[Bug tree-optimization/35642] New: heisenbug in tree vectorizer

[Bug tree-optimization/35642] New: heisenbug in tree vectorizer

[zadeck at naturalbridge dot com]

with svn version 133159 the following bugs show up in the tree vectorizer if
you do not bootstrap the compiler:

FAIL: gcc.dg/vect/vect-strided-store-u16-i4.c scan-tree-dump-times vect
"vectorized 1 loops" 2
FAIL: gcc.dg/vect/vect-vfa-03.c scan-tree-dump-times vect "vectorized 1 loops"
1
XPASS: gcc.dg/vect/vect-vfa-03.c scan-tree-dump-times vect "vectorized 1 loops"
0
FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorized 4 loops" 1
FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 2

and the bugs go away when you bootstrap.   I have made modifications to
register allocator, and these downstream changes cause the compiler to be
enough different so that the bugs become hard failures.    

however the bugs are easily reproducible by 

--enable-languages=c --with-cpu=default-32 --disable-multilib 
--disable-bootstrap 

with this version of the compiler.  I have not played around with this in other
revisions.  I know that this combo will reproduce the issue.  Since this is
most likely a storage leak or an uninitialized variable, it is important to
recreate the environment exactly.

Kenny


-- 
           Summary: heisenbug in tree vectorizer
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: zadeck at naturalbridge dot com
 GCC build triplet: powerpc-unknown-linux-gnu  32 bit compiler


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2008-03-19 19:15 -------
On PowerPC-darwin, I started to get the following failures between
LAST_UPDATED: Wed Mar 12 03:41:52 UTC 2008 (revision 133132)
and
LAST_UPDATED: Tue Mar 18 21:37:03 UTC 2008 (revision 133327)


 +FAIL: gcc.dg/vect/vect-strided-store-u16-i4.c scan-tree-dump-times vect
"vectorized 1 loops" 2
+FAIL: gcc.dg/vect/vect-vfa-03.c scan-tree-dump-times vect "vectorized 1 loops"
1
+XPASS: gcc.dg/vect/vect-vfa-03.c scan-tree-dump-times vect "vectorized 1
loops" 0
+FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorized 4 loops" 1
+FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 2
+FAIL: gcc.dg/vect/slp-7.c scan-tree-dump-times vect "vectorized 3 loops" 1
+FAIL: gcc.dg/vect/slp-7.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 3

I wonder if this is related to this bug.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2008-03-19 19:17 -------
I wonder if the following patch could have caused the use of a variable after a
free:
2008-03-12  Victor Kaplansky  <victork@il.ibm.com>
+           Ira Rosen  <irar@il.ibm.com>
+
+       * tree-vectorizer.c (free_stmt_vec_info): New function.
+       (destroy_loop_vec_info): Move code to free_stmt_vec_info().
+       Call free_stmt_vec_info(). Free LOOP_VINFO_STRIDED_STORES..
+       * tree-vectorizer.h (free_stmt_vec_info): Declare.
+       * tree-vect-transform.c (vectorizable_conversion): Free
+       vec_oprnds0 if it was allocated.
+       (vect_permute_store_chain): Remove unused VECs.
+       (vectorizable_store): Free VECs that are allocated in the..
+       function.
+       (vect_transform_strided_load, vectorizable_load): Likewise.
+       (vect_remove_stores): Simplify the code.
+       (vect_transform_loop): Move code to vect_remove_stores().
+       Call vect_remove_stores() and free_stmt_vec_info().
+


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[zadeck at naturalbridge dot com]

------- Comment #3 from zadeck at naturalbridge dot com  2008-03-19 19:26 -------
I forgot to mention that valgrind does not find anything.

kenny


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[irar at il dot ibm dot com]

------- Comment #4 from irar at il dot ibm dot com  2008-03-20 09:30 -------
I reproduced the failures with revision 133362 (and without
--disable-multilib). Reverting our patch (revision 133134) didn't help, I still
see the failures even without it.

