Fix type instability pr28

Project.toml

@@ -5,14 +5,12 @@ version = "0.2.2"
 
 [deps]
 BitTwiddlingConvenienceFunctions = "62783981-4cbd-42fc-bca8-16325de8dc4b"
-CPUSummary = "2a0fbf3d-bb9c-48f3-b0a9-814d99fd7ab9"
 Static = "aedffcd0-7271-4cad-89d0-dc628f76c6d3"
 ThreadingUtilities = "8290d209-cae3-49c0-8002-c8c24d57dab5"
 
 [compat]
 Aqua = "0.8"
 BitTwiddlingConvenienceFunctions = "0.1"
-CPUSummary = "0.1.2, 0.2"
 Static = "1"
 Test = "<0.1, 1.10"
 ThreadingUtilities = "0.5"

src/request.jl

@@ -1,13 +1,12 @@
-import CPUSummary
-
 function worker_bits()
-  wts = nextpow2(CPUSummary.sys_threads()) # Typically sys_threads (i.e. Sys.CPU_THREADS) does not change between runs, thus it will precompile well.
+  wts = nextpow2(Threads.nthreads()) # Typically sys_threads (i.e. Sys.CPU_THREADS) does not change between runs, thus it will precompile well.
   ws = static(8sizeof(UInt))               # For testing purposes it can be overridden by JULIA_CPU_THREADS,
-  ifelse(Static.lt(wts, ws), ws, wts)
+  # Always return Int to avoid type instability with high thread counts
+  ifelse(wts < 64, 64, wts)
 end
 function worker_mask_count()
   bits = worker_bits()
-  (bits + StaticInt{63}()) ÷ StaticInt{64}() # cld not defined on `StaticInt`
+  cld(bits, 64)
 end
 
 worker_pointer() = Base.unsafe_convert(Ptr{UInt}, pointer_from_objref(WORKERS))
@@ -63,6 +62,30 @@ end
   (ui,), (ft,)
 end
 
+# Handle regular Int for type stability with high thread counts
+@inline function _request_threads(
+  num_requested::UInt32,
+  wp::Ptr,
+  N::Int,
+  threadmask
+)
+  if N == 1
+    ui, ft, num_requested, wp =
+      __request_threads(num_requested, wp, _first(threadmask))
+    return (ui,), (ft,)
+  else
+    ui, ft, num_requested, wp =
+      __request_threads(num_requested, wp, _first(threadmask))
+    uit, ftt = _request_threads(
+      num_requested,
+      wp,
+      N - 1,
+      _remaining(threadmask)
+    )
+    return (ui, uit...), (ft, ftt...)
+  end
+end
+
 @inline function _exchange_mask!(wp, ::Nothing)
   all_threads = _atomic_xchg!(wp, zero(UInt))
   all_threads, all_threads

test/runtests.jl

@@ -28,3 +28,4 @@ using Test
   end
 end
 Aqua.test_all(PolyesterWeave)
+include("test_high_thread_count.jl")

test/test_high_thread_count.jl

@@ -0,0 +1,32 @@
+using Test
+using PolyesterWeave
+using BitTwiddlingConvenienceFunctions: nextpow2
+
+@testset "High thread count compatibility" begin
+    # Test worker_bits returns Int for all thread counts
+    @test isa(PolyesterWeave.worker_bits(), Int)
+
+    # Test worker_mask_count returns Int
+    @test isa(PolyesterWeave.worker_mask_count(), Int)
+
+    # Test that request_threads works with high thread counts
+    # This simulates the case where worker_mask_count() > 1
+    if Threads.nthreads() > 64
+        # With > 64 threads, worker_mask_count() should be 2 or more
+        @test PolyesterWeave.worker_mask_count() >= 2
+
+        # Test that request_threads doesn't throw
+        threads, torelease = PolyesterWeave.request_threads(10)
+        @test length(threads) >= 0  # May get 0 if no threads available
+
+        # Free the threads
+        PolyesterWeave.free_threads!(torelease)
+    else
+        # With <= 64 threads, worker_mask_count() should be 1
+        @test PolyesterWeave.worker_mask_count() == 1
+    end
+
+    # Test specific values
+    @test PolyesterWeave.worker_bits() == max(64, nextpow2(Threads.nthreads()))
+    @test PolyesterWeave.worker_mask_count() == cld(PolyesterWeave.worker_bits(), 64)
+end