Andreas Gal
00b12bbaca
Always trash TreeInfo when we recompile. This avoids the much dreaded JS_ASSERT(insInt32) errors and we can do this now because the demotion information is provided by the oracle and no longer stored in the typemap directly.
2008-08-07 15:38:54 -07:00
Andreas Gal
95f5262e4f
Use an oracle to predict when slots are demotable. This allows trashing the TreeInfo* object in the root fragment more aggressively. Rebuild the stack type map at every compile to avert type map mismatches. Purge all global slot info when we re-record a tree.
2008-08-07 15:28:43 -07:00
Andreas Gal
e09041e2a5
If we run into an error during compilation, blacklist that fragment. If we run out of memory, flush the cache.
2008-08-06 22:26:20 -07:00
Andreas Gal
0fe0cf9a98
Flush JIT cache for all contexts. Clear nanojit error state when we start compiling.
2008-08-06 22:18:33 -07:00
Andreas Gal
e9733ceed9
Remove dead code.
2008-08-06 22:10:30 -07:00
Andreas Gal
9385e7842c
Add a helper to flush the JIT code cache (and the fragment lookup quick cache). On a global shape mismatch trash the entire cache (might be a bit overly aggressive). Similary, during GC flush the code cache.
2008-08-06 21:56:25 -07:00
Andreas Gal
d1a1a76121
Tracker outer trees for every tree and merge globals of inner trees into all outer trees as we register inner trees with outer trees.
2008-08-06 19:25:24 -07:00
Andreas Gal
3785f2a5af
Fix misleading debug text.
2008-08-06 19:04:41 -07:00
Andreas Gal
ff22344594
Fix tracing of code inside methods that were called with arity mismatch.
2008-08-06 18:51:26 -07:00
Brendan Eich
f23efc77b2
Replace bogus assertion with runtime test for correct abort test (getter or setter, given sprop hit from propcache).
2008-08-06 17:54:36 -07:00
Andreas Gal
d41941073e
Disable outerlining, to be replaced with nesting.
2008-08-06 17:34:06 -07:00
Andreas Gal
1bf2697257
Add a helper to emit a tree-call into the currently recording trace.
2008-08-06 16:24:29 -07:00
Andreas Gal
27de48d1f9
Cleanup stack offset calculation and eliminate the stack offset fiddling that was necessary for nanojit prior to adding the getTop hook.
2008-08-06 15:04:02 -07:00
Brendan Eich
0f1f29e9a4
Merge again, I lost to Andreas\!
2008-08-06 13:44:07 -07:00
Andreas Gal
2acb2afdec
Always trashing the tree on a type mismatch can't work since the first iteration comes in as undefined for loop-outputs and then kills the main tree that deals with the proper stable types. Have to find a different way to deal with this.
2008-08-06 13:41:04 -07:00
Andreas Gal
fcaab6d007
Pull recompile flag out of the state and pass in as argument. Trash the tree if a secondary trace can't be connected to the loop header. This is very aggressive and might need more tinkering. Trashing the tree on every mismatch doesn't seem to work well, so thats currently disabled.
2008-08-06 13:37:29 -07:00
Brendan Eich
6e215765df
- Add builtins to support for-in loops, both iterating and getting/setter properties by name using o[i] instead of o.p where i is 'p'.
...
- Record JSOP_ITER and JSOP_ENDITER, which must be handled since tracing could start in between them (and always will, until we nest or unless we outerline).
- Add a null state guard in JSOP_FORVAR's recorder method.
- Pick include order, indentation, and trailing space nits.
2008-08-06 13:36:29 -07:00
Andreas Gal
a253b66a09
If we stop recording, immediately attempt to trigger the tree.
2008-08-06 13:18:30 -07:00
Andreas Gal
141f71186c
Properly deal with trace entry with non-empty stack.
2008-08-06 12:38:17 -07:00
Andreas Gal
0ed65746a7
Split side exit handling from js_ExecuteTree.
2008-08-06 06:34:16 -07:00
Brendan Eich
03b511f9e4
- Use JSStackFrame* fp over, don't declare another JSStackFrame* f, in the FORALL macros and clones (f is canonical variable name for nanojit::Fragment*, fp for JSStackFrame*).
...
- Nit-pick include order (alphabetical within each group).
- Trailing space and indentation no-knock raid.
2008-08-03 22:55:28 -07:00
Brendan Eich
da42a8c6f7
Strength-reduced unsigned modulus in the fragment quick cache, don't use signed % which requires a branch and less-than-zero test.
2008-08-03 22:35:15 -07:00
Brendan Eich
df36d9e36f
Use INS_CONST to addName a few lir->insImm immediates.
2008-08-02 20:01:36 -07:00
Brendan Eich
5b9f299115
Fix TraceRecorder::ifop backward logic bug in the OBJECT case, and implement the STRING case. Add tests for truthy and falsy strings.
2008-08-02 16:51:22 -07:00
Brendan Eich
607e259b69
Fix return NULL in bool to return false.
2008-08-02 16:05:29 -07:00
Andreas Gal
3106dc8845
Use quick cache to bypass fragmento is possible.
