This commit splits gc.c into two files:
- gc.c now only contains code not specific to Ruby GC. This includes
code to mark objects (which the GC implementation may choose not to
use) and wrappers for internal APIs that the implementation may need
to use (e.g. locking the VM).
- gc_impl.c now contains the implementation of Ruby's GC. This includes
marking, sweeping, compaction, and statistics. Most importantly,
gc_impl.c only uses public APIs in Ruby and a limited set of functions
exposed in gc.c. This allows us to build gc_impl.c independently of
Ruby and plug Ruby's GC into itself.
This patch optimizes forwarding callers and callees. It only optimizes methods that only take `...` as their parameter, and then pass `...` to other calls.
Calls it optimizes look like this:
```ruby
def bar(a) = a
def foo(...) = bar(...) # optimized
foo(123)
```
```ruby
def bar(a) = a
def foo(...) = bar(1, 2, ...) # optimized
foo(123)
```
```ruby
def bar(*a) = a
def foo(...)
list = [1, 2]
bar(*list, ...) # optimized
end
foo(123)
```
All variants of the above but using `super` are also optimized, including a bare super like this:
```ruby
def foo(...)
super
end
```
This patch eliminates intermediate allocations made when calling methods that accept `...`.
We can observe allocation elimination like this:
```ruby
def m
x = GC.stat(:total_allocated_objects)
yield
GC.stat(:total_allocated_objects) - x
end
def bar(a) = a
def foo(...) = bar(...)
def test
m { foo(123) }
end
test
p test # allocates 1 object on master, but 0 objects with this patch
```
```ruby
def bar(a, b:) = a + b
def foo(...) = bar(...)
def test
m { foo(1, b: 2) }
end
test
p test # allocates 2 objects on master, but 0 objects with this patch
```
How does it work?
-----------------
This patch works by using a dynamic stack size when passing forwarded parameters to callees.
The caller's info object (known as the "CI") contains the stack size of the
parameters, so we pass the CI object itself as a parameter to the callee.
When forwarding parameters, the forwarding ISeq uses the caller's CI to determine how much stack to copy, then copies the caller's stack before calling the callee.
The CI at the forwarded call site is adjusted using information from the caller's CI.
I think this description is kind of confusing, so let's walk through an example with code.
```ruby
def delegatee(a, b) = a + b
def delegator(...)
delegatee(...) # CI2 (FORWARDING)
end
def caller
delegator(1, 2) # CI1 (argc: 2)
end
```
Before we call the delegator method, the stack looks like this:
```
Executing Line | Code | Stack
---------------+---------------------------------------+--------
1| def delegatee(a, b) = a + b | self
2| | 1
3| def delegator(...) | 2
4| # |
5| delegatee(...) # CI2 (FORWARDING) |
6| end |
7| |
8| def caller |
-> 9| delegator(1, 2) # CI1 (argc: 2) |
10| end |
```
The ISeq for `delegator` is tagged as "forwardable", so when `caller` calls in
to `delegator`, it writes `CI1` on to the stack as a local variable for the
`delegator` method. The `delegator` method has a special local called `...`
that holds the caller's CI object.
Here is the ISeq disasm fo `delegator`:
```
== disasm: #<ISeq:delegator@-e:1 (1,0)-(1,39)>
local table (size: 1, argc: 0 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 1] "..."@0
0000 putself ( 1)[LiCa]
0001 getlocal_WC_0 "..."@0
0003 send <calldata!mid:delegatee, argc:0, FCALL|FORWARDING>, nil
0006 leave [Re]
```
The local called `...` will contain the caller's CI: CI1.
Here is the stack when we enter `delegator`:
```
Executing Line | Code | Stack
---------------+---------------------------------------+--------
1| def delegatee(a, b) = a + b | self
2| | 1
3| def delegator(...) | 2
-> 4| # | CI1 (argc: 2)
5| delegatee(...) # CI2 (FORWARDING) | cref_or_me
6| end | specval
7| | type
8| def caller |
9| delegator(1, 2) # CI1 (argc: 2) |
10| end |
```
The CI at `delegatee` on line 5 is tagged as "FORWARDING", so it knows to
memcopy the caller's stack before calling `delegatee`. In this case, it will
memcopy self, 1, and 2 to the stack before calling `delegatee`. It knows how much
memory to copy from the caller because `CI1` contains stack size information
(argc: 2).
