http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50221
--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-05-12
12:10:33 UTC ---
The following program illustrates some of the problems:
a) If the comment lines are removed (i.e. a module is used), there is no
valgrind failure and the result is correct. (Note: It requires the patch from
PR 53329 with "ns" replaced by "sym->ns".)
b) The program (as is) shows no valgrind failure, but the assignment is wrong:
"c3", "c3", "c3" instead of "a1", "b2", "c3".
c) If one removes the "save,", the result is as with (b) but valgrind shows
many errors of the form:
Conditional jump or move depends on uninitialised value(s)
at 0x4C2C3A9: memcpy@@GLIBC_2.14
(The same failures one gets for the original program of comment 0.)
Looking at the dump for (b) - also in comparison with (a) -, I fail to see why
one get's ["c1","c1","c1"] - the code looks correct ("S.0" goes from 1 to 3):
__builtin_memcpy ((void *) &(*D.1881)[(S.0 + D.1885) + D.1882],
(void *) &const[S.0 + -1], (unsigned long) D.1887);
In principle, accessing the second argument wrongly should cause that problem.
But that one looks okay. I wonder more about the left as
(*D.1881)[...]
assumes that the compiler knows the size of one element - I am not sure that
that works as ".str" is not yet the right value before the line:
character(kind=1)[0:][1:.str] * restrict D.1881;
!module m
character(len=:), save, allocatable :: str(:)
character(len=2), parameter :: const(3) = ["a1", "b2", "c3"]
!end
!use m
call test()
if(allocated(str)) deallocate(str)
contains
subroutine test()
call doit()
print *, 'strlen=',len(str),' / array size =',size(str)
print '(3a)', '>',str(1),'<'
print '(3a)', '>',str(2),'<'
print '(3a)', '>',str(3),'<'
end subroutine test
subroutine doit()
str = const
end subroutine doit
end
