[Bug tree-optimization/23376] New: ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[Bug tree-optimization/23376] New: ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[drab at kepler dot fjfi dot cvut dot cz]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 2684 bytes --]

The example (attached below), when compiled by following gcc

-------------------
$ gcc -v
Using built-in specs.
Target: i686-pc-linux-gnu
Configured with: ../../../gcc-CVS-20050812/gcc-CVS-20050812/configure
--host=i686-pc-linux-gnu --prefix=/usr/local/opt/gcc-4.1
--exec-prefix=/usr/local/opt/gcc-4.1 --sysconfdir=/etc
--libdir=/usr/local/opt/gcc-4.1/lib --libexecdir=/usr/local/opt/gcc-4.1/libexec
--sharedstatedir=/var --localstatedir=/var --program-suffix=-4.1
--with-x-includes=/usr/X11R6/include --with-x-libraries=/usr/X11R6/lib
--enable-shared --enable-static --with-gnu-as --with-gnu-ld --with-stabs
--enable-threads=posix --enable-version-specific-runtime-libs --disable-coverage
--disable-libgcj --disable-checking --enable-multilib --with-x --enable-cmath
--enable-libstdcxx-debug --enable-fast-character --enable-hash-synchronization
--with-system-zlib --with-libbanshee --with-demangler-in-ld
--with-arch=athlon-xp --enable-libada --enable-languages=c,c++,f95,objc,ada
Thread model: posix
gcc version 4.1.0 20050812 (experimental)
---------------------

with

---------------------
g++ -O1 -funroll-loops -fvariable-expansion-in-unroller -c me_utils_mmx.cpp -o
me_utils_mmx.o
---------------------

results in this:

---------------------
me_utils_mmx.cpp: In function ‘dirac::CalcValueType
dirac::simple_block_diff_up_mmx_4(const dirac::PicArray&, const
dirac::PicArray&, const dirac::ImageCoords&, const dirac::ImageCoords&, const
dirac::ImageCoords&, const dirac::ValueType*)’:
me_utils_mmx.cpp:23841: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.
---------------------

Tested on i686-pc-linux-gnu.

gcc version 4.0.2 20050804 (prerelease) fails the same way.
gcc version 3.4.3 (Mandrakelinux 10.2 3.4.3-7mdk) doesn't compile this at all
because of errors (should be because the included headers are from 4.1), perhaps
after reduction it might.

For x86_64 it fails as well, though it requires a patch (also attached) to
compile, again due to the includes from i686 GCC this time.

-- 
           Summary: ICE on GCC 4.x with -O1 -funroll-loops -fvariable-
                    expansion-in-unroller
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: drab at kepler dot fjfi dot cvut dot cz
                CC: gcc-bugs at gcc dot gnu dot org
  GCC host triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug tree-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[drab at kepler dot fjfi dot cvut dot cz]

------- Additional Comments From drab at kepler dot fjfi dot cvut dot cz  2005-08-13 14:49 -------
Created an attachment (id=9491)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=9491&action=view)
Triggers the bug


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug tree-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[drab at kepler dot fjfi dot cvut dot cz]

------- Additional Comments From drab at kepler dot fjfi dot cvut dot cz  2005-08-13 14:51 -------
Created an attachment (id=9492)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=9492&action=view)
Patch to the original example to compile (and also trigger the bug) for x86_64


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug tree-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   GCC host triplet|i686-pc-linux-gnu           |
 GCC target triplet|                            |i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug rtl-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-08-13 18:04 -------
Reducing, this is not a regression as variable-expansion-in-unroller is new in 4.0.0.  This is also a rtl 
issue and not a tree issue.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |rtl-optimization
           Keywords|                            |ice-on-valid-code
      Known to fail|                            |4.0.0 4.1.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug rtl-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-08-13 19:18 -------
Confirmed, reduced testcase:
typedef int __m64 __attribute__ ((__vector_size__ (8)));
typedef int __v2si __attribute__ ((__vector_size__ (8)));
static __inline __m64 __attribute__((__always_inline__)) _mm_add_pi32 (__m64 __m1, __m64 __m2) {
  return (__m64) __builtin_ia32_paddd ((__v2si)__m1, (__v2si)__m2);
};
__m64 simple_block_diff_up_mmx_4(  const int width ,__m64 ref1 ) {
  __m64 sum;
  int count = width >>1;
  while (count--)
    sum = _mm_add_pi32 (sum, ref1);
  return sum;
}

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2005-08-13 19:18:05
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug rtl-optimization/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-14 11:36 -------
GCC tries to unroll and expand this insn: 
(insn 28 26 30 2 (set (reg/v:V2SI 65 [ sum ]) 
        (plus:V2SI (reg/v:V2SI 65 [ sum ]) 
            (reg/v:V2SI 68 [ ref1 ]))) 1023 {mmx_addv2si3} (nil) 
    (nil)) 
 
But for some reason the force_operand call in combine_var_copies_in_loop_exit 
returns NULL, in other words it fails to synthesize this insn in the unrolled 
copies.  At that point we have: 
 
2088      expr = force_operand (sum, ve->reg); 
(gdb) p debug_rtx(ve->insn) 
(insn 28 26 30 2 (set (reg/v:V2SI 65 [ sum ]) 
        (plus:V2SI (reg/v:V2SI 65 [ sum ]) 
            (reg/v:V2SI 68 [ ref1 ]))) -1 (nil) 
    (nil)) 
$42 = void 
(gdb) p debug_rtx(ve->reg) 
(reg/v:V2SI 65 [ sum ]) 
$43 = void 
(gdb) p debug_rtx(sum) 
(plus:V2SI (reg:V2SI 76) 
    (reg/v:V2SI 65 [ sum ])) 
$44 = void 
(gdb) 
2089      if (expr != ve->reg) 
(gdb) p expr 
$45 = 0x0 
(gdb)  
 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-14 11:40 -------
My uneducated guess is that expand_binop can't find an appropriate obtab.  
That means that this problem is target specific (e.g. missing named pattern). 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|rtl-optimization            |target


