Skip to content
Permalink
Browse files

Self-Profiling: Make names of existing events more consistent and use…

… new API.
  • Loading branch information...
michaelwoerister committed Sep 27, 2019
1 parent b0b073c commit d94262272bbdc700689a012a2e8a34adbe2a0f18
@@ -112,7 +112,7 @@ impl<'a, 'tcx, Q: QueryDescription<'tcx>> JobOwner<'a, 'tcx, Q> {
let mut lock = cache.get_shard_by_value(key).lock();
if let Some(value) = lock.results.get(key) {
profq_msg!(tcx, ProfileQueriesMsg::CacheHit);
tcx.sess.profiler(|p| p.record_query_hit(Q::NAME));
tcx.prof.query_cache_hit(Q::NAME);
let result = (value.value.clone(), value.index);
#[cfg(debug_assertions)]
{
@@ -128,7 +128,7 @@ impl<'a, 'tcx, Q: QueryDescription<'tcx>> JobOwner<'a, 'tcx, Q> {
// in another thread has completed. Record how long we wait in the
// self-profiler.
#[cfg(parallel_compiler)]
tcx.sess.profiler(|p| p.query_blocked_start(Q::NAME));
tcx.prof.query_blocked_start(Q::NAME);

job.clone()
},
@@ -170,7 +170,7 @@ impl<'a, 'tcx, Q: QueryDescription<'tcx>> JobOwner<'a, 'tcx, Q> {
#[cfg(parallel_compiler)]
{
let result = job.r#await(tcx, span);
tcx.sess.profiler(|p| p.query_blocked_end(Q::NAME));
tcx.prof.query_blocked_end(Q::NAME);

if let Err(cycle) = result {
return TryGetJob::Cycle(Q::handle_cycle_error(tcx, cycle));
@@ -382,8 +382,9 @@ impl<'tcx> TyCtxt<'tcx> {
}

if Q::ANON {

profq_msg!(self, ProfileQueriesMsg::ProviderBegin);
self.sess.profiler(|p| p.start_query(Q::NAME));
let prof_timer = self.prof.query_provider(Q::NAME);

let ((result, dep_node_index), diagnostics) = with_diagnostics(|diagnostics| {
self.start_query(job.job.clone(), diagnostics, |tcx| {
@@ -393,7 +394,7 @@ impl<'tcx> TyCtxt<'tcx> {
})
});

self.sess.profiler(|p| p.end_query(Q::NAME));
drop(prof_timer);
profq_msg!(self, ProfileQueriesMsg::ProviderEnd);

self.dep_graph.read_index(dep_node_index);
@@ -451,9 +452,8 @@ impl<'tcx> TyCtxt<'tcx> {
// First we try to load the result from the on-disk cache.
let result = if Q::cache_on_disk(self, key.clone(), None) &&
self.sess.opts.debugging_opts.incremental_queries {
self.sess.profiler(|p| p.incremental_load_result_start(Q::NAME));
let _prof_timer = self.prof.incr_cache_loading(Q::NAME);
let result = Q::try_load_from_disk(self, prev_dep_node_index);
self.sess.profiler(|p| p.incremental_load_result_end(Q::NAME));

// We always expect to find a cached result for things that
// can be forced from `DepNode`.
@@ -469,21 +469,17 @@ impl<'tcx> TyCtxt<'tcx> {

let result = if let Some(result) = result {
profq_msg!(self, ProfileQueriesMsg::CacheHit);
self.sess.profiler(|p| p.record_query_hit(Q::NAME));

result
} else {
// We could not load a result from the on-disk cache, so
// recompute.

self.sess.profiler(|p| p.start_query(Q::NAME));
let _prof_timer = self.prof.query_provider(Q::NAME);

// The dep-graph for this computation is already in-place.
let result = self.dep_graph.with_ignore(|| {
Q::compute(self, key)
});

self.sess.profiler(|p| p.end_query(Q::NAME));
result
};

@@ -551,7 +547,7 @@ impl<'tcx> TyCtxt<'tcx> {
key, dep_node);

profq_msg!(self, ProfileQueriesMsg::ProviderBegin);
self.sess.profiler(|p| p.start_query(Q::NAME));
let prof_timer = self.prof.query_provider(Q::NAME);

let ((result, dep_node_index), diagnostics) = with_diagnostics(|diagnostics| {
self.start_query(job.job.clone(), diagnostics, |tcx| {
@@ -571,7 +567,7 @@ impl<'tcx> TyCtxt<'tcx> {
})
});

self.sess.profiler(|p| p.end_query(Q::NAME));
drop(prof_timer);
profq_msg!(self, ProfileQueriesMsg::ProviderEnd);

if unlikely!(self.sess.opts.debugging_opts.query_dep_graph) {
@@ -619,7 +615,7 @@ impl<'tcx> TyCtxt<'tcx> {
let _ = self.get_query::<Q>(DUMMY_SP, key);
} else {
profq_msg!(self, ProfileQueriesMsg::CacheHit);
self.sess.profiler(|p| p.record_query_hit(Q::NAME));
self.prof.query_cache_hit(Q::NAME);
}
}

