Translate MIR to clif ir in parallel with parallel rustc

On dev-desktop the advantage of cg_clif over cg_llvm on simple-raytracer
is 15% when parallel rustc is disabled. With -Zthreads=16 the advantage
goes from 5% to 22% with this change.
This commit is contained in:
bjorn3 2024-05-11 18:51:59 +00:00
parent 50b34279c3
commit a167142946
2 changed files with 31 additions and 31 deletions

View File

@ -6,7 +6,7 @@
// FIXME don't panic when a worker thread panics // FIXME don't panic when a worker thread panics
pub(super) struct ConcurrencyLimiter { pub(super) struct ConcurrencyLimiter {
helper_thread: Option<HelperThread>, helper_thread: Option<Mutex<HelperThread>>,
state: Arc<Mutex<state::ConcurrencyLimiterState>>, state: Arc<Mutex<state::ConcurrencyLimiterState>>,
available_token_condvar: Arc<Condvar>, available_token_condvar: Arc<Condvar>,
finished: bool, finished: bool,
@ -39,14 +39,14 @@ pub(super) fn new(sess: &Session, pending_jobs: usize) -> Self {
}) })
.unwrap(); .unwrap();
ConcurrencyLimiter { ConcurrencyLimiter {
helper_thread: Some(helper_thread), helper_thread: Some(Mutex::new(helper_thread)),
state, state,
available_token_condvar, available_token_condvar,
finished: false, finished: false,
} }
} }
pub(super) fn acquire(&mut self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLimiterToken { pub(super) fn acquire(&self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLimiterToken {
let mut state = self.state.lock().unwrap(); let mut state = self.state.lock().unwrap();
loop { loop {
state.assert_invariants(); state.assert_invariants();
@ -73,7 +73,7 @@ pub(super) fn acquire(&mut self, dcx: &rustc_errors::DiagCtxt) -> ConcurrencyLim
} }
} }
self.helper_thread.as_mut().unwrap().request_token(); self.helper_thread.as_ref().unwrap().lock().unwrap().request_token();
state = self.available_token_condvar.wait(state).unwrap(); state = self.available_token_condvar.wait(state).unwrap();
} }
} }

View File

@ -15,6 +15,7 @@
use rustc_codegen_ssa::{CodegenResults, CompiledModule, CrateInfo, ModuleKind}; use rustc_codegen_ssa::{CodegenResults, CompiledModule, CrateInfo, ModuleKind};
use rustc_data_structures::profiling::SelfProfilerRef; use rustc_data_structures::profiling::SelfProfilerRef;
use rustc_data_structures::stable_hasher::{HashStable, StableHasher}; use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
use rustc_data_structures::sync::{par_map, IntoDynSyncSend};
use rustc_metadata::fs::copy_to_stdout; use rustc_metadata::fs::copy_to_stdout;
use rustc_metadata::EncodedMetadata; use rustc_metadata::EncodedMetadata;
use rustc_middle::dep_graph::{WorkProduct, WorkProductId}; use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
@ -611,13 +612,10 @@ pub(crate) fn run_aot(
CguReuse::PreLto | CguReuse::PostLto => false, CguReuse::PreLto | CguReuse::PostLto => false,
}); });
let mut concurrency_limiter = ConcurrencyLimiter::new(tcx.sess, todo_cgus.len()); let concurrency_limiter = IntoDynSyncSend(ConcurrencyLimiter::new(tcx.sess, todo_cgus.len()));
let modules = let modules = tcx.sess.time("codegen mono items", || {
tcx.sess.time("codegen mono items", || { let mut modules: Vec<_> = par_map(todo_cgus, |(_, cgu)| {
todo_cgus
.into_iter()
.map(|(_, cgu)| {
let dep_node = cgu.codegen_dep_node(tcx); let dep_node = cgu.codegen_dep_node(tcx);
tcx.dep_graph tcx.dep_graph
.with_task( .with_task(
@ -633,11 +631,13 @@ pub(crate) fn run_aot(
Some(rustc_middle::dep_graph::hash_result), Some(rustc_middle::dep_graph::hash_result),
) )
.0 .0
}) });
.chain(done_cgus.into_iter().map(|(_, cgu)| { modules.extend(
OngoingModuleCodegen::Sync(reuse_workproduct_for_cgu(tcx, cgu)) done_cgus
})) .into_iter()
.collect::<Vec<_>>() .map(|(_, cgu)| OngoingModuleCodegen::Sync(reuse_workproduct_for_cgu(tcx, cgu))),
);
modules
}); });
let mut allocator_module = make_module(tcx.sess, &backend_config, "allocator_shim".to_string()); let mut allocator_module = make_module(tcx.sess, &backend_config, "allocator_shim".to_string());
@ -706,6 +706,6 @@ pub(crate) fn run_aot(
metadata_module, metadata_module,
metadata, metadata,
crate_info: CrateInfo::new(tcx, target_cpu), crate_info: CrateInfo::new(tcx, target_cpu),
concurrency_limiter, concurrency_limiter: concurrency_limiter.0,
}) })
} }