These show gains from the recent optimization commits:
```
arg_splat
miniruby: 7346039.9 i/s
miniruby-before: 4692240.8 i/s - 1.57x slower
arg_splat_block
miniruby: 6539749.6 i/s
miniruby-before: 4358063.6 i/s - 1.50x slower
splat_kw_splat
miniruby: 5433641.5 i/s
miniruby-before: 3851048.6 i/s - 1.41x slower
splat_kw_splat_block
miniruby: 4916137.1 i/s
miniruby-before: 3477090.1 i/s - 1.41x slower
splat_kw_block
miniruby: 2912829.5 i/s
miniruby-before: 2465611.7 i/s - 1.18x slower
arg_splat_post
miniruby: 2195208.2 i/s
miniruby-before: 1860204.3 i/s - 1.18x slower
```
zsuper only speeds up in the post argument case, because
it was already set to use splatarray false in cases where
there were no post arguments.
Thanks to the new semantics from [ruby-core:115808], `**nil` is now
equivalent to `**{}`. Since the only thing one could do with anonymous
keyword rest parameter is to delegate it with `**`, nil is just as good
as an empty hash. Using nil avoids allocating an empty hash.
This is particularly important for `...` methods since they now use
`**kwrest` under the hood after 4f77d8d328. Most calls don't pass
keywords.
Comparison:
fw_no_kw
post: 9816800.9 i/s
pre: 8570297.0 i/s - 1.15x slower
The following code previously caused a crash:
```ruby
h = {}
1000000.times{|i| h[i.to_s.to_sym] = i}
def f(kw: 1, **kws) end
f(**h)
```
Inside a thread or fiber, the size of the keyword splat could be much smaller
and still cause a crash.
I found this issue while optimizing method calling by reducing implicit
allocations. Given the following code:
```ruby
def f(kw: , **kws) end
kw = {kw: 1}
f(**kw)
```
The `f(**kw)` call previously allocated two hashes callee side instead of a
single hash. This is because `setup_parameters_complex` would extract the
keywords from the keyword splat hash to the C stack, to attempt to mirror
the case when literal keywords are passed without a keyword splat. Then,
`make_rest_kw_hash` would build a new hash based on the extracted keywords
that weren't used for literal keywords.
Switch the implementation so that if a keyword splat is passed, literal keywords
are deleted from the keyword splat hash (or a copy of the hash if the hash is
not mutable).
In addition to avoiding the crash, this new approach is much more
efficient in all cases. With the included benchmark:
```
1
miniruby: 5247879.9 i/s
miniruby-before: 2474050.2 i/s - 2.12x slower
1_mutable
miniruby: 1797036.5 i/s
miniruby-before: 1239543.3 i/s - 1.45x slower
10
miniruby: 1094750.1 i/s
miniruby-before: 365529.6 i/s - 2.99x slower
10_mutable
miniruby: 407781.7 i/s
miniruby-before: 225364.0 i/s - 1.81x slower
100
miniruby: 100992.3 i/s
miniruby-before: 32703.6 i/s - 3.09x slower
100_mutable
miniruby: 40092.3 i/s
miniruby-before: 21266.9 i/s - 1.89x slower
1000
miniruby: 21694.2 i/s
miniruby-before: 4949.8 i/s - 4.38x slower
1000_mutable
miniruby: 5819.5 i/s
miniruby-before: 2995.0 i/s - 1.94x slower
```
To avoid stack overflow, Ruby splits compilation of large arrays
into smaller arrays, and concatenates the small arrays together.
It previously used newarray/concatarray for this, which is
inefficient. This switches the compilation to use pushtoarray,
which is much faster. This makes almost all literal arrays only
allocate a single array.
For cases where there is a large amount of static values in the
array, Ruby will statically compile subarrays, and previously
added them using concatarray. This switches to concattoarray,
avoiding an array allocation for the append.
