[Bug fortran/51904] [4.6/4.7 Regression] ICE on SIZE function evaluation

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51904

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-01-19 
16:50:33 UTC ---
Fails:
  gcc version 4.6.0 20100828 (experimental) [trunk revision 163612] (GCC)
Works:
  gcc version 4.6.0 20100716 (experimental) [trunk revision 162255] (GCC)
(Note: The builds might have local patches.)


I think it's due to the patch Rev. 162648 of 2010-07-28
  http://gcc.gnu.org/viewcvs?view=revision&revision=162648

That patch tries to compile-time simplify SIZE but also changes for
nonconstant-bound arrays, e.g., SIZE(X ** 2) to  SIZE(X).

Possibly the additional resolution only happens if also a symtree is available?
That would be a change to gfc_build_intrinsic_call.

If one replaces
  print *, size(f*x)
by
  i = size(f*x)
  print *, i
the program works and one finds in the dump:
  i = (integer(kind=4))
        MAX_EXPR <(D.1879->dim[0].ubound - D.1879->dim[0].lbound) + 1, 0>;