| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
|
|
Lint
|
|
|
|
The accuracy is greatly improved by using the C preprocessor to scan C
sources for presym. C preprocessor can perfectly interpret all comments and
preprocessor directives, so it can detect all symbols defined, for example
`mrbgems/mruby-socket/src/const.cstub`.
Also, as described later, this change will greatly improve the accuracy of
presym detection from Ruby sources.
## Result
The number of lines in the `presym` file for all gems is as follows:
```console
Previous: 999 (false positive = 89, undetected = 297)
New: 1207
```
## Build process
The new build process (with presym) is as follows:
1. Build `mrbc` without presym (more on building without presym later).
2. Compile Ruby sources to C struct format with the `mrbc` created in
step 1, and create` mrblib.c` and `gem_init.c`. Note that the symbols
in the created files are output as `MRB_SYM` family macros or
`mrb_intern_lit` instead of IDs (details will be described later).
3. C preprocessor processes C sources including the created files of
step 2 and outputs them as `.i` files. In these files, for example,
`MRB_IVSYM(foo)` is converted to `<@! "@" "foo" !@>` and
`mrb_define_module(mrb, "Foo")` is converted to `<@! "Foo" !@>`.
4. Scan the files created in step 3 and create `presym` and` presym.inc`
files.
The files created in step 2 should output all static symbols defined in Ruby
sources, including local variables, so we can detect all presyms by just
scanning C sources without scanning Ruby sources directly.
Further, by this process, the files to be scanned becomes the same as the
files to be compiled, so that there is no excess or deficiency.
## Related changes
The following changes have been made in relation to realizing this feature.
### Allow build without presym
It enables build without presym to achieve the "Build process: 1". This
incorporates #5202, see its issue for details.
Note that when presym is enabled, even adding a local variable to a Ruby
source may change contents of presym and require recompilation of almost
all C sources. This is inconvenient, especially during trial and error in
development, but this feature is also useful because it does not cause
this problem if presym is disabled.
### Automatically create build target for `mrbc` without presym
The `mrbc` used in the "Build process: 1" will be built by automatically
creating a build target for it. The build name is `SOURCE_BUILD_NAME/mrbc`.
### Constantize output of C struct format by `mrbc`
To realizing the "Build process: 2", as mentioned above, symbol IDs are not
output directly in C struct format output by `mrbc`. As a result, the output
becomes constant regardless of the state of presym at the time of `mrbc`
build, and it is possible to detect symbols of Ruby sources in the same way
as other C sources.
Note that `mrb_intern_lit` is used for symbols that do not become presym,
but in this state, the corresponding element in the symbol array cannot be
statically initialized, so it is initialized at run time (therefore, in this
case, the `const` qualifier is not added to the symbol array).
### Specify arbitrary `mrbc` file
To realizing the "Build process: 2", enabled to specify `mrbc` created by
another build target or pre-built` mrbc`. Use `MRuby::Build#mrbcfile =` to
specify it explicitly. You can omit the "Build process: 1" by specifying
pre-built `mrbc`, and you can always use an optimized build to compile Ruby
sources faster. I think changes that affect the output of `mrbc` are rare,
so in many cases it helps to improve efficiency.
With presym, the build will be a little slower due to more build steps, but
this feature will improve it a bit.
### Create presym files for each build target
This feature was proposed at #5194 and merged once, but was reverted in
5c205e6e due to problems especially with cross-compilation. It has been
introduced again because this change solves the problem.
The presym files will be created below.
* `build/NAME/presym`
* `build/NAME/include/mruby/presym.inc`
### Other changes
* Because presym detection accuracy is greatly improved as mentioned above,
`MRuby::Gem::Specification#cdump?` is set to true by default, and
`disable_cdump` is added instead of `enable_cdump`. Also, support for gem
specific presym files has been discontinued (https://github.com/mruby/mruby/issues/5151#issuecomment-730967232).
* Previously, `mrbc` was automatically created for the `host` build, but it
will not be created if the build target for `mrbc` mentioned above is
automatically created. At this time, `mrbc` file of the `mrbc` build is
copied to` bin/`.
* Two types of `.d` files will be created, `.o.d` and `.i.d`. oThis is
because if `.i` depends on `presym.inc`, the dependency will circulate, so
the `.d` file cannot be shared.
* Changed file created with `enable_cxx_exception` to `X-cxx.cxx` from
`X.cxx` to use the mruby standard Rake rule.
