yjit

yjit copied to clipboard

Avoid discarding type information at every call

19

At the moment, we discard type information at every call, whether it be a call to a C function or another iseq, because it could be that a bindings environment...

maximecb

Ability to discard/free machine code (code GC)

27

At the moment, we generated code in a linear array of executable memory until we run out of space. In practice, this isn't a problem for most applications, because we...

maximecb

Possible simplification in iseq-to-iseq argument passing

2

Here is a review of the argument setup details in `gen_send_iseq()` cause I think there is room for simplification, both in terms of generator code and generated code. It's not...

XrXr

Code pages refactoring

5

Before we can proceed with the actual GC part of the code GC, we need to refactor our codegen so that code is allocated and generated into small regions (code...

maximecb

Optimize `Module#===`

4

This is also used very often in SFR, because of `case...when` patterns. It seems to be used most often with arrays and hashes, so I'm wondering if we could peek...

maximecb

Block invalidation doesn't always work (depends on invalidation order)

1

This is a bug found by @XrXr with repro shown below: ``` # Repro for buggy invalidation. Run with --yjit-call-threshold=1 and compare with interpreter only mode # The gist of...

maximecb

Avoid moving REG_SP (make it a base pointer)

5

This is an idea to improve how we handle stack addresses to make codegen simpler to write and possibly improve performance. Currently, `REG_SP` always has the same value of `reg_cfp->sp`,...

jhawthorn

Memory leak with const_set

1

The current `opt_setinlinecache` implementation invalidates the block for the corresponding `opt_getinlinecache`, and one could make `opt_setinlinecache` run by setting any constant. Apps that use `Module#const_set` at runtime will leak memory...

XrXr