padding on 64 bit platforms. Patch by Eric Wong.
[Feature #9441] [ruby-core:59993]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@44686 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
RClass->m_tbl with serial. This prevents double marking method
tables, since many classes/modules can share the same method table.
This improves minor mark time in a large application by 30%.
* internal.h (struct method_table_wrapper): Define new
wrapper struct with additional serial.
* internal.h (RCLASS_M_TBL_INIT): New macro for initializing method
table wrapper and st_table.
* method.h (void rb_sweep_method_entry): Rename rb_free_m_table to
rb_free_m_tbl for consistentcy
* .gdbinit (define rb_method_entry): Update rb_method_entry gdb helper
for new method table structure.
* class.c: Use RCLASS_M_TBL_WRAPPER and
RCLASS_M_TBL_INIT macros.
* class.c (rb_include_class_new): Share WRAPPER between module and
iclass, so serial can prevent double marking.
* eval.c (rb_prepend_module): ditto.
* eval.c (rb_using_refinement): ditto.
* gc.c: Mark and free new wrapper struct.
* gc.c (obj_memsize_of): Count size of additional wrapper struct.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@43973 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* gc.c (obj_free): ditto
* internal.h (struct rb_classext_struct): ditto
* method.h (rb_method_entry): remove ent param
* vm_method.c: restore the global method cache. Per class cache tables
turned out to be far too slow.
[ruby-core:57289] [Bug #8930]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@43027 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* vm_method.c (rb_method_entry_at): return the method entry for id at
klass, without ancestors.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@40683 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* proc.c (rb_method_entry_location, rb_{mod,obj}_method_location): new
functions to obtain source location of method definition.
* vm_method.c (rb_obj_respond_to): show the location of old style
respond_to? method.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@39089 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
This method type is for optimized CFUNC such as Fixnum#+ and so on.
This feature is half-baked and no way to use them.
[Background]
Now, VM has opt_plus instructions to optimize `+' methods for
some Classes (such as Fixnum, Float (flonum)). We call this
type of instructions as `specialized instructions'.
This simple technique improve simple program dramatically.
However, we can make specialized instructions for only several
types (classes) and selectors (method names) because a large
instruction will be slow. In other words, this technique has no
extensibility.
To overcome this problem, VM_METHOD_TYPE_CFUNC_FRAMELESS was
introduced (r37198). This type is a variant of CFUNC, but called
their functiions directly without building a method frame.
Any CFUNC method can be defined as frameless methods if a method
is not needed to make method frame. Frameless methods are faster
as specialized instructions (a bit slower, but no need to care).
No problem described at
http://charlie.bz/blog/why-do-singleton-methods-make-ruby-slow
because this technique doesn't see class, but see method body
itself. Alias is also no problem.
[Problem]
However, we can't set frameless method type for polymorphic methods
such as Array#[]. Necessity for method frame depends on which
parameter type. For example, Fixnum#+ needs method frame if
coerce is needed. Current VM_METHOD_TYPE_CFUNC_FRAMELESS is not
flexible and need more tuning to introduce it.
Expected behavior of frameless method type may be:
result = optimized_cfunc(params); /* call optimized cfunc */
if (result == Qundef) { result = normal_cfunc(); }
This is why I say this feature is half-baked.
We need to learn primitive method in Smalltalk more.
(I heard this name at RubyConf Taiwan this month. Thanks!)
[Conclusion]
Nobody may use this feature and there is no compatibility issue.
This feature goes to next minor (2.1?).
* proc.c (rb_method_entry_arity): ditto.
* vm_eval.c, vm_insnhelper.c, vm_method.c: ditto.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@38431 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
rb_method_type_t is not a number and "_MAX" causes misunderstanding.
* proc.c (rb_method_entry_arity): ditto.
* vm_eval.c (vm_call0_body): ditto.
