This patch contains several ideas:
(1) Disposable inline method cache (IMC) for race-free inline method cache
* Making call-cache (CC) as a RVALUE (GC target object) and allocate new
CC on cache miss.
* This technique allows race-free access from parallel processing
elements like RCU.
(2) Introduce per-Class method cache (pCMC)
* Instead of fixed-size global method cache (GMC), pCMC allows flexible
cache size.
* Caching CCs reduces CC allocation and allow sharing CC's fast-path
between same call-info (CI) call-sites.
(3) Invalidate an inline method cache by invalidating corresponding method
entries (MEs)
* Instead of using class serials, we set "invalidated" flag for method
entry itself to represent cache invalidation.
* Compare with using class serials, the impact of method modification
(add/overwrite/delete) is small.
* Updating class serials invalidate all method caches of the class and
sub-classes.
* Proposed approach only invalidate the method cache of only one ME.
See [Feature #16614] for more details.
Now, rb_call_info contains how to call the method with tuple of
(mid, orig_argc, flags, kwarg). Most of cases, kwarg == NULL and
mid+argc+flags only requires 64bits. So this patch packed
rb_call_info to VALUE (1 word) on such cases. If we can not
represent it in VALUE, then use imemo_callinfo which contains
conventional callinfo (rb_callinfo, renamed from rb_call_info).
iseq->body->ci_kw_size is removed because all of callinfo is VALUE
size (packed ci or a pointer to imemo_callinfo).
To access ci information, we need to use these functions:
vm_ci_mid(ci), _flag(ci), _argc(ci), _kwarg(ci).
struct rb_call_info_kw_arg is renamed to rb_callinfo_kwarg.
rb_funcallv_with_cc() and rb_method_basic_definition_p_with_cc()
is temporary removed because cd->ci should be marked.
This behavior was deprecated in 2.7 and scheduled to be removed
in 3.0.
Calling yield in a class definition outside a method is now a
SyntaxError instead of a LocalJumpError, as well.
We were inefficient in cases where there are a lot of duplicates due to
the use of linear search. Use a hash table instead.
These cases are not that rare in the wild.
[Feature #16505]
Saves comitters' daily life by avoid #include-ing everything from
internal.h to make each file do so instead. This would significantly
speed up incremental builds.
We take the following inclusion order in this changeset:
1. "ruby/config.h", where _GNU_SOURCE is defined (must be the very
first thing among everything).
2. RUBY_EXTCONF_H if any.
3. Standard C headers, sorted alphabetically.
4. Other system headers, maybe guarded by #ifdef
5. Everything else, sorted alphabetically.
Exceptions are those win32-related headers, which tend not be self-
containing (headers have inclusion order dependencies).
On USE_LAZY_LOAD=1, the iseq should be loaded. So rb_iseq_check()
is needed. Furthermore, now lazy loading with builtin_function_table
is not supported, so it should cancel lazy loading.
`foo(*rest, post, **empty_kw)` is compiled like
`foo(*rest + [post, **empty_kw])`, and `**empty_kw` is removed by
"newarraykwsplat" instruction.
However, the method call still has a flag of KW_SPLAT, so "post" is
considered as a keyword hash, which caused a segfault.
Note that the flag cannot be removed if "empty_kw" is not always empty.
This change fixes the issue by compiling arguments with "newarray"
instead of "newarraykwsplat".
[Bug #16442]
Now, C functions written by __builtin_cexpr!(code) and others are
named as "__builtin_inline#{n}". However, it is difficult to know
what the function is. This patch rename them into
"__builtin_foo_#{lineno}" when cexpr! is in 'foo' method.
%p is for void *. Becuase fprintf is a function with variadic arguments
automatic cast from any pointer to void * does not work. We have to be
explicit.
(This is the second try of 036bc1da6c6c9b0fa9b7f5968d897a9554dd770e.)