@@ -62,11 +62,13 @@ fn prepare_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
};
let exported_symbols = cgcx.exported_symbols
.as_ref().expect("needs exported symbols for LTO");
let mut symbol_white_list = exported_symbols[&LOCAL_CRATE]
.iter()
.filter_map(symbol_filter)
.collect::<Vec<CString>>();
let _timer = cgcx.profile_activity("generate_symbol_white_list_for_thinlto");
let mut symbol_white_list = {
let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbol_white_list");
exported_symbols[&LOCAL_CRATE]
.iter()
.filter_map(symbol_filter)
.collect::<Vec<CString>>()
};
info!("{} symbols to preserve in this crate", symbol_white_list.len());

// If we're performing LTO for the entire crate graph, then for each of our
@@ -95,14 +97,17 @@ fn prepare_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
}

for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
let _timer = cgcx.profile_activity(format!("load: {}", path.display()));
let exported_symbols = cgcx.exported_symbols
.as_ref().expect("needs exported symbols for LTO");
symbol_white_list.extend(
exported_symbols[&cnum]
.iter()
.filter_map(symbol_filter));
{
let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbol_white_list");
symbol_white_list.extend(
exported_symbols[&cnum]
.iter()
.filter_map(symbol_filter));
}

let _timer = cgcx.prof.generic_activity("LLVM_lto_load_upstream_bitcode");
let archive = ArchiveRO::open(&path).expect("wanted an rlib");
let bytecodes = archive.iter().filter_map(|child| {
child.ok().and_then(|c| c.name().map(|name| (name, c)))
@@ -189,6 +194,7 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
symbol_white_list: &[*const libc::c_char])
-> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError>
{
let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
info!("going for a fat lto");

// Sort out all our lists of incoming modules into two lists.
@@ -287,6 +293,7 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
// save and persist everything with the original module.
let mut linker = Linker::new(llmod);
for (bc_decoded, name) in serialized_modules {
let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_link_module");
info!("linking {:?}", name);
time_ext(cgcx.time_passes, None, &format!("ll link {:?}", name), || {
let data = bc_decoded.data();
@@ -388,6 +395,7 @@ fn thin_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
symbol_white_list: &[*const libc::c_char])
-> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError>
{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
unsafe {
info!("going for that thin, thin LTO");

@@ -601,16 +609,6 @@ impl ModuleBuffer {
llvm::LLVMRustModuleBufferCreate(m)
})
}

pub fn parse<'a>(
&self,
name: &str,
cx: &'a llvm::Context,
handler: &Handler,
) -> Result<&'a llvm::Module, FatalError> {
let name = CString::new(name).unwrap();
parse_module(cx, &name, self.data(), handler)
}
}

impl ModuleBufferMethods for ModuleBuffer {
@@ -723,7 +721,7 @@ pub unsafe fn optimize_thin_module(
// Like with "fat" LTO, get some better optimizations if landing pads
// are disabled by removing all landing pads.
if cgcx.no_landing_pads {
let _timer = cgcx.profile_activity("LLVM_remove_landing_pads");
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_remove_landing_pads");
llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
save_temp_bitcode(&cgcx, &module, "thin-lto-after-nounwind");
}
@@ -736,26 +734,41 @@ pub unsafe fn optimize_thin_module(
//
// You can find some more comments about these functions in the LLVM
// bindings we've got (currently `PassWrapper.cpp`)
if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))
{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_rename");
if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))

{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_resolve_weak");
if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))

{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_internalize");
if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))

{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_import");
if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod) {
let msg = "failed to prepare thin LTO module";
return Err(write::llvm_err(&diag_handler, msg))
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
}
save_temp_bitcode(cgcx, &module, "thin-lto-after-import");

// Ok now this is a bit unfortunate. This is also something you won't
// find upstream in LLVM's ThinLTO passes! This is a hack for now to
@@ -786,18 +799,24 @@ pub unsafe fn optimize_thin_module(
// not too much) but for now at least gets LLVM to emit valid DWARF (or
// so it appears). Hopefully we can remove this once upstream bugs are
// fixed in LLVM.
llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_patch_debuginfo");
llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
}

