Redo of 34a2acdac788602c14bf05fb616215187badd504 and
931138b00696419945dc03e10f033b1f53cd50f3 which were reverted.
GitHub PR #4340.
This change implements a cache for class variables. Previously there was
no cache for cvars. Cvar access is slow due to needing to travel all the
way up th ancestor tree before returning the cvar value. The deeper the
ancestor tree the slower cvar access will be.
The benefits of the cache are more visible with a higher number of
included modules due to the way Ruby looks up class variables. The
benchmark here includes 26 modules and shows with the cache, this branch
is 6.5x faster when accessing class variables.
```
compare-ruby: ruby 3.1.0dev (2021-03-15T06:22:34Z master 9e5105c) [x86_64-darwin19]
built-ruby: ruby 3.1.0dev (2021-03-15T12:12:44Z add-cache-for-clas.. c6be009) [x86_64-darwin19]
| |compare-ruby|built-ruby|
|:--------|-----------:|---------:|
|vm_cvar | 5.681M| 36.980M|
| | -| 6.51x|
```
Benchmark.ips calling `ActiveRecord::Base.logger` from within a Rails
application. ActiveRecord::Base.logger has 71 ancestors. The more
ancestors a tree has, the more clear the speed increase. IE if Base had
only one ancestor we'd see no improvement. This benchmark is run on a
vanilla Rails application.
Benchmark code:
```ruby
require "benchmark/ips"
require_relative "config/environment"
Benchmark.ips do |x|
x.report "logger" do
ActiveRecord::Base.logger
end
end
```
Ruby 3.0 master / Rails 6.1:
```
Warming up --------------------------------------
logger 155.251k i/100ms
Calculating -------------------------------------
```
Ruby 3.0 with cvar cache / Rails 6.1:
```
Warming up --------------------------------------
logger 1.546M i/100ms
Calculating -------------------------------------
logger 14.857M (± 4.8%) i/s - 74.198M in 5.006202s
```
Lastly we ran a benchmark to demonstate the difference between master
and our cache when the number of modules increases. This benchmark
measures 1 ancestor, 30 ancestors, and 100 ancestors.
Ruby 3.0 master:
```
Warming up --------------------------------------
1 module 1.231M i/100ms
30 modules 432.020k i/100ms
100 modules 145.399k i/100ms
Calculating -------------------------------------
1 module 12.210M (± 2.1%) i/s - 61.553M in 5.043400s
30 modules 4.354M (± 2.7%) i/s - 22.033M in 5.063839s
100 modules 1.434M (± 2.9%) i/s - 7.270M in 5.072531s
Comparison:
1 module: 12209958.3 i/s
30 modules: 4354217.8 i/s - 2.80x (± 0.00) slower
100 modules: 1434447.3 i/s - 8.51x (± 0.00) slower
```
Ruby 3.0 with cvar cache:
```
Warming up --------------------------------------
1 module 1.641M i/100ms
30 modules 1.655M i/100ms
100 modules 1.620M i/100ms
Calculating -------------------------------------
1 module 16.279M (± 3.8%) i/s - 82.038M in 5.046923s
30 modules 15.891M (± 3.9%) i/s - 79.459M in 5.007958s
100 modules 16.087M (± 3.6%) i/s - 81.005M in 5.041931s
Comparison:
1 module: 16279458.0 i/s
100 modules: 16087484.6 i/s - same-ish: difference falls within error
30 modules: 15891406.2 i/s - same-ish: difference falls within error
```
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
This change implements a cache for class variables. Previously there was
no cache for cvars. Cvar access is slow due to needing to travel all the
way up th ancestor tree before returning the cvar value. The deeper the
ancestor tree the slower cvar access will be.
The benefits of the cache are more visible with a higher number of
included modules due to the way Ruby looks up class variables. The
benchmark here includes 26 modules and shows with the cache, this branch
is 6.5x faster when accessing class variables.
```
compare-ruby: ruby 3.1.0dev (2021-03-15T06:22:34Z master 9e5105ca45) [x86_64-darwin19]
built-ruby: ruby 3.1.0dev (2021-03-15T12:12:44Z add-cache-for-clas.. c6be0093ae) [x86_64-darwin19]
| |compare-ruby|built-ruby|
|:--------|-----------:|---------:|
|vm_cvar | 5.681M| 36.980M|
| | -| 6.51x|
```
Benchmark.ips calling `ActiveRecord::Base.logger` from within a Rails
application. ActiveRecord::Base.logger has 71 ancestors. The more
ancestors a tree has, the more clear the speed increase. IE if Base had
only one ancestor we'd see no improvement. This benchmark is run on a
vanilla Rails application.