* vm_insnhelper.c (vm_call_method): ditto.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@38349 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* method.h (rb_method_flag_t): name a magic number for NOEX_SAFE and
NOEX_WITH as NOEX_SAFE_SHIFT_OFFSET.
* method.h (rb_method_type_t, method_optimized_type): C89 forbids a
comma after the last element in enum.
* proc.c (rb_method_entry_arity), vm_eval.c (vm_call0_body),
vm_insnhelper.c (vm_call_method): add VM_METHOD_TYPE__MAX case.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@38340 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
refinements.
* proc.c (mnew): Kernel#method, Kernel#public_method,
Module#instance_method, and Module#public_instance_method should
not use refinements.
* vm_method.c (rb_method_boundp): Kernel#respond_to? should not use
refinements.
* test/ruby/test_refinement.rb: related test.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@38279 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
entry with VM_METHOD_TYPE_REFINED holds only the original method
definition, so ci->me is set to a method entry allocated in the
stack, and it causes SEGV/ILL. In this commit, a method entry
with VM_METHOD_TYPE_REFINED holds the whole original method entry.
Furthermore, rb_thread_mark() is changed to mark cfp->klass to
avoid GC for iclasses created by copy_refinement_iclass().
* vm_method.c (rb_method_entry_make): add a method entry with
VM_METHOD_TYPE_REFINED to the class refined by the refinement if
the target module is a refinement. When a method entry with
VM_METHOD_TYPE_UNDEF is invoked by vm_call_method(), a method with
the same name is searched in refinements. If such a method is
found, the method is invoked. Otherwise, the original method in
the refined class (rb_method_definition_t::body.orig_me) is
invoked. This change is made to simplify the normal method lookup
and to improve the performance of normal method calls.
* vm_method.c (EXPR1, search_method, rb_method_entry),
vm_eval.c (rb_call0, rb_search_method_entry): do not use
refinements for method lookup.
* vm_insnhelper.c (vm_call_method): search methods in refinements if
ci->me is VM_METHOD_TYPE_REFINED. If the method is called by
super (i.e., ci->call == vm_call_super_method), skip the same
method entry as the current method to avoid infinite call of the
same method.
* class.c (include_modules_at): add a refined method entry for each
method defined in a module included in a refinement.
* class.c (rb_prepend_module): set an empty table to
RCLASS_M_TBL(klass) to add refined method entries, because
refinements should have priority over prepended modules.
* proc.c (mnew): use rb_method_entry_with_refinements() to get
a refined method.
* vm.c (rb_thread_mark): mark cfp->klass for iclasses created by
copy_refinement_iclass().
* vm.c (Init_VM), cont.c (fiber_init): initialize th->cfp->klass.
* test/ruby/test_refinement.rb (test_inline_method_cache): do not skip
the test because it should pass successfully.
* test/ruby/test_refinement.rb (test_redefine_refined_method): new
test for the case a refined method is redefined.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@38236 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
VM_METHOD_TYPE_REFINED to the class refined by the refinement if
the target module is a refinement. When a method entry with
VM_METHOD_TYPE_UNDEF is invoked by vm_call_method(), a method with
the same name is searched in refinements. If such a method is
found, the method is invoked. Otherwise, the original method in
the refined class (rb_method_definition_t::body.orig_def) is
invoked. This change is made to simplify the normal method lookup
and to improve the performance of normal method calls.
* vm_method.c (EXPR1, search_method, rb_method_entry),
vm_eval.c (rb_call0, rb_search_method_entry): do not use
refinements for method lookup.
* vm_insnhelper.c (vm_call_method): search methods in refinements if
ci->me is VM_METHOD_TYPE_REFINED. If the method is called by
super (i.e., ci->call == vm_call_super_method), skip the same
method entry as the current method to avoid infinite call of the
same method.
* class.c (include_modules_at): add a refined method entry for each
method defined in a module included in a refinement.
