[Bug fortran/40037] New: gfortran -O3 optimization generates code that seg faults on unaligned array data

[Bug fortran/40037] New: gfortran -O3 optimization generates code that seg faults on unaligned array data

[mick at nag dot co dot uk]

The program below compiled with gfortran 4.4.0 (and 4.3.2) and using the
-O3 optimization flag generates, in subroutine SMALL, sse instructions
which assume that the data array C is aligned on a 16 byte boundary.
For example:  mulpd  (%rax,%rcx,1),%xmm0
where register %rax may not be 16-byte aligned.

This leads to a segmentation violation at run time.

The test program deliberately forces array C to be unaligned before
being passed to subroutine SMALL. In a real situation one would
not do this, of course, but when the memory holding array C comes from
somewhere else (e.g. a program written in a different language) there
is no guarantee of alignment.

gfortran version 4.2.0 did not display this problem - the code
generated for SMALL did not assume alignment.

C     Declare a REAL array as workspace. We pass it to a routine
C     expecting a DOUBLE PRECISION array. This allows us to pass
C     data aligned or not aligned on a 16 byte boundary.
      real c(100)
      integer m
      external tt
      m = 2
C     Call tt with array C aligned on 16 byte boundary
      call tt(m, c(1))
C     Call tt with array C not aligned on 16 byte boundary
      call tt(m, c(2))
      end

      subroutine tt(m, c)
      double precision c(m,m)
      external small
      do i = 1, m
         do j = 1, m
            c(i,j) = 1.0d0
         end do
      end do
      call small(m, c)
      write (*,*) 'After call small(m, c), c(1,1) = ', c(1,1)
      end

      subroutine small(m, c)
      integer          m
      double precision c(m,*)
      integer          i
      do 70, i = 1, m
         c(i,1) = c(i,1) + c(i,1)
  70  continue
      return
      end

Here's the output from the -v flag of the compiler:

% gfortran -v -O3 align.f
Driving: gfortran -v -O3 align.f -lgfortranbegin -lgfortran -lm -shared-libgcc
Using built-in specs.
Target: x86_64-unknown-linux-gnu
Configured with: /opt/gcc-4.4.0-build/gcc-4.4.0/configure
--enable-languages=c,fortran,c++ --prefix=/opt/gcc-4.4.0
--with-mpfr=/opt/mpfr-2.4.1 --with-gmp=/opt/gmp-4.1.4
Thread model: posix
gcc version 4.4.0 (GCC) 
COLLECT_GCC_OPTIONS='-v' '-O3' '-shared-libgcc' '-mtune=generic'
 /opt/gcc-4.4.0/libexec/gcc/x86_64-unknown-linux-gnu/4.4.0/f951 align.f
-ffixed-form -quiet -dumpbase align.f -mtune=generic -auxbase align -O3
-version -fintrinsic-modules-path
/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/finclude -o
/tmp/cc9eHSWv.s
GNU Fortran (GCC) version 4.4.0 (x86_64-unknown-linux-gnu)
        compiled by GNU C version 4.4.0, GMP version 4.1.4, MPFR version 2.4.1.
warning: GMP header version 4.1.4 differs from library version 4.2.
warning: MPFR header version 2.4.1 differs from library version 2.2.1.
GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
COLLECT_GCC_OPTIONS='-v' '-O3' '-shared-libgcc' '-mtune=generic'
 as -V -Qy -o /tmp/cc05mKcU.o /tmp/cc9eHSWv.s
GNU assembler version 2.19 (x86_64-unknown-linux-gnu) using BFD version (GNU
Binutils) 2.19
COMPILER_PATH=/opt/gcc-4.4.0/libexec/gcc/x86_64-unknown-linux-gnu/4.4.0/:/opt/gcc-4.4.0/libexec/gcc/x86_64-unknown-linux-gnu/4.4.0/:/opt/gcc-4.4.0/libexec/gcc/x86_64-unknown-linux-gnu/:/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/:/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/
LIBRARY_PATH=/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/:/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/../../../../lib64/:/lib/../lib64/:/usr/lib/../lib64/:/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/../../../:/lib/:/usr/lib/
COLLECT_GCC_OPTIONS='-v' '-O3' '-shared-libgcc' '-mtune=generic'
 /opt/gcc-4.4.0/libexec/gcc/x86_64-unknown-linux-gnu/4.4.0/collect2
--eh-frame-hdr -m elf_x86_64 -dynamic-linker /lib64/ld-linux-x86-64.so.2
/usr/lib/../lib64/crt1.o /usr/lib/../lib64/crti.o
/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/crtbegin.o
-L/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0
-L/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/../../../../lib64
-L/lib/../lib64 -L/usr/lib/../lib64
-L/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/../../..
/tmp/cc05mKcU.o -lgfortranbegin -lgfortran -lm -lgcc_s -lgcc -lc -lgcc_s -lgcc
/opt/gcc-4.4.0/lib/gcc/x86_64-unknown-linux-gnu/4.4.0/crtend.o
/usr/lib/../lib64/crtn.o


-- 
           Summary: gfortran -O3 optimization generates code that seg faults
                    on unaligned array data
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: mick at nag dot co dot uk
 GCC build triplet: x86_64-unknown-linux-gnu
  GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: x86_64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40037

[Bug fortran/40037] gfortran -O3 optimization generates code that seg faults on unaligned array data

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2009-05-06 08:34 -------
I think this is not a valid Fortran program.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40037

[Bug fortran/40037] gfortran -O3 optimization generates code that seg faults on unaligned array data

[mick at nag dot co dot uk]

------- Comment #2 from mick at nag dot co dot uk  2009-05-06 08:45 -------
(In reply to comment #1)
> I think this is not a valid Fortran program.
> 

The main program is illegal because it passes a REAL array argument
to a subroutine expecting DOUBLE PRECISION array. However, the subroutine
SMALL is not illegal, but causes bad code to be generated. The main
program is only present for easy demonstration of the bug in SMALL.

Mick Pont


-- 

mick at nag dot co dot uk changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |UNCONFIRMED
         Resolution|INVALID                     |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40037

[Bug fortran/40037] gfortran -O3 optimization generates code that seg faults on unaligned array data

[jakub at gcc dot gnu dot org]

------- Comment #3 from jakub at gcc dot gnu dot org  2009-05-06 09:18 -------
The code generated for SMALL is correct, any caller that passes argument not
aligned on 8 byte boundary (you are mistaken that it requires 16 byte
alignment,
try calling it with c(3) which is 8 byte aligned, but not 16 byte aligned and
it will work too) is invalid.  All DOUBLE PRECISION variables/arrays must be
properly aligned on 8 byte boundaries.


-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40037

[Bug fortran/40037] gfortran -O3 optimization generates code that seg faults on unaligned array data

[hjl dot tools at gmail dot com]

------- Comment #4 from hjl dot tools at gmail dot com  2009-05-06 17:39 -------
(In reply to comment #3)
> The code generated for SMALL is correct, any caller that passes argument not
> aligned on 8 byte boundary (you are mistaken that it requires 16 byte
> alignment,
> try calling it with c(3) which is 8 byte aligned, but not 16 byte aligned and
> it will work too) is invalid.  All DOUBLE PRECISION variables/arrays must be
> properly aligned on 8 byte boundaries.
> 

I believe x86-64 psABI requires that all arrays >= 16byte must be aligned
at 16byte.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40037