This callcache is on stack, must not be GCed. However its contents are
copied from other materials, which can be an ordinal object. Should
set a flag to make sure it is properly skipped by the GC.
This changeset reduces the generated binary of
vm_call_method_missing_body from 604 bytes to 532 bytes on my machine.
Should reduce GC pressure as well.
VM stack could overflow here. The condition is when a symbol is passed
to a block-taking method via &variable, and that symbol has never been
used for actual method names (thus yielding that results in calling
method_missing), and the VM stack is full (no single word left). This
is a once-in-a-blue-moon event. Yet there is a very tiny room of stack
overflow. We need to check that.
This commit changes the number of calls of MEMCPY from...
| send | &:sym
-------------------------|-------|-------
Symbol already interned | once | twice
Symbol not pinned yet | none | once
to:
| send | &:sym
-------------------------|-------|-------
Symbol already interned | once | none
Symbol not pinned yet | twice | once
So it sacrifices exceptional situation for normal path.
Symbol#to_proc and Object#send are closely related each other. Why not
share their implementations. By doing so we can skip recursive call of
vm_exec(), which could benefit for speed.
This changeset slightly speeds up on my machine.
Calculating -------------------------------------
before after
Optcarrot Lan_Master.nes 38.33488426546287 40.89825082589147 fps
40.91288557922081 41.48687465359386
40.96591995270991 41.98499064664184
41.20461943032173 43.67314690779162
42.38344888176518 44.02777536251875
43.43563728880915 44.88695892714136
43.88082889062643 45.11226186242523
This makes it possible for vm_invoke_block to pass its passed arguments
verbatimly to calling functions. Because they are tail-called the
function calls can be strength-recuced into indirect jumps, which is a
huge win.
These two function were almost identical, except in case of
T_STRING/T_FLOAT. Why not merge them into one, and let the difference be
handled in normal method calls (slowpath). This does not improve
runtime performance for me, but at least reduces for instance rb_eql_opt
from 653 bytes to 86 bytes on my machine, according to nm(1).
If a module has an origin, and that module is included in another
module or class, previously the iclass created for the module had
an origin pointer to the module's origin instead of the iclass's
origin.
Setting the origin pointer correctly requires using a stack, since
the origin iclass is not created until after the iclass itself.
Use a hidden ruby array to implement that stack.
Correctly assigning the origin pointers in the iclass caused a
use-after-free in GC. If a module with an origin is included
in a class, the iclass shares a method table with the module
and the iclass origin shares a method table with module origin.
Mark iclass origin with a flag that notes that even though the
iclass is an origin, it shares a method table, so the method table
should not be garbage collected. The shared method table will be
garbage collected when the module origin is garbage collected.
I've tested that this does not introduce a memory leak.
This change caused a VM assertion failure, which was traced to callable
method entries using the incorrect defined_class. Update
rb_vm_check_redefinition_opt_method and find_defined_class_by_owner
to treat iclass origins different than class origins to avoid this
issue.
This also includes a fix for Module#included_modules to skip
iclasses with origins.
Fixes [Bug #16736]
rb_callable_method_entry() creates ccs entry in cc_tbl, but this
code overwrite by insert newly created ccs and overwrote ccs never
freed.
[Bug #16900]
We started to use fastpath on invokesuper when a method is not refinements
since 5c27681813, but we shouldn't have used fastpath for attr_writer either.
`cc->aux_.attr_index` is for an actual receiver class, while we store
its superclass in `cc->klass` and therefore there's no way to properly
invalidate attr_writer's inline cache when it's called by super.
[Bug #16785]
I suspect the same bug also exists in attr_reader. I'll address that in
another commit.
Fastpath has not been used for invokesuper since it has set vm_call_super_method on every invocation.
Because it seems to be blocked only by refinements, try enabling fastpath on invokesuper when cme is not for refinements.
While this patch itself should be helpful for VM performance, a part of this patch's motivation is to unblock inlining invokesuper on JIT.
$ benchmark-driver -v --rbenv 'before;after' benchmark/vm2_super.yml --repeat-count=4
before: ruby 2.8.0dev (2020-04-11T15:19:58Z master a01bda5949) [x86_64-linux]
after: ruby 2.8.0dev (2020-04-12T02:00:08Z invokesuper-fastpath c171984ee3) [x86_64-linux]
Calculating -------------------------------------
before after
vm2_super 20.031M 32.860M i/s - 6.000M times in 0.299534s 0.182593s
Comparison:
vm2_super
after: 32859885.2 i/s
before: 20031097.3 i/s - 1.64x slower
This changes the following warnings:
* warning: class variable access from toplevel
* warning: class variable @foo of D is overtaken by C
into RuntimeErrors. Handle defined?(@@foo) at toplevel
by returning nil instead of raising an exception (the previous
behavior warned before returning nil when defined? was used).