Ira


-- 

irar at il dot ibm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2008-03-20 09:30:22
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[victork at gcc dot gnu dot org]

------- Comment #5 from victork at gcc dot gnu dot org  2008-03-20 09:42 -------
I see same failures using compiler boostrapped with vectorization enabled using
BOOT_CFLAGS="-g -O2 -ftree-vectorize -maltivec -mabi=altivec"

Tests fail because some loops were not vectorized:
home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
vect_is_simple_use: operand 2
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note: op
not supported by target.
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
vect_is_simple_use: operand D.2114_11 * 2
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
not ssa-name.
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
use not simple.
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
not vectorized: relevant stmt not supported: D.2115_12 = D.2114_11 * 2
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:26: note:
bad operation or unsupported loop bound.
/home/victork/mainline/gcc/gcc/testsuite/gcc.dg/vect/vect-vfa-03.c:23: note:
vectorized 0 loops in function.

In above case vect_is_simple_use() has a problem with constant operand, which
is very strange.  Looks like a memory overwrite.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[victork at gcc dot gnu dot org]

------- Comment #6 from victork at gcc dot gnu dot org  2008-03-20 13:43 -------
After some regression hunting, I see that failures of vectorizer testcases
started after this patch:
------------------------------------------------------------------------
r133144 | bonzini | 2008-03-12 17:33:45 +0200 (Wed, 12 Mar 2008) | 14 lines

2008-03-12  Paolo Bonzini  <bonzini@gnu.org>

        PR tree-opt/35422
        * fold-const.c (fold_unary) <NOP_EXPR>: Distribute a narrowing
        conversion to the operands of a multiplication.

testsuite:
2008-03-12  Paolo Bonzini  <bonzini@gnu.org>

        PR tree-opt/35422
        * gcc.dg/vect/slp-7.c: Change target keywords required for vectorizing
        third loop.
        * gcc.target/i386/pr35422.c: New.


-- 

victork at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |victork at gcc dot gnu dot
                   |                            |org, bonzini at gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[bonzini at gnu dot org]

------- Comment #7 from bonzini at gnu dot org  2008-03-20 13:51 -------
Indeed my patch exposes additional vectorization abilities (which was
unexpected).

Kenny, can you run the failing testcase under valgrind?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[zadeck at naturalbridge dot com]

------- Comment #8 from zadeck at naturalbridge dot com  2008-03-20 13:59 -------
Subject: Re:  heisenbug in tree vectorizer

bonzini at gnu dot org wrote:
> ------- Comment #7 from bonzini at gnu dot org  2008-03-20 13:51 -------
> Indeed my patch exposes additional vectorization abilities (which was
> unexpected).
>
> Kenny, can you run the failing testcase under valgrind?
>
>
>   
I did run it under valgrind and found no problems. 

However, i am a little skeptical about regression hunting for these 
failures.   I have found that they come and go based on small changes to 
other parts of the compiler, i.e. they are symptomatic of some 
uninitialized variable problem or a storage overwriting.  

in particular, with the release that i specified in message 1, the 
failures were there if the compiler was not bootstrapped and were there 
if it was bootstrapped.  Also, changes that i made to that release in 
the register allocator, which is far downstream of the tree vectorizer 
caused the problems to hit for both bootstrapped and not bootstrapped 
builds.   

I would suggest that you go roll back to that release mentioned in 
message 1, where you have exactly the same code base for bootstrapped 
and non bootstrapped compiler and run the two compilers in parallel 
until you find the place where they diverge.

It could very well be that this may be that this is a problem with 
someone elses code, but you are the occam's razor first choice.

kenny


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] heisenbug in tree vectorizer

[bonzini at gnu dot org]

------- Comment #9 from bonzini at gnu dot org  2008-03-20 14:07 -------
Subject: Re:  heisenbug in tree vectorizer

> I would suggest that you go roll back to that release mentioned in 
> message 1, where you have exactly the same code base for bootstrapped 
> and non bootstrapped compiler and run the two compilers in parallel 
> until you find the place where they diverge.

Definitely agreed.

