* 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.
This allows x86_64 based YJIT to run on Docker Desktop on Apple silicon (arm64)
Mac because it will avoid a subtle behavior difference in `mprotect` system call
between the Linux kernel and `qemu-x86_64` user space emulator.
* Implement optimize send in yjit
This successfully makes all our benchmarks exit way less for optimize send reasons.
It makes some benchmarks faster, but not by as much as I'd like. I think this implementation
works, but there are definitely more optimial arrangements. For example, what if we compiled
send to a jump table? That seems like perhaps the most optimal we could do, but not obvious (to me)
how to implement give our current setup.
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* Attempt at fixing the issues raised by @XrXr
* fix allowlist
* returns 0 instead of nil when not found
* remove comment about encoding exception
* Fix up c changes
* Update assert
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* get rid of unneeded code and fix the flags
* Apply suggestions from code review
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* rename and fix typo
Co-authored-by: Alan Wu <XrXr@users.noreply.github.com>
* YJIT: fix a parameter name
* YJIT: add support for calling bmethods
This commit adds support for the VM_METHOD_TYPE_BMETHOD method type in
YJIT. You can get these type of methods from facilities like
Kernel#define_singleton_method and Module#define_method.
Even though the body of these methods are blocks, the parameter setup
for them is exactly the same as VM_METHOD_TYPE_ISEQ, so we can reuse
the same logic in gen_send_iseq(). You can see this from how
vm_call_bmethod() eventually calls setup_parameters_complex() with
arg_setup_method.
Bmethods do need their frame environment to be setup differently. We
handle this by allowing callers of gen_send_iseq() to control the iseq,
the frame flag, and the prev_ep. The `prev_ep` goes into the same
location as the block handler would go into in an iseq method frame.
Co-authored-by: John Hawthorn <john@hawthorn.email>
Co-authored-by: John Hawthorn <john@hawthorn.email>
* Initial support for VM_CALL_ARGS_SPLAT
This implements support for calls with splat (*) for some methods. In
benchmarks this made very little difference for most benchmarks, but a
large difference for binarytrees. Looking at side exits, many
benchmarks now don't exit for splat, but exit for some other
reason. Binarytrees however had a number of calls that used splat args
that are now much faster. In my non-scientific benchmarking this made
splat args performance on par with not using splat args at all.
* Fix wording and whitespace
Co-authored-by: Maxime Chevalier-Boisvert <maximechevalierb@gmail.com>
* Get rid of side_effect reassignment
Co-authored-by: Maxime Chevalier-Boisvert <maximechevalierb@gmail.com>
Previously we cleared the cache for all the code in the system when we
flip memory protection, which was prohibitively expensive since the
operation is not constant time. Instead, only clear the cache for the
memory region of newly written code when we write out new code.
This brings the runtime for the 30k_if_else test down to about 6 seconds
from the previous 45 seconds on my laptop.
We have a large extern block in cruby.rs leftover from the port. We can
use bindgen for it now and reserve the manual declaration for just a
handful of vm_insnhelper.c functions.
Fixup a few minor discrepencies bindgen found between the C declaration
and the manual declaration. Mostly missing `const` on the C side.
Teach getblockparamproxy to handle the no-block case without exiting
Co-authored-by: John Hawthorn <john@hawthorn.email>
Co-authored-by: John Hawthorn <john@hawthorn.email>
In a small script the speed of this feature isn't really noticeable but
on Rails it's very noticeable how slow this can be. This PR aims to
speed up two parts of the functionality.
1) The Rust exit recording code
Instead of adding all samples as we see them to the yjit_raw_samples and
yjit_line_samples, we can increment the counter on the ones we've seen
before. This will be faster on traces where we are hitting the same
stack often. In a crude measurement of booting just the active record
base test (`test/cases/base_test.rb`) we found that this improved the
speed by 1 second.
This also results in a smaller marshal dump file which sped up the test
boot time by 4 seconds with trace exits on.
2) The Ruby parsing code
Previously we were allocating new arrays using `shift` and
`each_with_index`. This change avoids allocating new arrays by using an
index. This change saves us the most amount of time, gaining 11 seconds.
Before this change the test boot time took 62 seconds, after it took 47
seconds. This is still too long but it's a step closer to faster
functionality. Next we're going to tackle allowing you to collect trace
exits for a specific instruction. There is also some potential slowness
in the GC code that I'd like to take a second look at.
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
This commit makes YJIT allocate memory for generated code gradually as
needed. Previously, YJIT allocates all the memory it needs on boot in
one go, leading to higher than necessary resident set size (RSS) and
time spent on boot initializing the memory with a large memset().
Users should no longer need to search for a magic number to pass to
`--yjit-exec-mem` since physical memory consumption should now more
accurately reflect the requirement of the workload.
YJIT now reserves a range of addresses on boot. This region start out
with no access permission at all so buggy attempts to jump to the region
crashes like before this change. To get this hardening at finer
granularity than the page size, we fill each page with trapping
instructions when we first allocate physical memory for the page.
Most of the time applications don't need 256 MiB of executable code, so
allocating on-demand ends up doing less total work than before. Case in
point, a simple `ruby --yjit-call-threshold=1 -eitself` takes about
half as long after this change. In terms of memory consumption, here is
a table to give a rough summary of the impact:
| Peak RSS in MiB | -eitself example | railsbench once |
| :-------------: | ---------------: | --------------: |
| before | 265 | 377 |
| after | 11 | 143 |
| no YJIT | 10 | 101 |
A new module is introduced to handle allocation bookkeeping.
`CodePtr` is moved into the module since it has a close relationship
with the new `VirtualMemory` struct. This new interface has a slightly
smaller surface than before in that marking a region as writable is no
longer a public operation.
Peter Zhu
Nobuyoshi Nakada
John Hawthorn
Matt Valentine-House
Takashi Kokubun
Maxime Chevalier-Boisvert
Alan Wu
Jimmy Miller
Noah Gibbs
Matthew Draper
Tatsuya Kawano
Noah Gibbs
Eileen M. Uchitelle
Dave Schwantes
Alan Wu