* class.c (rb_prepend_module): set an empty table to
RCLASS_M_TBL(klass) to add refined method entries, because
refinements should have priority over prepended modules.
* proc.c (mnew): use rb_method_entry_with_refinements() to get
a refined method.
* test/ruby/test_refinement.rb (test_inline_method_cache): do not skip
the test because it should pass successfully.
* test/ruby/test_refinement.rb (test_redefine_refined_method): new
test for the case a refined method is redefined.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37993 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
`rb_add_method_cfunc_frameless()' API.
This API is not mature to become an offical API.
For example, we can not use this API with
`rb_define_private_method()'.
* method.h, vm_method.c (rb_add_method_cfunc_frameless): removed.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37320 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
add a new otpimization and its macro `OPT_CALL_CFUNC_WITHOUT_FRAME'.
This optimization makes all cfunc method calls `frameless', which
is fster than ordinal cfunc method call.
If `frame' is needed (for example, it calls another method with
`rb_funcall()'), then build a frame. In other words, this
optimization delays frame building.
However, to delay the frame building, we need additional overheads:
(1) Store the last call information.
(2) Check the delayed frame buidling before the frame is needed.
(3) Overhead to build a delayed frame.
rb_thread_t::passed_ci is storage of delayed cfunc call information.
(1) is lightweight because it is only 1 assignment to `passed_ci'.
To achieve (2), we modify GET_THREAD() to check `passed_ci' every
time. It causes 10% overhead on my envrionment.
This optimization only works for cfunc methods which do not need
their `frame'.
After evaluation on my environment, this optimization does not
effective every time. Because of this evaluation results, this
optimization is disabled at default.
* vm_insnhelper.c, vm.c: add VM_PROFILE* macros to measure behaviour
of VM internals. I will extend this feature.
* vm_method.c, method.h: change parameters of the `invoker' function.
Receive `func' pointer as the first parameter.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37293 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
`invoker'. `invoker' function invoke cfunc body
(rb_method_cfunc_t::func).
`invoker' is set at method definition timing.
With this change, the big `switch' (branch) in `call_cfunc()'
is no longer needed.
However, the performance benefit is only a bit.
* vm_core.h (rb_call_info_t::aux::func): add a new field to store
cfunc body function pointer.
* vm_method.c (call_cfunc_invoker_func): add a new function which
returns a suitable invoke function.
* vm_method.c (setup_method_cfunc_struct): added.
* vm_method.c (rb_add_method): fix to set `invoker'.
* vm_eval.c (vm_call0_body): catch up above changes.
* vm_insnhelper.c (call_cfunc): removed.
* vm_insnhelper.c (vm_call_cfunc): fix to call cfunc body
with `invoker' function.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37268 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
rb_define_method_fast(). Defined method with this C API
does not make a method frame. It is bit lightweight than
ordinal C functions. Now only 0 or 1 argc are permitted.
* method.h (VM_METHOD_TYPE_CFUNC_FRAMELESS): rename macro name
from VM_METHOD_TYPE_CFUNC_FAST.
* vm_insnhelper.c, vm_method.c: rename related functions.
* proc.c (rb_method_entry_arity): catch up above changes.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37252 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
This method is similar to VM_METHOD_TYPE_CFUNC methods, but
called cfunc without building new frame (does not push new control
frame). If error is occured in cfunc, the backtrace only shows
caller frame and upper.
This kind of methods can be added by rb_define_method_fast().
This feature is similar to specialized instructions (opt_plus, etc),
but more flexible (but a bit slower).
* class.c (rb_define_method_fast): added.
Maybe it will be renamed soon.
* vm_insnhelper.c (vm_call_method): support method type
VM_METHOD_TYPE_CFUNC_FAST.
* proc.c (rb_method_entry_arity): catch up new method type.