Paolo


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|heisenbug in tree vectorizer|[4.4 Regression] heisenbug
                   |                            |in tree vectorizer
   Target Milestone|---                         |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[rguenth at gcc dot gnu dot org]

------- Comment #10 from rguenth at gcc dot gnu dot org  2008-04-20 20:39 -------
What is this bug?  A missed-optimization?  wrong-code? ...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[zadeck at naturalbridge dot com]

------- Comment #11 from zadeck at naturalbridge dot com  2008-04-20 21:21 -------
Subject: Re:  [4.4 Regression] heisenbug in tree
 vectorizer

rguenth at gcc dot gnu dot org wrote:
> ------- Comment #10 from rguenth at gcc dot gnu dot org  2008-04-20 20:39 -------
> What is this bug?  A missed-optimization?  wrong-code? ...
>
>
>   
the testcases that randomly pass or fail check the dump files to make 
sure that specific transformations have been performed.  So i guess you 
could say that it is a missed optimization but given that this is most 
likely caused by either storage being overwritten or an uninitialized 
variable, it is hard to assure that this does not effect correctness.

kenny


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[rguenth at gcc dot gnu dot org]

------- Comment #12 from rguenth at gcc dot gnu dot org  2008-04-20 21:54 -------
P1 until this is properly analyzed then.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Priority|P3                          |P1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[irar at il dot ibm dot com]

------- Comment #13 from irar at il dot ibm dot com  2008-05-22 05:41 -------
*** Bug 36295 has been marked as a duplicate of this bug. ***


-- 

irar at il dot ibm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |janis at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[irar at il dot ibm dot com]

------- Comment #14 from irar at il dot ibm dot com  2008-07-06 11:54 -------
All those failures occur because vector multiplication of shorts to shorts is
not supported on power. Paolo's patch changed the order of type conversion and
multiplication, and removed type conversions. 

All those loops contain multiplication either of type short = short * constant
or short = int * constant. For the first case we used to have conversion to
int, multiplication, and conversion to short, and now there are no type
conversions. For the second case, before the patch there was first
multiplication of ints and then the result was converted to short, and now the
conversion to short comes first and then there is multiplication of shorts. In
other words, before Paolo's patch those loops contained multiplications of ints
(supported) and now they contain multiplications of shorts (not supported).

As for the heisenbug, first of all, the failures occur with and without
bootstrap. Also Victor tried to reproduced the behavior that Kenny described in
comment #3 (on the same revision and the same configuration), but he saw the
failures with a bootstrapped version too. I.e., we can't reproduce the random
behavior - the failures are deterministic and well understood.

I suggest to open a missed-optimization PR for vector multiplication of shorts
to shorts (seems easy to fix), to xfail the tests, and to close this PR.

Ira


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[irar at il dot ibm dot com]

------- Comment #15 from irar at il dot ibm dot com  2008-07-06 11:55 -------
(In reply to comment #14)
> Also Victor tried to reproduced the behavior that Kenny described in
> comment #3 

comment #1

Sorry,
Ira


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] heisenbug in tree vectorizer

[irar at il dot ibm dot com]

------- Comment #16 from irar at il dot ibm dot com  2008-07-06 11:57 -------
(In reply to comment #15)
> (In reply to comment #14)
> > Also Victor tried to reproduced the behavior that Kenny described in
> > comment #3 
> 
> comment #1

Grrr, it was in the problem description not in the comments.

Sorry again,
Ira


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] short * short multiplication not vectorized on Power

[bonzini at gnu dot org]

------- Comment #17 from bonzini at gnu dot org  2008-07-10 14:05 -------
changing subject then, thanks for the analysis!


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[4.4 Regression] heisenbug  |[4.4 Regression] short *
                   |in tree vectorizer          |short multiplication not
                   |                            |vectorized on Power


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] short * short multiplication not vectorized on Power

[irar at il dot ibm dot com]

-- 

irar at il dot ibm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P1                          |P3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] short * short multiplication not vectorized on Power

[rguenth at gcc dot gnu dot org]

-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] short * short multiplication not vectorized on Power

[irar at gcc dot gnu dot org]

------- Comment #18 from irar at gcc dot gnu dot org  2008-09-07 08:55 -------
Subject: Bug 35642

Author: irar
Date: Sun Sep  7 08:54:00 2008
New Revision: 140083

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=140083
Log:

        PR tree-optimization/35642
        * config/rs6000/altivec.md (mulv8hi3): Implement.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/altivec.md
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/lib/target-supports.exp


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642

[Bug tree-optimization/35642] [4.4 Regression] short * short multiplication not vectorized on Power

[irar at il dot ibm dot com]

------- Comment #19 from irar at il dot ibm dot com  2008-09-07 11:05 -------
Fixed.


-- 

irar at il dot ibm dot com changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35642