| Age | Commit message (Collapse) | Author |
|---|
|
These instructions call methods of the receiver.
|
|
* `R(a)` -> `R[a]`
* `Pool(a)` -> `Pool[a]`
* `Syms(a)` -> `Syms[a]`
* `Irep(a)` -> `Irep[a]`
|
|
|
|
take-cheeze-exc_mesg
|
|
The Difference
Since Ruby1.9, the keyword arguments were emulated by Ruby using the hash
object at the bottom of the arguments. But we have gradually moved toward
keyword arguments separated from normal (positinal) arguments.
At the same time, we value compatibility, so that Ruby3.0 keyword
arguments are somewhat compromise. Basically, keyword arguments are
separated from positional arguments, except when the method does not
take any formal keyword arguments, given keyword arguments (packed
in the hash object) are considered as the last argument.
And we also allow non symbol keys in the keyword arguments. In that
case, those keys are just passed in the `**` hash (or raise
`ArgumentError` for unknown keys).
The Instruction Changes
We have changed `OP_SEND` instruction. `OP_SEND` instruction used to
take 3 operands, the register, the symbol, the number of (positional)
arguments. The meaning of the third operand has been changed. It is now
considered as `n|(nk<<4)`, where `n` is the number of positional
arguments, and `nk` is the number of keyword arguments, both occupies
4 bits in the operand.
The number `15` in both `n` and `nk` means variable sized arguments are
packed in the object. Positional arguments will be packed in the array,
and keyword arguments will be packed in the hash object. That means
arguments more than 14 values are always packed in the object.
Arguments information for other instructions (`OP_SENDB` and `OP_SUPER`)
are also changed. It works as the third operand of `OP_SEND`. the
difference between `OP_SEND` and `OP_SENDB` is just trivial. It assigns
`nil` to the block hidden arguments (right after arguments).
The instruction `OP_SENDV` and `OP_SENDVB` are removed. Those
instructions are replaced by `OP_SEND` and `OP_SENDB` respectively with
the `15` (variable sized) argument information.
Calling Convention
When calling a method, the stack elements shall be in the order of the
receiver of the method, positional arguments, keyword arguments and the
block argument. If the number of positional or keyword arugument (`n` or
`nk`) is zero, corresponding arguments will be empty. So when `n=0` and
`nk=0` the stack layout (from bottom to top) will be:
+-----------------------+
| recv | block (or nil) |
+-----------------------+
The last elements `block` should be explicitly filled before `OP_SEND`
or assigned to `nil` by `OP_SENDB` internally. In other words, the
following have exactly same behavior:
OP_SENDB clears `block` implicitly:
```
OP_SENDB reg sym 0
```
OP_SEND clears `block` implicitly:
```
OP_LOADNIL R2
OP_SEND R2 sym 0
```
When calling a method with only positional arguments (n=0..14) without
keyword arguments, the stack layout will be like following:
+--------------------------------------------+
| recv | arg1 | ... | arg_n | block (or nil) |
+--------------------------------------------+
When calling a method with arguments packed in the array (n=15) which
means argument splat (*) is used in the actual arguments, or more than
14 arguments are passed the stack layout will be like following:
+-------------------------------+
| recv | array | block (or nil) |
+-------------------------------+
The number of the actual arguments is determined by the length of the
argument array.
When keyword arguments are given (nk>0), keyword arguments are passed
between positional arguments and the block argument. For example, when
we pass one positional argument `1` and one keyword argument `a: 2`,
the stack layout will be like:
+------------------------------------+
| recv | 1 | :a | 2 | block (or nil) |
+------------------------------------+
Note that keyword arguments consume `2*nk` elements in the stack when
`nk=0..14` (unpacked).
