Property names were always required a string reference which consumed
a large amount of memory for arrays. This patch reduces this consumption
by directly storing the value part of certain strings.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
This change allows easier access to array length which improves
the performance of inserting new items into an array and there
is no need to allocate length strings anymore. The trade-of is
that array length cannot be cached anymore.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
This patch reduces memory consumption for strings such as "0" or "123"
by 8 bytes and "4294967295" by 16 bytes. The hash computation is changed
for using the lower 16 bits for these strings which is much faster than
converting the value to string first and compute the hash. The trade-of
is a small overhead when strings are created or concatenated.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
Legacy implementation made incorrect assumptions on how many bits
of useful information is returned by libc's `rand()`. If `RAND_MAX`
had more than 16 useful bits, it assumed that `rand()` could return
32 useful random bits. However, e.g., jerry-libc's `RAND_MAX` is
`0x7fffffff`, which denotes 31 useful bits only. The consequence
was that `Math.random()` covered only the lower half of the
standard-mandated `[0,1)` interval.
This path fixes the error and always uses the exact value of
`RAND_MAX` to compute the random value, which will thus fully cover
`[0,1)`.
Fixes#1414
JerryScript-DCO-1.0-Signed-off-by: Akos Kiss akiss@inf.u-szeged.hu
Some test cases were failed with enabled show-opcodes build option.
Example source:
"98765".replace(76)
print ("\ud801\udc00".replace("\ud801", "#") === "#\udc00");
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
The new code does not use value collections which reduces the argument array size by half.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
Class and value internal properties are always exists for primitive
types (e.g. Boolean, Regex) so they can be stored right after the
object. This improve property access (since internal properties are
searched by a slow linear algorithm) and reduces memory consumption,
since only 8 byte is allocated for these two properties instead of
16 which is the size of a property pair.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
"ecma_builtin_number_prototype_helper_round" is used by following
functions:
* ecma_builtin_number_prototype_object_to_fixed
* ecma_builtin_number_prototype_object_to_exponential
* ecma_builtin_number_prototype_object_to_precision
The current implementation does not support currying numbers and will
produce illegal digit for some test cases. For example, the result of
"0.95.toFixed(1)" is "0.:".
This patch fixs the issue, however the implementation of "toFixed" is
still problematic, at least not meet section 15.7.4.5 8.a of the
specification. The related test case is:
* assert((0.995).toFixed(2) === "0.99");
* assert((9.995).toFixed(2) === "9.99");
Reference:
* http://www.ecma-international.org/ecma-262/5.1/#sec-15.7.4.5
JerryScript-DCO-1.0-Signed-off-by: Yanhui Shen shen.elf@gmail.com
Properties are changed to a type and value pair instead of a pointer to an internal
representation. Functions such as ecma_op_object_get_[own_]property do not
return with property pointers anymore.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
The ecma_op_object_get_[own_]property calls should be phased out from
the project eventually and virtual properties should be introduced instead.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
Don't increment the current position in case of a continuation byte.
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
OS X build regularly reports some 39 warnings falling in 3 major
categories:
* "static function '...' is used in an inline function with
external linkage [-Wstatic-in-inline]": Some semantics around
`inline` have changed between C89 and C99, and gcc and clang seem
to disagree on how strict they should be about them. Solution
chosen is to use `-Wnostatic-in-inline` command line option for
clang.
* "implicit conversion turns floating-point number into integer:
'double' to 'bool' [-Wfloat-conversion]": `if (fmod (..., ...))`
was used at different places, which is not nice anyway, thus the
return value is compared explicitly against `ECMA_NUMBER_ZERO`.
* "format string is not a string literal [-Wformat-nonliteral]":
Console and log port I/O functions have a printf-like interface,
and the default implementations actually pass both format string
and the remaining arguments to a vfprintf. However, clang is
strict about the format string parameter of vfprintf and expects
a literal there. By annotating the port I/O functions with
`__attribute__ ((format (printf, ..., ...)))`, clang will check
the format string being a literal string earlier, when the port
functions are called, and will not complain within them when
vfprintf is called.
