[Bug c/25371] New: -ftree-vectorize results in internal compiler error on AMD64

[Bug c/25371] New: -ftree-vectorize results in internal compiler error on AMD64

[David dot Monniaux at ens dot fr]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1792 bytes --]

Same problem occurs in 4.0.1, 4.1 (Subversion) but not in 4.2 (Subversion).

$ gcc -c -ftree-vectorize -O2 oct_makes_gcc_explode.i
oct_makes_gcc_explode.i: In function ‘slow_close’:
oct_makes_gcc_explode.i:8: internal compiler error: in gimplify_expr, at
gimplify.c:4182

$ cat oct_makes_gcc_explode.i
typedef double num_t;
typedef long size_t;
void *octfag_mm_malloc (size_t t);
void octfag_mm_free (void* p);

void
slow_close(num_t* m, int n)
{
  int i,j,k;
  num_t* mm = ((num_t*) octfag_mm_malloc (sizeof(num_t)*(4*n*n)));
  num_t a,b,c,d;
  ;
  ; ; ; ;
  for (k=0;k<2*n;k+=2) {
    for (i=0;i<2*n;i++)
      for (j=0;j<2*n;j++) {
 *(&a) = *(m+i*2*n+k) + *(m+k*2*n+j);
 *(&b) = *(m+i*2*n+k+1) + *(m+(k+1)*2*n+j);
 *(&c) = *(m+i*2*n+k) + *(m+k*2*n+k+1); *(&c) = *(&c) + *(m+(k+1)*2*n+j);
 *(&d) = *(m+i*2*n+k+1) + *(m+(k+1)*2*n+k); *(&d) = *(&d) + *(m+k*2*n+j);
 num_min(&a,&a,&b);
 num_min(&c,&c,&d);
 num_min(&a,&a,&c);
 num_min(mm+i*2*n+j,mm+i*2*n+j,&a);
      }
    for (i=0;i<2*n;i++) *(mm+i*2*n+i) = (num_t)(0);
    for (i=0;i<2*n;i++)
      for (j=0;j<2*n;j++) {
 *(&a) = *(mm+i*2*n+(i^1)) / 2.;
 *(&b) = *(mm+(j^1)*2*n+j) / 2.;
 *(&a) = *(&a) + *(&b);
 num_min(mm+i*2*n+j,mm+i*2*n+j,&a);
      }
  }
  ;
  ; ; ; ;
  octfag_mm_free(mm);
}


-- 
           Summary: -ftree-vectorize results in internal compiler error on
                    AMD64
           Product: gcc
           Version: 4.0.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: David dot Monniaux at ens dot fr
 GCC build triplet: x86_64-redhat-linux
  GCC host triplet: x86_64-redhat-linux
GCC target triplet: x86_64-redhat-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[David dot Monniaux at ens dot fr]

------- Comment #1 from David dot Monniaux at ens dot fr  2005-12-12 16:24 -------
The above error message was for 4.0.1. On 4.0.2 it's gimplify.c:4186.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2005-12-12 18:47 -------
We have a polynomial_chrec here for some reason.
(gdb) p debug_generic_expr(*expr_p)
{0, +, nD.1614_23}_4


-- 

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         |2005-12-12 18:47:27
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[irar at il dot ibm dot com]

------- Comment #3 from irar at il dot ibm dot com  2005-12-18 08:15 -------
I failed to reproduce this ICE on ppc and i686.
Vectorizer's dump file can help.


-- 

irar at il dot ibm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |irar at il dot ibm dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[sebastian dot pop at cri dot ensmp dot fr]

------- Comment #4 from sebastian dot pop at cri dot ensmp dot fr  2006-01-05 16:32 -------
(In reply to comment #3)
> Vectorizer's dump file can help.
> 

vect dump ends with:

/home/seb/ex/pr25371.c:26: note: Access function of PHI: {0, +,
1}_4(get_loop_exit_condition 
  if (i_99 < D.1700_228) goto <L54>; else goto <L55>;)

/home/seb/ex/pr25371.c:26: note: === vec_transform_loop ===
/home/seb/ex/pr25371.c:26: note: === vect_do_peeling_for_alignment ===

and the problem seems to come from the fact that in
vect_create_addr_base_for_vector_ref, we're calling 
base_offset = force_gimple_operand (base_offset, &new_stmt, false, dest);  
with a base_offset that contains a chrec:
((num_tD.1607 *) (long unsigned intD.4) {0, +, nD.1614_23}_4 * 16B);


