[Bug c/24376] New: wrong-code unless -fno-sched-interblock

[Bug c/24376] New: wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

The cross-compiler gcc-3.4.4 (or 3.4.3) for v850 generates wrong-code for
example source bug2.c (see below).
The cmp instruction in expression (ps1 ? 1 : 2) was lost in the first
scheduling pass (see dump-file bug2.c.23.sched generated by compilation with
-da option). The return value of sh_dir() (see below) is random and depends on
state of PSW register at return from previous bprintf() call.

Shorted asm listing of wrong-code:
...
        jarl _x_bprintf,r31     #
        setf z,r10      # tmp73
        add 1,r10       #, <result>
...

The wrong-code may be work-arounded by -fno-sched-interblock option.
The good-code - shorted asm listing:
...
        jarl _x_bprintf,r31     #
        cmp r0,r28      # ps           <---(This is the lost instruction)
        setf z,r10      # tmp73
        add 1,r10       #, <result>
...


The problem was tested on several linux instalation (Redhat 7.2 on i686, RedHat
Enterprise release 4 on x86_64, ...).

The source file bug2.c follows:
/*****************************************************************/
#define NULL 0

extern const char *f_cwd(const char *s);
extern int x_strcmp(const char *s1, const char *s2);
extern int x_strncmp(const char *s1, const char *s2, unsigned n);
extern int x_bprintf(int f, const char *fmt, ...);


typedef struct
{
  char cline[202];
  unsigned char out;
  unsigned short dir;
} t_sloc;


int sh_dir(t_sloc *sl)
{
  const char *ps=NULL, *ps1;

  if (x_strcmp(sl->cline, "pwd") == 0)
    goto label01;

  if (x_strncmp(sl->cline, "cd ", 3) == 0 && sl->cline[3] > ' ') {
    ps = sl->cline + 3;
label01:
    if ((ps1 = ps = f_cwd(ps)) == NULL)
      ps = f_cwd(NULL);
    if (!ps)
      return 2;

    x_bprintf(sl->out, "wd: %s\r\n", ps);
    sl->dir = ps[1] != 'M';
    x_bprintf(1, "ps1=%p, %d\r\n", ps1, 2);

/* This occurs wrong code: The state of Z flag of PSW register (cc0)
 * at end of previous call x_bprintf()
 * is evaluated instead of ps1 in next line. */

    return ps1 ? 1 : 2;
  }

  return 0;
}
/* End of source file*/
/*****************************************************************/


The compiler command line:
/usr/local/bin/gcc.v -Wall -Os -mspace -ep -mghs -mno-app-regs -mzda=8 -c
-Wa,-a=bug2.lst -fverbose-asm -o bug2.o bug2.c

The cross-compiler configure line:
../gcc-3.4.4/configure --prefix=/usr/local/ --host=i586-pc-linux-gnu
--target=v850-linux-elf --program-suffix=.v


-- 
           Summary: wrong-code unless -fno-sched-interblock
           Product: gcc
           Version: 3.4.4
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: seidl at feld dot cvut dot cz
 GCC build triplet: i586-pc-linux-gnu
  GCC host triplet: i586-pc-linux-gnu
GCC target triplet: v850-linux-elf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug c/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #1 from seidl at feld dot cvut dot cz  2005-10-14 21:52 -------
Created an attachment (id=9987)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=9987&action=view)
source code for wrong-code problem demonstration


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[steven at gcc dot gnu dot org]

------- Comment #2 from steven at gcc dot gnu dot org  2005-12-29 21:54 -------
Are you seeing the same problem with more recent compilers, like GCC 4.0 or a
GCC 4.1 snapshot?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at fel dot cvut dot cz]

------- Comment #3 from seidl at fel dot cvut dot cz  2006-01-01 18:38 -------
Subject: Re:  wrong-code unless -fno-sched-interblock

The problem was detected with gcc-3.4.3 and gcc-3.4.4. I not tested the older
gcc versions. I tested gcc-4.0.1, and the problem on the small sample of code
was not occurs. It is chance (but not certainty) that the problem may be solved
by update gcc. Unfortunately the gcc-4.x generate a little bigger output code
size (about 2-3 percents) and I can not check the all program functionality,
because the 128 kB limit of physical program space of my V850 system was
exceeded. My application is limited strictly by size of physical memory and
using of gcc-3.4.4 is optimal for me for now. Therefore, I can not confirm yet,
whether the gcc-4.x generate correct code - I can not support the confirm via
long-term using of the compiler. The gcc-3.4.4 with -fno-sched-interblock
option works fine (by my opinion).
    Libor Seidl

Cituji z emailu od steven at gcc dot gnu dot org <gcc-bugzilla@gcc.gnu.org>:

>
>
> ------- Comment #2 from steven at gcc dot gnu dot org  2005-12-29 21:54
> -------
> Are you seeing the same problem with more recent compilers, like GCC 4.0 or a
> GCC 4.1 snapshot?
>
>
> --
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376
>
> ------- You are receiving this mail because: -------
> You reported the bug, or are watching the reporter.
>
>


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[steven at gcc dot gnu dot org]

------- Comment #4 from steven at gcc dot gnu dot org  2006-01-01 18:49 -------
Adding Nick to the CC list, because he is the listed v850 port maintainer.


