[clang] [Clang][CMake] Support perf, LBR, and Instrument CLANG_BOLT options (PR #69133)
Amir Ayupov via cfe-commits
cfe-commits at lists.llvm.org
Tue Nov 7 20:42:07 PST 2023
https://github.com/aaupov updated https://github.com/llvm/llvm-project/pull/69133
>From 1ca9ce9ef3e90e3d24aea3945ae2d5129682621d Mon Sep 17 00:00:00 2001
From: Amir Aupov <amir.aupov at gmail.com>
Date: Mon, 16 Oct 2023 01:08:28 +0200
Subject: [PATCH 1/3] [Clang][CMake] Support perf, LBR, and Instrument
CLANG_BOLT options
Split up and refactor CLANG_BOLT_INSTRUMENT into support for
perf no-LBR and perf with LBR profiling modes.
Differential Revision: https://reviews.llvm.org/D143617
---
clang/CMakeLists.txt | 44 ++++++++-----
clang/cmake/caches/BOLT.cmake | 2 +-
clang/utils/perf-training/CMakeLists.txt | 29 ++++++++-
clang/utils/perf-training/bolt.lit.cfg | 53 +++++++++++++---
.../utils/perf-training/bolt.lit.site.cfg.in | 2 +
clang/utils/perf-training/perf-helper.py | 63 +++++++++++++++++++
6 files changed, 167 insertions(+), 26 deletions(-)
diff --git a/clang/CMakeLists.txt b/clang/CMakeLists.txt
index 9b52c58be41e7f7..8f64d95cc394ffe 100644
--- a/clang/CMakeLists.txt
+++ b/clang/CMakeLists.txt
@@ -850,23 +850,38 @@ if (CLANG_ENABLE_BOOTSTRAP)
endforeach()
endif()
-if (CLANG_BOLT_INSTRUMENT AND NOT LLVM_BUILD_INSTRUMENTED)
+set(CLANG_BOLT "INSTRUMENT" CACHE STRING "Apply BOLT optimization to Clang. \
+ May be specified as Instrument or Perf or LBR to use a particular profiling \
+ mechanism.")
+string(TOUPPER "${CLANG_BOLT}" uppercase_CLANG_BOLT)
+
+if (CLANG_BOLT AND NOT LLVM_BUILD_INSTRUMENTED)
set(CLANG_PATH ${LLVM_RUNTIME_OUTPUT_INTDIR}/clang)
- set(CLANG_INSTRUMENTED ${CLANG_PATH}-bolt.inst)
+ set(CLANG_INSTRUMENTED ${LLVM_RUNTIME_OUTPUT_INTDIR}/${CLANG_BOLT_INSTRUMENTED})
set(BOLT_FDATA ${CMAKE_CURRENT_BINARY_DIR}/utils/perf-training/prof.fdata)
- # Instrument clang with BOLT
- add_custom_target(clang-instrumented
- DEPENDS ${CLANG_INSTRUMENTED}
- )
- add_custom_command(OUTPUT ${CLANG_INSTRUMENTED}
- DEPENDS clang llvm-bolt
- COMMAND llvm-bolt ${CLANG_PATH} -o ${CLANG_INSTRUMENTED}
- -instrument --instrumentation-file-append-pid
- --instrumentation-file=${BOLT_FDATA}
- COMMENT "Instrumenting clang binary with BOLT"
- VERBATIM
- )
+ # Pass extra flag in no-LBR mode
+ if (uppercase_CLANG_BOLT STREQUAL "PERF")
+ set(BOLT_NO_LBR "-nl")
+ endif()
+
+ if (uppercase_CLANG_BOLT STREQUAL "INSTRUMENT")
+ # Instrument clang with BOLT
+ add_custom_target(clang-instrumented
+ DEPENDS ${CLANG_INSTRUMENTED}
+ )
+ add_custom_command(OUTPUT ${CLANG_INSTRUMENTED}
+ DEPENDS clang llvm-bolt
+ COMMAND llvm-bolt ${CLANG_PATH} -o ${CLANG_INSTRUMENTED}
+ -instrument --instrumentation-file-append-pid
+ --instrumentation-file=${BOLT_FDATA}
+ COMMENT "Instrumenting clang binary with BOLT"
+ VERBATIM
+ )
+ add_custom_target(clang-bolt-training-deps DEPENDS clang-instrumented)
+ else() # perf or LBR
+ add_custom_target(clang-bolt-training-deps DEPENDS clang)
+ endif()
# Optimize original (pre-bolt) Clang using the collected profile
set(CLANG_OPTIMIZED ${CMAKE_CURRENT_BINARY_DIR}/clang.bolt)
