1 ----------------------------------------------------------------------------- 2 Notes on performance 3 ----------------------------------------------------------------------------- 4 The intent of this file is to record progress in improving performance. 5 6 ----------------------------------------------------------------------------- 7 Just before 3.1.0: 8 - Julian made LibVEX_Alloc() inlinable. Saved a couple of percent. 9 - Julian started building Vex at -O2. Saved up to 8% or so(?) in some 10 cases. 11 12 Post 3.1.0: 13 - Julian made the tree builder linear. Saved 2--13% on a range of programs. 14 - Nick improved vg_SP_update_pass() to identify more small constant 15 increments/decrements of SP, so the fast cases can be used more often. 16 Saved 1--3% on a few programs. 17 - r5345,r5346,r5352: Julian improved the dispatcher so that x86 and 18 AMD64 use jumps instead of call/return for calling translations. 19 Also, on x86, amd64, ppc32 and ppc64, --profile-flags style profiling was 20 removed from the despatch loop unless --profile-flags is being used. 21 Improved Nulgrind performance typically by 10--20%, and Memcheck 22 performance typically by 2--20%. 23 - Julian changed findSb to slowly move superblocks to the front of the list 24 as they were accessed. This sped up perf/heap by 25--50%, and some big 25 programs (eg. ktuberling) programs by a couple of percent. 26 - Nick reduced the iteration count of the loop in swizzle() from 20 to 5, 27 which gave almost identical results while saving 2% in perf/tinycc and 10% 28 in perf/heap on a 3GHz Prescott P4. 29 - Nick changed ExeContext gathering to not record/save extra zeroes at the 30 end. Saved 7% on perf/heap with --num-callers=50, and about 1% on 31 perf/tinycc. 32 - Julian vectorised copy_address_range_perms for common cases, which 33 gives about 40% speedup on artificial programs which just do 34 realloc() and nothing else, and about a 3-4% speedup on starting 35 kpresenter-1.5.0 and loading a 16-slide presentation. 36 37 COMPVBITS branch: 38 - Nick converted to compress V bits, initial version saved 0--5% on most 39 cases, with a 30% improvement on one case (tsim_arch) which calls 40 set_address_range_perms() a lot. 41 - Nick rewrote set_address_range_perms(), which gained 0--3% typically, 42 and 22% on tsim_arch. 43 44
