view docs/DEVELOPER_README @ 290:fdfb4edd78d6

8011630: JSON parsing performance issue Reviewed-by: lagergren, sundar
author hannesw
date Fri, 24 May 2013 13:54:18 +0200
parents 43e40c08e7f8
children eb7b8340ce3a
line wrap: on
line source

This document describes system properties that are used for internal
debugging and instrumentation purposes, along with the system loggers,
which are used for the same thing.

This document is intended as a developer resource, and it is not
needed as Nashorn documentation for normal usage. Flags and system
properties described herein are subject to change without notice.

=====================================
1. System properties used internally
=====================================

This documentation of the system property flags assume that the
default value of the flag is false, unless otherwise specified.

SYSTEM PROPERTY: -Dnashorn.args=<string>

This property takes as its value a space separated list of Nashorn
command line options that should be passed to Nashorn. This might be useful 
in environments where it is hard to tell how a nashorn.jar is launched.

Example:

> java -Dnashorn.args="--lazy-complation --log=compiler" large-java-app-with-nashorn.jar 
> ant -Dnashorn.args="--log=codegen" antjob

SYSTEM PROPERTY: -Dnashorn.unstable.relink.threshold=x

This property controls how many call site misses are allowed before a 
callsite is relinked with "apply" semantics to never change again. 
In the case of megamorphic callsites, this is necessary, or the 
program would spend all its time swapping out callsite targets. Dynalink 
has a default value (currently 8 relinks) for this property if it 
is not explicitly set.


SYSTEM PROPERTY: -Dnashorn.compiler.splitter.threshold=x

This will change the node weight that requires a subgraph of the IR to
be split into several classes in order not to run out of bytecode space.
The default value is 0x8000 (32768).


SYSTEM PROPERTY: -Dnashorn.compiler.intarithmetic

Arithmetic operations in Nashorn (except bitwise ones) typically
coerce the operands to doubles (as per the JavaScript spec). To switch
this off and remain in integer mode, for example for "var x = a&b; var
y = c&d; var z = x*y;", use this flag. This will force the
multiplication of variables that are ints to be done with the IMUL
bytecode and the result "z" to become an int.

WARNING: Note that is is experimental only to ensure that type support
exists for all primitive types. The generated code is unsound. This
will be the case until we do optimizations based on it. There is a CR
in Nashorn to do better range analysis, and ensure that this is only
done where the operation can't overflow into a wider type. Currently
no overflow checking is done, so at the moment, until range analysis
has been completed, this option is turned off.

We've experimented by using int arithmetic for everything and putting
overflow checks afterwards, which would recompute the operation with
the correct precision, but have yet to find a configuration where this
is faster than just using doubles directly, even if the int operation
does not overflow. Getting access to a JVM intrinsic that does branch
on overflow would probably alleviate this.

There is also a problem with this optimistic approach if the symbol
happens to reside in a local variable slot in the bytecode, as those
are strongly typed. Then we would need to split large sections of
control flow, so this is probably not the right way to go, while range
analysis is. There is a large difference between integer bytecode
without overflow checks and double bytecode. The former is
significantly faster.


SYSTEM PROPERTY: -Dnashorn.codegen.debug, -Dnashorn.codegen.debug.trace=<x>

See the description of the codegen logger below.


SYSTEM_PROPERTY: -Dnashorn.fields.debug

See the description on the fields logger below.


SYSTEM PROPERTY: -Dnashorn.fields.dual

When this property is true, Nashorn will attempt to use primitive
fields for AccessorProperties (currently just AccessorProperties, not
spill properties). Memory footprint for script objects will increase,
as we need to maintain both a primitive field (a long) as well as an
Object field for the property value. Ints are represented as the 32
low bits of the long fields. Doubles are represented as the
doubleToLongBits of their value. This way a single field can be used
for all primitive types. Packing and unpacking doubles to their bit
representation is intrinsified by the JVM and extremely fast.

While dual fields in theory runs significantly faster than Object
fields due to reduction of boxing and memory allocation overhead,
there is still work to be done to make this a general purpose
solution. Research is ongoing.

In the future, this might complement or be replaced by experimental
feature sun.misc.TaggedArray, which has been discussed on the mlvm
mailing list. TaggedArrays are basically a way to share data space
between primitives and references, and have the GC understand this.

