diff mbox

[Fortran,committed] Fix Fortran bug affecting MinGW-w64, only

Message ID 53C18936.1070001@net-b.de
State New
Headers show

Commit Message

Tobias Burnus July 12, 2014, 7:15 p.m. UTC 
64bit Windows is special as "long" is only 32bit wide - while pointers 
are 64bit and, hence, wider.

Another special code is gfortran's array descriptor. Its "dtype" field 
is of type ptrdiff_t and it stores the array rank and type in the first 
6 bits and uses the remaining size to store the byte size for an array. 
In case of 32bit systems, one has 32bits - 6 bits = 24 bits (or


There seems to be a single piece of code in gfortran, which assumes that 
a pointer and long have the same width. Namely, the code which checks 
that one can squeeze the array-element size together with the data type 
and the rank into the "dtype" field of the array descriptor, which has 
the type ptrdiff_t.

For a system with 32bit pointer size (= type precision of size_type_node 
or for ptrdiff_t), one has 32 bits minus 6 bits = 26 bits space. Thus, a 
single array element might "only" be 64 mega bytes large. For 64bit 
pointers, 58 bits remain.

For some odd reason, the code does not use size_type_node or something 
equivalent but it uses an unsigned long. On 32bit Windows and on Linux, 
everything is fine. But on 64bit Windows (i.e. MinGW-w64), one tries to 
squeeze 58 bits into a 32bit pointer – which will cause that 
gfc_max_array_element_size has the wrong result.

This patch does the obvious: It masks out the required bits on a 
size_type_node instead.

Committed as obvious as Rev. 212485.

* * *

Unfortunately, that patch does not solve the original problem: On 32bit 
– and now also on 64 bit – Windows, the array and the trailing "done" is 
not printed. One has still to investigate why that fails.

Tobias
diff mbox

Patch

Index: gcc/fortran/ChangeLog
===================================================================
--- gcc/fortran/ChangeLog	(Revision 212484)
+++ gcc/fortran/ChangeLog	(Arbeitskopie)
@@ -1,5 +1,11 @@ 
 2014-07-12  Tobias Burnus  <burnus@net-b.de>
 
+	PR fortran/61628
+	* trans-types.c (gfc_init_types): Fix data-type bug
+	with gfc_max_array_element_size.
+
+2014-07-12  Tobias Burnus  <burnus@net-b.de>
+
 	* libgfortran.h (libcaf_atomic_codes): Add.
 	* trans-decl.c (gfor_fndecl_caf_atomic_def,
 	gfor_fndecl_caf_atomic_ref, gfor_fndecl_caf_atomic_cas,
Index: gcc/fortran/trans-types.c
===================================================================
--- gcc/fortran/trans-types.c	(Revision 212483)
+++ gcc/fortran/trans-types.c	(Arbeitskopie)
@@ -955,9 +955,9 @@  gfc_init_types (void)
 
   n = TYPE_PRECISION (gfc_array_index_type) - GFC_DTYPE_SIZE_SHIFT;
   gfc_max_array_element_size
-    = wide_int_to_tree (long_unsigned_type_node,
+    = wide_int_to_tree (size_type_node,
 			wi::mask (n, UNSIGNED,
-				  TYPE_PRECISION (long_unsigned_type_node)));
+				  TYPE_PRECISION (size_type_node)));
 
   boolean_type_node = gfc_get_logical_type (gfc_default_logical_kind);
   boolean_true_node = build_int_cst (boolean_type_node, 1);