summaryrefslogtreecommitdiff
path: root/runtime
diff options
context:
space:
mode:
Diffstat (limited to 'runtime')
-rw-r--r--runtime/arch/arm/quick_entrypoints_arm.S36
-rw-r--r--runtime/arch/arm64/quick_entrypoints_arm64.S39
-rw-r--r--runtime/arch/x86/quick_entrypoints_x86.S39
-rw-r--r--runtime/arch/x86_64/quick_entrypoints_x86_64.S33
-rw-r--r--runtime/asm_support.h3
-rw-r--r--runtime/debugger.cc17
-rw-r--r--runtime/gc/collector/concurrent_copying.cc2
-rw-r--r--runtime/gc/heap.cc10
-rw-r--r--runtime/generated/asm_support_gen.h10
-rw-r--r--runtime/simulator/Android.mk4
-rw-r--r--runtime/thread-inl.h1
-rw-r--r--runtime/thread.cc12
-rw-r--r--runtime/thread.h38
-rw-r--r--runtime/thread_list.cc36
14 files changed, 237 insertions, 43 deletions
diff --git a/runtime/arch/arm/quick_entrypoints_arm.S b/runtime/arch/arm/quick_entrypoints_arm.S
index c4ec72685f..881bebe002 100644
--- a/runtime/arch/arm/quick_entrypoints_arm.S
+++ b/runtime/arch/arm/quick_entrypoints_arm.S
@@ -191,7 +191,7 @@
.cfi_rel_offset r11, 44
.cfi_rel_offset ip, 48
.cfi_rel_offset lr, 52
- vpush {d0-d15} @ 32 words of float args.
+ vpush {d0-d15} @ 32 words, 2 for each of the 16 saved doubles.
.cfi_adjust_cfa_offset 128
sub sp, #8 @ 2 words of space, alignment padding and Method*
.cfi_adjust_cfa_offset 8
@@ -1030,11 +1030,37 @@ ENTRY art_quick_set64_instance
END art_quick_set64_instance
/*
- * Entry from managed code to resolve a string, this stub will allocate a String and deliver an
- * exception on error. On success the String is returned. R0 holds the string index. The fast
- * path check for hit in strings cache has already been performed.
+ * Entry from managed code to resolve a string, this stub will
+ * check the dex cache for a matching string (the fast path), and if not found,
+ * it will allocate a String and deliver an exception on error.
+ * On success the String is returned. R0 holds the string index.
*/
-ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+
+ENTRY art_quick_resolve_string
+ ldr r1, [sp] @ load referrer
+ ldr r1, [r1, #ART_METHOD_DECLARING_CLASS_OFFSET] @ load declaring class
+ ldr r1, [r1, #DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET] @ load string dex cache
+ ubfx r2, r0, #0, #STRING_DEX_CACHE_HASH_BITS
+ add r1, r1, r2, LSL #STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT
+ ldrd r2, r3, [r1] @ load index into r3 and pointer into r2
+ cmp r0, r3
+ bne .Lart_quick_resolve_string_slow_path
+#ifdef USE_READ_BARRIER
+ ldr r3, [r2, MIRROR_OBJECT_LOCK_WORD_OFFSET]
+ tst r3, #LOCK_WORD_MARK_BIT_MASK_SHIFTED
+ beq .Lart_quick_resolve_string_slow_path
+#endif
+ mov r0, r2
+ bx lr
+
+.Lart_quick_resolve_string_slow_path:
+ SETUP_SAVE_REFS_ONLY_FRAME r2 @ save callee saves in case of GC
+ mov r1, r9 @ pass Thread::Current
+ mov r3, sp
+ bl artResolveStringFromCode @ (uint32_t type_idx, Method* method, Thread*)
+ RESTORE_SAVE_REFS_ONLY_FRAME
+ RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+END art_quick_resolve_string
// Generate the allocation entrypoints for each allocator.