-- 

sebastian dot pop at cri dot ensmp dot fr changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |sebastian dot pop at cri dot
                   |                            |ensmp dot fr


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[harsha dot jagasia at amd dot com]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1916 bytes --]



------- Comment #5 from harsha dot jagasia at amd dot com  2007-03-01 23:18 -------
I could reproduce this with 4.1, 4.2, 4.3 and the autovect branch (from svn) on
x86-64 (but not on x86):

obj-4.1/bin/gcc -c -ftree-vectorize -O2 oct_makes_gcc_explode.i
oct_makes_gcc_explode.i: In function âslow_closeâ:
oct_makes_gcc_explode.i:8: internal compiler error: in gimplify_expr, at
gimplify.c:4519
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.

obj-4.2/bin/gcc -c -ftree-vectorize -O2 oct_makes_gcc_explode.i
oct_makes_gcc_explode.i: In function âslow_closeâ:
oct_makes_gcc_explode.i:8: internal compiler error: in gimplify_expr, at
gimplify.c:5860
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.

obj-4.3/bin/gcc -c -ftree-vectorize -O2 oct_makes_gcc_explode.i
oct_makes_gcc_explode.i: In function âslow_closeâ:
oct_makes_gcc_explode.i:8: internal compiler error: in gimplify_expr, at
gimplify.c:5917
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.

obj-autovect/bin/gcc -c -ftree-vectorize -O2 oct_makes_gcc_explode.i
oct_makes_gcc_explode.i: In function âslow_closeâ:
oct_makes_gcc_explode.i:8: internal compiler error: in gimplify_expr, at
gimplify.c:5933
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.


-- 

harsha dot jagasia at amd dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |harsha dot jagasia at amd
                   |                            |dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[irar at il dot ibm dot com]

------- Comment #6 from irar at il dot ibm dot com  2007-03-11 10:33 -------
Harsha, could you please attach vectorizer's dump file (produced with
-fdump-tree-vect-details)?

Thanks,
Ira


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[spop at gcc dot gnu dot org]

------- Comment #7 from spop at gcc dot gnu dot org  2007-05-08 21:19 -------
Subject: Re:  -ftree-vectorize results in internal compiler error on AMD64

On 5/8/07, Sjodin, Jan <Jan.Sjodin@amd.com> wrote:
> > Okay, I looked at the code, and the problem is that we have to pass to
> > force_gimple_operand an expression that is not a chain of recurrence,
> > ie. a real expression that can be used to be decomposed in 3 addr code
> > by the gimplifier.
>
> Thanks! That was the answer I was looking for. Sounds like the
> vectorizer is biting off more than it can chew :) Perhaps there is a
> check missing for chrecs somewhere.

Yes, I also think that the quick fix is to add to the vectorizer the missing
checks for simple enough expressions.

I went back to look at the testcase, and I saw these nested loops,
and we are trying to vectorize the innermost ones.  So for avoiding a
failed vectorization, and making the vectorizer to produce the right code,
we want to generate some code for the base address in the outer loop.

Sebastian


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[Jan dot Sjodin at amd dot com]

------- Comment #8 from Jan dot Sjodin at amd dot com  2007-05-08 21:30 -------
Subject: RE:  -ftree-vectorize results in
 internal compiler error on AMD64

I would prefer to make it work instead of disabling the vectorizer for
these cases.

- Jan

> -----Original Message-----
> From: sebpop@gmail.com [mailto:sebpop@gmail.com] On Behalf Of
Sebastian
> Pop
> Sent: Tuesday, May 08, 2007 3:19 PM
> To: gcc-bugzilla@gcc.gnu.org
> Cc: Sjodin, Jan
> Subject: Re: [Bug tree-optimization/25371] -ftree-vectorize results in
> internal compiler error on AMD64
> 
> On 5/8/07, Sjodin, Jan <Jan.Sjodin@amd.com> wrote:
> > > Okay, I looked at the code, and the problem is that we have to
pass to
> > > force_gimple_operand an expression that is not a chain of
recurrence,
> > > ie. a real expression that can be used to be decomposed in 3 addr
code
> > > by the gimplifier.
> >
> > Thanks! That was the answer I was looking for. Sounds like the
> > vectorizer is biting off more than it can chew :) Perhaps there is a
> > check missing for chrecs somewhere.
> 
> Yes, I also think that the quick fix is to add to the vectorizer the
> missing
> checks for simple enough expressions.
> 
> I went back to look at the testcase, and I saw these nested loops,
> and we are trying to vectorize the innermost ones.  So for avoiding a
> failed vectorization, and making the vectorizer to produce the right
code,
> we want to generate some code for the base address in the outer loop.
> 
> Sebastian
> 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[spop at gcc dot gnu dot org]

