Add BOLT Makefile

Make.inc

@@ -516,6 +516,11 @@ SHIPFLAGS_COMMON  := -O3
 SHIPFLAGS_CLANG   := $(SHIPFLAGS_COMMON) -g
 SHIPFLAGS_GCC     := $(SHIPFLAGS_COMMON) -ggdb2 -falign-functions
 
+BOLT_LDFLAGS :=
+
+BOLT_CFLAGS_GCC    :=
+BOLT_CFLAGS_CLANG  :=
+
 ifeq ($(OS), Darwin)
 JCPPFLAGS_CLANG   += -D_LARGEFILE_SOURCE -D_DARWIN_USE_64_BIT_INODE=1
 endif
@@ -532,7 +537,8 @@ JCFLAGS    := $(JCFLAGS_GCC)
 JCPPFLAGS  := $(JCPPFLAGS_GCC)
 JCXXFLAGS  := $(JCXXFLAGS_GCC)
 DEBUGFLAGS := $(DEBUGFLAGS_GCC)
-SHIPFLAGS  := $(SHIPFLAGS_GCC)
+SHIPFLAGS  := $(SHIPFLAGS_GCC) $(BOLT_CFLAGS_GCC)
+BOLT_CFLAGS  := $(BOLT_CFLAGS_GCC)
 endif
 
 ifeq ($(USECLANG),1)
@@ -542,7 +548,8 @@ JCFLAGS    := $(JCFLAGS_CLANG)
 JCPPFLAGS  := $(JCPPFLAGS_CLANG)
 JCXXFLAGS  := $(JCXXFLAGS_CLANG)
 DEBUGFLAGS := $(DEBUGFLAGS_CLANG)
-SHIPFLAGS  := $(SHIPFLAGS_CLANG)
+SHIPFLAGS  := $(SHIPFLAGS_CLANG) $(BOLT_CFLAGS_CLANG)
+BOLT_CFLAGS  := $(BOLT_CFLAGS_CLANG)
 
 ifeq ($(OS), Darwin)
 CC += -mmacosx-version-min=$(MACOSX_VERSION_MIN)
@@ -1295,7 +1302,7 @@ CSL_NEXT_GLIBCXX_VERSION=GLIBCXX_3\.4\.33|GLIBCXX_3\.5\.|GLIBCXX_4\.
 # Note: we explicitly _do not_ define `CSL` here, since it requires some more
 # advanced techniques to decide whether it should be installed from a BB source
 # or not.  See `deps/csl.mk` for more detail.
-BB_PROJECTS := BLASTRAMPOLINE OPENBLAS LLVM LIBSUITESPARSE OPENLIBM GMP MBEDTLS LIBSSH2 NGHTTP2 MPFR CURL LIBGIT2 PCRE LIBUV LIBUNWIND DSFMT OBJCONV ZLIB P7ZIP LLD LIBTRACYCLIENT
+BB_PROJECTS := BLASTRAMPOLINE OPENBLAS LLVM LIBSUITESPARSE OPENLIBM GMP MBEDTLS LIBSSH2 NGHTTP2 MPFR CURL LIBGIT2 PCRE LIBUV LIBUNWIND DSFMT OBJCONV ZLIB P7ZIP LLD LIBTRACYCLIENT BOLT
 define SET_BB_DEFAULT
 # First, check to see if BB is disabled on a global setting
 ifeq ($$(USE_BINARYBUILDER),0)

contrib/bolt/.gitignore

@@ -0,0 +1,10 @@
+profiles-bolt*
+optimized.build
+toolchain
+
+bolt
+bolt_instrument
+merge_data
+copy_originals
+stage0
+stage1