GENERATE_ALLOC_ENTRYPOINTS_FOR_EACH_ALLOCATOR
diff --git a/runtime/arch/arm64/quick_entrypoints_arm64.S b/runtime/arch/arm64/quick_entrypoints_arm64.S
index 4289cabbc6..7b71983724 100644
--- a/runtime/arch/arm64/quick_entrypoints_arm64.S
+++ b/runtime/arch/arm64/quick_entrypoints_arm64.S
@@ -331,6 +331,7 @@
#endif
// Save FP registers.
+ // For better performance, store d0 and d31 separately, so that all STPs are 16-byte aligned.
str d0, [sp, #8]
stp d1, d2, [sp, #16]
stp d3, d4, [sp, #32]
@@ -431,6 +432,7 @@
.macro RESTORE_SAVE_EVERYTHING_FRAME
// Restore FP registers.
+ // For better performance, load d0 and d31 separately, so that all LDPs are 16-byte aligned.
ldr d0, [sp, #8]
ldp d1, d2, [sp, #16]
ldp d3, d4, [sp, #32]
@@ -1784,11 +1786,40 @@ ENTRY art_quick_set64_static
END art_quick_set64_static
/*
- * Entry from managed code to resolve a string, this stub will allocate a String and deliver an
- * exception on error. On success the String is returned. w0 holds the string index. The fast
- * path check for hit in strings cache has already been performed.
+ * Entry from managed code to resolve a string, this stub will
+ * check the dex cache for a matching string (the fast path), and if not found,
+ * it will allocate a String and deliver an exception on error.
+ * On success the String is returned. R0 holds the string index.
*/
-ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+
+ENTRY art_quick_resolve_string
+ ldr x1, [sp] // load referrer
+ ldr w2, [x1, #ART_METHOD_DECLARING_CLASS_OFFSET] // load declaring class
+ ldr x1, [x2, #DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET] // load string dex cache
+ mov x2, #STRING_DEX_CACHE_SIZE_MINUS_ONE
+ and x2, x0, x2
+ lsl x2, x2, #STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT
+ add x1, x2, x1
+ ldr x2, [x1] // load index and pointer into x2
+ lsr x3, x2, #32 // get upper 32 bits
+ ubfx x2, x2, #0, #32 // get lower 32 bits
+ cmp x0, x3
+ bne .Lart_quick_resolve_string_slow_path
+#ifdef USE_READ_BARRIER
+ ldr x3, [x2, #MIRROR_OBJECT_LOCK_WORD_OFFSET]
+ tbz x3, #LOCK_WORD_MARK_BIT_SHIFT, .Lart_quick_resolve_string_slow_path
+#endif
+ mov x0, x2
+ ret
+
+.Lart_quick_resolve_string_slow_path:
+ SETUP_SAVE_REFS_ONLY_FRAME // save callee saves in case of GC
+ mov x1, xSELF // pass Thread::Current
+ bl artResolveStringFromCode // (uint32_t type_idx, Method* method, Thread*)
+ RESTORE_SAVE_REFS_ONLY_FRAME
+ RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+
+END art_quick_resolve_string
// Generate the allocation entrypoints for each allocator.