Keyword splats are also supported in arrays, and ignored if the
keyword splat is empty. Previously, this used newarraykwsplat and
concatarray. This still uses newarraykwsplat, but switches to
concattoarray to save an allocation. So large arrays with keyword
splats can allocate 2 arrays instead of 1.
Previously, for the following array sizes (assuming local variable
access for each element), Ruby allocated the following number of
arrays:
1000 elements: 7 arrays
10000 elements: 79 arrays
100000 elements: 781 arrays
With these changes, only a single array is allocated (or 2 for a
large array with a keyword splat.
Results using the included benchmark:
```
array_1000
miniruby: 34770.0 i/s
./miniruby-before: 10511.7 i/s - 3.31x slower
array_10000
miniruby: 4938.8 i/s
./miniruby-before: 483.8 i/s - 10.21x slower
array_100000
miniruby: 727.2 i/s
./miniruby-before: 4.1 i/s - 176.98x slower
```
Co-authored-by: Nobuyoshi Nakada <nobu@ruby-lang.org>
This follows the same approach used for attr_reader/attr_writer in
2d98593bf5, skipping the checking for
tracing after the first call using the call cache, and clearing the
call cache when tracing is turned on/off.
Fixes [Bug #18886]
`String#+@` is 2-3 times faster than `String#dup` because it can
directly go through `rb_str_dup` instead of using the generic
much slower `rb_obj_dup`.
This fact led to the existance of the ugly `Performance/UnfreezeString`
rubocop performance rule that encourage users to rewrite the much
more readable and convenient `"foo".dup` into the ugly `(+"foo")`.
Let's make that rubocop rule useless.
```
compare-ruby: ruby 3.3.0dev (2023-11-20T02:02:55Z master 701b0650de) [arm64-darwin22]
last_commit=[ruby/prism] feat: add encoding for IBM865 (https://github.com/ruby/prism/pull/1884)
built-ruby: ruby 3.3.0dev (2023-11-20T12:51:45Z faster-str-lit-dup 6b745bbc5d) [arm64-darwin22]
warming up..
| |compare-ruby|built-ruby|
|:------|-----------:|---------:|
|uplus | 16.312M| 16.332M|
| | -| 1.00x|
|dup | 5.912M| 16.329M|
| | -| 2.76x|
```
When an inline cache misses, it is very likely that the stale shape_id
and the current instance shape_id have a close common ancestor.
For example if the instance variable is sometimes frozen sometimes
not, one of the two shape will be the direct parent of the other.
Another pattern that commonly cause IC misses is "memoization",
in such case the object will have a "base common shape" and then
a number of close descendants.
In addition, when we find a common ancestor, we store it in the
inline cache instead of the current shape. This help prevent the
cache from flip-flopping, ensuring the next lookup will be marginally
faster and more generally avoid writing in memory too much.
However, now that shapes have an ancestors index, we only check
for a few ancestors before falling back to use the index.
So overall this change speeds up what is assumed to be the more common
case, but makes what is assumed to be the less common case a bit slower.
```
compare-ruby: ruby 3.3.0dev (2023-10-26T05:30:17Z master 701ca070b4) [arm64-darwin22]
built-ruby: ruby 3.3.0dev (2023-10-26T09:25:09Z shapes_double_sear.. a723a85235) [arm64-darwin22]
warming up......
| |compare-ruby|built-ruby|
|:------------------------------------|-----------:|---------:|
|vm_ivar_stable_shape | 11.672M| 11.679M|
| | -| 1.00x|
|vm_ivar_memoize_unstable_shape | 7.551M| 10.506M|
| | -| 1.39x|
|vm_ivar_memoize_unstable_shape_miss | 11.591M| 11.624M|
| | -| 1.00x|
|vm_ivar_unstable_undef | 9.037M| 7.981M|
| | 1.13x| -|
|vm_ivar_divergent_shape | 8.034M| 6.657M|
| | 1.21x| -|
|vm_ivar_divergent_shape_imbalanced | 10.471M| 9.231M|
| | 1.13x| -|
```
Co-Authored-By: John Hawthorn <john@hawthorn.email>
This is an experimental commit that uses a functional red-black tree to
create an index of the ancestor shapes. It uses an Okasaki style
functional red black tree:
https://www.cs.tufts.edu/comp/150FP/archive/chris-okasaki/redblack99.pdf
This tree is advantageous because:
* It offers O(n log n) insertions and O(n log n) lookups.