As long as only primitive values are written to the fields and enough
type information exists to make sure that any reads don't have to be
uselessly boxed and unboxed, this is significantly faster than the
standard "Objects only" approach that currently is the default. See
test/examples/dual-fields-micro.js for an example that runs twice as
fast with dual fields as without them. Here, the compiler, can
determine that we are dealing with numbers only throughout the entire
property life span of the properties involved.

If a "real" object (not a boxed primitive) is written to a field that
has a primitive representation, its callsite is relinked and an Object
field is used forevermore for that particular field in that
PropertyMap and its children, even if primitives are later assigned to
it.

As the amount of compile time type information is very small in a
dynamic language like JavaScript, it is frequently the case that
something has to be treated as an object, because we don't know any
better. In reality though, it is often a boxed primitive is stored to
an AccessorProperty. The fastest way to handle this soundly is to use
a callsite typecheck and avoid blowing the field up to an Object. We
never revert object fields to primitives. Ping-pong:ing back and forth
between primitive representation and Object representation would cause
fatal performance overhead, so this is not an option.

For a general application the dual fields approach is still slower
than objects only fields in some places, about the same in most cases,
and significantly faster in very few. This is due the program using
primitives, but we still can't prove it. For example "local_var a =
call(); field = a;" may very well write a double to the field, but the
compiler dare not guess a double type if field is a local variable,
due to bytecode variables being strongly typed and later non
interchangeable. To get around this, the entire method would have to
be replaced and a continuation retained to restart from. We believe
that the next steps we should go through are instead:

1) Implement method specialization based on callsite, as it's quite
frequently the case that numbers are passed around, but currently our
function nodes just have object types visible to the compiler. For
example "var b = 17; func(a,b,17)" is an example where two parameters
can be specialized, but the main version of func might also be called
from another callsite with func(x,y,"string").

2) This requires lazy jitting as the functions have to be specialized
per callsite.

Even though "function square(x) { return x*x }" might look like a
trivial function that can always only take doubles, this is not
true. Someone might have overridden the valueOf for x so that the
toNumber coercion has side effects. To fulfil JavaScript semantics,
the coercion has to run twice for both terms of the multiplication
even if they are the same object. This means that call site
specialization is necessary, not parameter specialization on the form
"function square(x) { var xd = (double)x; return xd*xd; }", as one
might first think.

Generating a method specialization for any variant of a function that
we can determine by types at compile time is a combinatorial explosion
of byte code (try it e.g. on all the variants of am3 in the Octane
benchmark crypto.js). Thus, this needs to be lazy

3) Possibly optimistic callsite writes, something on the form

x = y; //x is a field known to be a primitive. y is only an object as
far as we can tell

turns into

try {
  x = (int)y;
} catch (X is not an integer field right now | ClassCastException e) {
  x = y;
}

Mini POC shows that this is the key to a lot of dual field performance
in seemingly trivial micros where one unknown object, in reality
actually a primitive, foils it for us. Very common pattern. Once we
are "all primitives", dual fields runs a lot faster than Object fields
only.

We still have to deal with objects vs primitives for local bytecode
slots, possibly through code copying and versioning.


SYSTEM PROPERTY: -Dnashorn.compiler.symbol.trace=[<x>[,*]], 
  -Dnashorn.compiler.symbol.stacktrace=[<x>[,*]]

When this property is set, creation and manipulation of any symbol
named "x" will show information about when the compiler changes its
type assumption, bytecode local variable slot assignment and other
data. This is useful if, for example, a symbol shows up as an Object,
when you believe it should be a primitive. Usually there is an
explanation for this, for example that it exists in the global scope
and type analysis has to be more conservative. 

Several symbols names to watch can be specified by comma separation.

If no variable name is specified (and no equals sign), all symbols
will be watched

By using "stacktrace" instead of or together with "trace", stack
traces will be displayed upon symbol changes according to the same
semantics.


SYSTEM PROPERTY: nashorn.lexer.xmlliterals

If this property it set, it means that the Lexer should attempt to
parse XML literals, which would otherwise generate syntax
errors. Warning: there are currently no unit tests for this
functionality.

XML literals, when this is enabled, end up as standard LiteralNodes in
the IR.


SYSTEM_PROPERTY: nashorn.debug

If this property is set to true, Nashorn runs in Debug mode. Debug
mode is slightly slower, as for example statistics counters are enabled
during the run. Debug mode makes available a NativeDebug instance
called "Debug" in the global space that can be used to print property
maps and layout for script objects, as well as a "dumpCounters" method
that will print the current values of the previously mentioned stats
counters.