GENERATE_ALLOC_ENTRYPOINTS_FOR_NON_REGION_TLAB_ALLOCATORS
diff --git a/runtime/arch/x86/quick_entrypoints_x86.S b/runtime/arch/x86/quick_entrypoints_x86.S
index 2e9682e563..d685ace90f 100644
--- a/runtime/arch/x86/quick_entrypoints_x86.S
+++ b/runtime/arch/x86/quick_entrypoints_x86.S
@@ -1108,7 +1108,44 @@ DEFINE_FUNCTION art_quick_alloc_object_region_tlab
ALLOC_OBJECT_TLAB_SLOW_PATH artAllocObjectFromCodeRegionTLAB
END_FUNCTION art_quick_alloc_object_region_tlab
-ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+DEFINE_FUNCTION art_quick_resolve_string
+ SETUP_SAVE_REFS_ONLY_FRAME ebx, ebx
+ movl FRAME_SIZE_SAVE_REFS_ONLY(%esp), %ecx // get referrer
+ movl ART_METHOD_DECLARING_CLASS_OFFSET(%ecx), %ecx // get declaring class
+ movl DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET(%ecx), %ecx // get string dex cache
+ movl LITERAL(STRING_DEX_CACHE_SIZE_MINUS_ONE), %edx
+ andl %eax, %edx
+ shl LITERAL(STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT), %edx
+ addl %ecx, %edx
+ movlps (%edx), %xmm0 // load string idx and pointer to xmm0
+ movd %xmm0, %ecx // extract pointer
+ pshufd LITERAL(0x55), %xmm0, %xmm0 // shuffle index into lowest bits
+ movd %xmm0, %edx // extract index
+ cmp %edx, %eax
+ jne .Lart_quick_resolve_string_slow_path
+#ifdef USE_READ_BARRIER
+ testl LITERAL(LOCK_WORD_MARK_BIT_MASK_SHIFTED), MIRROR_OBJECT_LOCK_WORD_OFFSET(%ecx)
+ jz .Lart_quick_resolve_string_slow_path
+#endif
+ movl %ecx, %eax
+ RESTORE_SAVE_REFS_ONLY_FRAME
+ ret
+
+.Lart_quick_resolve_string_slow_path:
+ // Outgoing argument set up
+ subl LITERAL(8), %esp // push padding
+ CFI_ADJUST_CFA_OFFSET(8)
+ pushl %fs:THREAD_SELF_OFFSET // pass Thread::Current()
+ CFI_ADJUST_CFA_OFFSET(4)
+ PUSH eax // pass arg1
+ call SYMBOL(artResolveStringFromCode)
+ addl LITERAL(16), %esp // pop arguments
+ CFI_ADJUST_CFA_OFFSET(-16)
+ RESTORE_SAVE_REFS_ONLY_FRAME
+ RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+
+END_FUNCTION art_quick_resolve_string
+
ONE_ARG_DOWNCALL art_quick_initialize_static_storage, artInitializeStaticStorageFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
ONE_ARG_DOWNCALL art_quick_initialize_type, artInitializeTypeFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
ONE_ARG_DOWNCALL art_quick_initialize_type_and_verify_access, artInitializeTypeAndVerifyAccessFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
diff --git a/runtime/arch/x86_64/quick_entrypoints_x86_64.S b/runtime/arch/x86_64/quick_entrypoints_x86_64.S
index 32768b0263..3266d86595 100644
--- a/runtime/arch/x86_64/quick_entrypoints_x86_64.S
+++ b/runtime/arch/x86_64/quick_entrypoints_x86_64.S
@@ -1330,7 +1330,38 @@ DEFINE_FUNCTION art_quick_alloc_object_initialized_region_tlab
ALLOC_OBJECT_TLAB_SLOW_PATH artAllocObjectFromCodeInitializedRegionTLAB
END_FUNCTION art_quick_alloc_object_initialized_region_tlab
-ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+DEFINE_FUNCTION art_quick_resolve_string
+ movq 8(%rsp), %rcx // get referrer
+ movl ART_METHOD_DECLARING_CLASS_OFFSET(%rcx), %ecx // get declaring class
+ movq DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET(%ecx), %rcx // get string dex cache
+ movq LITERAL(STRING_DEX_CACHE_SIZE_MINUS_ONE), %rdx
+ andq %rdi, %rdx
+ shlq LITERAL(STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT), %rdx
