[Bug c/46315] New: gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[Bug c/46315] New: gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[zorry at gentoo dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

           Summary: gcc-4.5.1 -O2 -fno-strict-overflow causes bad code
                    generation
           Product: gcc
           Version: 4.5.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zorry@gentoo.org


Created attachment 22294
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22294
testcase

If we disable -fstrict-overflow in the -O2 optimization the testcase segfault
but works fine with -O1 or -fstrict-overflow on in the -O2 optimization.
flame@yamato mytmpfs % gcc -O2 -fno-strict-overflow testcase.c -o testcase
flame@yamato mytmpfs % ./testcase
0x400770
0x2
zsh: segmentation fault (core dumped)  ./testcase

[Bug tree-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.11.05 16:43:32
      Known to work|                            |4.4.4, 4.6.0
   Target Milestone|---                         |4.5.2
            Summary|gcc-4.5.1 -O2               |[4.5 Regression] gcc-4.5.1
                   |-fno-strict-overflow causes |-O2 -fno-strict-overflow
                   |bad code generation         |causes bad code generation
     Ever Confirmed|0                           |1
      Known to fail|                            |4.5.0, 4.5.1

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-05 16:43:32 UTC ---
Confirmed.

[Bug tree-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-05 16:47:12 UTC ---
ipa-cp triggers this but is not neccessarily at fault.  Workaround:
-fno-ipa-cp.

[Bug tree-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[dirtyepic at gentoo dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

Ryan Hill <dirtyepic at gentoo dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dirtyepic at gentoo dot org

--- Comment #3 from Ryan Hill <dirtyepic at gentoo dot org> 2010-11-09 04:49:54 UTC ---
Not to rush you, but I was wondering if there will be anyone looking at this in
the near future.  I was planning on releasing 4.5.1 into the unstable branch of
our repo this week but if a patch may be soon available I'll hold off.  If
there isn't any interest I'll do the release and mask it on our hardened
profiles (which use -fno-strict-overflow by default) for now.

[Bug tree-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-12 14:56:54 UTC ---
loop-header copying is where things start to be different.

In the end the asm doesn't make much sense:

parse_ranged.clone.0:
.LFB2:
        .cfi_startproc
        leaq    2(%rdi), %rdx
        movq    %rdi, %rax
...
        leal    (%rdx,%rdx,4), %edx
        movb    $2, %al
        leal    (%rcx,%rdx,2), %edx
.L2:
        testq   %rax, %rax
        je      .L7
        cmpl    $60, %edx
        movl    $0, %edx
        cmovge  %rdx, %rax
        ret

so we might end up returning literal 2.


Reduced testcase:

static char const *
parse_ranged (char const *s, int digits)
{
  int n = 0;
  char const *lim = s + digits;
  while (s < lim)
    {
      unsigned d = *s++ - '0';
      if (9 < d)
        return 0;
      n = 10 * n + d;
    }
  return s && 0 <= n && n <= 59 ? s : 0;
}
extern void abort (void);
int main()
{
  const char *s = "10092240";

  s = parse_ranged (s, 2);
  s = parse_ranged (s, 2);
  s = parse_ranged (s, 2);
  s = parse_ranged (s, 2);
  if (!s || *s != '\0')
    abort();
  return 0;
}

[Bug rtl-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org, rguenth at gcc dot
                   |                            |gnu.org, stevenb.gcc at
                   |                            |gmail dot com
          Component|tree-optimization           |rtl-optimization

--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-13 19:08:02 UTC ---
RTL constant propagation (cprop3) changes

(insn 75 26 29 5 t.c:8 (set (reg/v/f:DI 83 [ s ])
        (reg/f:DI 95 [ s ])) -1 (expr_list:REG_EQUAL (plus:DI (reg/v/f:DI 83 [
s ])
            (const_int 2 [0x2]))
        (nil)))

to

(insn 75 26 29 5 t.c:8 (set (reg/v/f:DI 83 [ s ])
        (reg/f:DI 95 [ s ])) -1 (expr_list:REG_EQUAL (const_int 2 [0x2])
        (nil)))

note that the original REG_EQUAL note is suspicious anyway as it references
the set destination.  The note is correct though, reg:DI 95 is set as

(insn 11 3 12 2 t.c:5 (parallel [
            (set (reg/f:DI 95 [ s ])
                (plus:DI (reg/v/f:DI 83 [ s ])
                    (const_int 2 [0x2])))
            (clobber (reg:CC 17 flags))
        ]) 253 {*adddi_1} (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

RTL PRE generates insn 75 and also adds this REG_EQUAL note.

