Rather than exposing that an imemo has a flag and four fields, this
changes the implementation to only expose one field (the klass) and
fills the rest with 0. The type will have to fill in the values themselves.
Before object shapes, we were using class serial to invalidate
inline caches. Now that we use shape_id for inline cache keys,
the class serial is unnecessary.
Co-Authored-By: Aaron Patterson <tenderlove@ruby-lang.org>
Before the new constant cache behavior, caches were invalidated by a
single global variable. You could inspect the value of this variable
with RubyVM.stat(:global_constant_state). This was mostly useful to
verify the behavior of the VM or to test constant loading like in Rails.
With the new constant cache behavior, we introduced
RubyVM.stat(:constant_cache) which returned a hash with symbol keys and
integer values that represented the number of live constant caches
associated with the given symbol. Additionally, we removed the old
RubyVM.stat(:global_constant_state).
This was proven to be not very useful, so it doesn't help you diagnose
constant loading issues. So, instead we added the global constant state
back into the RubyVM output. However, that number can be misleading as
now when you invalidate something like `Foo::Bar::Baz` you're actually
invalidating 3 different lists of inline caches.
This commit attempts to get the best of both worlds. We remove
RubyVM.stat(:global_constant_state) like we did originally, as it
doesn't have the same semantic meaning and it could be confusing going
forward. Instead we add RubyVM.stat(:constant_cache_invalidations) and
RubyVM.stat(:constant_cache_misses). These two metrics should provide
enough information to diagnose any constant loading issues, as well as
provide a replacement for the old global constant state.
This was removed as part of [Feature #18589]. But some applications were relying on this behavior. So bringing this back to make it better for backward compatibility going forward.
This commit reintroduces finer-grained constant cache invalidation.
After 8008fb7 got merged, it was causing issues on token-threaded
builds (such as on Windows).
The issue was that when you're iterating through instruction sequences
and using the translator functions to get back the instruction structs,
you're either using `rb_vm_insn_null_translator` or
`rb_vm_insn_addr2insn2` depending if it's a direct-threading build.
`rb_vm_insn_addr2insn2` does some normalization to always return to
you the non-trace version of whatever instruction you're looking at.
`rb_vm_insn_null_translator` does not do that normalization.
This means that when you're looping through the instructions if you're
trying to do an opcode comparison, it can change depending on the type
of threading that you're using. This can be very confusing. So, this
commit creates a new translator function
`rb_vm_insn_normalizing_translator` to always return the non-trace
version so that opcode comparisons don't have to worry about different
configurations.
[Feature #18589]
This reverts commits for [Feature #18589]:
* 8008fb7352
"Update formatting per feedback"
* 8f6eaca2e1
"Delete ID from constant cache table if it becomes empty on ISEQ free"
* 629908586b
"Finer-grained inline constant cache invalidation"
MSWin builds on AppVeyor have been crashing since the merger.
Current behavior - caches depend on a global counter. All constant mutations cause caches to be invalidated.
```ruby
class A
B = 1
end
def foo
A::B # inline cache depends on global counter
end
foo # populate inline cache
foo # hit inline cache
C = 1 # global counter increments, all caches are invalidated
foo # misses inline cache due to `C = 1`
```
Proposed behavior - caches depend on name components. Only constant mutations with corresponding names will invalidate the cache.
```ruby
class A
B = 1
end
def foo
A::B # inline cache depends constants named "A" and "B"
end
foo # populate inline cache
foo # hit inline cache
C = 1 # caches that depend on the name "C" are invalidated
foo # hits inline cache because IC only depends on "A" and "B"
```
Examples of breaking the new cache:
```ruby
module C
# Breaks `foo` cache because "A" constant is set and the cache in foo depends
# on "A" and "B"
class A; end
end
B = 1
```
We expect the new cache scheme to be invalidated less often because names aren't frequently reused. With the cache being invalidated less, we can rely on its stability more to keep our constant references fast and reduce the need to throw away generated code in YJIT.
The interpreter instruction count was enabled based on RUBY_DEBUG as
opposed to YJIT_STATS. In builds with YJIT_STATS=1 but RUBY_DEBUG=0,
the count was not available.
Move YJIT_STATS in yjit.h where declarations are expoed to code outside
of YJIT. Also reduce the changes made to the interpreter for calling
into YJIT's instruction counting function.
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.
Aaron Patterson
Peter Zhu
Takashi Kokubun
Jemma Issroff
John Hawthorn
Kevin Newton
Nobuyoshi Nakada
Alan Wu
Jose Narvaez
eileencodes
Koichi Sasada
Kenta Murata
卜部昌平
Jeremy Evans