Mike Pall
e7f8cc964e
Split off fast function recording to lj_ffrecord.c.
2010-12-05 17:12:34 +01:00
Mike Pall
18fbf86c9e
Improve unrolling heuristics for non-looping inner loops.
2010-11-25 11:42:17 +01:00
Mike Pall
57cd5026eb
Disable Lua 5.2 features by default. See -DLUAJIT_ENABLE_LUA52COMPAT.
2010-11-19 17:00:11 +01:00
Mike Pall
ba602c9578
Add support for __pairs and __ipairs metamethods (from Lua 5.2).
2010-11-18 00:23:24 +01:00
Mike Pall
5391a5f5b1
Fix recording of y = select(n, ...) for non-int indexes.
2010-10-11 21:52:22 +02:00
Mike Pall
b3cf2c70f4
Decouple SLOAD type and optional conversion.
2010-10-11 21:13:37 +02:00
Mike Pall
ac76493df1
Specialize bytecode for pairs()/next() iterator. Speedup: 3.5x.
...
Parser predict pairs/next and emits specialized bytecode.
Bytecode is descpecialized at runtime if the prediction was wrong.
Store slot index in hidden control var to avoid key lookups.
2010-09-30 21:55:32 +02:00
Mike Pall
8dc76ee327
Add IR_VLOAD for vararg loads.
...
Also fixes the broken AA improvement in the last commit.
2010-09-14 19:58:27 +02:00
Mike Pall
b69c02eb68
Avoid unnecessary vararg loads.
2010-09-14 16:48:25 +02:00
Mike Pall
c88169dc46
Record y = select(x, ...) idiom.
2010-09-13 01:23:19 +02:00
Mike Pall
4ba0eb5f80
Record select().
2010-09-13 01:21:44 +02:00
Mike Pall
847b9cf253
Record vararg expressions with varargs defined off-trace.
...
Add SLOAD variant to access the frame type/size.
2010-09-13 01:17:38 +02:00
Mike Pall
e32f7d96c1
Fix tailcalls from vararg functions.
2010-09-13 00:55:05 +02:00
Mike Pall
52f75e9a5c
Fix off-by-one errors in maxslot calculation of trace recorder.
2010-09-12 03:14:17 +02:00
Mike Pall
4da03965c0
Record vararg expressions with known fixed number of results.
2010-09-12 02:42:02 +02:00
Mike Pall
c2c08ba9b3
Record calls to vararg functions.
...
This loop is now roughly 1000x faster than the Lua interpreter:
local function f(a,b,...) end; for i=1,2e8 do f(1,2,i) end
Yet another silly microbenchmark -- I know.
2010-09-12 01:44:13 +02:00
Mike Pall
96957a4551
Turn some lua_State fields into 32 bit pointers.
...
lua_State now fits into one cache line on x64.
2010-09-09 12:28:17 +02:00
Mike Pall
844e40d700
PPC: Clean up masked shift/rotate target settings.
2010-08-29 12:52:20 +02:00
Mike Pall
d05873ee0a
Abstract out pointer hash to hashrot(). Tune hash constants.
2010-07-21 22:06:38 +02:00
Mike Pall
5ff994fa37
Add weak guards. Emit TNEW/TDUP with a guard bit.
2010-05-08 18:59:59 +02:00
Mike Pall
41379126a2
Treat the tag of a TValue as unsigned everywhere.
2010-04-25 23:21:15 +02:00
Mike Pall
ef0904e80b
Make metamethod names proper GC roots.
2010-04-25 18:35:47 +02:00
Mike Pall
2e24770ed3
Simplify management of current trace. Drop lazy save.
2010-04-25 13:53:33 +02:00
Mike Pall
721b73fecb
Turn traces into true GC objects (GCtrace).
2010-04-25 03:32:29 +02:00
Mike Pall
ab45481199
No longer let the GC replace dead keys with the LJ_TDEADKEY tag.
...
Important: this changes the semantics of the write barrier!