+ addq %rcx, %rdx
+ movq %rax, %rcx
+ movq (%rdx), %rdx
+ movq %rdx, %rax
+ movl %eax, %eax
+ shrq LITERAL(32), %rdx
+ cmp %rdx, %rdi
+ jne .Lart_quick_resolve_string_slow_path
+#ifdef USE_READ_BARRIER
+ testl LITERAL(LOCK_WORD_MARK_BIT_MASK_SHIFTED), MIRROR_OBJECT_LOCK_WORD_OFFSET(%rax)
+ jz .Lart_quick_resolve_string_slow_path
+#endif
+ ret
+
+.Lart_quick_resolve_string_slow_path:
+ SETUP_SAVE_REFS_ONLY_FRAME
+ movq %rcx, %rax
+ // Outgoing argument set up
+ movq %gs:THREAD_SELF_OFFSET, %rsi // pass Thread::Current()
+
+ call artResolveStringFromCode // artResolveStringFromCode(arg0, referrer, Thread*)
+ RESTORE_SAVE_REFS_ONLY_FRAME // restore frame up to return address
+ RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+END_FUNCTION art_quick_resolve_string
+
ONE_ARG_DOWNCALL art_quick_initialize_static_storage, artInitializeStaticStorageFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
ONE_ARG_DOWNCALL art_quick_initialize_type, artInitializeTypeFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
ONE_ARG_DOWNCALL art_quick_initialize_type_and_verify_access, artInitializeTypeAndVerifyAccessFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
diff --git a/runtime/asm_support.h b/runtime/asm_support.h
index 848f8e532c..102b993686 100644
--- a/runtime/asm_support.h
+++ b/runtime/asm_support.h
@@ -19,12 +19,15 @@
#if defined(__cplusplus)
#include "art_method.h"
+#include "base/bit_utils.h"
#include "gc/allocator/rosalloc.h"
#include "gc/heap.h"
#include "jit/jit.h"
#include "lock_word.h"
#include "mirror/class.h"
+#include "mirror/dex_cache.h"
#include "mirror/string.h"
+#include "utils/dex_cache_arrays_layout.h"
#include "runtime.h"
#include "thread.h"
#endif
diff --git a/runtime/debugger.cc b/runtime/debugger.cc
index 2a5198bf01..89bebb4ccb 100644
--- a/runtime/debugger.cc
+++ b/runtime/debugger.cc
@@ -4059,7 +4059,7 @@ void Dbg::ExecuteMethodWithoutPendingException(ScopedObjectAccess& soa, DebugInv
// Prepare JDWP ids for the reply.
JDWP::JdwpTag result_tag = BasicTagFromDescriptor(m->GetShorty());
const bool is_object_result = (result_tag == JDWP::JT_OBJECT);
- StackHandleScope<2> hs(soa.Self());
+ StackHandleScope<3> hs(soa.Self());
Handle<mirror::Object> object_result = hs.NewHandle(is_object_result ? result.GetL() : nullptr);
Handle<mirror::Throwable> exception = hs.NewHandle(soa.Self()->GetException());
soa.Self()->ClearException();
@@ -4098,10 +4098,17 @@ void Dbg::ExecuteMethodWithoutPendingException(ScopedObjectAccess& soa, DebugInv
// unless we threw, in which case we return null.
DCHECK_EQ(JDWP::JT_VOID, result_tag);
if (exceptionObjectId == 0) {
- // TODO we could keep the receiver ObjectId in the DebugInvokeReq to avoid looking into the
- // object registry.
- result_value = GetObjectRegistry()->Add(pReq->receiver.Read());
- result_tag = TagFromObject(soa, pReq->receiver.Read());
+ if (m->GetDeclaringClass()->IsStringClass()) {
+ // For string constructors, the new string is remapped to the receiver (stored in ref).
+ Handle<mirror::Object> decoded_ref = hs.NewHandle(soa.Self()->DecodeJObject(ref.get()));
+ result_value = gRegistry->Add(decoded_ref);
+ result_tag = TagFromObject(soa, decoded_ref.Get());
+ } else {
+ // TODO we could keep the receiver ObjectId in the DebugInvokeReq to avoid looking into the
+ // object registry.
