threading: support more than nthreads at runtime

Hook a couple functions (notably cfunction) to handle adopting
foreign threads automatically when used.

n.b. If returning an object pointer, we do not gc_unsafe_leave
afterwards as that would render the pointer invalid. However, this means
that it can be a long time before the next safepoint (if ever). We
should look into ways of improving this bad situation, such as pinning
only that specific object temporarily.

n.b. There are some remaining issues to clean up. For example, we may
trap pages in the ptls after GC to keep them "warm", and trap other
pages in the unwind buffer, etc.

base/deprecated.jl

@@ -327,4 +327,6 @@ function setproperty!(ci::CodeInfo, s::Symbol, v)
     return setfield!(ci, s, convert(fieldtype(CodeInfo, s), v))
 end
 
+@eval Threads nthreads() = threadpoolsize()
+
 # END 1.9 deprecations

base/partr.jl

@@ -2,7 +2,7 @@
 
 module Partr
 
-using ..Threads: SpinLock, nthreads, threadid
+using ..Threads: SpinLock, maxthreadid, threadid
 
 # a task minheap
 mutable struct taskheap

base/pcre.jl

@@ -28,7 +28,7 @@ THREAD_MATCH_CONTEXTS::Vector{Ptr{Cvoid}} = [C_NULL]
 PCRE_COMPILE_LOCK = nothing
 
 _tid() = Int(ccall(:jl_threadid, Int16, ())) + 1
-_nth() = Int(unsafe_load(cglobal(:jl_n_threads, Cint)))
+_mth() = Int(Core.Intrinsics.atomic_pointerref(cglobal(:jl_n_threads, Cint), :acquire))
 
 function get_local_match_context()
     tid = _tid()
@@ -40,7 +40,7 @@ function get_local_match_context()
         try
             ctxs = THREAD_MATCH_CONTEXTS
             if length(ctxs) < tid
-                global THREAD_MATCH_CONTEXTS = ctxs = copyto!(fill(C_NULL, _nth()), ctxs)
+                global THREAD_MATCH_CONTEXTS = ctxs = copyto!(fill(C_NULL, length(ctxs) + _mth()), ctxs)
             end
         finally
             unlock(l)

base/task.jl

@@ -733,7 +733,7 @@ function workqueue_for(tid::Int)
     @lock l begin
         qs = Workqueues
         if length(qs) < tid
-            nt = Threads.nthreads()
+            nt = Threads.maxthreadid()
             @assert tid <= nt
             global Workqueues = qs = copyto!(typeof(qs)(undef, length(qs) + nt - 1), qs)
         end
@@ -746,7 +746,7 @@ end
 
 function enq_work(t::Task)
     (t._state === task_state_runnable && t.queue === nothing) || error("schedule: Task not runnable")
-    if t.sticky || Threads.nthreads() == 1
+    if t.sticky || Threads.threadpoolsize() == 1
         tid = Threads.threadid(t)
         if tid == 0
             # Issue #41324

base/threadingconstructs.jl

@@ -1,7 +1,6 @@
 # This file is a part of Julia. License is MIT: https://julialang.org/license
 
-export threadid, nthreads, @threads, @spawn,
-       threadpool, nthreadpools
+export threadid, @threads, @spawn, threadpool, nthreadpools
 
 """
     Threads.threadid() -> Int
@@ -11,20 +10,25 @@ ID `1`.
 """
 threadid() = Int(ccall(:jl_threadid, Int16, ())+1)
 
+# lower bound on the largest threadid()
 """
-    Threads.nthreads([:default|:interactive]) -> Int
+    Threads.maxthreadid() -> Int
 
-Get the number of threads (across all thread pools or within the specified
-thread pool) available to Julia. The number of threads across all thread
-pools is the inclusive upper bound on [`threadid()`](@ref).
+Get a lower bound on the number of threads (across all thread pools or within the specified
+thread pool) available to the Julia process >= [`threadid()`](@ref), with atomic-acquire semantics.
+"""
+maxthreadid() = Int(Core.Intrinsics.atomic_pointerref(cglobal(:jl_n_threads, Cint), :acquire))
 
