Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

GC.gc() unable to free memory in the scope it is called from? #51818

Closed
Tortar opened this issue Oct 22, 2023 · 10 comments · Fixed by #52935
Closed

GC.gc() unable to free memory in the scope it is called from? #51818

Tortar opened this issue Oct 22, 2023 · 10 comments · Fixed by #52935
Labels
GC Garbage collector

Comments

@Tortar
Copy link
Contributor

Tortar commented Oct 22, 2023

I opened some days ago this issue here: apache/arrow-julia#492. But after experimenting quite a bit I think this is a GC issue.

MWE (which I runned in REPL mode) similar as the one reported here https://discourse.julialang.org/t/how-to-force-an-object-to-be-freed-by-the-garbage-collector/77985:

function testme()
    X = rand(1_000_000_00)
    Y = sum(X)
    X = nothing
    GC.gc(); GC.gc(); GC.gc()
    return Y
end

function tester()
    Y = testme()
    return Y
end

tester()
# GC.gc() here frees it

the memory occupied by X is not freed, if I put after the call to tester a GC call instead it is.

Version Info:

julia> versioninfo()
Julia Version 1.9.3
Commit bed2cd540a (2023-08-24 14:43 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 12 × AMD Ryzen 5 5600H with Radeon Graphics
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, znver3)
  Threads: 1 on 12 virtual cores
@Tortar Tortar changed the title GC unable to free memory with .arrow files on Windows GC unable to free memory in the scope it is called from? Oct 22, 2023
@nsajko
Copy link
Contributor

nsajko commented Oct 22, 2023

Your examples rely on user packages, so I don't see why would you think this is a Julia bug. Apart from that, it's not even clear what behavior you consider buggy.

Are you aware that languages that use garbage collection (except simple forms like reference counting, like with C++, for example) don't provide guarantees about when some pointer will be freed, or when the finalizer of some object will be ran. The precise moment is an implementation detail. Wikipedia, furthermore, says:

Notably, both Java and Python do not guarantee that finalizers will ever be called, and thus they cannot be relied on for cleanup.

@Tortar
Copy link
Contributor Author

Tortar commented Oct 22, 2023

I changed it to one not relying on user packages. Thanks for the info about the garbage collection in other languages, I had the impression that a GC.gc() call (or multiple ones) would have guaranteed to free the unused memory. This could be nonetheless useful sometimes, but I can't judge how much of a hard task is it to make this an option, probably a lot

@Tortar Tortar changed the title GC unable to free memory in the scope it is called from? GC.gc() unable to free memory in the scope it is called from? Oct 22, 2023
@Tortar
Copy link
Contributor Author

Tortar commented Oct 22, 2023

Then given that it is probably like so for good reasons, maybe changing the docstring of GC.gc() could be a good idea? For now it is:

  Perform garbage collection. The argument full determines the kind of collection: A full collection (default) sweeps
  all objects, which makes the next GC scan much slower, while an incremental collection may only sweep so-called
  young objects.

maybe saying that it is a suggestion is better?

  Suggest to perform garbage collection. The argument full determines the kind of collection: A full collection (default) sweeps
  all objects, which makes the next GC scan much slower, while an incremental collection may only sweep so-called
  young objects.

@PallHaraldsson
Copy link
Contributor

See also my long reply on discourse on using Bumper.jl that WILL guarantee deallocation. GC languages in general can't do it, but in exception cases it could be done, with the help of the compiler, as I explained. I think Julia should actually try to to that, or even guarantee in all possible cases, as Mojo does.

@vchuravy
Copy link
Member

Julia uses a precise GC together with a GC shadow stack.

julia> function testme()
           X = @noinline rand(1_000_000_00)
           Y = @noinline sum(X)
           X = nothing
           GC.gc()
           return Y
       end

Let's take a look at the IR generated:

julia> @code_llvm  testme()
;  @ REPL[7]:1 within `testme`
define double @julia_testme_190() #0 {
top:
  %gcframe2 = alloca [3 x {}*], align 16
  %gcframe2.sub = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe2, i64 0, i64 0
  %0 = bitcast [3 x {}*]* %gcframe2 to i8*
  call void @llvm.memset.p0i8.i64(i8* align 16 %0, i8 0, i64 24, i1 true)
  %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #8
  %tls_ppgcstack = getelementptr i8, i8* %thread_ptr, i64 -8
  %1 = bitcast i8* %tls_ppgcstack to {}****
  %tls_pgcstack = load {}***, {}**** %1, align 8
;  @ REPL[7]:2 within `testme`
  %2 = bitcast [3 x {}*]* %gcframe2 to i64*
  store i64 4, i64* %2, align 16
  %3 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe2, i64 0, i64 1
  %4 = bitcast {}** %3 to {}***
  %5 = load {}**, {}*** %tls_pgcstack, align 8
  store {}** %5, {}*** %4, align 8
  %6 = bitcast {}*** %tls_pgcstack to {}***
  store {}** %gcframe2.sub, {}*** %6, align 8
  %7 = call nonnull {}* @j_rand_192(i64 signext 100000000)
  %8 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe2, i64 0, i64 2
  store {}* %7, {}** %8, align 16
;  @ REPL[7]:3 within `testme`
  %9 = call double @j_sum_193({}* nonnull %7)
;  @ REPL[7]:5 within `testme`
; ┌ @ gcutils.jl:129 within `gc` @ gcutils.jl:129
   call void inttoptr (i64 140443355088864 to void (i32)*)(i32 1)
   %10 = load {}*, {}** %3, align 8
   %11 = bitcast {}*** %tls_pgcstack to {}**
   store {}* %10, {}** %11, align 8
; └
;  @ REPL[7]:6 within `testme`
  ret double %9
}
  %gcframe2 = alloca [3 x {}*], align 16

Is the creation of the shadow stack. Simplifying a bit every allocated object inside a function get's assinged a slot in this gcframe and GC scans the data structure (which forms a linked list) to find objects that are used.

  %7 = call nonnull {}* @j_rand_192(i64 signext 100000000)
  %8 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe2, i64 0, i64 2
  store {}* %7, {}** %8, align 16

Now this is the call to rand and store of X to the gcframe, so that it is rooted
for the next step the call to sum.

  %9 = call double @j_sum_193({}* nonnull %7)

Now immediately after we call GC.gc

 ┌ @ gcutils.jl:129 within `gc` @ gcutils.jl:129
   call void inttoptr (i64 140443355088864 to void (i32)*)(i32 1)

But X is still rooted on the gcstack. Once we exit the function we pop the gcframe from the pgcstack and X can be freed.

There might be an open issue flying around, but if I recall correctly we are currently missing the step where we zero out the GC slot when the lifetime of the object ends,
since that is often not necessary to do (e.g. if you exit the function the memory will be freed).

Probably the place to change is around

void LateLowerGCFrame::PlaceGCFrameStores(State &S, unsigned MinColorRoot,

@PallHaraldsson
Copy link
Contributor

PallHaraldsson commented Oct 26, 2023

I think you're saying GC is called. But is GC guaranteed to free large objects, even for full (the default for GC.gc()? I mean is it promised or a known limitation. Should is be promised, at least for full GC? This can be avoided with Bumper.jl, i.e. it eliminates the problem as opt-in. Julia should maybe add Bumber.jl as stdlib, similar to Mojo. I.e. a compile time guarantee. But until there, there is a large and small distinction for objects in the GC, right? Just pointing out in case it explains the (perceived) problem.

@vchuravy
Copy link
Member

I see no reason why Bumper is part of this discussion?

Julia's GC is precise and generational.

  1. A full GC will sweep all "garbage" objects.
  2. An incremental GC will sweep all "garbage" & young objects.

Now the challenge is what is "garbage". Garbage is all objects that are detached from the object-graph and are not in one of the root-sets.

Due to Julia being precise the root-sets include "all objects that are used within a function".

So when we see behavior that surprises us like the original post the question we need to answer:
Why is this object live? It isn't referenced anywhere so the only remaining bit is that it must be part of root-set.

My hope was to demystify Julia a bit by walking through the reason why X is live until the end of the function, which is a wider lifetime than expected. There is a simple technical reason (e.g. we didn't think it was worth it) and maybe we should revisit that.

So yes the GC is guaranteed to free objects, but it needs to see them as garbage.

@PallHaraldsson
Copy link
Contributor

So when we see behavior that surprises us like the original post the question we need to answer:
Why is this object live?

The topic here is why GC.gc() is "unable to free", and as soon as "X = nothing" happens in the code there, X is no longer live, so it's a good question (and should be answered), and yes this should get fixed. But I mean X could be freed right there, with Libc.free implicitly, before the next line, or any other (with allocation) triggering it. The GC without help couldn't. My understanding is that is what Mojo does, why I brought up Bumper. I believe it's similar, just explicit, not implicit (and doesn't use the heap; uses its own stack for allocated objects, I'm not sure Mojo does the same or works for the heap, maybe both).

@gbaraldi
Copy link
Member

The GC could free that object, the issue is that the generated code mark that object as dead until the end of the function, so when GC.gc() runs it still thinks the array is live.

@brenhinkeller brenhinkeller added the GC Garbage collector label Oct 26, 2023
@elextr
Copy link

elextr commented Oct 27, 2023

@gbaraldi just to check my understanding, you are saying because X is unused after the call to sum() the compiler determines that it can optimise away X = nothing (its not in the IR @vchuravy posted) because there is no use of X after that point.

