win,Profile: fix timeout units (JuliaLang#38208)

Previously, we were using units of seconds (rounded down) but the
function expected milliseconds. So this was effectively always zero
(ignored) in the computation.

Empirically, the documentation for Sleep is fundamentally incorrect, and
the stated advice to add `wPeriodMin` is bad. So here we remove that.
(Microsoft has given the thumbs up to this observation also now.)

Also update the Profile code defaults to sample for longer, and faster,
since only the Win32 stack-unwinder (dbghelp) is absurdly slow.

Co-authored-by: kimikage <kimikage.ceo@gmail.com>

src/signals-win.c

@@ -337,9 +337,8 @@ static DWORD WINAPI profile_bt( LPVOID lparam )
     }
     timeBeginPeriod(tc.wPeriodMin);
     while (1) {
-        DWORD timeout = nsecprof / GIGA;
-        timeout += tc.wPeriodMin;
-        Sleep(timeout);
+        DWORD timeout_ms = nsecprof / (GIGA / 1000);
+        Sleep(timeout_ms > 0 ? timeout_ms : 1);
         if (bt_size_cur < bt_size_max && running) {
             JL_LOCK_NOGC(&jl_in_stackwalk);
             jl_lock_profile();

stdlib/Profile/src/Profile.jl

@@ -64,11 +64,14 @@ function init(n::Integer, delay::Real)
 end
 
 # init with default values
-# Use a max size of 1M profile samples, and fire timer every 1ms
-if Sys.iswindows()
+# Use a max size of 10M profile samples, and fire timer every 1ms
+# (that should typically give around 100 seconds of record)
+if Sys.iswindows() && Sys.WORD_SIZE == 32
+    # The Win32 unwinder is 1000x slower than elsewhere (around 1ms/frame),
+    # so we don't want to slow the program down by quite that much
     __init__() = init(1_000_000, 0.01)
 else
-    __init__() = init(1_000_000, 0.001)
+    __init__() = init(10_000_000, 0.001)
 end
 
 """

stdlib/Profile/test/runtests.jl

@@ -84,14 +84,15 @@ end
 
 @testset "setting sample count and delay in init" begin
     n_, delay_ = Profile.init()
-    @test n_ == 1_000_000
-    def_delay = Sys.iswindows() ? 0.01 : 0.001
+    def_n = Sys.iswindows() && Sys.WORD_SIZE == 32 ? 1_000_000 : 10_000_000
+    @test n_ == def_n
+    def_delay = Sys.iswindows() && Sys.WORD_SIZE == 32 ? 0.01 : 0.001
     @test delay_ == def_delay
     Profile.init(n=1_000_001, delay=0.0005)
     n_, delay_ = Profile.init()
     @test n_ == 1_000_001
     @test delay_ == 0.0005
-    Profile.init(n=1_000_000, delay=def_delay)
+    Profile.init(n=def_n, delay=def_delay)
 end
 
 @testset "Line number correction" begin
@@ -131,5 +132,5 @@ let cmd = Base.julia_cmd()
     s = read(p, String)
     close(t)
     @test success(p)
-    @test parse(Int, s) > 1000
+    @test parse(Int, s) > 100
 end