-See also: `BLAS.get_num_threads` and `BLAS.set_num_threads` in the
-[`LinearAlgebra`](@ref man-linalg) standard library, and `nprocs()` in the
-[`Distributed`](@ref man-distributed) standard library.
 """
-function nthreads end
+    Threads.nthreads(:default | :interactive) -> Int
 
-nthreads() = Int(unsafe_load(cglobal(:jl_n_threads, Cint)))
+Get the current number of threads within the specified thread pool. The threads in default
+have id numbers `1:nthreads(:default)`.
+
+See also `BLAS.get_num_threads` and `BLAS.set_num_threads` in the [`LinearAlgebra`](@ref
+man-linalg) standard library, and `nprocs()` in the [`Distributed`](@ref man-distributed)
+standard library and [`Threads.maxthreadid()`](@ref).
+"""
 function nthreads(pool::Symbol)
     if pool === :default
         tpid = Int8(0)
@@ -35,6 +39,7 @@ function nthreads(pool::Symbol)
     end
     return _nthreads_in_pool(tpid)
 end
+
 function _nthreads_in_pool(tpid::Int8)
     p = unsafe_load(cglobal(:jl_n_threads_per_pool, Ptr{Cint}))
     return Int(unsafe_load(p, tpid + 1))
@@ -57,10 +62,21 @@ Returns the number of threadpools currently configured.
 """
 nthreadpools() = Int(unsafe_load(cglobal(:jl_n_threadpools, Cint)))
 
+"""
+    Threads.threadpoolsize()
+
+Get the number of threads available to the Julia default worker thread-pool.
+
+See also: `BLAS.get_num_threads` and `BLAS.set_num_threads` in the
+[`LinearAlgebra`](@ref man-linalg) standard library, and `nprocs()` in the
+[`Distributed`](@ref man-distributed) standard library.
+"""
+threadpoolsize() = Threads._nthreads_in_pool(Int8(0))
+#threadpoolsize() = max(Int(JLOptions().nthreads), 1)
 
 function threading_run(fun, static)
     ccall(:jl_enter_threaded_region, Cvoid, ())
-    n = nthreads()
+    n = threadpoolsize()
     tasks = Vector{Task}(undef, n)
     for i = 1:n
         t = Task(() -> fun(i)) # pass in tid
@@ -93,7 +109,7 @@ function _threadsfor(iter, lbody, schedule)
                 tid = 1
                 len, rem = lenr, 0
             else
-                len, rem = divrem(lenr, nthreads())
+                len, rem = divrem(lenr, threadpoolsize())
             end
             # not enough iterations for all the threads?
             if len == 0
@@ -185,7 +201,7 @@ assumption may be removed in the future.
 This scheduling option is merely a hint to the underlying execution mechanism. However, a
 few properties can be expected. The number of `Task`s used by `:dynamic` scheduler is
 bounded by a small constant multiple of the number of available worker threads
-([`nthreads()`](@ref Threads.nthreads)). Each task processes contiguous regions of the
+([`Threads.threadpoolsize()`](@ref)). Each task processes contiguous regions of the
 iteration space. Thus, `@threads :dynamic for x in xs; f(x); end` is typically more
 efficient than `@sync for x in xs; @spawn f(x); end` if `length(xs)` is significantly
 larger than the number of the worker threads and the run-time of `f(x)` is relatively
@@ -222,15 +238,15 @@ julia> function busywait(seconds)
 
 julia> @time begin
             Threads.@spawn busywait(5)
-            Threads.@threads :static for i in 1:Threads.nthreads()
+            Threads.@threads :static for i in 1:Threads.threadpoolsize()
                 busywait(1)
             end
         end
 6.003001 seconds (16.33 k allocations: 899.255 KiB, 0.25% compilation time)
 
 julia> @time begin
             Threads.@spawn busywait(5)
-            Threads.@threads :dynamic for i in 1:Threads.nthreads()
+            Threads.@threads :dynamic for i in 1:Threads.threadpoolsize()
                 busywait(1)
             end
         end

base/threads.jl

@@ -11,25 +11,4 @@ include("threadingconstructs.jl")
 include("atomics.jl")
 include("locks-mt.jl")
 
