[Patch, fortran] PR22038, PR25072 and PR28119 - things broken in forall statements

[Paul Thomas]

:ADDPATCH fortran:

The point of this contribution from Joost VandeVondele

  INTEGER :: I, A(10)
  FORALL (I = 1:10, T (I)) A(I) = I
CONTAINS
  LOGICAL FUNCTION T (I)
    INTEGER :: I
    T = (MOD (I, 2) == 0)
  END FUNCTION T
END

was that the function referenced in the forall mask should be pure.

However, gfortran responds with

 Internal Error at (1):
 free_expr0(): Bad expr type

whether or not T is pure!

This turns out to be due to the attempt to match the mask to a forall
iterator.  Since T is internal and not referenced, match_variable
fixes it up as having flavor FL_VARIABLE, which causes the mask not
to match.  Two regtestable fixes were developed in match_variable
and gfc_match_rvalue but I decided that slightly ad-hoc fixes in these
functions for a very specific problem was not a good idea.  Instead,
the proposed patch is to the cleanup in match_forall_iterator. If a
symtree is created for the iterator variable, which is detected by
the expression pointing to the symtree and the symbol having only 1
reference, this is set to FL_UNKNOWN, so that the matching of the
mask works.

Having fixed this, we find that the original problem is still
present; ie. we need to signal an error if T is not pure. This is
done in resolve.c by a small extension to the flagging system for
non-pure references in the forall-block.

I have not touched subroutine references since I do not think that
non-PURE subroutines can be referenced from the mask expression
(If contained in PURE functions, an error is signalled for a 
non-PURE reference from within a PURE procedure.)

Finally, 

  FORALL (I = 1:10, T (I)) A(I) = I ; any_stmt_at_all

generated

  1
 Internal Error at (1):
 free_expr0(): Bad expr type

and this was caused by an extraneous gfc_match_eos.  This function
is already carried out by the match to the assignment.  Removing
this cures the problem.

Regtested on FC5/Athlon - OK for trunk and 4.1?

Paul

2006-06-22  Paul Thomas  <pault@gcc.gnu.org>

	PR fortran/22038
	* match.c (match_forall_iterator): Mark new variables as
	FL_UNKNOWN if the match fails.

	PR fortran/28119
	* match.c (gfc_match_forall): Remove extraneous call to
	gfc_match_eos.

	PR fortran/25072
	* resolve.c (resolve_code, resolve_function): Rework
	forall_flag scheme so that it is set and has a value of
	2, when the code->expr (ie. the forall mask) is resolved.
	This is used to change "block" to "mask" in the non-PURE
	error message.

2006-06-22  Paul Thomas  <pault@gcc.gnu.org>

	PR fortran/22038
	PR fortran/28119
	* gfortran.dg/forall_4.f90: New test.

	PR fortran/25072
	* gfortran.dg/forall_5.f90: New test.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr25072_2.diff
Type: text/x-patch
Size: 2052 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060622/15c61cad/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: forall_4.f90
Type: text/x-fortran
Size: 1703 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060622/15c61cad/attachment-0001.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: forall_5.f90
Type: text/x-fortran
Size: 967 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060622/15c61cad/attachment-0002.bin>