On 22/05/14 22:18, Richard Earnshaw wrote: > On 22/05/14 00:44, Kugan wrote: >> Compiling some applications with -mgeneral-regs-only produces better >> code (runs faster) compared to not using it. The difference here is that >> when -mgeneral-regs-only is not used, floating point register are also >> used in register allocation. Then IRA/LRA has to move them to core >> registers before performing operations. >> >> I experimented with TARGET_SPILL_CLASS (as in attached patch) to make >> floating point register class as just spill class for integer pseudos. >> Though this benefits the application which had this issue. Overall >> performance with speck2k is neutral (some of the benchmarks benefits a >> lot but others regress). I am looking to see if I can make it perform >> better overall. Any suggestions welcome. >> >> Attached experimental patch passes regression but 168.wupwise and >> 187.facerec miscompares now. I am looking at fixing this as well. >> > > While I'll be the first to admit that the generic costs are currently > little more than estimates, I'm worried about changing them based on the > results of one benchmark run on a single (unspecified) implementation. > The generic costs are intended to be a "best blend" of costs across a > large range of implementations. Generically tuned code will not > necessarily be optimial on every processor -- what's more important is > that it's not severely sub-optimal on any processor.
Thanks Richard for the comments. My primary intention here is to use TARGET_SPILL_CLASS to make FP_REGS as spill registers. Do you think AArch64 can benefit from TARGET_SPILL_CLASS hook. I agree that just increasing GP2FP and FP2GP for all the modes as I am doing is not the right think to do. Thanks, Kugan