contrib/bolt/Makefile

@@ -0,0 +1,134 @@
+.PHONY: clean clean_profiles restore_originals
+
+# Settings taken from https://github.com/rust-lang/rust/blob/master/src/tools/opt-dist/src/bolt.rs
+BOLT_ARGS :=
+# Reorder basic blocks within functions
+BOLT_ARGS += -reorder-blocks=ext-tsp
+# Reorder functions within the binary
+BOLT_ARGS += -reorder-functions=cdsort
+# Split function code into hot and code regions
+BOLT_ARGS += -split-functions
+# Split as many basic blocks as possible
+BOLT_ARGS += -split-all-cold
+# Move jump tables to a separate section
+BOLT_ARGS += -jump-tables=move
+# Use regular size pages for code alignment
+BOLT_ARGS += -no-huge-pages
+# Fold functions with identical code
+BOLT_ARGS += -icf=1
+# Split using best available strategy (three-way splitting, Cache-Directed Sort)
+# Disabled for libjulia-internal till https://github.com/llvm/llvm-project/issues/89508 is fixed
+# BOLT_ARGS += -split-strategy=cdsplit
+# Update DWARF debug info in the final binary
+BOLT_ARGS += -update-debug-sections
+# Print optimization statistics
+BOLT_ARGS += -dyno-stats
+# BOLT doesn't fully support computed gotos, https://github.com/llvm/llvm-project/issues/89117
+# Use escaped regex as the name BOLT recognises is often a bit different, e.g. apply_cl/1(*2)
+# This doesn't actually seem to do anything, the actual mitigation is not using --use-old-text
+# which we do in the bolt target
+BOLT_ARGS += -skip-funcs=.\*apply_cl.\*
+
+# -fno-reorder-blocks-and-partition is needed on gcc >= 8.
+BOLT_FLAGS := $\
+	"BOLT_CFLAGS_GCC+=-fno-reorder-blocks-and-partition" $\
+	"BOLT_LDFLAGS=-Wl,--emit-relocs"
+
+STAGE0_BUILD:=$(CURDIR)/toolchain
+STAGE1_BUILD:=$(CURDIR)/optimized.build
+
+STAGE0_BINARIES:=$(STAGE0_BUILD)/usr/bin/
+
+PROFILE_DIR:=$(CURDIR)/profiles-bolt
+JULIA_ROOT:=$(CURDIR)/../..
+
+LLVM_BOLT:=$(STAGE0_BINARIES)llvm-bolt
+LLVM_MERGEFDATA:=$(STAGE0_BINARIES)merge-fdata
+
+# If you add new files to optimize, you need to add BOLT_LDFLAGS and BOLT_CFLAGS to the build of your new file.
+SYMLINKS_TO_OPTIMIZE := libLLVM.so libjulia-internal.so libjulia-codegen.so
+FILES_TO_OPTIMIZE := $(shell for file in $(SYMLINKS_TO_OPTIMIZE); do readlink $(STAGE1_BUILD)/usr/lib/$$file; done)
+
+AFTER_INSTRUMENT_MESSAGE:='Run `make finish_stage1` to finish off the build. $\
+	You can now optionally collect more profiling data by running Julia with an appropriate workload, $\
+	if you wish, run `make clean_profiles` before doing so to remove any profiling data generated by `make finish_stage1`. $\
+	You should end up with some data in $(PROFILE_DIR). Afterwards run `make merge_data && make bolt`. $\
+
+$(STAGE0_BUILD) $(STAGE1_BUILD):
+	$(MAKE) -C $(JULIA_ROOT) O=$@ configure
+
+stage0: | $(STAGE0_BUILD)
+	$(MAKE) -C $(STAGE0_BUILD)/deps install-BOLT && \
+	touch $@
+
+# Build with our custom flags, binary builder doesn't use them so we need to build LLVM for now.
+# We manually skip package image creation so that we can profile it
+$(STAGE1_BUILD): stage0
+stage1: export USE_BINARYBUILDER_LLVM=0
+stage1: | $(STAGE1_BUILD)
+	$(MAKE) -C $(STAGE1_BUILD) $(BOLT_FLAGS) julia-src-release julia-symlink julia-libccalltest \
+								julia-libccalllazyfoo julia-libccalllazybar julia-libllvmcalltest && \
+	touch $@
+
+copy_originals: stage1
+	for file in $(FILES_TO_OPTIMIZE); do \
+		abs_file=$(STAGE1_BUILD)/usr/lib/$$file; \
+		cp $$abs_file "$$abs_file.original"; \
