add atomic operators for globals, memory, and setonce #52868

vtjnash · 2024-01-12T02:31:14Z

This implements AtomicMemory, atomic operations for globals, and the class of atomic function for setting a field or global or memory exactly once (setting undef => value). It is quite similar to an @atomicreplace, but where there was not a previous value. This is not needed for pointers, since it is simply calling C_NULL => pointer in that case, since there is no "undef" value that throws after load.

plenty of future work still, for a later PR:

syntax lowering for @atomic global x = y and @atomic closedover = y
adding Core.AtomicBox for closure capture
generic functions for atomicsetindex & friends
macro for @atomic x = y and @atomic x[] = y, etc
design for @atomiconce f()

vtjnash · 2024-01-27T17:47:22Z

32-bit failing with

Error in testset atomics:
Test Failed at C:\buildkite-agent\builds\win2k22-amdci6-4\julialang\julia-master\julia-48bc10d274\share\julia\test\atomics.jl:76
  Expression: sizeof(mr) == length(mr) * sizeof(ar)
   Evaluated: 1600 == 1440

but everything else looks ready for review

oscardssmith · 2024-01-27T21:18:04Z

base/genericmemory.jl

@@ -145,10 +151,9 @@ end

 ## Conversions ##

-convert(::Type{T}, a::AbstractArray) where {T<:GenericMemory} = a isa T ? a : T(a)::T
+convert(::Type{T}, a::AbstractArray) where {T<:Memory} = a isa T ? a : T(a)::T


should the second argument here be AbstracVector?

apparently not?

julia> @which convert(Vector{Int}, [1 2; 3 4]) convert(::Type{T}, a::AbstractArray) where T<:Array @ Base array.jl:607 julia> convert(Vector{Int}, [1 2; 3 4]) ERROR: MethodError: no method matching Vector{Int64}(::Matrix{Int64})

test/compiler/inference.jl

vtjnash · 2024-01-31T18:37:58Z

I will merge tomorrow, and take any post-merge feedback instead then, if not gotten to before then.

topolarity · 2024-01-31T19:36:15Z

Do you have a list of all the new and deleted intrinsics/builtins?

It's a bit hard to glean from the diff, since you also did some (appreciated) clean-up on things.

gbaraldi · 2024-01-31T19:41:32Z

base/expr.jl

+ @atomiconce :sequentially_consistent :monotonic a.b.x = value
+
+Perform the conditional assignment of value atomically if it was previously
+unset, returning the value `success::Bool`. Where `success` indicates whether


So one thing is that unset isn't the nomenclature we seem to usually use, we mostly seem to use defined/undefined

I guess that should say the field was not set or the field was undef?

I think that might be a bit more inline to what we call it in other places?

We use unset at lot for this also (e.g. unsetindex), since we use undefined to mean isbits types that are not initialized, while this is not undef->defined but rather unset->set

src/cgutils.cpp

doc/src/devdocs/builtins.md

vtjnash · 2024-02-01T18:18:19Z

Here's the list of new functions, from the docs (though your comment made me realize I had missed adding the rest of the memoryref builtins there):

Core.modifyglobal!
Core.swapglobal!
Core.replaceglobal!
Core.setglobalonce!

Core.memoryrefswap!
Core.memoryrefmodify!
Core.memoryrefreplace!
Core.memoryrefsetonce!

Core.setfieldonce!

vchuravy · 2024-02-01T21:42:55Z

base/docs/basedocs.jl

@@ -2281,6 +2282,21 @@ instruction, otherwise it'll use a loop.
 """
 replacefield!

+"""
+ setfieldonce!(value, name::Union{Int,Symbol}, desired,
+ [success_order::Symbol, [fail_order::Symbol=success_order]) -> success::Bool


What is the default success_order? (Same for the other docstrings here)

The default is not to have a success_order (in some places, equivalent to :not_atomic, but not identical)

vchuravy · 2024-02-01T21:43:50Z

base/docs/basedocs.jl

+ setfield!(value, name, desired, success_order)
+ end
+ return ok
+"""


Probably should also get a compat note?