Carefully read the big comment block in lj_obj.h
This helps HREFK key slot specialization and allows safely hoisting
HREF/HREFK across GC steps, too (fix for a barely reproducible bug).
Dead keys are only removed during a table resize (as before).
2010-04-21 01:45:58 +02:00
Mike Pall
932cda0fe3
Replace on-trace GC frame syncing with interpreter exit.
...
Need to sync GC objects to stack only during atomic GC phase.
Need to setup a proper frame structure only for calling finalizers.
Force an exit to the interpreter and let it handle the uncommon cases.
Finally solves the "NYI: gcstep sync with frames" issue.
2010-04-19 00:43:35 +02:00
Mike Pall
e4bca95451
Generate EQ(HREF, niltv) for load path, too (better CSE).
2010-03-28 06:49:09 +02:00
Mike Pall
69df01efb2
Fix precondition check for NEWREF.
...
A check for nil value is not enough. Must check for ptr == niltv.
2010-03-28 03:34:18 +02:00
Mike Pall
51c14bf1c8
Avoid snapshots for returns to known callers.
2010-03-21 16:01:44 +01:00
Mike Pall
c4727220e8
Add array bounds check elimination (-Oabc, on by default).
2010-03-15 17:02:53 +01:00
Mike Pall
24402ede04
Reorganize scalar evolution analysis.
2010-03-15 16:23:02 +01:00
Mike Pall
c56d791316
Fix handling of bad argument types in recorder.
2010-03-09 19:03:43 +01:00
Mike Pall
f49649d005
Fix recording of getmetatable() for non-tables.
2010-03-09 17:49:13 +01:00
Mike Pall
6e6034e809
Avoid tracing the nil return case of tonumber().
2010-03-09 03:47:49 +01:00
Mike Pall
e38fbb3d0a
Fix assertion in rec_check_slots.
2010-03-07 14:32:50 +01:00
Mike Pall
5fdb6e2e20
Fix 64 bit conversion warning.
2010-03-03 04:25:48 +01:00
Mike Pall
e7b737aa12
Implement down-recursion.
2010-03-01 06:45:30 +01:00
Mike Pall
4e73488985
Fix 64 bit conversion warnings.
2010-02-24 20:02:20 +01:00
Mike Pall
e46f4c8a11
Fix 64 bit portability problem in rec_ret().
2010-02-24 05:29:46 +01:00
Mike Pall
8ae2f9feaa
Randomize penalties for aborts and add blacklisting.
2010-02-23 19:41:32 +01:00
Mike Pall
659ea9de7b
Ensure function and all args have a reference for call recording.
...
In practice this is only needed after a return to a lower frame.
2010-02-22 17:37:26 +01:00
Mike Pall
3336434aa0
Fix TRef for (dummy) 2nd arg of __len metamethod.
2010-02-22 17:33:39 +01:00
Mike Pall
86494c783d
Back out history buffer for tailcall counts.
...
Use an aggregate counter independent of frame depth.
2010-02-22 16:57:59 +01:00
Mike Pall
19af483166
Add region selection for up-recursion and tail-recursion.
2010-02-22 14:35:47 +01:00
Mike Pall
bbe7d818d9
Allow linking to already compiled functions.
2010-02-18 19:37:30 +01:00
Mike Pall
b11eeab906
Use a limited history buffer for tailcall counts while recording.
2010-02-18 19:32:13 +01:00
Mike Pall
59f54be9d3
Update trace recorder infrastructure for hot calls.
2010-02-18 06:07:31 +01:00
Mike Pall
2b22b9e710
Drop obsolete shadow frame link stack.
2010-02-18 03:19:46 +01:00
Mike Pall
265a24b978
Add more assertions to compare the recorder state and the VM state.
2010-02-18 02:49:08 +01:00
Mike Pall
00ad80c9e4
Clear frame gaps in recorder to avoid resurrecting previous refs.
2010-02-18 02:45:03 +01:00
