On Monday, July 06, 2015 02:45:59 AM Francisco Jerez wrote: > Matt Turner <matts...@gmail.com> writes: > > On Fri, Jul 3, 2015 at 3:46 AM, Francisco Jerez <curroje...@riseup.net> > > wrote: [snip] > Yeah. I did in fact try to implement uaddCarry last Friday without > using the accumulator by doing something like: > > | CMP.o tmp, src0, -src1 > | MOV dst, -tmp > > ...what of course didn't work because of the extra argument precision > post-source modifiers and also because the .o condmod doesn't work at > all on CMP, but... > > > Ideally, we'd recognize merge the addition and carry operations into a > > single ADDC instruction, but it's pretty unimportant. It's all pretty > > academic -- I've never seen an application use either operation (or > > [iu]mulExtended either). > > ...if we did the following instead: > > | ADD tmp, src0, src1 > | CMP.l tmp, tmp, src0 > | MOV dst, -tmp > > the ADD could be easily CSE'ed with the original ADD instruction (and > the source modifier of the last MOV can also be easily propagated into > some other instruction), so even though it seems like one instruction > more than what we emit now it might be a net win (aside from it working > on SIMD16). usubBorrow is even easier: > > | CMP.l tmp, src0, src1 > | MOV dst, -tmp > > I was planning to run it through shader-db tomorrow but if you say > you've never seen them used I guess I shouldn't get my hopes too high? :P
Yeah, there's nothing in shader-db that uses them, so I wouldn't bother. I'm definitely a huge fan of avoiding the accumulator - it's always been a total pain to deal with. If you guys come up with a solution that uses normal registers and avoids MACH, that sounds great to me.
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