* vm_method.c (rb_add_method_cfunc_fast): added.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@37198 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
vm_insnhelper.c, vm_insnhelper.h, vm_method.c: add klass to
rb_control_frame_t to implement super correctly.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@36595 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
[Bug #6048]. Isolate hash computation for proc
* internal.h: Declaration for above
* vm_method.c (rb_method_definition_hash): Computation for
hash part of a method definition
* method.h: Declaration for above
* test/ruby/test_method.rb: Test for above
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@34715 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* compile.c, eval.c, iseq.c, object.c, parse.y, proc.c, process.c,
thread.c, vm.c, vm_eval.c, vm_insnhelper.c, vm_method.c: don't
declare internal functions.
Note that rb_method_entry_eq() is defined in vm_method.c but
there was a declaration in proc.c with different const-ness.
Now it is declared in method.h with same const-ness to the
definition.
* object.c (rb_mod_module_exec): don't declare functions declared in
include/ruby/intern.h.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@32163 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
vm_insnhelper.c: use struct rb_constant_entry_t as entry of
RCLASS_CONST_TBL. RCLASS_CONST_TBL has contained VALUE of constant
directly. Now instead rb_const_entry_t is contained in
RCLASS_CONST_TBL, rb_const_entry_t is managed by malloc, and
have not only the value itself but also visibility flag.
This is another preparation for private constant (see
[ruby-dev:39685][ruby-core:32698]).
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@29602 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
added. Unlinked method entries are collected to
vm->unlinked_method_entry_list. On the GC timing, mark all method
entries which are on all living threads. Only non-marked method
entries are collected. This hack prevents releasing living method
entry.
[Performance Consideration] Since this Method Entry GC (MEGC)
doesn't occuer frequently, MEGC will not be a performance bottleneck.
However, to traverse living method entries, every control frame push
needs to clear cfp->me field. This will be a performance issue
(because pushing control frame is occurred frequently).
Bug #2777 [ruby-dev:40457]
* cont.c (fiber_init): init cfp->me.
* gc.c (garbage_collect): kick rb_sweep_method_entry().
* method.h (rb_method_entry_t): add a mark field.
* vm.c (invoke_block_from_c): set passed me.
* vm.c (rb_thread_mark): mark cfp->me.
* vm_core.h (rb_thread_t): add a field passed_me.
* vm_core.h (rb_vm_t): add a field unlinked_method_entry_list.
* vm_insnhelper.c (vm_push_frame): clear cfp->me at all times.
* vm_insnhelper.c (vm_call_bmethod): pass me.
* bootstraptest/test_method.rb: add a test.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@27634 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
#respond_to_missing? for not implemented methods.
[ruby-core:25909]
* vm_method.c (rb_method_boundp): returns exceptional value 2 for
not-implemented methods when called from #respond_to? (specifies
by new contant NOEX_RESPONDS).
* method.h (enum): new constant NOEX_RESPONDS added.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@25234 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* method.h (rb_get_method_entry): added a prototype of the function.
(rb_method_entry_without_cache): more friendly name.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@24702 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
method management. This change affect some VM control stack structure.
* vm.c, vm_insnhelper.c, vm_method.c, vm_eval.c: ditto. and make some
refactoring.
* insns.def, class.c, eval.c, proc.c, vm_dump.c : ditto.
* vm_core.h, compile.c (iseq_specialized_instruction): remove
VM_CALL_SEND_BIT. use another optimization tech for Kernel#send.
* node.h: remove unused node types.
* ext/objspace/objspace.c (count_nodes): ditto.
* gc.c: add mark/free functions for method entry.
* include/ruby/intern.h: remove decl of
rb_define_notimplement_method_id(). nobody can use it
because noex is not opend.
* iseq.c (iseq_mark): fix to check ic_method is available.
* iseq.c (rb_iseq_disasm): fix to use rb_method_get_iseq().
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@24128 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
charliesome
tmm1
shyouhei
nobu
ko1
shugo
nagachika
marcandre
akr
mame
wanabe
matz
naruse
yugui