Benchmark code:
```ruby
require "benchmark/ips"
require_relative "config/environment"
Benchmark.ips do |x|
x.report "logger" do
ActiveRecord::Base.logger
end
end
```
Ruby 3.0 master / Rails 6.1:
```
Warming up --------------------------------------
logger 155.251k i/100ms
Calculating -------------------------------------
```
Ruby 3.0 with cvar cache / Rails 6.1:
```
Warming up --------------------------------------
logger 1.546M i/100ms
Calculating -------------------------------------
logger 14.857M (± 4.8%) i/s - 74.198M in 5.006202s
```
Lastly we ran a benchmark to demonstate the difference between master
and our cache when the number of modules increases. This benchmark
measures 1 ancestor, 30 ancestors, and 100 ancestors.
Ruby 3.0 master:
```
Warming up --------------------------------------
1 module 1.231M i/100ms
30 modules 432.020k i/100ms
100 modules 145.399k i/100ms
Calculating -------------------------------------
1 module 12.210M (± 2.1%) i/s - 61.553M in 5.043400s
30 modules 4.354M (± 2.7%) i/s - 22.033M in 5.063839s
100 modules 1.434M (± 2.9%) i/s - 7.270M in 5.072531s
Comparison:
1 module: 12209958.3 i/s
30 modules: 4354217.8 i/s - 2.80x (± 0.00) slower
100 modules: 1434447.3 i/s - 8.51x (± 0.00) slower
```
Ruby 3.0 with cvar cache:
```
Warming up --------------------------------------
1 module 1.641M i/100ms
30 modules 1.655M i/100ms
100 modules 1.620M i/100ms
Calculating -------------------------------------
1 module 16.279M (± 3.8%) i/s - 82.038M in 5.046923s
30 modules 15.891M (± 3.9%) i/s - 79.459M in 5.007958s
100 modules 16.087M (± 3.6%) i/s - 81.005M in 5.041931s
Comparison:
1 module: 16279458.0 i/s
100 modules: 16087484.6 i/s - same-ish: difference falls within error
30 modules: 15891406.2 i/s - same-ish: difference falls within error
```
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
This commit introduces Ractor mechanism to run Ruby program in
parallel. See doc/ractor.md for more details about Ractor.
See ticket [Feature #17100] to see the implementation details
and discussions.
[Feature #17100]
This commit does not complete the implementation. You can find
many bugs on using Ractor. Also the specification will be changed
so that this feature is experimental. You will see a warning when
you make the first Ractor with `Ractor.new`.
I hope this feature can help programmers from thread-safety issues.
According to MSVC manual (*1), cl.exe can skip including a header file
when that:
- contains #pragma once, or
- starts with #ifndef, or
- starts with #if ! defined.
GCC has a similar trick (*2), but it acts more stricter (e. g. there
must be _no tokens_ outside of #ifndef...#endif).
Sun C lacked #pragma once for a looong time. Oracle Developer Studio
12.5 finally implemented it, but we cannot assume such recent version.
This changeset modifies header files so that each of them include
strictly one #ifndef...#endif. I believe this is the most portable way
to trigger compiler optimizations. [Bug #16770]
*1: https://docs.microsoft.com/en-us/cpp/preprocessor/once
*2: https://gcc.gnu.org/onlinedocs/cppinternals/Guard-Macros.html
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
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.
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.
Avoids genereating a "throwaway" sentinel class serial. There wasn't any
read harm in doing so (we're at no risk of exhaustion and there'd be no
measurable performance impact), but if feels cleaner that all class
serials actually end up assigned and used (especially now that we won't
overwrite them in a single method definition).
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.
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]
As `struct vm_throw_data::throw_state` is initialized as `VALUE`
by rb_imemo_new, extended MSW part is assigned to it on LP64
big-endian platforms.
Fix up 1feda1c2b0
Same as last commit, make some fields `const`.
include/ruby/ruby.h:
* Rasic::klass
* RArray::heap::aux::shared_root
* RRegexp::src
internal.h:
* rb_classext_struct::origin_, redefined_class
* vm_svar::cref_or_me, lastline, backref, others
* vm_throw_data::throw_obj
* vm_ifunc::data
* MEMO::v1, v2, u3::value
While modifying this patch, I found write-barrier miss on
rb_classext_struct::redefined_class.