@@ -880,6 +895,7 @@ if (CLANG_BOLT_INSTRUMENT AND NOT LLVM_BUILD_INSTRUMENTED)
-data ${BOLT_FDATA}
-reorder-blocks=ext-tsp -reorder-functions=hfsort+ -split-functions
-split-all-cold -split-eh -dyno-stats -icf=1 -use-gnu-stack
+ ${BOLT_NO_LBR}
COMMAND ${CMAKE_COMMAND} -E rename ${CLANG_OPTIMIZED} $<TARGET_FILE:clang>
COMMENT "Optimizing Clang with BOLT"
VERBATIM
diff --git a/clang/cmake/caches/BOLT.cmake b/clang/cmake/caches/BOLT.cmake
index 0442f73e5426ac7..eba2346b2f4ca12 100644
--- a/clang/cmake/caches/BOLT.cmake
+++ b/clang/cmake/caches/BOLT.cmake
@@ -1,5 +1,5 @@
set(CMAKE_BUILD_TYPE Release CACHE STRING "")
-set(CLANG_BOLT_INSTRUMENT ON CACHE BOOL "")
+set(CLANG_BOLT "INSTRUMENT" CACHE STRING "")
set(CMAKE_EXE_LINKER_FLAGS "-Wl,--emit-relocs,-znow" CACHE STRING "")
set(LLVM_ENABLE_PROJECTS "bolt;clang" CACHE STRING "")
diff --git a/clang/utils/perf-training/CMakeLists.txt b/clang/utils/perf-training/CMakeLists.txt
index c6d51863fb1b5c2..48fbee62a8636d1 100644
--- a/clang/utils/perf-training/CMakeLists.txt
+++ b/clang/utils/perf-training/CMakeLists.txt
@@ -62,7 +62,9 @@ if(APPLE AND DTRACE AND NOT LLVM_TOOL_LLVM_DRIVER_BUILD)
DEPENDS generate-dtrace-logs)
endif()
-if(CLANG_BOLT_INSTRUMENT AND NOT LLVM_BUILD_INSTRUMENTED)
+if(CLANG_BOLT AND NOT LLVM_BUILD_INSTRUMENTED)
+ set(CLANG_BOLT_INSTRUMENTED "clang-bolt.inst" CACHE STRING
+ "Name of BOLT-instrumented Clang binary")
configure_lit_site_cfg(
${CMAKE_CURRENT_SOURCE_DIR}/bolt.lit.site.cfg.in
${CMAKE_CURRENT_BINARY_DIR}/bolt-fdata/lit.site.cfg
@@ -71,16 +73,37 @@ if(CLANG_BOLT_INSTRUMENT AND NOT LLVM_BUILD_INSTRUMENTED)
add_lit_testsuite(generate-bolt-fdata "Generating BOLT profile for Clang"
${CMAKE_CURRENT_BINARY_DIR}/bolt-fdata/
EXCLUDE_FROM_CHECK_ALL
- DEPENDS clang-instrumented clear-bolt-fdata
+ DEPENDS clang-bolt-training-deps clear-bolt-fdata clear-perf-data
)
add_custom_target(clear-bolt-fdata
COMMAND "${Python3_EXECUTABLE}" ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py clean ${CMAKE_CURRENT_BINARY_DIR} fdata
COMMENT "Clearing old BOLT fdata")
+ add_custom_target(clear-perf-data
+ COMMAND "${Python3_EXECUTABLE}" ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py clean ${CMAKE_CURRENT_BINARY_DIR} perf.data
+ COMMENT "Clearing old perf data")
+
+ string(TOUPPER "${CLANG_BOLT}" uppercase_CLANG_BOLT)
+ if (uppercase_CLANG_BOLT STREQUAL "LBR")
+ set(BOLT_LBR "--lbr")
+ endif()
+
+ add_custom_target(merge-fdata-deps)
+ if (uppercase_CLANG_BOLT STREQUAL "INSTRUMENT")
+ add_dependencies(merge-fdata-deps generate-bolt-fdata)
+ else()
+ # Convert perf profiles into fdata
+ add_custom_target(convert-perf-fdata
+ COMMAND "${Python3_EXECUTABLE}" ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py perf2bolt $<TARGET_FILE:llvm-bolt> ${CMAKE_CURRENT_BINARY_DIR} $<TARGET_FILE:clang> ${BOLT_LBR}
+ COMMENT "Converting perf files to BOLT fdata"
+ DEPENDS llvm-bolt generate-bolt-fdata)
+ add_dependencies(merge-fdata-deps convert-perf-fdata)
+ endif()
+
# Merge profiles into one using merge-fdata
add_custom_target(clang-bolt-profile