* It shares memory with previous "versions" of the tree
When we insert a node in the tree, only the parts of the tree that need
to be rebalanced are newly allocated. Parts of the tree that don't need
to be rebalanced are not reallocated, so "new trees" are able to share
memory with old trees. This is in contrast to a sorted set where we
would have to duplicate the set, and also resort the set on each
insertion.
I've added a new stat to RubyVM.stat so we can understand how the red
black tree increases.
On Range#bsearch for endless ranges, we try positions at `begin + 2**i` (i = 0, 1, 2, ...)
to find a point that satisfies a given condition.
Subsequently, we perform binary searching with the interval `[begin, begin + 2**n]`.
However, the interval `[begin + 2**(n-1), begin + 2**n]` is sufficient for binary search
because `begin + 2**(n-1)` does not satisfy the condition.
The same applies to beginless ranges.
Leave callers to convert byte index to char index, as well as
`rb_str_index`, so that `rb_str_rpartition` does not need to
re-convert char index to byte index.
In most of case `sort_by` works on primitive type.
Using `qsort_r` with function pointer is much slower than compare data directly.
I implement an intro sort which compare primitive data directly for `sort_by`.
We can even afford an O(n) type check before primitive data sort.
It still go faster.
CALLER_ARG_SPLAT is not necessary for method_missing. We just need
to unshift the method name into the arguments.
This optimizes all method_missing calls:
* mm(recv) ~9%
* mm(recv, *args) ~215% for args.length == 200
* mm(recv, *args, **kw) ~55% for args.length == 200
* mm(recv, **kw) ~22%
* mm(recv, kw: 1) ~100%
Note that empty argument splats do get slower with this approach,
by about 30-40%. Other than non-empty argument splats, other
argument splats are faster, with the speedup depending on the
number of arguments.
Similar to the bmethod/send optimization, this avoids using
CALLER_ARG_SPLAT if not necessary. As long as the receiver argument
can be shifted off, other arguments are passed through as-is.
This optimizes the following types of calls:
* symproc.(recv) ~5%
* symproc.(recv, *args) ~65% for args.length == 200
* symproc.(recv, *args, **kw) ~45% for args.length == 200
* symproc.(recv, **kw) ~30%
* symproc.(recv, kw: 1) ~100%
Note that empty argument splats do get slower with this approach,
by about 2-3%. This is probably because iseq argument setup is
slower for empty argument splats than CALLER_SETUP_ARG is. Other
than non-empty argument splats, other argument splats are faster,
with the speedup depending on the number of arguments.
The following types of calls are not optimized:
* symproc.(*args)
* symproc.(*args, **kw)
This is because the you cannot shift the receiver argument off
without first splatting the arg.
Similar to the bmethod optimization, this avoids using
CALLER_ARG_SPLAT if not necessary. As long as the method argument
can be shifted off, other arguments are passed through as-is.
This optimizes the following types of calls:
* send(meth, arg) ~5%
* send(meth, *args) ~75% for args.length == 200
* send(meth, *args, **kw) ~50% for args.length == 200
* send(meth, **kw) ~25%
* send(meth, kw: 1) ~115%
Note that empty argument splats do get slower with this approach,
by about 20%. This is probably because iseq argument setup is
slower for empty argument splats than CALLER_SETUP_ARG is. Other
than non-empty argument splats, other argument splats are faster,
with the speedup depending on the number of arguments.