+ result_value = GetObjectRegistry()->Add(pReq->receiver.Read());
+ result_tag = TagFromObject(soa, pReq->receiver.Read());
+ }
} else {
result_value = 0;
result_tag = JDWP::JT_OBJECT;
diff --git a/runtime/gc/collector/concurrent_copying.cc b/runtime/gc/collector/concurrent_copying.cc
index 7afe6f9ab4..42816a04f1 100644
--- a/runtime/gc/collector/concurrent_copying.cc
+++ b/runtime/gc/collector/concurrent_copying.cc
@@ -435,10 +435,8 @@ void ConcurrentCopying::FlipThreadRoots() {
gc_barrier_->Init(self, 0);
ThreadFlipVisitor thread_flip_visitor(this, heap_->use_tlab_);
FlipCallback flip_callback(this);
- heap_->ThreadFlipBegin(self); // Sync with JNI critical calls.
size_t barrier_count = Runtime::Current()->FlipThreadRoots(
&thread_flip_visitor, &flip_callback, this);
- heap_->ThreadFlipEnd(self);
{
ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
gc_barrier_->Increment(self, barrier_count);
diff --git a/runtime/gc/heap.cc b/runtime/gc/heap.cc
index 39f26e7fe2..638c1d841a 100644
--- a/runtime/gc/heap.cc
+++ b/runtime/gc/heap.cc
@@ -878,9 +878,13 @@ void Heap::IncrementDisableThreadFlip(Thread* self) {
MutexLock mu(self, *thread_flip_lock_);
bool has_waited = false;
uint64_t wait_start = NanoTime();
- while (thread_flip_running_) {
- has_waited = true;
- thread_flip_cond_->Wait(self);
+ if (thread_flip_running_) {
+ TimingLogger::ScopedTiming split("IncrementDisableThreadFlip",
+ GetCurrentGcIteration()->GetTimings());
+ while (thread_flip_running_) {
+ has_waited = true;
+ thread_flip_cond_->Wait(self);
+ }
}
++disable_thread_flip_count_;
if (has_waited) {
diff --git a/runtime/generated/asm_support_gen.h b/runtime/generated/asm_support_gen.h
index 716c23d1b0..40b71c40a4 100644
--- a/runtime/generated/asm_support_gen.h
+++ b/runtime/generated/asm_support_gen.h
@@ -70,6 +70,16 @@ DEFINE_CHECK_EQ(static_cast<int32_t>(ART_METHOD_JNI_OFFSET_64), (static_cast<int
DEFINE_CHECK_EQ(static_cast<int32_t>(ART_METHOD_QUICK_CODE_OFFSET_32), (static_cast<int32_t>(art::ArtMethod:: EntryPointFromQuickCompiledCodeOffset(art::PointerSize::k32).Int32Value())))
#define ART_METHOD_QUICK_CODE_OFFSET_64 48
DEFINE_CHECK_EQ(static_cast<int32_t>(ART_METHOD_QUICK_CODE_OFFSET_64), (static_cast<int32_t>(art::ArtMethod:: EntryPointFromQuickCompiledCodeOffset(art::PointerSize::k64).Int32Value())))
+#define ART_METHOD_DECLARING_CLASS_OFFSET 0
+DEFINE_CHECK_EQ(static_cast<int32_t>(ART_METHOD_DECLARING_CLASS_OFFSET), (static_cast<int32_t>(art::ArtMethod:: DeclaringClassOffset().Int32Value())))
+#define DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET 40
+DEFINE_CHECK_EQ(static_cast<int32_t>(DECLARING_CLASS_DEX_CACHE_STRINGS_OFFSET), (static_cast<int32_t>(art::mirror::Class:: DexCacheStringsOffset().Int32Value())))
+#define STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT 3
+DEFINE_CHECK_EQ(static_cast<int32_t>(STRING_DEX_CACHE_ELEMENT_SIZE_SHIFT), (static_cast<int32_t>(art::WhichPowerOf2(sizeof(art::mirror::StringDexCachePair)))))