-
-"""
-    resize_nthreads!(A, copyvalue=A[1])
-
-Resize the array `A` to length [`nthreads()`](@ref).   Any new
-elements that are allocated are initialized to `deepcopy(copyvalue)`,
-where `copyvalue` defaults to `A[1]`.
-
-This is typically used to allocate per-thread variables, and
-should be called in `__init__` if `A` is a global constant.
-"""
-function resize_nthreads!(A::AbstractVector, copyvalue=A[1])
-    nthr = nthreads()
-    nold = length(A)
-    resize!(A, nthr)
-    for i = nold+1:nthr
-        A[i] = deepcopy(copyvalue)
-    end
-    return A
-end
-
 end

base/threads_overloads.jl

@@ -3,7 +3,7 @@
 """
     Threads.foreach(f, channel::Channel;
                     schedule::Threads.AbstractSchedule=Threads.FairSchedule(),
-                    ntasks=Threads.nthreads())
+                    ntasks=Base.threadpoolsize())
 
 Similar to `foreach(f, channel)`, but iteration over `channel` and calls to
 `f` are split across `ntasks` tasks spawned by `Threads.@spawn`. This function
@@ -40,7 +40,7 @@ collect(d) = [1, 4, 9, 16, 25, 36, 49, 64, 81, 100, 121, 144, 169, 196, 225, 256
 """
 function Threads.foreach(f, channel::Channel;
                          schedule::Threads.AbstractSchedule=Threads.FairSchedule(),
-                         ntasks=Threads.nthreads())
+                         ntasks=Threads.threadpoolsize())
     apply = _apply_for_schedule(schedule)
     stop = Threads.Atomic{Bool}(false)
     @sync for _ in 1:ntasks

cli/loader_exe.c

@@ -15,7 +15,7 @@ extern "C" {
 JULIA_DEFINE_FAST_TLS
 
 #ifdef _COMPILER_ASAN_ENABLED_
-JL_DLLEXPORT const char* __asan_default_options()
+JL_DLLEXPORT const char* __asan_default_options(void)
 {
     return "allow_user_segv_handler=1:detect_leaks=0";
     // FIXME: enable LSAN after fixing leaks & defining __lsan_default_suppressions(),

contrib/generate_precompile.jl

@@ -1,7 +1,7 @@
 # This file is a part of Julia. License is MIT: https://julialang.org/license
 
-if Threads.nthreads() != 1
-    @warn "Running this file with multiple Julia threads may lead to a build error" Threads.nthreads()
+if Threads.maxthreadid() != 1
+    @warn "Running this file with multiple Julia threads may lead to a build error" Base.maxthreadid()
 end
 
 if Base.isempty(Base.ARGS) || Base.ARGS[1] !== "0"
@@ -340,7 +340,7 @@ function generate_precompile_statements()
                 # wait for the next prompt-like to appear
                 readuntil(output_copy, "\n")
                 strbuf = ""
-                while true
+                while !eof(output_copy)
                     strbuf *= String(readavailable(output_copy))
                     occursin(JULIA_PROMPT, strbuf) && break
                     occursin(PKG_PROMPT, strbuf) && break