The following types of calls are not optimized:
* send(*args)
* send(*args, **kw)
This is because the you cannot shift the method argument off
without first splatting the arg.
This optimizes the following calls:
* ~10-15% for f(*a) when a does not end with a flagged keywords hash
* ~10-15% for f(*a) when a ends with an empty flagged keywords hash
* ~35-40% for f(*a, **kw) if kw is empty
This still copies the array contents to the VM stack, but avoids some
overhead. It would be faster to use the array pointer directly,
but that could cause problems if the array was modified during
the call to the function. You could do that optimization for frozen
arrays, but as splatting frozen arrays is uncommon, and the speedup
is minimal (<5%), it doesn't seem worth it.
The vm_send_cfunc benchmark has been updated to test additional cfunc
call types, and the numbers above were taken from the benchmark results.
Currently, bmethod arguments are copied from the VM stack to the
C stack in vm_call_bmethod, then copied from the C stack to the VM
stack later in invoke_iseq_block_from_c. This is inefficient.
This adds vm_call_iseq_bmethod and vm_call_noniseq_bmethod.
vm_call_iseq_bmethod is an optimized method that skips stack
copies (though there is one copy to remove the receiver from
the stack), and avoids calling vm_call_bmethod_body,
rb_vm_invoke_bmethod, invoke_block_from_c_proc,
invoke_iseq_block_from_c, and vm_yield_setup_args.
Th vm_call_iseq_bmethod argument handling is similar to the
way normal iseq methods are called, and allows for similar
performance optimizations when using splats or keywords.
However, even in the no argument case it's still significantly
faster.
A benchmark is added for bmethod calling. In my environment,
it improves bmethod calling performance by 38-59% for simple
bmethod calls, and up to 180% for bmethod calls passing
literal keywords on both sides.
```
./miniruby-iseq-bmethod: 18159792.6 i/s
./miniruby-m: 13174419.1 i/s - 1.38x slower
bmethod_simple_1
./miniruby-iseq-bmethod: 15890745.4 i/s
./miniruby-m: 10008972.7 i/s - 1.59x slower
bmethod_simple_0_splat
./miniruby-iseq-bmethod: 13142804.3 i/s
./miniruby-m: 11168595.2 i/s - 1.18x slower
bmethod_simple_1_splat
./miniruby-iseq-bmethod: 12375791.0 i/s
./miniruby-m: 8491140.1 i/s - 1.46x slower
bmethod_no_splat
./miniruby-iseq-bmethod: 10151258.8 i/s
./miniruby-m: 8716664.1 i/s - 1.16x slower
bmethod_0_splat
./miniruby-iseq-bmethod: 8138802.5 i/s
./miniruby-m: 7515600.2 i/s - 1.08x slower
bmethod_1_splat
./miniruby-iseq-bmethod: 8028372.7 i/s
./miniruby-m: 5947658.6 i/s - 1.35x slower
bmethod_10_splat
./miniruby-iseq-bmethod: 6953514.1 i/s
./miniruby-m: 4840132.9 i/s - 1.44x slower
bmethod_100_splat
./miniruby-iseq-bmethod: 5287288.4 i/s
./miniruby-m: 2243218.4 i/s - 2.36x slower
bmethod_kw
./miniruby-iseq-bmethod: 8931358.2 i/s
./miniruby-m: 3185818.6 i/s - 2.80x slower
bmethod_no_kw
./miniruby-iseq-bmethod: 12281287.4 i/s
./miniruby-m: 10041727.9 i/s - 1.22x slower
bmethod_kw_splat
./miniruby-iseq-bmethod: 5618956.8 i/s
./miniruby-m: 3657549.5 i/s - 1.54x slower
```
Prior to this commit the `OPTIMIZED_CMP` macro relied on a method lookup
to determine whether `<=>` was overridden. The result of the lookup was
cached, but only for the duration of the specific method that
initialized the cmp_opt_data cache structure.
