Auto merge of #90617 - tmiasko:time-trace-threads, r=wesleywiser
Initialize LLVM time trace profiler on each code generation thread In https://reviews.llvm.org/D71059 LLVM 11, the time trace profiler was extended to support multiple threads. `timeTraceProfilerInitialize` creates a thread local profiler instance. When a thread finishes `timeTraceProfilerFinishThread` moves a thread local instance into a global collection of instances. Finally when all codegen work is complete `timeTraceProfilerWrite` writes data from the current thread local instance and the instances in global collection of instances. Previously, the profiler was intialized on a single thread only. Since this thread performs no code generation on its own, the resulting profile was empty. Update LLVM codegen to initialize & finish time trace profiler on each code generation thread. cc `@tmandry` r? `@wesleywiser`
This commit is contained in:
commit
3cd3bbecc5
@ -76,6 +76,27 @@ pub mod llvm {
|
||||
#[derive(Clone)]
|
||||
pub struct LlvmCodegenBackend(());
|
||||
|
||||
struct TimeTraceProfiler {
|
||||
enabled: bool,
|
||||
}
|
||||
|
||||
impl TimeTraceProfiler {
|
||||
fn new(enabled: bool) -> Self {
|
||||
if enabled {
|
||||
unsafe { llvm::LLVMTimeTraceProfilerInitialize() }
|
||||
}
|
||||
TimeTraceProfiler { enabled }
|
||||
}
|
||||
}
|
||||
|
||||
impl Drop for TimeTraceProfiler {
|
||||
fn drop(&mut self) {
|
||||
if self.enabled {
|
||||
unsafe { llvm::LLVMTimeTraceProfilerFinishThread() }
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
impl ExtraBackendMethods for LlvmCodegenBackend {
|
||||
fn new_metadata(&self, tcx: TyCtxt<'_>, mod_name: &str) -> ModuleLlvm {
|
||||
ModuleLlvm::new_metadata(tcx, mod_name)
|
||||
@ -119,6 +140,34 @@ fn target_cpu<'b>(&self, sess: &'b Session) -> &'b str {
|
||||
fn tune_cpu<'b>(&self, sess: &'b Session) -> Option<&'b str> {
|
||||
llvm_util::tune_cpu(sess)
|
||||
}
|
||||
|
||||
fn spawn_thread<F, T>(time_trace: bool, f: F) -> std::thread::JoinHandle<T>
|
||||
where
|
||||
F: FnOnce() -> T,
|
||||
F: Send + 'static,
|
||||
T: Send + 'static,
|
||||
{
|
||||
std::thread::spawn(move || {
|
||||
let _profiler = TimeTraceProfiler::new(time_trace);
|
||||
f()
|
||||
})
|
||||
}
|
||||
|
||||
fn spawn_named_thread<F, T>(
|
||||
time_trace: bool,
|
||||
name: String,
|
||||
f: F,
|
||||
) -> std::io::Result<std::thread::JoinHandle<T>>
|
||||
where
|
||||
F: FnOnce() -> T,
|
||||
F: Send + 'static,
|
||||
T: Send + 'static,
|
||||
{
|
||||
std::thread::Builder::new().name(name).spawn(move || {
|
||||
let _profiler = TimeTraceProfiler::new(time_trace);
|
||||
f()
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
impl WriteBackendMethods for LlvmCodegenBackend {
|
||||
|
@ -1737,6 +1737,8 @@ pub fn LLVMRustBuildAtomicFence(
|
||||
|
||||
pub fn LLVMTimeTraceProfilerInitialize();
|
||||
|
||||
pub fn LLVMTimeTraceProfilerFinishThread();
|
||||
|
||||
pub fn LLVMTimeTraceProfilerFinish(FileName: *const c_char);
|
||||
|
||||
pub fn LLVMAddAnalysisPasses(T: &'a TargetMachine, PM: &PassManager<'a>);
|
||||
|
@ -113,11 +113,6 @@ fn llvm_arg_to_arg_name(full_arg: &str) -> &str {
|
||||
}
|
||||
|
||||
if sess.opts.debugging_opts.llvm_time_trace {
|
||||
// time-trace is not thread safe and running it in parallel will cause seg faults.
