python · erlend-aasland · May 22, 2023 · May 20, 2023 · May 21, 2023 · May 21, 2023
diff --git a/.gitignore b/.gitignore
@@ -23,6 +23,10 @@
 *.gc??
 *.profclang?
 *.profraw
+# Copies of binaries before BOLT optimizations.
+*.prebolt
+# BOLT profile data.
+*.fdata
 *.dyn
 .gdb_history
 .purify
@@ -124,6 +128,7 @@ Tools/unicode/data/
 /platform
 /profile-clean-stamp
 /profile-run-stamp
+/profile-bolt-stamp
 /Python/deepfreeze/*.c
 /pybuilddir.txt
 /pyconfig.h

@@ -314,6 +314,13 @@ also be used to improve performance.
    is dependent on a combination of the build environment + the other
    optimization configure args + the CPU architecture, and not all combinations
    are supported.
+   BOLT versions before LLVM 16 are known to crash BOLT under some scenarios.
+   Use of LLVM 16 or newer for BOLT optimization is strongly encouraged.
+
+   The :envvar:`!BOLT_INSTRUMENT_FLAGS` and :envvar:`!BOLT_APPLY_FLAGS`
+   :program:`configure` variables can be defined to override the default set of
+   arguments for :program:`llvm-bolt` to instrument and apply BOLT data to
+   binaries, respectively.
 
    .. versionadded:: 3.12
 

@@ -672,21 +672,55 @@ profile-opt: profile-run-stamp
 	-rm -f profile-clean-stamp
 	$(MAKE) @DEF_MAKE_RULE@ CFLAGS_NODIST="$(CFLAGS_NODIST) $(PGO_PROF_USE_FLAG)" LDFLAGS_NODIST="$(LDFLAGS_NODIST)"
 
-.PHONY: bolt-opt
-bolt-opt: @PREBOLT_RULE@
+# List of binaries that BOLT runs on.
+BOLT_BINARIES := @BOLT_BINARIES@
+
+BOLT_INSTRUMENT_FLAGS := @BOLT_INSTRUMENT_FLAGS@
+BOLT_APPLY_FLAGS := @BOLT_APPLY_FLAGS@
+
+.PHONY: clean-bolt
+clean-bolt:
+	# Profile data.
 	rm -f *.fdata
-	@if $(READELF) -p .note.bolt_info $(BUILDPYTHON) | grep BOLT > /dev/null; then\
-		echo "skip: $(BUILDPYTHON) is already BOLTed."; \
-	else \
-		@LLVM_BOLT@ ./$(BUILDPYTHON) -instrument -instrumentation-file-append-pid -instrumentation-file=$(abspath $(BUILDPYTHON).bolt) -o $(BUILDPYTHON).bolt_inst; \
-		./$(BUILDPYTHON).bolt_inst $(PROFILE_TASK) || true; \
-		@MERGE_FDATA@ $(BUILDPYTHON).*.fdata > $(BUILDPYTHON).fdata; \
-		@LLVM_BOLT@ ./$(BUILDPYTHON) -o $(BUILDPYTHON).bolt -data=$(BUILDPYTHON).fdata -update-debug-sections -reorder-blocks=ext-tsp -reorder-functions=hfsort+ -split-functions -icf=1 -inline-all -split-eh -reorder-functions-use-hot-size -peepholes=none -jump-tables=aggressive -inline-ap -indirect-call-promotion=all -dyno-stats -use-gnu-stack -frame-opt=hot; \
-		rm -f *.fdata; \
-		rm -f $(BUILDPYTHON).bolt_inst; \
-		mv $(BUILDPYTHON).bolt $(BUILDPYTHON); \
-	fi
+	# Pristine binaries before BOLT optimization.
+	rm -f *.prebolt
+	# BOLT instrumented binaries.
+	rm -f *.bolt_inst
+
+profile-bolt-stamp: $(BUILDPYTHON)
+	# Ensure a pristine, pre-BOLT copy of the binary and no profile data from last run.
+	for bin in $(BOLT_BINARIES); do \
+	  prebolt="$${bin}.prebolt"; \
+	  if [ -e "$${prebolt}" ]; then \
+	    echo "Restoring pre-BOLT binary $${prebolt}"; \
+	    mv "$${bin}.prebolt" "$${bin}"; \
+	  fi; \
+	  cp "$${bin}" "$${prebolt}"; \
+	  rm -f $${bin}.bolt.*.fdata $${bin}.fdata; \
+	done
+	# Instrument each binary.
+	for bin in $(BOLT_BINARIES); do \
+	  @LLVM_BOLT@ "$${bin}" -instrument -instrumentation-file-append-pid -instrumentation-file=$(abspath $${bin}.bolt) -o $${bin}.bolt_inst $(BOLT_INSTRUMENT_FLAGS); \
+	  mv "$${bin}.bolt_inst" "$${bin}"; \
+	done
+	# Run instrumented binaries to collect data.
+	$(RUNSHARED) ./$(BUILDPYTHON) $(PROFILE_TASK) || true
+	# Merge all the data files together.
+	for bin in $(BOLT_BINARIES); do \
+	  @MERGE_FDATA@ $${bin}.*.fdata > "$${bin}.fdata"; \
+	  rm -f $${bin}.*.fdata; \
+	done
+	# Run bolt against the merged data to produce an optimized binary.
+	for bin in $(BOLT_BINARIES); do \
+	  @LLVM_BOLT@ "$${bin}.prebolt" -o "$${bin}.bolt" -data="$${bin}.fdata" $(BOLT_APPLY_FLAGS); \
+	  mv "$${bin}.bolt" "$${bin}"; \
+	done
+	touch $@
 
