[Bug c/47947] New: Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[Bug c/47947] New: Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[tkarkha at us dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

           Summary: Varibles of type vector double are not copied
                    correctly in gcc-4.5.1 and gcc-4.6.0
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: critical
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: tkarkha@us.ibm.com
                CC: meissner@linux.vnet.ibm.com


Created attachment 23505
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23505
Tarfile for recreating the bug

When copying one vector double variable to another vector double variable, for
some variables the lower half of the variable is zeroed out.  The zeroing
results in incorrect computations.  This happens for all -O optimization
levels.

A small testcase is provided to recreate the bug.  In the testcase a simple
vector addition with vec_add() results in an incorrect result due to incorrect
copy.  Here is the ouptut from the testcase:
$ ./gccBugTest_GCC               
DEBUG:: u_in0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG:: u_in1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: l_temp0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG::foo:: l_temp1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: in0 HI = <1.000000,0.000000>
DEBUG::foo:: in0 LO = <3.000000,0.000000>
DEBUG::foo:: in1 HI = <5.000000,0.000000>
DEBUG::foo:: in1 LO = <7.000000,0.000000>
DEBUG::foo:: l_result HI = <6.000000,0.000000>
DEBUG::foo:: l_result LO = <10.000000,12.000000>
DEBUG::foo:: result:: l_temp2 = <6.000000,0.000000,10.000000,12.000000>
DEBUG:: u_out = <6.000000,0.000000,10.000000,12.000000>
$ 

The last DEBUG line should be DEBUG:: u_out =
<6.000000,8.000000,10.000000,12.000000>
          ^^^^^^^^

[Bug c/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |powerpc-*-*
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2011.03.01 20:10:28
                 CC|                            |dje at gcc dot gnu.org
               Host|                            |powerpc-*-*
     Ever Confirmed|0                           |1
   Target Milestone|---                         |4.6.0
              Build|                            |powerpc-*-*

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

Pat Haugen <pthaugen at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pthaugen at gcc dot gnu.org

--- Comment #1 from Pat Haugen <pthaugen at gcc dot gnu.org> 2011-03-01 21:51:09 UTC ---
I checked to see if this is the same issue as PR47862, and it doesn't appear to
be so. I was unable to duplicate the problem on Linux with trunk and ibm-4.5
branch, so looking AIX specific.

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.6.0                       |---

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #2 from David Edelsohn <dje at gcc dot gnu.org> 2011-03-02 22:08:51 UTC ---
Did you build the application as a 32 bit or 64 bit application?  If you used
the default 32 bit on AIX did you try using -maix64?

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #3 from David Edelsohn <dje at gcc dot gnu.org> 2011-03-02 22:29:43 UTC ---
I can reproduce this bug.  It occurs in AIX 32 bit mode, but not in AIX 64 bit
mode.

Linux 32 bit mode ABI is different than AIX 32 bit ABI, so it is more difficult
for LTC to reproduce.  They are going to try Linux 32 bit -fpic/-fPIC.

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[tkarkha at us dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #4 from Tejas Karkhanis <tkarkha at us dot ibm.com> 2011-03-02 22:35:01 UTC ---
David,

I tried building with -maix64 flag.  Here is the output after compiling with
-maix64 flag:

gcc-4.6.0 -Wall -O3 -I ./ -std=gnu99 -maltivec -mabi=altivec
-flax-vector-conversions -mvsx -mcpu=power7 -mtune=power7 -maix64 gccBugTest.c
-o gccBugTest_GCC
$ ./gccBugTest_GCC
DEBUG:: u_in0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG:: u_in1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: l_temp0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG::foo:: l_temp1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: in0 HI = <1.000000,0.000000>
DEBUG::foo:: in0 LO = <3.000000,0.000000>
DEBUG::foo:: in1 HI = <5.000000,0.000000>
DEBUG::foo:: in1 LO = <7.000000,0.000000>
DEBUG::foo:: l_result HI = <6.000000,8.000000>
DEBUG::foo:: l_result LO = <10.000000,0.000000>
DEBUG::foo:: result:: l_temp2 = <6.000000,8.000000,10.000000,0.000000>
DEBUG:: u_out = <6.000000,8.000000,10.000000,0.000000>
$ 

Notice that this time the error is in the right-most element of the variable
u_out -- the right-most element should have been 12.000000.

