On Thu, 22 Apr 2004, Bradley D. LaRonde wrote:
> gcc 3.4 complians about:
>
> include/asm/unaligned.h:66: error: inconsistent operand constraints in an
> `asm'
>
> from linux CVS 2.4 branch. That's:
>
> /*
> * Store doubleword ununaligned.
> */
> static inline void __stq_u(unsigned long __val, unsigned long long * __addr)
> {
> __asm__("usw\t%1, %0\n\t"
> "usw\t%D1, 4+%0"
> : "=m" (*__addr)
> : "r" (__val));
> }
As usually recent gcc is invaluable in finding dubious constructs. ;-)
> I finally decided to punt and write:
>
> static inline void __stq_u(unsigned long long __val, unsigned long long *
> __addr)
> {
> *__addr = __val;
> }
>
> Is this OK? Is there a better solution?
No. Yes.
Reviving old tricks about unaligned data I've come with the following
implementation:
void __stq_u(unsigned long long __val, unsigned long long *__addr)
{
typedef struct {
unsigned long long v __attribute__((packed));
} ull_u_t;
ull_u_t *a = (ull_u_t *)__addr;
a->v = __val;
}
which yields a nice and desirable code:
unaligned.o: file format elf32-tradlittlemips
Disassembly of section .text:
00000000 <__stq_u>:
0: a8c40003 swl a0,3(a2)
4: b8c40000 swr a0,0(a2)
8: a8c50007 swl a1,7(a2)
c: 03e00008 jr ra
10: b8c50004 swr a1,4(a2)
...
I'm pretty sure it works fine with gcc 2.95.x as well -- for Alpha it used
to, even with such antiques as egcs 1.1.2.
Ralf, I can see 2.6 already does the right thing -- I suppose you won't
mind me backporting (copying?) it?
--
+ Maciej W. Rozycki, Technical University of Gdansk, Poland +
+--------------------------------------------------------------+
+ e-mail: macro@xxxxxxxxxxxxx, PGP key available +