COMMAND "${Python3_EXECUTABLE}" ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py merge-fdata $<TARGET_FILE:merge-fdata> ${CMAKE_CURRENT_BINARY_DIR}/prof.fdata ${CMAKE_CURRENT_BINARY_DIR}
COMMENT "Merging BOLT fdata"
- DEPENDS merge-fdata generate-bolt-fdata)
+ DEPENDS merge-fdata merge-fdata-deps)
endif()
diff --git a/clang/utils/perf-training/bolt.lit.cfg b/clang/utils/perf-training/bolt.lit.cfg
index 234ac855bd67c65..d2b6042a1627e19 100644
--- a/clang/utils/perf-training/bolt.lit.cfg
+++ b/clang/utils/perf-training/bolt.lit.cfg
@@ -6,15 +6,52 @@ import lit.util
import os
import subprocess
-config.clang = os.path.realpath(lit.util.which('clang-bolt.inst', config.clang_tools_dir)).replace('\\', '/')
+clang_binary = "clang"
+perf_wrapper = ""
+if config.clang_bolt_mode.lower() == "instrument":
+ clang_binary = config.clang_bolt_name
+else: # perf or LBR
+ perf_wrapper = "%s %s/perf-helper.py perf" % (
+ config.python_exe,
+ config.perf_helper_dir,
+ )
+ if config.clang_bolt_mode.lower() == "lbr":
+ perf_wrapper += " --lbr"
+ perf_wrapper += " -- "
-config.name = 'Clang Perf Training'
-config.suffixes = ['.c', '.cc', '.cpp', '.m', '.mm', '.cu', '.ll', '.cl', '.s', '.S', '.modulemap', '.test']
+config.clang = os.path.realpath(
+ lit.util.which(clang_binary, config.clang_tools_dir)
+).replace("\\", "/")
+
+config.name = "Clang Perf Training"
+config.suffixes = [
+ ".c",
+ ".cc",
+ ".cpp",
+ ".m",
+ ".mm",
+ ".cu",
+ ".ll",
+ ".cl",
+ ".s",
+ ".S",
+ ".modulemap",
+ ".test",
+]
use_lit_shell = os.environ.get("LIT_USE_INTERNAL_SHELL")
config.test_format = lit.formats.ShTest(use_lit_shell == "0")
-config.substitutions.append( ('%clang_cpp_skip_driver', ' %s --driver-mode=g++ ' % (config.clang)))
-config.substitutions.append( ('%clang_cpp', ' %s --driver-mode=g++ ' % (config.clang)))
-config.substitutions.append( ('%clang_skip_driver', ' %s ' % (config.clang)))
-config.substitutions.append( ('%clang', ' %s ' % (config.clang) ) )
-config.substitutions.append( ('%test_root', config.test_exec_root ) )
+config.substitutions.append(
+ (
+ "%clang_cpp_skip_driver",
+ " %s %s --driver-mode=g++ " % (perf_wrapper, config.clang),
+ )
+)
+config.substitutions.append(
+ ("%clang_cpp", " %s %s --driver-mode=g++ " % (perf_wrapper, config.clang))
+)
+config.substitutions.append(
+ ("%clang_skip_driver", " %s %s " % (perf_wrapper, config.clang))
+)
+config.substitutions.append(("%clang", " %s %s " % (perf_wrapper, config.clang)))
+config.substitutions.append(("%test_root", config.test_exec_root))
diff --git a/clang/utils/perf-training/bolt.lit.site.cfg.in b/clang/utils/perf-training/bolt.lit.site.cfg.in
index 3029319673fc26c..54de12701c1ae91 100644
--- a/clang/utils/perf-training/bolt.lit.site.cfg.in
+++ b/clang/utils/perf-training/bolt.lit.site.cfg.in
@@ -9,6 +9,8 @@ config.test_source_root = "@CLANG_PGO_TRAINING_DATA@"
config.target_triple = "@LLVM_TARGET_TRIPLE@"
config.python_exe = "@Python3_EXECUTABLE@"
config.clang_obj_root = path(r"@CLANG_BINARY_DIR@")
+config.clang_bolt_mode = "@CLANG_BOLT@"
+config.clang_bolt_name = "@CLANG_BOLT_INSTRUMENTED@"
