[Bug target/37633] New: [4.4 Regression] wrong register use on sh64

[Bug target/37633] New: [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

Currently many tests for string builtin functions fail on
sh64-unknown-linux-gnu.  Here is a small test case:

extern void abort (void);

char buf[64];

int
main (void)
{
  __builtin_strcpy (buf, "mystring");
  if (strcmp (buf, "mystring") != 0)
    abort ();

  __builtin_strcpy (buf, "mystring");
  if (strcmp (buf, "mystring") != 0)
    abort ();

  return 0;
}

With -O3, the register R11 is used to hold 8-byte of "mystring"
for the first __builtin_strcpy call.  The value in R11 is reused
with the 2nd __builtin_strcpy call.
When strcmp is dynamically linked, the upper 32-bit of R11 is
clobbered by the first call of strcmp.  Then the 2nd strcmp
returns false.
The backend defines HARD_REGNO_CALL_PART_CLOBBERED so to notify
that the upper 32-bit of R11 may be clobbered for SHmedia:

/* Only the lower 32-bits of R10-R14 are guaranteed to be preserved
   across SHcompact function calls.  We can't tell whether a called
   function is SHmedia or SHcompact, so we assume it may be when
   compiling SHmedia code with the 32-bit ABI, since that's the only
   ABI that can be linked with SHcompact code.  */
#define HARD_REGNO_CALL_PART_CLOBBERED(REGNO,MODE) \
  (TARGET_SHMEDIA32 \
   && GET_MODE_SIZE (MODE) > 4 \
   && (((REGNO) >= FIRST_GENERAL_REG + 10 \
        && (REGNO) <= FIRST_GENERAL_REG + 15) \
       || TARGET_REGISTER_P (REGNO) \
       || (REGNO) == PR_MEDIA_REG))

but it seems that now HARD_REGNO_CALL_PART_CLOBBERED doesn't affect
the register allocation.


-- 
           Summary: [4.4 Regression] wrong register use on sh64
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: kkojima at gcc dot gnu dot org
GCC target triplet: sh64-unknow-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

-- 

kkojima at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P4


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
   Target Milestone|---                         |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

------- Comment #1 from kkojima at gcc dot gnu dot org  2008-09-26 03:31 -------
I've tried

--- ORIG/trunk/gcc/ira-color.c  Wed Sep 17 09:48:49 2008
+++ LOCAL/trunk/gcc/ira-color.c Thu Sep 25 12:09:30 2008
@@ -514,7 +514,9 @@ assign_hard_reg (ira_allocno_t allocno, 
 #endif
       if (! ira_hard_reg_not_in_set_p (hard_regno, mode, conflicting_regs)
          || TEST_HARD_REG_BIT (prohibited_class_mode_regs[cover_class][mode],
-                               hard_regno))
+                               hard_regno)
+         || (ALLOCNO_CALLS_CROSSED_NUM (allocno) != 0
+             && HARD_REGNO_CALL_PART_CLOBBERED (hard_regno, mode)))
        continue;
       cost = costs[i];
       full_cost = full_costs[i];
@@ -3064,7 +3066,9 @@ ira_fast_allocation (void)
 #endif
          if (!ira_hard_reg_not_in_set_p (hard_regno, mode, conflict_hard_regs)
              || (TEST_HARD_REG_BIT
-                 (prohibited_class_mode_regs[cover_class][mode], hard_regno)))
+                 (prohibited_class_mode_regs[cover_class][mode], hard_regno))
+             || (ALLOCNO_CALLS_CROSSED_NUM (a) != 0
+                 && HARD_REGNO_CALL_PART_CLOBBERED (hard_regno, mode)))
            continue;
          ALLOCNO_HARD_REGNO (a) = hard_regno;
          for (r = ALLOCNO_LIVE_RANGES (a); r != NULL; r = r->next)


It works for me, though it could be completely wrong.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[vmakarov at redhat dot com]

------- Comment #2 from vmakarov at redhat dot com  2008-10-08 21:55 -------
Thanks for working on the problem.  I am afraid that preventing to use
partially clobbered hard regs for allocnos living through calls is wrong.  In
most general case I believe the problem should be fixed in caller-saves.c: they
should saved and restored around calls.  As temporary workaround would be
correct calculation of costs in IRA for partially clobbered hard regs (as it
was done in global.c) in order to force IRA avoid the hard regs.

But without fixing the problem in caller-saves.c IRA and the old RA are still
error-prune for this problem in general case.

I'll work on the patch for correct partially clobbered hard register
calculation.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[vmakarov at redhat dot com]

------- Comment #3 from vmakarov at redhat dot com  2008-10-14 19:42 -------
I've checked the old RA.  It does not assigned partially clobbered hard
register because it is done only when non partially clobbered hard regs were
tried first.  Sh has a lot of such registers therefore the chance to generate
wrong code is small.