+#define STRING_DEX_CACHE_SIZE_MINUS_ONE 1023
+DEFINE_CHECK_EQ(static_cast<int32_t>(STRING_DEX_CACHE_SIZE_MINUS_ONE), (static_cast<int32_t>(art::mirror::DexCache::kDexCacheStringCacheSize - 1)))
+#define STRING_DEX_CACHE_HASH_BITS 10
+DEFINE_CHECK_EQ(static_cast<int32_t>(STRING_DEX_CACHE_HASH_BITS), (static_cast<int32_t>(art::LeastSignificantBit(art::mirror::DexCache::kDexCacheStringCacheSize))))
#define MIN_LARGE_OBJECT_THRESHOLD 0x3000
DEFINE_CHECK_EQ(static_cast<size_t>(MIN_LARGE_OBJECT_THRESHOLD), (static_cast<size_t>(art::gc::Heap::kMinLargeObjectThreshold)))
#define LOCK_WORD_STATE_SHIFT 30
diff --git a/runtime/simulator/Android.mk b/runtime/simulator/Android.mk
index a34a84100a..e39af2dcd9 100644
--- a/runtime/simulator/Android.mk
+++ b/runtime/simulator/Android.mk
@@ -22,6 +22,9 @@ LIBART_SIMULATOR_SRC_FILES := \
code_simulator.cc \
code_simulator_arm64.cc
+LIBART_SIMULATOR_CFLAGS := \
+ -DVIXL_INCLUDE_SIMULATOR_AARCH64
+
# $(1): target or host
# $(2): ndebug or debug
define build-libart-simulator
@@ -54,6 +57,7 @@ define build-libart-simulator
LOCAL_MODULE_CLASS := SHARED_LIBRARIES
LOCAL_SRC_FILES := $$(LIBART_SIMULATOR_SRC_FILES)
+ LOCAL_CFLAGS := $$(LIBART_SIMULATOR_CFLAGS)
ifeq ($$(art_target_or_host),target)
$(call set-target-local-clang-vars)
diff --git a/runtime/thread-inl.h b/runtime/thread-inl.h
index 3aa1fc256d..216d8a7194 100644
--- a/runtime/thread-inl.h
+++ b/runtime/thread-inl.h
@@ -224,6 +224,7 @@ inline ThreadState Thread::TransitionFromSuspendedToRunnable() {
thread_to_pass = this;
}
MutexLock mu(thread_to_pass, *Locks::thread_suspend_count_lock_);
+ ScopedTransitioningToRunnable scoped_transitioning_to_runnable(this);
old_state_and_flags.as_int = tls32_.state_and_flags.as_int;
DCHECK_EQ(old_state_and_flags.as_struct.state, old_state);
while ((old_state_and_flags.as_struct.flags & kSuspendRequest) != 0) {
diff --git a/runtime/thread.cc b/runtime/thread.cc
index b35a614e99..79b9f02991 100644
--- a/runtime/thread.cc
+++ b/runtime/thread.cc
@@ -1217,10 +1217,8 @@ void Thread::FullSuspendCheck() {
ScopedTrace trace(__FUNCTION__);
VLOG(threads) << this << " self-suspending";
// Make thread appear suspended to other threads, release mutator_lock_.
- tls32_.suspended_at_suspend_check = true;
// Transition to suspended and back to runnable, re-acquire share on mutator_lock_.
ScopedThreadSuspension(this, kSuspended);
- tls32_.suspended_at_suspend_check = false;
VLOG(threads) << this << " self-reviving";
}
@@ -1433,6 +1431,12 @@ struct StackDumpVisitor : public StackVisitor {
if (o == nullptr) {
os << "an unknown object";
} else {
+ if (kUseReadBarrier && Thread::Current()->GetIsGcMarking()) {
+ // We may call Thread::Dump() in the middle of the CC thread flip and this thread's stack
+ // may have not been flipped yet and "o" may be a from-space (stale) ref, in which case the
+ // IdentityHashCode call below will crash. So explicitly mark/forward it here.
