| 8b095215a4d5bde723819087f3455bdcc250a78f |
|
20-Jun-2012 |
David Butcher <david.butcher@arm.com> |
Switched code to blx <reg>
ldr ip,<addr> blx ip is preferred over mov lr,pc ldr pc,<addr> from armv5te,
and will typically perform better on later ARM processors.
Change-Id: I8f2e5e794c644faafd767037ad56579f2934de47
|
| 30bc0d46ae730d78c42c39cfa56a59ba3025380b |
|
22-Apr-2011 |
buzbee <buzbee@google.com> |
Consolidate curFrame fields in thread storage
We ended up with two locations in the Thread structure for saved
Dalvik frame pointer. This change consolidates them.
Change-Id: I78f288e4e57e232f29663be930101e775bfe370f
|
| 9a3147c7412f4794434b4c2604aa2ba784867774 |
|
03-Mar-2011 |
buzbee <buzbee@google.com> |
Interpreter restructuring
This is a restructuring of the Dalvik ARM and x86 interpreters:
o Combine the old portstd and portdbg interpreters into a single
portable interpreter.
o Add debug/profiling support to the fast (mterp) interpreters.
o Delete old mechansim of switching between interpreters. Now, once
you choose an interpreter at startup, you stick with it.
o Allow JIT to co-exist with profiling & debugging (necessary for
first-class support of debugging with the JIT active).
o Adds single-step capability to the fast assembly interpreters without
slowing them down (and, in fact, measurably improves their performance).
o Remove old "polling for safe point" mechanism. Breakouts now achieved
via modifying base of interpreter handler table.
o Simplify interpeter control mechanism.
o Allow thread-granularity control for profiling & debugging
The primary motivation behind this change was to improve the responsiveness
of debugging and profiling and to make it easier to add new debugging and
profiling capabilities in the future. Instead of always bailing out to the
slow debug portable interpreter, we can now stay in the fast interpreter.
A nice side effect of the change is that the fast interpreters
got a healthy speed boost because we were able to replace the
polling safepoint check that involved a dozen or so instructions
with a single table-base reload. When combined with the two earlier CLs
related to this restructuring, we show a 5.6% performance improvement
using libdvm_interp.so on the Checkers benchmark relative to Honeycomb.
Change-Id: I8d37e866b3618def4e582fc73f1cf69ffe428f3c
|
| 9f601a917c8878204482c37aec7005054b6776fa |
|
12-Feb-2011 |
buzbee <buzbee@google.com> |
Interpreter restructuring: eliminate InterpState
The key datastructure for the interpreter is InterpState.
This change eliminates it, merging its data with the Thread structure.
Here's why:
In principio creavit Fadden Thread et InterpState. And it was good.
Thread holds thread-private state, while InterpState captures data
associated with a Dalvik interpreter activation. Because JNI calls
can result in nested interpreter invocations, we can have more than one
InterpState for each actual thread. InterpState was relatively small,
and it all worked well. It was used enough that in the Arm version
a register (rGLUE) was dedicated to it.
Then, along came the JIT guys, who saw InterpState as a convenient place
to dump all sorts of useful data that they wanted quick access to through
that dedicated register. InterpState grew and grew. In terms of
space, this wasn't a big problem - but it did mean that the initialization
cost of each interpreter activation grew as well. For applications
that do a lot of callbacks from native code into Dalvik, this is
measurable. It's also mostly useless cost because much of the JIT-related
InterpState initialization was setting up useful constants - things that
don't need to be saved and restored all the time.
The biggest problem, though, deals with thread control. When something
interesting is happening that needs all threads to be stopped (such as
GC and debugger attach), we have access to all of the Thread structures,
but we don't have access to all of the InterpState structures (which
may be buried/nested on the native stack). As a result, polling for
thread suspension is done via a one-indirection pointer chase. InterpState
itself can't hold the stop bits because we can't always find it, so
instead it holds a pointer to the global or thread-specific stop control.
Yuck.
With this change, we eliminate InterpState and merge all needed data
into Thread. Further, we replace the decidated rGLUE register with a
pointer to the Thread structure (rSELF). The small subset of state
data that needs to be saved and restored across nested interpreter
activations is collected into a record that is saved to the interpreter
frame, and restored on exit. Further, these small records are linked
together to allow tracebacks to show nested activations. Old InterpState
variables that simply contain useful constants are initialized once at
thread creation time.
This CL is large enough by itself that the new ability to streamline
suspend checks is not done here - that will happen in a future CL. Here
we just focus on consolidation.
Change-Id: Ide6b2fb85716fea454ac113f5611263a96687356
|
| 18fba346582c08d81aa96d9508c0e935bad5f36f |
|
20-Jan-2011 |
buzbee <buzbee@google.com> |
Support traceview-style profiling in all builds
This change builds on an earlier bccheng change that allowed JIT'd code
to avoid reverting to the debug portable interpeter when doing traceview-style
method profiling. That CL introduced a new traceview build (libdvm_traceview)
because the performance delta was too great to enable the capability for
all builds.
