mirror of https://github.com/svaarala/duktape.git
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
112 lines
3.7 KiB
112 lines
3.7 KiB
==================================
|
|
Performance sensitive environments
|
|
==================================
|
|
|
|
Overview
|
|
========
|
|
|
|
This document describes suggested feature options for optimizing Duktape
|
|
performance for performance sensitive environments.
|
|
|
|
The following genconfig option file template enables most performance
|
|
related options: ``config/examples/performance_sensitive.yaml``.
|
|
|
|
Compiler optimization level
|
|
===========================
|
|
|
|
Size optimization using ``-Os`` is a good default when performance is
|
|
not critical. However, it's not ideal when performance matters for
|
|
several reasons:
|
|
|
|
* Although ``-Os`` optimized code performs reasonably well, even
|
|
``-O2`` will yield significantly better results.
|
|
|
|
* Code performance with ``-Os`` can vary a great deal even when source
|
|
code changes are innocent. It's not uncommon for some performance
|
|
test result to change +/- 10-30% with unrelated changes. Presumably
|
|
this is caused by changes in code alignment etc.
|
|
|
|
Because of this, ``-Os`` is definitely a bad idea for measuring
|
|
performance.
|
|
|
|
* Overall suggestion is to use ``-O2`` and try ``-O3`` if the end result
|
|
is better. Note that ``-O3`` is not always better because the code is
|
|
larger and may not fit in caches as well as with ``-O2``.
|
|
|
|
Profile guided optimization (PGO)
|
|
=================================
|
|
|
|
Duktape source files contain some performance attributes like forced inline
|
|
forced noinline, and hot/cold attributes.
|
|
|
|
A better alternative is to use profile guided optimization (PGO) which is
|
|
highly recommended for performance sensitive environments. For example,
|
|
GCC -O2 with PGO can be around 20% faster than GCC -O2 without PGO.
|
|
|
|
See for example the following:
|
|
|
|
* http://stackoverflow.com/questions/13881292/gcc-profile-guided-optimization-pgo
|
|
|
|
* https://msdn.microsoft.com/en-us/library/e7k32f4k.aspx
|
|
|
|
With GCC PGO is relatively simple:
|
|
|
|
* Use ``-fprofile-generate`` to compile Duktape and your application.
|
|
|
|
* Execute the result with representative (this is important) source files.
|
|
|
|
* Use ``-fprofile-use`` to recompile Duktape and your application.
|
|
|
|
Suggested feature options
|
|
=========================
|
|
|
|
* On some platforms ``setjmp/longjmp`` store the signal mask and may be
|
|
much slower than alternative like ``_setjmp/_longjmp`` or
|
|
``sigsetjmp/siglongjmp``:
|
|
|
|
- Check the current provider from ``duk_config.h`` or ``config/platforms/``
|
|
header snippets.
|
|
|
|
- Edit ``DUK_SETJMP``, ``DUK_LONGJMP``, and ``DUK_JMPBUF_TYPE`` to change
|
|
the setjmp provider.
|
|
|
|
- On some platforms (e.g. OSX/iPhone) Duktape will automatically use
|
|
``_setjmp()``.
|
|
|
|
* Consider enabling "fastints":
|
|
|
|
- ``#define DUK_USE_FASTINT``
|
|
|
|
Fastints are often useful on platforms with soft floats, but they can also
|
|
speed up execution on some hard float platforms (even on x64). The benefit
|
|
(or penalty) depends on the kind of Ecmascript code executed, e.g. code
|
|
heavy on integer loops benefits.
|
|
|
|
* Enable specific fast paths:
|
|
|
|
- ``#define DUK_USE_JSON_STRINGIFY_FASTPATH``
|
|
|
|
- ``#define DUK_USE_JSON_QUOTESTRING_FASTPATH``
|
|
|
|
- ``#define DUK_USE_JSON_DECSTRING_FASTPATH``
|
|
|
|
- ``#define DUK_USE_JSON_DECNUMBER_FASTPATH``
|
|
|
|
- ``#define DUK_USE_JSON_EATWHITE_FASTPATH``
|
|
|
|
* If you don't need debugging support or execution timeout support, ensure
|
|
the following are **not enabled**:
|
|
|
|
- ``#define DUK_USE_INTERRUPT_COUNTER``
|
|
|
|
- ``#define DUK_USE_DEBUGGER_SUPPORT``
|
|
|
|
Especially interrupt counter option will have a measurable performance
|
|
impact because it includes code executed for every bytecode instruction
|
|
dispatch.
|
|
|
|
* Disable safety check for value stack resizing so that if calling code
|
|
fails to ``duk_check_stack()`` value stack, the result is memory unsafe
|
|
behavior rather than an explicit error, but stack operations are faster:
|
|
|
|
- ``#undef DUK_USE_VALSTACK_UNSAFE``
|
|
|