+ o = ReadBarrier::Mark(o);
+ }
if ((o->GetLockWord(false).GetState() == LockWord::kThinLocked) &&
Locks::mutator_lock_->IsExclusiveHeld(Thread::Current())) {
// Getting the identity hashcode here would result in lock inflation and suspension of the
@@ -1635,7 +1639,7 @@ Thread::Thread(bool daemon) : tls32_(daemon), wait_monitor_(nullptr), interrupte
}
tlsPtr_.flip_function = nullptr;
tlsPtr_.thread_local_mark_stack = nullptr;
- tls32_.suspended_at_suspend_check = false;
+ tls32_.is_transitioning_to_runnable = false;
}
bool Thread::IsStillStarting() const {
@@ -1773,7 +1777,7 @@ Thread::~Thread() {
CHECK(tlsPtr_.checkpoint_function == nullptr);
CHECK_EQ(checkpoint_overflow_.size(), 0u);
CHECK(tlsPtr_.flip_function == nullptr);
- CHECK_EQ(tls32_.suspended_at_suspend_check, false);
+ CHECK_EQ(tls32_.is_transitioning_to_runnable, false);
// Make sure we processed all deoptimization requests.
CHECK(tlsPtr_.deoptimization_context_stack == nullptr) << "Missed deoptimization";
diff --git a/runtime/thread.h b/runtime/thread.h
index 840b7817f8..1c2d4ab533 100644
--- a/runtime/thread.h
+++ b/runtime/thread.h
@@ -1085,8 +1085,12 @@ class Thread {
return tlsPtr_.nested_signal_state;
}
- bool IsSuspendedAtSuspendCheck() const {
- return tls32_.suspended_at_suspend_check;
+ bool IsTransitioningToRunnable() const {
+ return tls32_.is_transitioning_to_runnable;
+ }
+
+ void SetIsTransitioningToRunnable(bool value) {
+ tls32_.is_transitioning_to_runnable = value;
}
void PushVerifier(verifier::MethodVerifier* verifier);
@@ -1264,7 +1268,7 @@ class Thread {
suspend_count(0), debug_suspend_count(0), thin_lock_thread_id(0), tid(0),
daemon(is_daemon), throwing_OutOfMemoryError(false), no_thread_suspension(0),
thread_exit_check_count(0), handling_signal_(false),
- suspended_at_suspend_check(false), ready_for_debug_invoke(false),
+ is_transitioning_to_runnable(false), ready_for_debug_invoke(false),
debug_method_entry_(false), is_gc_marking(false), weak_ref_access_enabled(true),
disable_thread_flip_count(0) {
}
@@ -1306,10 +1310,10 @@ class Thread {
// True if signal is being handled by this thread.
bool32_t handling_signal_;
- // True if the thread is suspended in FullSuspendCheck(). This is
- // used to distinguish runnable threads that are suspended due to
- // a normal suspend check from other threads.
- bool32_t suspended_at_suspend_check;
+ // True if the thread is in TransitionFromSuspendedToRunnable(). This is used to distinguish the
+ // non-runnable threads (eg. kNative, kWaiting) that are about to transition to runnable from
+ // the rest of them.
+ bool32_t is_transitioning_to_runnable;
// True if the thread has been suspended by a debugger event. This is
// used to invoke method from the debugger which is only allowed when
@@ -1588,6 +1592,26 @@ class ScopedDebugDisallowReadBarriers {
Thread* const self_;
};
+class ScopedTransitioningToRunnable : public ValueObject {
+ public:
+ explicit ScopedTransitioningToRunnable(Thread* self)
+ : self_(self) {
+ DCHECK_EQ(self, Thread::Current());
+ if (kUseReadBarrier) {
+ self_->SetIsTransitioningToRunnable(true);
+ }
+ }
+
+ ~ScopedTransitioningToRunnable() {
+ if (kUseReadBarrier) {
+ self_->SetIsTransitioningToRunnable(false);
+ }
+ }
+
+ private:
+ Thread* const self_;
+};
+
std::ostream& operator<<(std::ostream& os, const Thread& thread);
std::ostream& operator<<(std::ostream& os, const StackedShadowFrameType& thread);
diff --git a/runtime/thread_list.cc b/runtime/thread_list.cc
index 419ecec696..688514cd76 100644
--- a/runtime/thread_list.cc
+++ b/runtime/thread_list.cc
@@ -405,6 +405,8 @@ size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
Locks::thread_suspend_count_lock_->AssertNotHeld(self);
CHECK_NE(self->GetState(), kRunnable);
+ collector->GetHeap()->ThreadFlipBegin(self); // Sync with JNI critical calls.