These functions currently exists for Debug:

"map" - print(Debug.map(x)) will dump the PropertyMap for object x to
stdout (currently there also exist functions called "embedX", where X
is a value from 0 to 3, that will dump the contents of the embed pool
for the first spill properties in any script object and "spill", that
will dump the contents of the growing spill pool of spill properties
in any script object. This is of course subject to change without
notice, should we change the script object layout.

"methodHandle" - this method returns the method handle that is used
for invoking a particular script function.

"identical" - this method compares two script objects for reference
equality. It is a == Java comparison

"dumpCounters" - will dump the debug counters' current values to
stdout.

Currently we count number of ScriptObjects in the system, number of
Scope objects in the system, number of ScriptObject listeners added,
removed and dead (without references).

We also count number of ScriptFunctions, ScriptFunction invocations
and ScriptFunction allocations.

Furthermore we count PropertyMap statistics: how many property maps
exist, how many times were property maps cloned, how many times did
the property map history cache hit, prevent new allocations, how many
prototype invalidations were done, how many time the property map
proto cache hit.

Finally we count callsite misses on a per callsite bases, which occur
when a callsite has to be relinked, due to a previous assumption of
object layout being invalidated.


SYSTEM PROPERTY: nashorn.methodhandles.debug,
nashorn.methodhandles.debug=create

If this property is enabled, each MethodHandle related call that uses
the java.lang.invoke package gets its MethodHandle intercepted and an
instrumentation printout of arguments and return value appended to
it. This shows exactly which method handles are executed and from
where. (Also MethodTypes and SwitchPoints). This can be augmented with
more information, for example, instance count, by subclassing or
further extending the TraceMethodHandleFactory implementation in
MethodHandleFactory.java.

If the property is specialized with "=create" as its option,
instrumentation will be shown for method handles upon creation time
rather than at runtime usage.


SYSTEM PROPERTY: nashorn.methodhandles.debug.stacktrace

This does the same as nashorn.methodhandles.debug, but when enabled
also dumps the stack trace for every instrumented method handle
operation. Warning: This is enormously verbose, but provides a pretty
decent "grep:able" picture of where the calls are coming from.

See the description of the codegen logger below for a more verbose
description of this option


SYSTEM PROPERTY: nashorn.scriptfunction.specialization.disable

There are several "fast path" implementations of constructors and
functions in the NativeObject classes that, in their original form,
take a variable amount of arguments. Said functions are also declared
to take Object parameters in their original form, as this is what the
JavaScript specification mandates.

However, we often know quite a lot more at a callsite of one of these
functions. For example, Math.min is called with a fixed number (2) of
integer arguments. The overhead of boxing these ints to Objects and
folding them into an Object array for the generic varargs Math.min
function is an order of magnitude slower than calling a specialized
implementation of Math.min that takes two integers. Specialized
functions and constructors are identified by the tag
@SpecializedFunction and @SpecializedConstructor in the Nashorn
code. The linker will link in the most appropriate (narrowest types,
right number of types and least number of arguments) specialization if
specializations are available.

Every ScriptFunction may carry specializations that the linker can
choose from. This framework will likely be extended for user defined
functions. The compiler can often infer enough parameter type info
from callsites for in order to generate simpler versions with less
generic Object types. This feature depends on future lazy jitting, as
there tend to be many calls to user defined functions, some where the
callsite can be specialized, some where we mostly see object
parameters even at the callsite.

If this system property is set to true, the linker will not attempt to
use any specialized function or constructor for native objects, but
just call the generic one.


SYSTEM PROPERTY: nashorn.tcs.miss.samplePercent=<x>

When running with the trace callsite option (-tcs), Nashorn will count
and instrument any callsite misses that require relinking. As the
number of relinks is large and usually produces a lot of output, this
system property can be used to constrain the percentage of misses that
should be logged. Typically this is set to 1 or 5 (percent). 1% is the
default value.


SYSTEM_PROPERTY: nashorn.profilefile=<filename>

When running with the profile callsite options (-pcs), Nashorn will
dump profiling data for all callsites to stderr as a shutdown hook. To
instead redirect this to a file, specify the path to the file using
this system property.