Before executing the `send` instruction, we push `...` on the stack. The
`send` instruction pops `...`, and because it is tagged with `FORWARDING`, it
knows to memcopy (using the information in the CI it just popped):
```
== disasm: #<ISeq:delegator@-e:1 (1,0)-(1,39)>
local table (size: 1, argc: 0 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 1] "..."@0
0000 putself ( 1)[LiCa]
0001 getlocal_WC_0 "..."@0
0003 send <calldata!mid:delegatee, argc:0, FCALL|FORWARDING>, nil
0006 leave [Re]
```
Instruction 001 puts the caller's CI on the stack. `send` is tagged with
FORWARDING, so it reads the CI and _copies_ the callers stack to this stack:
```
Executing Line | Code | Stack
---------------+---------------------------------------+--------
1| def delegatee(a, b) = a + b | self
2| | 1
3| def delegator(...) | 2
4| # | CI1 (argc: 2)
-> 5| delegatee(...) # CI2 (FORWARDING) | cref_or_me
6| end | specval
7| | type
8| def caller | self
9| delegator(1, 2) # CI1 (argc: 2) | 1
10| end | 2
```
The "FORWARDING" call site combines information from CI1 with CI2 in order
to support passing other values in addition to the `...` value, as well as
perfectly forward splat args, kwargs, etc.
Since we're able to copy the stack from `caller` in to `delegator`'s stack, we
can avoid allocating objects.
I want to do this to eliminate object allocations for delegate methods.
My long term goal is to implement `Class#new` in Ruby and it uses `...`.
I was able to implement `Class#new` in Ruby
[here](https://github.com/ruby/ruby/pull/9289).
If we adopt the technique in this patch, then we can optimize allocating
objects that take keyword parameters for `initialize`.
For example, this code will allocate 2 objects: one for `SomeObject`, and one
for the kwargs:
```ruby
SomeObject.new(foo: 1)
```
If we combine this technique, plus implement `Class#new` in Ruby, then we can
reduce allocations for this common operation.
Co-Authored-By: John Hawthorn <john@hawthorn.email>
Co-Authored-By: Alan Wu <XrXr@users.noreply.github.com>
* Implement BitVector data structure for variable-length context encoding
* Rename method to make intent clearer
* Rename write_uint => push_uint to make intent clearer
* Implement debug trait for BitVector
* Fix bug in BitVector::read_uint_at(), enable more tests
* Add one more test for good measure
* Start sketching Context::encode()
* Progress on variable length context encoding
* Add tests. Fix bug.
* Encode stack state
* Add comments. Try to estimate context encoding size.
* More compact encoding for stack size
* Commit before rebase
* Change Context::encode() to take a BitVector as input
* Refactor BitVector::read_uint(), add helper read functions
* Implement Context::decode() function. Add test.
* Fix bug, add tests
* Rename methods
* Add Context::encode() and decode() methods using global data
* Make encode and decode methods use u32 indices
* Refactor YJIT to use variable-length context encoding
* Tag functions as allow unused
* Add a simple caching mechanism and stats for bytes per context etc
* Add comments, fix formatting
* Grow vector of bytes by 1.2x instead of 2x
* Add debug assert to check round-trip encoding-decoding
* Take some rustfmt formatting
* Add decoded_from field to Context to reuse previous encodings
* Remove olde context stats
* Re-add stack_size assert
* Disable decoded_from optimization for now
Previously, the update was done in the ISEQ callback. That effectively
never updated anything because the callback itself is given an intact
reference, so it could update its content, and `rb_gc_location(iseq)`
never returned a new address. Update the whole table once in the YJIT
root instead.
* Revert "Revert "YJIT: Optimize local variables when EP == BP" (#10584)"
This reverts commit c878344195.
* YJIT: Take care of GC references in ISEQ invariants
Co-authored-by: Alan Wu <alansi.xingwu@shopify.com>
---------
Co-authored-by: Alan Wu <alansi.xingwu@shopify.com>
This reverts commit 4cc58ea0b8.
Since the change landed call-threshold=1 CI runs have been timing out.
There has also been `verify-ctx` violations. Revert for now while we debug.
This `st_table` is used to both mark and pin classes
defined from the C API. But `vm->mark_object_ary` already
does both much more efficiently.
