static
void addProfileEntry(ProfileStack *aStack, ThreadProfile &aProfile, int i)
{
// First entry has tagName 's' (start)
// Check for magic pointer bit 1 to indicate copy
const char* sampleLabel = aStack->mStack[i].mLabel;
if (aStack->mStack[i].isCopyLabel()) {
// Store the string using 1 or more 'd' (dynamic) tags
// that will happen to the preceding tag
aProfile.addTag(ProfileEntry('c', ""));
// Add one to store the null termination
size_t strLen = strlen(sampleLabel) + 1;
for (size_t j = 0; j < strLen;) {
// Store as many characters in the void* as the platform allows
char text[sizeof(void*)];
for (size_t pos = 0; pos < sizeof(void*) && j+pos < strLen; pos++) {
text[pos] = sampleLabel[j+pos];
}
j += sizeof(void*)/sizeof(char);
// Cast to *((void**) to pass the text data to a void*
aProfile.addTag(ProfileEntry('d', *((void**)(&text[0]))));
}
} else {
aProfile.addTag(ProfileEntry('c', sampleLabel));
}
}
void TableTicker::doBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
#ifndef XP_MACOSX
uintptr_t thread = GetThreadHandle(platform_data());
MOZ_ASSERT(thread);
#endif
void* pc_array[1000];
PCArray array = {
pc_array,
mozilla::ArrayLength(pc_array),
0
};
// Start with the current function.
StackWalkCallback(aSample->pc, &array);
#ifdef XP_MACOSX
pthread_t pt = GetProfiledThread(platform_data());
void *stackEnd = reinterpret_cast<void*>(-1);
if (pt)
stackEnd = static_cast<char*>(pthread_get_stackaddr_np(pt));
nsresult rv = FramePointerStackWalk(StackWalkCallback, 0, &array, reinterpret_cast<void**>(aSample->fp), stackEnd);
#else
nsresult rv = NS_StackWalk(StackWalkCallback, 0, &array, thread);
#endif
if (NS_SUCCEEDED(rv)) {
aProfile.addTag(ProfileEntry('s', "(root)"));
for (size_t i = array.count; i > 0; --i) {
aProfile.addTag(ProfileEntry('l', (void*)array.array[i - 1]));
}
}
}
void TableTicker::Tick(TickSample* sample)
{
// Marker(s) come before the sample
ProfileStack* stack = mPrimaryThreadProfile.GetStack();
for (int i = 0; stack->getMarker(i) != NULL; i++) {
mPrimaryThreadProfile.addTag(ProfileEntry('m', stack->getMarker(i)));
}
stack->mQueueClearMarker = true;
bool recordSample = true;
if (mJankOnly) {
// if we are on a different event we can discard any temporary samples
// we've kept around
if (sLastSampledEventGeneration != sCurrentEventGeneration) {
// XXX: we also probably want to add an entry to the profile to help
// distinguish which samples are part of the same event. That, or record
// the event generation in each sample
mPrimaryThreadProfile.erase();
}
sLastSampledEventGeneration = sCurrentEventGeneration;
recordSample = false;
// only record the events when we have a we haven't seen a tracer event for 100ms
if (!sLastTracerEvent.IsNull()) {
TimeDuration delta = sample->timestamp - sLastTracerEvent;
if (delta.ToMilliseconds() > 100.0) {
recordSample = true;
}
}
}
#if defined(USE_BACKTRACE) || defined(USE_NS_STACKWALK) || defined(USE_LIBUNWIND)
if (mUseStackWalk) {
doBacktrace(mPrimaryThreadProfile, sample);
} else {
doSampleStackTrace(stack, mPrimaryThreadProfile, sample);
}
#else
doSampleStackTrace(stack, mPrimaryThreadProfile, sample);
#endif
if (recordSample)
mPrimaryThreadProfile.flush();
if (!sLastTracerEvent.IsNull() && sample) {
TimeDuration delta = sample->timestamp - sLastTracerEvent;
mPrimaryThreadProfile.addTag(ProfileEntry('r', delta.ToMilliseconds()));
}
if (sample) {
TimeDuration delta = sample->timestamp - mStartTime;
mPrimaryThreadProfile.addTag(ProfileEntry('t', delta.ToMilliseconds()));
}
if (sLastFrameNumber != sFrameNumber) {
mPrimaryThreadProfile.addTag(ProfileEntry('f', sFrameNumber));
sLastFrameNumber = sFrameNumber;
}
}
static
void addPseudoEntry(volatile StackEntry &entry, ThreadProfile &aProfile,
PseudoStack *stack, void *lastpc)
{
// Pseudo-frames with the ASMJS flag are just annotations and should not be
// recorded in the profile.
if (entry.hasFlag(StackEntry::ASMJS))
return;
int lineno = -1;
// First entry has tagName 's' (start)
// Check for magic pointer bit 1 to indicate copy
const char* sampleLabel = entry.label();
if (entry.isCopyLabel()) {
// Store the string using 1 or more 'd' (dynamic) tags
// that will happen to the preceding tag
addDynamicTag(aProfile, 'c', sampleLabel);
if (entry.isJs()) {
if (!entry.pc()) {
// The JIT only allows the top-most entry to have a nullptr pc
MOZ_ASSERT(&entry == &stack->mStack[stack->stackSize() - 1]);
// If stack-walking was disabled, then that's just unfortunate
if (lastpc) {
jsbytecode *jspc = js::ProfilingGetPC(stack->mRuntime, entry.script(),
lastpc);
if (jspc) {
lineno = JS_PCToLineNumber(entry.script(), jspc);
}
}
} else {
lineno = JS_PCToLineNumber(entry.script(), entry.pc());
}
} else {
lineno = entry.line();
}
} else {
aProfile.addTag(ProfileEntry('c', sampleLabel));
// XXX: Bug 1010578. Don't assume a CPP entry and try to get the
// line for js entries as well.