I can simulate the same behaviour for IRA by increasing costs for partially
clobbered hard registers.  Currently rs6000 and sh define
HARD_REGNO_CALL_PART_CLOBBERED.  So even the problem is solved for sh in
different way, the patch increasing cost would be useful for rs6000.

Still, as I wrote, the complete solution (the mentioned cost increase will be
still necessary in any case) would be save and restore partially clobbered
hard-registers in caller-saves.c.

The patch increasing the costs will be sent later today.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[hjl at gcc dot gnu dot org]

------- Comment #4 from hjl at gcc dot gnu dot org  2008-10-14 22:24 -------
Subject: Bug 37633

Author: hjl
Date: Tue Oct 14 22:23:23 2008
New Revision: 141120

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141120
Log:
2008-10-14  Vladimir Makarov  <vmakarov@redhat.com>

        PR target/37633
        * ira-costs.c (ira_tune_allocno_costs_and_cover_classes): Check
        HARD_REGNO_CALL_PART_CLOBBERED.

Modified:
    branches/ira-merge/gcc/ChangeLog.ira
    branches/ira-merge/gcc/ira-costs.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

------- Comment #5 from kkojima at gcc dot gnu dot org  2008-10-15 23:14 -------
Thanks for taking a look at this problem.  Your patch mentioned
in #3 fixes the test case in #1 on sh64.  Even with it, there
are similar failures with -O0 in the result of the gcc testsuite
on sh64-linux.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[vmakarov at gcc dot gnu dot org]

------- Comment #6 from vmakarov at gcc dot gnu dot org  2008-10-16 00:54 -------
Subject: Bug 37633

Author: vmakarov
Date: Thu Oct 16 00:52:59 2008
New Revision: 141161

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141161
Log:
2008-10-14  Vladimir Makarov  <vmakarov@redhat.com>

        PR target/37633
        * ira-costs.c (ira_tune_allocno_costs_and_cover_classes): Check
        HARD_REGNO_CALL_PART_CLOBBERED.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/ira-costs.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

------- Comment #7 from kkojima at gcc dot gnu dot org  2008-10-21 23:42 -------
Subject: Bug 37633

Author: kkojima
Date: Tue Oct 21 23:40:41 2008
New Revision: 141282

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141282
Log:
        PR target/37633
        * config/sh/sh.c (sh_hard_regno_mode_ok): New function.
        * config/sh/sh.h (HARD_REGNO_MODE_OK): Use it.
        * config/sh/sh-protos.h (sh_hard_regno_mode_ok): Declare.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/sh/sh-protos.h
    trunk/gcc/config/sh/sh.c
    trunk/gcc/config/sh/sh.h


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[andreasmeier80 at gmx dot de]

------- Comment #8 from andreasmeier80 at gmx dot de  2009-01-02 17:08 -------
What is the status of this bug?


-- 

andreasmeier80 at gmx dot de changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |andreasmeier80 at gmx dot de


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[kkojima at gcc dot gnu dot org]

------- Comment #9 from kkojima at gcc dot gnu dot org  2009-01-02 22:51 -------
The failure itself went away but requested to stay open.
See the thread

http://gcc.gnu.org/ml/gcc-patches/2008-10/msg00606.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4 Regression] wrong register use on sh64

[amylaar at gcc dot gnu dot org]

------- Comment #10 from amylaar at gcc dot gnu dot org  2009-03-08 16:35 -------
(In reply to comment #3)
> I've checked the old RA.  It does not assigned partially clobbered hard
> register because it is done only when non partially clobbered hard regs were
> tried first.  Sh has a lot of such registers therefore the chance to generate
> wrong code is small.
> 
> I can simulate the same behaviour for IRA by increasing costs for partially
> clobbered hard registers.  Currently rs6000 and sh define
> HARD_REGNO_CALL_PART_CLOBBERED.  So even the problem is solved for sh in
> different way, the patch increasing cost would be useful for rs6000.
> 
> Still, as I wrote, the complete solution (the mentioned cost increase will be
> still necessary in any case) would be save and restore partially clobbered
> hard-registers in caller-saves.c.

To avoid pessimizing code where partially clobbered registers are used
to allocate pseudos which are not affected by the partial clobbers, the
register allocator will have to provide caller_save.c with more information.
It needs to know which registers are live across a call in excess of the part
that is saved.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4/4.5 Regression] wrong register use on sh64

[jakub at gcc dot gnu dot org]

-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.4.0                       |4.4.1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4/4.5 Regression] wrong register use on sh64

[jakub at gcc dot gnu dot org]

-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.4.1                       |4.4.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4/4.5 Regression] wrong register use on sh64

[jakub at gcc dot gnu dot org]

-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.4.2                       |4.4.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4/4.5 Regression] wrong register use on sh64

[jakub at gcc dot gnu dot org]

-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.4.3                       |4.4.4


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633

[Bug target/37633] [4.4/4.5/4.6 Regression] wrong register use on sh64

[jakub at gcc dot gnu dot org]

-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.4.4                       |4.4.5


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37633