+
SuspendAllInternal(self, self, nullptr);
// Run the flip callback for the collector.
@@ -414,26 +416,31 @@ size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
collector->RegisterPause(NanoTime() - start_time);
// Resume runnable threads.
- std::vector<Thread*> runnable_threads;
+ size_t runnable_thread_count = 0;
std::vector<Thread*> other_threads;
{
+ TimingLogger::ScopedTiming split2("ResumeRunnableThreads", collector->GetTimings());
MutexLock mu(self, *Locks::thread_list_lock_);
MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
--suspend_all_count_;
for (const auto& thread : list_) {
+ // Set the flip function for all threads because Thread::DumpState/DumpJavaStack() (invoked by
+ // a checkpoint) may cause the flip function to be run for a runnable/suspended thread before
+ // a runnable thread runs it for itself or we run it for a suspended thread below.
+ thread->SetFlipFunction(thread_flip_visitor);
if (thread == self) {
continue;
}
- // Set the flip function for both runnable and suspended threads
- // because Thread::DumpState/DumpJavaStack() (invoked by a
- // checkpoint) may cause the flip function to be run for a
- // runnable/suspended thread before a runnable threads runs it
- // for itself or we run it for a suspended thread below.
- thread->SetFlipFunction(thread_flip_visitor);
- if (thread->IsSuspendedAtSuspendCheck()) {
+ // Resume early the threads that were runnable but are suspended just for this thread flip or
+ // about to transition from non-runnable (eg. kNative at the SOA entry in a JNI function) to
+ // runnable (both cases waiting inside Thread::TransitionFromSuspendedToRunnable), or waiting
+ // for the thread flip to end at the JNI critical section entry (kWaitingForGcThreadFlip),
+ ThreadState state = thread->GetState();
+ if (state == kWaitingForGcThreadFlip ||
+ thread->IsTransitioningToRunnable()) {
// The thread will resume right after the broadcast.
thread->ModifySuspendCount(self, -1, nullptr, false);
- runnable_threads.push_back(thread);
+ ++runnable_thread_count;
} else {
other_threads.push_back(thread);
}
@@ -441,8 +448,11 @@ size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
Thread::resume_cond_->Broadcast(self);
}
+ collector->GetHeap()->ThreadFlipEnd(self);
+
// Run the closure on the other threads and let them resume.
{
+ TimingLogger::ScopedTiming split3("FlipOtherThreads", collector->GetTimings());
ReaderMutexLock mu(self, *Locks::mutator_lock_);
for (const auto& thread : other_threads) {
Closure* flip_func = thread->GetFlipFunction();
@@ -451,11 +461,15 @@ size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
}
}
// Run it for self.
- thread_flip_visitor->Run(self);
+ Closure* flip_func = self->GetFlipFunction();
+ if (flip_func != nullptr) {
+ flip_func->Run(self);
+ }
}
// Resume other threads.
{
+ TimingLogger::ScopedTiming split4("ResumeOtherThreads", collector->GetTimings());
MutexLock mu2(self, *Locks::thread_suspend_count_lock_);
for (const auto& thread : other_threads) {
thread->ModifySuspendCount(self, -1, nullptr, false);
@@ -463,7 +477,7 @@ size_t ThreadList::FlipThreadRoots(Closure* thread_flip_visitor,
Thread::resume_cond_->Broadcast(self);
}
- return runnable_threads.size() + other_threads.size() + 1; // +1 for self.
+ return runnable_thread_count + other_threads.size() + 1; // +1 for self.
}
void ThreadList::SuspendAll(const char* cause, bool long_suspend) {
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2025-12-10 21:21:02 +0000