if (entry.isCpp()) {
lineno = entry.line();
}
}
if (lineno != -1) {
aProfile.addTag(ProfileEntry('n', lineno));
}
uint32_t category = entry.category();
MOZ_ASSERT(!(category & StackEntry::IS_CPP_ENTRY));
MOZ_ASSERT(!(category & StackEntry::FRAME_LABEL_COPY));
if (category) {
aProfile.addTag(ProfileEntry('y', (int)category));
}
}
void TableTicker::doBacktrace(ThreadProfile &aProfile, Address pc)
{
void *array[100];
int count = backtrace (array, 100);
aProfile.addTag(ProfileEntry('s', "(root)", 0));
for (int i = 0; i < count; i++) {
if( (intptr_t)array[i] == -1 ) break;
aProfile.addTag(ProfileEntry('l', (const char*)array[i]));
}
}
void TableTicker::doBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
void *array[100];
int count = backtrace (array, 100);
aProfile.addTag(ProfileEntry('s', "(root)"));
for (int i = 0; i < count; i++) {
if( (intptr_t)array[i] == -1 ) break;
aProfile.addTag(ProfileEntry('l', (void*)array[i]));
}
}
static
void addProfileEntry(volatile StackEntry &entry, ThreadProfile &aProfile,
ProfileStack *stack, void *lastpc)
{
int lineno = -1;
// First entry has tagName 's' (start)
// Check for magic pointer bit 1 to indicate copy
const char* sampleLabel = entry.label();
if (entry.isCopyLabel()) {
// Store the string using 1 or more 'd' (dynamic) tags
// that will happen to the preceding tag
aProfile.addTag(ProfileEntry('c', ""));
// Add one to store the null termination
size_t strLen = strlen(sampleLabel) + 1;
for (size_t j = 0; j < strLen;) {
// Store as many characters in the void* as the platform allows
char text[sizeof(void*)];
for (size_t pos = 0; pos < sizeof(void*) && j+pos < strLen; pos++) {
text[pos] = sampleLabel[j+pos];
}
j += sizeof(void*)/sizeof(char);
// Cast to *((void**) to pass the text data to a void*
aProfile.addTag(ProfileEntry('d', *((void**)(&text[0]))));
}
if (entry.js()) {
if (!entry.pc()) {
// The JIT only allows the top-most entry to have a NULL pc
MOZ_ASSERT(&entry == &stack->mStack[stack->stackSize() - 1]);
// If stack-walking was disabled, then that's just unfortunate
if (lastpc) {
jsbytecode *jspc = js::ProfilingGetPC(stack->mRuntime, entry.script(),
lastpc);
if (jspc) {
lineno = JS_PCToLineNumber(NULL, entry.script(), jspc);
}
}
} else {
lineno = JS_PCToLineNumber(NULL, entry.script(), entry.pc());
}
} else {
lineno = entry.line();
}
} else {
aProfile.addTag(ProfileEntry('c', sampleLabel));
lineno = entry.line();
}
if (lineno != -1) {
aProfile.addTag(ProfileEntry('n', lineno));
}
}
static
void doSampleStackTrace(ThreadProfile &aProfile, TickSample *aSample, bool aAddLeafAddresses)
{
NativeStack nativeStack = { nullptr, nullptr, 0, 0 };
mergeStacksIntoProfile(aProfile, aSample, nativeStack);
#ifdef ENABLE_SPS_LEAF_DATA
if (aSample && aAddLeafAddresses) {
aProfile.addTag(ProfileEntry('l', (void*)aSample->pc));
#ifdef ENABLE_ARM_LR_SAVING
aProfile.addTag(ProfileEntry('L', (void*)aSample->lr));
#endif
}
#endif
}
void TableTicker::doBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
void* pc_array[1000];
size_t count = 0;
unw_cursor_t cursor; unw_context_t uc;
unw_word_t ip;
unw_getcontext(&uc);
// Dirty hack: replace the registers with values from the signal handler
// We do this in order to avoid the overhead of walking up to reach the
// signal handler frame, and the possibility that libunwind fails to
// handle it correctly.
unw_tdep_context_t *unw_ctx = reinterpret_cast<unw_tdep_context_t*> (&uc);
mcontext_t& mcontext = reinterpret_cast<ucontext_t*> (aSample->context)->uc_mcontext;
#define REPLACE_REG(num) unw_ctx->regs[num] = mcontext.gregs[R##num]
REPLACE_REG(0);
REPLACE_REG(1);
REPLACE_REG(2);
REPLACE_REG(3);
REPLACE_REG(4);
REPLACE_REG(5);
REPLACE_REG(6);
REPLACE_REG(7);
REPLACE_REG(8);
REPLACE_REG(9);
REPLACE_REG(10);
REPLACE_REG(11);
REPLACE_REG(12);
REPLACE_REG(13);
REPLACE_REG(14);
REPLACE_REG(15);
#undef REPLACE_REG
unw_init_local(&cursor, &uc);
while (count < ArrayLength(pc_array) &&
unw_step(&cursor) > 0) {
unw_get_reg(&cursor, UNW_REG_IP, &ip);
pc_array[count++] = reinterpret_cast<void*> (ip);
}
aProfile.addTag(ProfileEntry('s', "(root)"));
for (size_t i = count; i > 0; --i) {
aProfile.addTag(ProfileEntry('l', reinterpret_cast<void*>(pc_array[i - 1])));
}
}
static
void doSampleStackTrace(ProfileStack *aStack, ThreadProfile &aProfile, TickSample *sample)
{
// Sample
// 's' tag denotes the start of a sample block
// followed by 0 or more 'c' tags.