# Let the main config do the real work.
lit_config.load_config(config, "@CLANG_SOURCE_DIR@/utils/perf-training/bolt.lit.cfg")
diff --git a/clang/utils/perf-training/perf-helper.py b/clang/utils/perf-training/perf-helper.py
index 99d6a3333b6ef08..647bf09ffd24b51 100644
--- a/clang/utils/perf-training/perf-helper.py
+++ b/clang/utils/perf-training/perf-helper.py
@@ -67,6 +67,67 @@ def merge_fdata(args):
return 0
+def perf(args):
+ parser = argparse.ArgumentParser(
+ prog="perf-helper perf", description="perf wrapper for BOLT profile collection"
+ )
+ parser.add_argument(
+ "--lbr", action="store_true", help="Use perf with branch stacks"
+ )
+ parser.add_argument("cmd", nargs="*", help="")
+
+ # Use python's arg parser to handle all leading option arguments, but pass
+ # everything else through to perf
+ first_cmd = next(arg for arg in args if not arg.startswith("--"))
+ last_arg_idx = args.index(first_cmd)
+
+ opts = parser.parse_args(args[:last_arg_idx])
+ cmd = args[last_arg_idx:]
+
+ perf_args = [
+ "perf",
+ "record",
+ "--event=cycles:u",
+ "--freq=max",
+ "--output=%d.perf.data" % os.getpid(),
+ ]
+ if opts.lbr:
+ perf_args += ["--branch-filter=any,u"]
+ perf_args.extend(cmd)
+
+ start_time = time.time()
+ subprocess.check_call(perf_args)
+
+ elapsed = time.time() - start_time
+ print("... data collection took %.4fs" % elapsed)
+ return 0
+
+
+def perf2bolt(args):
+ parser = argparse.ArgumentParser(
+ prog="perf-helper perf2bolt",
+ description="perf2bolt conversion wrapper for perf.data files",
+ )
+ parser.add_argument("bolt", help="Path to llvm-bolt")
+ parser.add_argument("path", help="Path containing perf.data files")
+ parser.add_argument("binary", help="Input binary")
+ parser.add_argument("--lbr", action="store_true", help="Use LBR perf2bolt mode")
+ opts = parser.parse_args(args)
+
+ p2b_args = [
+ opts.bolt,
+ opts.binary,
+ "--aggregate-only",
+ "--profile-format=yaml",
+ ]
+ if not opts.lbr:
+ p2b_args += ["-nl"]
+ p2b_args += ["-p"]
+ for filename in findFilesWithExtension(opts.path, "perf.data"):
+ subprocess.check_call(p2b_args + [filename, "-o", filename + ".fdata"])
+ return 0
+
+
def dtrace(args):
parser = argparse.ArgumentParser(
prog="perf-helper dtrace",
@@ -507,6 +568,8 @@ def genOrderFile(args):
"cc1": cc1,
"gen-order-file": genOrderFile,
"merge-fdata": merge_fdata,
+ "perf": perf,
+ "perf2bolt": perf2bolt,
}
>From 4949c9bd950f309910f84a3affab7c9ab109fea8 Mon Sep 17 00:00:00 2001
From: Amir Ayupov <aaupov at fb.com>
Date: Tue, 7 Nov 2023 09:21:20 -0800
Subject: [PATCH 2/3] Address review comments
---
clang/CMakeLists.txt | 6 ++--
clang/utils/perf-training/CMakeLists.txt | 6 ++--
clang/utils/perf-training/bolt.lit.cfg | 36 ++++++++++--------------
clang/utils/perf-training/perf-helper.py | 11 ++------
4 files changed, 24 insertions(+), 35 deletions(-)
diff --git a/clang/CMakeLists.txt b/clang/CMakeLists.txt
index 8f64d95cc394ffe..052171d29f9973d 100644
--- a/clang/CMakeLists.txt
+++ b/clang/CMakeLists.txt
@@ -853,7 +853,7 @@ endif()
set(CLANG_BOLT "INSTRUMENT" CACHE STRING "Apply BOLT optimization to Clang. \
May be specified as Instrument or Perf or LBR to use a particular profiling \
mechanism.")