Also vm_throw_data::throw_state is only `int` so change the type.
In addition to detect dead canary, we try to detect the very moment
when we smash the stack top. Requested by k0kubun:
https://twitter.com/k0kubun/status/1085180749899194368
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@66981 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
This changeset modifies the VM generator so that vm.inc is written in
C99. Also added some comments in _insn_entry.erb so that the
intention of each parts to be made more clear. I think this improves
overall readability of the generated VM.
Confirmed that the exact same binary is generated before/after this
changeset.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@66923 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
- FIXNUM_2_P: moved to vm_insnhelper.c because that is the only
place this macro is used.
- FLONUM_2_P: ditto.
- FLOAT_HEAP_P: not used anywhere.
- FLOAT_INSTANCE_P: ditto.
- GET_TOS: ditto.
- USE_IC_FOR_SPECIALIZED_METHOD: ditto.
- rb_obj_hidden_p: ditto.
- REG_A: ditto.
- REG_B: ditto.
- GET_CONST_INLINE_CACHE: ditto.
- vm_regan_regtype: moved inside of VM_COLLECT_USAGE_DETAILS
because that os the only place this enum is used.
- vm_regan_acttype: ditto.
- GET_GLOBAL: used only once. Removed with replacing that usage.
- SET_GLOBAL: ditto.
- rb_method_definition_create: declaration moved to
vm_insnhelper.c because that is the only place this declaration
makes sense.
- rb_method_definition_set: ditto.
- rb_method_definition_eq: ditto.
- rb_make_no_method_exception: ditto.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@66597 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
because it's actually setting fastpath to cc instead of ci since r51903.
vm_insnhelper.c: ditto
mjit_compile.c: ditto
tool/ruby_vm/views/_mjit_compile_send.erb: ditto
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64772 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Now that we can say for sure if an instruction calls a method or
not internally, it is now possible to reroute the bugs that
forced us to revert the "move PC around" optimization.
First try: r62051
Reverted: r63763
See also: r63999
----
trunk: ruby 2.6.0dev (2018-09-13 trunk 64736) [x86_64-darwin15]
ours: ruby 2.6.0dev (2018-09-13 trunk 64736) [x86_64-darwin15]
last_commit=move ADD_PC around (take 2)
Calculating -------------------------------------
trunk ours
so_ackermann 1.884 2.278 i/s - 1.000 times in 0.530926s 0.438935s
so_array 1.178 1.157 i/s - 1.000 times in 0.848786s 0.864467s
so_binary_trees 0.176 0.177 i/s - 1.000 times in 5.683895s 5.657707s
so_concatenate 0.220 0.221 i/s - 1.000 times in 4.546896s 4.518949s
so_count_words 6.729 6.470 i/s - 1.000 times in 0.148602s 0.154561s
so_exception 3.324 3.688 i/s - 1.000 times in 0.300872s 0.271147s
so_fannkuch 0.546 0.968 i/s - 1.000 times in 1.831328s 1.033376s
so_fasta 0.541 0.547 i/s - 1.000 times in 1.849923s 1.827091s
so_k_nucleotide 0.800 0.777 i/s - 1.000 times in 1.250635s 1.286295s
so_lists 2.101 1.848 i/s - 1.000 times in 0.475954s 0.541095s
so_mandelbrot 0.435 0.408 i/s - 1.000 times in 2.299328s 2.450535s
so_matrix 1.946 1.912 i/s - 1.000 times in 0.513872s 0.523076s
so_meteor_contest 0.311 0.317 i/s - 1.000 times in 3.219297s 3.152052s
so_nbody 0.746 0.703 i/s - 1.000 times in 1.339815s 1.423441s
so_nested_loop 0.899 0.901 i/s - 1.000 times in 1.111767s 1.109555s
so_nsieve 0.559 0.579 i/s - 1.000 times in 1.787763s 1.726552s
so_nsieve_bits 0.435 0.428 i/s - 1.000 times in 2.296282s 2.333852s
so_object 1.368 1.442 i/s - 1.000 times in 0.731237s 0.693684s
so_partial_sums 0.616 0.546 i/s - 1.000 times in 1.623592s 1.833097s
so_pidigits 0.831 0.832 i/s - 1.000 times in 1.203117s 1.202334s
so_random 2.934 2.724 i/s - 1.000 times in 0.340791s 0.367150s
so_reverse_complement 0.583 0.866 i/s - 1.000 times in 1.714144s 1.154615s
so_sieve 1.829 2.081 i/s - 1.000 times in 0.546607s 0.480562s
so_spectralnorm 0.524 0.558 i/s - 1.000 times in 1.908716s 1.792382s
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64737 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
because cc->call is NULL by default and it is not overridden by
vm_search_super_method if OPT_CALL_FASTPATH is 0. So this macro is not
just a switch for optimization but now it's mandatory.