for (int i = 0; i < aStack->mStackPointer; i++) {
if (i == 0) {
Address pc = 0;
if (sample) {
pc = sample->pc;
}
aProfile.addTag(ProfileEntry('s', aStack->mStack[i], pc));
} else {
aProfile.addTag(ProfileEntry('c', aStack->mStack[i]));
}
}
}
static
void doSampleStackTrace(PseudoStack *aStack, ThreadProfile &aProfile, TickSample *sample)
{
// Sample
// 's' tag denotes the start of a sample block
// followed by 0 or more 'c' tags.
aProfile.addTag(ProfileEntry('s', "(root)"));
for (uint32_t i = 0; i < aStack->stackSize(); i++) {
addProfileEntry(aStack->mStack[i], aProfile, aStack, nullptr);
}
#ifdef ENABLE_SPS_LEAF_DATA
if (sample) {
aProfile.addTag(ProfileEntry('l', (void*)sample->pc));
#ifdef ENABLE_ARM_LR_SAVING
aProfile.addTag(ProfileEntry('L', (void*)sample->lr));
#endif
}
#endif
}
static
void addDynamicTag(ThreadProfile &aProfile, char aTagName, const char *aStr)
{
aProfile.addTag(ProfileEntry(aTagName, ""));
// Add one to store the null termination
size_t strLen = strlen(aStr) + 1;
for (size_t j = 0; j < strLen;) {
// Store as many characters in the void* as the platform allows
char text[sizeof(void*)];
size_t len = sizeof(void*)/sizeof(char);
if (j+len >= strLen) {
len = strLen - j;
}
memcpy(text, &aStr[j], len);
j += sizeof(void*)/sizeof(char);
// Cast to *((void**) to pass the text data to a void*
aProfile.addTag(ProfileEntry('d', *((void**)(&text[0]))));
}
}
static
void doSampleStackTrace(ProfileStack *aStack, ThreadProfile &aProfile, TickSample *sample)
{
// Sample
// 's' tag denotes the start of a sample block
// followed by 0 or more 'c' tags.
aProfile.addTag(ProfileEntry('s', "(root)"));
for (mozilla::sig_safe_t i = 0;
i < aStack->mStackPointer && i < mozilla::ArrayLength(aStack->mStack);
i++) {
addProfileEntry(aStack, aProfile, i);
}
#ifdef ENABLE_SPS_LEAF_DATA
if (sample) {
aProfile.addTag(ProfileEntry('l', (void*)sample->pc));
#ifdef ENABLE_ARM_LR_SAVING
aProfile.addTag(ProfileEntry('L', (void*)sample->lr));
#endif
}
#endif
}
static
void addProfileEntry(volatile StackEntry &entry, ThreadProfile &aProfile,
PseudoStack *stack, void *lastpc)
{
int lineno = -1;
// First entry has tagName 's' (start)
// Check for magic pointer bit 1 to indicate copy
const char* sampleLabel = entry.label();
if (entry.isCopyLabel()) {
// Store the string using 1 or more 'd' (dynamic) tags
// that will happen to the preceding tag
addDynamicTag(aProfile, 'c', sampleLabel);
if (entry.js()) {
if (!entry.pc()) {
// The JIT only allows the top-most entry to have a NULL pc
MOZ_ASSERT(&entry == &stack->mStack[stack->stackSize() - 1]);
// If stack-walking was disabled, then that's just unfortunate
if (lastpc) {
jsbytecode *jspc = js::ProfilingGetPC(stack->mRuntime, entry.script(),
lastpc);
if (jspc) {
lineno = JS_PCToLineNumber(NULL, entry.script(), jspc);
}
}
} else {
lineno = JS_PCToLineNumber(NULL, entry.script(), entry.pc());
}
} else {
lineno = entry.line();
}
} else {
aProfile.addTag(ProfileEntry('c', sampleLabel));
lineno = entry.line();
}
if (lineno != -1) {
aProfile.addTag(ProfileEntry('n', lineno));
}
}
static void mergeNativeBacktrace(ThreadProfile &aProfile, const PCArray &array) {
aProfile.addTag(ProfileEntry('s', "(root)"));
PseudoStack* stack = aProfile.GetPseudoStack();
uint32_t pseudoStackPos = 0;
/* We have two stacks, the native C stack we extracted from unwinding,
* and the pseudostack we managed during execution. We want to consolidate
* the two in order. We do so by merging using the approximate stack address
* when each entry was push. When pushing JS entry we may not now the stack
* address in which case we have a NULL stack address in which case we assume
* that it follows immediatly the previous element.
