LibretroAdmin
9140befd81
dynarec: room for the largest block, and a block that is already gone
...
A block is compiled if there are 16 KB left in the code cache. A block
is up to 511 operations, and with the MMU on a load or store can be over
a hundred bytes: upstream went to 32 KB after Tomb Raider: The Last
Revelation ran past 16 (f11a5981f). The largest block seen here is
10 KB, in Sega Rally 2, on x86-64 and on ARM64 alike. 64 KB now, of
16 MB: CODE_MARGIN, for the cache, the one for rewritten code and the
checks the recompilers make.
A build with NO_MMU had no cache for rewritten code at all
(TEMP_CODE_SIZE 0), and since such code is kept there with the MMU off
too would have been given a block to compile into it. No Makefile
defines NO_MMU; it has the megabyte the others have.
rdv_BlockCheckFail(), with the MMU on, used a block it had looked up
without asking whether there was one - there is none if what failed was
run once more after it was discarded. Upstream has the same check.
2026-10-10 14:38:17 +02:00
twinaphex
cb3bd71cc5
Backport https://github.com/flyinghead/flycast/commit/e99aac357567302e1d1daf4b59ec8e4139c879da#diff-7194c04007dd96b8c496118886760725f8eb7b8bda6916d63bf6e3ea967e44a7
2021-04-12 02:32:26 +02:00
twinaphex
352536f977
Get rid of core/rec.cpp to make it more in line with upstream
2021-03-01 12:08:12 +01:00
twinaphex
5e821261e2
(Dynarec) Get rid of function pointers
2021-02-28 22:35:44 +01:00
Flyinghead
05da8c9c2f
rec-cpp: Add missing opcode call classes and allow for immediate args
2019-12-12 19:37:22 +01:00
Flyinghead
32c2bae64f
dynarecs,reios: use new logging
2019-08-30 12:58:52 +02:00
Flyinghead
8f41faa035
rec-arm64: use SSA, generate main loop
...
Fix xtrct canonical implementation
2019-06-14 13:45:11 +02:00
Flyinghead
fc14e9f2f8
rec-cpp: crash when running rom
2019-05-31 12:50:55 +02:00
twinaphex
cb25abd9e8
Update rec_cpp
2018-08-11 13:46:09 +02:00
twinaphex
9e9a09fa42
Rebase code
2018-06-15 05:03:54 +02:00
twinaphex
f740bca1e9
Move audiostream file to core/hw/aica and get rid of oslib/ in
...
general
2015-12-28 16:17:03 +01:00
twinaphex
2b5ee8a0b5
(rec_cpp) Take out emitter/x86_emitter.h
2015-12-24 16:08:36 +01:00
twinaphex
206fc20f61
Uniquely name BlockCompiler
2015-12-24 14:48:05 +01:00
twinaphex
2f5bcba70d
Create interface so we can easily switch between dynarecs
...
in the future
2015-12-24 14:28:21 +01:00
twinaphex
d6ba6f90e9
(rec_cpp.cpp) Get rid of some of the auto type deduction
2015-12-24 12:27:09 +01:00
twinaphex
291263d3ad
Make a centralized rec.cpp from which all other dynarecs originate
...
from
2015-12-24 09:38:10 +01:00
twinaphex
8fa5df554d
Take out profiler
2015-12-18 07:47:52 +01:00
twinaphex
96335ca884
Cleanups
2015-10-18 06:54:55 +02:00
twinaphex
4e9b9238b1
Make the mainloops somewhat semi-comprehensible
2015-10-18 04:15:26 +02:00
Stefanos Kornilios Mitsis Poiitidis
2beb23978b
rec-cpp: Fix non-emscripten build
2015-08-03 11:12:24 +02:00
Stefanos Kornilios Mitsis Poiitidis
d8ccc9d2e7
emscripten: More rec-cpp progress
...
It almost builds now, except emscripten randomly gives up and crashes. No one likes my code these days...
2015-07-29 06:45:02 +02:00
~skmp
f1a6f04dce
rec-cpp: verify() that we have an executer w/ shop_count
2015-07-29 02:29:07 +02:00
Stefanos Kornilios Mitsis Poiitidis
ec5bb09917
rec-cpp: Add sig aCaCcCdC for mul_u64, mul_s64
2015-07-29 02:28:52 +02:00
Stefanos Kornilios Mitsis Poiitidis
1e0224dd1b
rec-cpp: Add sig aCaCcCdC for rocl, rocr
2015-07-29 02:28:29 +02:00
Stefanos Kornilios Mitsis Poiitidis
4b97f468b5
rec-cpp: Implement block ends and a few more decoders
...
- Full IL mode now works!
- Dispatch up to 270 mips on 2.2ghz i7, on a complex 3d scene
- Will run many games fullspeed on modern pc, thanks to idleskip
2015-07-25 20:38:48 +02:00
Stefanos Kornilios Mitsis Poiitidis
b1f7015a98
rec/cpp: Mostly fully direct specialized dispatchers
...
- Nest shil implementations to shilop_##name::form::impl for easier template matching
- Add more direct handlers
- Add prints if a direct handler isn't found
- Remove multiple indirect handlers, leave them for fallback
2015-07-25 14:36:04 +02:00
Stefanos Kornilios Mitsis Poiitidis
51a6ed2146
rec/all: optional mips counter on recs
...
Also de-unroll rec-cpp for debug builds
2015-07-25 14:31:46 +02:00
Stefanos Kornilios Mitsis Poiitidis
00dbd6222f
rec-cpp: Massage for gcc compatibility
2015-07-25 14:25:51 +02:00
Stefanos Kornilios Mitsis Poiitidis
1710f76ab1
unroll the execution loop
2015-07-25 14:24:11 +02:00
Stefanos Kornilios Mitsis Poiitidis
7303a046be
rec-cpp: Direct calls to shil_opcl_* handlers for many opcodes
2015-07-25 14:23:59 +02:00
Stefanos Kornilios Mitsis Poiitidis
e4138f9a60
rec-cpp: Wastefully generate more dispatchers for better BTB cache locality
...
The real solution is to have inlining between the (now static) dispatchers and the impls.
It's gonna be hard to convince the compiler on that.
2015-07-25 14:23:07 +02:00
Stefanos Kornilios Mitsis Poiitidis
fc01428c30
rec-cpp: Force resets when running out of blocks
2015-07-25 14:22:28 +02:00
Stefanos Kornilios Mitsis Poiitidis
bfef969626
rec-cpp: Implement stuff up to the corejit layer
2015-07-25 14:21:53 +02:00
Stefanos Kornilios Mitsis Poiitidis
cd80850acb
First idea for a fully portable "jit"
2015-07-25 14:20:54 +02:00