(Actually, this has revealed an incorrect format string, which
has been fixed as well.)
(There were also some single conversion errors not listed above.)
The patch was tested on OS X (where all warnings disappeared), but
it should help clang compilation on other OS's as well.
JerryScript-DCO-1.0-Signed-off-by: Akos Kiss akiss@inf.u-szeged.hu
Removing a lot of ECMA_PROPERTY_VALUE_PTR macro calls. The only drawback
is free callbacks for native objects cannot be deleted anymore. Redefining
a free callback is a rare case, so this trade-of is acceptable.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
The Math.pow implementation relies on libm's pow. However, the ISO C
and ES5.1 standards differ on pow:
* `x ** NAN` is NAN in ES but `+1 ** y` is 1 in C
* `+-1 ** +-INF` is NAN in ES but 1 in C
This patch:
* Modifies the Math.pow implementation to handle the special cases
instead calling pow.
* Adds a test case to jerry-test-suite as it did not test
`Math.pow(1,NaN)`.
* Fixes jerry-libm's pow, as it was not standard conforming, which
helped hiding the error in Math.pow.
* Updates the unit test for libm.
JerryScript-DCO-1.0-Signed-off-by: Akos Kiss akiss@inf.u-szeged.hu
Until now, jerry had 3 different assert-like routines:
`jerry_assert_fail`, `jerry_unreachable`, and `jerry_unimplemented`,
and 3 corresponding macros (`JERRY_ASSERT`, `JERRY_UNREACHABLE`,
and `JERRY_UNIMPLEMENTED`). They had some irregularities, namely:
* All of them had a string parameter, although `jerry_unreachable`
never got anything there but NULL.
* Both `jerry_unreachable` and `jerry_unimplemented` checked its
string parameter for NULL, although it was always NULL for the
first one and never NULL for the second.
* `jerry_unreachable` is just a regular assert with a fixed error
message (i.e., control should not have got here), however, the
expansion of its corresponding macro in debug and release modes
differs from the behaviour of `JERRY_ASSERT`: `JERRY_ASSERT` is
a no-op in release, however, `JERRY_UNREACHABLE` was triggering
a crash even there.
* Moreover, `JERRY_UNIMPLEMENTED` was almost never used anymore but
in a few places (where often an `#ifdef` selected between
`JERRY_UNIMPLEMENTED` and `JERRY_UNREACHABLE`).
Because of the above, this patch makes the following changes:
* Drops `JERRY_UNIMPLEMENTED` completely and whereever it was still
used, replaces it with `JERRY_UNREACHABLE`. As a consequence, the
`jerry_unimplemented` function and the `ERR_UNIMPLEMENTED_CASE`
fatal error code are also removed.
* Makes `JERRY_UNREACHABLE` expand to no-op in release builds.
(Actually, to `__builtin_unreachable ()` to avoid warnings.) As
a consequence, makes both `jerry_assert_fail` and
`jerry_unreachable` be guarded by `#ifndef JERRY_NDEBUG`. Also,
changes `jerry_unreachable` not to expect a string parameter.
* Rewrites `TEST_ASSERT` not to rely on `jerry_assert_fail` as
`TEST_ASSERT` has to work in release builds as well. This also
allows changing the error message not to mention "ICE", which
would misleadingly suggest an assert within the engine, but
"TEST" instead.
As a side-effect of the cleanup, some refactorings happened in
jrt.h:
* Removed the definition of the unnecessary `__extension__` macro.
* Re-used `JERRY_UNUSED` and `unlikely` where possible.
* Moved some parts of the file around.
* Fixed some comments (`/**` should only be used for the docstring
of a single entity, for groups header comments, the regular `/*`
should be used).
JerryScript-DCO-1.0-Signed-off-by: Akos Kiss akiss@inf.u-szeged.hu
The standard doesn't defines ECMAScript Compact Profile as a subset of Ecma-262 Edition 5.1.
Profile modes can be added easily like the minimal profile if required.
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
Zero out all globals (and remove unnecessary init() functions).