But of course GC.gc() in a sense "uses" everything, but that information is probably not available to the compiler.

@vtjnash vtjnash closed this as completed in efc43cb Nov 6, 2024
udesou added a commit to mmtk/julia that referenced this issue Dec 6, 2024
* Implement faster `issubset` for `CartesianIndices{N}` (#56282)

Co-authored-by: xili <xili@phas.ubc.ca>

* Improve doc example: Extracting the type parameter from a super-type (#55983)

Documentation describes the correct way of extracting the element type
of a supertype:

https://docs.julialang.org/en/v1/manual/methods/#Extracting-the-type-parameter-from-a-super-type

However, one of the examples to showcase this is nonsensical since it is
a union of multiple element types.
I have replaced this example with a union over the dimension.
Now, the `eltype_wrong` function still gives a similar error, yet the
correct way returns the unambiguous answer.

---------

Co-authored-by: Lilith Orion Hafner <lilithhafner@gmail.com>

* llvmpasses: force vector width for compatibility with non-x86 hosts. (#56300)

The pipeline-prints test currently fails when running on an
aarch64-macos device:

```
/Users/tim/Julia/src/julia/test/llvmpasses/pipeline-prints.ll:309:23: error: AFTERVECTORIZATION: expected string not found in input
; AFTERVECTORIZATION: vector.body
                      ^
<stdin>:2:40: note: scanning from here
; *** IR Dump Before AfterVectorizationMarkerPass on julia_f_199 ***
                                       ^
<stdin>:47:27: note: possible intended match here
; *** IR Dump Before AfterVectorizationMarkerPass on jfptr_f_200 ***
                          ^

Input file: <stdin>
Check file: /Users/tim/Julia/src/julia/test/llvmpasses/pipeline-prints.ll

-dump-input=help explains the following input dump.

Input was:
<<<<<<
             1: opt: WARNING: failed to create target machine for 'x86_64-unknown-linux-gnu': unable to get target for 'x86_64-unknown-linux-gnu', see --version and --triple.
             2: ; *** IR Dump Before AfterVectorizationMarkerPass on julia_f_199 ***
check:309'0                                            X~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ error: no match found
             3: define i64 @julia_f_199(ptr addrspace(10) noundef nonnull align 16 dereferenceable(40) %0) #0 !dbg !4 {
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
             4: top:
check:309'0     ~~~~~
             5:  %1 = call ptr @julia.get_pgcstack()
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
             6:  %ptls_field = getelementptr inbounds ptr, ptr %1, i64 2
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
             7:  %ptls_load45 = load ptr, ptr %ptls_field, align 8, !tbaa !8
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
             .
             .
             .
            42:
check:309'0     ~
            43: L41: ; preds = %L41.loopexit, %L17, %top
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            44:  %value_phi10 = phi i64 [ 0, %top ], [ %7, %L17 ], [ %.lcssa, %L41.loopexit ]
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            45:  ret i64 %value_phi10, !dbg !52
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            46: }
check:309'0     ~~
            47: ; *** IR Dump Before AfterVectorizationMarkerPass on jfptr_f_200 ***
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
check:309'1                               ?                                           possible intended match
            48: ; Function Attrs: noinline optnone
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            49: define nonnull ptr addrspace(10) @jfptr_f_200(ptr addrspace(10) %0, ptr noalias nocapture noundef readonly %1, i32 %2) #1 {
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            50: top:
check:309'0     ~~~~~
            51:  %3 = call ptr @julia.get_pgcstack()
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            52:  %4 = getelementptr inbounds ptr addrspace(10), ptr %1, i32 0
check:309'0     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
             .
             .
             .
>>>>>>

--

********************
Failed Tests (1):
  Julia :: pipeline-prints.ll
```

The problem is that these tests assume x86_64, which fails because the
target isn't available, so it presumably uses the native target which
has different vectorization characteristics:

```
❯ ./usr/tools/opt --load-pass-plugin=libjulia-codegen.dylib -passes='julia' --print-before=AfterVectorization -o /dev/null ../../test/llvmpasses/pipeline-prints.ll
./usr/tools/opt: WARNING: failed to create target machine for 'x86_64-unknown-linux-gnu': unable to get target for 'x86_64-unknown-linux-gnu', see --version and --triple.
```

There's other tests that assume this (e.g. the `fma` cpufeatures one),
but they don't fail, so I've left them as is.

* Reduce generic matrix*vector latency (#56289)

```julia
julia> using LinearAlgebra

julia> A = rand(Int,4,4); x = rand(Int,4); y = similar(x);

julia> @time mul!(y, A, x, 2, 2);
  0.330489 seconds (792.22 k allocations: 41.519 MiB, 8.75% gc time, 99.99% compilation time) # master
  0.134212 seconds (339.89 k allocations: 17.103 MiB, 15.23% gc time, 99.98% compilation time) # This PR
```
Main changes:
- `generic_matvecmul!` and `_generic_matvecmul!` now accept `alpha` and
`beta` arguments instead of `MulAddMul(alpha, beta)`. The methods that
accept a `MulAddMul(alpha, beta)` are also retained for backward
compatibility, but these now forward `alpha` and `beta`, instead of the
other way around.
- Narrow the scope of the `@stable_muladdmul` applications. We now
construct the `MulAddMul(alpha, beta)` object only where it is needed in
a function call, and we annotate the call site with `@stable_muladdmul`.
This leads to smaller branches.
- Create a new internal function with methods for the `'N'`, `'T'` and
`'C'` cases, so that firstly, there's less code duplication, and
secondly, the `_generic_matvecmul!` method is now simple enough to
enable constant propagation. This eliminates the unnecessary branches,
and only the one that is taken is compiled.

Together, this reduces the TTFX substantially.

* Type `Base.is_interactive` as `Bool` (#56303)

Before, typing `Base.is_interactive = 7` would cause weird internal REPL
failures down the line. Now, it throws an InexactError and has no
impact.

* REPL: don't complete str and cmd macros when the input matches the internal name like `r_` to `r"` (#56254)

* fix REPL test if a "juliadev" directory exists in home (#56218)

* Fix trampoline warning on x86 as well (#56280)

* typeintersect: more fastpath to skip intersect under circular env (#56304)

fix #56040

* Preserve type in `first` for `OneTo` (#56263)

With this PR,
```julia
julia> first(Base.OneTo(10), 4)
Base.OneTo(4)
```
Previously, this would have used indexing to return a `UnitRange`. This
is probably the only way to slice a `Base.OneTo` and obtain a
`Base.OneTo` back.

* Matmul: dispatch on specific blas paths using an enum  (#55002)

This expands on the approach taken by
https://github.com/JuliaLang/julia/pull/54552.

We pass on more type information to `generic_matmatmul_wrapper!`, which
lets us convert the branches to method dispatches. This helps spread the
latency around, so that instead of compiling all the branches in the
first call, we now compile the branches only when they are actually
taken. While this reduces the latency in individual branches, there is
no reduction in latency if all the branches are reachable.

```julia
julia> A = rand(2,2);

julia> @time A * A;
  0.479805 seconds (809.66 k allocations: 40.764 MiB, 99.93% compilation time) # 1.12.0-DEV.806
  0.346739 seconds (633.17 k allocations: 31.320 MiB, 99.90% compilation time) # This PR

julia> @time A * A';
  0.030413 seconds (101.98 k allocations: 5.359 MiB, 98.54% compilation time) # v1.12.0-DEV.806
  0.148118 seconds (219.51 k allocations: 11.652 MiB, 99.72% compilation time) # This PR
```
The latency is spread between the two calls here.

In fresh sessions:
```julia
julia> A = rand(2,2);

julia> @time A * A';
  0.473630 seconds (825.65 k allocations: 41.554 MiB, 99.91% compilation time) # v1.12.0-DEV.806
  0.490305 seconds (774.87 k allocations: 38.824 MiB, 99.90% compilation time) # This PR
```
In this case, both the `syrk` and `gemm` branches are reachable, so
there is no reduction in latency.

Analogously, there is a reduction in latency in the second set of matrix
multiplications where we call `symm!/hemm!` or `_generic_matmatmul`:

```julia
julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time Symmetric(A) * A;
  0.711178 seconds (2.06 M allocations: 103.878 MiB, 2.20% gc time, 99.98% compilation time) # v1.12.0-DEV.806
  0.540669 seconds (904.12 k allocations: 43.576 MiB, 2.60% gc time, 97.36% compilation time) # This PR
```

* Scaling `mul!` for generic `AbstractArray`s (#56313)

This improves performance in the scaling `mul!` for `StridedArray`s by
using loops instead of broadcasting.
```julia
julia> using LinearAlgebra

julia> A = zeros(200,200); C = similar(A);

julia> @btime mul!($C, $A, 1, 2, 2);
  19.180 μs (0 allocations: 0 bytes) # nightly v"1.12.0-DEV.1479"
  11.361 μs (0 allocations: 0 bytes) # This PR
```
The latency is reduced as well for the same reason.
```julia
julia> using LinearAlgebra

julia> A = zeros(2,2); C = similar(A);

julia> @time mul!(C, A, 1, 2, 2);
  0.203034 seconds (522.94 k allocations: 27.011 MiB, 14.95% gc time, 99.97% compilation time) # nightly
  0.034713 seconds (59.16 k allocations: 2.962 MiB, 99.91% compilation time) # This PR
```
Thirdly, I've replaced the `.*ₛ` calls by explicit branches. This fixes
the following:
```julia
julia> A = [zeros(2), zeros(2)]; C = similar(A);

julia> mul!(C, A, 1)
ERROR: MethodError: no method matching +(::Vector{Float64}, ::Bool)
```
After this,
```julia
julia> mul!(C, A, 1)
2-element Vector{Vector{Float64}}:
 [0.0, 0.0]
 [0.0, 0.0]
```
Also, I've added `@stable_muladdmul` annotations to the `generic_mul!`
call, but moved it within the loop to narrow its scope. This doesn't
increase the latency, while making the call type-stable.

```julia
julia> D = Diagonal(1:2); C = similar(D);

julia> @time mul!(C, D, 1, 2, 2);
  0.248385 seconds (898.18 k allocations: 47.027 MiB, 12.30% gc time, 99.96% compilation time) # nightly
  0.249940 seconds (919.80 k allocations: 49.128 MiB, 11.36% gc time, 99.99% compilation time) # This PR
```

* InteractiveUtils.jl: fixes issue where subtypes resolves bindings and causes deprecation warnings  (#56306)

The current version of `subtypes` will throw deprecation errors even if
no one is using the deprecated bindings.

