We set the block address as soon as we make the block, so there is no
point in making it `Option<CodePtr>`. No memory saving, unfortunately,
as `mem::size_of::<Block>() = 176` before and after this change. Still
a simplification for the logic, though.
Previously, with Code GC, YJIT panicked while trying to emit a B.cond
instruction with an offset that is not encodable in 19 bits. This only
happens when the code in an assembler instance straddles two pages.
To fix this, when we detect that a jump to a label can land on a
different page, we switch to a fresh new page and regenerate all the
code in the assembler there. We still assume that no one assembler has
so much code that it wouldn't fit inside a fresh new page.
[Bug #19385]
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
Co-authored-by: Maxime Chevalier-Boisvert <maxime.chevalierboisvert@shopify.com>
* use correct svar
Without this patch, svar location is used "nearest Ruby frame".
It is almost correct but it doesn't correct when the `each` method
is written in Ruby.
```ruby
class C
include Enumerable
def each
%w(bar baz).each{|e| yield e}
end
end
C.new.grep(/(b.)/){|e| p [$1, e]}
```
This patch fix this issue by traversing ifunc's cfp.
Note that if cfp doesn't specify this Thread's cfp stack, reserved
svar location (`ec->root_svar`) is used.
* make yjit-bindgen
---------
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
Grouping these together helps with finding all of the unimplemented
method types. It was interleaved with some other match arm long and
short previously.
Rust 1.58.0 unfortunately doesn't provide facilities to control symbol
visibility/presence, but we care about controlling the list of
symbols exported from libruby-static.a and libruby.so.
This commit uses `ld -r` to make a single object out of rustc's
staticlib output, libyjit.a. This moves libyjit.a out of MAINLIBS and adds
libyjit.o into COMMONOBJS, which obviates the code for merging libyjit.a
into libruby-static.a. The odd appearance of libyjit.a in SOLIBS is also
gone.
To filter out symbols we do not want to export on ELF platforms, we use
objcopy after the partial link. On darwin, we supply a symbol list to
the linker which takes care of hiding unprefixed symbols.
[Bug #19255]
Co-authored-by: Nobuyoshi Nakada <nobu@ruby-lang.org>
Note: On the new code of yjit/src/core.rs:2178, we no longer leave the state `.block=None` but `.address=Some...`, which might be important.
We assume it's actually not needed and take a risk here to minimize heap allocations, but in case it turns out to be necessary, we could signal/resurrect that state by introducing a new BranchTarget (or BranchShape) variant dedicated to it.
YJIT: Implement splat for cfuncs. Split exit cases
This also implements a new check for ruby2keywords as the last
argument of a splat. This does mean that we generate more code, but in
actual benchmarks where we gained speed from this (binarytrees) I
don't see any significant slow down. I did have to struggle here with
the register allocator to find code that didn't allocate too many
registers. It's a bit hard when everything is implicit. But I think I
got to the minimal amount of copying and stuff given our current
allocation strategy.
This reduces the code size of libyjit.a by a lot. On darwin it went from
23 MiB to 12 MiB for me. I chose ThinLTO over fat LTO for the relatively
fast build time; in case we need to debug release-build-only problems
it won't be painful.
* Add job to check clippy lints in CI
* Address all remaining clippy lints
* Check lints on arm64 as well
* Apply latest clippy lints
* Do not exit 0 on clippy warnings
* Differentiate T_ARRAY and array subclasses
This commit teaches the YJIT context the difference between Arrays
(objects with type T_ARRAY and class rb_cArray) vs Array subclasses
(objects with type T_ARRAY but _not_ class rb_cArray). It uses this
information to reduce the number of guards emitted when using
`jit_guard_known_klass` with rb_cArray, notably opt_aref
* Update yjit/src/core.rs
Co-authored-by: Maxime Chevalier-Boisvert <maximechevalierb@gmail.com>
Previously, we did not update `cfp->sp` before calling the C function of
ISEQs marked with `Primitive.attr! "inline"` (leaf builtins). This
caused the GC to miss temporary values on the stack in case the function
allocates and triggers a GC run. Right now, there is only a few leaf
builtins in numeric.rb on Integer methods such as `Integer#~`. Since
these methods only allocate when operating on big numbers, we missed
this issue.
Fix by saving PC and SP before calling the functions -- our usual
protocol for calling C functions that may allocate on the GC heap.
[Bug #19316]
* YJIT: Make iseq_get_location consistent with iseq.c
* YJIT: Call it "YJIT entry point"
Co-authored-by: Maxime Chevalier-Boisvert <maximechevalierb@gmail.com>
Co-authored-by: Maxime Chevalier-Boisvert <maximechevalierb@gmail.com>
We should differentiate between set and get for megamorphic exits. This
patch fixes the megamorphic exit name in gen_setinstancevariable so that
we can tell the difference between megamorphic get / set sites
On a hash miss we need to call default if it is redefined in order to
return the default value to be used. Previously we checked this with
rb_method_basic_definition_p, which avoids the method call but requires
a method lookup.
This commit replaces the previous check with BASIC_OP_UNREDEFINED_P and
a new BOP_DEFAULT. We still need to fall back to
rb_method_basic_definition_p when called on a subclasss of hash.
| |compare-ruby|built-ruby|
|:---------------|-----------:|---------:|
|hash_aref_miss | 2.692| 3.531|
| | -| 1.31x|
Co-authored-by: Daniel Colson <danieljamescolson@gmail.com>
Co-authored-by: "Ian C. Anderson" <ian@iancanderson.com>
Co-authored-by: Jack McCracken <me@jackmc.xyz>
All the method call types need to handle argument shifting in case they're
called by `.send`, and we weren't handling that in `OPTIMIZED_METHOD_TYPE_CALL`.
Lack of shifting caused the stack size assertion in gen_leave() to fail.
Discovered by Rails CI: https://buildkite.com/rails/rails/builds/91705#018516c4-f8f8-469e-bc2d-ddeb25ca8317/1920-2067
Diagnosed with help from `@eileencodes` and `@k0kubun`.
SIZE_POOL_COUNT is a GC macro, it should belong in gc.h and not shape.h.
SIZE_POOL_COUNT doesn't depend on shape.h so we can have shape.h depend
on gc.h.
Co-Authored-By: Matt Valentine-House <matt@eightbitraptor.com>
Stubs we generate for invalidation don't necessarily co-locate with the
code that jump to the stub. Since we rely on co-location to keep stubs
alive as they are in the outlined code block, it used to be possible for
code GC inside branch_stub_hit() to free the stub that's its direct
caller, leading us to return to freed code after.
Stubs used to look like:
```
mov arg0, branch_ptr
mov arg1, target_idx
mov arg2, ec
call branch_stub_hit
jmp return_reg
```
Since the call and the jump after the call is the same for all stubs, we
can extract them and use a static trampoline for them. That makes
branch_stub_hit() always return to static code. Stubs now look like:
```
mov arg0, branch_ptr
mov arg1, target_idx
jmp trampoline
```
Where the trampoline is:
```
mov arg2, ec
call branch_stub_hit
jmp return_reg
```
Code GC can now free stubs without problems since we'll always return
to the trampoline, which we generate once on boot and lives forever.
This might save a small bit of memory due to factoring out the static
part of stubs, but it's probably minor.
[Bug #19234]
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