|
||||
if !sess.opts.debugging_opts.no_parallel_llvm {
|
||||
bug!("`-Z llvm-time-trace` requires `-Z no-parallel-llvm")
|
||||
}
|
||||
|
||||
llvm::LLVMTimeTraceProfilerInitialize();
|
||||
}
|
||||
|
||||
|
@ -310,6 +310,7 @@ pub struct CodegenContext<B: WriteBackendMethods> {
|
||||
pub no_landing_pads: bool,
|
||||
pub save_temps: bool,
|
||||
pub fewer_names: bool,
|
||||
pub time_trace: bool,
|
||||
pub exported_symbols: Option<Arc<ExportedSymbols>>,
|
||||
pub opts: Arc<config::Options>,
|
||||
pub crate_types: Vec<CrateType>,
|
||||
@ -1039,6 +1040,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
|
||||
no_landing_pads: sess.panic_strategy() == PanicStrategy::Abort,
|
||||
fewer_names: sess.fewer_names(),
|
||||
save_temps: sess.opts.cg.save_temps,
|
||||
time_trace: sess.opts.debugging_opts.llvm_time_trace,
|
||||
opts: Arc::new(sess.opts.clone()),
|
||||
prof: sess.prof.clone(),
|
||||
exported_symbols,
|
||||
@ -1198,7 +1200,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
|
||||
// Each LLVM module is automatically sent back to the coordinator for LTO if
|
||||
// necessary. There's already optimizations in place to avoid sending work
|
||||
// back to the coordinator if LTO isn't requested.
|
||||
return thread::spawn(move || {
|
||||
return B::spawn_thread(cgcx.time_trace, move || {
|
||||
let mut worker_id_counter = 0;
|
||||
let mut free_worker_ids = Vec::new();
|
||||
let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
|
||||
@ -1615,59 +1617,57 @@ fn maybe_start_llvm_timer<'a>(
|
||||
pub struct WorkerFatalError;
|
||||
|
||||
fn spawn_work<B: ExtraBackendMethods>(cgcx: CodegenContext<B>, work: WorkItem<B>) {
|
||||
let builder = thread::Builder::new().name(work.short_description());
|
||||
builder
|
||||
.spawn(move || {
|
||||
// Set up a destructor which will fire off a message that we're done as
|
||||
// we exit.
|
||||
struct Bomb<B: ExtraBackendMethods> {
|
||||
coordinator_send: Sender<Box<dyn Any + Send>>,
|
||||
result: Option<Result<WorkItemResult<B>, FatalError>>,
|
||||
worker_id: usize,
|
||||
}
|
||||
impl<B: ExtraBackendMethods> Drop for Bomb<B> {
|
||||
fn drop(&mut self) {
|
||||
let worker_id = self.worker_id;
|
||||
let msg = match self.result.take() {
|
||||
Some(Ok(WorkItemResult::Compiled(m))) => {
|
||||
Message::Done::<B> { result: Ok(m), worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsLink(m))) => {
|
||||
Message::NeedsLink::<B> { module: m, worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsFatLTO(m))) => {
|
||||
Message::NeedsFatLTO::<B> { result: m, worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))) => {
|
||||
Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id }
|
||||
}
|
||||
Some(Err(FatalError)) => {
|
||||
Message::Done::<B> { result: Err(Some(WorkerFatalError)), worker_id }
|
||||
}
|
||||
None => Message::Done::<B> { result: Err(None), worker_id },
|
||||
};
|
||||
drop(self.coordinator_send.send(Box::new(msg)));
|
||||
}
|
||||
B::spawn_named_thread(cgcx.time_trace, work.short_description(), move || {
|
||||
// Set up a destructor which will fire off a message that we're done as
|
||||
// we exit.