### Note
Almost all C sources will need to be recompiled if there are any changes to
`persym.inc` (if not recompiled properly, it will often result in run-time
error). If `gcc` toolchain is used, dependencies are resolved by the `.d`
file, so it become automatically recompile target, but if not (e.g. MSVC),
it is necessary to manually make it recompile target.
Also, even if `gcc` toolchain is used, it may not become recompile target if
external gems does not use the mruby standard Rake rule. In particular, if
the standard rule is overwritten, such as
https://github.com/mruby/mruby/pull/5112/files, `.d` file will not be read,
so be careful.
|
|
They used to be `size_t`, `uint32_t` and `ptrdiff_t`. Now all of them
made to be `uint32_t`.
|
|
|
|
Due to the above changes, it may not work with the existing build
configurations in cross-compilation (even if we can build without presym),
therefore revert it once (ref https://github.com/mruby/mruby/pull/5202#issuecomment-735412643). Sorry for the lack of consideration.
|
|
Follow commit 7150c6753933f12a2ba63769fb7b3a44cfcddd3d .
|
|
Create presym files for each build target
|
|
|
|
Previously, presym files were always created in `build/{presym,presym.inc}`.
However, this constraint is inconvenient because it is common to use
multiple build configurations and build targets in a single mruby tree.
Therefore, change to create presym file for each build target.
|
|
Jump target address is `operand (16bit)` + `address of next instruction`.
In addition, `ilen` was made `uint32_t` so that `iseq` length limitation
of 65536 is removed. Only jump target address should be within signed
16bit (-32768 .. 32767).
|
|
|
|
|
|
|
|
|
|
If it gives clang-11.0 `-Wembedded-directive`, a warning will be reported in `include/mruby/boxing_nan.h`.
```
include/mruby/boxing_nan.h:48:2: warning: embedding a directive within macro arguments has undefined behavior [-Wembedded-directive]
include/mruby/boxing_nan.h:52:2: warning: embedding a directive within macro arguments has undefined behavior [-Wembedded-directive]
```
The cause of this is #5117.
ref. e993b83c509912f2d90ffece32c969a642f5df01
|
|
|
|
| Previous Name | New Name |
|------------------------------|-------------------------|
| MRB_ENABLE_ALL_SYMBOLS | MRB_USE_ALL_SYMBOLS |
| MRB_ENABLE_SYMBOLL_ALL | MRB_USE_ALL_SYMBOLS |
| MRB_ENABLE_CXX_ABI | MRB_USE_CXX_ABI |
| MRB_ENABLE_CXX_EXCEPTION | MRB_USE_CXX_EXCEPTION |
| MRB_ENABLE_DEBUG_HOOK | MRB_USE_DEBUG_HOOK |
| MRB_DISABLE_DIRECT_THREADING | MRB_NO_DIRECT_THREADING |
| MRB_DISABLE_STDIO | MRB_NO_STDIO |
| ENABLE_LINENOISE | MRB_USE_LINENOISE |
| ENABLE_READLINE | MRB_USE_READLINE |
| DISABLE_MIRB_UNDERSCORE | MRB_NO_MIRB_UNDERSCORE |
| DISABLE_GEMS | MRB_NO_GEMS |
* `MRB_ENABLE_SYMBOLL_ALL` seems to be a typo, so it is fixed.
* `MRB_` prefix is added to those without.
* The previous names can also be used for compatibility.
|
|
It is not decides by the extension.
In order to be recognized as a `.mrb` file, the following three points must be satisfied:
- File starts with "RITE"
- At least `sizeof(struct rite_binary_header)` bytes can be read
- `NUL` is included in the first 64 bytes of the file
If these are not met, it is judged as a text file and it is processed as a Ruby script.
The `bin/mruby -b` switch is still available which treats the given file as a `.mrb` file.
New `MRB_API` function:
- `include/mruby/compile.h` and `mrbgems/mruby-compiler/core/parse.y`
- `mrb_load_detect_file_cxt()` (remove with `MRB_DISABLE_STDIO`)
NOTE:
- Even script files now always open in binary mode for `bin/mruby`.
The `\r\n` is handled by the `nextc()` function already, so there is no problem even on Windows.
- The `nextc0()` function in `mrbgems/mruby-compiler/core/parse.y` can now specify a string buffer and a file pointer at the same time.
