* YJIT: Allow dev_nodebug to disasm release-mode code
* Revert "YJIT: Squash canary before falling back"
This reverts commit f05ad373d8.
The stray canary issue should have been solved by
def7023ee4, alleviating this codegen
accommodation.
* s/runtime_assertions/runtime_checks/
---------
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* YJIT: Local variable register allocation
* locals are not stack temps
* Rename RegTemps to RegMappings
* Rename RegMapping to RegOpnd
* Rename local_size to num_locals
* s/stack value/operand/
* Rename spill_temps() to spill_regs()
* Clarify when num_locals becomes None
* Mention that InsnOut uses different registers
* Rename get_reg_mapping to get_reg_opnd
* Resurrect --yjit-temp-regs capability
* Use MAX_CTX_TEMPS and MAX_CTX_LOCALS
* YJIT: increase context cache size to 1024 redux
* Move context hashing code outside of unsafe block
* Avoid allocating large table on the stack, which would cause a stack overflow
Co-authored by Alan Wu @XrXr
* YJIT: increase context cache size to 1024
The other day I ran into a mysterious bug while increasing the
cache size to 1024. I was not able to reproduce this locally.
Opening this PR for testing/debugging.
* Add extra debug assertions
* Add more comments to context code
* Update yjit/src/core.rs
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* Update yjit/src/core.rs
* Comment out potentially problematic assertion
* Revert cache size to 512 so we can merge other changes
---------
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
This change implements a fallback mode for the `--yjit-dump-disasm`
development command-line option to make it usable in release builds.
Previously, using the option with release builds of YJIT yielded only
a warning asking the user to build with `--enable-yjit=dev`.
While builds that use the `disasm` feature still give the best output,
just having the comments is useful enough for many kinds of debugging.
Having it usable in release builds is nice for new hackers, too, since
this allows for tinkering without having to learn how to build YJIT in
development mode.
Sample output on A64:
```
# regenerate_branch
# Insn: 0001 opt_send_without_block (stack_size: 1)
# guard known object with singleton class
0x11f7e0034: 4b 00 00 58 03 00 00 14 08 ce 9c 04 01 00 00
0x11f7e0043: 00 3f 00 0b eb 81 06 01 54 1f 20 03 d5
# RUBY_VM_CHECK_INTS(ec)
0x11f7e0050: 8b 02 42 b8 cb 07 01 35
# stack overflow check
0x11f7e0058: ab 62 02 91 7f 02 0b eb 69 07 01 54
# save PC to CFP
0x11f7e0064: 0b 3b 9a d2 2b 2f a0 f2 0b 00 cc f2 6b 02 00
0x11f7e0073: f8 ab 82 00 91
```
To ensure this feature doesn't incur too much cost when running without
the `--yjit-dump-disasm` option, I checked that there is no significant
impact to compile time and memory usage with the `compile_time_ns` and
`yjit_alloc_size` entry in `RubyVM::YJIT.runtime_stats`. For each
sample, I ran 3 iterations of the `lobsters` YJIT benchmark. The
statistics summary and done with the `summary` function in R.
Compile time, sample size of 60, lower is better:
```
Before After
Min. :2.054e+09 Min. :2.028e+09
1st Qu.:2.069e+09 1st Qu.:2.044e+09
Median :2.081e+09 Median :2.060e+09
Mean :2.089e+09 Mean :2.066e+09
3rd Qu.:2.109e+09 3rd Qu.:2.085e+09
Max. :2.146e+09 Max. :2.144e+09
```
Allocation size, sample size of 20, lower is better:
```
Before After
Min. :21804742 Min. :21794082
1st Qu.:21826682 1st Qu.:21816282
Median :21844042 Median :21826814
Mean :21960664 Mean :22026291
3rd Qu.:21861228 3rd Qu.:22040439
Max. :22587426 Max. :22930614
```
The `yjit_alloc_size` samples are noisy, but since the average increased
by only 0.3%, and the median is lower, I feel safe saying that there is
no significant change.
Use a special breakpoint address if one isn't explicitly supplied in order to support natural line stepping.
ARM64 will not increment the program counter (PC) upon hitting a breakpoint instruction. Consequently, stepping through code with a debugger ends up looping back to the breakpoint instruction. LLDB has a special breakpoint address of 0xf000 that will increment the PC and allow the debugger to work as expected. This change makes it possible to debug YJIT generated code on ARM64.
