bjoernager/rust
bjoernager/rust
1
Fork 0
Code Activity

Auto merge of #112012 - Kobzol:try-build-llvm-rebuild, r=nikic

Browse source 
Avoid one `rustc` rebuild in the optimized build pipeline

This PR changes the optimized build pipeline to avoid one `rustc` rebuild, inspired by [this comment](https://github.com/rust-lang/rust/issues/112011#issuecomment-1564991175). This speeds up the pipeline by 5-10 minutes. After this change, we **no longer gather LLVM PGO profiles from compiling stage 2 of `rustc`**.

Now we build `rustc` two times (1x PGO instrumented, 1x PGO optimized) and LLVM three times (1x normal, 1x PGO instrumented, 1x PGO optimized). It should be possible to cache the normal LLVM build, but I'll leave that for another PR.
This commit is contained in:
bors 2023-06-08 00:46:58 +00:00
commit 50f2176721
2 changed files with 100 additions and 56 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/bootstrap/compile.rs
View file

@ -118,6 +118,10 @@ impl Step for Std {
|| builder.config.keep_stage_std.contains(&compiler.stage) || builder.config.keep_stage_std.contains(&compiler.stage)
{ {
builder.info("Warning: Using a potentially old libstd. This may not behave well."); builder.info("Warning: Using a potentially old libstd. This may not behave well.");
copy_third_party_objects(builder, &compiler, target);
copy_self_contained_objects(builder, &compiler, target);
builder.ensure(StdLink::from_std(self, compiler)); builder.ensure(StdLink::from_std(self, compiler));
return; return;
} }