Refactor the specs to avoid the warnings even in older versions.
The specs were checking for the warnings, but the purpose of
the related specs as evidenced from their description is to
test for behavior, not for warnings.
Fixes [Bug #14541]
Previously, passing a keyword splat to a method always allocated
a hash on the caller side, and accepting arbitrary keywords in
a method allocated a separate hash on the callee side. Passing
explicit keywords to a method that accepted a keyword splat
did not allocate a hash on the caller side, but resulted in two
hashes allocated on the callee side.
This commit makes passing a single keyword splat to a method not
allocate a hash on the caller side. Passing multiple keyword
splats or a mix of explicit keywords and a keyword splat still
generates a hash on the caller side. On the callee side,
if arbitrary keywords are not accepted, it does not allocate a
hash. If arbitrary keywords are accepted, it will allocate a
hash, but this commit uses a callinfo flag to indicate whether
the caller already allocated a hash, and if so, the callee can
use the passed hash without duplicating it. So this commit
should make it so that a maximum of a single hash is allocated
during method calls.
To set the callinfo flag appropriately, method call argument
compilation checks if only a single keyword splat is given.
If only one keyword splat is given, the VM_CALL_KW_SPLAT_MUT
callinfo flag is not set, since in that case the keyword
splat is passed directly and not mutable. If more than one
splat is used, a new hash needs to be generated on the caller
side, and in that case the callinfo flag is set, indicating
the keyword splat is mutable by the callee.
In compile_hash, used for both hash and keyword argument
compilation, if compiling keyword arguments and only a
single keyword splat is used, pass the argument directly.
On the caller side, in vm_args.c, the callinfo flag needs to
be recognized and handled. Because the keyword splat
argument may not be a hash, it needs to be converted to a
hash first if not. Then, unless the callinfo flag is set,
the hash needs to be duplicated. The temporary copy of the
callinfo flag, kw_flag, is updated if a hash was duplicated,
to prevent the need to duplicate it again. If we are
converting to a hash or duplicating a hash, we need to update
the argument array, which can including duplicating the
positional splat array if one was passed. CALLER_SETUP_ARG
and a couple other places needs to be modified to handle
similar issues for other types of calls.
This includes fairly comprehensive tests for different ways
keywords are handled internally, checking that you get equal
results but that keyword splats on the caller side result in
distinct objects for keyword rest parameters.
Included are benchmarks for keyword argument calls.
Brief results when compiled without optimization:
def kw(a: 1) a end
def kws(**kw) kw end
h = {a: 1}
kw(a: 1) # about same
kw(**h) # 2.37x faster
kws(a: 1) # 1.30x faster
kws(**h) # 2.19x faster
kw(a: 1, **h) # 1.03x slower
kw(**h, **h) # about same
kws(a: 1, **h) # 1.16x faster
kws(**h, **h) # 1.14x faster
send() has special method launcher in VM and it has special
method_missing caller. This path doesn't set
ec->method_missing_reason which is used at exception creation,
so setup this information. Without this setting, NoMethodError
exception becomes NameError.
This patch will fix:
http://ci.rvm.jp/results/trunk-random1@phosphorus-docker/2761643
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 removes the warnings added in 2.7, and changes the behavior
so that a final positional hash is not treated as keywords or
vice-versa.
To handle the arg_setup_block splat case correctly with keyword
arguments, we need to check if we are taking a keyword hash.
That case didn't have a test, but it affects real-world code,
so add a test for it.
This removes rb_empty_keyword_given_p() and related code, as
that is not needed in Ruby 3. The empty keyword case is the
same as the no keyword case in Ruby 3.
This changes rb_scan_args to implement keyword argument
separation for C functions when the : character is used.
For backwards compatibility, it returns a duped hash.
This is a bad idea for performance, but not duping the hash
breaks at least Enumerator::ArithmeticSequence#inspect.
Instead of having RB_PASS_CALLED_KEYWORDS be a number,
simplify the code by just making it be rb_keyword_given_p().
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).
With these macros implemented we can write codes just like we can assume
the compiler being clang. MSC_VERSION_SINCE is defined to implement
those macros, but turned out to be handy for other places. The -fdeclspec
compiler flag is necessary for clang to properly handle __has_declspec().
Methods and their definitions can be allocated/deallocated on-the-fly.
One pathological situation is when a method is deallocated then another
one is allocated immediately after that. Address of those old/new method
entries/definitions can be the same then, depending on underlying
malloc/free implementation.
So pointer comparison is insufficient. We have to check the contents.