-string(TOUPPER "${CLANG_BOLT}" uppercase_CLANG_BOLT)
+string(TOUPPER "${CLANG_BOLT}" CLANG_BOLT)
if (CLANG_BOLT AND NOT LLVM_BUILD_INSTRUMENTED)
set(CLANG_PATH ${LLVM_RUNTIME_OUTPUT_INTDIR}/clang)
@@ -861,11 +861,11 @@ if (CLANG_BOLT AND NOT LLVM_BUILD_INSTRUMENTED)
set(BOLT_FDATA ${CMAKE_CURRENT_BINARY_DIR}/utils/perf-training/prof.fdata)
# Pass extra flag in no-LBR mode
- if (uppercase_CLANG_BOLT STREQUAL "PERF")
+ if (CLANG_BOLT STREQUAL "PERF")
set(BOLT_NO_LBR "-nl")
endif()
- if (uppercase_CLANG_BOLT STREQUAL "INSTRUMENT")
+ if (CLANG_BOLT STREQUAL "INSTRUMENT")
# Instrument clang with BOLT
add_custom_target(clang-instrumented
DEPENDS ${CLANG_INSTRUMENTED}
diff --git a/clang/utils/perf-training/CMakeLists.txt b/clang/utils/perf-training/CMakeLists.txt
index 48fbee62a8636d1..601f40902fa34ea 100644
--- a/clang/utils/perf-training/CMakeLists.txt
+++ b/clang/utils/perf-training/CMakeLists.txt
@@ -84,13 +84,13 @@ if(CLANG_BOLT AND NOT LLVM_BUILD_INSTRUMENTED)
COMMAND "${Python3_EXECUTABLE}" ${CMAKE_CURRENT_SOURCE_DIR}/perf-helper.py clean ${CMAKE_CURRENT_BINARY_DIR} perf.data
COMMENT "Clearing old perf data")
- string(TOUPPER "${CLANG_BOLT}" uppercase_CLANG_BOLT)
- if (uppercase_CLANG_BOLT STREQUAL "LBR")
+ string(TOUPPER "${CLANG_BOLT}" CLANG_BOLT)
+ if (CLANG_BOLT STREQUAL "LBR")
set(BOLT_LBR "--lbr")
endif()
add_custom_target(merge-fdata-deps)
- if (uppercase_CLANG_BOLT STREQUAL "INSTRUMENT")
+ if (CLANG_BOLT STREQUAL "INSTRUMENT")
add_dependencies(merge-fdata-deps generate-bolt-fdata)
else()
# Convert perf profiles into fdata
diff --git a/clang/utils/perf-training/bolt.lit.cfg b/clang/utils/perf-training/bolt.lit.cfg
index d2b6042a1627e19..0e81a5501e9fcfc 100644
--- a/clang/utils/perf-training/bolt.lit.cfg
+++ b/clang/utils/perf-training/bolt.lit.cfg
@@ -6,20 +6,21 @@ import lit.util
import os
import subprocess
+clang_bolt_mode = config.clang_bolt_mode.lower()
clang_binary = "clang"
-perf_wrapper = ""
-if config.clang_bolt_mode.lower() == "instrument":
+perf_wrapper = f"{config.python_exe} {config.perf_helper_dir}/perf-helper.py perf "
+
+if clang_bolt_mode == "instrument":
+ perf_wrapper = ""
clang_binary = config.clang_bolt_name
-else: # perf or LBR
- perf_wrapper = "%s %s/perf-helper.py perf" % (
- config.python_exe,
- config.perf_helper_dir,
- )
- if config.clang_bolt_mode.lower() == "lbr":
- perf_wrapper += " --lbr"
+elif clang_bolt_mode == "lbr":
+ perf_wrapper += " --lbr -- "
+elif clang_bolt_mode == "perf":
perf_wrapper += " -- "
+else:
+ assert 0, "Unsupported CLANG_BOLT_MODE variable"
-config.clang = os.path.realpath(
+config.clang = perf_wrapper + os.path.realpath(