A similar bug was fixed in Aqua.jl -
https://github.com/JuliaTesting/Aqua.jl/pull/89/files

See discussion here: 

- https://github.com/JuliaIO/ImageMagick.jl/issues/235 (for identifying
the problem)
- https://github.com/simonster/Reexport.jl/issues/42 (for pointing to
the issue in Aqua.jl)
- https://github.com/JuliaTesting/Aqua.jl/pull/89/files (for the fix in
Aqua.jl)

This adds the `isbindingresolved` test to the `subtypes` function to
avoid throwing deprecation warnings. It also adds a test to check that
this doesn't happen.

---

On the current master branch (before the fix), the added test shows: 
 
```
WARNING: using deprecated binding InternalModule.MyOldType in OuterModule.
, use MyType instead.
Subtypes and deprecations: Test Failed at /home/dgleich/devextern/julia/usr/share/julia/stdlib/v1.12/Test/src/Test.jl:932
  Expression: isempty(stderr_content)
   Evaluated: isempty("WARNING: using deprecated binding InternalModule.MyOldType in OuterModule.\n, use MyType instead.\n")
Test Summary:             | Fail  Total  Time
Subtypes and deprecations |    1      1  2.8s
ERROR: LoadError: Some tests did not pass: 0 passed, 1 failed, 0 errored, 0 broken.
in expression starting at /home/dgleich/devextern/julia/stdlib/InteractiveUtils/test/runtests.jl:841
ERROR: Package InteractiveUtils errored during testing
```

---

Using the results of this pull request:

```
@test_nowarn subtypes(Integer);
```

passes without error. The other tests pass too.

* [CRC32c] Support AbstractVector{UInt8} as input (#56164)

This is a similar PR to https://github.com/JuliaIO/CRC32.jl/pull/12

I added a generic fallback method for `AbstractVector{UInt8}` similar to
the existing generic `IO` method.

Co-authored-by: Steven G. Johnson <stevenj@mit.edu>

* Put `jl_gc_new_weakref` in a header file again (#56319)

* use textwidth for string display truncation (#55442)

It makes a big difference when displaying strings that have width-2 or
width-0 characters.

* Use `pwd()` as the default directory to walk in `walkdir` (#55550)

* Reset mtime of BOLTed files to prevent make rebuilding targets (#55587)

This simplifies the `finish_stage` rule.

Co-authored-by: Zentrik <Zentrik@users.noreply.github.com>

* add docstring note about `displaysize` and `IOContext` with `context` (#55510)

* LinearAlgebra: replace some hardcoded loop ranges with axes (#56243)

These are safer in general, as well as easier to read.

Also, narrow the scopes of some `@inbounds` annotations.

* inference: fix `[modifyfield!|replacefield!]_tfunc`s (#56310)

Currently the following code snippet results in an internal error:
```julia
julia> func(x) = @atomic :monotonic x[].count += 1;

julia> let;Base.Experimental.@force_compile
           x = Ref(nothing)
           func(x)
       end
Internal error: during type inference of
...
```

This issue is caused by the incorrect use of `_fieldtype_tfunc(𝕃, o, f)`
within `modifyfield!_tfunc`, specifically because `o` should be
`widenconst`ed, but it isn’t. By using `_fieldtype_tfunc` correctly, we
can avoid the error through error-catching in `abstract_modifyop!`. This
commit also includes a similar fix for `replacefield!_tfunc` as well.

* inference: don't allow `SSAValue`s in assignment lhs (#56314)

In `InferenceState` the lhs of a `:=` expression should only contain
`GlobalRef` or `SlotNumber` and no other IR elements. Currently when
`SSAValue` appears in `lhs`, the invalid assignment effect is somehow
ignored, but this is incorrect anyway, so this commit removes that
check. Since `SSAValue` should not appear in `lhs` in the first place,
this is not a significant change though.

* Fix `unsafe_read` for `IOBuffer` with non dense data (#55776)

Fixes one part of #54636 

It was only safe to use the following if `from.data` was a dense vector
of bytes.
```julia
GC.@preserve from unsafe_copyto!(p, pointer(from.data, from.ptr), adv)
```

This PR adds a fallback suggested by @matthias314 in
https://discourse.julialang.org/t/copying-bytes-from-abstractvector-to-ptr/119408/7

* support `isless` for zero-dimensional `AbstractArray`s (#55772)

Fixes #55771

* inference: don't add backdge when `applicable` inferred to return `Bool` (#56316)

Also just as a minor backedge reduction optimization, this commit avoids
adding backedges when `applicable` is inferred to return `::Bool`.

* Mark `require_one_based_indexing` and `has_offset_axes` as public (#56196)

The discussion here mentions `require_one_based_indexing` being part of
the public API: https://github.com/JuliaLang/julia/pull/43263

Both functions are also documented (albeit in the dev docs): 
* `require_one_based_indexing`:
https://docs.julialang.org/en/v1/devdocs/offset-arrays/#man-custom-indices
* `has_offset_axes`:
https://docs.julialang.org/en/v1/devdocs/offset-arrays/#For-objects-that-mimic-AbstractArray-but-are-not-subtypes

Towards https://github.com/JuliaLang/julia/issues/51335.

---------

Co-authored-by: Matt Bauman <mbauman@gmail.com>

* Avoid some allocations in various `println` methods (#56308)

* Add a developer documentation section to the `LinearAlgebra` docs (#56324)

Functions that are meant for package developers may go here, instead of
the main section that is primarily for users.

* drop require lock when not needed during loading to allow parallel precompile loading (#56291)

Fixes `_require_search_from_serialized` to first acquire all
start_loading locks (using a deadlock-free batch-locking algorithm)
before doing stalechecks and the rest, so that all the global
computations happen behind the require_lock, then the rest can happen
behind module-specific locks, then (as before) extensions can be loaded
in parallel eventually after `require` returns.

* Make `String(::Memory)` copy (#54457)

A more targeted fix of #54369 than #54372

Preserves the performance improvements added in #53962 by creating a new
internal `_unsafe_takestring!(v::Memory{UInt8})` function that does what
`String(::Memory{UInt8})` used to do.

* 🤖 [master] Bump the Pkg stdlib from 799dc2d54 to 116ba910c (#56336)

Stdlib: Pkg
URL: https://github.com/JuliaLang/Pkg.jl.git
Stdlib branch: master
Julia branch: master
Old commit: 799dc2d54
New commit: 116ba910c
Julia version: 1.12.0-DEV
Pkg version: 1.12.0
Bump invoked by: @IanButterworth
Powered by:
[BumpStdlibs.jl](https://github.com/JuliaLang/BumpStdlibs.jl)

Diff:
https://github.com/JuliaLang/Pkg.jl/compare/799dc2d54c4e809b9779de8c604564a5b3befaa0...116ba910c74ab565d348aa8a50d6dd10148f11ab

```
$ git log --oneline 799dc2d54..116ba910c
116ba910c fix Base.unreference_module call (#4057)
6ed1d2f40 do not show right hand progress without colors (#4047)
```

Co-authored-by: Dilum Aluthge <dilum@aluthge.com>

* Wall-time/all tasks profiler (#55889)

One limitation of sampling CPU/thread profiles, as is currently done in
Julia, is that they primarily capture samples from CPU-intensive tasks.

If many tasks are performing IO or contending for concurrency primitives
like semaphores, these tasks won’t appear in the profile, as they aren't
scheduled on OS threads sampled by the profiler.

A wall-time profiler, like the one implemented in this PR, samples tasks
regardless of OS thread scheduling. This enables profiling of IO-heavy
tasks and detecting areas of heavy contention in the system.

Co-developed with @nickrobinson251.

* recommend explicit `using Foo: Foo, ...` in package code (was: "using considered harmful") (#42080)

I feel we are heading up against a "`using` crisis" where any new
feature that is implemented by exporting a new name (either in Base or a
package) becomes a breaking change. This is already happening
(https://github.com/JuliaGPU/CUDA.jl/pull/1097,
https://github.com/JuliaWeb/HTTP.jl/pull/745) and as projects get bigger
and more names are exported, the likelihood of this rapidly increases.

The flaw in `using Foo` is fundamental in that you cannot lexically see
where a name comes from so when two packages export the same name, you
are screwed. Any code that relies on `using Foo` and then using an
exported name from `Foo` is vulnerable to another dependency exporting
the same name.
Therefore, I think we should start to strongly discourage the use of
`using Foo` and only recommend `using Foo` for ephemeral work (e.g. REPL
work).

