Preivously, [1] failed to undo the effect of applying
add_block_version() to a block, leaving dangling pointers in the iseq
when compilation fails.
[1]: d0772632bf
Verify that the iseq idx pair for the block is valid in
invalidate_block_version(). While we are at it, bound loop
iterating over instructions to `iseq_body->iseq_size`.
Some platforms don't want memory to be marked as writeable and
executable at the same time. When we write to the code block, we
calculate the OS page that the buffer position maps to. Then we call
`mprotect` to allow writes on that particular page. As an optimization,
we cache the "last written" aligned page which allows us to amortize the
cost of the `mprotect` call. In other words, sequential writes to the
same page will only call `mprotect` on the page once.
When we're done writing, we call `mprotect` on the entire JIT buffer.
This means we don't need to keep track of which pages were marked as
writeable, we let the OS take care of that.
Co-authored-by: John Hawthorn <john@hawthorn.email>
Previously, YJIT crashes with rb_bug() when asked to compile new methods
while out of executable memory.
To handle this situation gracefully, this change keeps track of all the
blocks compiled each invocation in case YJIT runs out of memory in the
middle of a compliation sequence. The list is used to free all blocks in
case compilation fails.
yjit_gen_block() is renamed to gen_single_block() to make it distinct from
gen_block_version(). Call to limit_block_version() and block_t
allocation is moved into the function to help tidy error checking in the
outer loop.
limit_block_version() now returns by value. I feel that an out parameter
with conditional mutation is unnecessarily hard to read in code that
does not need to go for last drop performance. There is a good chance
that the optimizer is able to output identical code anyways.
Previously, YJIT assumed that it's always possible to generate a new
basic block when servicing a stub in branch_stub_hit(). When YJIT is out
of executable memory, for example, this assumption doesn't hold up.
Add handling to branch_stub_hit() for servicing stubs without consuming
more executable memory by adding a code path that exits to the
interpreter at the location the branch stub represents. The new code
path reconstructs interpreter state in branch_stub_hit() and then exits
with a new snippet called `code_for_exit_from_stub` that returns
`Qundef` from the YJIT native stack frame.
As this change adds another place where we regenerate code from
`branch_t`, extract the logic for it into a new function and call it
regenerate_branch(). While we are at it, make the branch shrinking code
path in branch_stub_hit() more explicit.
This new functionality is hard to test without full support for out of
memory conditions. To verify this change, I ran
`RUBY_YJIT_ENABLE=1 make check -j12` with the following patch to stress
test the new code path:
```diff
diff --git a/yjit_core.c b/yjit_core.c
index 4ab63d9806..5788b8c5ed 100644
--- a/yjit_core.c
+++ b/yjit_core.c
@@ -878,8 +878,12 @@ branch_stub_hit(branch_t *branch, const uint32_t target_idx, rb_execution_contex
cb_set_write_ptr(cb, branch->end_addr);
}
+if (rand() < RAND_MAX/2) {
// Compile the new block version
p_block = gen_block_version(target, target_ctx, ec);
+}else{
+ p_block = NULL;
+}
if (!p_block && branch_modified) {
// We couldn't generate a new block for the branch, but we modified the branch.
```
We can enable the new test along with other OOM tests once full support
lands.
Other small changes:
* yjit_utils.c (print_str): Update to work with new native frame shape.
Follow up for 8fa0ee4d40.
* yjit_iface.c (rb_yjit_init): Run yjit_init_core() after
yjit_init_codegen() so `cb` and `ocb` are available.
This commit adds an entry_exit field to block_t for use in
invalidate_block_version(). By patching the start of the block while
invalidating it, invalidate_block_version() can function correctly
while there is no executable memory left for new branch stubs.
This change additionally fixes correctness for situations where we
cannot patch incoming jumps to the invalidated block. In situations
such as Shopify/yjit#226, the address to the start of the block
is saved and used later, possibly after the block is invalidated.
The assume_* family of function now generate block->entry_exit before
remembering blocks for invalidation.
