Gabriel Baraldi

Avoid potential race between read and upload_data

8

`upload_data` might set `easy.input` to nothing before it's actually read if `read_callbacks` runs before it.

https://github.com/JuliaGPU/CUDA.jl/pull/2304 added optimistic garbage_collection checks and apparently that had a good impact on performance. While that is a new implementation and not very battle tested, it seems like it would...

Move globals in top level exprs to the old generation of the GC.

4

This fixes #54422, the code here assumes that top level exprs are always rooted, but I don't see that referenced anywhere else, or guaranteed, so conservatively always root objects that...

With the memory changes the UB surface of String(Vector{UInt8}) is larger

1

Observing/modifying the data of a Vector{UInt8} after one has called `String(vec)` on it has always been UB, this was usually fine because it's quite hard on a pre-Memory{T} world to...

Views on arrays should become views on a slice of the Memory that was accssible at view construction

1

This has historically being a possible unchecked OOB access if a user made a vector and pushed/popped/sizehinted it while keeping a view into the array. Views do not do a...

Some uses of `@__DIR__` should be relocatable.

4

~The proposed use I have for this is for when packages are looking for their Artifacts.toml file, they have to annoying things, and non relocatable ones even more annyoingly. (Maybe...

Julia's target parsing/detecting deserves a refactor

1

Julia currently handrolls it's target detection with a very convoluted script that deserves a refactor. - We should start using LLVMs TargetParser lib, we already statically link julia to it...

Actually setup jit targets when compiling packageimages instead of targeting only one

14

I haven't assembled a test for this just yet, but this works Fixes https://github.com/JuliaLang/julia/issues/54464

In the current version, we show all allocations a function and it's callees, might do. But that can unnecessarily expose internals that the user might not care about. So maybe...

Optimize jl_tls_world_age ccall

1

This makes it so we avoid a ccall boundary and potentially optimize away repeated world age checks