workaround for memory error in GAS

[Bob Wilson <bwilson at tensilica dot com>]

I've committed this patch to avoid an infrequent memory error in the  
Xtensa port of GAS.  It looks like the i960 and v850 ports may suffer  
from the same problem, since it occurs when HANDLE_ALIGN is defined  
but MAX_MEM_FOR_RS_ALIGN_CODE is not.  Unfortunately, I do not have a  
testcase to demonstrate the problem, since Sterling tracked this down  
in an older version of GAS and told me that it was difficult to  
reproduce.  (This was happening from subsegs_finish where writing one  
byte past the end of a frag is usually only fatal when it is at the  
end of an obstack chunk.)

The default definition of MAX_MEM_FOR_RS_ALIGN_CODE in this case is:

#  define MAX_MEM_FOR_RS_ALIGN_CODE  ((1 << alignment) - 1)

If the value of "alignment" is zero, the value of  
MAX_MEM_FOR_RS_ALIGN_CODE is also zero, but it needs to be at least  
one byte for the call to frag_var in frag_align_code.  There are at  
least some cases where frag_align_code is called from subsegs_finish  
with an alignment of zero.  A more general solution might be to  
change the default to:

(alignment == 0 ? 1 : (1 << alignment) - 1)

but the Xtensa port should really have MAX_MEM_FOR_RS_ALIGN_CODE set  
to 1 regardless.

2005-12-29  Sterling Augustine  <sterling@tensilica.com>

        * config/tc-xtensa.h (MAX_MEM_FOR_RS_ALIGN_CODE): Define.

Attachment: gas-memerror-fix.diff
Description: Binary data