2008-08-01 22:56:52 -07:00
Andreas Gal
12f1c06952
Factor out most of the remaining code in js_LoopEdge into js_RecordTree and move the code to attach new branches into js_ExecuteTree.
2008-08-01 22:47:15 -07:00
Andreas Gal
746cb91802
Factor out js_ContinueRecording from js_LoopEdge.
2008-08-01 22:39:52 -07:00
Andreas Gal
3fea3f5d53
Split up FORALL_SLOTS_IN_PENDING_FRAMES macro in a macro that processes a frame, and a macro that uses that macro to process all pending frames.
2008-08-01 22:27:39 -07:00
Brendan Eich
3691962a2f
Extend ifop to handle undefined tests; implement JSOP_OR and JSOP_AND (so much work\!).
2008-08-01 12:06:04 -07:00
Brendan Eich
dc1072fe55
Style police raid: function names start in column 1, with one blank line between functions.
2008-08-01 11:03:52 -07:00
Brendan Eich
2cfdd9a55a
1. Don't store thisp literally in state.rp, get it from argv[-1]. 2. Fix DEBUG localNames code.
2008-08-01 10:40:38 -07:00
Brendan Eich
8f837e0784
- Implement interpreter frame reconstruction (js_SynthesizeFrame).
...
- Fix MONITOR_BRANCH to restore frame-dependent interpreter locals (we want fewer of these if they can be removed with no perf loss).
- Fix FORALL_SLOTS_IN_PENDING_FRAMES not to scan argv in callee when it has scanned operand stack in caller.
- Fix import to take the localFrame from whose fun its localNames parameter was computed -- it was using cx->fp which is wrong when FORALL_SLOTS_IN_PENDING_FRAMES iterates other than the top of stack frame.
- A few interval tests that were double-ended are single-ended now.
- Add call.js mini-test (more cases needed).
2008-08-01 08:26:32 -07:00
Andreas Gal
4eba8e4012
Remove entryStackDepth. Calculate sp_adj relative to entryNativeStackSlots.
2008-08-01 02:33:54 -07:00
shaver@mozilla.org
3221386525
add specialized StringToInt32 and filter for it
2008-07-31 11:35:08 -07:00
shaver@mozilla.org
12614d0334
Coerce strings to numbers for appropriate ops.
...
(Wants a specialized StringToInt32 and a filter to put it in place.)
2008-07-31 07:41:58 -07:00
shaver@mozilla.org
d200eb863e
merge backout of d24e6005ee4c to fix the world
2008-07-31 06:36:53 -07:00
shaver@mozilla.org
f2c0c93637
Backed out changeset d24e6005ee4c (causing major array-fail).
2008-07-31 06:35:11 -07:00
Andreas Gal
e5a87ade9f
Merge.
2008-07-31 01:35:56 -07:00
Andreas Gal
6f4cfc2628
Remove ANY_TYPE and move debug printf around to print entry point even if we can't enter due to type mismatch.
2008-07-31 01:35:18 -07:00
shaver@mozilla.org
a3bc7296d9
Make guard return expected, so we can trace alternate cases easily.
...
Better diagnostic for non-global scope chain head.
2008-07-30 22:59:13 -07:00
Andreas Gal
86aa1f7ff7
If trees are not enabled, don't try to reuse state and param1.
2008-07-30 21:40:21 -07:00
David Anderson
ad3e892e92
Re-use initial parameters on tree fragments.
2008-07-30 23:30:58 -05:00
shaver@mozilla.org
959dd9184f
Implement JSOP_EQ and JSOP_NE over strings, plus JSOP_STRING.
2008-07-30 17:28:59 -07:00
Andreas Gal
e91eb4bc47
Merge.
2008-07-30 16:37:35 -07:00
Brendan Eich
fe3a6127c5
- Export JSSLOT_ITER_* from jsiter.cpp to jsiter.h, for jstracer.cpp to use.
...
- Fix OBJ_GET_SLOT to be STOBJ_GET_SLOT in jsiter.cpp, no thread safety here (bug on file).
- Move JSNativeEnumerator from jsobj.cpp to jsobj.h for jstracer.cpp as well.
- Rename JOF_2BYTE JOF_UINT8 for consistency, and actually decompile it.
- Trace JSOP_FORVAR (can't do anything in JSOP_ITER, it comes before the loop).
- Shortened some guardMyLongSummerVacationWithinBounds names ;-).
- Removed/refactored to avoid dslots_ins obligation on all callers.
2008-07-30 16:32:33 -07:00
Andreas Gal
698d3fb5cc
Make the global frame layout match the slot layout in the global object. This will allow leaving global values in place when switching trees as long both inner and outer tree use the same value.
2008-07-30 16:28:48 -07:00
Andreas Gal
e8043788a7
Merge.
2008-07-30 16:06:11 -07:00
Andreas Gal
9b9d60583c
Use isGlobal to distinguish whether a value is a global slot instead of scanning the table every time.
2008-07-30 16:05:51 -07:00
shaver@mozilla.org
124e9ae44b
trace Math.random
2008-07-30 15:51:44 -07:00