With this method lookup, `[x,y].max` is slower than doing `x > y ?
x : y` even though there's an optimized instruction for "new array max".
(John noticed somebody a proposed micro-optimization based on this fact
in https://github.com/mastodon/mastodon/pull/19903.)
```rb
a, b = 1, 2
Benchmark.ips do |bm|
bm.report('conditional') { a > b ? a : b }
bm.report('method') { [a, b].max }
bm.compare!
end
```
Before:
```
Comparison:
conditional: 22603733.2 i/s
method: 19820412.7 i/s - 1.14x (± 0.00) slower
```
This commit replaces the method lookup with a new CMP basic op, which
gives the examples above equivalent performance.
After:
```
Comparison:
method: 24022466.5 i/s
conditional: 23851094.2 i/s - same-ish: difference falls within
error
```
Relevant benchmarks show an improvement to Array#max and Array#min when
not using the optimized newarray_max instruction as well. They are
noticeably faster for small arrays with the relevant types, and the same
or maybe a touch faster on larger arrays.
```
$ make benchmark COMPARE_RUBY=<master@5958c305> ITEM=array_min
$ make benchmark COMPARE_RUBY=<master@5958c305> ITEM=array_max
```
The benchmarks added in this commit also look generally improved.
Co-authored-by: John Hawthorn <jhawthorn@github.com>
Previously YARV bytecode implemented constant caching by having a pair
of instructions, opt_getinlinecache and opt_setinlinecache, wrapping a
series of getconstant calls (with putobject providing supporting
arguments).
This commit replaces that pattern with a new instruction,
opt_getconstant_path, handling both getting/setting the inline cache and
fetching the constant on a cache miss.
This is implemented by storing the full constant path as a
null-terminated array of IDs inside of the IC structure. idNULL is used
to signal an absolute constant reference.
$ ./miniruby --dump=insns -e '::Foo::Bar::Baz'
== disasm: #<ISeq:<main>@-e:1 (1,0)-(1,13)> (catch: FALSE)
0000 opt_getconstant_path <ic:0 ::Foo::Bar::Baz> ( 1)[Li]
0002 leave
The motivation for this is that we had increasingly found the need to
disassemble the instructions between the opt_getinlinecache and
opt_setinlinecache in order to determine the constant we are fetching,
or otherwise store metadata.
This disassembly was done:
* In opt_setinlinecache, to register the IC against the constant names
it is using for granular invalidation.
* In rb_iseq_free, to unregister the IC from the invalidation table.
* In YJIT to find the position of a opt_getinlinecache instruction to
invalidate it when the cache is populated
* In YJIT to register the constant names being used for invalidation.
With this change we no longe need disassemly for these (in fact
rb_iseq_each is now unused), as the list of constant names being
referenced is held in the IC. This should also make it possible to make
more optimizations in the future.
This may also reduce the size of iseqs, as previously each segment
required 32 bytes (on 64-bit platforms) for each constant segment. This
implementation only stores one ID per-segment.
There should be no significant performance change between this and the
previous implementation. Previously opt_getinlinecache was a "leaf"
instruction, but it included a jump (almost always to a separate cache
line). Now opt_getconstant_path is a non-leaf (it may
raise/autoload/call const_missing) but it does not jump. These seem to
even out.
I'm planning to introduce mjit_compiler.rb, and I want to make this
consistent with it. Consistency with compile.c doesn't seem important
for MJIT anyway.
Takashi Kokubun
Jeremy Evans
Alan Wu
Jean Boussier
Aaron Patterson
Aaron Patterson
Nobuyoshi Nakada
Kouhei Yanagita
nekoyama32767
John Bampton
Matt Valentine-House
Daniel Colson
S.H
Samuel Williams
Jemma Issroff
HParker
Jean Boussier
John Hawthorn