More details at: https://discourse.llvm.org/t/stepping-over-a-brk-instruction-on-arm64/69766/8
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
This commit expands inlining for simple ISeqs to accept
callees that have unused keyword parameters and callers
that specify unused keywords. The following shows 2 new
callsites that will be inlined:
```ruby
def let(a, checked: true) = a
let(1)
let(1, checked: false)
```
Co-authored-by: Kaan Ozkan <kaan.ozkan@shopify.com>
Many functions take an outlined code block but do nothing more than
passing it along; only a couple of functions actually make use of it.
So, in most cases the `ocb` parameter is just boilerplate.
Most functions that take `ocb` already also take a `JITState` and this
commit moves `ocb` into `JITState` to remove the visual noise of the
`ocb` parameter.
This commit fixes splat and block handling when calling in to a
forwarding iseq. In the case of a splat we need to avoid expanding the
array to the stack. We need to also ensure the CI write is flushed to
the SP, otherwise it's possible for a block handler to clobber the CI
[ruby-core:118360]
This commit adds `sendforward` and `invokesuperforward` for forwarding
parameters to calls
Co-authored-by: Matt Valentine-House <matt@eightbitraptor.com>
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>
This mainly aims to make `--yjit-dump-disasm=<relative_path>` more
usable. Previously, it crashed if the program did chdir(2), since it
opened the dump file every time when appending.
Tested with:
./miniruby --yjit-dump-disasm=. --yjit-call-threshold=1 -e 'Dir.chdir("/") {}'
And the `lobsters` benchmark.
Calls to defer_compilation() leave behind a stub and a `struct Block`
that we retain. If the block is empty, it only exits to hold the
`struct Branch` that the stub needs.
This patch transplants the branch out of the empty block into the newly
generated block when the defer_compilation() stub is hit, and deletes
the empty block to save memory.
To assist the transplantation, `Block::outgoing` is now a
`MutableBranchList`, and `Branch::Block` now in a `Cell`. These types
don't incur a size cost.
On the `lobsters` benchmark, `yjit_alloc_size` is roughly 98% of what
it was before the change.
Co-authored-by: Kevin Menard <kevin.menard@shopify.com>
Co-authored-by: Randy Stauner <randy@r4s6.net>
Co-authored-by: Maxime Chevalier-Boisvert <maxime.chevalierboisvert@shopify.com>
* YJIT: add context cache hits stat
This stat should make more sense when it comes to interpreting
the effectiveness of the cache on large deployed apps.
* 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
If you start Ruby with `--yjit-disable`, the `+YJIT` shouldn't be
added until `RubyVM::YJIT.enable` is actually called. Otherwise
it's confusing in crash reports etc.
No plan about migrating to the 2024 edition yet (it's not even
available yet), but this is a simple enough suggestion so we can just
take it.
```
warning: this method call resolves to `<&Box<[T]> as IntoIterator>::into_iter` (due to backwards compatibility), but will resolve to `<Box<[T]> as IntoIterator>::into_iter` in Rust 2024
--> ../yjit/src/core.rs:1003:49
|
1003 | formatter.debug_list().entries(branches.into_iter()).finish()
| ^^^^^^^^^
|
= warning: this changes meaning in Rust 2024
= note: `#[warn(boxed_slice_into_iter)]` on by default
help: use `.iter()` instead of `.into_iter()` to avoid ambiguity
|
1003 | formatter.debug_list().entries(branches.iter()).finish()
| ~~~~
help: or use `IntoIterator::into_iter(..)` instead of `.into_iter()` to explicitly iterate by value
|
1003 | formatter.debug_list().entries(IntoIterator::into_iter(branches)).finish()
| ++++++++++++++++++++++++ ~
```
* YJIT: limit size of call count stats dict
Someone reported that logs were getting bloated because the
ISEQ and C call count dicts were huge, since they include all
of the call sites. I wrote code on the Rust size to limit the
size of the dict to avoid this problem. The size limit is
hardcoded at 20, but I figure this is probably fine?
* Fix bug reported by Kokubun.
They were initially made frozen to avoid false positives for cases such
as:
str = str.dup if str.frozen?
But this may cause bugs and is generally confusing for users.
[Feature #20205]
Co-authored-by: Jean Boussier <byroot@ruby-lang.org>
* YJIT: Fix `Struct` accessors not firing tracing events
Reading and writing to structs should fire `c_call` and `c_return`, but
YJIT wasn't correctly dropping those calls when tracing.
This has been missing since this functionality was added in 3081c83169,
but the added test only fails when ran in isolation with
`--yjit-call-threshold=1`. The test sometimes failed on CI.