152
src/ci/stage-build.py
View file

@ -620,11 +620,17 @@ def get_files(directory: Path, filter: Optional[Callable[[Path], bool]] = None)
yield path yield path
def build_rustc( def bootstrap_build(
pipeline: Pipeline, pipeline: Pipeline,
args: List[str], args: List[str],
env: Optional[Dict[str, str]] = None env: Optional[Dict[str, str]] = None,
targets: Iterable[str] = ("library/std", )
): ):
if env is None:
env = {}
else:
env = dict(env)
env["RUST_BACKTRACE"] = "1"
arguments = [ arguments = [
sys.executable, sys.executable,
pipeline.checkout_path() / "x.py", pipeline.checkout_path() / "x.py",
@ -632,8 +638,7 @@ def build_rustc(
"--target", PGO_HOST, "--target", PGO_HOST,
"--host", PGO_HOST, "--host", PGO_HOST,
"--stage", "2", "--stage", "2",
"library/std" ] + list(targets) + args
] + args
cmd(arguments, env=env) cmd(arguments, env=env)
@ -776,18 +781,18 @@ def record_metrics(pipeline: Pipeline, timer: Timer):
if metrics is None: if metrics is None:
return return
llvm_steps = tuple(metrics.find_all_by_type("bootstrap::llvm::Llvm")) llvm_steps = tuple(metrics.find_all_by_type("bootstrap::llvm::Llvm"))
assert len(llvm_steps) > 0
llvm_duration = sum(step.duration for step in llvm_steps) llvm_duration = sum(step.duration for step in llvm_steps)
rustc_steps = tuple(metrics.find_all_by_type("bootstrap::compile::Rustc")) rustc_steps = tuple(metrics.find_all_by_type("bootstrap::compile::Rustc"))
assert len(rustc_steps) > 0
rustc_duration = sum(step.duration for step in rustc_steps) rustc_duration = sum(step.duration for step in rustc_steps)
# The LLVM step is part of the Rustc step # The LLVM step is part of the Rustc step
rustc_duration -= llvm_duration rustc_duration = max(0, rustc_duration - llvm_duration)
timer.add_duration("LLVM", llvm_duration) if llvm_duration > 0:
timer.add_duration("Rustc", rustc_duration) timer.add_duration("LLVM", llvm_duration)
if rustc_duration > 0:
timer.add_duration("Rustc", rustc_duration)
log_metrics(metrics) log_metrics(metrics)
@ -872,79 +877,114 @@ download-ci-llvm = true
)) ))
def execute_build_pipeline(timer: Timer, pipeline: Pipeline, runner: BenchmarkRunner, final_build_args: List[str]): def execute_build_pipeline(timer: Timer, pipeline: Pipeline, runner: BenchmarkRunner, dist_build_args: List[str]):
# Clear and prepare tmp directory # Clear and prepare tmp directory
shutil.rmtree(pipeline.opt_artifacts(), ignore_errors=True) shutil.rmtree(pipeline.opt_artifacts(), ignore_errors=True)
os.makedirs(pipeline.opt_artifacts(), exist_ok=True) os.makedirs(pipeline.opt_artifacts(), exist_ok=True)
pipeline.build_rustc_perf() pipeline.build_rustc_perf()
# Stage 1: Build rustc + PGO instrumented LLVM """
with timer.section("Stage 1 (LLVM PGO)") as stage1: Stage 1: Build PGO instrumented rustc
with stage1.section("Build rustc and LLVM") as rustc_build:
build_rustc(pipeline, args=[
"--llvm-profile-generate"
], env=dict(
LLVM_PROFILE_DIR=str(pipeline.llvm_profile_dir_root() / "prof-%p")
))
record_metrics(pipeline, rustc_build)
with stage1.section("Gather profiles"): We use a normal build of LLVM, because gathering PGO profiles for LLVM and `rustc` at the same time
gather_llvm_profiles(pipeline, runner) can cause issues.
print_free_disk_space(pipeline) """
with timer.section("Stage 1 (rustc PGO)") as stage1:
clear_llvm_files(pipeline) with stage1.section("Build PGO instrumented rustc and LLVM") as rustc_pgo_instrument:
final_build_args += [ bootstrap_build(pipeline, args=[
"--llvm-profile-use",
pipeline.llvm_profile_merged_file()
]
# Stage 2: Build PGO instrumented rustc + LLVM
with timer.section("Stage 2 (rustc PGO)") as stage2:
with stage2.section("Build rustc and LLVM") as rustc_build:
build_rustc(pipeline, args=[
"--rust-profile-generate", "--rust-profile-generate",
pipeline.rustc_profile_dir_root() pipeline.rustc_profile_dir_root()
]) ])
record_metrics(pipeline, rustc_build) record_metrics(pipeline, rustc_pgo_instrument)
with stage2.section("Gather profiles"): with stage1.section("Gather profiles"):
gather_rustc_profiles(pipeline, runner) gather_rustc_profiles(pipeline, runner)
print_free_disk_space(pipeline) print_free_disk_space(pipeline)
clear_llvm_files(pipeline) with stage1.section("Build PGO optimized rustc") as rustc_pgo_use:
final_build_args += [ bootstrap_build(pipeline, args=[
"--rust-profile-use", "--rust-profile-use",
pipeline.rustc_profile_merged_file() pipeline.rustc_profile_merged_file()
] ])
record_metrics(pipeline, rustc_pgo_use)
dist_build_args += [
"--rust-profile-use",
pipeline.rustc_profile_merged_file()
]
# Stage 3: Build rustc + BOLT instrumented LLVM """
Stage 2: Gather LLVM PGO profiles
"""
with timer.section("Stage 2 (LLVM PGO)") as stage2:
# Clear normal LLVM artifacts
clear_llvm_files(pipeline)
with stage2.section("Build PGO instrumented LLVM") as llvm_pgo_instrument:
bootstrap_build(pipeline, args=[
"--llvm-profile-generate",
# We want to keep the already built PGO-optimized `rustc`.
"--keep-stage", "0",
"--keep-stage", "1"
], env=dict(
LLVM_PROFILE_DIR=str(pipeline.llvm_profile_dir_root() / "prof-%p")
))
record_metrics(pipeline, llvm_pgo_instrument)
with stage2.section("Gather profiles"):
gather_llvm_profiles(pipeline, runner)
dist_build_args += [
"--llvm-profile-use",
pipeline.llvm_profile_merged_file(),
]
print_free_disk_space(pipeline)
# Clear PGO-instrumented LLVM artifacts
clear_llvm_files(pipeline)
"""
Stage 3: Build BOLT instrumented LLVM
We build a PGO optimized LLVM in this step, then instrument it with BOLT and gather BOLT profiles.
Note that we don't remove LLVM artifacts after this step, so that they are reused in the final dist build.
BOLT instrumentation is performed "on-the-fly" when the LLVM library is copied to the sysroot of rustc,
therefore the LLVM artifacts on disk are not "tainted" with BOLT instrumentation and they can be reused.
"""
if pipeline.supports_bolt(): if pipeline.supports_bolt():
with timer.section("Stage 3 (LLVM BOLT)") as stage3: with timer.section("Stage 3 (LLVM BOLT)") as stage3:
with stage3.section("Build rustc and LLVM") as rustc_build: with stage3.section("Build BOLT instrumented LLVM") as llvm_bolt_instrument:
build_rustc(pipeline, args=[ bootstrap_build(pipeline, args=[
"--llvm-profile-use", "--llvm-profile-use",
pipeline.llvm_profile_merged_file(), pipeline.llvm_profile_merged_file(),
"--llvm-bolt-profile-generate", "--llvm-bolt-profile-generate",
"--rust-profile-use", # We want to keep the already built PGO-optimized `rustc`.
pipeline.rustc_profile_merged_file() "--keep-stage", "0",
"--keep-stage", "1"
]) ])
record_metrics(pipeline, rustc_build) record_metrics(pipeline, llvm_bolt_instrument)
with stage3.section("Gather profiles"): with stage3.section("Gather profiles"):
gather_llvm_bolt_profiles(pipeline, runner) gather_llvm_bolt_profiles(pipeline, runner)
# LLVM is not being cleared here, we want to reuse the previous build dist_build_args += [
print_free_disk_space(pipeline) "--llvm-bolt-profile-use",
final_build_args += [ pipeline.llvm_bolt_profile_merged_file()
"--llvm-bolt-profile-use", ]
pipeline.llvm_bolt_profile_merged_file() print_free_disk_space(pipeline)
]
# Stage 4: Build PGO optimized rustc + PGO/BOLT optimized LLVM # We want to keep the already built PGO-optimized `rustc`.
with timer.section("Stage 4 (final build)") as stage4: dist_build_args += [
cmd(final_build_args) "--keep-stage", "0",
record_metrics(pipeline, stage4) "--keep-stage", "1"
]
"""
Final stage: Build PGO optimized rustc + PGO/BOLT optimized LLVM
"""
with timer.section("Final stage (dist build)") as final_stage:
cmd(dist_build_args)
record_metrics(pipeline, final_stage)
# Try builds can be in various broken states, so we don't want to gatekeep them with tests # Try builds can be in various broken states, so we don't want to gatekeep them with tests
if not is_try_build(): if not is_try_build():