Currently a Ruby process starts with 252 rooted classes,
which uses `7224B` in an `st_table` or `2016B` in an `RArray`.
So a baseline of 5kB saved, but since `mark_object_ary` is
preallocated with `1024` slots but only use `405` of them,
it's a net `7kB` save.
`vm->mark_object_ary` is also being refactored.
Prior to this changes, `mark_object_ary` was a regular `RArray`, but
since this allows for references to be moved, it was marked a second
time from `rb_vm_mark()` to pin these objects.
This has the detrimental effect of marking these references on every
minors even though it's a mostly append only list.
But using a custom TypedData we can save from having to mark
all the references on minor GC runs.
Addtionally, immediate values are now ignored and not appended
to `vm->mark_object_ary` as it's just wasted space.
Usually we deal with splats by speculating that they're of a specific
size. In this case, the C method takes a pointer and a length, so
we can support changing sizes just fine.
This is the same optimization as e4272fd29 ("Avoid allocation when
passing no keywords to anonymous kwrest methods") but for YJIT. For
anonymous kwrest parameters, nil is just as good as an empty hash.
On the usage side, update `splatkw` to handle `nil` with a leaner path.
YJIT didn't guard for ruby2_keywords hash in case of splat calls that
land in methods with a rest parameter, creating incorrect results.
The compile-time checks didn't correspond to any actual effects of
ruby2_keywords, so it was masking this bug and YJIT was needlessly
refusing to compile some code. About 16% of fallback reasons in
`lobsters` was due to the ISeq check.
We already handle the tagging part with
exit_if_supplying_kw_and_has_no_kw() and should now have a dynamic guard
for all splat cases.
Note for backporting: You also need 7f51959ff1.
[Bug #20195]
```
warning: unused import: `condition::Condition`
--> src/asm/arm64/arg/mod.rs:13:9
|
13 | pub use condition::Condition;
| ^^^^^^^^^^^^^^^^^^^^
|
= note: `#[warn(unused_imports)]` on by default
warning: unused import: `rb_yjit_fix_mul_fix as rb_fix_mul_fix`
--> src/cruby.rs:188:9
|
188 | pub use rb_yjit_fix_mul_fix as rb_fix_mul_fix;
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
warning: unused import: `rb_insn_len as raw_insn_len`
--> src/cruby.rs:142:9
|
142 | pub use rb_insn_len as raw_insn_len;
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
= note: `#[warn(unused_imports)]` on by default
```
Make asm public so it stops warning about unused public stuff in there.
* Unify length field for embedded and heap strings
The length field is of the same type and position in RString for both
embedded and heap allocated strings, so we can unify it.
* Remove RSTRING_EMBED_LEN
yjit-trace-exits appends a synthetic sample for the instruction being
exited, but we didn't increment the size of the stack. Fixing this count
correctly lets us successfully generate a flamegraph from the exits.
I also replaced the line number for instructions with 0, as I don't
think the previous value had meaning.
Co-authored-by: Adam Hess <HParker@github.com>
* YJIT: Add --yjit-pause and RubyVM::YJIT.resume
This allows booting YJIT in a suspended state. We chose to add a new
command line option as opposed to simply allowing YJIT.resume to work
without any command line option because it allows for combining with
YJIT tuning command line options. It also simpifies implementation.
Paired with Kokubun and Maxime.
* Update yjit.rb
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
---------
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
If the previous instruction is not a leaf instruction, then the PC was
incremented before the instruction was ran (meaning the currently
executing instruction is actually the previous instruction), so we
should not increment the PC otherwise we will calculate the source
line for the next instruction.
This bug can be reproduced in the following script:
```
require "objspace"
ObjectSpace.trace_object_allocations_start
a =
1.0 / 0.0
p [ObjectSpace.allocation_sourceline(a), ObjectSpace.allocation_sourcefile(a)]
```
Which outputs: [4, "test.rb"]
This is incorrect because the object was allocated on line 10 and not
line 4. The behaviour is correct when we use a leaf instruction (e.g.
if we replaced `1.0 / 0.0` with `"hello"`), then the output is:
[10, "test.rb"].
[Bug #19456]
This dispatches to a c func for doing the dynamic lookup. I experimented with chain on the proc but wasn't able to detect which call sites would be monomorphic vs polymorphic. There is definitely room for optimization here, but it does reduce exits.