# Mostly-header library for threaded-programming sync primitives & helpers `fx::[shared::]{Mutex,CondVar}`, `fx::Queue<T>`, & namespace `fx::signal`.
#
# Likely prefer static-linking. Non-header elements are required for `fx::error` & certain required features of `fx::signal::`
#
# Contains targets for `release', `debug', and `clean'.
PROJECT=fx-sync
AUTHOR=Avril (Flanchan) <flanchan@cumallover.me>
DESCRIPTION=sync module of libfx
VERSION=0.1.1

SRC = src

SRC_C   = $(shell find $(SRC)/ -type f -name \*.c -and -not -path $(SRC)/test/\* )
SRC_CXX = $(shell find -O2 $(SRC)/ -type f \( -name \*.cpp -or -name \*.cxx \) -and -not -path $(SRC)/test/\* )

INCLUDE=include
# If PCH should be auto-included for all TUs, set to 1.
INCLUDE_PCH_GLOBAL?=0
# Files to be auto-included by all TUs (after PCH global, if `INCLUDE_PCH_GLOBAL == 1`.)
INCLUDE_GLOBAL=

# Link to these libraries dynamicalls
SHARED_LIBS=
# Link to these libraries statically
STATIC_LIBS=

# Prefix for un/install targets, if it's blank, ensure it is set (default to `/usr/local` instead of `/usr`.)
ifeq ($(PREFIX),)
	PREFIX := /usr/local
endif

# Prefix for un/install target headers (default: all of `$(INCLUDE)`)
# Make blank for default install to PREFIX/include
# TODO: Do we want to turn this to `fx/sync/` instead of `fx-sync/`? What if `fx` is taken in usr/include..? (XXX: Or if other fx-* projects have things of the same name (esp. error.hh which is currently not standardised...?) Leave it as fx-sync for now i think...)
INCLUDE_PREFIX := $(PROJECT)/

# Default archivers
AR?=ar
RANLIB?=ranlib

# Default linker
LINKER?=gold

# Use gcc-{ar,ranlib} when using gcc
ifeq ($(CXX),g++)
	AR:=gcc-ar
	RANLIB:=gcc-ranlib
endif

# Pre-compile these headers
PCH_HEADERS+=
# PCH_HEADERS depend on these header files
PCH_INCLUDES+=

#TODO: Add in the configurable opt levels from `work/map-rc/Makefile`

# Compile-time default program features (see `Features application` below.)
FEATURES?=

# Build constants
CONSTANTS+=_GNU_SOURCE __API_EXPORT__=__attribute__\(\(\visibility\(\"default\"\)\)\)

# Default visibility ("default" / "", "hidden", "protected", "internal".)
# If not exported, symbols must be manually exported.
SYMBOL_VISIBILITY?=default

# Reduce binary size by removing inline address identity & define symbols relevant only to exported symbols
#
# NOTE: This requires either `SYMBOL_VISIBILITY=hidden` or manual non-exported symbol hiding to be effective. 
# 	Only disable this if `inline`s need address identity.
STRICT_EXPORT?=yes

# Testing
# Can be {debug,release}.{a,so}
TARGET?=debug.a

TEST_LDFLAGS+= -lfmt -lstdc++ -Wl,-z -Wl,now -Wl,-z -Wl,relro
TEST_CFLAGS+= -Og -g -fwhole-program

override __COMMA=,
override __VERSION_SPLIT:= $(subst ., ,$(VERSION))
override __VERSION_REVISION:=$(word 3,$(__VERSION_SPLIT)) 0

VERSION_MAJOR:= $(word 1,$(__VERSION_SPLIT))
VERSION_MINOR:= $(word 2,$(__VERSION_SPLIT))
VERSION_BUGFIX:= $(word 3,$(__VERSION_SPLIT))
VERSION_REVISION:= $(word 2,$(subst r, ,$(__VERSION_REVISION)))

override __VERSION_SPLIT:= MAJOR:$(word 1,$(__VERSION_SPLIT)) MINOR:$(word 2,$(__VERSION_SPLIT)) BUGFIX:$(word 1,$(subst r, ,$(__VERSION_REVISION))) REVISION:$(word 2,$(subst r, ,$(__VERSION_REVISION))) REVISION_STRING:$(word 3,$(__VERSION_SPLIT))

