> > Personally, I've always wanted to see a compiler which would do smart
> > things with the eight condition registers that each PowerPC has at its
> > disposal. Right now everybody just uses cr0 for everything. (I think
> > I've seen some rare instances where gcc will use cr7 as well.)
> > Originally the PowerPC was going to minimize on branching by allowing
> > you to do things like:
> > cmplwi cr1, something
> > cmplwi cr2, something
> > cmplwi cr3, something
> > cmplwi cr4, something
> > cror/crand something
> > bc something
> >
> > That way you can stack a lot of comparisons together by or'ing and
> > and'ing the condition bits. But no compiler actually generates it.
XLC and XLF do. It's not a sure win to make this optimization over a long
stretch of conditionals though. You end up making all paths equal length/time
if you use CR logic as above. By doing individual compare/branches, you have
an opportunity to speed up the common case. (You are putting your most likely
condition first, right? :)
Here's a test case where we're checking 3 conditionals. XLC groups the 1st
two together, but it leaves the 3rd condition on its own:
int foo(int a, int b, int c) {
if (a<0 || b<0 || c<0) return 1;
return 0;
}
00000000 cmpwi r4,0x0
00000004 cmpwi cr1,r3,0x0
00000008 cror 2,4,0
0000000c beq 0x20
00000010 li r3,0x0
00000014 cmpwi r5,0x0
00000018 blt 0x20
0000001c blr
00000020 li r3,0x1
00000024 blr
Note that you can 'optimize' this function by converting logical ops into
arithmetic ops. (I didn't test the speed of this code, so I'm not sure that
this is actually faster.)
int foo(int a, int b, int c) {
unsigned int boolSum = (a<0) + (b<0) + (c<0) ;
if (boolSum>0) return 1;
else return 0;
}
00000040 srwi r0,r3,31
00000044 srwi r3,r4,31
00000048 srwi r2,r5,31
0000004c add r2,r3,r2
00000050 add r0,r0,r2
00000054 addic r2,r0,0xffff
00000058 subfe r3,r2,r0
0000005c blr
--Sanjay
_______________________________
Do you Yahoo!?
Declare Yourself - Register online to vote today!
http://vote.yahoo.com
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/perfoptimization-dev/email@hidden
This email sent to email@hidden
