* Improve perfomance for Integer#size method [Feature #17135]
* re-run ci
* Let MJIT frame skip work for Integer#size
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
Performance is probably improved?
$ benchmark-driver -v --rbenv 'before --jit;after --jit' --repeat-count=12 --alternate --output=all benchmark.yml
before --jit: ruby 3.0.0dev (2020-11-27T04:37:47Z master 69e77e81dc) +JIT [x86_64-linux]
after --jit: ruby 3.0.0dev (2020-11-27T05:28:19Z master df6b05c6dd) +JIT [x86_64-linux]
last_commit=Set VM_FRAME_FLAG_FINISH at once
Calculating -------------------------------------
before --jit after --jit
Optcarrot Lan_Master.nes 80.89292998533379 82.19497327502751 fps
80.93130641142331 85.13943315260148
81.06214830270119 87.43757879797808
82.29172808453910 87.89942441487113
84.61206450455929 87.91309779491075
85.44545883567997 87.98026086648694
86.02923132404449 88.03081060383973
86.07411817365879 88.14650206137341
86.34348799602836 88.32791633649961
87.90257338977324 88.57599644892220
88.58006509876580 88.67426384743277
89.26611118140011 88.81669430874207
This should have no bad impact on VM because this function is ALWAYS_INLINE.
iv_index_tbl manages instance variable indexes (ID -> index).
This data structure should be synchronized with other ractors
so introduce some VM locks.
This patch also introduced atomic ivar cache used by
set/getinlinecache instructions. To make updating ivar cache (IVC),
we changed iv_index_tbl data structure to manage (ID -> entry)
and an entry points serial and index. IVC points to this entry so
that cache update becomes atomically.
for opt_* insns.
opt_eq handles rb_obj_equal inside opt_eq, and all other cfunc is
handled by opt_send_without_block. Therefore we can't decide which insn
should be generated by checking whether it's cfunc cc or not.
```
$ benchmark-driver -v --rbenv 'before --jit;after --jit' benchmark/mjit_opt_cc_insns.yml --repeat-count=4
before --jit: ruby 2.8.0dev (2020-06-26T05:21:43Z master 9dbc2294a6) +JIT [x86_64-linux]
after --jit: ruby 2.8.0dev (2020-06-26T06:30:18Z master 75cece1b0b) +JIT [x86_64-linux]
last_commit=Decide JIT-ed insn based on cached cfunc
Calculating -------------------------------------
before --jit after --jit
mjit_nil?(1) 73.878M 74.021M i/s - 40.000M times in 0.541432s 0.540391s
mjit_not(1) 72.635M 74.601M i/s - 40.000M times in 0.550702s 0.536187s
mjit_eq(1, nil) 7.331M 7.445M i/s - 8.000M times in 1.091211s 1.074596s
mjit_eq(nil, 1) 49.450M 64.711M i/s - 8.000M times in 0.161781s 0.123627s
Comparison:
mjit_nil?(1)
after --jit: 74020528.4 i/s
before --jit: 73878185.9 i/s - 1.00x slower
mjit_not(1)
after --jit: 74600882.0 i/s
before --jit: 72634507.6 i/s - 1.03x slower
mjit_eq(1, nil)
after --jit: 7444657.4 i/s
before --jit: 7331304.3 i/s - 1.02x slower
mjit_eq(nil, 1)
after --jit: 64710790.6 i/s
before --jit: 49449507.4 i/s - 1.31x slower
```
only for opt_nil_p and opt_not.
While vm_method_cfunc_is is used for opt_eq too, many fast paths of it
don't call it. So if it's populated, it should generate opt_send,
regardless of cfunc or not. And again, opt_neq isn't relevant due to the
difference in operands.
So opt_nil_p and opt_not are the only variants using vm_method_cfunc_is
like they use.
```
$ benchmark-driver -v --rbenv 'before2 --jit::ruby --jit;before --jit;after --jit' benchmark/mjit_opt_cc_insns.yml --repeat-count=4
before2 --jit: ruby 2.8.0dev (2020-06-22T08:37:37Z master 3238641750) +JIT [x86_64-linux]
before --jit: ruby 2.8.0dev (2020-06-23T01:01:24Z master 9ce2066209) +JIT [x86_64-linux]
after --jit: ruby 2.8.0dev (2020-06-23T06:58:37Z master 17e9df3157) +JIT [x86_64-linux]
last_commit=Avoid generating opt_send with cfunc cc with JIT
Calculating -------------------------------------
before2 --jit before --jit after --jit
mjit_nil?(1) 54.204M 75.536M 75.031M i/s - 40.000M times in 0.737947s 0.529548s 0.533110s
mjit_not(1) 53.822M 70.921M 71.920M i/s - 40.000M times in 0.743195s 0.564007s 0.556171s
mjit_eq(1, nil) 7.367M 6.496M 7.331M i/s - 8.000M times in 1.085882s 1.231470s 1.091327s
Comparison:
mjit_nil?(1)
before --jit: 75536059.3 i/s
after --jit: 75031409.4 i/s - 1.01x slower
before2 --jit: 54204431.6 i/s - 1.39x slower
mjit_not(1)
after --jit: 71920324.1 i/s
before --jit: 70921063.1 i/s - 1.01x slower
before2 --jit: 53821697.6 i/s - 1.34x slower
mjit_eq(1, nil)
before2 --jit: 7367280.0 i/s
after --jit: 7330527.4 i/s - 1.01x slower
before --jit: 6496302.8 i/s - 1.13x slower
```
because opt_nil/opt_not/opt_eq populates cc even when it doesn't
fallback to opt_send_without_block because of vm_method_cfunc_is.
```
$ benchmark-driver -v --rbenv 'before --jit;after --jit' benchmark/mjit_opt_cc_insns.yml --repeat-count=4
before --jit: ruby 2.8.0dev (2020-06-22T08:11:24Z master d231b8f95b) +JIT [x86_64-linux]
after --jit: ruby 2.8.0dev (2020-06-22T08:53:27Z master e1125879ed) +JIT [x86_64-linux]
last_commit=Compile opt_send for opt_* only when cc has ISeq
Calculating -------------------------------------
before --jit after --jit
mjit_nil?(1) 54.106M 73.693M i/s - 40.000M times in 0.739288s 0.542795s
mjit_not(1) 53.398M 74.477M i/s - 40.000M times in 0.749090s 0.537075s
mjit_eq(1, nil) 7.427M 6.497M i/s - 8.000M times in 1.077136s 1.231326s
Comparison:
mjit_nil?(1)
after --jit: 73692594.3 i/s
before --jit: 54106108.4 i/s - 1.36x slower
mjit_not(1)
after --jit: 74477487.9 i/s
before --jit: 53398125.0 i/s - 1.39x slower
mjit_eq(1, nil)
before --jit: 7427105.9 i/s
after --jit: 6497063.0 i/s - 1.14x slower
```
Actually opt_eq becomes slower by this. Maybe it's indeed using
opt_send_without_block, but I'll approach that one in another commit.
by calling combined functions specialized for each cancel type.
I'm hoping to improve locality of hot code, but this patch's impact should
be insignificant.
jit_unit to avoid marking wrong cc entries when inlined iseq is compiled
multiple times, resolving the TODO added by daf7c48d88.
This obviates pseudo jit_unit in inlined iseq introduced by 7ec2359374
and fixes memory leak of the adhoc unit.
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.
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).
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
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]
for ISeq including only leaf and no-handles_sp insns except leaf.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@67574 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Nobuyoshi Nakada
S.H
Takashi Kokubun
Koichi Sasada
卜部昌平
Alan Wu
k0kubun
usa