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-14 12:22 -------
>From config/i386/mmx.md: 
 
;; Note!  Except for the basic move instructions, *all* of these 
;; patterns are outside the normal optabs namespace.  This is because 
;; use of these registers requires the insertion of emms or femms 
;; instructions to return to normal fpu mode.  The compiler doesn't 
;; know how to do that itself, which means it's up to the user.  Which 
;; means that we should never use any of these patterns except at the 
;; direction of the user via a builtin. 
 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-14 12:26 -------
There is a patch t fix the issue mentioned in that mmx.md comment, see 
http://gcc.gnu.org/ml/gcc-patches/2005-07/msg01128.html 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rth at gcc dot gnu dot org,
                   |                            |uros at gcc dot gnu dot org
  BugsThisDependsOn|                            |19161


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-14 12:55 -------
This patch is a bit paradoxical: There is an insn that we want to expand in 
the unroller, so we know that have_insn_for *should* return true if the MD is 
sane.  But the MD can't be sane for an insane ISA combination like x87/mmx.  
So indeed, the insn that src comes from is valid but the insn does not exist 
according to have_insn_for.  This happens because have_insn_for checks that 
there is an optab for this combination of operation and mode.  The patch is 
obviously suboptimal (Uros' patch would fix the problem for real, someone 
review it please! :-) but probably this kind of precaution is a good idea 
anyway... 
 
Index: loop-unroll.c 
=================================================================== 
RCS file: /cvs/gcc/gcc/gcc/loop-unroll.c,v 
retrieving revision 1.37 
diff -u -3 -p -r1.37 loop-unroll.c 
--- loop-unroll.c       3 Aug 2005 13:34:35 -0000       1.37 
+++ loop-unroll.c       14 Aug 2005 12:51:16 -0000 
@@ -1574,6 +1574,9 @@ analyze_insn_to_expand_var (struct loop 
       && GET_CODE (src) != MINUS 
       && GET_CODE (src) != MULT) 
     return NULL; 
+ 
+  if (!have_insn_for (GET_CODE (src), GET_MODE (src))) 
+    return NULL; 
 
   if (!XEXP (src, 0)) 
     return NULL; 
 
 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2005-08-16 22:24 -------
Subject: Bug 23376

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	steven@gcc.gnu.org	2005-08-16 22:24:30

Modified files:
	gcc            : ChangeLog loop-unroll.c 
	gcc/fortran    : ChangeLog 
	gcc/testsuite  : ChangeLog 
Added files:
	gcc/testsuite/gcc.target/x86_64: pr23376.c 

Log message:
	PR target/23376
	* loop-unroll.c (analyze_insn_to_expand_var): Make sure that
	force_operand will work later on using have_insn_for.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.9750&r2=2.9751
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/loop-unroll.c.diff?cvsroot=gcc&r1=1.37&r2=1.38
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/fortran/ChangeLog.diff?cvsroot=gcc&r1=1.525&r2=1.526
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/ChangeLog.diff?cvsroot=gcc&r1=1.5931&r2=1.5932
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.target/x86_64/pr23376.c.diff?cvsroot=gcc&r1=NONE&r2=1.1



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[steven at gcc dot gnu dot org]

------- Additional Comments From steven at gcc dot gnu dot org  2005-08-16 22:27 -------
Begone, bug. 

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.1.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[drab at kepler dot fjfi dot cvut dot cz]

------- Additional Comments From drab at kepler dot fjfi dot cvut dot cz  2005-08-18 02:51 -------
(In reply to comment #11)
> Begone, bug. 

Have you applied/fixed this also in the 4.0 branch? Because I don't see it
applied there and it segfaults there as well.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-08-18 02:53 -------
(In reply to comment #12)
> Have you applied/fixed this also in the 4.0 branch? Because I don't see it
> applied there and it segfaults there as well.

This was not a regression as -fvariable-expansion-in-unroller is a new option in 4.0.0 so closing it as 
fixed for 4.1.0 is the correct thing to do.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[drab at kepler dot fjfi dot cvut dot cz]

------- Additional Comments From drab at kepler dot fjfi dot cvut dot cz  2005-08-18 10:48 -------
(In reply to comment #13)
> This was not a regression as -fvariable-expansion-in-unroller is a new option in
> 4.0.0 so closing it as fixed for 4.1.0 is the correct thing to do.

I don't say it's a bad thing to close it. I just say it's a bad thing that 4.0
segfaults. So either it should be fixed there as well or the
-fvariable-expansion-in-unroller option should be removed/disabled there.
Especially if you want to make a release, it shouldn't segfault no matter what.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[bonzini at gcc dot gnu dot org]

------- Additional Comments From bonzini at gcc dot gnu dot org  2005-08-18 11:51 -------
Mark, should this patch go into 4.0.2 too?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mark at codesourcery dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376

[Bug target/23376] ICE on GCC 4.x with -O1 -funroll-loops -fvariable-expansion-in-unroller

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2005-08-22 17:32 -------
Yes, I think it's a good idea to put this patch in 4.0.2, given that it is in
mainline, and given that it is so very simple.   However, I don't think it's
critical either way.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23376