diff --git a/contrib/pgo-lto/.gitignore b/contrib/pgo-lto/.gitignore
new file mode 100644
index 0000000000000..978d8f2ca86dd
--- /dev/null
+++ b/contrib/pgo-lto/.gitignore
@@ -0,0 +1,4 @@
+profiles
+stage0*
+stage1*
+stage2*
diff --git a/contrib/pgo-lto/Makefile b/contrib/pgo-lto/Makefile
new file mode 100644
index 0000000000000..a73825e182561
--- /dev/null
+++ b/contrib/pgo-lto/Makefile
@@ -0,0 +1,80 @@
+.PHONY: top clean clean-profiles
+
+STAGE0_BUILD:=$(CURDIR)/stage0.build
+STAGE1_BUILD:=$(CURDIR)/stage1.build
+STAGE2_BUILD:=$(CURDIR)/stage2.build
+
+STAGE0_TOOLS:=$(STAGE0_BUILD)/usr/tools/
+
+PROFILE_DIR:=$(CURDIR)/profiles
+PROFILE_FILE:=$(PROFILE_DIR)/merged.prof
+PROFRAW_FILES:=$(wildcard $(PROFILE_DIR)/*.profraw)
+JULIA_ROOT:=$(CURDIR)/../..
+
+LLVM_CXXFILT:=$(STAGE0_TOOLS)llvm-cxxfilt
+LLVM_PROFDATA:=$(STAGE0_TOOLS)llvm-profdata
+LLVM_OBJCOPY:=$(STAGE0_TOOLS)llvm-objcopy
+
+# When building a single libLLVM.so we need to increase -vp-counters-per-site
+# significantly
+COUNTERS_PER_SITE:=6
+
+AFTER_STAGE1_MESSAGE:='Run `make clean-profiles` to start with a clean slate. $\
+    Then run Julia to collect realistic profile data, for example: `$(STAGE1_BUILD)/julia -O3 -e $\
+    '\''using Pkg; Pkg.add("LoopVectorization"); Pkg.test("LoopVectorization")'\''`. This $\
+	should produce about 15MB of data in $(PROFILE_DIR). Note that running extensive $\
+	scripts may result in counter overflows, which can be detected by running $\
+	`make top`. Afterwards run `make stage2`.'
+
+TOOLCHAIN_FLAGS = $\
+	"CC=$(STAGE0_TOOLS)clang" $\
+	"CXX=$(STAGE0_TOOLS)clang++" $\
+	"LD=$(STAGE0_TOOLS)ld.lld" $\
+	"AR=$(STAGE0_TOOLS)llvm-ar" $\
+	"RANLIB=$(STAGE0_TOOLS)llvm-ranlib" $\
+	"CFLAGS+=$(PGO_CFLAGS)" $\
+	"CXXFLAGS+=$(PGO_CXXFLAGS)" $\
+	"LDFLAGS+=$(PGO_LDFLAGS)"
+
+$(STAGE0_BUILD) $(STAGE1_BUILD) $(STAGE2_BUILD):
+	$(MAKE) -C $(JULIA_ROOT) O=$@ configure
+
+stage0: export USE_BINARYBUILDER_LLVM=1
+stage0: | $(STAGE0_BUILD)
+	# Turn [cd]tors into init/fini_array sections in libclang_rt, since lld
+	# doesn't do that, and otherwise the profile constructor is not executed
+	$(MAKE) -C $(STAGE0_BUILD)/deps install-clang install-llvm install-lld install-llvm-tools && \
+	find $< -name 'libclang_rt.profile-*.a' -exec $(LLVM_OBJCOPY) --rename-section .ctors=.init_array --rename-section .dtors=.fini_array {} + && \
+	touch $@
+
+$(STAGE1_BUILD): stage0
+stage1: PGO_CFLAGS:=-fprofile-generate=$(PROFILE_DIR) -Xclang -mllvm -Xclang -vp-counters-per-site=$(COUNTERS_PER_SITE)
+stage1: PGO_CXXFLAGS:=-fprofile-generate=$(PROFILE_DIR) -Xclang -mllvm -Xclang -vp-counters-per-site=$(COUNTERS_PER_SITE)
+stage1: PGO_LDFLAGS:=-fuse-ld=lld -flto=thin -fprofile-generate=$(PROFILE_DIR)
+stage1: export USE_BINARYBUILDER_LLVM=0
+stage1: | $(STAGE1_BUILD)
+	$(MAKE) -C $(STAGE1_BUILD) $(TOOLCHAIN_FLAGS) && touch $@
+	@echo $(AFTER_STAGE1_MESSAGE)
+
+stage2: PGO_CFLAGS:=-fprofile-use=$(PROFILE_FILE)
+stage2: PGO_CXXFLAGS:=-fprofile-use=$(PROFILE_FILE)
+stage2: PGO_LDFLAGS:=-fuse-ld=lld -flto=thin -fprofile-use=$(PROFILE_FILE) -Wl,--icf=safe
+stage2: export USE_BINARYBUILDER_LLVM=0
+stage2: $(PROFILE_FILE) | $(STAGE2_BUILD)
+	$(MAKE) -C $(STAGE2_BUILD) $(TOOLCHAIN_FLAGS) && touch $@
+
+install: stage2
+	$(MAKE) -C $(STAGE2_BUILD) USE_BINARYBUILDER_LLVM=0 install
+
+$(PROFILE_FILE): stage1 $(PROFRAW_FILES)
+	$(LLVM_PROFDATA) merge -output=$@ $(PROFRAW_FILES)
+
+# show top 50 functions
+top: $(PROFILE_FILE)
+	$(LLVM_PROFDATA) show --topn=50 $< | $(LLVM_CXXFILT)
+
+clean-profiles:
+	rm -rf $(PROFILE_DIR)
+
+clean:
+	rm -f stage0 stage1 stage2 $(PROFILE_FILE)