[PATCH] PR fortran/78033 -- This was a REAL pain

[Steve Kargl]

All,

The attached patch fixes PR fortran/78033.  This was a REAL pain
to fix because Fortran overloads REAL as an intrinsic type and
an intrinsic subprogram.

gfc_match_type_spec() in match.c is used to match Fortran 2003
type-specs in things like array constructors and TYPE IS statements.
At some point in time, PR fortran/54730 was submitted because an ICE
occurred for

  subroutine s
    implicit none
    intrinsic :: real
    real :: vec(1:2)
    vec = (/ real(a = 1), 1. /)
  end subroutine s

where a symbol for 'a' was create while parsing for a validate
typespec.  The invalid 'a' was causing an ICE during translation.
Mikael fixed the ICE by introducing checkpointing of the symbols in
gfc_match_array_constructor() in array.c, which allowed 'a' to be
removed.

Fast-forward to PR fortran/78033, submitted a few days ago.
Code like 

  subroutine f(n, x)
     integer, intent(in) :: n 
     complex, intent(in) :: x(1:n)
     real :: y(2*n)
     y = [real(x(1:n), aimag(x(1:n))]
  end subroutine f

was now ICE'ing due what appears to be a tangling checkpoint.

f951: internal compiler error: in enforce_single_undo_checkpoint,
at fortran/symbol.c:3514

If I disabled, Mikael's fix for PR fortran/54730 then PR fortran/78033
would compile with the expected regression with PR fortran/54730.  Having
spent to much time looking for a mismatch in checkpoints, I decided to
remove Mikael's fix in gfc_match_array_constructor() and fix the issue
in gfc_match_type_spec() where I special case the parsing of
REAL([KIND]=scalar-int-initialization-expr).

An early version of the patch passed regression except for gomp/udr3.f90.
Note that gfortran never vists gfc_match_type_spec while compling udr3.f90.
I've deleted obj/ and started a clean bootstrap to see if this failure
was collateral damage for my tinkering.  If regression testing is 
successfull, OK to commit?

2016-10-21  Steven G. Kargl  <kargl@gcc.gnu.org>

	PR fortran/78033
	* array.c (gfc_match_array_constructor): Remove checkpointing
	introduced in r196416.  Move initialization to top of function.
	* match.c (gfc_match_type_spec): Special case matching for REAL.

2016-10-21  Steven G. Kargl  <kargl@gcc.gnu.org>

	PR fortran/78033
	* gfortran.dg/pr78033.f90: New test.

-- 
Steve
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr78033.diff
Type: text/x-diff
Size: 6414 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20161022/02001f5e/attachment.bin>