---------

Co-authored-by: Dilum Aluthge <dilum@aluthge.com>
Co-authored-by: Mason Protter <mason.protter@icloud.com>
Co-authored-by: Max Horn <max@quendi.de>
Co-authored-by: Matt Bauman <mbauman@juliahub.com>
Co-authored-by: Alex Arslan <ararslan@comcast.net>
Co-authored-by: Ian Butterworth <i.r.butterworth@gmail.com>
Co-authored-by: Neven Sajko <s@purelymail.com>

* Change some hardcoded loop ranges to axes in dense linalg functions (#56348)

These should be safer in general, and are also easier to reason about.

* Make `LinearAlgebra.haszero` public (#56223)

The trait `haszero` is used to check if a type `T` has a unique zero
defined using `zero(T)`. This lets us dispatch to optimized paths
without losing generality. This PR makes the function public so that
this may be extended by packages (such as `StaticArrays`).

* remove spurious parens in profiler docs (#56357)

* Fix `log_quasitriu` for internal scaling `s=0` (#56311)

This PR is a potential fix for #54833.

## Description
The function
https://github.com/JuliaLang/julia/blob/2a06376c18afd7ec875335070743dcebcd85dee7/stdlib/LinearAlgebra/src/triangular.jl#L2220
computes $\boldsymbol{A}^{\dfrac{1}{2^s}} - \boldsymbol{I}$ for a
real-valued $2\times 2$ matrix $\boldsymbol{A}$ using Algorithm 5.1 in
[R1]. However, the algorithm in [R1] as well as the above function do
not handle the case $s=0.$ This fix extends the function to compute
$\boldsymbol{A}^{\dfrac{1}{2^s}} - \boldsymbol{I} \Bigg|_{s=0} =
\boldsymbol{A} - \boldsymbol{I}.$

## Checklist
- [X] Fix code: `stdlib\LinearAlgebra\src\triangular.jl` in function
`_sqrt_pow_diag_block_2x2!(A, A0, s)`.
- [X] Add test case: `stdlib\LinearAlgebra\test\triangular.jl`.
- [X] Update `NEWS.md`.
- [X] Testing and self review.

|  Tag  | Reference |
| --- | --- |
| <nobr>[R1]</nobr> | Al-Mohy, Awad H. and Higham, Nicholas J. "Improved
Inverse Scaling and Squaring Algorithms for the Matrix Logarithm", 2011,
url: https://eprints.maths.manchester.ac.uk/1687/1/paper11.pdf |

---------

Co-authored-by: Daniel Karrasch <daniel.karrasch@posteo.de>
Co-authored-by: Oscar Smith <oscardssmith@gmail.com>

* loading: clean up more concurrency issues (#56329)

Guarantee that `__init__` runs before `using` returns. Could be slightly
breaking for people that do crazy things inside `__init__`, but just
don't do that. Since extensions then probably load after `__init__` (or
at least, run their `__init__` after), this is a partial step towards
changing things so that extensions are guaranteed to load if using all
of their triggers before the corresponding `using` returns

Fixes #55556

* make `_unsetindex` fast for isbits eltype (#56364)

fixes
https://github.com/JuliaLang/julia/issues/56359#issuecomment-2441537634
```
using Plots

function f(n)
    a = Vector{Int}(undef, n)
    s = time_ns()
    resize!(a, 8)
    time_ns() - s
end

x = 8:10:1000000
y = f.(x)

plot(x, y)
```

![image](https://github.com/user-attachments/assets/5a1fb963-7d44-4cac-bedd-6f0733d4cf56)

* improved `eltype` for `flatten` with tuple argument (#55946)

We have always had
```
julia> t = (Int16[1,2], Int32[3,4]); eltype(Iterators.flatten(t))
Any
```
With this PR, the result is `Signed` (`promote_typejoin` applied to the
element types of the tuple elements).

The same applies to `NamedTuple`:
```
julia> nt = (a = [1,2], b = (3,4)); eltype(Iterators.flatten(nt))
Any     # old
Int64   # new
```

* Reland "Reroute (Upper/Lower)Triangular * Diagonal through __muldiag #55984" (#56270)

This relands #55984 which was reverted in #56267. Previously, in #55984,
the destination in multiplying triangular matrices with diagonals was
also assumed to be triangular, which is not necessarily the case in
`mul!`. Tests for this case, however, were being run
non-deterministically, so this wasn't caught by the CI runs.

This improves performance:
```julia
julia> U = UpperTriangular(rand(100,100)); D = Diagonal(rand(size(U,2))); C = similar(U);

julia> @btime mul!($C, $D, $U);
  1.517 μs (0 allocations: 0 bytes) # nightly
  1.116 μs (0 allocations: 0 bytes) # This PR
```

* Add one-arg `norm` method (#56330)

This reduces the latency of `norm` calls, as the single-argument method
lacks branches and doesn't use aggressive constant propagation, and is
therefore simpler to compile. Given that a lot of `norm` calls use
`p==2`, it makes sense for us to reduce the latency on this call.
```julia
julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time norm(A);
  0.247515 seconds (390.09 k allocations: 19.993 MiB, 33.57% gc time, 99.99% compilation time) # master
  0.067201 seconds (121.24 k allocations: 6.067 MiB, 99.98% compilation time) # this PR
```
An example of an improvement in ttfx because of this:
```julia
julia> A = rand(2,2);

julia> @time A ≈ A;
  0.556475 seconds (1.16 M allocations: 59.949 MiB, 24.14% gc time, 100.00% compilation time) # master
  0.333114 seconds (899.85 k allocations: 46.574 MiB, 8.11% gc time, 99.99% compilation time) # this PR
```

* fix a forgotten rename `readuntil`  -> `copyuntil` (#56380)

Fixes https://github.com/JuliaLang/julia/issues/56352, with the repro in
that issue:

```
Master:
  1.114874 seconds (13.01 M allocations: 539.592 MiB, 3.80% gc time)

After:
   0.369492 seconds (12.99 M allocations: 485.031 MiB, 10.73% gc time)

1.10:
    0.341114 seconds (8.36 M allocations: 454.242 MiB, 2.69% gc time)
```

* remove unnecessary operations from `typejoin_union_tuple` (#56379)

Removes the unnecessary call to `unwrap_unionall` and type assertion.

* precompile: fix performance issues with IO (#56370)

The string API here rapidly becomes unusably slow if dumping much debug
output during precompile. Fix the design here to use an intermediate IO
instead to prevent that.

* cache the `find_all_in_cache_path` call during parallel precompilation (#56369)

Before (in an environment with DifferentialEquations.jl):

```julia
julia> @time Pkg.precompile()
  0.733576 seconds (3.44 M allocations: 283.676 MiB, 6.24% gc time)

julia> isfile_calls[1:10]
10-element Vector{Pair{String, Int64}}:
        "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/Printf/3FQLY_zHycD.ji" => 178
        "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/Printf/3FQLY_xxrt3.ji" => 178
         "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/Dates/p8See_xxrt3.ji" => 158
         "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/Dates/p8See_zHycD.ji" => 158
          "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/TOML/mjrwE_zHycD.ji" => 155
          "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/TOML/mjrwE_xxrt3.ji" => 155
                                     "/home/kc/.julia/compiled/v1.12/Preferences/pWSk8_4Qv86.ji" => 152
                                     "/home/kc/.julia/compiled/v1.12/Preferences/pWSk8_juhqb.ji" => 152
 "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/StyledStrings/UcVoM_zHycD.ji" => 144
 "/home/kc/.julia/juliaup/julia-nightly/share/julia/compiled/v1.12/StyledStrings/UcVoM_xxrt3.ji" => 144
 ```  


After:

```julia
julia> @time Pkg.precompile()
  0.460077 seconds (877.59 k allocations: 108.075 MiB, 4.77% gc time)

julia> isfile_calls[1:10]
  10-element Vector{Pair{String, Int64}}:
"/tmp/jl_a5xFWK/Project.toml" => 15
"/tmp/jl_a5xFWK/Manifest.toml" => 7
"/home/kc/.julia/registries/General.toml" => 6

"/home/kc/.julia/juliaup/julia-nightly/share/julia/stdlib/v1.12/Markdown/src/Markdown.jl"
=> 3

"/home/kc/.julia/juliaup/julia-nightly/share/julia/stdlib/v1.12/Serialization/src/Serialization.jl"
=> 3

"/home/kc/.julia/juliaup/julia-nightly/share/julia/stdlib/v1.12/Distributed/src/Distributed.jl"
=> 3

"/home/kc/.julia/juliaup/julia-nightly/share/julia/stdlib/v1.12/UUIDs/src/UUIDs.jl"
=> 3

"/home/kc/.julia/juliaup/julia-nightly/share/julia/stdlib/v1.12/LibCURL/src/LibCURL.jl"
=> 3
```

Performance is improved and we are not calling `isfile` on a bunch of the same ji files hundreds times.

Benchmark is made on a linux machine so performance diff should be a lot better on Windows where these `isfile_casesensitive` call is much more expensive.