RubyVM::YJIT.simulate_oom! is introduced for testing out of memory
conditions. The test for it is disabled for now because OOM triggers
other failure conditions not addressed by this commit.
FixesShopify/yjit#226
* Add --yjit-no-type-prop so we can test YJIT without type propagation
* Fix typo in command line option
* Leave just two test workflows enable for YJIT
* New code page allocation logic
* Fix leaked globals
* Fix leaked symbols, yjit asm tests
* Make COUNTED_EXIT take a jit argument, so we can eliminate global ocb
* Remove extra whitespace
* Change block start_pos/end_pos to be pointers instead of uint32_t
* Change branch end_pos and start_pos to end_addr, start_addr
For upstreaming, we want functions we export either prefixed with "rb_"
or made static. Historically we haven't been following this rule, so we
were "leaking" a lot of symbols as `make leak-globals` would tell us.
This change unifies everything YJIT into a single compilation unit,
yjit.o, and makes everything unprefixed static to pass `make leak-globals`.
This manual "unified build" setup is similar to that of vm.o.
Having everything in one compilation unit allows static functions to
be visible across YJIT files and removes the need for declarations in
headers in some cases. Unnecessary declarations were removed.
Other changes of note:
- switched to MJIT_SYMBOL_EXPORT_BEGIN which indicates stuff as being
off limits for native extensions
- the first include of each YJIT file is change to be "internal.h"
- undefined MAP_STACK before explicitly redefining it since it
collide's with a definition in system headers. Consider renaming?
We should only clear the JIT function when the entry point is
invalidated. Right now we only support compiling functions with a PC
offset of zero (functions that take optional parameters can start at
non-zero PC), so this patch just checks that the index is 0 before
clearing the jit function
This change fixes some cases where YJIT fails to fire tracing events.
Most of the situations YJIT did not handle correctly involves enabling
tracing while running inside generated code.
A new operation to invalidate all generated code is added, which uses
patching to make generated code exit at the next VM instruction
boundary. A new routine called `jit_prepare_routine_call()` is
introduced to facilitate this and should be used when generating code
that could allocate, or could otherwise use `RB_VM_LOCK_ENTER()`.
The `c_return` event is fired in the middle of an instruction as opposed
to at an instruction boundary, so it requires special handling. C method
call return points are patched to go to a fucntion which does everything
the interpreter does, including firing the `c_return` event. The
generated code for C method calls normally does not fire the event.
Invalided code should not change after patching so the exits are not
clobbered. A new variable is introduced to track the region of code that
should not change.
RUBY_DEBUG have a very significant performance overhead. Enough that
YJIT with RUBY_DEBUG is noticeably slower than the interpreter without
RUBY_DEBUG.
This makes it hard to collect yjit-stats in production environments.
By allowing to collect JIT statistics without the RUBy_DEBUG overhead,
I hope to make such use cases smoother.
Previously all stack operands would become unknown once the stack grew
beyond a certain size. This worked, but unnecessarily hid available
information.
If we hit this at PC > 0 (ie. with an optional argument) the provided
types and context are likely incorrect and it is likely this block can't
be used.
If `--disable-jit-support` is passed to configure, then `jit_func` is
removed from the iseq body and we can't compile YJIT. This commit
detects when the JIT function pointer is gone and disables YJIT in that
case.
Methods with optional parameters don't always start executing at the
first PC, but we compile all methods assuming that they do. This commit
adds a guard to ensure that we're actually starting at the first PC for
methods with optional params
We clear locals when we know their values might change (ex. when
performing a method call). However previously values on the stack which
were originally pushed from a local would still point back to that
local.
With this commit, when clearing locals, we'll now iterate over the
mappings of the stack and copy the known type from the local to the
stack mapping, removing the association to the local.
This should mean both that we'll retain any information we already know
about the local type, and that if a local is modified we won't
incorrectly infer it's new type from the existing value on the stack.
Nobuyoshi Nakada
Alan Wu
Aaron Patterson
Maxime Chevalier-Boisvert
John Hawthorn
Jean Boussier
Noah Gibbs
Maxime Chevalier-Boisvert