmove seems to change:
(insn 107 105 816 9 (set (reg:QI 236)
        (mem/s/u/j:QI (const:DI (plus:DI (symbol_ref:DI ("A.4.1573") [flags
0x2]  <var_decl # A.4>)
                    (const_int 3 [0x3]))) [0 A.4 S1 A8])) pr46804.f90:15 66
{*movqi_internal}
     (expr_list:REG_EQUAL (mem/s/u/j:QI (const:DI (plus:DI (symbol_ref:DI
("A.4.1573") [flags 0x2]  <var_decl # A.4>)
                    (const_int 3 [0x3]))) [0 A.4 S1 A8])
        (nil)))
...
(insn 108 824 643 9 (set (mem/s/j:QI (plus:DI (plus:DI (mult:DI (reg:DI 235 [
S.5 ])
                        (const_int 2 [0x2]))
                    (reg/f:DI 20 frame))
                (const_int -781 [0xfffffffffffffcf3])) [0 FRAME.22.shift1 S1
A8])
        (reg:QI 236)) pr46804.f90:15 66 {*movqi_internal}
     (nil))
...
(insn 117 116 832 9 (set (reg:HI 272)
        (sign_extend:HI (reg:QI 236))) pr46804.f90:16 134 {extendqihi2}
     (nil))
...
(insn 133 510 134 9 (set (reg:SI 310)
        (sign_extend:SI (reg:QI 236))) pr46804.f90:17 132 {extendqisi2}
     (expr_list:REG_DEAD (reg:QI 236)
        (nil)))

into:

(insn 107 105 816 9 (set (reg:SI 236)
        (sign_extend:SI (mem/s/u/j:QI (const:DI (plus:DI (symbol_ref:DI
("A.4.1573") [flags 0x2]  <var_decl # A.4>)
                        (const_int 3 [0x3]))) [0 A.4 S1 A8]))) pr46804.f90:15
132 {extendqisi2}
     (expr_list:REG_EQUAL (mem/s/u/j:QI (const:DI (plus:DI (symbol_ref:DI
("A.4.1573") [flags 0x2]  <var_decl # A.4>)
                    (const_int 3 [0x3]))) [0 A.4 S1 A8])
        (nil)))
...
(insn 108 824 643 9 (set (mem/s/j:QI (plus:DI (plus:DI (mult:DI (reg:DI 235 [
S.5 ])
                        (const_int 2 [0x2]))
                    (reg/f:DI 20 frame))
                (const_int -781 [0xfffffffffffffcf3])) [0 FRAME.22.shift1 S1
A8])
        (subreg:QI (reg:SI 236) 0)) pr46804.f90:15 66 {*movqi_internal}
     (nil))
...
(insn 117 116 832 9 (set (reg:HI 272)
        (sign_extend:HI (subreg:QI (reg:SI 236) 0))) pr46804.f90:16 134
{extendqihi2}
     (nil))
...
(insn 133 510 134 9 (set (reg:SI 310)
        (reg:SI 236)) pr46804.f90:17 64 {*movsi_internal}
     (expr_list:REG_DEAD (reg:SI 236)
        (nil)))

The problem is that REG_EQUAL note on insn 107 isn't adjusted and then with
high register pressure just results in wrong code.

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-08 11:17:00 UTC ---
Ah, it seems Bernd committed a patch for this 9 years ago:
http://gcc.gnu.org/ml/gcc-patches/2001-06/msg01250.html
but probably had just testcase for REG_EQUIV and not REG_EQUAL, so didn't
discover that p at that point is not the p it wants, as there has been a
forward loop too.  I guess if the REG_EQUAL note is identical to the source
operand, it could be sign/zero extended instead of removing it, for anything
else of course we want to remove it, but from the right insn.

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-08 11:53:31 UTC ---
Created attachment 22682
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22682
gcc46-pr46804.patch

Untested fix.

[Bug rtl-optimization/46804] [4.5/4.6 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-10 16:40:57 UTC ---
Author: jakub
Date: Fri Dec 10 16:40:51 2010
New Revision: 167695

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=167695
Log:
    PR rtl-optimization/46804
    * regmove.c (optimize_reg_copy_3): Look for REG_EQUAL note
    on the setter of src_reg rather than on insn.  If it is
    equal to the setter's original SET_SRC, replace it with its
    zero or sign extension instead of dropping it.

    * gfortran.dg/pr46804.f90: New test.

Added:
    trunk/gcc/testsuite/gfortran.dg/pr46804.f90
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/regmove.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/46804] [4.5 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.6.0
            Summary|[4.5/4.6 Regression]        |[4.5 Regression]
                   |gfortran.dg/char_cshift_2.f |gfortran.dg/char_cshift_2.f
                   |90 FAILs with -fregmove     |90 FAILs with -fregmove
      Known to fail|4.6.0                       |

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-10 16:47:35 UTC ---
Fixed on the trunk so far.

[Bug rtl-optimization/46804] [4.5 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.5.2                       |4.5.3

--- Comment #8 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-12-16 13:03:18 UTC ---
GCC 4.5.2 is being released, adjusting target milestone.

[Bug rtl-optimization/46804] [4.5 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-01-16 20:16:34 UTC ---
Author: jakub
Date: Sun Jan 16 20:16:30 2011
New Revision: 168863

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=168863
Log:
    Backport from mainline
    2010-12-10  Jakub Jelinek  <jakub@redhat.com>

    PR rtl-optimization/46804
    * regmove.c (optimize_reg_copy_3): Look for REG_EQUAL note
    on the setter of src_reg rather than on insn.  If it is
    equal to the setter's original SET_SRC, replace it with its
    zero or sign extension instead of dropping it.

    * gfortran.dg/pr46804.f90: New test.

Added:
    branches/gcc-4_5-branch/gcc/testsuite/gfortran.dg/pr46804.f90
Modified:
    branches/gcc-4_5-branch/gcc/ChangeLog
    branches/gcc-4_5-branch/gcc/regmove.c
    branches/gcc-4_5-branch/gcc/testsuite/ChangeLog

[Bug rtl-optimization/46804] [4.5 Regression] gfortran.dg/char_cshift_2.f90 FAILs with -fregmove

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46804

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

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

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-01-17 08:08:05 UTC ---
Fixed.