If iseq is GC'ed, the pointer of iseq may be reused, which may hide a
deprecation warning of keyword argument change.
http://ci.rvm.jp/results/trunk-test1@phosphorus-docker/2474221
```
1) Failure:
TestKeywordArguments#test_explicit_super_kwsplat [/tmp/ruby/v2/src/trunk-test1/test/ruby/test_keyword.rb:549]:
--- expected
+++ actual
@@ -1 +1 @@
-/The keyword argument is passed as the last hash parameter.* for `m'/m
+""
```
This change ad-hocly adds iseq_unique_id for each iseq, and use it
instead of iseq pointer. This covers the case where caller is GC'ed.
Still, the case where callee is GC'ed, is not covered.
But anyway, it is very rare that iseq is GC'ed. Even when it occurs, it
just hides some warnings. It's no big deal.
This commit introduces an "inline ivar cache" struct. The reason we
need this is so compaction can differentiate from an ivar cache and a
regular inline cache. Regular inline caches contain references to
`VALUE` and ivar caches just contain references to the ivar index. With
this new struct we can easily update references for inline caches (but
not inline var caches as they just contain an int)
jump-jump optimization ignores the event flags of the jump instruction
being skipped, which leads to overlook of line events.
This changeset stops the wrong optimization when coverage measurement is
neabled and when the jump instruction has any event flag.
Note that this issue is not only for coverage but also for TracePoint,
and this change does not fix TracePoint.
However, fixing it fundamentally is tough (which requires revamp of
the compiler). This issue is critical in terms of coverage measurement,
but minor for TracePoint (ko1 said), so we here choose a stopgap
measurement.
[Bug #15980] [Bug #16397]
Note for backporters: this changeset can be viewed by `git diff -w`.
rename __builtin_inline!(code) to __builtin_cstmt(code).
Also this commit introduce the following inlining C code features.
* __builtin_cstmt!(STMT)
(renamed from __builtin_inline!)
Define a function which run STMT implicitly and call this function at
evatuation time. Note that you need to return some value in STMT.
If there is a local variables (includes method parameters), you can
read these values.
static VALUE func(ec, self) {
VALUE x = ...;
STMT
}
Usage:
def double a
# a is readable from C code.
__builtin_cstmt! 'return INT2FIX(FIX2INT(a) * 2);'
end
* __builtin_cexpr!(EXPR)
Define a function which invoke EXPR implicitly like `__builtin_cstmt!`.
Different from cstmt!, which compiled with `return EXPR;`.
(`return` and `;` are added implicitly)
static VALUE func(ec, self) {
VALUE x = ...;
return EXPPR;
}
Usage:
def double a
__builtin_cexpr! 'INT2FIX(FIX2INT(a) * 2)'
end
* __builtin_cconst!(EXPR)
Define a function which invoke EXPR implicitly like cexpr!.
However, the function is called once at compile time, not evaluated time.
Any local variables are not accessible (because there is no local variable
at compile time).
Usage:
GCC = __builtin_cconst! '__GNUC__'
* __builtin_cinit!(STMT)
STMT are writtein in auto-generated code.
This code does not return any value.
Usage:
__builtin_cinit! '#include <zlib.h>'
def no_compression?
__builtin_cconst! 'Z_NO_COMPRESSION ? Qtrue : Qfalse'
end
These functions are used from within a compilation unit so we can
make them static, for better binary size. This changeset reduces
the size of generated ruby binary from 26,590,128 bytes to
26,584,472 bytes on my macihne.
opt_invokebuiltin_delegate and opt_invokebuiltin_delegate_leave
invokes builtin functions with same parameters of the method.
This technique eliminate stack push operations. However, delegation
parameters should be completely same as given parameters.
(e.g. `def foo(a, b, c) __builtin_foo(a, b, c)` is okay, but
__builtin_foo(b, c) is not allowed)
This patch relaxes this restriction. ISeq has a local variables
table which includes parameters. For example, the method defined
as `def foo(a, b, c) x=y=nil`, then local variables table contains
[a, b, c, x, y]. If calling builtin-function with arguments which
are sub-array of the lvar table, use opt_invokebuiltin_delegate
instruction with start index. For example, `__builtin_foo(b, c)`,
`__builtin_bar(c, x, y)` is okay, and so on.