vm_insnhelper.c: cosmetic change. Use boolean-ish `TRUE` instead of 1 to
specify `enabled` flag.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64735 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
This reverts commit r64711, because EXEC_EC_CFP on JIT-ed code does not
call jit_func with the patch when catch_except_p is true. It wasn't intentional.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64730 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
This is mostly cosmetic. Should generate a slightly readable
vm.inc output.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64709 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
By this commit's changes in other files, now MJIT started to work on VC++.
Unfortunately some features are still broken and they'll be fixed later.
This also suppresses cl.exe's default output to stdout because there
seems to be no option to do it. Tweaking some log messages as well.
vm_core.h: declare `__declspec(dllimport)` to export them correctly on mswin.
vm_insnhelper.h: ditto
mjit.h: ditto
test_jit.rb: skipped some pending tests.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64221 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* vm_insnhelper.h (PUSH): r63763 increase "PUSH()" op and it reveals that
there are hidden objects (at least T_IMEMO/iseq objects) are located
on VM stack. Remove this check and I'll revisit it later.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@63765 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
I introduced this mechanism in r62051 to speed things up. Later it
was reported that the change causes problems. I searched for
workarounds but nothing seemed appropriate. I hereby officially
give it up. The idea to move ADD_PC around was a mistake.
Fixes [Bug #14809] and [Bug #14834].
Signed-off-by: Urabe, Shyouhei <shyouhei@ruby-lang.org>
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@63763 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* vm_insnhelper.h (PUSH): hidden objects (klass == 0) should not be pushed
to a VM value stack. Add assertion for it.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@63687 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
if catch_except_p is FALSE. If catch_except_p is TRUE, stack values
should be on VM's stack when exception is thrown and the JIT-ed frame
is re-executed by VM's exception handler. If it's FALSE, the JIT-ed
frame won't be re-executed and don't need to keep values on VM's stack.
Using local variables allows us to reduce cfp->sp motion. Moving cfp->sp
is needed only for insns whose handles_frame? is false. So it improves
performance.
_mjit_compile_insn.erb: Prepare `stack_size` variable for GET_SP,
STACK_ADDR_FROM_TOP, TOPN macros. Share pc and sp motion partial view.
Use cancel handler created in mjit_compile.c.
_mjit_compile_send.erb: ditto. Also, when iseq->body->catch_except_p is
TRUE, this stops to call mjit_exec directly. I described the reason in
vm_insnhelper.h's comment for EXEC_EC_CFP.
_mjit_compile_pc_and_sp.erb: Shared logic for moving sp and pc. As you
can see from thsi file, when status->local_stack_p is TRUE and
insn.handles_frame? is false, moving sp is skipped. But if
insn.handles_frame? is true, values should be rolled back to VM's stack.
common.mk: add dependency for the file
_mjit_compile_insn_body.erb: Set sp value before canceling JIT on
DISPATCH_ORIGINAL_INSN. Replace GET_SP, STACK_ADDR_FROM_TOP, TOPN macros
for the case ocal_stack_p is TRUE and insn.handles_frame? is false.
In that case, values are not available on VM's stack and those macros
should be replaced.
mjit_compile.inc.erb: updated comments of macros which are supported by
JIT compiler. All references to `cfp->sp` should be replaced and thus
INC_SP, SET_SV, PUSH are no longer supported for now, because they are
not used now.
vm_exec.h: moved EXEC_EC_CFP definition to vm_insnhelper.h because it's
tighly coupled to CALL_METHOD.
vm_insnhelper.h: Have revised EXEC_EC_CFP definition moved from vm_exec.h.
Now it triggers mjit_exec for VM, and has the guard for catch_except_p
on JIT-ed code. See comments for details. CALL_METHOD delegates
triggering mjit_exec to EXEC_EC_CFP.
insns.def: Stopped using EXEC_EC_CFP for the case we don't want to
trigger mjit_exec. Those insns (defineclass, opt_call_c_function) are
not supported by JIT and it's safe to use RESTORE_REGS(), NEXT_INSN().
expandarray is changed to pass GET_SP() to replace the macro in
_mjit_compile_insn_body.erb.
vm_insnhelper.c: change to take sp for the above reason.
