1cb7965cb1
Jit feature preparation: Introduce "proxy blocks".
...
When these are invalidated, the block they point to gets invalidated too.
Will be useful to implement various types of block merging and function inlining
without affecting correctness of cache clears etc.
Also, with this commit we can now fully inline replaced functions. fabsf() boils
down to 1-2 instructions and the block continues, for example.
2013-12-19 00:39:49 +01:00
763eff181d
Fix handling of jalr when delay slot changes rd.
2013-11-14 23:39:13 -08:00
ef8631c57f
Cache VFPU_CTRL_CC in a register
2013-11-12 17:58:29 +01:00
17074f5a7f
Cache fpcond in a register to avoid store/load between compare and branch
2013-11-12 10:33:38 +01:00
bb960480c8
x86/armjit: Stop compiling on a jump to invalid.
2013-11-10 21:59:50 -08:00
359110f010
x86/armjit: Add jump following (off by default.)
...
Inlines function calls up to a certain extent. Allows us to get
immediates all the way to a syscall, for example, usually.
Not sure if faster.
2013-11-10 21:59:49 -08:00
aacb31bc18
armjit: Copy over (disabled) immbranch optim.
...
This does a little loop unrolling. Costs a bit more cache space, but
avoids flushing regs for longer.
Not enabled.
2013-11-10 21:59:48 -08:00
7e46ee0b0f
armjit: Replace MOVI2R with using the regcache.
...
So that it can optimize the value with existing imms.
Not actually optimizing yet.
2013-11-10 15:50:45 -08:00
1cc68f50ca
armjit: Small optimization to syscall instr.
2013-11-10 14:38:10 -08:00
b30928036e
armjit: Avoid flushing an imm in beq/bne/etc.
...
We might be able to STMIA it instead.
2013-11-10 14:38:10 -08:00
5a95e267fb
Add an optimization to discard registers at the end of functions when possible.
...
Works in some games but crashes many so hiding it for now. Do not add UI.
2013-11-08 12:43:48 +01:00
91393093bc
Re-enable the "nice delay slot" optimization on ARM
2013-11-07 15:29:12 +01:00
1e158fa652
ARM vtx dec: Preserving our FP scratch register appears to improve
...
stability.
Also added some logging.
2013-11-06 11:47:26 +01:00
6483c0c2cd
Two minor armjit optimizations
2013-11-05 16:25:01 +01:00
732ae13ebb
Fast path CallSyscall where possible.
...
It seems we're spending a decent amount of time there, which isn't
entirely unexpected. We can eliminate some things easily.
2013-11-04 07:59:37 -08:00
9a707f6c00
Fix typo, ARM build.
2013-10-17 15:28:50 -07:00
2e8ef3027f
Write the retaddr to rd, not always ra, in jalr.
...
Thanks go entirely to @Kingcom for pointing this out.
Don't know of any games not using RA as the rd.
2013-10-17 07:39:33 -07:00
3a1b6fb269
Fix some sign comparison warnings.
2013-10-05 11:13:41 -07:00
97aa1a631e
Improve typesafety in the x86 regalloc.
2013-08-24 19:41:10 -07:00
109ad17ac6
Use a typesafe struct for opcodes.
...
Also, correctly read delayslots using Read_Instruction on ARM.
2013-08-24 15:36:24 -07:00
8327cd0f8e
Clean up some inconsistency in jit branches on arm.
2013-08-16 02:02:56 -07:00
2758634a0f
Avoid overhead calling __KernelIdle().
...
~1.8% improvement in Zettai Hero Project.
2013-08-15 01:35:17 -07:00
5387e5c717
Oops, typo.
2013-08-15 00:38:53 -07:00
df50e03146
Add a safety log for cases we don't handle right.
2013-08-14 23:14:25 -07:00
e639f8d15f
Handle branches in VFPU delay slots better.
...
Based on tests on a PSP, all branches are attempted. The behavior is
technically undefined.
It seems to take the delay slot's target if they differ and both pass.
This is the behavior the interpreter has, but it's more work in jit.
Since only a couple games seem to do this, and clearly expect this
behavior, this fixes all known cases of #1926 .
2013-08-14 22:56:02 -07:00
76a937f489
ARMJIT Experiment: Keep downcount in a register. Needs benchmarking.
2013-07-27 17:27:26 +02:00
c9c3bc83e4
Log more info about branches in delay slots.
2013-06-30 13:19:27 -07:00
a640ac4001
Bail out on a delay slot in VFPU on arm also.
...
Otherwise we loop infinitely.
Workaround for #1926 . Should fix #1679 .
2013-06-03 00:06:24 -07:00
5595146f56
Add reporting for jumps in delay slots.
2013-05-26 20:30:14 -07:00
9eace8a80e
Combine the two JitCache implementations (x86, ARM) into one.
2013-04-27 01:32:03 +02:00
ed76563973
Don't bother checking nice, just do it after.
2013-03-11 02:18:27 -07:00
b5fe67eb3d
If the out is RA, delay slot isn't nice for jal.
...
Fixes Phantasy Star Portable 2 in jit.
2013-03-11 02:14:38 -07:00
d525abbcad
Cleanup some indentation, more compares to 0.
2013-03-07 02:08:45 -08:00
028e85dc92
Cleanup some differences between the two jits.
2013-03-07 02:08:44 -08:00
268d16bd24
Use correct args for STR(..) throughout armjit.
2013-03-07 00:59:07 +10:00
81589b67e5
Save one instruction on jal, shorten block exits by one instruction (moved to dispatcher).
2013-02-15 23:37:59 +01:00
048cf35922
More ARMJit FPU work - some instructions and optimizations.
2013-02-14 00:02:09 +01:00
44b5adeaac
Properly jit the break instruction.
...
Otherwise, it just keeps on going past it.
We never want to hit this anyway, but it's good to know if we do.
2013-02-01 00:49:14 -08:00
d8f4e27926
Rename ARMABI_MOVI2R to MOVI2R
2013-01-31 23:41:05 +01:00
c97f63a9d9
Minor armjit opt
2013-01-30 20:01:42 +01:00
739b76a55a
Armjit: branch code cleanup #1
2013-01-30 01:05:36 +01:00
dbe597ce64
Buildfix Android, darn.
2013-01-22 08:23:27 -08:00
c324983340
Make the jit support bltzal and friends.
...
Fixes problems with jit in games. Android changes completely untested.
2013-01-22 08:04:01 -08:00
bc9c3db303
Armjit: Add option for fastmem. Small optimization.
2013-01-11 17:25:54 +01:00
dafe2c389c
More regalloc fixing and tweaks. Still not working the way I want it.
2013-01-09 13:38:44 +01:00
dafc9f62df
Regcache fixes, etc. thing still don't work when I turn on addiu :(
2013-01-09 11:20:48 +01:00
f5c94775b9
Cleanup
2013-01-09 00:12:02 +01:00
76481a300c
Icache must be invalidated. Jit now starts to run, but there's no cube in cube.elf!
2013-01-08 23:52:11 +01:00
8915677241
More progress but it weirds out...
2013-01-08 21:24:42 +01:00
8c06edc47b
It's getting close to the first totally unoptimized jit run.
2013-01-08 17:03:17 +01:00
Henrik Rydgard