Fixes https://github.com/JuliaLang/julia/issues/56366

---------

Co-authored-by: KristofferC <kristoffer.carlsson@juliacomputing.com>
Co-authored-by: Ian Butterworth <i.r.butterworth@gmail.com>

* [docs] Fix note admonition in llvm-passes.md (#56392)

At the moment this is rendered incorrectly:
https://docs.julialang.org/en/v1.11.1/devdocs/llvm-passes/#JuliaLICM

* structure-preserving broadcast for `SymTridiagonal` (#56001)

With this PR, certain broadcasting operations preserve the structure of
a `SymTridiagonal`:
```julia
julia> S = SymTridiagonal([1,2,3,4], [1,2,3])
4×4 SymTridiagonal{Int64, Vector{Int64}}:
 1  1  ⋅  ⋅
 1  2  2  ⋅
 ⋅  2  3  3
 ⋅  ⋅  3  4

julia> S .* 2
4×4 SymTridiagonal{Int64, Vector{Int64}}:
 2  2  ⋅  ⋅
 2  4  4  ⋅
 ⋅  4  6  6
 ⋅  ⋅  6  8
```
This was deliberately disabled on master, but I couldn't find any test
that fails if this is enabled.

* 🤖 [master] Bump the Pkg stdlib from 116ba910c to 9f8e11a4c (#56386)

Stdlib: Pkg
URL: https://github.com/JuliaLang/Pkg.jl.git
Stdlib branch: master
Julia branch: master
Old commit: 116ba910c
New commit: 9f8e11a4c
Julia version: 1.12.0-DEV
Pkg version: 1.12.0
Bump invoked by: @IanButterworth
Powered by:
[BumpStdlibs.jl](https://github.com/JuliaLang/BumpStdlibs.jl)

Diff:
https://github.com/JuliaLang/Pkg.jl/compare/116ba910c74ab565d348aa8a50d6dd10148f11ab...9f8e11a4c0efb3b68a1e25a33f372f398c89cd66

```
$ git log --oneline 116ba910c..9f8e11a4c
9f8e11a4c strip out tree_hash for stdlibs that have have been freed in newer julia versions (#4062)
c0df25a47 rm dead code (#4061)
```

Co-authored-by: Dilum Aluthge <dilum@aluthge.com>

* load extensions with fewer triggers earlier (#49891)

Aimed to support the use case in
https://github.com/JuliaLang/julia/issues/48734#issuecomment-1554626135.

https://github.com/KristofferC/ExtSquared.jl is an example, see
specifically
https://github.com/KristofferC/ExtSquared.jl/blob/ded7c57d6f799674e3310b8174dfb07591bbe025/ext/BExt.jl#L4.

I think this makes sense, happy for a second pair of eyes though.

cc @termi-official

---------

Co-authored-by: KristofferC <kristoffer.carlsson@juliacomputing.com>
Co-authored-by: Cody Tapscott <84105208+topolarity@users.noreply.github.com>

* Dispatch in generic_matmatmul (#56384)

Replacing the branches by dispatch reduces latency, presumably because
there's less dead code in the method.
```julia
julia> using LinearAlgebra

julia> A = rand(Int,2,2); B = copy(A); C = similar(A);

julia> @time mul!(C, A, B, 1, 2);
  0.363944 seconds (1.65 M allocations: 84.584 MiB, 37.57% gc time, 99.99% compilation time) # master
  0.102676 seconds (176.55 k allocations: 8.904 MiB, 27.04% gc time, 99.97% compilation time) # this PR
```
The latency is now distributed between the different branches:
```julia
julia> @time mul!(C, A, B, 1, 2);
  0.072441 seconds (176.55 k allocations: 8.903 MiB, 99.97% compilation time)

julia> @time mul!(C, A', B, 1, 2);
  0.085817 seconds (116.44 k allocations: 5.913 MiB, 99.96% compilation time: 4% of which was recompilation)

julia> @time mul!(C, A', B', 1, 2);
  0.345337 seconds (1.07 M allocations: 54.773 MiB, 25.77% gc time, 99.99% compilation time: 40% of which was recompilation)
```
It would be good to look into why there's recompilation in the last
case, but the branch is less commonly taken than the others that have
significantly lower latency after this PR.

* Add `atol` to addmul tests (#56210)

This avoids the issues as in
https://github.com/JuliaLang/julia/issues/55781 and
https://github.com/JuliaLang/julia/issues/55779 where we compare small
numbers using a relative tolerance. Also, in this PR, I have added an
extra test, so now we compare both `A * B * alpha + C * beta` and `A * B
* alpha - C * beta` with the corresponding in-place versions. The idea
is that if the terms `A * B * alpha` and ` C * beta` have similar
magnitudes, at least one of the two expressions will usually result in a
large enough number that may be compared using a relative tolerance.

I am unsure if the `atol` chosen here is optimal, as I have ballparked
it to use the maximum `eps` by looking at all the `eltype`s involved.

Fixes #55781
Fixes #55779

* Export jl_gc_new_weakref again via julia.h (#56373)

This is how it used for at least Julia 1.0 - 1.11

Closes #56367

* InteractiveUtils: define `InteractiveUtils.@code_ircode` (#56390)

* Fix some missing write barriers and add some helpful comments (#56396)

I was trying some performance optimization which didn't end up working
out, but in the process I found two missing write barriers and added
some helpful comments for future readers, so that part is probably still
useful.

* compiler: fix specialization mistake introduced by #40985 (#56404)

Hopefully there aren't any others like this hiding around? Not useful to
make a new closure for every method that we inline, since we just called
`===` inside it

* Avoid racy double-load of binding restriction in `import_module` (#56395)

Fixes #56333

* define `InteractiveUtils.@infer_[return|exception]_type` (#56398)

Also simplifies the definitions of `@code_typed` and the other similar
macros.

* irinterp: set `IR_FLAG_REFINED` for narrowed `PhiNode`s (#56391)

`adce_pass!` can transform a `Union`-type `PhiNode` into a narrower
`PhiNode`, but in such cases, the `IR_FLAG_REFINED` flag isn’t set on
that `PhiNode` statement. By setting this flag, irinterp can perform
statement reprocessing using the narrowed `PhiNode`, enabling type
stability in cases like JuliaLang/julia#56387.

- fixes JuliaLang/julia#56387

* document isopen(::Channel) (#56376)

This PR has two purposes -- 
1) Add some documentation for public API
2) Add a small note about a footgun I've hit a few times: `!isopen(ch)`
does not mean that you are "done" with the channel because buffered
channels can still have items left in them that need to be taken.

---------

Co-authored-by: CY Han <cyhan.dev@outlook.com>

* Make build system respect `FORCE_COLOR` and `NO_COLOR` settings (#56346)

Follow up to #53742, but for the build system.  CC: @omus.

* Add `edges` vector to CodeInstance/CodeInfo to keep backedges as edges (#54894)

Appears to add about 11MB (128MB to 139MB) to the system image, and to 
decrease the stdlib size by 55 MB (325MB to 270MB), so seems overall 
favorable right now. The edges are computed following the encoding 
<https://hackmd.io/sjPig55kS4a5XNWC6HmKSg?both#Edges-Encoding> to
correctly reflect the backedges.

Co-authored-by: Shuhei Kadowaki <aviatesk@gmail.com>

* docs: remove `dirname.c` from THIRDPARTY file (#56413)

- `dirname.c` was removed by
https://github.com/JuliaLang/julia/commit/c2cec7ad57102e4fbb733b8fb79d617a9524f0ae

* Allow ext → ext dependency if triggers are a strict superset (#56368) (#56402)

Forward port of #56368 - this was a pretty clean port, so it should be
good to go once tests pass.

* [docs] Fix rendering of warning admonition in llvm passes page (#56412)

Follow up to #56392: also the warning in
https://docs.julialang.org/en/v1.11.1/devdocs/llvm-passes/#Multiversioning
is rendered incorrectly because of a missing space.

* Fix dispatch for `rdiv!` with `LU` (#55764)

* Remove overwritten method of OffsetArray (#56414)

This is overwritten three definitions later in
`Base.reshape(A::OffsetArray, inds::Colon)`.

Should remove warnings I saw when testing a package that uses it.

* Add a missing GC root in constant declaration (#56408)

As pointed out in
https://github.com/JuliaLang/julia/pull/56224#discussion_r1816974147.

* Teach compiler about partitioned bindings (#56299)

This commit teaches to compiler to update its world bounds whenever it
looks at a binding partition, making the compiler sound in the presence
of a partitioned binding. The key adjustment is that the compiler is no
longer allowed to directly query the binding table without recording the
world bounds, so all the various abstract evaluations that look at
bindings need to be adjusted and are no longer pure tfuncs. We used to
look at bindings a lot more, but thanks to earlier prep work to remove
unnecessary binding-dependent code (#55288, #55289 and #55271), these
changes become relatively straightforward.

Note that as before, we do not create any binding partitions by default,
so this commit is mostly preperatory.

---------

Co-authored-by: Shuhei Kadowaki <40514306+aviatesk@users.noreply.github.com>

* Restore JL_NOTSAFEPOINT in jl_stderr_obj (#56407)

This is not a function we're really using, but it's used in the
embedding examples, so I'm sure somebody would complain if I deleted it
or made it a safepoint, so let's just give the same best-effort result
as before.

* reland "Inlining: Remove outdated code path for GlobalRef movement (#46880)" (#56382)

From the description of the original PR:
> We used to not allow `GlobalRef` in `PhiNode` at all (because they
> could have side effects). However, we then change the IR to make
> side-effecting `GlobalRef`s illegal in statement position in general,
> so now `PhiNode`s values are just regular value position, so there's
> no reason any more to try to move `GlobalRef`s out to statement
> position in inlining. Moreover, doing so introduces a bunch of
> unnecessary `GlobalRef`s that weren't being moved back. We could fix
> that separately by setting appropriate flags, but it's simpler to just
> get rid of this special case entirely.