*
* C Stack | Address -- Pseudo Stack | Address
* main() | 0x100 run_js() | 0x40
* start() | 0x80 jsCanvas() | NULL
* timer() | 0x50 drawLine() | NULL
* azure() | 0x10
*
* Merged: main(), start(), timer(), run_js(), jsCanvas(), drawLine(), azure()
*/
// i is the index in C stack starting at main and decreasing
// pseudoStackPos is the position in the Pseudo stack starting
// at the first frame (run_js in the example) and increasing.
for (size_t i = array.count; i > 0; --i) {
while (pseudoStackPos < stack->stackSize()) {
volatile StackEntry& entry = stack->mStack[pseudoStackPos];
if (entry.stackAddress() < array.sp_array[i-1] && entry.stackAddress())
break;
addProfileEntry(entry, aProfile, stack, array.array[0]);
pseudoStackPos++;
}
aProfile.addTag(ProfileEntry('l', (void*)array.array[i-1]));
}
}
static
void doSampleStackTrace(ProfileStack *aStack, ThreadProfile &aProfile, TickSample *sample)
{
// Sample
// 's' tag denotes the start of a sample block
// followed by 0 or more 'c' tags.
aProfile.addTag(ProfileEntry('s', "(root)"));
for (mozilla::sig_safe_t i = 0; i < aStack->mStackPointer; i++) {
// First entry has tagName 's' (start)
// Check for magic pointer bit 1 to indicate copy
const char* sampleLabel = aStack->mStack[i].mLabel;
if (aStack->mStack[i].isCopyLabel()) {
// Store the string using 1 or more 'd' (dynamic) tags
// that will happen to the preceding tag
aProfile.addTag(ProfileEntry('c', ""));
// Add one to store the null termination
size_t strLen = strlen(sampleLabel) + 1;
for (size_t j = 0; j < strLen;) {
// Store as many characters in the void* as the platform allows
char text[sizeof(void*)];
for (size_t pos = 0; pos < sizeof(void*) && j+pos < strLen; pos++) {
text[pos] = sampleLabel[j+pos];
}
j += sizeof(void*);
// Take '*((void**)(&text[0]))' to pass the char[] as a single void*
aProfile.addTag(ProfileEntry('d', *((void**)(&text[0]))));
}
} else {
aProfile.addTag(ProfileEntry('c', sampleLabel));
}
}
#ifdef ENABLE_SPS_LEAF_DATA
if (sample) {
aProfile.addTag(ProfileEntry('l', (void*)sample->pc));
}
#endif
}
void TableTicker::doNativeBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
#ifdef XP_WIN
/* For the purpose of SPS it is currently too expensive to call NS_StackWalk()
on itself on Windows. For each call to NS_StackWalk(), we currently have to
use another thread to suspend the calling thread, obtain its context, and
walk its stack. We shouldn't do that every time we need a thread to obtain
its own backtrace in SPS.*/
if (aSample->isSamplingCurrentThread) {
void* ptrs[100];
USHORT count = RtlCaptureStackBackTrace(0, 100, ptrs, nullptr);
aProfile.addTag(ProfileEntry('s', "(root)"));
if (count) {
USHORT i = count;
do {
--i;
aProfile.addTag(ProfileEntry('l', ptrs[i]));
} while (i != 0);
}
return;
}
#endif
#ifndef XP_MACOSX
uintptr_t thread = GetThreadHandle(aSample->threadProfile->GetPlatformData());
MOZ_ASSERT(thread);
#endif
void* pc_array[1000];
void* sp_array[1000];
PCArray array = {
pc_array,
sp_array,
mozilla::ArrayLength(pc_array),
0
};
// Start with the current function.
StackWalkCallback(aSample->pc, aSample->sp, &array);
uint32_t maxFrames = uint32_t(array.size - array.count);
#ifdef XP_MACOSX
pthread_t pt = GetProfiledThread(aSample->threadProfile->GetPlatformData());
void *stackEnd = reinterpret_cast<void*>(-1);
if (pt)
stackEnd = static_cast<char*>(pthread_get_stackaddr_np(pt));
nsresult rv = NS_OK;
if (aSample->fp >= aSample->sp && aSample->fp <= stackEnd)
rv = FramePointerStackWalk(StackWalkCallback, /* skipFrames */ 0,
maxFrames, &array,
reinterpret_cast<void**>(aSample->fp), stackEnd);
#else
void *platformData = nullptr;
#ifdef XP_WIN
platformData = aSample->context;
#endif // XP_WIN
nsresult rv = NS_StackWalk(StackWalkCallback, /* skipFrames */ 0, maxFrames,
&array, thread, platformData);
#endif
if (NS_SUCCEEDED(rv))
mergeNativeBacktrace(aProfile, array);
}
static
void mergeStacksIntoProfile(ThreadProfile& aProfile, TickSample* aSample, NativeStack& aNativeStack)
{
PseudoStack* pseudoStack = aProfile.GetPseudoStack();
volatile StackEntry *pseudoFrames = pseudoStack->mStack;
uint32_t pseudoCount = pseudoStack->stackSize();
// Make a copy of the JS stack into a JSFrame array. This is necessary since,
// like the native stack, the JS stack is iterated youngest-to-oldest and we
// need to iterate oldest-to-youngest when adding entries to aProfile.