COMMON_FLAGS+= -W -Wall 
COMMON_FLAGS+= $(addprefix -D,$(CONSTANTS))
COMMON_FLAGS+= -pipe -Wstrict-aliasing -fno-strict-aliasing $(addprefix -I,$(INCLUDE))
COMMON_FLAGS+= $(addprefix -D_VERSION_,$(subst :,=,$(__VERSION_SPLIT))) '-D_PROJECT_VERSION="$(VERSION)"' '-D_PROJECT_AUTHOR="$(AUTHOR)"' '-D_PROJECT_NAME="$(PROJECT)"' 
COMMON_FLAGS+= $(addprefix -fvisibility=,$(SYMBOL_VISIBILITY))
ifeq ($(STRICT_EXPORT),yes)
	COMMON_FLAGS+= -fvisibility-inlines-hidden -ffunction-sections -fdata-sections
endif

# PCH targets for `%_p.hh` -> `%_p.hh.gch`, add `-include $(PCH_HEADERS)` to general `COMMON_FLAGS`, and add $(PCH_OUT) as a requirement for all `%.c/pp` targets as well. (where `PCH_OUT= $(addsuffix .gch,$(PCH_HEADERS))`.)

# Target arch & CPU. Set to blank for generic
ARCH?=native
CPU?=native
# Enable OpenMP and loop parallelisation? (dyn-links to openmp, allows C++STL algorithms to use parallel execution mode when provided, enables auto-parallelise optimisation of loops in `release` build only.)
#
# To enable in CXX without auto-optimisation, set to `maybe` instead of `yes`. (__NOTE__: When set to `maybe`, if C++ parallel execution modes are used (from the `<execution>` header,) you will need to manually add `tbb` to `SHARED_LIBS`. `yes` and `force` add `-ltbb` automatically alongside `-fopenmp`.)
# To enable throughout, *regardless of* release-only optimisations, set to `force`, (`yes` only enables to always in CXX, and only enables it for CC in release builds.)
# To *only ever* enable in release builds, *regardless of optimisations*, set to `release` (which acts as `yes` but only in release builds.)
PARALLEL?=no

# Set exception mode to:
# - Empty	- Exceptions are left as the default for CXX, and nothing about them is communicated to CC either.
# - 'debug' (default)	- Same as above, but `-fexceptions` is passed to CC in debug builds only. (XXX: Is this useful, to not pass it to CXX as well in debug...?)
# - 'yes'		- Pass `-fexceptions` to CC.
# - 'no'		- Pass `-fno-exceptions` to CXX.
# - 'force'		- Pass `-fexceptions` to CC and CXX always.
# - 'never'		- Pass `-fno-exceptions` to CC and CXX always.
EXCEPTIONS?=force
# Enable CPU-specific features
CPU_FLAGS?=

BINFLAGS+=
DEBUG_BINFLAGS+=
RELEASE_BINFLAGS+= -fuse-linker-plugin

OPT_FLAGS?= -fgraphite \
	    -floop-interchange -ftree-loop-distribution -floop-strip-mine -floop-block \
	    -fno-stack-check

# Static and shared common flags
SHARED_FLAGS+=-fPIC
SHARED_RELEASE_FLAGS+=
SHARED_DEBUG_FLAGS+=

STATIC_FLAGS+=
STATIC_RELEASE_FLAGS+=-ffat-lto-objects
STATIC_DEBUG_FLAGS+=

# Features application

## Tell program which features are enabled via `FEATURE_<feature name in UPPER_SNAKE_CASE>`.
override __FEATURES=$(shell echo "$(FEATURES)" | tr '[:lower:]' '[:upper:]' | tr '-' '_')
COMMON_FLAGS+=$(addsuffix =1,$(addprefix -DFEATURE_,$(__FEATURES)))

## Specific feature additions

# ## Example:
#ifneq (,$(findstring fast-math,$(FEATURES)))
#	# fast-math feature: Apply `-ffast-math`
#	COMMON_FLAGS+=-ffast-math
#endif

# Add linker
ifneq ($(LINKER),)
	COMMON_FLAGS+=-fuse-ld=$(LINKER)
endif

# Arch and optimisation
ifneq ($(ARCH),)
	OPT_FLAGS+= $(addprefix -march=,$(ARCH))
endif

ifneq ($(CPU),)
	OPT_FLAGS+= $(addprefix -mtune=,$(CPU))
endif

# Add parallelisation flags
override __PARALLEL_OPT_FLAGS=-floop-parallelize-all -ftree-parallelize-loops=4

ifeq ($(PARALLEL),yes)
	# NOTE: `-fopenmp` adds parallelised STL algorithms by default, so we want C++ (debug & release) TUs to always be compiled with it
	# XXX: Should we add this in for all `COMMON_FLAGS` to enable to C TUs too? Will this make the interop / debugging any better to *only* enable it in C for release builds??
	