+	done && \
+	touch $@
+
+# I don't think there's any particular reason to have -no-huge-pages here, perhaps slightly more accurate profile data
+# as the final build uses -no-huge-pages
+bolt_instrument: copy_originals
+	for file in $(FILES_TO_OPTIMIZE); do \
+		abs_file=$(STAGE1_BUILD)/usr/lib/$$file; \
+		$(LLVM_BOLT) "$$abs_file.original" -o $$abs_file --instrument --instrumentation-file-append-pid --instrumentation-file="$(PROFILE_DIR)/$$file-prof" -no-huge-pages; \
+		mkdir -p $$(dirname "$(PROFILE_DIR)/$$file-prof"); \
+		printf "\n"; \
+	done && \
+	touch $@
+	@echo $(AFTER_INSTRUMENT_MESSAGE)
+
+# We don't want to rebuild julia-src as then we lose the bolt instrumentation
+# So we have to manually build the sysimage and package image
+finish_stage1: stage1
+	$(MAKE) -C $(STAGE1_BUILD) julia-base-cache && \
+	$(MAKE) -C $(STAGE1_BUILD) -f sysimage.mk sysimg-release && \
+	$(MAKE) -C $(STAGE1_BUILD) -f pkgimage.mk release
+
+merge_data: bolt_instrument
+	for file in $(FILES_TO_OPTIMIZE); do \
+		profiles=$(PROFILE_DIR)/$$file-prof.*.fdata; \
+		$(LLVM_MERGEFDATA) $$profiles > "$(PROFILE_DIR)/$$file-prof.merged.fdata"; \
+	done && \
+	touch $@
+
+# The --use-old-text saves about 16 MiB of libLLVM.so size.
+# However, the rust folk found it succeeds very non-deterministically for them.
+# It tries to reuse old text segments to reduce binary size
+# BOLT doesn't fully support computed gotos https://github.com/llvm/llvm-project/issues/89117, so we cannot use --use-old-text on libjulia-internal
+# That flag saves less than 1 MiB for libjulia-internal so oh well.
+bolt: merge_data
+	for file in $(FILES_TO_OPTIMIZE); do \
+        abs_file=$(STAGE1_BUILD)/usr/lib/$$file; \
+		$(LLVM_BOLT) "$$abs_file.original" -data "$(PROFILE_DIR)/$$file-prof.merged.fdata" -o $$abs_file $(BOLT_ARGS) $$(if [ "$$file" != $(shell readlink $(STAGE1_BUILD)/usr/lib/libjulia-internal.so) ]; then echo "--use-old-text -split-strategy=cdsplit"; fi); \
+    done && \
+    touch $@
+
+clean_profiles:
+	rm -rf $(PROFILE_DIR)
+
+clean:
+	rm -f stage0 stage1 bolt copy_originals merge_data bolt_instrument
+
+restore_originals: copy_originals
+	for file in $(FILES_TO_OPTIMIZE); do \
+		abs_file=$(STAGE1_BUILD)/usr/lib/$$file; \
+		cp -P "$$abs_file.original" $$abs_file; \
+	done
+
+delete_originals: copy_originals
+	for file in $(FILES_TO_OPTIMIZE); do \
+		abs_file=$(STAGE1_BUILD)/usr/lib/$$file; \
+		rm "$$abs_file.original"; \
+	done

contrib/bolt/README.md

@@ -0,0 +1,17 @@
+BOLT only works on x86_64 and arch64 on Linux.
+
+DO NOT STRIP THE RESULTING .so FILES, https://github.com/llvm/llvm-project/issues/56738.
+If you really need to, try adding `-use-gnu-stack` to `BOLT_ARGS`.
+
+To build a BOLT-optimized version of Julia run the following commands (`cd` into this directory first)
+```bash
+make stage1
+make copy_originals
+make bolt_instrument
+make finish_stage1
+make merge_data
+make bolt
+```
+After these commands finish, the optimized version of Julia will be built in the `optimized.build` directory.
+
+This doesn't align the code to support huge pages as it doesn't seem that we do that currently, this decreases the size of the .so files by 2-4mb.

contrib/pgo-lto-bolt/.gitignore

@@ -0,0 +1,14 @@
+stage0*
+stage1*
+stage2*
+bolt
+bolt_instrument
+merge_data
+copy_originals
+
+profiles
+profiles-bolt
+
+toolchain
+pgo-instrumented.build
+optimized.build