// Synchronous sampling reports an invalid buffer generation to
// ProfilingFrameIterator to avoid incorrectly resetting the generation of
// sampled JIT entries inside the JS engine. See note below concerning 'J'
// entries.
uint32_t startBufferGen;
if (aSample->isSamplingCurrentThread) {
startBufferGen = UINT32_MAX;
} else {
startBufferGen = aProfile.bufferGeneration();
}
uint32_t jsCount = 0;
#ifndef SPS_STANDALONE
JS::ProfilingFrameIterator::Frame jsFrames[1000];
// Only walk jit stack if profiling frame iterator is turned on.
if (pseudoStack->mRuntime && JS::IsProfilingEnabledForRuntime(pseudoStack->mRuntime)) {
AutoWalkJSStack autoWalkJSStack;
const uint32_t maxFrames = mozilla::ArrayLength(jsFrames);
if (aSample && autoWalkJSStack.walkAllowed) {
JS::ProfilingFrameIterator::RegisterState registerState;
registerState.pc = aSample->pc;
registerState.sp = aSample->sp;
#ifdef ENABLE_ARM_LR_SAVING
registerState.lr = aSample->lr;
#endif
JS::ProfilingFrameIterator jsIter(pseudoStack->mRuntime,
registerState,
startBufferGen);
for (; jsCount < maxFrames && !jsIter.done(); ++jsIter) {
// See note below regarding 'J' entries.
if (aSample->isSamplingCurrentThread || jsIter.isAsmJS()) {
uint32_t extracted = jsIter.extractStack(jsFrames, jsCount, maxFrames);
jsCount += extracted;
if (jsCount == maxFrames)
break;
} else {
mozilla::Maybe<JS::ProfilingFrameIterator::Frame> frame =
jsIter.getPhysicalFrameWithoutLabel();
if (frame.isSome())
jsFrames[jsCount++] = frame.value();
}
}
}
}
#endif
// Start the sample with a root entry.
aProfile.addTag(ProfileEntry('s', "(root)"));
// While the pseudo-stack array is ordered oldest-to-youngest, the JS and
// native arrays are ordered youngest-to-oldest. We must add frames to
// aProfile oldest-to-youngest. Thus, iterate over the pseudo-stack forwards
// and JS and native arrays backwards. Note: this means the terminating
// condition jsIndex and nativeIndex is being < 0.
uint32_t pseudoIndex = 0;
int32_t jsIndex = jsCount - 1;
int32_t nativeIndex = aNativeStack.count - 1;
uint8_t *lastPseudoCppStackAddr = nullptr;
// Iterate as long as there is at least one frame remaining.
while (pseudoIndex != pseudoCount || jsIndex >= 0 || nativeIndex >= 0) {
// There are 1 to 3 frames available. Find and add the oldest.
uint8_t *pseudoStackAddr = nullptr;
uint8_t *jsStackAddr = nullptr;
uint8_t *nativeStackAddr = nullptr;
if (pseudoIndex != pseudoCount) {
volatile StackEntry &pseudoFrame = pseudoFrames[pseudoIndex];
if (pseudoFrame.isCpp())
lastPseudoCppStackAddr = (uint8_t *) pseudoFrame.stackAddress();
#ifndef SPS_STANDALONE
// Skip any pseudo-stack JS frames which are marked isOSR
// Pseudostack frames are marked isOSR when the JS interpreter
// enters a jit frame on a loop edge (via on-stack-replacement,
// or OSR). To avoid both the pseudoframe and jit frame being
// recorded (and showing up twice), the interpreter marks the
// interpreter pseudostack entry with the OSR flag to ensure that
// it doesn't get counted.
if (pseudoFrame.isJs() && pseudoFrame.isOSR()) {
pseudoIndex++;
continue;
}
#endif
MOZ_ASSERT(lastPseudoCppStackAddr);
pseudoStackAddr = lastPseudoCppStackAddr;
}
#ifndef SPS_STANDALONE
if (jsIndex >= 0)
jsStackAddr = (uint8_t *) jsFrames[jsIndex].stackAddress;
#endif
if (nativeIndex >= 0)
nativeStackAddr = (uint8_t *) aNativeStack.sp_array[nativeIndex];
// If there's a native stack entry which has the same SP as a
// pseudo stack entry, pretend we didn't see the native stack
// entry. Ditto for a native stack entry which has the same SP as
// a JS stack entry. In effect this means pseudo or JS entries
// trump conflicting native entries.
if (nativeStackAddr && (pseudoStackAddr == nativeStackAddr || jsStackAddr == nativeStackAddr)) {
nativeStackAddr = nullptr;
nativeIndex--;
MOZ_ASSERT(pseudoStackAddr || jsStackAddr);
}
// Sanity checks.
MOZ_ASSERT_IF(pseudoStackAddr, pseudoStackAddr != jsStackAddr &&
pseudoStackAddr != nativeStackAddr);
MOZ_ASSERT_IF(jsStackAddr, jsStackAddr != pseudoStackAddr &&
jsStackAddr != nativeStackAddr);
MOZ_ASSERT_IF(nativeStackAddr, nativeStackAddr != pseudoStackAddr &&
nativeStackAddr != jsStackAddr);
// Check to see if pseudoStack frame is top-most.
if (pseudoStackAddr > jsStackAddr && pseudoStackAddr > nativeStackAddr) {
MOZ_ASSERT(pseudoIndex < pseudoCount);
volatile StackEntry &pseudoFrame = pseudoFrames[pseudoIndex];
addPseudoEntry(pseudoFrame, aProfile, pseudoStack, nullptr);
pseudoIndex++;
continue;
}
#ifndef SPS_STANDALONE
// Check to see if JS jit stack frame is top-most
if (jsStackAddr > nativeStackAddr) {
MOZ_ASSERT(jsIndex >= 0);
const JS::ProfilingFrameIterator::Frame& jsFrame = jsFrames[jsIndex];
// Stringifying non-asm.js JIT frames is delayed until streaming
// time. To re-lookup the entry in the JitcodeGlobalTable, we need to
// store the JIT code address ('J') in the circular buffer.