This change itself does not sound to have any issues, and in fact, it is
very useful for keeping the IR slim, especially in code generated by
Cassette-like systems, so I would like to reland it.

However, the original PR was reverted in JuliaLang/julia#46951 due to
bugs like JuliaLang/julia#46940 and JuliaLang/julia#46943. I could not
reproduce these bugs on my end (maybe they have been fixed on some
GC-side fixes?), so I believe relanding the original PR’s changes would
not cause any issues, but it is necessary to confirm that similar
problems do not arise before merging this PR.

* copy effects key to `Base.infer_effects` (#56363)

Copied from the docstring of `Core.Compiler.Effects`, this makes it
easier to figure out what the output of `Base.infer_effects` is actually
telling you.

* Fix `make install` for asan build (#56347)

Now the makescript finds libclang_rt.asan-x86_64.so for example.

The change from `-0` to `-1` is as with `-1`, `libclang_rt.asan-*` is
searched for in `usr/lib/julia` instead of `usr/lib`.

* Add dims check to triangular mul (#56393)

This adds a dimension check to triangular matrix multiplication methods.
While such checks already exist in the individual branches (occasionally
within `BLAS` methods), having these earlier would permit certain
optimizations, as we are assured that the axes are compatible. This
potentially duplicates the checks, but this is unlikely to be a concern
given how cheap the checks are.

I've also reused the `check_A_mul_B!_sizes` function that is defined in
`bidiag.jl`, instead of hard-coding the checks.

Further, I've replaced some hard-coded loop ranges by the corresponding
`axes` and `first/lastindex` calls. These are identical under the
1-based indexing assumption, but the `axes` variants are easier to read
and reason about.

* clarify short-circuit && and || docs (#56420)

This clarifies the docs to explain that `a && b` is equivalent to `a ? b
: false` and that `a || b` is equivalent to `a ? true : b`.

In particular, this explains why the second argument does not need to be
a boolean value, which is a common point of confusion. (See e.g. [this
discourse
thread](https://discourse.julialang.org/t/internals-of-assignment-when-doing-short-circuit-evaluation/122178/2?u=stevengj).)

* docs: replace 'leaf types' with 'concrete types' (#56418)

Fixes #55044

---------

Co-authored-by: inkydragon <inkydragon@users.noreply.github.com>

* Remove aggressive constprop annotation on generic_matmatmul_wrapper! (#56400)

This annotation seems unnecessary, as the method gets inlined and
there's no computation being carried out using the value of the
constant.

* Clarify the FieldError docstring (#55222)

* Allow `Time`s to be rounded to `Period`s (#52629)

Co-authored-by: CyHan <git@wo-class.cn>
Co-authored-by: Curtis Vogt <curtis.vogt@gmail.com>

* Replace unconditional store with cmpswap to avoid deadlocking in jl_fptr_wait_for_compiled_addr (#56444)

That unconditional store could overwrite the actual compiled code in
that pointer, so make it a cmpswap

* Correct nothrow modeling of `get_binding_type` (#56430)

As pointed out in
https://github.com/JuliaLang/julia/pull/56299#discussion_r1826509185,
although the bug predates that PR.

* add tip for module docstrings before load (#56445)

* compiler: Strengthen some assertions and fix a couple small bugs (#56449)

* inference: minor follow-ups to JuliaLang/julia#56299 (#56450)

* Ensure that String(::Memory) returns only a String, not any owner (#56438)

Fixes #56435

* Take safepoint lock before going to sleep in the scheduler. (#56443)

This avoids a deadlock during exit. Between a thread going to sleep and
the thread exiting.

* Profile: mention `kill -s SIGUSR1 julia_pid` for Linux (#56441)

currentlu this route is mentioned in docs
https://docs.julialang.org/en/v1/stdlib/Profile/#Triggered-During-Execution
but missing from the module docstring, this should help users who have
little idea how to "send a kernel signal to a process" to get started

---------

Co-authored-by: Ian Butterworth <i.r.butterworth@gmail.com>

* Fix and test an overflow issue in `searchsorted` (#56464)

And remove `searchsorted` special cases for offset arrays in tests that
had the impact of bypassing actually testing `searchsorted` behavior on
offset arrays

To be clear, after this bugfix the function is still broken, just a little bit less so.

* Update docs of calling convention arg in `:foreigncall` AST node (#56417)

* `step(::AbstractUnitRange{Bool})` should return `Bool` (#56405)

The issue was introduced by #27302 , as
```julia
julia> true-false
1
```

By definitions below, `AbstractUnitRange{Bool} <: OrdinalRange{Bool,
Bool}` whose step type is `Bool`.


https://github.com/JuliaLang/julia/blob/da74ef1933b12410b217748e0f7fbcbe52e10d29/base/range.jl#L280-L299

---------

Co-authored-by: Matt Bauman <mbauman@gmail.com>
Co-authored-by: Matt Bauman <mbauman@juliahub.com>

* fixup! JuliaLang/julia#56028, fix up the type-level escapability check

In JuliaLang/julia#56028, the type-level escapability check was changed
to use `is_mutation_free_argtype`, but this was a mistake because EA no
longer runs for structs like
`mutable struct ForeignBuffer{T}; const ptr::Ptr{T}; end`.
This commit changes it to use `is_identity_free_argtype` instead, which
can be used to detect whether a type may contain any mutable allocations
or not.

* add `show(::IO, ::ArgEscapeInfo)`

* EA: disable finalizer inlining for allocations that are edges of `PhiNode`s (#56455)

The current EA-based finalizer inlining implementation can create
invalid IR when the target object is later aliased as a `PhiNode`, which
was causing #56422.
In such cases, finalizer inlining for the allocations that are edges of
each `PhiNode` should be avoided, and instead, finalizer inlining should
ideally be applied to the `PhiNode` itself, but implementing that is
somewhat complex. As a temporary fix, this commit disables inlining in
those cases.

- fixes #56422

* make `verify_ir` error messages more informative (#56452)

Currently, when `verify_ir` finds an error, the `IRCode` is printed, but
it's not easy to determine which method instance generated that
`IRCode`. This commit adds method instance and code location information
to the error message, making it easier to identify the problematic code.

E.g.:
```julia
[...]
610 │    %95 =   builtin Core.tuple(%48, %94)::Tuple{GMT.Gdal.IGeometry, GMT.Gdal.IGeometry}
    └───       return %95

ERROR: IR verification failed.
  Code location:   ~/julia/packages/GMT/src/gdal_extensions.jl:606
  Method instance: MethodInstance for GMT.Gdal.helper_2geoms(::Matrix{Float64}, ::Matrix{Float64})
Stacktrace:
  [1] error(::String, ::String, ::String, ::Symbol, ::String, ::Int32, ::String, ::String, ::Core.MethodInstance)
    @ Core.Compiler ./error.jl:53
  [...]
```

* [GHA] Explicitly install Julia for whitespace workflow (#56468)

So far we relied on the fact that Julia comes in the default Ubuntu
images on GitHub Actions runners, but this may change in the future
(although there's apparently no plan in this direction for the time
being). To make the workflow more future-proof, we now explicitly
install Julia using a dedicated workflow.

* Allow taking Matrix slices without an extra allocation (#56236)

Since changing Array to use Memory as the backing, we had the option of
making non-Vector arrays more flexible, but had instead preserved the
restriction that they must be zero offset and equal in length to the
Memory. This results in extra complexity, restrictions, and allocations
however, but doesn't gain many known benefits. Nanosoldier shows a
decrease in performance on linear eachindex loops, which we theorize is
due to a minor failure to CSE before SCEV or a lack of NUW/NSW on the
length multiplication calculation.

* [late-gc-lowering] null-out GC frame slots for dead objects (#52935)

Should fix https://github.com/JuliaLang/julia/issues/51818.

MWE:

```julia
function testme()
     X = @noinline rand(1_000_000_00)
     Y = @noinline sum(X)
     X = nothing
     GC.gc()
     return Y
 end
```

Note that it now stores a `NULL` in the GC frame before calling
`jl_gc_collect`.