lit.util.which(clang_binary, config.clang_tools_dir)
).replace("\\", "/")
@@ -42,16 +43,9 @@ config.suffixes = [
use_lit_shell = os.environ.get("LIT_USE_INTERNAL_SHELL")
config.test_format = lit.formats.ShTest(use_lit_shell == "0")
config.substitutions.append(
- (
- "%clang_cpp_skip_driver",
- " %s %s --driver-mode=g++ " % (perf_wrapper, config.clang),
- )
-)
-config.substitutions.append(
- ("%clang_cpp", " %s %s --driver-mode=g++ " % (perf_wrapper, config.clang))
-)
-config.substitutions.append(
- ("%clang_skip_driver", " %s %s " % (perf_wrapper, config.clang))
+ ("%clang_cpp_skip_driver", f" {config.clang} --driver-mode=g++ ")
)
-config.substitutions.append(("%clang", " %s %s " % (perf_wrapper, config.clang)))
+config.substitutions.append(("%clang_cpp", f" {config.clang} --driver-mode=g++ "))
+config.substitutions.append(("%clang_skip_driver", config.clang))
+config.substitutions.append(("%clang", config.clang))
config.substitutions.append(("%test_root", config.test_exec_root))
diff --git a/clang/utils/perf-training/perf-helper.py b/clang/utils/perf-training/perf-helper.py
index 647bf09ffd24b51..8db6e99bbbe1778 100644
--- a/clang/utils/perf-training/perf-helper.py
+++ b/clang/utils/perf-training/perf-helper.py
@@ -74,15 +74,10 @@ def perf(args):
parser.add_argument(
"--lbr", action="store_true", help="Use perf with branch stacks"
)
- parser.add_argument("cmd", nargs="*", help="")
+ parser.add_argument("cmd", nargs=argparse.REMAINDER, help="")
- # Use python's arg parser to handle all leading option arguments, but pass
- # everything else through to perf
- first_cmd = next(arg for arg in args if not arg.startswith("--"))
- last_arg_idx = args.index(first_cmd)
-
- opts = parser.parse_args(args[:last_arg_idx])
- cmd = args[last_arg_idx:]
+ opts = parser.parse_args()
+ cmd = opts["cmd"][1:]
perf_args = [
"perf",
>From 4f2d0c27478b52bcc530c899bfe20baa0eb01087 Mon Sep 17 00:00:00 2001
From: Amir Ayupov <aaupov at fb.com>
Date: Tue, 7 Nov 2023 20:25:44 -0800
Subject: [PATCH 3/3] Bugfix
---
clang/utils/perf-training/perf-helper.py | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/clang/utils/perf-training/perf-helper.py b/clang/utils/perf-training/perf-helper.py
index 8db6e99bbbe1778..959bdba5c98ccdb 100644
--- a/clang/utils/perf-training/perf-helper.py
+++ b/clang/utils/perf-training/perf-helper.py
@@ -76,8 +76,8 @@ def perf(args):
)
parser.add_argument("cmd", nargs=argparse.REMAINDER, help="")
- opts = parser.parse_args()
- cmd = opts["cmd"][1:]
+ opts = parser.parse_args(args)
+ cmd = opts.cmd[1:]
perf_args = [
"perf",
More information about the cfe-commits
mailing list