Stopping here with my investigation.

[Bug rtl-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

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

--- Comment #6 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-13 21:31:18 UTC ---
Investigating.

[Bug rtl-optimization/46315] [4.5 Regression] gcc-4.5.1 -O2 -fno-strict-overflow causes bad code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #7 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-14 11:34:42 UTC ---
The problem is indeed the note, but I don't think it comes from it being sort
of self-referential.  The use of reg 83 in the note doesn't make it live so ce1
merges BB 4 just before insn 75:


(note 19 18 22 4 [bb 4] NOTE_INSN_BASIC_BLOCK)

(insn 22 19 23 4 pr43615.c:10 (set (reg:SI 86)
        (sign_extend:SI (mem:QI (plus:SI (reg/v/f:SI 83 [ s ])
                    (const_int 1 [0x1])) [0 S1 A8]))) 96 {extendqisi2}
(expr_list:REG_DEAD (reg/v/f:SI 73 [ s ])
        (nil)))

(insn 23 22 24 4 pr43615.c:10 (parallel [
            (set (reg/v:SI 81 [ d ])
                (plus:SI (reg:SI 86)
                    (const_int -48 [0xffffffd0])))
            (clobber (reg:CC 17 flags))
        ]) 197 {*addsi_1} (expr_list:REG_DEAD (reg:SI 86)
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))

(insn 24 23 25 4 pr43615.c:11 (set (reg:CC 17 flags)
        (compare:CC (reg/v:SI 81 [ d ])
            (const_int 9 [0x9]))) 6 {*cmpsi_1} (nil))

(jump_insn 25 24 26 4 pr43615.c:11 (set (pc)
        (if_then_else (gtu (reg:CC 17 flags)
                (const_int 0 [0x0]))
            (label_ref:SI 68)
            (pc))) 464 {*jcc_1} (expr_list:REG_DEAD (reg:CC 17 flags)
        (expr_list:REG_BR_PROB (const_int 450 [0x1c2])
            (nil)))
 -> 68)

(note 26 25 75 5 [bb 5] NOTE_INSN_BASIC_BLOCK)

(insn 75 26 29 5 pr43615.c:10 (set (reg/v/f:SI 83 [ s ])
        (reg/f:SI 95 [ s ])) -1 (expr_list:REG_EQUAL (plus:SI (reg/v/f:SI 83 [
s ])
            (const_int 2 [0x2]))
        (nil)))


with BB 11:


(note 67 68 5 11 [bb 11] NOTE_INSN_BASIC_BLOCK)

(insn 5 67 72 11 pr43615.c:12 (set (reg/v/f:SI 83 [ s ])
        (const_int 0 [0x0])) 44 {*movsi_1} (nil))


yielding:


(note 19 18 22 4 [bb 4] NOTE_INSN_BASIC_BLOCK)

(insn 22 19 23 4 pr43615.c:10 (set (reg:SI 86)
        (sign_extend:SI (mem:QI (plus:SI (reg/v/f:SI 83 [ s ])
                    (const_int 1 [0x1])) [0 S1 A8]))) 96 {extendqisi2}
(expr_list:REG_DEAD (reg/v/f:SI 73 [ s ])
        (nil)))

(insn 23 22 5 4 pr43615.c:10 (parallel [
            (set (reg/v:SI 81 [ d ])
                (plus:SI (reg:SI 86)
                    (const_int -48 [0xffffffd0])))
            (clobber (reg:CC 17 flags))
        ]) 197 {*addsi_1} (expr_list:REG_DEAD (reg:SI 86)
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))

(insn 5 23 24 4 pr43615.c:12 (set (reg/v/f:SI 83 [ s ])
        (const_int 0 [0x0])) 44 {*movsi_1} (nil))

(insn 24 5 25 4 pr43615.c:11 (set (reg:CC 17 flags)
        (compare:CC (reg/v:SI 81 [ d ])
            (const_int 9 [0x9]))) 6 {*cmpsi_1} (nil))

(jump_insn 25 24 26 4 pr43615.c:11 (set (pc)
        (if_then_else (gtu (reg:CC 17 flags)
                (const_int 0 [0x0]))
            (label_ref:SI 46)
            (pc))) 464 {*jcc_1} (expr_list:REG_DEAD (reg:CC 17 flags)
        (expr_list:REG_BR_PROB (const_int 450 [0x1c2])
            (nil)))
 -> 46)

(insn 75 26 29 5 pr43615.c:10 (set (reg/v/f:SI 83 [ s ])
        (reg/f:SI 95 [ s ])) -1 (expr_list:REG_EQUAL (plus:SI (reg/v/f:SI 83 [
s ])
            (const_int 2 [0x2]))
        (nil)))


and insn 5 now feeds the note.