// Alright now that we've done everything related to the ThinLTO
// analysis it's time to run some optimizations! Here we use the same
// `run_pass_manager` as the "fat" LTO above except that we tell it to
// populate a thin-specific pass manager, which presumably LLVM treats a
// little differently.
info!("running thin lto passes over {}", module.name);
let config = cgcx.config(module.kind);
run_pass_manager(cgcx, &module, config, true);
save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
{
let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_optimize");
info!("running thin lto passes over {}", module.name);
let config = cgcx.config(module.kind);
run_pass_manager(cgcx, &module, config, true);
save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
}
}
Ok(module)
}
@@ -306,6 +306,8 @@ pub(crate) unsafe fn optimize(cgcx: &CodegenContext<LlvmCodegenBackend>,
config: &ModuleConfig)
-> Result<(), FatalError>
{
let _timer = cgcx.prof.generic_activity("LLVM_module_optimize");

let llmod = module.module_llvm.llmod();
let llcx = &*module.module_llvm.llcx;
let tm = &*module.module_llvm.tm;
@@ -423,7 +425,7 @@ pub(crate) unsafe fn optimize(cgcx: &CodegenContext<LlvmCodegenBackend>,

// Finally, run the actual optimization passes
{
let _timer = cgcx.profile_activity("LLVM_function_passes");
let _timer = cgcx.prof.generic_activity("LLVM_module_optimize_function_passes");
time_ext(config.time_passes,
None,
&format!("llvm function passes [{}]", module_name.unwrap()),
@@ -432,7 +434,7 @@ pub(crate) unsafe fn optimize(cgcx: &CodegenContext<LlvmCodegenBackend>,
});
}
{
let _timer = cgcx.profile_activity("LLVM_module_passes");
let _timer = cgcx.prof.generic_activity("LLVM_module_optimize_module_passes");
time_ext(config.time_passes,
None,
&format!("llvm module passes [{}]", module_name.unwrap()),
@@ -454,7 +456,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<LlvmCodegenBackend>,
config: &ModuleConfig)
-> Result<CompiledModule, FatalError>
{
let _timer = cgcx.profile_activity("codegen");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen");
{
let llmod = module.module_llvm.llmod();
let llcx = &*module.module_llvm.llcx;
@@ -505,25 +507,26 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<LlvmCodegenBackend>,


if write_bc || config.emit_bc_compressed || config.embed_bitcode {
let _timer = cgcx.profile_activity("LLVM_make_bitcode");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_make_bitcode");
let thin = ThinBuffer::new(llmod);
let data = thin.data();

if write_bc {
let _timer = cgcx.profile_activity("LLVM_emit_bitcode");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_emit_bitcode");
if let Err(e) = fs::write(&bc_out, data) {
let msg = format!("failed to write bytecode to {}: {}", bc_out.display(), e);
diag_handler.err(&msg);
}
}

if config.embed_bitcode {
let _timer = cgcx.profile_activity("LLVM_embed_bitcode");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_embed_bitcode");
embed_bitcode(cgcx, llcx, llmod, Some(data));
}

if config.emit_bc_compressed {
let _timer = cgcx.profile_activity("LLVM_compress_bitcode");
let _timer =
cgcx.prof.generic_activity("LLVM_module_codegen_emit_compressed_bitcode");
let dst = bc_out.with_extension(RLIB_BYTECODE_EXTENSION);
let data = bytecode::encode(&module.name, data);
if let Err(e) = fs::write(&dst, data) {
@@ -538,7 +541,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<LlvmCodegenBackend>,
time_ext(config.time_passes, None, &format!("codegen passes [{}]", module_name.unwrap()),
|| -> Result<(), FatalError> {
if config.emit_ir {
let _timer = cgcx.profile_activity("LLVM_emit_ir");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_emit_ir");
let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
let out_c = path_to_c_string(&out);

@@ -585,7 +588,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<LlvmCodegenBackend>,
}

if config.emit_asm || asm_to_obj {
let _timer = cgcx.profile_activity("LLVM_emit_asm");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_emit_asm");
let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);

// We can't use the same module for asm and binary output, because that triggers
@@ -603,13 +606,13 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<LlvmCodegenBackend>,
}

if write_obj {
let _timer = cgcx.profile_activity("LLVM_emit_obj");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_emit_obj");
with_codegen(tm, llmod, config.no_builtins, |cpm| {
write_output_file(diag_handler, tm, cpm, llmod, &obj_out,
llvm::FileType::ObjectFile)
})?;
} else if asm_to_obj {
let _timer = cgcx.profile_activity("LLVM_asm_to_obj");
let _timer = cgcx.prof.generic_activity("LLVM_module_codegen_asm_to_obj");
let assembly = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
run_assembler(cgcx, diag_handler, &assembly, &obj_out);

0 comments on commit d942622

Please sign in to comment.
You can’t perform that action at this time.