//
// Note that we cannot do this when we are sychronously sampling the
// current thread; that is, when called from profiler_get_backtrace. The
// captured backtrace is usually externally stored for an indeterminate
// amount of time, such as in nsRefreshDriver. Problematically, the
// stored backtrace may be alive across a GC during which the profiler
// itself is disabled. In that case, the JS engine is free to discard
// its JIT code. This means that if we inserted such 'J' entries into
// the buffer, nsRefreshDriver would now be holding on to a backtrace
// with stale JIT code return addresses.
if (aSample->isSamplingCurrentThread ||
jsFrame.kind == JS::ProfilingFrameIterator::Frame_AsmJS) {
addDynamicTag(aProfile, 'c', jsFrame.label);
} else {
MOZ_ASSERT(jsFrame.kind == JS::ProfilingFrameIterator::Frame_Ion ||
jsFrame.kind == JS::ProfilingFrameIterator::Frame_Baseline);
aProfile.addTag(ProfileEntry('J', jsFrames[jsIndex].returnAddress));
}
jsIndex--;
continue;
}
#endif
// If we reach here, there must be a native stack entry and it must be the
// greatest entry.
if (nativeStackAddr) {
MOZ_ASSERT(nativeIndex >= 0);
aProfile
.addTag(ProfileEntry('l', (void*)aNativeStack.pc_array[nativeIndex]));
}
if (nativeIndex >= 0) {
nativeIndex--;
}
}
#ifndef SPS_STANDALONE
// Update the JS runtime with the current profile sample buffer generation.
//
// Do not do this for synchronous sampling, which create their own
// ProfileBuffers.
if (!aSample->isSamplingCurrentThread && pseudoStack->mRuntime) {
MOZ_ASSERT(aProfile.bufferGeneration() >= startBufferGen);
uint32_t lapCount = aProfile.bufferGeneration() - startBufferGen;
JS::UpdateJSRuntimeProfilerSampleBufferGen(pseudoStack->mRuntime,
aProfile.bufferGeneration(),
lapCount);
}
#endif
}
static
void mergeStacksIntoProfile(ThreadProfile& aProfile, TickSample* aSample, NativeStack& aNativeStack)
{
PseudoStack* pseudoStack = aProfile.GetPseudoStack();
volatile StackEntry *pseudoFrames = pseudoStack->mStack;
uint32_t pseudoCount = pseudoStack->stackSize();
// Make a copy of the JS stack into a JSFrame array. This is necessary since,
// like the native stack, the JS stack is iterated youngest-to-oldest and we
// need to iterate oldest-to-youngest when adding entries to aProfile.
JSFrame jsFrames[1000];
uint32_t jsCount = 0;
if (aSample && pseudoStack->mRuntime) {
JS::ProfilingFrameIterator::RegisterState registerState;
registerState.pc = aSample->pc;
registerState.sp = aSample->sp;
#ifdef ENABLE_ARM_LR_SAVING
registerState.lr = aSample->lr;
#endif
JS::ProfilingFrameIterator jsIter(pseudoStack->mRuntime, registerState);
for (; jsCount < mozilla::ArrayLength(jsFrames) && !jsIter.done(); ++jsCount, ++jsIter) {
jsFrames[jsCount].stackAddress = jsIter.stackAddress();
jsFrames[jsCount].label = jsIter.label();
}
}
// Start the sample with a root entry.
aProfile.addTag(ProfileEntry('s', "(root)"));
// While the pseudo-stack array is ordered oldest-to-youngest, the JS and
// native arrays are ordered youngest-to-oldest. We must add frames to
// aProfile oldest-to-youngest. Thus, iterate over the pseudo-stack forwards
// and JS and native arrays backwards. Note: this means the terminating
// condition jsIndex and nativeIndex is being < 0.
uint32_t pseudoIndex = 0;
int32_t jsIndex = jsCount - 1;
int32_t nativeIndex = aNativeStack.count - 1;
// Iterate as long as there is at least one frame remaining.
while (pseudoIndex != pseudoCount || jsIndex >= 0 || nativeIndex >= 0) {
// There are 1 to 3 frames available. Find and add the oldest. Handle pseudo
// frames first, since there are two special cases that must be considered
// before everything else.
if (pseudoIndex != pseudoCount) {
volatile StackEntry &pseudoFrame = pseudoFrames[pseudoIndex];
// isJs pseudo-stack frames assume the stackAddress of the preceding isCpp
// pseudo-stack frame. If we arrive at an isJs pseudo frame, we've already
// encountered the preceding isCpp stack frame and it was oldest, we can
// assume the isJs frame is oldest without checking other frames.
if (pseudoFrame.isJs()) {
addPseudoEntry(pseudoFrame, aProfile, pseudoStack, nullptr);
pseudoIndex++;
continue;
}
// Currently, only asm.js frames use the JS stack and Ion/Baseline/Interp
// frames use the pseudo stack. In the optimized asm.js->Ion call path, no
// isCpp frame is pushed, leading to the callstack:
// old | pseudo isCpp | asm.js | pseudo isJs | new
// Since there is no interleaving isCpp pseudo frame between the asm.js
// and isJs pseudo frame, the above isJs logic will render the callstack:
// old | pseudo isCpp | pseudo isJs | asm.js | new
// which is wrong. To deal with this, a pseudo isCpp frame pushed right
// before entering asm.js flagged with StackEntry::ASMJS. When we see this
// flag, we first push all the asm.js frames (up to the next frame with a
// stackAddress) before pushing the isJs frames. There is no Ion->asm.js
// fast path, so we don't have to worry about asm.js->Ion->asm.js.