Fixes [Bug #16332]
Constant access was changed to no longer allow top-level constant access
through `nil`, but `defined?` wasn't changed at the same time to stay
consistent.
Use a separate defined type to distinguish between a constant
referenced from the current lexical scope and one referenced from
another namespace.
Add an experimental `__builtin_inline!(c_expression)` special intrinsic
which run a C code snippet.
In `c_expression`, you can access the following variables:
* ec (rb_execution_context_t *)
* self (const VALUE)
* local variables (const VALUE)
Not that you can read these variables, but you can not write them.
You need to return from this expression and return value will be a
result of __builtin_inline!().
Examples:
`def foo(x) __builtin_inline!('return rb_p(x);'); end` calls `p(x)`.
`def double(x) __builtin_inline!('return INT2NUM(NUM2INT(x) * 2);')`
returns x*2.
Support loading builtin features written in Ruby, which implement
with C builtin functions.
[Feature #16254]
Several features:
(1) Load .rb file at boottime with native binary.
Now, prelude.rb is loaded at boottime. However, this file is contained
into the interpreter as a text format and we need to compile it.
This patch contains a feature to load from binary format.
(2) __builtin_func() in Ruby call func() written in C.
In Ruby file, we can write `__builtin_func()` like method call.
However this is not a method call, but special syntax to call
a function `func()` written in C. C functions should be defined
in a file (same compile unit) which load this .rb file.
Functions (`func` in above example) should be defined with
(a) 1st parameter: rb_execution_context_t *ec
(b) rest parameters (0 to 15).
(c) VALUE return type.
This is very similar requirements for functions used by
rb_define_method(), however `rb_execution_context_t *ec`
is new requirement.
(3) automatic C code generation from .rb files.
tool/mk_builtin_loader.rb creates a C code to load .rb files
needed by miniruby and ruby command. This script is run by
BASERUBY, so *.rb should be written in BASERUBY compatbile
syntax. This script load a .rb file and find all of __builtin_
prefix method calls, and generate a part of C code to export
functions.
tool/mk_builtin_binary.rb creates a C code which contains
binary compiled Ruby files needed by ruby command.
Get rid of these redundant and useless warnings.
```
$ ruby -e 'def bar(a) a; end; def foo(...) bar(...) end; foo({})'
-e:1: warning: The last argument is used as the keyword parameter
-e:1: warning: for `foo' defined here
-e:1: warning: The keyword argument is passed as the last hash parameter
-e:1: warning: for `bar' defined here
```
To perform a regular method call, the VM needs two structs,
`rb_call_info` and `rb_call_cache`. At the moment, we allocate these two
structures in separate buffers. In the worst case, the CPU needs to read
4 cache lines to complete a method call. Putting the two structures
together reduces the maximum number of cache line reads to 2.
Combining the structures also saves 8 bytes per call site as the current
layout uses separate two pointers for the call info and the call cache.
This saves about 2 MiB on Discourse.
This change improves the Optcarrot benchmark at least 3%. For more
details, see attached bugs.ruby-lang.org ticket.
Complications:
- A new instruction attribute `comptime_sp_inc` is introduced to
calculate SP increase at compile time without using call caches. At
compile time, a `TS_CALLDATA` operand points to a call info struct, but
at runtime, the same operand points to a call data struct. Instruction
that explicitly define `sp_inc` also need to define `comptime_sp_inc`.
- MJIT code for copying call cache becomes slightly more complicated.
- This changes the bytecode format, which might break existing tools.
[Misc #16258]
This changeset basically replaces `ruby_xmalloc(x * y)` into
`ruby_xmalloc2(x, y)`. Some convenient functions are also
provided for instance `rb_xmalloc_mul_add(x, y, z)` which allocates
x * y + z byes.
The parser needs to determine whether a local varaiable is defined or
not in outer scope. For the sake, "base_block" field has kept the outer
block.
However, the whole block was actually unneeded; the parser used only
base_block->iseq.
So, this change lets parser_params have the iseq directly, instead of
the whole block.
