On Thursday, April 26, 2001 I wrote:
>The bug is related to the -MD option, which normally produces a file with
extension .d holding
>the dependencies for the compiled file. This option, as stated in the gcc
sources, is deprecated and
>will be removed at a later time; however, there's always some software that
depends on it :-(
>
>When you specify -MD compiling, say, filename.c into filename, /usr/bin/gcc
should translate -MD (by
>means of the specs file) to
> -M -MF filename.d -MQ filename
>and pass these three options to cc1. What happens instead is that gcc emits
> -M -MF filename.d -MQ .filename filename
>and cc1 gets confused.
>
>The bug should be in gcc.c, maybe when parsing the "%o*" string from specs.
I found no workaround
>until now.
The bug is indeed in gcc.c, around line 5250. There is code which should
replace the extension in the output argument; however, when the output
argument
has no extension and no path separator, the counter goes to -1 and a dot is
written
outside the array bounds, with unpredictable results. It seems that only the
expansion of
-MD uses this code.
Here is a patch, which avoids writing to arg[-1]. It is against
gcc-20010430,
but works also in the gcc-2.96 Mandrake uses:
--- gcc.c Tue May 1 18:37:58 2001
+++ gcc.c Wed May 2 16:19:36 2001
@@ -5271,7 +5271,7 @@
do_spec_1 (" ", 0, NULL_PTR);
if (suffix_subst)
{
- unsigned length = strlen (arg);
+ int length = strlen (arg);
while (length-- && !IS_DIR_SEPARATOR (arg[length]))
if (arg[length] == '.')
@@ -5280,11 +5280,11 @@
break;
}
do_spec_1 (arg, 1, NULL_PTR);
- if (!arg[length])
+ if (length>=0 && !arg[length])
{
((char *)arg)[length] = '.';
- do_spec_1 (suffix_subst, 1, NULL_PTR);
}
+ do_spec_1 (suffix_subst, 1, NULL_PTR);
}
else
do_spec_1 (arg, 1, NULL_PTR);
Alberto