Before:

```llvm
; Function Signature: testme()
;  @ /Users/dnetto/Personal/test.jl:3 within `testme`
define double @julia_testme_535() #0 {
top:
  %gcframe1 = alloca [3 x ptr], align 16
  call void @llvm.memset.p0.i64(ptr align 16 %gcframe1, i8 0, i64 24, i1 true)
  %pgcstack = call ptr inttoptr (i64 6595051180 to ptr)(i64 262) #10
  store i64 4, ptr %gcframe1, align 16
  %task.gcstack = load ptr, ptr %pgcstack, align 8
  %frame.prev = getelementptr inbounds ptr, ptr %gcframe1, i64 1
  store ptr %task.gcstack, ptr %frame.prev, align 8
  store ptr %gcframe1, ptr %pgcstack, align 8
;  @ /Users/dnetto/Personal/test.jl:4 within `testme`
  %0 = call nonnull ptr @j_rand_539(i64 signext 100000000)
  %gc_slot_addr_0 = getelementptr inbounds ptr, ptr %gcframe1, i64 2
  store ptr %0, ptr %gc_slot_addr_0, align 16
;  @ /Users/dnetto/Personal/test.jl:5 within `testme`
  %1 = call double @j_sum_541(ptr nonnull %0)
;  @ /Users/dnetto/Personal/test.jl:7 within `testme`
; ┌ @ gcutils.jl:132 within `gc` @ gcutils.jl:132
   call void @jlplt_ijl_gc_collect_543_got.jit(i32 1)
   %frame.prev4 = load ptr, ptr %frame.prev, align 8
   store ptr %frame.prev4, ptr %pgcstack, align 8
; └
;  @ /Users/dnetto/Personal/test.jl:8 within `testme`
  ret double %1
}
```

After:

```llvm
; Function Signature: testme()
;  @ /Users/dnetto/Personal/test.jl:3 within `testme`
define double @julia_testme_752() #0 {
top:
  %gcframe1 = alloca [3 x ptr], align 16
  call void @llvm.memset.p0.i64(ptr align 16 %gcframe1, i8 0, i64 24, i1 true)
  %pgcstack = call ptr inttoptr (i64 6595051180 to ptr)(i64 262) #10
  store i64 4, ptr %gcframe1, align 16
  %task.gcstack = load ptr, ptr %pgcstack, align 8
  %frame.prev = getelementptr inbounds ptr, ptr %gcframe1, i64 1
  store ptr %task.gcstack, ptr %frame.prev, align 8
  store ptr %gcframe1, ptr %pgcstack, align 8
;  @ /Users/dnetto/Personal/test.jl:4 within `testme`
  %0 = call nonnull ptr @j_rand_756(i64 signext 100000000)
  %gc_slot_addr_0 = getelementptr inbounds ptr, ptr %gcframe1, i64 2
  store ptr %0, ptr %gc_slot_addr_0, align 16
;  @ /Users/dnetto/Personal/test.jl:5 within `testme`
  %1 = call double @j_sum_758(ptr nonnull %0)
  store ptr null, ptr %gc_slot_addr_0, align 16
;  @ /Users/dnetto/Personal/test.jl:7 within `testme`
; ┌ @ gcutils.jl:132 within `gc` @ gcutils.jl:132
   call void @jlplt_ijl_gc_collect_760_got.jit(i32 1)
   %frame.prev6 = load ptr, ptr %frame.prev, align 8
   store ptr %frame.prev6, ptr %pgcstack, align 8
; └
;  @ /Users/dnetto/Personal/test.jl:8 within `testme`
  ret double %1
}
```

* Added test for resolving array references in exprresolve (#56471)

added test to take care of non-real-index handling while resolving array
references in exprresolve to test julia/base/cartesian.jl - line 427 to
432

* Fix and test searchsorted for arrays whose first index is `typemin(Int)` (#56474)

This fixes the issue reported in
https://github.com/JuliaLang/julia/issues/56457#issuecomment-2457223264
which, combined with #56464 which fixed the issue in the OP, fixes #56457.

`searchsortedfirst` was fine all along, but I added it to tests regardless.

* Move Core.Compiler into Base

This is the first step in what I am hoping will eventually result in making
the compiler itself and upgradable stdlib. Over time, we've gained several
non-Base consumers of `Core.Compiler`, and we've reached a bit of a breaking
point where maintaining those downstream dependencies is getting more difficult
than the close coupling of Core.Compiler to the runtime is worth.

In this first step, I am moving Core.Compiler into Base, ending the duplication
of common data structure and generic functions between Core.Compiler and Base.
This split goes back quite far (although not all the way) to the early days of
Julia and predates the world-age mechanism.

The extant Base and Core.Compiler environments have some differences
(other than the duplication). I think the primary ones are (but I will add
more here if somebody points one out).

- `Core.Compiler` does not use `getproperty`
- `Core.Compiler` does not have extensible `==` equality

In this, I decided to retain the former by setting `getproperty = getfield`
for Core.Compiler itself (though of course not for the datatstructures shared
with Base). I don't think it's strictly necessary, but might as well.

For equality, I decided the easiest thing to do would be to try to merge
the equalities and see what happens. In general, Core.Compiler is relatively
restricted in the kinds of equality comparisons it can make, so I think it'll
work out fine, but we can revisit this.

This seems to be fully working and most of this is just moving code around.
I think most of that refactoring is independently useful, so I'll pull some
of it out into separate PRs to make this PR more manageable.

* Delete buggy `stat(::Integer)` method (#54855)

"Where did someone get a RawFD as an integer anyway?" -@stefankarpinski

See also #51711

Fixes #51710

* missing gc-root store in subtype (#56472)

Fixes #56141
Introduced by #52228 (a624d445c02c)

* further defer jl_insert_backedges after loading (#56447)

Finish fully breaking the dependency between method insertions and
inferring whether the cache is valid. The cache should be inferable in
parallel and in aggregate after all loading is finished. This prepares
us for moving this code into Julia (Core.Compiler) next.

* count bytes allocated through malloc more precisely (#55223)

Should make the accounting for memory allocated through malloc a bit
more accurate.

Should also simplify the accounting code by eliminating the use of
`jl_gc_count_freed` in `jl_genericmemory_to_string`.

* Fix external IO loop thead interaction and add function to Base.Experimental to facilitate it's use. Also add a test. (#55529)

While looking at https://github.com/JuliaLang/julia/issues/55525 I found
that the implementation wasn't working correctly.
I added it to Base.Experimental so people don't need to handroll their
own and am also testing a version of what the issue was hitting.

* [REPL] raise default implicit `show` limit to 1MiB (#56297)

https://github.com/JuliaLang/julia/pull/53959#issuecomment-2426946640

I would like to understand more where these issues are coming from; it
would be easy to exempt some types from Base or Core with
```julia
REPL.show_limited(io::IO, mime::MIME, x::SomeType) = show(io, mime, x)
```
but I'm not sure which are causing problems in practice.

But meanwhile I think raising the limit makes sense.

* Add a docstring for `Base.divgcd` (#53769)

Co-authored-by: Sukera <11753998+Seelengrab@users.noreply.github.com>

* Fix compilation warning on aarch64-linux (#56480)

This fixes the warning:
```
/cache/build/default-aws-aarch64-ci-1-3/julialang/julia-master/src/stackwalk.c: In function 'jl_simulate_longjmp':
/cache/build/default-aws-aarch64-ci-1-3/julialang/julia-master/src/stackwalk.c:995:22: warning: initialization of 'mcontext_t *' {aka 'struct sigcontext *'} from incompatible pointer type 'struct unw_sigcontext *' [-Wincompatible-pointer-types]
  995 |     mcontext_t *mc = &c->uc_mcontext;
      |                      ^
```

This is the last remaining warning during compilation on aarch64-linux.

* Make Compiler an independent package

This is a further extension to #56128 to make the compiler into a proper
independent, useable outside of `Base` as `using Compiler` in the same way
that `JuliaSyntax` works already. InteractiveUtils gains a new `@activate`
macro that can be used to activate an outside Compiler package, either for
reflection only or for codegen also.

* Make heap size hint available as an env variable (#55631)

This makes `JULIA_HEAP_SIZE_HINT` the environment variable version of
the `--heap-size-hint` command-line flag. Seems like there was interest
in
https://github.com/JuliaLang/julia/pull/45369#issuecomment-1544204022.

The same syntax is used as for the command-line version with, for
example, `2G` => 2 GB and `200M` => 200 MB.

@oscardssmith want to take a look?

* Allow indexing `UniformScaling` with `CartesianIndex{2}` (#56461)

Since indexing with two `Integer`s is defined, we might as well define
indexing with a `CartesianIndex`. This makes certain loops convenient
where the index is obtained using `eachindex`.

* Simplify first index in `FastContiguousSubArray` definition (#56491)

Since `Slice <: AbstractUnitRange` and `Union{Slice, AbstractUnitRange}
== AbstractUnitRange`, we may simplify the first index.

* Make `popat!` support `@inbounds` (#56323)

Co-authored-by: Jishnu Bhattacharya <jishnub.github@gmail.com>

* NEWS.md: clarify `--trim` (#56460)

Co-authored-by: Matt Bauman <mbauman@gmail.com>

* Remove aggressive constprop annotation from 2x2 and 3x3 matmul (#56453)

Removing these annotations reduces ttfx slightly.
```julia
julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time mul!(similar(A), A, A, 1, 2);
  0.296096 seconds (903.49 k allocations: 44.313 MiB, 4.25% gc time, 99.98% compilation time) # nightly
  0.286009 seconds (835.88 k allocations: 40.732 MiB, 3.29% gc time, 99.98% compilation time) # this PR
```

* `sincos` for non-float symmetric matrices (#56484)

Ensures that the `eltype` of the array to which the result of `sincos`
is a floating-point one, even if the argument doesn't have a
floating-point `eltype`.

After this, the following works:
```julia
julia> A = diagm(0=>1:3)
3×3 Matrix{Int64}:
 1  0  0
 0  2  0
 0  0  3

julia> sincos(A)
([0.8414709848078965 0.0 0.0; 0.0 0.9092974268256817 0.0; 0.0 0.0 0.1411200080598672], [0.5403023058681398 0.0 0.0; 0.0 -0.4161468365471424 0.0; 0.0 0.0 -0.9899924966004454])
```

* Specialize 2-arg `show` for `LinearIndices` (#56482)

After this,
```julia
julia> l = LinearIndices((1:3, 1:4));

julia> show(l)
LinearIndices((1:3, 1:4))
```
The printed form is a valid constructor.