[Bug rtl-optimization/46315] [4.5 regression] -O2 -fno-strict-overflow causes wrong code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #8 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-15 23:44:08 UTC ---
Created attachment 22412
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22412
fix for the 4.5 branch

Not fully testes yet.

[Bug rtl-optimization/46315] [4.5 regression] -O2 -fno-strict-overflow causes wrong code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #9 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 22:13:56 UTC ---
Author: ebotcazou
Date: Tue Nov 16 22:13:52 2010
New Revision: 166827

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=166827
Log:
    PR rtl-optimization/46315
    * rtl.h (remove_reg_equal_equiv_notes_for_regno): Declare.
    * rtlanal.c (remove_reg_equal_equiv_notes_for_regno): New function
    extracted from...
    * dce.c (delete_corresponding_reg_eq_notes): ...here.  Rename into...
    (remove_reg_equal_equiv_notes_for_defs): ...this.
    (delete_unmarked_insns): Adjust to above renaming.
    * ifcvt.c (dead_or_predicable): Remove REG_EQUAL and REG_EQUIV notes
    referring to registers set in the insns being moved, if any.

    * df-core.c (df_ref_dump): New function extracted from...
    (df_refs_chain_dump): ...here.  Call it.
    (df_regs_chain_dump): Likewise.
    * df-problems.c (df_chain_dump): Print 'e' for uses in notes.
    * df-scan.c (df_scan_start_dump): Likewise.  Fix long line.

Added:
    trunk/gcc/testsuite/gcc.dg/pr46315.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/dce.c
    trunk/gcc/df-core.c
    trunk/gcc/df-problems.c
    trunk/gcc/df-scan.c
    trunk/gcc/ifcvt.c
    trunk/gcc/rtl.h
    trunk/gcc/rtlanal.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/46315] [4.5 regression] -O2 -fno-strict-overflow causes wrong code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

--- Comment #10 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 22:17:22 UTC ---
Author: ebotcazou
Date: Tue Nov 16 22:17:17 2010
New Revision: 166828

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=166828
Log:
    PR rtl-optimization/46315
    * rtl.h (remove_reg_equal_equiv_notes_for_regno): Declare.
    * rtlanal.c (remove_reg_equal_equiv_notes_for_regno): New function
    extracted from...
    * dce.c (delete_corresponding_reg_eq_notes): ...here.  Rename into...
    (remove_reg_equal_equiv_notes_for_defs): ...this.
    (delete_unmarked_insns): Adjust to above renaming.
    * ifcvt.c (dead_or_predicable): Remove REG_EQUAL and REG_EQUIV notes
    referring to registers set in the insns being moved, if any.

    * df-core.c (df_ref_dump): New function extracted from...
    (df_refs_chain_dump): ...here.  Call it.
    (df_regs_chain_dump): Likewise.
    * df-problems.c (df_chain_dump): Print 'e' for uses in notes.
    * df-scan.c (df_scan_start_dump): Likewise.  Fix long line.

Added:
    branches/gcc-4_5-branch/gcc/testsuite/gcc.dg/pr46315.c
      - copied unchanged from r166827, trunk/gcc/testsuite/gcc.dg/pr46315.c
Modified:
    branches/gcc-4_5-branch/gcc/ChangeLog
    branches/gcc-4_5-branch/gcc/dce.c
    branches/gcc-4_5-branch/gcc/df-core.c
    branches/gcc-4_5-branch/gcc/df-problems.c
    branches/gcc-4_5-branch/gcc/df-scan.c
    branches/gcc-4_5-branch/gcc/ifcvt.c
    branches/gcc-4_5-branch/gcc/rtl.h
    branches/gcc-4_5-branch/gcc/rtlanal.c
    branches/gcc-4_5-branch/gcc/testsuite/ChangeLog

[Bug rtl-optimization/46315] [4.5 regression] -O2 -fno-strict-overflow causes wrong code generation

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46315

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
                URL|                            |http://gcc.gnu.org/ml/gcc-p
                   |                            |atches/2010-11/msg01735.htm
                   |                            |l
         Resolution|                            |FIXED

--- Comment #11 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 22:22:21 UTC ---
Thanks for the reduced testcase.