Move snapshot globals to a temporary stack variable.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
The "length" property name is the most frequently used built-in string
and also frequently created by various hot-paths. New functions are
added to improve the speed of the "length" string creation.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
Several internal properties are removed and directly stored as
part of the object. Faster built-in and JS function processing.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
* Make constants static in `ecma_number_to_utf8_string`.
* Make `ecma_number_to_utf8_string` use early returns, and rewrite
its self-recursion in case of negative numbers.
* Make the stringification of decimal exponent in
`ecma_number_to_utf8_string` use `ecma_uint32_to_utf8_string`.
* Changed ERROL0 dtoa implementation to use the `double` type
instead of `ecma_number_t`. Thus, even is `ecma_number_t` is 32
bit wide, the algorithm works the same.
* Changed `ecma_number_to_decimal` to use the ERROL0 dtoa algorithm
for 32-bit floats as well.
* Changed `ecma_number_to_decimal` to generate the decimal string
representation of the mantissa instead of an `uint64_t` number.
* Changed `ecma_number_to_utf8_string` to make use of the already
available string representation of the mantissa, generated now by
`ecma_number_to_decimal`.
* Changed `ecma_number_to_utf8_string` not to use static arrays and
variables for digit, "e", etc. generation.
* Changed all `Number.prototype.toXXX` implementations and the
`ecma_builtin_number_prototype_helper_round` helper to make use
of the already available string representation of the mantissa,
generated now by `ecma_number_to_decimal`.
* Factored out the common stringification parts of all
`Number.prototype.toXXX` implementations into a new helper
`ecma_builtin_number_prototype_helper_to_string`.
JerryScript-DCO-1.0-Signed-off-by: Akos Kiss akiss@inf.u-szeged.hu
Changed 'ecma_copy_or_ref_ecma_string' to 'ecma_ref_ecma_string'. It does
not copy the string if the maximum number of reference counter is reached,
but bails out with an error like the 'ecma_ref_object' function does.
JerryScript-DCO-1.0-Signed-off-by: László Langó llango.u-szeged@partner.samsung.com
Correct behaviour of step 8, 15.9.3.1, ECMA-262 v5.1.
- Change ToInt32(y) to ToInteger(y).
- If ToInteger(y) is not between 0 and 99 then yr = y
Related issue: #1071
JerryScript-DCO-1.0-Signed-off-by: Hanjoung Lee hanjoung.lee@samsung.com
JerryScript-DCO-1.0-Signed-off-by: Hanjoung Lee hanjoung.lee@samsung.com
Parts:
* Rename ecma_string_to_utf8_string() to ecma_string_copy_to_utf8_buffer.
* Introduce ecma_string_to_utf8_bytes(), which wraps the usual 'function call-assertion' pair,
and check strict equality of size of the string and the buffer.
JerryScript-DCO-1.0-Signed-off-by: Zsolt Borbély zsborbely.u-szeged@partner.samsung.com
The ecma_op_to_boolean return value is changed to bool for faster
evaluation, and no need to swap operandos of relational compare.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
the three boolean arguments of ecma_create_named_data_property and the
two boolean arguments of ecma_create_named_accessor_property are combined
into one uint8_t argument. On ARM-32 it is preferred to have less than
four arguments, since these arguments can be passed in registers.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
The 'mem_' prefix is too general, so it might clash with
symbols in other libraries. Renamed the directory, file,
funtion and type names.
Related issue: #1052
JerryScript-DCO-1.0-Signed-off-by: László Langó llango.u-szeged@partner.samsung.com
String, Boolean) to use ecma-values. When the value of a Number object is
a small integer number, this change reduces the memory consumption, since no
double is allocated.
JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
This modificiation affects those conditions which check that
a value can be represented with a smaller type.
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
Those internal properties which have a primitive value field or where the property value type
is an ecma_object_t ptr were checked earlier.
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
This change affects only those internal properties which have a primitive value field.
JerryScript-DCO-1.0-Signed-off-by: Robert Sipka rsipka.uszeged@partner.samsung.com