* Avoid constprop in `syevd!` and `syev!` (#56442)

This improves compilation times slightly:
```julia
julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time eigen!(Hermitian(A));
  0.163380 seconds (180.51 k allocations: 8.760 MiB, 99.88% compilation time) # master
  0.155285 seconds (163.77 k allocations: 7.971 MiB, 99.87% compilation time) # This PR
```
The idea is that the constant propagation is only required to infer the
return type, and isn't necessary in the body of the method. We may
therefore annotate the body with a `@constprop :none`.

* make: define `basecompiler.ji` target (#56498)

For easier experimentation with just the bootstrap process.

Additionally, as a follow-up to JuliaLang/julia#56409, this commit also
includes some minor cosmetic changes.

* speed up bootstrapping by compiling few optimizer subroutines earlier (#56501)

Speeds up the bootstrapping process by about 30 seconds.

* remove top-level branches checking for Base (#56507)

These are no longer needed, now that the files are no longer included
twice.

* Undo the decision to publish incomplete types to the binding table (#56497)

This effectively reverts #36121 and replaces it with #36111, which was
the originally proposed alternative to fix #36104. To recap, the
question is what should happen for
```
module Foo
    struct F
        v::Foo.F
    end
end
```
i.e. where the type reference tries to refer to the newly defined type
via its global path. In #36121 we adjusted things so that we first
assign the type to its global binding and then evaluate the field type
(leaving the type in an incomplete state in the meantime). The primary
reason that this choice was that we would have to deal with incomplete
types assigned to global bindings anyway if we ever did #32658. However,
I think this was the wrong choice. There is a difference between
allowing incomplete types and semantically forcing incomplete types to
be globally observable every time a new type is defined.

The situation was a little different four years ago, but with more
extensive threading (which can observe the incompletely constructed
type) and the upcoming completion of bindings partition, the situation
is different. For bindings partition in particular, this would require
two invalidations on re-definition, one to the new incomplete type and
then back to the complete type. I don't think this is worth it, for the
(somewhat niche and possibly-should-be- deprecated-future) case of
refering to incompletely defined types by their global names.

So let's instead try the hack in #36111, which does a frontend rewrite
of the global path. This should be sufficient to at least address the
obvious cases.

* Merge identical methods for Symmetric/Hermitian and SymTridiagonal (#56434)

Since the methods do identical things, we may define each method once
for a union of types instead of defining methods for each type.

* Specialize findlast for integer AbstractUnitRanges and StepRanges (#54902)

For monotonic ranges, `findfirst` and `findlast` with `==(val)` as the
predicate should be identical, as each value appears only once in the
range. Since `findfirst` is specialized for some ranges, we may define
`findlast` as well analogously.

On v"1.12.0-DEV.770"
```julia
julia> @btime findlast(==(1), $(Ref(1:1_000))[])
  1.186 μs (0 allocations: 0 bytes)
1
```
This PR
```julia
julia> @btime findlast(==(1), $(Ref(1:1_000))[])
  3.171 ns (0 allocations: 0 bytes)
1
```

I've also specialized `findfirst(iszero, r::AbstractRange)` to make this
be equivalent to `findfirst(==(0), ::AbstractRange)` for numerical
ranges. Similarly, for `isone`. These now take the fast path as well.

Thirdly, I've added some `convert` calls to address issues like
```julia
julia> r = Int128(1):Int128(1):Int128(4);

julia> findfirst(==(Int128(2)), r) |> typeof
Int128

julia> keytype(r)
Int64
```
This PR ensures that the return type always corresponds to `keytype`,
which is what the docstring promises.

This PR also fixes
```julia
julia> findfirst(==(0), UnitRange(-0.5, 0.5))
ERROR: InexactError: Int64(0.5)
Stacktrace:
 [1] Int64
   @ ./float.jl:994 [inlined]
 [2] findfirst(p::Base.Fix2{typeof(==), Int64}, r::UnitRange{Float64})
   @ Base ./array.jl:2397
 [3] top-level scope
   @ REPL[1]:1
```
which now returns `nothing`, as expected.

* Loop over `Iterators.rest` in `_foldl_impl` (#56492)

For reasons that I don't understand, this improves performance in
`mapreduce` in the following example:
```julia
julia> function g(A)
           for col in axes(A,2)
               mapreduce(iszero, &, view(A, UnitRange(axes(A,1)), col), init=true) || return false
           end
           return true
       end
g (generic function with 2 methods)

julia> A = zeros(2, 10000);

julia> @btime g($A);
  28.021 μs (0 allocations: 0 bytes) # nightly v"1.12.0-DEV.1571"
  12.462 μs (0 allocations: 0 bytes) # this PR

julia> A = zeros(1000,1000);

julia> @btime g($A);
  372.080 μs (0 allocations: 0 bytes) # nightly
  321.753 μs (0 allocations: 0 bytes) # this PR
```
It would be good to understand what the underlying issue is, as the two
seem equivalent to me. Perhaps this form makes it clear that it's not,
in fact, an infinite loop?

* better error message for rpad/lpad with zero-width padding (#56488)

Closes #45339 — throw a more informative `ArgumentError` message from
`rpad` and `lpad` if a zero-`textwidth` padding is passed (not a
`DivideError`).

If the padding character has `ncodeunits == 1`, suggests that maybe they
want `str * pad^max(0, npad - ncodeunits(str))` instead.

* Safer indexing in dense linalg methods (#56451)

Ensure that `eachindex` is used consistently alongside `@inbounds`, and
use `diagind` to obtain indices along a diagonal.

* The `info` in LAPACK calls should be a Ref instead of a Ptr (#56511)

Co-authored-by: Viral B. Shah <ViralBShah@users.noreply.github.com>

* Scaling loop instead of broadcasting in strided matrix exp (#56463)

Firstly, this is easier to read. Secondly, this merges the two loops
into one. Thirdly, this avoids the broadcasting latency.
```julia
julia> using LinearAlgebra

julia> A = rand(2,2);

julia> @time LinearAlgebra.exp!(A);
  0.952597 seconds (2.35 M allocations: 116.574 MiB, 2.67% gc time, 99.01% compilation time) # master
  0.877404 seconds (2.17 M allocations: 106.293 MiB, 2.65% gc time, 99.99% compilation time) # this PR
```
The performance also improves as there are fewer allocations in the
first branch (`opnorm(A, 1) <= 2.1`):
```julia
julia> B = diagm(0=>im.*(float.(1:200))./200, 1=>(1:199)./400, -1=>(1:199)./400);

julia> opnorm(B,1)
1.9875

julia> @btime exp($B);
  5.066 ms (30 allocations: 4.89 MiB) # nightly v"1.12.0-DEV.1581"
  4.926 ms (27 allocations: 4.28 MiB) # this PR
```

* codegen: Respect binding partition (#56494)

Minor changes to make codegen correct in the face of partitioned
constant bindings. Does not yet handle the envisioned semantics for
globals that change restriction type, which will require a fair bit of
additional work.

* Profile: fix Compiler short path (#56515)

* Check `isdiag` in dense trig functions (#56483)

This improves performance for dense diagonal matrices, as we may apply
the function only to the diagonal elements.
```julia
julia> A = diagm(0=>rand(100));

julia> @btime cos($A);
  349.211 μs (22 allocations: 401.58 KiB) # nightly v"1.12.0-DEV.1571"
  16.215 μs (7 allocations: 80.02 KiB) # this PR
```

---------

Co-authored-by: Daniel Karrasch <daniel.karrasch@posteo.de>

* Profile: add helper method for printing profile report to file (#56505)

The IOContext part is isn't obvious, because otherwise the IO is assumed
to be 80 chars wide, which makes for bad reports.

* Change in-place exp to out-of-place in matrix trig functions (#56242)

This makes the functions work for arbitrary matrix types that support
`exp`, but not necessarily the in-place `exp!`. For example, the
following works after this:
```julia
julia> m = SMatrix{2,2}(1:4);

julia> cos(m)
2×2 SMatrix{2, 2, Float64, 4} with indices SOneTo(2)×SOneTo(2):
  0.855423  -0.166315
 -0.110876   0.689109
```
There's a slight performance improvement as well because we don't
compute `im*A` and `-im*A` separately, but we negate the first to obtain
the second.
```julia
julia> A = rand(ComplexF64,100,100);

julia> @btime sin($A);
  2.796 ms (48 allocations: 1.84 MiB) # nightly v"1.12.0-DEV.1571"
  2.304 ms (48 allocations: 1.84 MiB) # this PR
```

* Test: Don't change scope kind in `test_{warn,nowarn}` (#56524)

This was part of #56509, but is an independent bugfix. The basic issue
is that these macro were using `do` block internally. This is
undesirable for test macros, because we would like them not to affect
the behavior of what they're testing. E.g. right now:
```
julia> using Test

julia> const x = 1
1

julia> @test_nowarn const x = 1
ERROR: syntax: `global const` declaration not allowed inside function around /home/keno/julia/usr/share/julia/stdlib/v1.12/Test/src/Test.jl:927
Stacktrace:
 [1] top-level scope
   @ REPL[3]:1
```

This PR just writes out the try/finally manually, so the above works
fine after this PR.

* For loop instead of while in generic `copyto!` (#56517)

This appears to improve performance.
```julia
j…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
GC Garbage collector
Projects
None yet
Development

Successfully merging a pull request may close this issue.

7 participants