------- Comment #9 from spop at gcc dot gnu dot org  2007-05-10 08:24 -------
Subject: Re:  -ftree-vectorize results in internal compiler error on AMD64

I was still thinking about how to fix this, and the solution seems to me
to not instantiate the initial value of the induction variable if it is a
chrec,
i.e. keep the base_offset with references to the variable that varies in
the outer loop instead of analyzing that variable and replacing it with its
evolution.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[sebastian dot pop at cri dot ensmp dot fr]

------- Comment #10 from sebastian dot pop at cri dot ensmp dot fr  2007-05-10 09:32 -------
Subject: Re:  -ftree-vectorize results in internal compiler error on AMD64

Zdenek's patch for cleaning the dataref analysis is also fixing this bug.
http://gcc.gnu.org/ml/gcc-patches/2007-05/msg00634.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[rakdver at gcc dot gnu dot org]

------- Comment #11 from rakdver at gcc dot gnu dot org  2007-05-10 10:46 -------
Simplified testcase:

void
slow_close(int n)
{
  int i;
  double *mm;
  for (i=0;i<2*n;i++)
    for (i=0;i<2*n;i++)
      *(mm+i*2*n+i) = 0;
}


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[dorit at il dot ibm dot com]

------- Comment #12 from dorit at il dot ibm dot com  2007-07-01 09:30 -------
> Subject: Re:  -ftree-vectorize results in internal compiler error on AMD64
> Zdenek's patch for cleaning the dataref analysis is also fixing this bug.
> http://gcc.gnu.org/ml/gcc-patches/2007-05/msg00634.html

So now that Zdenek's patch went in, can someone confirm if this problem doesn't
occur anymore on x86_64?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[ubizjak at gmail dot com]

------- Comment #13 from ubizjak at gmail dot com  2007-07-01 09:52 -------
(In reply to comment #12)
> > Subject: Re:  -ftree-vectorize results in internal compiler error on AMD64
> > Zdenek's patch for cleaning the dataref analysis is also fixing this bug.
> > http://gcc.gnu.org/ml/gcc-patches/2007-05/msg00634.html
> 
> So now that Zdenek's patch went in, can someone confirm if this problem doesn't
> occur anymore on x86_64?

Compiles OK for original and reduced testcase (-O2 -ftree-vectorize) with:

Target: x86_64-unknown-linux-gnu
gcc version 4.3.0 20070622 (experimental)

Also works OK when (-m32 -msse3) is added to activate 32bit compilation.

BTW: I plan to add the testcase from Comment #11 to vectorizer testsuite.

Also note that this PR is not marked as a regression on 4.0/4.1/4.2 so it
should be either marked as a regression or should be closed as the testcase
compiles OK.
> 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[dorit at gcc dot gnu dot org]

------- Comment #14 from dorit at gcc dot gnu dot org  2007-07-01 10:12 -------
> > So now that Zdenek's patch went in, can someone confirm if this problem doesn't
> > occur anymore on x86_64?
> Compiles OK for original and reduced testcase (-O2 -ftree-vectorize) with:
> Target: x86_64-unknown-linux-gnu
> gcc version 4.3.0 20070622 (experimental)
> Also works OK when (-m32 -msse3) is added to activate 32bit compilation.

thanks for checking!

> BTW: I plan to add the testcase from Comment #11 to vectorizer testsuite.

good idea

> Also note that this PR is not marked as a regression on 4.0/4.1/4.2 so it
> should be either marked as a regression or should be closed as the testcase
> compiles OK.
> >

I vote for closing it... maybe you can close it after committing the testcase? 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[uros at gcc dot gnu dot org]

------- Comment #15 from uros at gcc dot gnu dot org  2007-07-01 10:30 -------
Subject: Bug 25371

Author: uros
Date: Sun Jul  1 10:30:31 2007
New Revision: 126163

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=126163
Log:
       PR tree-optimization/25371
       * gcc.dg/vect/pr25371.c: New test.


Added:
    trunk/gcc/testsuite/gcc.dg/vect/pr25371.c
Modified:
    trunk/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371

[Bug tree-optimization/25371] -ftree-vectorize results in internal compiler error on AMD64

[ubizjak at gmail dot com]

------- Comment #16 from ubizjak at gmail dot com  2007-07-01 10:32 -------
Closed, not marked as a regression.


-- 

ubizjak at gmail dot com changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25371