+.PHONY: bolt-opt
+bolt-opt:
+	$(MAKE) @PREBOLT_RULE@
+	$(MAKE) profile-bolt-stamp
 
 # Compile and run with gcov
 .PHONY: coverage
@@ -2623,10 +2657,11 @@ profile-removal:
 	rm -f $(COVERAGE_INFO)
 	rm -rf $(COVERAGE_REPORT)
 	rm -f profile-run-stamp
+	rm -f profile-bolt-stamp
 
 .PHONY: clean
-clean: clean-retain-profile
-	@if test @DEF_MAKE_ALL_RULE@ = profile-opt; then \
+clean: clean-retain-profile clean-bolt
+	@if test @DEF_MAKE_ALL_RULE@ = profile-opt -o @DEF_MAKE_ALL_RULE@ = bolt-opt; then \
 		rm -f profile-gen-stamp profile-clean-stamp; \
 		$(MAKE) profile-removal; \
 	fi

diff --git a/Misc/NEWS.d/next/Build/2023-05-20-16-09-59.gh-issue-101282.FvRARb.rst b/Misc/NEWS.d/next/Build/2023-05-20-16-09-59.gh-issue-101282.FvRARb.rst
@@ -0,0 +1,4 @@
+BOLT optimization is now applied to the libpython shared library if building
+a shared library. BOLT instrumentation and application settings can now be
+influenced via the ``BOLT_INSTRUMENT_FLAGS`` and ``BOLT_APPLY_FLAGS``
+configure variables.
@@ -2028,13 +2028,6 @@ if test "$Py_BOLT" = 'true' ; then
   DEF_MAKE_ALL_RULE="bolt-opt"
   DEF_MAKE_RULE="build_all"
 
-  AC_SUBST(READELF)
-  AC_CHECK_TOOLS(READELF, [readelf], "notfound")
-  if test "$READELF" == "notfound"
-  then
-    AC_MSG_ERROR([readelf is required for a --enable-bolt build but could not be found.])
-  fi
-
   # -fno-reorder-blocks-and-partition is required for bolt to work.
   # Possibly GCC only.
   AX_CHECK_COMPILE_FLAG([-fno-reorder-blocks-and-partition],[
@@ -2067,6 +2060,54 @@ if test "$Py_BOLT" = 'true' ; then
   fi
 fi
 
+dnl Enable BOLT of libpython if built.
+AC_SUBST(BOLT_BINARIES)
+BOLT_BINARIES='$(BUILDPYTHON)'
+AS_VAR_IF([enable_shared], [yes], [
+  BOLT_BINARIES="${BOLT_BINARIES} \$(INSTSONAME)"
+])
+
+AC_ARG_VAR(
+  [BOLT_INSTRUMENT_FLAGS],
+  [Arguments to llvm-bolt when instrumenting binaries]
+)
+AC_MSG_CHECKING([BOLT_INSTRUMENT_FLAGS])
+if test -z "${BOLT_INSTRUMENT_FLAGS}"
+then
+  BOLT_INSTRUMENT_FLAGS=
+fi
+AC_MSG_RESULT([$BOLT_INSTRUMENT_FLAGS])
+
+AC_ARG_VAR(
+  [BOLT_APPLY_FLAGS],
+  [Arguments to llvm-bolt when creating a BOLT optimized binary]
+)
+AC_MSG_CHECKING([BOLT_APPLY_FLAGS])
+if test -z "${BOLT_APPLY_FLAGS}"
+then
+  AS_VAR_SET(
+    [BOLT_APPLY_FLAGS],
+    [m4_join([ ],
+      [-update-debug-sections],
+      [-reorder-blocks=ext-tsp],
+      [-reorder-functions=hfsort+],
+      [-split-functions],
+      [-icf=1],
+      [-inline-all],
+      [-split-eh],
+      [-reorder-functions-use-hot-size],
+      [-peepholes=none],
+      [-jump-tables=aggressive],
+      [-inline-ap],
+      [-indirect-call-promotion=all],
+      [-dyno-stats],
+      [-use-gnu-stack],
+      [-frame-opt=hot]
+    )]
+  )
+fi
+AC_MSG_RESULT([$BOLT_APPLY_FLAGS])
+
 # XXX Shouldn't the code above that fiddles with BASECFLAGS and OPT be
 # merged with this chunk of code?