To do so we introduce def->method_serial, which is an integer unique to
that specific method definition.
PS: Note that method_serial being uintptr_t rather than rb_serial_t is
intentional. This is because rb_serial_t can be bigger than a pointer
on a 32bit system (rb_serial_t is at least 64bit). In order to preserve
old packing of struct rb_call_cache, rb_serial_t is inappropriate.
The equation shall hold for every call cache. However prior to this
changeset cc->me could be updated without also updating cc->def. Let's
make it sure by introducing new macro named CC_SET_ME which sets cc->me
and cc->def at once.
This makes behavior the same as super in instance_eval in method
in class. The reason this wasn't implemented before is that
there is a check to determine if the self in the current context
is of the expected class, and a module itself can be included
in multiple classes, so it doesn't have an expected class.
Implementing this requires giving iclasses knowledge of which
class created them, so that super call in the module method
knows the expected class for super calls. This reference
is called includer, and should only be set for iclasses.
Note that the approach Ruby uses in this check is not robust. If
you instance_eval another object of the same class and call super,
instead of an TypeError, you get super called with the
instance_eval receiver instead of the method receiver. Truly
fixing super would require keeping a reference to the super object
(method receiver) in each frame where scope has changed, and using
that instead of current self when calling super.
Fixes [Bug #11636]
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)
Asynchronous events such as signal trap, finalization timing,
thread switching and so on are managed by "interrupt_flag".
Ruby's threads check this flag periodically and if a thread
does not check this flag, above events doesn't happen.
This checking is CHECK_INTS() (related) macro and it is placed
at some places (laeve instruction and so on). However, at the end
of C methods, C blocks (IMEMO_IFUNC) etc there are no checking
and it can introduce uninterruptible thread.
To modify this situation, we decide to place CHECK_INTS() at
vm_pop_frame(). It increases interrupt checking points.
[Bug #16366]
This patch can introduce unexpected events...
By this change, the following code prints only one warning.
```
def foo(**opt); end
100.times { foo({kw:1}) }
```
A global variable `st_table *caller_to_callees` is a map from caller to
a set of callee methods. It remembers that a warning is already printed
for each pair of caller and callee.
[Feature #16289]
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.
rb_vm_lvar_exposed() is prepared for __builtin_inline!(), needed for
mini_builtin.c and builtin.c. However, it's only on builtin.c.
So move it to make it as a part of VM.
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.
vm_invoke_builtin() accesses VM stack via cfp->sp. However, MJIT
can use their own stack. To access them appropriately, we need to
use STACK_ADDR_FROM_TOP().
A method which has keyword parameters has an implicit local variable
to specify which keywords are (un)specified.
vm_call_iseq_setup_kwparm_nokwarg() is special function to invoke
a ISeq method without any keyword arguments. However, it should
also initialize the special local var. Without this initialization,
the implicit lvar can points a freed (T_NONE) object.
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.
Prior to this changeset, majority of inline cache mishits resulted
into the same method entry when rb_callable_method_entry() resolves
a method search. Let's not call the function at the first place on
such situations.
In doing so we extend the struct rb_call_cache from 44 bytes (in
case of 64 bit machine) to 64 bytes, and fill the gap with
secondary class serial(s). Call cache's class serials now behavies
as a LRU cache.
Calculating -------------------------------------
ours 2.7 2.6
vm2_poly_same_method 2.339M 1.744M 1.369M i/s - 6.000M times in 2.565086s 3.441329s 4.381386s
Comparison:
vm2_poly_same_method
ours: 2339103.0 i/s
2.7: 1743512.3 i/s - 1.34x slower
2.6: 1369429.8 i/s - 1.71x slower
Noticed that rb_method_basic_definition_p is frequently called.
Its callers include vm_caller_setup_args_block(),
rb_hash_default_value(), rb_num_neative_int_p(), and a lot more.
It seems worth caching the method resolution part. Majority of
rb_method_basic_definion_p() usages take fixed class and fixed
method id combinations.
Calculating -------------------------------------
ours trunk
so_matrix 2.379 2.115 i/s - 1.000 times in 0.420409s 0.472879s
Comparison:
so_matrix
ours: 2.4 i/s
trunk: 2.1 i/s - 1.12x slower
This mirrors the behavior when manually splatting a hash. This
mirrors the changes made in setup_parameters_complex in
6081ddd6e6, so that splatting to a
non-iseq method works the same as splatting to an iseq method.
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]
Takashi Kokubun
Nobuyoshi Nakada
卜部昌平
Jeremy Evans
Koichi Sasada
John Hawthorn
Aaron Patterson
Yusuke Endoh
git
Kazuhiro NISHIYAMA
Dylan Thacker-Smith
wanabe
Alan Wu