[Bug middle-end/37813] New: assert with IRA_COVER_CLASSES with singleton

[Bug middle-end/37813] New: assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

With the IRA_COVER_CLASSES in cris.h changed as follows (exposing one of the
GENERAL_REGS that has special properties as a singleton class), build fails
compiling libgcc:

/home/hp/combw/crisv32-obj/./gcc/xgcc -B/home/hp/combw/crisv32-obj/./gcc/
-nostdinc -B/home/hp/combw/crisv32-obj/crisv32-elf/newlib/ -isystem
/home/hp/combw/crisv32-obj/crisv32-elf/newlib/targ-include -isystem
/home/hp/combw/combined/newlib/libc/include
-B/home/hp/combw/crisv32-obj/crisv32-elf/libgloss/cris
-L/home/hp/combw/crisv32-obj/crisv32-elf/libgloss/libnosys
-L/home/hp/combw/combined/libgloss/cris -B/tmp/reg-crisv32/crisv32-elf/bin/
-B/tmp/reg-crisv32/crisv32-elf/lib/ -isystem
/tmp/reg-crisv32/crisv32-elf/include -isystem
/tmp/reg-crisv32/crisv32-elf/sys-include -L/home/hp/combw/crisv32-obj/./ld -g
-O2 -O2  -g -O2 -DIN_GCC -DCROSS_DIRECTORY_STRUCTURE  -W -Wall -Wwrite-strings
-Wstrict-prototypes -Wmissing-prototypes -Wcast-qual -Wold-style-definition 
-isystem ./include   -g  -DIN_LIBGCC2 -D__GCC_FLOAT_NOT_NEEDED -Dinhibit_libc 
-I. -I. -I../.././gcc -I/home/hp/combw/combined/libgcc
-I/home/hp/combw/combined/libgcc/. -I/home/hp/combw/combined/libgcc/../gcc
-I/home/hp/combw/combined/libgcc/../include  -DHAVE_CC_TLS -o _muldi3.o -MT
_muldi3.o -MD -MP -MF _muldi3.dep -DL_muldi3 -c
/home/hp/combw/combined/libgcc/../gcc/libgcc2.c \

/home/hp/combw/combined/libgcc/../gcc/libgcc2.c: In function '__muldi3':
/home/hp/combw/combined/libgcc/../gcc/libgcc2.c:566: internal compiler error:
in process_bb_node_lives, at ira-lives.c:677

Index: gcc/config/cris/cris.h
===================================================================
--- gcc/config/cris/cris.h      (revision 141074)
+++ gcc/config/cris/cris.h      (working copy)
@@ -597,7 +597,7 @@

 #define INDEX_REG_CLASS GENERAL_REGS

-#define IRA_COVER_CLASSES { GENERAL_REGS, SPECIAL_REGS, LIM_REG_CLASSES }
+#define IRA_COVER_CLASSES { ACR_REGS, GENNONACR_REGS, SPECIAL_REGS,
LIM_REG_CLASSES }

 #define REG_CLASS_FROM_LETTER(C)               \
   (                                            \

Preprocessed libgcc2.i to be attached.


-- 
           Summary: assert with IRA_COVER_CLASSES with singleton
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hp at gcc dot gnu dot org
  GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: crisv32-axis-elf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

------- Comment #1 from hp at gcc dot gnu dot org  2008-10-13 00:16 -------
Created an attachment (id=16487)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16487&action=view)
Preprocessed and reduced test-case

- Configure with --target=crisv32-axis-elf --enable-languages=c (the latter
optional.)
- Build with "make all-gcc"
- Run as gcc/cc1 -O2 -S g.i
- Observe:
/home/hp/g.i: In function '__muldi3':
/home/hp/g.i:22: internal compiler error: in process_bb_node_lives, at
ira-lives.c:677
NB: If you configure using --target=cris-axis-elf, you need to add -march=v32
to the cc1 options.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[vmakarov at redhat dot com]

------- Comment #2 from vmakarov at redhat dot com  2008-10-22 13:51 -------
The problem is in that hard regs (like r10) which have class GENERAL_REGS (from
REGNO_REG_CLASS) are translated into ACR_REGS.  Such translation results in
wrong register pressure calculation for ACR_REGS and failed assert.  IRA
translates classes which are not inside a cover class into the first
intersected cover class and that is ACR_REGS for GENERAL_REGS.  We could
translate to the biggest cover class but I am afraid it would be wrong for
other targets, e.g. INT_FLOAT_REGS for x86 would be translated into FLOAT_REGS
which is more costly than GENERAL_REGS.

So the right solution would be correct finding cover class of specific hard
register.  I hope I'll make a patch today and submit it for approval.

Thanks for reporting the problem and reducing the test.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

------- Comment #3 from hp at gcc dot gnu dot org  2008-10-22 18:02 -------
(In reply to comment #2)
> The problem is in that hard regs (like r10) which have class GENERAL_REGS (from
> REGNO_REG_CLASS) are translated into ACR_REGS.

