[pull] master from ruby:master

ext/-test-/thread/instrumentation/extconf.rb

@@ -1,2 +1,3 @@
 # frozen_string_literal: false
+$INCFLAGS << " -I$(topdir) -I$(top_srcdir)"
 create_makefile("-test-/thread/instrumentation")

ext/-test-/thread/instrumentation/instrumentation.c

@@ -1,6 +1,7 @@
 #include "ruby/ruby.h"
 #include "ruby/atomic.h"
 #include "ruby/thread.h"
+#include "internal/gc.h"
 
 #ifndef RB_THREAD_LOCAL_SPECIFIER
 #  define RB_THREAD_LOCAL_SPECIFIER
@@ -19,9 +20,11 @@ static rb_atomic_t timeline_cursor;
 
 static void
 event_timeline_gc_mark(void *ptr) {
+    /* Hook sees every Ractor's threads; validate membership before marking. */
+    rb_atomic_t n = RUBY_ATOMIC_LOAD(timeline_cursor);
     rb_atomic_t cursor;
-    for (cursor = 0; cursor < timeline_cursor; cursor++) {
-        rb_gc_mark(event_timeline[cursor].thread);
+    for (cursor = 0; cursor < n; cursor++) {
+        rb_gc_mark_maybe(event_timeline[cursor].thread);
     }
 }
 
@@ -42,14 +45,12 @@ reset_timeline(void)
 static rb_event_flag_t
 find_last_event(VALUE thread)
 {
-    rb_atomic_t cursor = timeline_cursor;
-    if (cursor) {
-        do {
-            if (event_timeline[cursor].thread == thread){
-                return event_timeline[cursor].event;
-            }
-            cursor--;
-        } while (cursor > 0);
+    rb_atomic_t cursor = RUBY_ATOMIC_LOAD(timeline_cursor);
+    while (cursor > 0) {
+        cursor--;
+        if (event_timeline[cursor].thread == thread) {
+            return event_timeline[cursor].event;
+        }
     }
     return 0;
 }
@@ -166,6 +167,7 @@ event_symbol(rb_event_flag_t event)
     }
 }
 
+// NOTE: only safe when there's a single active Ractor
 static VALUE
 thread_unregister_callback(VALUE thread)
 {
@@ -174,11 +176,15 @@ thread_unregister_callback(VALUE thread)
         single_hook = NULL;
     }
 
-    VALUE events = rb_ary_new_capa(timeline_cursor);
+    rb_atomic_t n = RUBY_ATOMIC_LOAD(timeline_cursor);
+    VALUE events = rb_ary_new_capa(n);
     rb_atomic_t cursor;
-    for (cursor = 0; cursor < timeline_cursor; cursor++) {
+    for (cursor = 0; cursor < n; cursor++) {
+        VALUE th = event_timeline[cursor].thread;
+        /* Skip foreign-objspace threads: pushing them into this Array would violate containment. */
+        if (rb_objspace_foreign_object_p(th)) continue;
         VALUE pair = rb_ary_new_capa(2);
-        rb_ary_push(pair, event_timeline[cursor].thread);
+        rb_ary_push(pair, th);
         rb_ary_push(pair, event_symbol(event_timeline[cursor].event));
         rb_ary_push(events, pair);
     }

ext/digest/blake3/extconf.rb

@@ -29,7 +29,9 @@ def blake3_disable(macro)
 end
 
 # Probe used to confirm the compiler both accepts +flag+ and can compile the
-# intrinsics the backend relies on.
+# intrinsics the backend relies on.  It compiles the backend source itself:
+# a small snippet misses assemblers that reject what the compiler emits (e.g.
+# binutils 2.30 on RHEL 8 rejects AVX-512 code from gcc 8.5).
 def blake3_have_isa?(name, flag, snippet)
   checking_for("#{name} intrinsics" + (flag ? " (#{flag})" : "")) do
     try_compile(snippet, flag)
@@ -40,23 +42,15 @@ def blake3_have_isa?(name, flag, snippet)
 when /\A(x86_64|amd64|x64)\z/i
   # Try to detect which SIMD features this x86 machine and compiler has
   x86_backends = [
-    ["blake3_sse2",   "SSE2",    ["-msse2", "-arch:SSE2"],
-     "#include <immintrin.h>\nint main(void){ volatile __m128i x = _mm_setzero_si128(); (void)x; return 0; }\n",
-     "BLAKE3_NO_SSE2"],
-    ["blake3_sse41",  "SSE4.1",  ["-msse4.1", "-arch:AVX"],
-     "#include <immintrin.h>\nint main(void){ volatile __m128i x = _mm_setzero_si128(); return _mm_testz_si128(x, x); }\n",
-     "BLAKE3_NO_SSE41"],
-    ["blake3_avx2",   "AVX2",    ["-mavx2", "-arch:AVX2"],
-     "#include <immintrin.h>\nint main(void){ volatile __m256i x = _mm256_setzero_si256(); (void)x; return 0; }\n",
-     "BLAKE3_NO_AVX2"],
-    ["blake3_avx512", "AVX-512", ["-mavx512f -mavx512vl", "-arch:AVX512"],
-     "#include <immintrin.h>\nint main(void){ volatile __m512i x = _mm512_setzero_si512(); (void)x; return 0; }\n",
-     "BLAKE3_NO_AVX512"],
+    ["blake3_sse2",   "SSE2",    ["-msse2", "-arch:SSE2"],                 "BLAKE3_NO_SSE2"],
+    ["blake3_sse41",  "SSE4.1",  ["-msse4.1", "-arch:AVX"],                "BLAKE3_NO_SSE41"],
+    ["blake3_avx2",   "AVX2",    ["-mavx2", "-arch:AVX2"],                 "BLAKE3_NO_AVX2"],
+    ["blake3_avx512", "AVX-512", ["-mavx512f -mavx512vl", "-arch:AVX512"], "BLAKE3_NO_AVX512"],
   ]
 
-  x86_backends.each do |obj, name, flags, snippet, no_macro|
+  x86_backends.each do |obj, name, flags, no_macro|
     [nil, *flags].any? do |flag|
-      if blake3_have_isa?(name, flag, snippet)
+      if blake3_have_isa?(name, flag, %{#include "#{$srcdir}/#{obj}.c"\n})
         objs << obj
         simd_cflags[obj] = flag
         true

test/-ext-/thread/test_instrumentation_api.rb

@@ -134,8 +134,6 @@ def test_queue_releases_gvl
   end
 
   def test_blocking_on_ractor
-    # TODO: turn back on and fix when ractor-local GC (rlgc) lands
-    omit "this is failing CI and rlgc will change how this works so let's wait until it lands to fix it" if ENV["GITHUB_WORKFLOW"]
     assert_ractor(<<-"RUBY", require_relative: "helper", require: "-test-/thread/instrumentation")
       include ThreadInstrumentation::TestHelper
 
@@ -163,9 +161,6 @@ def test_blocking_on_ractor
   end
 
   def test_sleeping_inside_ractor
-    # TODO: turn back and fix when ractor-local GC (rlgc) lands
-    omit "This test is flaky and intermittently failing" if ENV['GITHUB_WORKFLOW']
-
     assert_ractor(<<-"RUBY", require_relative: "helper", require: "-test-/thread/instrumentation")
       include ThreadInstrumentation::TestHelper