Il 05/09/2013 17:17, Richard Henderson ha scritto: > On 09/05/2013 01:22 AM, Paolo Bonzini wrote: >> These use a 32-bit load-of-immediate to save a mflr+addi+mtlr sequence. >> Tested with a Windows 98 guest (pretty much the most recent thing I >> could run on my PPC machine) and kvm-unit-tests's sieve.flat. The >> speed up for sieve.flat is as high as 10% for qemu-system-i386, 25% >> (no kidding) for qemu-system-x86_64 on my PowerBook G4. > > See also the series beginning at > > http://lists.nongnu.org/archive/html/qemu-devel/2013-09/msg00025.html > > The major difference is that I use a conditional call out of the fast > path, which lets me later just use one mflr to pass the parameter. I > also, perhaps foolishly, got rid of the trampolines. E.g. > > 0xf57a1838: rlwinm r3,r15,24,20,27 > 0xf57a183c: rlwinm r0,r15,0,30,19 > 0xf57a1840: add r3,r3,r27 > 0xf57a1844: lwz r4,6436(r3) > 0xf57a1848: cmpw cr7,r0,r4 > 0xf57a184c: lwz r3,6444(r3) > 0xf57a1850: bnel- cr7,0xf57a1910 > 0xf57a1854: stwx r16,r3,r15 > ... > 0xf57a1910: mr r3,r27 > 0xf57a1914: mr r4,r15 > 0xf57a1918: mr r5,r16 > 0xf57a191c: li r6,1 > 0xf57a1920: mflr r7 > 0xf57a1924: lis r0,4120 > 0xf57a1928: ori r0,r0,45040 > 0xf57a192c: mtctr r0 > 0xf57a1930: bctrl > 0xf57a1934: b 0xf57a1858 > > I don't see anything technically wrong with your patch. But I'd be > interested to compare vs mine.
Sure, I'll give it a try tomorrow or in the weekend. The G4 in my computer must simply hate the mflr/add/mtlr sequence in the trampoline; there's no other explanation for such a huge performance improvement. So even though I suspect that there won't be much difference between our patches it's good to check what's better in case your sequences are triggering something as bad. The bnel/mflr is a nice trick to save one instruction, though! Regarding removal of the trampolines, the extra icache cost should be a wash now that they are half the size, but I'd still prefer it to be a separate patch. Paolo