//
// (This and the above isJs special cases can be removed once all JS
// execution modes switch from the pseudo stack to the JS stack.)
if (pseudoFrame.hasFlag(StackEntry::ASMJS)) {
void *stopStackAddress = nullptr;
for (uint32_t i = pseudoIndex + 1; i != pseudoCount; i++) {
if (pseudoFrames[i].isCpp()) {
stopStackAddress = pseudoFrames[i].stackAddress();
break;
}
}
if (nativeIndex >= 0) {
stopStackAddress = std::max(stopStackAddress, aNativeStack.sp_array[nativeIndex]);
}
while (jsIndex >= 0 && jsFrames[jsIndex].stackAddress > stopStackAddress) {
addDynamicTag(aProfile, 'c', jsFrames[jsIndex].label);
jsIndex--;
}
pseudoIndex++;
continue;
}
// Finally, consider the normal case of a plain C++ pseudo-frame.
if ((jsIndex < 0 || pseudoFrame.stackAddress() > jsFrames[jsIndex].stackAddress) &&
(nativeIndex < 0 || pseudoFrame.stackAddress() > aNativeStack.sp_array[nativeIndex]))
{
// The (C++) pseudo-frame is the oldest.
addPseudoEntry(pseudoFrame, aProfile, pseudoStack, nullptr);
pseudoIndex++;
continue;
}
}
if (jsIndex >= 0) {
// Test whether the JS frame is the oldest.
JSFrame &jsFrame = jsFrames[jsIndex];
if ((pseudoIndex == pseudoCount || jsFrame.stackAddress > pseudoFrames[pseudoIndex].stackAddress()) &&
(nativeIndex < 0 || jsFrame.stackAddress > aNativeStack.sp_array[nativeIndex]))
{
// The JS frame is the oldest.
addDynamicTag(aProfile, 'c', jsFrame.label);
jsIndex--;
continue;
}
}
// If execution reaches this point, there must be a native frame and it must
// be the oldest.
MOZ_ASSERT(nativeIndex >= 0);
aProfile.addTag(ProfileEntry('l', (void*)aNativeStack.pc_array[nativeIndex]));
nativeIndex--;
}
}
void TableTicker::doBacktrace(ThreadProfile &aProfile, TickSample* aSample)
{
#ifndef XP_MACOSX
uintptr_t thread = GetThreadHandle(platform_data());
MOZ_ASSERT(thread);
#endif
void* pc_array[1000];
void* sp_array[1000];
PCArray array = {
pc_array,
sp_array,
mozilla::ArrayLength(pc_array),
0
};
// Start with the current function.
StackWalkCallback(aSample->pc, aSample->sp, &array);
#ifdef XP_MACOSX
pthread_t pt = GetProfiledThread(platform_data());
void *stackEnd = reinterpret_cast<void*>(-1);
if (pt)
stackEnd = static_cast<char*>(pthread_get_stackaddr_np(pt));
nsresult rv = FramePointerStackWalk(StackWalkCallback, 0, &array, reinterpret_cast<void**>(aSample->fp), stackEnd);
#else
nsresult rv = NS_StackWalk(StackWalkCallback, 0, &array, thread);
#endif
if (NS_SUCCEEDED(rv)) {
aProfile.addTag(ProfileEntry('s', "(root)"));
ProfileStack* stack = aProfile.GetStack();
int pseudoStackPos = 0;
/* We have two stacks, the native C stack we extracted from unwinding,
* and the pseudostack we managed during execution. We want to consolidate
* the two in order. We do so by merging using the approximate stack address
* when each entry was push. When pushing JS entry we may not now the stack
* address in which case we have a NULL stack address in which case we assume
* that it follows immediatly the previous element.
*
* C Stack | Address -- Pseudo Stack | Address
* main() | 0x100 run_js() | 0x40
* start() | 0x80 jsCanvas() | NULL
* timer() | 0x50 drawLine() | NULL
* azure() | 0x10
*
* Merged: main(), start(), timer(), run_js(), jsCanvas(), drawLine(), azure()
*/
// i is the index in C stack starting at main and decreasing
// pseudoStackPos is the position in the Pseudo stack starting
// at the first frame (run_js in the example) and increasing.
for (size_t i = array.count; i > 0; --i) {
while (pseudoStackPos < stack->stackSize()) {
volatile StackEntry& entry = stack->mStack[pseudoStackPos];
if (entry.stackAddress() < array.sp_array[i-1] && entry.stackAddress())
break;
addProfileEntry(entry, aProfile, stack, array.array[0]);
pseudoStackPos++;
}
aProfile.addTag(ProfileEntry('l', (void*)array.array[i-1]));
}
}
}
static
void mergeStacksIntoProfile(ThreadProfile& aProfile, TickSample* aSample, NativeStack& aNativeStack)
{
PseudoStack* pseudoStack = aProfile.GetPseudoStack();
volatile StackEntry *pseudoFrames = pseudoStack->mStack;
uint32_t pseudoCount = pseudoStack->stackSize();
// Make a copy of the JS stack into a JSFrame array. This is necessary since,
// like the native stack, the JS stack is iterated youngest-to-oldest and we
// need to iterate oldest-to-youngest when adding entries to aProfile.
uint32_t startBufferGen = aProfile.bufferGeneration();
uint32_t jsCount = 0;
JS::ProfilingFrameIterator::Frame jsFrames[1000];
// Only walk jit stack if profiling frame iterator is turned on.
