cpu/drcbex64.cpp: Directly dispatch memory accesses to specific handlers #13325
Conversation
|
@MooglyGuy if you could see if this yields improvements on your AMD CPU, I’d appreciate it. |
|
I’ve updated this pull request to also inline reads smaller than native width. However, in my testing this actually hurts performance for If/when people try this out, please test and benchmark the two commits separately. Only the first commit requires an emu.h rebuild – If you can check out the first commit, build, copy the binary, check out the second commit, and build again, the second build will only need to recompile src/devices/cpu/drcbex64.cpp. |
|
I'll have a look this coming weekend, just remind me. |
emu/emumem_aspace.cpp: Got rid of static accessors structure, added a helper for obtaining info for dispatching specific accessors. cpu/drcbex64.cpp: Directly dispatch specific memory accessors for native width reads and writes no larger than native width.
|
I sorted out the performance issues with my initial implementation of reads smaller than the native width. I’m pretty confident this is a modest but useful improvement now. |
At the urging of @galibert, this makes the x86-64 recompiler back-end dispatch memory accesses to specific handlers directly, avoiding one level of out-of-line function calls. From my testing on an Intel 8th generation i7, the performance improvement is measurable, but less than 5% at best.
I’d like people to test this with other x86-64 CPUs (e.g. AMD or different generations of Intel CPUs) to see what kind of difference it makes (if any).
If it does seem to produce a useful improvement, I can follow up by adding code to directly dispatch reads smaller than native width, and add scaffolding for carrying over the same optimisations to AArch64. If it doesn’t seem to be a worthwhile improvement, I’ll just move across the removal of the static address space accessors and ditch the recompiler changes.