Oops, I guess I should make that NONACR_REGS; the documentation says the class
should be minimal.  I'll report back on this.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

------- Comment #4 from hp at gcc dot gnu dot org  2008-10-22 22:05 -------
Created an attachment (id=16530)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16530&action=view)
Updated cris.h patch, reflecting corrected REGNO_REG_CLASS together with the
IRA_COVER_CLASSES.

I tried this updated cris.h patch as per previous comment together with
r141228, but it ICE's compiling newlib's ldtoa.c.  See next attachment.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

------- Comment #5 from hp at gcc dot gnu dot org  2008-10-22 22:17 -------
Created an attachment (id=16531)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16531&action=view)
Reduced test-case for updated cris.h at r141228. Compile with -O2.

After fixing REGNO_REG_CLASS (but still with the singleton version of
IRA_COVER_CLASSES), see the cris.h patch, I still get an ICE in IRA, during
build, compiling ldtoa.c from newlib:

dtoa2.c: In function 'emovo':
ldtoa2.c:24: error: unrecognizable insn:
(insn 144 57 140 4 ldtoa2.c:12 (set (reg:HI 13 r13)
        (mem:HI (post_inc:SI (reg/v/f:SI 15 acr [orig:63 p.37 ] [63])) [2 S2
A8])) -1 (expr_list:REG_INC (reg/v/f:SI 15\
 acr [orig:63 p.37 ] [63])
        (nil)))
ldtoa2.c:24: internal compiler error: in extract_insn, at recog.c:2027

The last dump file is from IRA.  It appears IRA (or reload) allocates ACR (reg
15) for an address with post-increment.  It's just that this register can't be
used for that addressing mode - which is one of the reasons it's in a separate
class.  This doesn't happen with the non-singleton IRA_COVER_CLASSES.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hjl at gcc dot gnu dot org]

------- Comment #6 from hjl at gcc dot gnu dot org  2008-10-26 07:16 -------
Subject: Bug 37813

Author: hjl
Date: Sun Oct 26 07:14:49 2008
New Revision: 141372

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141372
Log:
2008-10-25  Vladimir Makarov  <vmakarov@redhat.com>

        PR middle-end/37813
        * ira-conflicts.c (process_regs_for_copy): Remove class subset
        check.

        * ira-int.h (ira_hard_regno_cover_class): New.

        * ira-lives.c (mark_reg_live, mark_reg_dead,
        process_bb_node_lives): Use ira_hard_regno_cover_class.

        * ira.c (reg_class ira_hard_regno_cover_class): New global
        variable.
        (setup_hard_regno_cover_class): New function.
        (ira_init): Call setup_hard_regno_cover_class.

        * ira-costs.c (cost_class_nums): Add comment.
        (find_allocno_class_costs): Initiate cost_class_nums.
        (setup_allocno_cover_class_and_costs): Check cost_class_nums.

Modified:
    branches/ira-merge/gcc/ChangeLog.ira
    branches/ira-merge/gcc/ira-conflicts.c
    branches/ira-merge/gcc/ira-costs.c
    branches/ira-merge/gcc/ira-int.h
    branches/ira-merge/gcc/ira-lives.c
    branches/ira-merge/gcc/ira.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[vmakarov at gcc dot gnu dot org]

------- Comment #7 from vmakarov at gcc dot gnu dot org  2008-10-27 16:25 -------
Subject: Bug 37813

Author: vmakarov
Date: Mon Oct 27 16:24:19 2008
New Revision: 141384

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141384
Log:
2008-10-27  Vladimir Makarov  <vmakarov@redhat.com>

        PR middle-end/37813
        * ira-conflicts.c (process_regs_for_copy): Remove class subset
        check.

        * ira-int.h (ira_hard_regno_cover_class): New.

        * ira-lives.c (mark_reg_live, mark_reg_dead,
        process_bb_node_lives): Use ira_hard_regno_cover_class.

        * ira.c (reg_class ira_hard_regno_cover_class): New global
        variable.
        (setup_hard_regno_cover_class): New function.
        (ira_init): Call setup_hard_regno_cover_class.

        * ira-costs.c (cost_class_nums): Add comment.
        (find_allocno_class_costs): Initiate cost_class_nums.
        (setup_allocno_cover_class_and_costs): Check cost_class_nums.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/ira-conflicts.c
    trunk/gcc/ira-costs.c
    trunk/gcc/ira-int.h
    trunk/gcc/ira-lives.c
    trunk/gcc/ira.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[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=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

------- Comment #9 from hp at gcc dot gnu dot org  2009-01-02 18:44 -------
See comment #5.  The subsequent commits were AFAIK not addressing it.  I don't
have time to experiment right now, and as the bug only appeared (at the time)
with separate patches, I'm making it SUSPENDED.  CC:ers, feel free to upgrade.


-- 

hp 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         |2009-01-02 18:44:36
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813

[Bug middle-end/37813] assert with IRA_COVER_CLASSES with singleton

[hp at gcc dot gnu dot org]

-- 

hp at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |SUSPENDED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37813