if (pseudoStack->mRuntime && JS::IsProfilingEnabledForRuntime(pseudoStack->mRuntime)) {
AutoWalkJSStack autoWalkJSStack;
const uint32_t maxFrames = mozilla::ArrayLength(jsFrames);
if (aSample && autoWalkJSStack.walkAllowed) {
JS::ProfilingFrameIterator::RegisterState registerState;
registerState.pc = aSample->pc;
registerState.sp = aSample->sp;
#ifdef ENABLE_ARM_LR_SAVING
registerState.lr = aSample->lr;
#endif
JS::ProfilingFrameIterator jsIter(pseudoStack->mRuntime,
registerState,
startBufferGen);
for (; jsCount < maxFrames && !jsIter.done(); ++jsIter) {
uint32_t extracted = jsIter.extractStack(jsFrames, jsCount, maxFrames);
MOZ_ASSERT(extracted <= (maxFrames - jsCount));
jsCount += extracted;
if (jsCount == maxFrames)
break;
}
}
}
// Start the sample with a root entry.
aProfile.addTag(ProfileEntry('s', "(root)"));
// While the pseudo-stack array is ordered oldest-to-youngest, the JS and
// native arrays are ordered youngest-to-oldest. We must add frames to
// aProfile oldest-to-youngest. Thus, iterate over the pseudo-stack forwards
// and JS and native arrays backwards. Note: this means the terminating
// condition jsIndex and nativeIndex is being < 0.
uint32_t pseudoIndex = 0;
int32_t jsIndex = jsCount - 1;
int32_t nativeIndex = aNativeStack.count - 1;
uint8_t *lastPseudoCppStackAddr = nullptr;
// Iterate as long as there is at least one frame remaining.
while (pseudoIndex != pseudoCount || jsIndex >= 0 || nativeIndex >= 0) {
// There are 1 to 3 frames available. Find and add the oldest.
uint8_t *pseudoStackAddr = nullptr;
uint8_t *jsStackAddr = nullptr;
uint8_t *nativeStackAddr = nullptr;
if (pseudoIndex != pseudoCount) {
volatile StackEntry &pseudoFrame = pseudoFrames[pseudoIndex];
if (pseudoFrame.isCpp())
lastPseudoCppStackAddr = (uint8_t *) pseudoFrame.stackAddress();
// Skip any pseudo-stack JS frames which are marked isOSR
// Pseudostack frames are marked isOSR when the JS interpreter
// enters a jit frame on a loop edge (via on-stack-replacement,
// or OSR). To avoid both the pseudoframe and jit frame being
// recorded (and showing up twice), the interpreter marks the
// interpreter pseudostack entry with the OSR flag to ensure that
// it doesn't get counted.
if (pseudoFrame.isJs() && pseudoFrame.isOSR()) {
pseudoIndex++;
continue;
}
MOZ_ASSERT(lastPseudoCppStackAddr);
pseudoStackAddr = lastPseudoCppStackAddr;
}
if (jsIndex >= 0)
jsStackAddr = (uint8_t *) jsFrames[jsIndex].stackAddress;
if (nativeIndex >= 0)
nativeStackAddr = (uint8_t *) aNativeStack.sp_array[nativeIndex];
// Sanity checks.
MOZ_ASSERT_IF(pseudoStackAddr, pseudoStackAddr != jsStackAddr &&
pseudoStackAddr != nativeStackAddr);
MOZ_ASSERT_IF(jsStackAddr, jsStackAddr != pseudoStackAddr &&
jsStackAddr != nativeStackAddr);
MOZ_ASSERT_IF(nativeStackAddr, nativeStackAddr != pseudoStackAddr &&
nativeStackAddr != jsStackAddr);
// Check to see if pseudoStack frame is top-most.
if (pseudoStackAddr > jsStackAddr && pseudoStackAddr > nativeStackAddr) {
MOZ_ASSERT(pseudoIndex < pseudoCount);
volatile StackEntry &pseudoFrame = pseudoFrames[pseudoIndex];
addPseudoEntry(pseudoFrame, aProfile, pseudoStack, nullptr);
pseudoIndex++;
continue;
}
// Check to see if JS jit stack frame is top-most
if (jsStackAddr > nativeStackAddr) {
MOZ_ASSERT(jsIndex >= 0);
addDynamicTag(aProfile, 'c', jsFrames[jsIndex].label);
// Stringifying optimization information is delayed until streaming
// time. To re-lookup the entry in the JitcodeGlobalTable, we need to
// store the JIT code address ('J') in the circular buffer.
if (jsFrames[jsIndex].hasTrackedOptimizations) {
aProfile.addTag(ProfileEntry('J', jsFrames[jsIndex].returnAddress));
}
jsIndex--;
continue;
}
// If we reach here, there must be a native stack entry and it must be the
// greatest entry.
MOZ_ASSERT(nativeStackAddr);
MOZ_ASSERT(nativeIndex >= 0);
aProfile.addTag(ProfileEntry('l', (void*)aNativeStack.pc_array[nativeIndex]));
nativeIndex--;
}
MOZ_ASSERT(aProfile.bufferGeneration() >= startBufferGen);
uint32_t lapCount = aProfile.bufferGeneration() - startBufferGen;
// Update the JS runtime with the current profile sample buffer generation.
if (pseudoStack->mRuntime) {
JS::UpdateJSRuntimeProfilerSampleBufferGen(pseudoStack->mRuntime,
aProfile.bufferGeneration(),
lapCount);
}
}
