[Bug libfortran/56737] [4.6/4.7/4.8/4.9 Regression] Wrong I/O result with format cache for Hollerith strings

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56737

--- Comment #3 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2013-03-28 02:34:32 UTC ---
Note the comment at line 728 of format.c where we must disable format caching
when we encounter FMT_STRING.  The problem is related to saving a pointer to a
string rather than the string itself.  That pointer can go out of scope on
subsequent I/O operations so caching the format data will not work.

A similar thing happens with FMT_H. Try this not regression tested yet

Index: format.c
===================================================================
--- format.c    (revision 196516)
+++ format.c    (working copy)
@@ -807,6 +807,7 @@
       goto data_desc;

     case FMT_H:
+      saveit = false;
       get_fnode (fmt, &head, &tail, FMT_STRING);
       if (fmt->format_string_len < 1)
     {
@@ -984,6 +985,7 @@
       break;

     case FMT_H:
+      saveit = false;
       if (repeat > fmt->format_string_len)
     {
       fmt->error = bad_hollerith;


I get:

$ gfortran -g -o gfort_bug gfort_bug-2.f90 && valgrind ./gfort_bug
==22746== Memcheck, a memory error detector
==22746== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==22746== Using Valgrind-3.8.1 and LibVEX; rerun with -h for copyright info
==22746== Command: ./gfort_bug
==22746== 
 fmt = (   3(1H ),6h===== ,a  12,i4,6h =====)                                   
 title1= end of level 
 level =            1
   ===== end of level   1 =====
 fmt = (   3(1H ),6h===== ,a  12,i4,6h =====)                                   
 title1= end of level 
 level =            1
   ===== end of level   1 =====
==22746== 
==22746== HEAP SUMMARY:
==22746==     in use at exit: 0 bytes in 0 blocks
==22746==   total heap usage: 29 allocs, 29 frees, 27,978 bytes allocated
==22746== 
==22746== All heap blocks were freed -- no leaks are possible
==22746== 
==22746== For counts of detected and suppressed errors, rerun with: -v
==22746== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 2 from 2)

I see no burning need to cache these kinds of formats so just disable it.