`freeze_string` essentially called iseq_add_mark_object_compile_time. I
need to know where all writes occur on the `rb_iseq_t`, so this commit
separates the function calls so we can add write barriers in the right
place.
Move the "add mark object" function to the location where we should be
calling RB_OBJ_WRITTEN. I'm going to add verification code next so we
can make sure the objects we're adding to the array are also reachable
from the mark function.
This makes it consistent with calling private attribute assignment
methods, which currently is allowed (e.g. `self.value =`).
Calling a private method in this way can be useful when trying to
assign the return value to a local variable with the same name.
[Feature #11297] [Feature #16123]
The output of RubyVM::InstructionSequence#to_binary is extremely large.
We have reduced the output of #to_binary by more than 70%.
The execution speed of RubyVM::InstructionSequence.load_from_binary is about 7% slower, but when reading a binary from a file, it may be faster than the master.
Since Bootsnap gem uses #to_binary, this proposal reduces the compilation cache size of Rails projects to about 1/4.
See details: [Feature #16163]
RubyVM::InstructionSequence.to_binary generates a bytecode binary
representation. To check compatibility with binary and loading
MRI we prepared major/minor version and compare them at loading
time. However, development version of MRI can change this format
but we can not increment minor version to make them consistent
with Ruby's major/minor versions.
To solve this issue, we introduce new minor version scheme
(binary's minor_version = ruby's minor * 10000 + dev ver)
and we can check incompatibility with older dev version.
The original code looks unnecessarily complicated (to me).
Also, it creates a pre-allocated array only for the prefix of the array.
The new code optimizes not only the prefix but also the subsequence that
is longer than 0x40 elements.
# not optimized
10000000.times { [1+1, 1,2,3,4,...,63] } # 2.12 sec.
# (1+1; push 1; push 2; ...; puts 63; newarray 64; concatarray)
# optimized
10000000.times { [1+1, 1,2,3,4,...,63,64] } # 1.46 sec.
# (1+1; newarray 1; putobject [1,2,3,...,64]; concatarray)
"popped" case can be so simple, so this change moves the branch to the
first, instead of scattering `if (popped)` branches to the main part.
Also, the return value "len" is not used. So it returns just 0 or 1.
compile_list was for the compilation of Array literal and Hash literal.
I guess it was originally reasonable to handle them in one function, but
now, compilation of Array is very different from Hash. So the function
was complicated by many branches for Array and Hash.
This change separates the function to two ones for Array and Hash.
An array literal [1,2,...,301] was compiled to the following iseq:
duparray [1,2,...,300]
putobject [301]
concatarray
The Array literal optimization took every two elements maybe because it
must handle not only Array but also Hash.
Now the optimization takes each element if it is an Array literal. So
the new iseq is: duparray [1,2,...,301].
`[{}, {}, {}, ..., {}, *{}]` is wrongly created.
A big array literal is created and concatenated for every 256 elements.
The newarraykwsplat must be emitted only at the last chunk.
and NODE_ZARRAY to NODE_ZLIST.
NODE_ARRAY is used not only by an Array literal, but also the contents
of Hash literals, method call arguments, dynamic string literals, etc.
In addition, the structure of NODE_ARRAY is a linked list, not an array.
This is very confusing, so I believe `NODE_LIST` is a better name.
Previously, **{} was removed by the parser:
```
$ ruby --dump=parse -e '{**{}}'
@ NODE_SCOPE (line: 1, location: (1,0)-(1,6))
+- nd_tbl: (empty)
+- nd_args:
| (null node)
+- nd_body:
@ NODE_HASH (line: 1, location: (1,0)-(1,6))*
+- nd_brace: 1 (hash literal)
+- nd_head:
(null node)
```
Since it was removed by the parser, the compiler did not know
about it, and `m(**{})` was therefore treated as `m()`.
This modifies the parser to not remove the `**{}`. A simple
approach for this is fairly simple by just removing a few
lines from the parser, but that would cause two hash
allocations every time it was used. The approach taken here
modifies both the parser and the compiler, and results in `**{}`
not allocating any hashes in the usual case.