[close https://github.com/ruby/ruby/pull/1828]
This patch resurrects the performance which was attached in
[Feature #14235].
* Benchmark
Optcarrot (with configuration for benchmark_driver.gem)
https://github.com/benchmark-driver/optcarrot
$ benchmark-driver benchmark.yml --verbose 1 --rbenv 'before;before+JIT::before,--jit;after;after+JIT::after,--jit' --repeat-count 10
before: ruby 2.6.0dev (2018-03-04 trunk 62652) [x86_64-linux]
before+JIT: ruby 2.6.0dev (2018-03-04 trunk 62652) +JIT [x86_64-linux]
after: ruby 2.6.0dev (2018-03-04 local-variable.. 62652) [x86_64-linux]
last_commit=mjit_compile.c: use local variables for stack
after+JIT: ruby 2.6.0dev (2018-03-04 local-variable.. 62652) +JIT [x86_64-linux]
last_commit=mjit_compile.c: use local variables for stack
Calculating -------------------------------------
before before+JIT after after+JIT
optcarrot 53.552 59.680 53.697 63.358 fps
Comparison:
optcarrot
after+JIT: 63.4 fps
before+JIT: 59.7 fps - 1.06x slower
after: 53.7 fps - 1.18x slower
before: 53.6 fps - 1.18x slower
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@62655 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
(IS_ARGS_SPLAT, IS_ARGS_KEYWORD) with vm_args.c.
vm_args.c: share them with mjit_compile.c.
vm_insnhelper.h: get those definitions, with CALLER_SETUP_ARG too
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@62254 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
which has been developed by Takashi Kokubun <takashikkbn@gmail> as
YARV-MJIT. Many of its bugs are fixed by wanabe <s.wanabe@gmail.com>.
This JIT compiler is designed to be a safe migration path to introduce
JIT compiler to MRI. So this commit does not include any bytecode
changes or dynamic instruction modifications, which are done in original
MJIT.
This commit even strips off some aggressive optimizations from
YARV-MJIT, and thus it's slower than YARV-MJIT too. But it's still
fairly faster than Ruby 2.5 in some benchmarks (attached below).
Note that this JIT compiler passes `make test`, `make test-all`, `make
test-spec` without JIT, and even with JIT. Not only it's perfectly safe
with JIT disabled because it does not replace VM instructions unlike
MJIT, but also with JIT enabled it stably runs Ruby applications
including Rails applications.
I'm expecting this version as just "initial" JIT compiler. I have many
optimization ideas which are skipped for initial merging, and you may
easily replace this JIT compiler with a faster one by just replacing
mjit_compile.c. `mjit_compile` interface is designed for the purpose.
common.mk: update dependencies for mjit_compile.c.
internal.h: declare `rb_vm_insn_addr2insn` for MJIT.
vm.c: exclude some definitions if `-DMJIT_HEADER` is provided to
compiler. This avoids to include some functions which take a long time
to compile, e.g. vm_exec_core. Some of the purpose is achieved in
transform_mjit_header.rb (see `IGNORED_FUNCTIONS`) but others are
manually resolved for now. Load mjit_helper.h for MJIT header.
mjit_helper.h: New. This is a file used only by JIT-ed code. I'll
refactor `mjit_call_cfunc` later.
vm_eval.c: add some #ifdef switches to skip compiling some functions
like Init_vm_eval.
win32/mkexports.rb: export thread/ec functions, which are used by MJIT.
include/ruby/defines.h: add MJIT_FUNC_EXPORTED macro alis to clarify
that a function is exported only for MJIT.
array.c: export a function used by MJIT.
bignum.c: ditto.
class.c: ditto.
compile.c: ditto.
error.c: ditto.
gc.c: ditto.
hash.c: ditto.
iseq.c: ditto.
numeric.c: ditto.
object.c: ditto.
proc.c: ditto.
re.c: ditto.
st.c: ditto.
string.c: ditto.
thread.c: ditto.
variable.c: ditto.
vm_backtrace.c: ditto.
vm_insnhelper.c: ditto.
vm_method.c: ditto.
I would like to improve maintainability of function exports, but I
believe this way is acceptable as initial merging if we clarify the
new exports are for MJIT (so that we can use them as TODO list to fix)
and add unit tests to detect unresolved symbols.
I'll add unit tests of JIT compilations in succeeding commits.
