Commit graph

775 commits

Author SHA1 Message Date
Henrik Rydgård
0f080aeaaa Merge pull request #492 from unknownbrackets/jit-minor
ALU jit optimizations
2013-01-25 01:01:34 -08:00
Unknown W. Brackets
a7c6f46829 Optimize and/or 0 to just a mov in x86 jit. 2013-01-25 00:25:40 -08:00
Unknown W. Brackets
ab9bea068c Jit reg+reg compile time, and avoid flushing EDX. 2013-01-25 00:16:55 -08:00
Unknown W. Brackets
ce5f393fb8 Hit immediates in the ALU better and more simply. 2013-01-25 00:16:55 -08:00
Unknown W. Brackets
d1909a1581 Add a quick disable define for nice delay slots. 2013-01-24 19:11:03 -08:00
Unknown W. Brackets
75cbe18afc Simplify nice delay slot detect, and yes for noop.
NOOP seems very common so this should already benefit speed a bit.
2013-01-24 08:29:32 -08:00
Unknown W. Brackets
2eba209f64 Move around the jit nice delay slot logic.
Nice delay slots don't not save flags, they run before the CMP.
2013-01-24 07:31:51 -08:00
Unknown W. Brackets
3444fc8981 Avoid some memory writes on jr.
Should improve tight mips function loops a bit.
2013-01-24 01:23:50 -08:00
Unknown W. Brackets
c1757ee166 Check downcount in jit after a syscall. 2013-01-23 22:25:35 -08:00
Unknown W. Brackets
0e33923844 Belt and suspenders check for branch ops. 2013-01-22 08:11:37 -08:00
Unknown W. Brackets
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
Unknown W. Brackets
a9d0390426 Adjust downcount before syscalls, not after.
This makes jit slightly slower for syscalls, but it's minor and makes
sure jit and interpreter timing are determistically the same.
2013-01-21 22:57:53 -08:00
Unknown W. Brackets
566b7a0910 A branch was missing inDelaySlot, refactor it. 2013-01-21 22:45:07 -08:00
Unknown W. Brackets
c897e6446a Don't over decr downcount when hitting a jit bp. 2013-01-21 19:41:12 -08:00
Unknown W. Brackets
8438371941 Read memory in the jit dispatcher same as lw.
Just for consistency.  One less op, maybe faster, probably same.
2013-01-21 17:51:14 -08:00
Unknown W. Brackets
1485b0865c Improve the speed of branch debugging a bit. 2013-01-20 19:48:55 -08:00
Unknown W. Brackets
dd69694302 Add some optional logging to debug jit branching. 2013-01-20 19:48:55 -08:00
Unknown W. Brackets
776eb8ab2e Simplify CompileDelaySlot(). 2013-01-20 19:48:54 -08:00
Unknown W. Brackets
df06bb5624 Add some checks to make sure ZERO is never set. 2013-01-20 19:48:53 -08:00
Unknown W. Brackets
a43078ab68 Same optimization for FPU load / store. 2013-01-20 13:16:41 -08:00
Unknown W. Brackets
f5963df0dc Optimize write to a single x64 op too. 2013-01-20 13:06:19 -08:00
Unknown W. Brackets
e8dc99328a Avoid using EAX as a temporary where possible.
All the regs should be indirect addressing compatible.  So if it's
in a reg, let's use that instead of EAX.
2013-01-20 12:57:14 -08:00
Unknown W. Brackets
eaa24ee047 Use EDX as a temporary for sb, and jit it. 2013-01-20 12:25:08 -08:00
Unknown W. Brackets
30f6a4ba87 Fix stupid stupid typo breaking slowmem jit. 2013-01-20 09:39:13 -08:00
Unknown W. Brackets
da22eb8adf Make swc1 and lwc1 fast even without fastmem. 2013-01-20 02:07:00 -08:00
Henrik Rydgård
2cb830510c Merge pull request #444 from unknownbrackets/jit-minor
Don't muck with currentMIPS->r directly in the slowmem jit
2013-01-20 01:58:48 -08:00
Unknown W. Brackets
53600161ba Don't write anything to a bad static pointer.