In this CL, we remove the libdvm_traceview build and provide full-speed
method tracing in all builds. This is done by introducing "_PROF"
versions of invoke and return templates used by the JIT. Normally, these
templates are not used, and performace in unaffected. However, when method
profiling is enabled, all existing translation are purged and new translations
are created using the _PROF templates. These templates introduce a
smallish performance penalty above and beyond the actual tracing cost, but
again are only used when tracing has been enabled.
Strictly speaking, there is a slight burden that is placed on invokes and
returns in the non-tracing case - on the order of an additional 3 or 4
cycles per invoke/return. Those operations are already heavyweight enough
that I was unable to measure the added cost in benchmarks.
Change-Id: Ic09baf4249f1e716e136a65458f4e06cea35fc18
|
| d88756df5b4dbc6fd450afd0019a5f64ebe4432d |
|
22-Oct-2010 |
Elliott Hughes <enh@google.com> |
Remove junk from platform.S now armv4t is gone.
Change-Id: I30079aacc753c89cfc3a3f64bd900a0cc858d65f
|
| c8293e7dfe856ca95e27aef1ac2e64d750d60662 |
|
12-Oct-2010 |
Ben Cheng <bccheng@android.com> |
Fine-tune the instructions on the method invocation path.
1) Initialize the register and out sizes for callee methods through
constant moves.
2) Eliminate an unnecessary load of Dalvik PC for chained and
native callees.
Improved method invocation performance by ~3%.
Change-Id: Iead1276eed0ba527e82eb876f08d169ab9b496b2
|
| 5cc61d70ec727aa22f58463bf7940cc717cf3eb1 |
|
31-Aug-2010 |
Ben Cheng <bccheng@android.com> |
Collect method traces with the fast interpreter and the JIT'ed code.
Insert inline code instead of switching to the debug interpreter in the hope
that the time stamps collected in traceview are more close to the real
world behavior with minimal profiling overhead.
Because the inline polling still introduces additional overhead (20% ~ 100%),
it is only enabled in the special VM build called "libdvm_traceview.so".
It won't work on the emulator because it is not implemented to collect the
detailed instruction traces.
Here are some performance numbers using the FibonacciSlow microbenchmark
(ie recursive workloads / the shorter the faster):
time: configuration
8,162,602: profiling off/libdvm.so/JIT off
2,801,829: profiling off/libdvm.so/JIT on
9,952,236: profiling off/libdvm_traceview.so/JIT off
4,465,701: profiling off/libdvm_traceview.so/JIT on
164,786,585: profiling on/libdvm.so/JIT off
164,664,634: profiling on/libdvm.so/JIT on
11,231,707: profiling on/libdvm_traceview.so/JIT off
8,427,846: profiling on/libdvm_traceview.so/JIT on
Comparing the 8,427,846 vs 164,664,634 numbers againt the true baseline
performance number of 2,801,829, the new libdvm_traceview.so improves the time
skew from 58x to 3x.
Change-Id: I48611a3a4ff9c4950059249e5503c26abd6b138e
|
| 7365493ad8d360c1dcf9cd8b6eee62747af01cae |
|
09-Jun-2010 |
Carl Shapiro <cshapiro@google.com> |
Remove repeated newlines at the end of files.
Change-Id: I1e3d103a7b932ef21acedb6438c0f26b315df28f
|
| fbdcfb9ea9e2a78f295834424c3f24986ea45dac |
|
29-May-2010 |
Brian Carlstrom <bdc@google.com> |
Merge remote branch 'goog/dalvik-dev' into dalvik-dev-to-master
Change-Id: I0c0edb3ebf0d5e040d6bbbf60269fab0deb70ef9
|
| 978738d2cbf9d08fa78c65762eaac3351ab76b9a |
|
13-May-2010 |
Ben Cheng <bccheng@android.com> |
Add counters to track JIT inline cache hit rate and code cache patch counts.
Also did some WITH_JIT_TUNING cleanup.
Change-Id: I8bb2d681a06b0f2af1f976a007326825a88cea38
|
| a62475ecfcc80c58add8f153c9605762dafb8227 |
|
30-Apr-2010 |
Ben Cheng <bccheng@android.com> |
Use unsigned comparison for stack pointers.
Bug: 2613607
Change-Id: I6a8abd69fbf9cb9f8ec9d9febf1ea42fd631fe9c
|
| 86717f79d9b018f4d69cc991075fa36611f234e5 |
|
06-Mar-2010 |
Ben Cheng <bccheng@android.com> |
Collect more JIT stats in the assert build.
New stuff includes breakdown of callsite types (ie monomorphic vs polymorphic
vs monoporphic resolved to native), total time spent in JIT'ing, and average
JIT time per compilation.