When calling a method with keyword arguments packed in the hash object
(nk=15) which means keyword argument splat (**) is used or more than
14 keyword arguments in the actual arguments, the stack layout will
be like:
+------------------------------+
| recv | hash | block (or nil) |
+------------------------------+
Note for mruby/c
When mruby/c authors try to support new keyword arguments, they need
to handle the new meaning of the argument information operand. If they
choose not to support keyword arguments in mruby/c, it just raise
error when `nk` (taken by `(c>>4)&0xf`) is not zero. And combine
`OP_SENDV` behavior with `OP_SEND` when `n` is `15`.
If they want to support keyword arguments seriously, contact me at
<[email protected]> or `@yukihiro_matz`. I can help you.
|
|
Which represent `obj[int]` and `obj[int]=val` respectively where `obj`
is either `string`, `array` or `hash`, so that index access could be
faster. When `obj` is not assumed type or `R(a+1)` is not integer, the
instructions fallback to method calls.
|
|
|
|
Rename `MRB_WORDBOX_USE_HEAP_FLOAT` to `MRB_WORDBOX_NO_FLOAT_TRUNCATE'.
|
|
- `OP_ARYPUSH` now takes operand for the number of pushing elements
- the code generator consume the stack no more than `64` for `mruby/c`
|
|
Add n elements at once. Reduces instructions for huge array
initialization. In addition, `gen_value` function in `codegen.c` was
refactored and clarified.
|
|
|
|
`acc` was used as an index of the receiver (if positive), or a flag for
methods implemented in C. We replace `regs[ci->acc]` by `ci[1].stack[0]`.
And renamed `acc` (originally meant accumulator position) to `cci`
(means callinfo for C implemented method).
|
|
It generates a symbol by interning from the pool string.
|
|
- `MRB_USE_ETEXT_RO_DATA_P`: use `etext` and `edata`
- `MRB_NO_DEFAULT_RO_DATA_P`: not use the default `mrb_ro_data_p()`
|
|
|
|
|
|
Consistent naming: `integer` to represent integer packed in `mrb_value`
instead of `int`.
|
|
Consistent naming: `integer` to represent integer packed in `mrb_value`
instead of `inum`.
|
|
- remove `mrb_jmpbuf` from `truct mrb_parser_state`
- unify exception handling of `mrb_state` and `mrb_parser_state`.
|
|
|
|
Replace them by `mrb_ensure_string_type()`.
|
|
And merged to `mrb_f_integer()` which is only usage of the function.
|
|
|
|
|
|
Allow `nil` for `c!` and `I!` specifiers of `mrb_get_args()`
|
|
|
|
It is `o`, `C`, `S`, `A` and `H` specifiers that are integrated.
As a side effect, the `C!` Specifier can now be used.
|
|
|
|
|
|
`MRB_WORDBOX_USE_HEAP_FLOAT` instead of `MRB_USE_FLOAT_FULL_PRECISION`.
|
|
* use predefined `mrb_ro_data_p()` for user-mode Linux and macOS
* define `MRB_LINK_TIME_RO_DATA_P` if predefined one is used
* configure macro `MRB_USE_LINK_TIME_RO_DATA_P` is no longer used
* contributions for new platforms are welcome
|
|
Checks the frozen object with `mrb_get_args()`
|
|
This now works with the `+` modifier that can be added after each specifier.
- `nil` is bypassed.
- The `s` and `z` specifiers are received in C as a `const char *`, so adding a `+` modifier will raise an exception.
- The `a` specifier is received in C as `const mrb_value *`, so adding a `+` modifier will raise an exception.
- The `|`, `*`, `&`, `?` and `:` specifiers with `+` modifier raises an exception.
If `!`/`+` exceeds one for each specifier, an exception will occur in the subsequent processing.
This is the same behavior as before.
|
|
- `#include <math.h>` is done in `mruby.h`.
Eliminate the need to worry about the `MRB_NO_FLOAT` macro.
- Include mruby header files before standard header files.
If the standard header file is already placed before `mruby.h`, the standard header file added in the future tends to be placed before `mruby.h`.