Not sure what came over me...
2013-01-20 00:19:18 -08:00
Unknown W. Brackets
385417effe Log jit misses at runtime instead of compile time.
Really, it could be very different after all... this shouldn't be
all that slow, I guess.
2013-01-19 20:11:17 -08:00
Unknown W. Brackets
75a3872923 Log missed jit ops for poor man profiling. 2013-01-19 19:58:25 -08:00
Unknown W. Brackets
e78223d2c0 Since flipping the op is easy, also do lb/lh. 2013-01-19 16:25:57 -08:00
Unknown W. Brackets
5e8f1917ee Fix 64-bit memory dereferencing. 2013-01-19 16:25:57 -08:00
Unknown W. Brackets
90e6f0b7df Optimize static memory read/writes in jit as well.
Like the arm jit does.
2013-01-19 16:25:56 -08:00
Unknown W. Brackets
c64966c16e Oops, lost the CONDITIONAL_DISABLE. 2013-01-19 16:25:56 -08:00
Unknown W. Brackets
37fb64ac83 Fast path scratchpad too, shouldn't be expensive. 2013-01-19 11:11:45 -08:00
Unknown W. Brackets
72e547420d Refactor jit slowmem, add lbu to jit since easy. 2013-01-19 11:11:45 -08:00
Unknown W. Brackets
5305017fc3 Properly save registers before the slowmem call. 2013-01-19 11:11:44 -08:00
Unknown W. Brackets
f1295f6262 Don't muck with currentMIPS->r in the slowmem jit. 2013-01-19 11:11:44 -08:00
Henrik Rydgard
229d4e9f32 Buildfix 2013-01-19 13:10:52 +01:00
Henrik Rydgård
c20cef2399 Merge pull request #440 from unknownbrackets/jit-minor
Micro optimizations to x86 jit mem when fastmem is off
2013-01-19 02:04:34 -08:00
Unknown W. Brackets
2ad77aa9c8 Gotta flush before the call, too. 2013-01-19 01:53:11 -08:00
Unknown W. Brackets
09422d5adb Avoid a func if possible when fastmem is off. 2013-01-19 01:02:47 -08:00
Unknown W. Brackets
bc75b68c36 Also jit sh and lhu, which are pretty common ops. 2013-01-18 23:10:51 -08:00
Unknown W. Brackets
d5ae85201c Optimize sw/lw even under safe memory.
They're very common instructions, so shaving cycles helps.
2013-01-18 23:10:50 -08:00
Unknown W. Brackets
11c5cdfdb0 Refactor out all the CheckJitBreakpoint()s.
Ah, much cleaner.
2013-01-18 21:33:23 -08:00
Unknown W. Brackets
5080285e54 Add breakpoints for delay slots.
Wanted to do this in CompileAt(), darn not nice delay slots.
2013-01-18 21:12:58 -08:00
Unknown W. Brackets
40ae3dfe45 Correctly break at branch points in x86 jit. 2013-01-18 21:12:53 -08:00
Unknown W. Brackets
beac991a9e Clear jit cache when changing breakpoints.
For now, only when paused.  I don't think clearing the cache while
running is an awesome idea.
2013-01-18 20:12:53 -08:00
Unknown W. Brackets
a9293c8923 Add breakpoints to x86 jit for easier debugging.
They should be really fast so leaving them on in release.
2013-01-18 20:10:37 -08:00
Henrik Rydgard
674911ddba Move downcount into MIPSState for efficiency, enable block linking.
On ARM JIT we can now reach it through the cpu context reg.
2013-01-12 00:44:18 +01:00
Henrik Rydgard
ea3055322c Oops 2013-01-09 00:45:54 +01:00