In this case, get it from the string buffer first.
This patch includes modifies by comment of https://github.com/mruby/mruby/pull/5157.
|
|
Fix documents for `mrb_ary_splice()` [ci skip]
|
|
|
|
|
|
|
|
I misunderstand the meaning of #4483. Sorry.
|
|
- Remove `mrb_ssize`
- Fix `MRB_FIXNUM_{MIN,MAX}` to 32 bits on `MRB_NAN_BOXING`
|
|
- `MRB_64BIT` and `MRB_INT32`
- `MRB_32BIT` and `MRB_INT64`
|
|
`mrb_bool` on LLP64 environment seems to become false if, for example,
`mrb_value` is a pointer whose lower 32 bits are 0.
|
|
|
|
|
|
|
|
|
|
Fix overflow detection in integer operations with `MRB_WORD_BOXING`.
This bug made `1073741824 == 1073741824+0` to be `false` on 32bit
platforms.
|
|
The `MRB_API` function `mrb_mt_foreach()` previously used the private structure `struct mt_elem`.
Therefore, use `mrb_method_t` instead.
|
|
* The term `hash_table` can be misleading because the return value of this
function includes memory usage of entire `Hash` object, including not only
hash table part but also entry list part, etc.
* This function takes a `Hash` object as a receiver and is defined in
`src/hash.c`, so it is natural to have a `mrb_hash_` prefix.
|
|
Change name and usage of presym macros
|
|
To be also able to build mruby without presym in the future. However,
`MRB_QSYM` has been removed and changed as follows:
### Example
| Type | Symbol | Previous Style | New Style |
|---------------------------|--------|------------------|----------------|
| Operator | & | MRB_QSYM(and) | MRB_OPSYM(and) |
| Class Variable | @@foo | MRB_QSYM(00_foo) | MRB_CVSYM(foo) |
| Instance Variable | @foo | MRB_QSYM(0_foo) | MRB_IVSYM(foo) |
| Method with Bang | foo! | MRB_QSYM(foo_b) | MRB_SYM_B(foo) |
| Method with Question mark | foo? | MRB_QSYM(foo_p) | MRB_SYM_Q(foo) |
| Mmethod with Equal | foo= | MRB_QSYM(foo_e) | MRB_SYM_E(foo) |
This change makes it possible to define, for example, `MRB_IVSYM(foo)` as
`mrb_intern_lit(mrb, "@" "foo")`, which is useful if we support building
without presym in the future.
|
|
|
|
Reduce memory usage of Hash object
|
|
|
|
## Implementation Summary
* Change entry list from segmented list to flat array.
* Change value of hash bucket from pointer to entry to index of entry list,
and represent it by variable length bits according to capacity of hash
buckets.
* Store management information about entry list and hash table to `struct
RHash` as much as possible.
## Benchmark Summary
Only the results of typical situations on 64-bit Word-boxing are present
here. For more detailed information, including consideration, see below
(although most of the body is written in Japanese).
* https://shuujii.github.io/mruby-hash-benchmark
### Memory Usage
Lower value is better.
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 344B | 256B | 0.74419x |
| 40 | 1,464B | 840B | 0.57377x |
| 200 | 8,056B | 3,784B | 0.46971x |
| 500 | 17,169B | 9,944B | 0.57949x |
### Performance
Higher value is better.