|
||||
struct Bomb<B: ExtraBackendMethods> {
|
||||
coordinator_send: Sender<Box<dyn Any + Send>>,
|
||||
result: Option<Result<WorkItemResult<B>, FatalError>>,
|
||||
worker_id: usize,
|
||||
}
|
||||
impl<B: ExtraBackendMethods> Drop for Bomb<B> {
|
||||
fn drop(&mut self) {
|
||||
let worker_id = self.worker_id;
|
||||
let msg = match self.result.take() {
|
||||
Some(Ok(WorkItemResult::Compiled(m))) => {
|
||||
Message::Done::<B> { result: Ok(m), worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsLink(m))) => {
|
||||
Message::NeedsLink::<B> { module: m, worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsFatLTO(m))) => {
|
||||
Message::NeedsFatLTO::<B> { result: m, worker_id }
|
||||
}
|
||||
Some(Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))) => {
|
||||
Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id }
|
||||
}
|
||||
Some(Err(FatalError)) => {
|
||||
Message::Done::<B> { result: Err(Some(WorkerFatalError)), worker_id }
|
||||
}
|
||||
None => Message::Done::<B> { result: Err(None), worker_id },
|
||||
};
|
||||
drop(self.coordinator_send.send(Box::new(msg)));
|
||||
}
|
||||
}
|
||||
|
||||
let mut bomb = Bomb::<B> {
|
||||
coordinator_send: cgcx.coordinator_send.clone(),
|
||||
result: None,
|
||||
worker_id: cgcx.worker,
|
||||
};
|
||||
let mut bomb = Bomb::<B> {
|
||||
coordinator_send: cgcx.coordinator_send.clone(),
|
||||
result: None,
|
||||
worker_id: cgcx.worker,
|
||||
};
|
||||
|
||||
// Execute the work itself, and if it finishes successfully then flag
|
||||
// ourselves as a success as well.
|
||||
//
|
||||
// Note that we ignore any `FatalError` coming out of `execute_work_item`,
|
||||
// as a diagnostic was already sent off to the main thread - just
|
||||
// surface that there was an error in this worker.
|
||||
bomb.result = {
|
||||
let _prof_timer = work.start_profiling(&cgcx);
|
||||
Some(execute_work_item(&cgcx, work))
|
||||
};
|
||||
})
|
||||
.expect("failed to spawn thread");
|
||||
// Execute the work itself, and if it finishes successfully then flag
|
||||
// ourselves as a success as well.
|
||||
//
|
||||
// Note that we ignore any `FatalError` coming out of `execute_work_item`,
|
||||
// as a diagnostic was already sent off to the main thread - just
|
||||
// surface that there was an error in this worker.
|
||||
bomb.result = {
|
||||
let _prof_timer = work.start_profiling(&cgcx);
|
||||
Some(execute_work_item(&cgcx, work))
|
||||
};
|
||||
})
|
||||
.expect("failed to spawn thread");
|
||||
}
|
||||
|
||||
enum SharedEmitterMessage {
|
||||
|
@ -142,4 +142,26 @@ fn target_machine_factory(
|
||||
) -> TargetMachineFactoryFn<Self>;
|
||||
fn target_cpu<'b>(&self, sess: &'b Session) -> &'b str;
|
||||
fn tune_cpu<'b>(&self, sess: &'b Session) -> Option<&'b str>;
|
||||
|
||||
fn spawn_thread<F, T>(_time_trace: bool, f: F) -> std::thread::JoinHandle<T>
|
||||
where
|
||||
F: FnOnce() -> T,
|
||||
F: Send + 'static,
|
||||
T: Send + 'static,
|
||||
{
|
||||
std::thread::spawn(f)
|
||||
}
|
||||
|
||||
fn spawn_named_thread<F, T>(
|
||||
_time_trace: bool,
|
||||
name: String,
|
||||
f: F,
|
||||
) -> std::io::Result<std::thread::JoinHandle<T>>
|
||||
where
|
||||
F: FnOnce() -> T,
|
||||
F: Send + 'static,
|
||||
T: Send + 'static,
|
||||
{
|
||||
std::thread::Builder::new().name(name).spawn(f)
|
||||
}
|
||||
}
|
||||
|
@ -75,6 +75,10 @@ extern "C" void LLVMTimeTraceProfilerInitialize() {
|
||||
/* ProcName */ "rustc");
|
||||
}
|
||||
|
||||
extern "C" void LLVMTimeTraceProfilerFinishThread() {
|
||||
timeTraceProfilerFinishThread();
|
||||
}
|
||||
|
||||
extern "C" void LLVMTimeTraceProfilerFinish(const char* FileName) {
|
||||
StringRef FN(FileName);
|
||||
std::error_code EC;
|
||||
|
Loading…
Reference in New Issue
Block a user