On Wed, Jun 25, 2008 at 11:17:45AM -0500, Scott Wood wrote:
> Gabriel Paubert wrote:
> >On Wed, Jun 25, 2008 at 10:34:32AM -0500, Scott Wood wrote:
> >>Kumar Gala wrote:
> >>>>+/* Macros to workout the correct index for the FPR in the thread
> >>>>struct */
> >>>>+#define FPRNUMBER(i) (((i) - PT_FPR0) >> 1)
> >>>>+#define FPRHALF(i) (((i) - PT_FPR0) % 2)
> >>>Have you looked at what the compiler spits out here to make sure we
> >>>aren't getting a divide? Seems like we could use '& 0x1'.
> >>GCC's not *that* dumb. However, you may get some unnecessary
> >>sign-twiddling if "i" is signed.
> >
> >Not for modulo 2, it's only an even/odd choice and GCC
> >implements that efficiently IIRC. For other powers of 2,
> >making the left hand side unsigned helps the compiler.
>
> From this:
>
> int foo(int x)
> {
> return x % 2;
> }
>
> I get this with -O3:
>
> foo:
> mr 0,3
> srawi 3,3,1
> addze 3,3
> slwi 3,3,1
> subf 3,3,0
> blr
> .size foo, .-foo
> .ident "GCC: (GNU) 4.1.2"
>
Indeed. Signed modulo results can be negative...
There are probably better ways to implement this case
on PPC, for example:
rlwinm tmp,input,4,27,28 ; make shift amount from LSB and MSB
lis result,0xff01
srw result,result,tmp
; result is now 0x00 for even, 0x01 for odd positive,
; and 0xff for odd negative
extsb result,result
No carry, shorter dependency length (although srw may be slow
on Cell it seems, but addze may be worse).
> Changing it to "x & 1", or to unsigned, gives this:
>
> foo:
> rlwinm 3,3,0,31,31
> blr
> .size foo, .-foo
> .ident "GCC: (GNU) 4.1.2"
>
> Maybe newer GCCs are better?
Nope, but unsigned is often better for the right shift.
Gabriel
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@ozlabs.org
https://ozlabs.org/mailman/listinfo/linuxppc-dev
