On Tue, 27 Feb 2001, Fabrice Bellard wrote:
> I found a serious bug in the assembler macros in asm-mips/uaccess.h. They
> all do something like that:
>
> __asm__ __volatile__( \
> "move\t$4, %1\n\t" \
> "move\t$5, %2\n\t" \
> "move\t$6, %3\n\t" \
> ".set\tnoreorder\n\t" \
> __MODULE_JAL(__copy_user) \
> ...
>
> The problem is that you cannot assume that gcc will not put %1, %2 or %3
> in registers different from $4, $5 or $6. For example, if %2 is put in $4,
> the code is incorrect. (With gcc-2.95.2 I got a bug in
> generic_file_write!).
Hmm, haven't looked through gcc sources, but docs state: "It is an error
for a clobber description to overlap an input or output operand (for
example, an operand describing a register class with one member, mentioned
in the clobber list)." I guess it implies clobbers are not used for input
or output. It's reasonable anyway and if gcc acts otherwise, you might
just have caught a bug in gcc.
> A possible fix would be to use asm registers:
>
> #define copy_from_user(to,from,n) ({ \
> register void *__cu_to asm("$4"); \
> register const void *__cu_from asm("$5"); \
> register long __cu_len asm("$6"); \
> \
> __cu_to = (to); \
> __cu_from = (from); \
> __cu_len = (n); \
> if (access_ok(VERIFY_READ, __cu_from, __cu_len)) \
> __asm__ __volatile__( \
> ".set\tnoreorder\n\t" \
> __MODULE_JAL(__copy_user) \
> ...
>
> But I am not sure that it is always correct. Any idea ?
This is fine and saves us three instructions. Go on, make a patch (I'd
suggest using "__asm__" for consistency, though)!
--
+ Maciej W. Rozycki, Technical University of Gdansk, Poland +
+--------------------------------------------------------------+
+ e-mail: macro@ds2.pg.gda.pl, PGP key available +
