mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
Alexey Milovidov
804666f02d
`Build (BOLT)` failed on master because the BOLT-pass build, which
runs with `-fprofile-use=clickhouse-pgo.profdata`, emits
`-Wbackend-plugin` warnings of the form
function control flow change detected (hash mismatch) <symbol>
Hash = ... up to N count discarded
for functions whose CFG hash at use time does not match what was
recorded during the instrumented run, and the project-wide `-Werror`
makes them fatal.
https://github.com/ClickHouse/ClickHouse/actions/runs/26198563312/job/77083370440
Job report:
https://s3.amazonaws.com/clickhouse-test-reports/json.html?REF=master&sha=8d1d74ce0bb6e52ff81e58af1b73a86e0b1f32be&name_0=OptimizeClickHouse&name_1=Collect%20ClickHouse%20Profiles%20%28PGO%2C%20BOLT%29%20%28aarch64%29
Affected files in the failure: `src/Common/ConcurrencyControl.cpp`,
`src/Common/Dwarf.cpp`, `src/Common/Exception.cpp` (the same kind of
warning also appears for `base/glibc-compatibility/musl/exp.c`,
`log.c`, `pow.c`, etc., but those are not under `-Werror` and so
did not fail the build).
The instrumented build is required to disable ThinLTO (clang IR
instrumentation is incompatible with it; the cmake check enforces it),
while the BOLT build enables ThinLTO. ThinLTO can perturb individual
function bodies before profile matching, so a non-zero number of hash
mismatches is expected. The compiler simply drops the stale counts
for those functions and the rest of the profile is still applied —
which is the right behaviour for a best-effort optimisation pass.
Demote the warning to a plain warning specifically for the PGO-use
build by adding `-Wno-error=backend-plugin` alongside the
`-fprofile-use=...` flags, so mismatches stay visible in build logs
without breaking the build.
74 lines
3.7 KiB
CMake
74 lines
3.7 KiB
CMake
# Support for Profile-Guided Optimization (PGO) and Binary Optimization and Layout Tool (BOLT)
|
|
# for the ClickHouse binary itself.
|
|
#
|
|
# PGO workflow:
|
|
# 1. Build with -DENABLE_CLICKHOUSE_PGO_GENERATE=ON (instrumented build)
|
|
# 2. Run representative workloads (performance tests)
|
|
# 3. Merge profiles: llvm-profdata merge -output=clickhouse.profdata *.profraw
|
|
# 4. Build with -DCLICKHOUSE_PGO_PROFILE_PATH=/path/to/clickhouse.profdata
|
|
#
|
|
# BOLT workflow (applied after linking):
|
|
# 1. Build with -DENABLE_CLICKHOUSE_BOLT=ON (adds --emit-relocs to linker flags)
|
|
# 2. Instrument: llvm-bolt clickhouse -instrument -o clickhouse.inst
|
|
# 3. Run workloads with instrumented binary
|
|
# 4. Merge: merge-fdata *.fdata > bolt.fdata
|
|
# 5. Optimize: llvm-bolt clickhouse -o clickhouse.bolt -data=bolt.fdata ...
|
|
|
|
# --- PGO: Profile Generation (instrumented build) ---
|
|
option(ENABLE_CLICKHOUSE_PGO_GENERATE
|
|
"Build ClickHouse with PGO instrumentation (-fprofile-generate). Incompatible with ThinLTO." OFF)
|
|
|
|
# --- PGO: Profile Use (optimized build) ---
|
|
set(CLICKHOUSE_PGO_PROFILE_PATH "" CACHE STRING
|
|
"Path to merged .profdata file for PGO-optimized build (-fprofile-use)")
|
|
|
|
# --- BOLT: Emit relocations for post-link optimization ---
|
|
option(ENABLE_CLICKHOUSE_BOLT
|
|
"Add --emit-relocs to linker flags so BOLT can post-process the binary" OFF)
|
|
|
|
# Validation
|
|
if (ENABLE_CLICKHOUSE_PGO_GENERATE AND CLICKHOUSE_PGO_PROFILE_PATH)
|
|
message(FATAL_ERROR "ENABLE_CLICKHOUSE_PGO_GENERATE and CLICKHOUSE_PGO_PROFILE_PATH are mutually exclusive")
|
|
endif()
|
|
|
|
if (ENABLE_CLICKHOUSE_PGO_GENERATE AND ENABLE_THINLTO)
|
|
message(FATAL_ERROR "ENABLE_CLICKHOUSE_PGO_GENERATE requires ENABLE_THINLTO=OFF (IR instrumentation is incompatible with ThinLTO)")
|
|
endif()
|
|
|
|
# Apply PGO generation flags
|
|
if (ENABLE_CLICKHOUSE_PGO_GENERATE)
|
|
message(STATUS "ClickHouse PGO: instrumented build enabled (-fprofile-generate)")
|
|
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fprofile-generate")
|
|
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-generate")
|
|
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fprofile-generate")
|
|
endif()
|
|
|
|
# Apply PGO use flags
|
|
if (CLICKHOUSE_PGO_PROFILE_PATH)
|
|
if (NOT EXISTS "${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
message(FATAL_ERROR "PGO profile not found: ${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
endif()
|
|
message(STATUS "ClickHouse PGO: using profile ${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fprofile-use=${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-use=${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fprofile-use=${CLICKHOUSE_PGO_PROFILE_PATH}")
|
|
# When a function's CFG hash at use time does not match what was recorded
|
|
# in the profile, clang's PGO backend emits `-Wbackend-plugin` with text
|
|
# like "function control flow change detected (hash mismatch)" and
|
|
# discards the stale counts for that function. This is expected to
|
|
# happen for some functions because the instrumented build runs without
|
|
# ThinLTO (required by IR instrumentation) while the use build enables
|
|
# ThinLTO, which can perturb individual function bodies before profile
|
|
# matching. Under the project-wide `-Werror` policy these informational
|
|
# warnings would otherwise abort the whole build, so demote them to
|
|
# plain warnings here.
|
|
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-error=backend-plugin")
|
|
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=backend-plugin")
|
|
endif()
|
|
|
|
# Apply BOLT linker flags
|
|
if (ENABLE_CLICKHOUSE_BOLT)
|
|
message(STATUS "ClickHouse BOLT: adding --emit-relocs for post-link optimization")
|
|
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--emit-relocs")
|
|
endif()
|