Example output:
D/dalvikvm( 840): 4042 compilations using 1976 + 329108 bytes
D/dalvikvm( 840): Compiler arena uses 10 blocks (8100 bytes each)
D/dalvikvm( 840): Compiler work queue length is 0/36
D/dalvikvm( 840): size if 8192, entries used is 4137
D/dalvikvm( 840): JIT: 4137 traces, 8192 slots, 1099 chains, 40 thresh, Non-blocking
D/dalvikvm( 840): JIT: Lookups: 1128780 hits, 168564 misses; 179520 normal, 6 punt
D/dalvikvm( 840): JIT: noChainExit: 528464 IC miss, 194708 interp callsite, 0 switch overflow
D/dalvikvm( 840): JIT: Invoke: 507 mono, 988 poly, 72 native, 1038 return
D/dalvikvm( 840): JIT: Total compilation time: 2342 ms
D/dalvikvm( 840): JIT: Avg unit compilation time: 579 us
D/dalvikvm( 840): JIT: 3357 Translation chains, 97 interp stubs
D/dalvikvm( 840): dalvik.vm.jit.op = 0-2,4-5,7-8,a-c,e-16,19-1a,1c-23,26,28-29,2b-2f,31-3d,44-4b,4d-51,60,62-63,68-69,70-72,76-78,7b,81-82,84,87,89,8d-93,95-98,a1,a3,a6,a8-a9,b0-b3,b5-b6,bb-bf,c6-c8,d0,d2-d6,d8,da-e2,ee-f0,f2-fb,
D/dalvikvm( 840): Code size stats: 50666/105126 (compiled/total Dalvik), 329108 (native)
|
| 40094c16d9727cc1e047a7d4bddffe04dd566211 |
|
25-Feb-2010 |
Ben Cheng <bccheng@android.com> |
Tweak the interpreter entries and 2nd level trace filter to capture more traces.
Real changes:
1) Add a new entry point from JIT to the interpreter to request hot traces w/o
doing chaining.
2) Increase the granularity of the secondary profile filter to match 64-byte
chunks using 64 entries.
The remaining are just cosmetic changes.
|
| 7a0bcd0de6c4da6499a088a18d1750e51204c2a6 |
|
23-Jan-2010 |
Ben Cheng <bccheng@android.com> |
Tighten the safe points for code cache resets to happen.
Add a new flag in the Thread struct to track the whereabout of the top frame
in each Java thread. It is not safe to blow away the code cache if any thread
is in the JIT'ed land.
|
| 60c24f436d603c564d5351a6f81821f12635733c |
|
04-Jan-2010 |
Ben Cheng <bccheng@google.com> |
Tear down the code cache when it is full and restart from scratch.
Because the code cache may be wiped out after safe points now the patching of
inline cache for predicted chains is done through the compiler thread's work
queue.
|
| 909b418219f63c0d0b2bde8a0835dbf27d5061b8 |
|
03-Dec-2009 |
Bill Buzbee <buzbee@google.com> |
Jit: Fix for 2187020, bad exception recovery from native invoke static
|
| 72e93344b4d1ffc71e9c832ec23de0657e5b04a5 |
|
13-Nov-2009 |
Jean-Baptiste Queru <jbq@google.com> |
eclair snapshot
|
| d5ab726b65d7271be261864c7e224fb90bfe06e0 |
|
25-Aug-2009 |
Andy McFadden <fadden@android.com> |
Another round of scary indirect ref changes.
This change adds a not-really-working implementation to Jni.c, with
various changes #ifdefed throughout the code. The ifdef is currently
disabled, so the old behavior should continue. Eventually the old
version will be stripped out and the ifdefs removed.
This renames the stack's "localRefTop" field, which nudged a bunch of
code. The name wasn't really right before (it's the *bottom* of the
local references), and it's even less right now. This and one other
mterp-visible constant were changed, which caused some ripples through
mterp and the JIT, but the ifdeffing was limited to one in
asm-constants.h (and the constant is the same both ways, so toggling the
ifdef won't require rebuilding asm sources).
Some comments and arg names in ReferenceTable were updated for the
correct orientation of bottom vs. top.
Some adjustments were made to the JNI code, e.g. dvmCallMethod now needs
to understand if it needs to convert reference arguments from
local/global refs to pointers (it's called from various places
throughout the VM).
|
| 97319a8a234e9fe1cf90ca39aa6eca37d729afd5 |
|
13-Aug-2009 |
Jeff Hao <jeffhao@google.com> |
New changes to enable self verification mode.
|
| 38329f5678fd7a4879528b02a0ab60322d38a897 |
|
07-Jul-2009 |
Ben Cheng <bccheng@android.com> |
Improved method invocation performance: 1.5x for virtual and 2.8x for interface.
- Implemented predicted chaining for invoke virtual and interface.
- Eliminated a little bit of fat for invoke native.
- Added 078-polymorphic-virtual for stress tests.
|