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #5 from David Edelsohn <dje at gcc dot gnu.org> 2011-03-02 22:48:10 UTC ---
The bug does not occur with -O1.  It does occur with -O2 and above, both 32 bit
and 64 bit AIX.

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #6 from David Edelsohn <dje at gcc dot gnu.org> 2011-03-02 22:51:04 UTC ---
Created attachment 23520
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23520
Assembly output from testcase

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

Pat Haugen <pthaugen at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.6.0

--- Comment #7 from Pat Haugen <pthaugen at gcc dot gnu.org> 2011-03-02 23:01:59 UTC ---
-fpic/-fPIC don't cause a failure on Linux either. I consistently get the
following output no matter what I've tried.

[pthaugen@ltcgranola3p01 hmmer]$ a.out
DEBUG:: u_in0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG:: u_in1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: l_temp0 = <1.000000,2.000000,3.000000,4.000000>
DEBUG::foo:: l_temp1 = <5.000000,6.000000,7.000000,8.000000>
DEBUG::foo:: in0 HI = <1.000000,2.000000>
DEBUG::foo:: in0 LO = <3.000000,4.000000>
DEBUG::foo:: in1 HI = <5.000000,6.000000>
DEBUG::foo:: in1 LO = <7.000000,8.000000>
DEBUG::foo:: l_result HI = <6.000000,8.000000>
DEBUG::foo:: l_result LO = <10.000000,12.000000>
DEBUG::foo:: result:: l_temp2 = <6.000000,8.000000,10.000000,12.000000>
DEBUG:: u_out = <6.000000,8.000000,10.000000,12.000000>

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #8 from Pat Haugen <pthaugen at gcc dot gnu.org> 2011-03-02 23:03:36 UTC ---
(In reply to comment #6)
> Created attachment 23520 [details]
> Assembly output from testcase

David,
Can you post your output you get from this run, since we've seen variations
now.

[Bug target/47947] Varibles of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

Pat Haugen <pthaugen at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.6.0                       |---

[Bug target/47947] Variables of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #9 from David Edelsohn <dje at gcc dot gnu.org> 2011-03-03 01:48:17 UTC ---
The assembly output produces the same results that Tejas shows in comment #4.

[Bug target/47947] Variables of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #10 from Pat Haugen <pthaugen at gcc dot gnu.org> 2011-03-03 22:12:12 UTC ---
This is looking like a dup of PR47862, note the following snippet of assembler.

    stfd 0,360(1)     #,
    stfd 12,344(1)     #,
    stfd 13,352(1)     #,
    stfd 11,336(1)     #,
...
    bl .printf     #
    nop
    lfd 11,336(1)     #,
    addi 9,1,384     #,,
    addi 3,31,176     #, tmp156,
    stxvd2x 11,0,9     # in0$hiVectorDouble,
...
    bl .printf     #
    nop
    lfd 13,352(1)     #,
    addi 9,1,384     #,,
    addi 3,31,208     #, tmp156,
    stxvd2x 13,0,9     # in0$loVectorDouble,

0,11,12,13 all hold vector values coming into this snippet, and you can see
we're just saving/restoring part of the vector across the printf calls.

[Bug target/47947] Variables of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

--- Comment #11 from Pat Haugen <pthaugen at gcc dot gnu.org> 2011-03-07 20:10:02 UTC ---
PR47862 has been fixed on trunk and 4.5 development branch. Can you verify if
that fixes this problem also?

[Bug target/47947] Variables of type vector double are not copied correctly in gcc-4.5.1 and gcc-4.6.0

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47947

David Edelsohn <dje at gcc dot gnu.org> changed:

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

--- Comment #12 from David Edelsohn <dje at gcc dot gnu.org> 2011-07-25 14:25:26 UTC ---
Fixed with patch for PR 47862