* RJIT: YJIT: Fix `Struct` readers not firing tracing events
Same issue as YJIT, but it looks like RJIT doesn't support writing to
structs, so only reading needs changing.
Since we often take the VM lock as the first thing we do when entering
YJIT, and that needs a `src_loc!()`, this removes a allocation from
that. The main trick here is `concat!(file!(), '\0')` to get a C string
statically baked into the binary.
* YJIT: Add specialized codegen function for `TrueClass#===`
TrueClass#=== is currently number 10 in the most frequent C calls list of the lobsters benchmark.
```
require "benchmark/ips"
def wrap
true === true
true === false
true === :x
end
Benchmark.ips do |x|
x.report(:wrap) do
wrap
end
end
```
```
before
Warming up --------------------------------------
wrap 1.791M i/100ms
Calculating -------------------------------------
wrap 17.806M (± 1.0%) i/s - 89.544M in 5.029363s
after
Warming up --------------------------------------
wrap 4.024M i/100ms
Calculating -------------------------------------
wrap 40.149M (± 1.1%) i/s - 201.223M in 5.012527s
```
Co-authored-by: Maxime Chevalier-Boisvert <maxime.chevalierboisvert@shopify.com>
Co-authored-by: Takashi Kokubun (k0kubun) <takashikkbn@gmail.com>
Co-authored-by: Kevin Menard <kevin.menard@shopify.com>
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* Fix the new test for RJIT
---------
Co-authored-by: Maxime Chevalier-Boisvert <maxime.chevalierboisvert@shopify.com>
Co-authored-by: Takashi Kokubun (k0kubun) <takashikkbn@gmail.com>
Co-authored-by: Kevin Menard <kevin.menard@shopify.com>
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
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.
Types like `Type::CString` really only assert that at one point the object had
its class field equal to `String`. Once a singleton class is created for any
strings, the type makes no assertion about any class field anymore, and becomes
the same as `Type::TString`.
Previously, the `--yjit-verify-ctx` option wasn't allowing objects of these
kind that have have singleton classes to pass verification even though the code
generators handle it just fine.
Found through `ruby/spec`.
* 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>
Add a specialized codegen function for `Class#superclass`.
Co-authored-by: Maxime Chevalier-Boisvert <maxime.chevalierboisvert@shopify.com>
Co-authored-by: Takashi Kokubun (k0kubun) <takashikkbn@gmail.com>
Co-authored-by: Randy Stauner <randy.stauner@shopify.com>
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* YJIT: Fix shrinking block with assumption too much
Under the very specific circumstances, discovered by a test case in
`ruby/spec`, an `expandarray` block can contain just a branch and carry
a method lookup assumption. Previously, when we regenerated the branch,
we allowed it to shrink to empty, since we put the code at the jump
target immediately after it. That was incorrect and caused a crash while
the block is invalidated, since that left no room to patch in an exit.
When regenerating a branch that makes up a block entirely, and the block
could be invalidated, we need to ensure there is room for invalidation.
When there is code before the branch, they should act as padding, so we
don't need to worry about those cases.
* skip on RJIT
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.
Previously, we got "We are killing the stack canary set by opt_ltlt"
from `$./miniruby --yjit-call-threshold=1 -e 'a = [].freeze; a << 1'`
Found by running ruby-spec with yjit-call-threshold=1.
* YJIT: A64: Add CBZ and CBNZ encoding functions
* YJIT: A64: Use CBZ/CBNZ to check for zero
Instead of emitting `cmp x0, #0` plus `b.z #target`, A64 offers Compare
and Branch on Zero for us to just do `cbz x0, #target`. This commit
utilizes that and the related CBNZ instruction when appropriate.
We check for zero most commonly in interrupt checks:
```diff
# Insn: 0003 leave (stack_size: 1)
# RUBY_VM_CHECK_INTS(ec)
ldur w11, [x20, #0x20]
-tst w11, w11
-b.ne #0x109002164
+cbnz w11, #0x1049021d0
```
* fix copy paste error
Co-authored-by: Randy Stauner <randy@r4s6.net>
---------
Co-authored-by: Randy Stauner <randy@r4s6.net>
Takashi Kokubun
Maxime Chevalier-Boisvert
Alan Wu
Alan Wu
Kevin Menard
Gabriel Lacroix
Aaron Patterson
Jean Boussier
Étienne Barrié
Nobuyoshi Nakada
Randy Stauner
Randy Stauner
Kevin Menard
careworry