The basic idea is we use a literal node in the parser containing
a frozen empty hash literal. In the compiler, we recognize when
that is used, and if it is the only keyword present, we just
push it onto the VM stack (no creation of a new hash or merging
of keywords). If it is the first keyword present, we push a
new empty hash onto the VM stack, so that later keywords can
merge into it. If it is not the first keyword present, we can
ignore it, since the there is no reason to merge an empty hash
into the existing hash.
Example instructions for `m(**{})`
Before (note ARGS_SIMPLE):
```
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,7)> (catch: FALSE)
0000 putself ( 1)[Li]
0001 opt_send_without_block <callinfo!mid:m, argc:0, FCALL|ARGS_SIMPLE>, <callcache>
0004 leave
```
After (note putobject and KW_SPLAT):
```
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,7)> (catch: FALSE)
0000 putself ( 1)[Li]
0001 putobject {}
0003 opt_send_without_block <callinfo!mid:m, argc:1, FCALL|KW_SPLAT>, <callcache>
0006 leave
```
Example instructions for `m(**h, **{})`
Before and After (no change):
```
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,12)> (catch: FALSE)
0000 putself ( 1)[Li]
0001 putspecialobject 1
0003 newhash 0
0005 putself
0006 opt_send_without_block <callinfo!mid:h, argc:0, FCALL|VCALL|ARGS_SIMPLE>, <callcache>
0009 opt_send_without_block <callinfo!mid:core#hash_merge_kwd, argc:2, ARGS_SIMPLE>, <callcache>
0012 opt_send_without_block <callinfo!mid:m, argc:1, FCALL|KW_SPLAT>, <callcache>
0015 leave
```
Example instructions for `m(**{}, **h)`
Before:
```
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,12)> (catch: FALSE)
0000 putself ( 1)[Li]
0001 putspecialobject 1
0003 newhash 0
0005 putself
0006 opt_send_without_block <callinfo!mid:h, argc:0, FCALL|VCALL|ARGS_SIMPLE>, <callcache>
0009 opt_send_without_block <callinfo!mid:core#hash_merge_kwd, argc:2, ARGS_SIMPLE>, <callcache>
0012 opt_send_without_block <callinfo!mid:m, argc:1, FCALL|KW_SPLAT>, <callcache>
0015 leave
```
After (basically the same except for the addition of swap):
```
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,12)> (catch: FALSE)
0000 putself ( 1)[Li]
0001 newhash 0
0003 putspecialobject 1
0005 swap
0006 putself
0007 opt_send_without_block <callinfo!mid:h, argc:0, FCALL|VCALL|ARGS_SIMPLE>, <callcache>
0010 opt_send_without_block <callinfo!mid:core#hash_merge_kwd, argc:2, ARGS_SIMPLE>, <callcache>
0013 opt_send_without_block <callinfo!mid:m, argc:1, FCALL|KW_SPLAT>, <callcache>
0016 leave
```
for simplicity and consistency.
Now SUPPORT_JOKE needs to be prefixed with OPT_ to make the config
visible in `RubyVM::VmOptsH`, and the inconsistency was introduced.
As it has never been available for override in configure (no #ifndef
guard), it should be fine to rename the config.
This syntax means the method should be treated as a method that
uses keyword arguments, but no specific keyword arguments are
supported, and therefore calling the method with keyword arguments
will raise an ArgumentError. It is still allowed to double splat
an empty hash when calling the method, as that does not pass
any keyword arguments.
After 5e86b005c0, I now think ANYARGS is
dangerous and should be extinct. This commit adds function prototypes
for rb_hash_foreach / st_foreach_safe. Also fixes some prototype
mismatches.
Jeremy Evans
Koichi Sasada
Nobuyoshi Nakada
卜部昌平
Masataka Pocke Kuwabara
NagayamaRyoga
Seiei Miyagi
Yusuke Endoh
Aaron Patterson
Dylan Thacker-Smith
git
Lourens Naudé
Alan Wu
NagayamaRyoga
Kazuhiro NISHIYAMA
Takashi Kokubun
Kazuki Tsujimoto