Author: Takashi Kokubun <takashikkbn@gmail.com>
Contributor: wanabe <s.wanabe@gmail.com>
Part of [Feature #14235]
---
* Known issues
* Code generated by gcc is faster than clang. The benchmark may be worse
in macOS. Following benchmark result is provided by gcc w/ Linux.
* Performance is decreased when Google Chrome is running
* JIT can work on MinGW, but it doesn't improve performance at least
in short running benchmark.
* Currently it doesn't perform well with Rails. We'll try to fix this
before release.
---
* Benchmark reslts
Benchmarked with:
Intel 4.0GHz i7-4790K with 16GB memory under x86-64 Ubuntu 8 Cores
- 2.0.0-p0: Ruby 2.0.0-p0
- r62186: Ruby trunk (early 2.6.0), before MJIT changes
- JIT off: On this commit, but without `--jit` option
- JIT on: On this commit, and with `--jit` option
** Optcarrot fps
Benchmark: https://github.com/mame/optcarrot
| |2.0.0-p0 |r62186 |JIT off |JIT on |
|:--------|:--------|:--------|:--------|:--------|
|fps |37.32 |51.46 |51.31 |58.88 |
|vs 2.0.0 |1.00x |1.38x |1.37x |1.58x |
** MJIT benchmarks
Benchmark: https://github.com/benchmark-driver/mjit-benchmarks
(Original: https://github.com/vnmakarov/ruby/tree/rtl_mjit_branch/MJIT-benchmarks)
| |2.0.0-p0 |r62186 |JIT off |JIT on |
|:----------|:--------|:--------|:--------|:--------|
|aread |1.00 |1.09 |1.07 |2.19 |
|aref |1.00 |1.13 |1.11 |2.22 |
|aset |1.00 |1.50 |1.45 |2.64 |
|awrite |1.00 |1.17 |1.13 |2.20 |
|call |1.00 |1.29 |1.26 |2.02 |
|const2 |1.00 |1.10 |1.10 |2.19 |
|const |1.00 |1.11 |1.10 |2.19 |
|fannk |1.00 |1.04 |1.02 |1.00 |
|fib |1.00 |1.32 |1.31 |1.84 |
|ivread |1.00 |1.13 |1.12 |2.43 |
|ivwrite |1.00 |1.23 |1.21 |2.40 |
|mandelbrot |1.00 |1.13 |1.16 |1.28 |
|meteor |1.00 |2.97 |2.92 |3.17 |
|nbody |1.00 |1.17 |1.15 |1.49 |
|nest-ntimes|1.00 |1.22 |1.20 |1.39 |
|nest-while |1.00 |1.10 |1.10 |1.37 |
|norm |1.00 |1.18 |1.16 |1.24 |
|nsvb |1.00 |1.16 |1.16 |1.17 |
|red-black |1.00 |1.02 |0.99 |1.12 |
|sieve |1.00 |1.30 |1.28 |1.62 |
|trees |1.00 |1.14 |1.13 |1.19 |
|while |1.00 |1.12 |1.11 |2.41 |
** Discourse's script/bench.rb
Benchmark: https://github.com/discourse/discourse/blob/v1.8.7/script/bench.rb
NOTE: Rails performance was somehow a little degraded with JIT for now.
We should fix this.
(At least I know opt_aref is performing badly in JIT and I have an idea
to fix it. Please wait for the fix.)
*** JIT off
Your Results: (note for timings- percentile is first, duration is second in millisecs)
categories_admin:
50: 17
75: 18
90: 22
99: 29
home_admin:
50: 21
75: 21
90: 27
99: 40
topic_admin:
50: 17
75: 18
90: 22
99: 32
categories:
50: 35
75: 41
90: 43
99: 77
home:
50: 39
75: 46
90: 49
99: 95
topic:
50: 46
75: 52
90: 56
99: 101
*** JIT on
Your Results: (note for timings- percentile is first, duration is second in millisecs)
categories_admin:
50: 19
75: 21
90: 25
99: 33
home_admin:
50: 24
75: 26
90: 30
99: 35
topic_admin:
50: 19
75: 20
90: 25
99: 30
categories:
50: 40
75: 44
90: 48
99: 76
home:
50: 42
75: 48
90: 51
99: 89
topic:
50: 49
75: 55
90: 58
99: 99
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@62197 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
eileencodes
Takashi Kokubun
Aaron Patterson
Nobuyoshi Nakada
Koichi Sasada
Kenta Murata
卜部昌平
Jeremy Evans
John Hawthorn
Alan Wu
k0kubun
shyouhei
ko1
nobu