	COMMON_FLAGS+= -fopenmp -DPARALLEL=1
	OPT_FLAGS+= $(__PARALLEL_OPT_FLAGS)

	# For C++ `std::execution` parallel execution models.
	LDFLAGS+= -ltbb -lgomp
endif
ifeq ($(PARALLEL),maybe)
	# Define `PARALLEL` constant, but to 0. So `#ifdef` and `#if` can distinguish.
	CFLAGS+= -DPARALLEL=0
	# Add openmp to C & CXX, without adding loop auto-parallelise optimisation
	CXXFLAGS+= -fopenmp -DPARALLEL=1 -DPARALLEL_CXX
endif
ifeq ($(PARALLEL),force)
	# Add openmp to C & CXX
	COMMON_FLAGS+= -fopenmp -DPARALLEL=1 -DPARALLEL_FORCE
	OPT_FLAGS+= $(__PARALLEL_OPT_FLAGS)

	# For C++ `std::execution` parallel execution models.
	LDFLAGS+= -ltbb -lgomp
endif
ifeq ($(PARALLEL),release)
	# Enable OMP & parallel opt in `release` mode only 
	
	# NOTE: Adds in non-opt flags *after* target-spec below. 
	OPT_FLAGS+= $(__PARALLEL_OPT_FLAGS)
endif

COMMON_FLAGS+=$(addprefix -m,$(CPU_FLAGS))

CXX_OPT_FLAGS?= $(OPT_FLAGS) -felide-constructors

COMMON_STD=23

CSTD?=gnu$(COMMON_STD)
CXXSTD?=gnu++$(COMMON_STD)

# Build Options

STRIP=strip

# XXX: Is `-fno-plt` actually helpful here?
RELEASE_COMMON_FLAGS+= -fno-plt -fno-bounds-check
DEBUG_COMMON_FLAGS+= -ggdb -gz -ftrapv -fbounds-check

ifneq ($(TARGET_SPEC_FLAGS),no)
	RELEASE_CFLAGS?=   -O3 -flto $(OPT_FLAGS)
	RELEASE_CXXFLAGS?= -O3 -flto $(CXX_OPT_FLAGS)
	RELEASE_LDFLAGS?=  -Wl,-O3 -Wl,-flto

	DEBUG_CFLAGS?= -Og	
	DEBUG_CXXFLAGS?= -Og

	DEBUG_LDFLAGS?=
endif

ifeq ($(PARALLEL),release)
	# Add openmp to C & CXX *only* in release mode.
	RELEASE_COMMON_FLAGS+= -fopenmp -DPARALLEL=1

	RELEASE_LDFLAGS+= -ltbb -lgomp
endif

# Set exception modes (See def `EXCEPTIONS`.)

ifneq ($(EXCEPTIONS),)
ifeq	($(EXCEPTIONS),debug)
	# XXX (see above) Should we change to `DEBUG_COMMON_FLAGS` here to pass to CXX as well?
	DEBUG_CFLAGS+=-fexceptions
endif

ifeq	($(EXCEPTIONS),yes)
	CFLAGS+=-fexceptions
endif

ifeq	($(EXCEPTIONS),no)
	CXXFLAGS+=-fno-exceptions
endif

ifeq	($(EXCEPTIONS),force)
	COMMON_FLAGS+=-fexceptions
endif

ifeq	($(EXCEPTIONS),never)
	COMMON_FLAGS+=-fno-exceptions
endif
endif

DEBUG_CFLAGS+=-DDEBUG $(DEBUG_COMMON_FLAGS)
DEBUG_CXXFLAGS+=-DDEBUG $(DEBUG_COMMON_FLAGS) -fasynchronous-unwind-tables

RELEASE_CFLAGS+=-DRELEASE $(RELEASE_COMMON_FLAGS)
RELEASE_CXXFLAGS+=-DRELEASE $(RELEASE_COMMON_FLAGS)

# Objects

OBJ_C   = $(addprefix obj/c/,$(SRC_C:.c=.o))
OBJ_CXX = $(addprefix obj/cxx/,$(SRC_CXX:.cpp=.o))
OBJ = $(OBJ_C) $(OBJ_CXX)