This change should some reduce the chances of macros that must be defined becoming undefined in C++ or including problematic header files in a particular mruby build configuration.
|
|
Embedding reduce memory consumption, sacrificing precision. It clips least
significant 2 bits from `mrb_float`, so if you need to keep float precision,
define `MRB_USE_FLOAT_FULL_PRECISION`.
`MRB_WORD_BOXING` and `MRB_INT64`:
`mrb_float` (`double`) is embedded in `mrb_value` clipped last 2 bits.
`MRB_WORD_BOXING` and `MRB_INT64` and `MRB_USE_FLOAT_FULL_PRECISION`:
`mrb_float` is allocated in the heaps wrapped by `struct RFloat`.
`MRB_WORD_BOXING` and `MRB_INT32` and `MRB_USE_FLOAT32`:
`mrb_float` (`float`) is embedded in `mrb_value` clipped last 2 bits.
In addition, to reserve bit space in the `mrb_value`, maximum inline
symbol length become 4 (instead of 5) in the configuration.
`MRB_WORD_BOXING` and `MRB_INT32`:
Assume `MRB_USE_FLOAT_FULL_PRECISION` and allocate Float values in heap.
|
|
In additions:
* use bitmap for flags
* BER integer compression for length
|
|
|
|
Probably we should add conditions for FreeBSD, etc.
|
|
|
|
Previously, the `I` specifier only checked if the object was `MRB_TT_ISTRUCT`.
So it was at risk of getting pointers to different C structs if multiple classes were to use the `MRB_TT_ISTRUCT` instance.
Change this behavior and change the C argument corresponding to the `I` specifier to `(void *, struct RClass)`.
This change is not compatible with the previous mruby.
Please note that if the user uses the previous specifications, `SIGSEGV` may occur or the machine stack may be destroyed.
resolve #5527
|
|
|
|
It rescans `s->iseq` so that peephole optimizer can take multiple
previous instructions for constant folding, etc.
|
|
|
|
- stdlib.h
- stddef.h
- stdint.h
- stdarg.h
- limits.h
- float.h
|
|
When doing `conf.enable_cxx_abi` and compiling with FreeBSD + clang or MinGW, such as `INTPTR_MAX` constant macro is not defined if `#include <stdint.h>` precedes `#include <mruby.h>`.
Currently I get a warning when I use an undefined macro, but if I don't notice it I get confused in a link error.
It can be expected that the problem will be easier to understand by making a clear error.
Adding `-Werror=undef` as a compiler flag can also result in an error, but this can be a problem if the system header file itself uses undefined macros, for example.
This patch does minimal confirmation only, but has no side effects.
|
|
|
|
It uses BER number compression of delta of instruction positions and line
numbers. BER compression is a variable length number representation.
* `mrb_debug_line_ary`: array of line numbers represented in `uint16_t`.
`[lineno, lineno, ...]`
* `mrb_debug_line_flat_map`: array of `mrb_irep_debug_info_line`, which
is `struct {uint32_t pos; uint16_t lineno}`, for each line.
* `mrb_debug_line_packed_map` [new]: sequence of BER compressed 2
numbers, `pos_delta, lineno_delta`. Deltas are differences from
previous values (starting `0`). `line_entry_counts` represents total
length of a packed map string for this type.
|
|
- `Lit` -> `Pool`
- `SEQ` -> `Irep`
|
|
The code generator no longer need to emit `OP_LOADSYM` after `OP_DEF`.
`doc/opcode.md` is also updated.
|
|
This reverts commit fd10c7231906ca48cb35892d2a86460004b62249.
I thought it was OK to restrict index value within 1 byte, but in some
cases index value could be 16 bits (2 bytes). I had several ideas to
address the issue, but reverting `fd10c72` is the easiest way. The
biggest reason is `mruby/c` still supports `OP_EXT[123]`, so that they
don't need any additional work.
|