Hopefully I got all those versions right

Adds an AtomicMemory{T} type (GenericMemory{:atomic,T,Core.CPU}) for doing element-wise atomic operations on an array (with a lock per-element if required). Adds the full complement of atomic operations for global variables. Also add a *setonce set of functions for atomically setting a value, but only if it is unset. The complementary operation (Atomic `_unsetindex!` for AtomicMemory) is not yet provided.

vtjnash added the domain:multithreading Base.Threads and related functionality label Jan 12, 2024

vtjnash requested a review from gbaraldi January 12, 2024 02:31

vtjnash force-pushed the jn/moreatomics branch 3 times, most recently from 4456be2 to cdd101e Compare January 27, 2024 00:34

vtjnash marked this pull request as ready for review January 27, 2024 00:34

vtjnash force-pushed the jn/moreatomics branch from cdd101e to 48bc10d Compare January 27, 2024 02:00

oscardssmith reviewed Jan 27, 2024

View reviewed changes

test/compiler/inference.jl Show resolved Hide resolved

vtjnash force-pushed the jn/moreatomics branch from 48bc10d to a872b19 Compare January 29, 2024 18:06

gbaraldi reviewed Jan 31, 2024

View reviewed changes

src/cgutils.cpp Show resolved Hide resolved

vtjnash commented Feb 1, 2024

View reviewed changes

doc/src/devdocs/builtins.md Show resolved Hide resolved

vchuravy reviewed Feb 1, 2024

View reviewed changes

vtjnash added 2 commits February 4, 2024 21:05

add docs compat

d8d17b8

vtjnash force-pushed the jn/moreatomics branch from a872b19 to d8d17b8 Compare February 4, 2024 21:56

vtjnash added the status:merge me PR is reviewed. Merge when all tests are passing label Feb 4, 2024

IanButterworth merged commit 8db1294 into master Feb 5, 2024
8 checks passed

IanButterworth deleted the jn/moreatomics branch February 5, 2024 00:51

IanButterworth removed the status:merge me PR is reviewed. Merge when all tests are passing label Feb 5, 2024

aviatesk added a commit that referenced this pull request Feb 12, 2024

follow up #52868 and fix up doc/src/base/base.md

3ed1066

aviatesk added a commit that referenced this pull request Feb 12, 2024

follow up #52868 and fix up doc/src/base/base.md (#53294)

61bf208

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add atomic operators for globals, memory, and setonce #52868

add atomic operators for globals, memory, and setonce #52868

vtjnash commented Jan 12, 2024 •

edited

Loading

vtjnash commented Jan 27, 2024

oscardssmith Jan 27, 2024

vtjnash Jan 27, 2024

vtjnash commented Jan 31, 2024

topolarity commented Jan 31, 2024

gbaraldi Jan 31, 2024

vtjnash Feb 1, 2024

gbaraldi Feb 1, 2024

vtjnash Feb 4, 2024 •

edited

Loading

vtjnash commented Feb 1, 2024

vchuravy Feb 1, 2024 •

edited

Loading

vtjnash Feb 1, 2024

vchuravy Feb 1, 2024

vtjnash Feb 4, 2024

add atomic operators for globals, memory, and setonce #52868

add atomic operators for globals, memory, and setonce #52868

Conversation

vtjnash commented Jan 12, 2024 • edited Loading

vtjnash commented Jan 27, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vtjnash commented Jan 31, 2024

topolarity commented Jan 31, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vtjnash Feb 4, 2024 • edited Loading

Choose a reason for hiding this comment

vtjnash commented Feb 1, 2024

vchuravy Feb 1, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vtjnash commented Jan 12, 2024 •

edited

Loading

vtjnash Feb 4, 2024 •

edited

Loading

vchuravy Feb 1, 2024 •

edited

Loading