#### `mrb_hash_set`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 1.41847M i/s | 1.36004M i/s | 0.95881x |
| 40 | 0.39224M i/s | 0.31888M i/s | 0.81296x |
| 200 | 0.03780M i/s | 0.04290M i/s | 1.13494x |
| 500 | 0.01225M i/s | 0.01314M i/s | 1.07275x |
#### `mrb_hash_get`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 26.05920M i/s | 30.19543M i/s | 1.15872x |
| 40 | 44.26420M i/s | 32.75781M i/s | 0.74005x |
| 200 | 44.55171M i/s | 31.56926M i/s | 0.70860x |
| 500 | 39.19250M i/s | 29.73806M i/s | 0.75877x |
#### `mrb_hash_each`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 25.11964M i/s | 30.34167M i/s | 1.20789x |
| 40 | 11.74253M i/s | 13.25539M i/s | 1.12884x |
| 200 | 2.01133M i/s | 2.97214M i/s | 1.47770x |
| 500 | 0.87411M i/s | 1.21178M i/s | 1.38631x |
#### `Hash#[]=`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 0.50095M i/s | 0.56490M i/s | 1.12764x |
| 40 | 0.19132M i/s | 0.18392M i/s | 0.96129x |
| 200 | 0.03624M i/s | 0.03256M i/s | 0.89860x |
| 500 | 0.01527M i/s | 0.01236M i/s | 0.80935x |
#### `Hash#[]`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 11.53211M i/s | 12.78806M i/s | 1.10891x |
| 40 | 15.26920M i/s | 13.37529M i/s | 0.87596x |
| 200 | 15.28550M i/s | 13.36410M i/s | 0.87430x |
| 500 | 14.57695M i/s | 12.75388M i/s | 0.87494x |
#### `Hash#each`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 0.30462M i/s | 0.27080M i/s | 0.88898x |
| 40 | 0.12912M i/s | 0.11704M i/s | 0.90642x |
| 200 | 0.02638M i/s | 0.02402M i/s | 0.91071x |
| 500 | 0.01066M i/s | 0.00959M i/s | 0.89953x |
#### `Hash#delete`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 7.84167M i/s | 6.96419M i/s | 0.88810x |
| 40 | 6.91292M i/s | 7.41427M i/s | 1.07252x |
| 200 | 3.75952M i/s | 7.32080M i/s | 1.94727x |
| 500 | 2.10754M i/s | 7.05963M i/s | 3.34970x |
#### `Hash#shift`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 14.66444M i/s | 13.18876M i/s | 0.89937x |
| 40 | 11.95124M i/s | 11.10420M i/s | 0.92913x |
| 200 | 5.53681M i/s | 7.88155M i/s | 1.42348x |
| 500 | 2.96728M i/s | 5.40405M i/s | 1.82121x |
#### `Hash#dup`
| Hash Size | Baseline | New | Factor |
|----------:|--------------:|--------------:|-----------:|
| 16 | 0.15063M i/s | 5.37889M i/s | 35.71024x |
| 40 | 0.06515M i/s | 3.38196M i/s | 51.91279x |
| 200 | 0.01359M i/s | 1.46538M i/s | 107.84056x |
| 500 | 0.00559M i/s | 0.75411M i/s | 134.88057x |
### Binary Size
Lower value is better.
| File | Baseline | New | Factor |
|:-----------|--------------:|--------------:|----------:|
| mruby | 730,408B | 734,176B | 1.00519x |
| libmruby.a | 1,068,134B | 1,072,846B | 1.00441x |
## Other Fixes
The following issues have also been fixed in the parts where there was some
change this time.
* [Heap use-after-free in `Hash#value?`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-heap-use-after-free-in-hash-value-md)
* [Heap use-after-free in `ht_hash_equal`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-heap-use-after-free-in-ht_hash_equal-md)
* [Heap use-after-free in `ht_hash_func`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-heap-use-after-free-in-ht_hash_func-md)
* [Heap use-after-free in `mrb_hash_merge`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-heap-use-after-free-in-mrb_hash_merge-md)
* [Self-replacement does not work for `Hash#replace`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-self-replacement-does-not-work-for-hash-replace-md)
* [Repeated deletes and inserts increase memory usage of `Hash`](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-repeated-deletes-and-inserts-increase-memory-usage-of-hash-md)
* [`Hash#rehash` does not reindex completely](https://gist.github.com/shuujii/30e4fcd5844a4112a0ecd4a5b3483101#file-hash-rehash-does-not-reindex-completely-md)
* `mrb_hash_delete_key` does not cause an error for frozen object
* `mrb_hash_new_capa` does not allocate required space first
* [`mrb_os_memsize_of_hash_table` result is incorrect](https://github.com/mruby/mruby/pull/5032#discussion_r457994075)
|
|
Define `mrb_value` in struct for Word/NaN boxing
|
|
`SET_CPTR_VALUE()` requires the `p` field on 32-bit CPU mode.
|
|
This is to make it possible to distinguish between `mrb_value` and `mrb_sym` in C++ "Function overloading" and C11 "Generic selection".
|
|
That loads 32 bit integer bypassing pool access.
|
|
|
|
New instructions:
* OP_LOADL16
* OP_LOADSYM16
* OP_STRING16
Size of pools, symbols are `int16_t` but offset representation in the
bytecode was 8 bits. Size of child `irep` array is `int16_t`, too but
this change does not address it.
|
|
|