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |nickc at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[nickc at redhat dot com]

------- Comment #5 from nickc at redhat dot com  2006-01-10 13:15 -------
Created an attachment (id=10605)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10605&action=view)
When computing region not to schedule, include earlier cc0 setters


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[nickc at redhat dot com]

------- Comment #6 from nickc at redhat dot com  2006-01-10 13:19 -------
Hi Libor,

  I have uploaded a patch which should fix the problem.  It is not the most
elegant solution but I believe that it should work for now.

  The patch amends the code in add_branch_dependencies() which searches back
from the end of a block to the first instruction which is not a branch. call,
cc0 setter, clobber or use.  The amendment is that once such an instruction is
found, the search is continued backwards until all cc0 setters have been found.
 In this way none of the cc0 setter/user pairs in the block are disturbed.

  Please let me know how you get on with this patch and if it solves the
problem for you.

Cheers
  Nick


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #7 from seidl at feld dot cvut dot cz  2006-01-11 01:24 -------
Hi Nick,
I tested patched compiler on my problematic sources and I check assembler
output of compiler carefully (but not execute destination code yet). I believe
that reported problem is solved. But the compilation of several other sources
is break by internal compiler error in schedule_block referencing at
sched-rgn.c:2572 or haifa-sched.c:2355.
I send you two selected problematic sources including command line and compiler
response in attachment for analyse. But, the compilation of internal gcc
library gcc-3.4.4/gcc/libgcc2.c is affected by the internal error also,
compilation of cross-compiler ends with error:
---- My commandline for cross-compiler make:
../gcc-3.4.4/configure --prefix=/usr/local/v850 --host=i586-pc-linux-gnu\
 --target=v850-linux-elf --program-suffix=.v
nice make -s
---- and the error response at end of compilation:
../../gcc-3.4.4/gcc/libgcc2.c: In function `__addvdi3':
../../gcc-3.4.4/gcc/libgcc2.c:109: internal compiler error: in schedule_block,
at haifa-sched.c:2355

I thank you for your help. I presume, that the error resolving is important for
other platform that v850 (may be for any platform with cc0 ?) and may impact
the wide quantity of GCC users. But your first patch has unneded side effect in
internal compiler errors and I ask you kindly for continue of the problem
elimination.

Regards
    Libor


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #8 from seidl at feld dot cvut dot cz  2006-01-11 01:26 -------
Created an attachment (id=10616)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10616&action=view)
Source generating internal error of patched compiler


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #9 from seidl at feld dot cvut dot cz  2006-01-11 01:30 -------
Created an attachment (id=10617)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10617&action=view)
Another source making internal error of patched compiler


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[nickc at redhat dot com]

------- Comment #10 from nickc at redhat dot com  2006-01-11 15:37 -------
Created an attachment (id=10622)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=10622&action=view)
Revised patch which should not walk over blocks


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[nickc at redhat dot com]

------- Comment #11 from nickc at redhat dot com  2006-01-11 15:38 -------
Hi Libor,

  Oops- there was a condition where the new code could walk backwards into a
previous block.  Sorry about that.  I have uploaded a revised patch which
should not fall into this trap.  It certainly works with the two other test
cases you provided.

Cheers
  Nick


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2006-01-11 23:46:57
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #12 from seidl at feld dot cvut dot cz  2006-01-13 02:11 -------
(In reply to comment #11)
Hi Nick,
the patched compiler works OK and cross-compiled program in target v850
platform seems to be functional and stable. I assume, the problem is solved,
but I plane to make a thorough program functionality tests in next days. I will
inform you about result of the tests in next week, but I hope that the problem
is solved. Thank you for your help and for GCC maintenance.
Cheers
   Libor


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[seidl at feld dot cvut dot cz]

------- Comment #13 from seidl at feld dot cvut dot cz  2006-01-30 00:00 -------
Hi Nick,
I tested the pathed compiler nearly two weaks. I not detect any problem or any
side-effect. Therefore, the bug seems to be fixed reliably. The patch is usable
and needful for next version gcc-3.4.5 too (I test it successfully). Thank you
for your maintenance of GCC.
Regards
    Libor


-- 

seidl at feld dot cvut dot cz changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[pinskia at gcc dot gnu dot org]

------- Comment #14 from pinskia at gcc dot gnu dot org  2006-01-30 12:45 -------
Actually this has not been fixed yet in the SVN (mainling), reopening.


-- 

pinskia at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] wrong-code unless -fno-sched-interblock

[nickc at gcc dot gnu dot org]

------- Comment #15 from nickc at gcc dot gnu dot org  2006-02-02 10:54 -------
Subject: Bug 24376

Author: nickc
Date: Thu Feb  2 10:54:53 2006
New Revision: 110510

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=110510
Log:
PR 24376
* sched-rgn.c (add_branch_dependences): For cc0 targets extend the 
dependencies inside a block back to the very first cc0 setter in the block.

Modified:
    branches/gcc-3_4-branch/gcc/ChangeLog
    branches/gcc-3_4-branch/gcc/sched-rgn.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376

[Bug rtl-optimization/24376] [3.4 Regression] wrong-code unless -fno-sched-interblock

[pinskia at gcc dot gnu dot org]

------- Comment #16 from pinskia at gcc dot gnu dot org  2006-02-02 12:41 -------
Fixed in 3.4.6 now. Thanks for your report and helping out testing the patches.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|REOPENED                    |RESOLVED
      Known to fail|                            |3.4.4
      Known to work|                            |3.3.3
         Resolution|                            |FIXED
            Summary|wrong-code unless -fno-     |[3.4 Regression] wrong-code
                   |sched-interblock            |unless -fno-sched-interblock
   Target Milestone|---                         |3.4.6


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24376