doc/src/base/multi-threading.md

@@ -5,9 +5,11 @@ Base.Threads.@threads
 Base.Threads.foreach
 Base.Threads.@spawn
 Base.Threads.threadid
+Base.Threads.maxthreadid
 Base.Threads.nthreads
 Base.Threads.threadpool
 Base.Threads.nthreadpools
+Base.Threads.threadpoolsize
 ```
 
 See also [Multi-Threading](@ref man-multithreading).

doc/src/manual/embedding.md

@@ -604,7 +604,7 @@ The second condition above implies that you can not safely call `jl_...()` funct
 void *func(void*)
 {
     // Wrong, jl_eval_string() called from thread that was not started by Julia
-    jl_eval_string("println(Threads.nthreads())");
+    jl_eval_string("println(Threads.threadid())");
     return NULL;
 }
 
@@ -630,7 +630,7 @@ void *func(void*)
     // Okay, all jl_...() calls from the same thread,
     // even though it is not the main application thread
     jl_init();
-    jl_eval_string("println(Threads.nthreads())");
+    jl_eval_string("println(Threads.threadid())");
     jl_atexit_hook(0);
     return NULL;
 }
@@ -670,7 +670,7 @@ int main()
     jl_eval_string("func(i) = ccall(:c_func, Float64, (Int32,), i)");
 
     // Call func() multiple times, using multiple threads to do so
-    jl_eval_string("println(Threads.nthreads())");
+    jl_eval_string("println(Base.threadpoolsize())");
     jl_eval_string("use(i) = println(\"[J $(Threads.threadid())] i = $(i) -> $(func(i))\")");
     jl_eval_string("Threads.@threads for i in 1:5 use(i) end");
 

doc/src/manual/multi-threading.md

@@ -6,10 +6,10 @@ of Julia multi-threading features.
 ## Starting Julia with multiple threads
 
 By default, Julia starts up with a single thread of execution. This can be verified by using the
-command [`Threads.nthreads()`](@ref):
+command [`Threads.threadpoolsize()`](@ref):
 
 ```jldoctest
-julia> Threads.nthreads()
+julia> Threads.threadpoolsize()
 1
 ```
 
@@ -38,7 +38,7 @@ $ julia --threads 4
 Let's verify there are 4 threads at our disposal.
 
 ```julia-repl
-julia> Threads.nthreads()
+julia> Threads.threadpoolsize()
 4
 ```
 
@@ -267,7 +267,7 @@ avoid the race:
 ```julia-repl
 julia> using Base.Threads
 
-julia> nthreads()
+julia> Threads.threadpoolsize()
 4
 
 julia> acc = Ref(0)

src/ccall.cpp

@@ -1540,7 +1540,8 @@ static jl_cgval_t emit_ccall(jl_codectx_t &ctx, jl_value_t **args, size_t nargs)
         assert(lrt == getVoidTy(ctx.builder.getContext()));
         assert(!isVa && !llvmcall && nccallargs == 0);
         JL_GC_POP();
-        emit_gc_safepoint(ctx);
+        ctx.builder.CreateCall(prepare_call(gcroot_flush_func));
+        emit_gc_safepoint(ctx.builder, get_current_ptls(ctx), ctx.tbaa().tbaa_const);
         return ghostValue(ctx, jl_nothing_type);
     }
     else if (is_libjulia_func("jl_get_ptls_states")) {
@@ -1643,7 +1644,8 @@ static jl_cgval_t emit_ccall(jl_codectx_t &ctx, jl_value_t **args, size_t nargs)
         ctx.builder.SetInsertPoint(checkBB);
         ctx.builder.CreateLoad(
                 getSizeTy(ctx.builder.getContext()),
-                ctx.builder.CreateConstInBoundsGEP1_32(getSizeTy(ctx.builder.getContext()), get_current_signal_page(ctx), -1),
+                ctx.builder.CreateConstInBoundsGEP1_32(getSizeTy(ctx.builder.getContext()),
+                    get_current_signal_page_from_ptls(ctx.builder, get_current_ptls(ctx), ctx.tbaa().tbaa_const), -1),
                 true);
         ctx.builder.CreateBr(contBB);
         ctx.f->getBasicBlockList().push_back(contBB);