SYSTEM_PROPERTY: nashorn.regexp.impl=[jdk|joni]

This property defines the regular expression engine to be used by
Nashorn. The default implementation is "jdk" which is based on the
JDK's java.util.regex package. Set this property to "joni" to install
an implementation based on Joni, the regular expression engine used by
the JRuby project.


===============
2. The loggers.
===============

It is very simple to create your own logger. Use the DebugLogger class
and give the subsystem name as a constructor argument.

The Nashorn loggers can be used to print per-module or per-subsystem
debug information with different levels of verbosity. The loggers for
a given subsystem are available are enabled by using

--log=<systemname>[:<level>]

on the command line.

Here <systemname> identifies the name of the subsystem to be logged
and the optional colon and level argument is a standard
java.util.logging.Level name (severe, warning, info, config, fine,
finer, finest). If the level is left out for a particular subsystem,
it defaults to "info". Any log message logged as the level or a level
that is more important will be output to stderr by the logger.

Several loggers can be enabled by a single command line option, by
putting a comma after each subsystem/level tuple (or each subsystem if
level is unspecified). The --log option can also be given multiple
times on the same command line, with the same effect.

For example: --log=codegen,fields:finest is equivalent to
--log=codegen:info --log=fields:finest

The subsystems that currently support logging are:


* compiler

The compiler is in charge of turning source code and function nodes
into byte code, and installs the classes into a class loader
controlled from the Context. Log messages are, for example, about
things like new compile units being allocated. The compiler has global
settings that all the tiers of codegen (e.g. Lower and CodeGenerator)
use.s


* codegen

The code generator is the emitter stage of the code pipeline, and
turns the lowest tier of a FunctionNode into bytecode. Codegen logging
shows byte codes as they are being emitted, line number information
and jumps. It also shows the contents of the bytecode stack prior to
each instruction being emitted. This is a good debugging aid. For
example:

[codegen] #41                       line:2 (f)_afc824e 
[codegen] #42                           load symbol x slot=2 
[codegen] #43  {1:O}                    load int 0 
[codegen] #44  {2:I O}                  dynamic_runtime_call GT:ZOI_I args=2 returnType=boolean 
[codegen] #45                              signature (Ljava/lang/Object;I)Z 
[codegen] #46  {1:Z}                    ifeq  ternary_false_5402fe28 
[codegen] #47                           load symbol x slot=2 
[codegen] #48  {1:O}                    goto ternary_exit_107c1f2f 
[codegen] #49                       ternary_false_5402fe28 
[codegen] #50                           load symbol x slot=2 
[codegen] #51  {1:O}                    convert object -> double 
[codegen] #52  {1:D}                    neg 
[codegen] #53  {1:D}                    convert double -> object 
[codegen] #54  {1:O}                ternary_exit_107c1f2f 
[codegen] #55  {1:O}                    return object 

shows a ternary node being generated for the sequence "return x > 0 ?
x : -x"

The first number on the log line is a unique monotonically increasing
emission id per bytecode. There is no guarantee this is the same id
between runs.  depending on non deterministic code
execution/compilation, but for small applications it usually is. If
the system variable -Dnashorn.codegen.debug.trace=<x> is set, where x
is a bytecode emission id, a stack trace will be shown as the
particular bytecode is about to be emitted. This can be a quick way to
determine where it comes from without attaching the debugger. "Who
generated that neg?"

The --log=codegen option is equivalent to setting the system variable
"nashorn.codegen.debug" to true.


* lower

This is the first lowering pass.

Lower is a code generation pass that turns high level IR nodes into
lower level one, for example substituting comparisons to RuntimeNodes
and inlining finally blocks.

Lower is also responsible for determining control flow information
like end points.


* attr

The lowering annotates a FunctionNode with symbols for each identifier
and transforms high level constructs into lower level ones, that the
CodeGenerator consumes.

Lower logging typically outputs things like post pass actions,
insertions of casts because symbol types have been changed and type
specialization information. Currently very little info is generated by
this logger. This will probably change.


* finalize

This --log=finalize log option outputs information for type finalization,
the third tier of the compiler. This means things like placement of 
specialized scope nodes or explicit conversions. 


* fields

The --log=fields option (at info level) is equivalent to setting the
system variable "nashorn.fields.debug" to true. At the info level it
will only show info about type assumptions that were invalidated. If
the level is set to finest, it will also trace every AccessorProperty
getter and setter in the program, show arguments, return values
etc. It will also show the internal representation of respective field
(Object in the normal case, unless running with the dual field
representation)