# Pre-compiled header objects

PCH_HEADERS+=$(shell find -O2 $(INCLUDE) -type f -name \*_p.hh -or -name \*_p.h)

PCH_OUTPUT_LOCATION=
PCH_OUTPUT=$(addprefix $(PCH_OUTPUT_LOCATION),$(addsuffix .gch,$(PCH_HEADERS)))

PCH_BUILD_COMMON_FLAGS+=-H 
#-D_PCH_BUILD=1

PCH_BUILD_CFLAGS+= -x c-header $(PCH_BUILD_COMMON_FLAGS)
PCH_BUILD_CXXFLAGS+= -x c++-header $(PCH_BUILD_COMMON_FLAGS)

PCH_CHEADERS=$(filter %.h,$(PCH_HEADERS))
PCH_CXXHEADERS=$(filter %.hh,$(PCH_HEADERS))

PCH_USE_COMMON_FLAGS+=-Winvalid-pch

ifeq ($(INCLUDE_PCH_GLOBAL),1)
# Globally included
PCH_USE_CFLAGS+= $(addprefix -include ,$(realpath $(PCH_CHEADERS)))
PCH_USE_CXXFLAGS+= $(addprefix -include ,$(realpath $(PCH_CXXHEADERS)))
else
# Seperately include them
ifneq ($(PCH_OUTPUT_LOCATION),)
override PCH_OUTPUT_LOCATION:=$(dir $(PCH_OUTPUT_LOCATION)) 
# XXX: We don't wan't to be using PCH_OUTPUT_LOCATION really...
COMMON_FLAGS:= $(addprefix -I,$(dir $(shell find -O3 $(PCH_OUTPUT_LOCATION) -mindepth 1 -type d -printf %p/ ))) $(COMMON_FLAGS) 
endif
endif

COMMON_FLAGS+=$(addprefix -include ,$(INCLUDE_GLOBAL))

PCH_USE_CFLAGS+= $(PCH_USE_COMMON_FLAGS) 
PCH_USE_CXXFLAGS+= $(PCH_USE_COMMON_FLAGS) 

# Compiler Flags

CFLAGS   += $(COMMON_FLAGS) --std=$(CSTD)
CXXFLAGS += $(COMMON_FLAGS) --std=$(CXXSTD) 
LDFLAGS  += $(addsuffix .a,$(addprefix -l:lib,$(STATIC_LIBS))) $(addprefix -l,$(SHARED_LIBS))

# PGO (unused for lib targets)

PROF_FLAGS= -D_PGO_GEN -fprofile-generate
PGO_OBJ_C= $(addprefix prof/c/,$(SRC_C:.c=.o))
PGO_OBJ_CXX= $(addprefix prof/cxx/,$(SRC_CXX:.c=.o))
PGO_OBJ= $(PGO_OBJ_C) $(PGO_OBJ_CXX)

PROF_ITERATIONS=10
PROF_LOCATION?=/tmp/$(PROJECT)-pgo
PROF_LARGE_BOUND= 10240
PROF_SMALL_BOUND= 1024

# Phonies

# XXX: This doesn't force them to run in series for some reason?

.PHONY: release
release: | dirs
	$(MAKE) lib$(PROJECT).a
	@$(MAKE) clean-rebuild >> /dev/null
	@$(MAKE) dirs >> /dev/null
	$(MAKE) lib$(PROJECT).so

.PHONY: debug
debug: | dirs
	$(MAKE) lib$(PROJECT)-debug.a
	@$(MAKE) clean-rebuild >> /dev/null
	@$(MAKE) dirs >> /dev/null
	$(MAKE) lib$(PROJECT)-debug.so

# Rebuild both release and debug targets from scratch
.PHONY: all
all: | clean
	@$(MAKE) release
	@$(MAKE) clean-rebuild
	@$(MAKE) debug

.PHONY: test
test: $(PROJECT)-test
	-strace ./$<
	-valgrind ./$<

# Targets

.PHONY: pch
pch: | dirs $(PROJECT)-pch
	@echo 'WARNING: When building target $(PROJECT)-pch, the PCH file(s) will not be built with the default auto-opts enabled by targets of kind `debug` or `release`. This should only be used when full manually overriding all compile and linker flags' >&2

$(PROJECT)-pch: $(PCH_OUTPUT)
	@echo ""
	@echo 'PCH: $(PCH_OUTPUT)' '<-' '$(PCH_HEADERS)'
	@echo " with: $(PCH_INCLUDES)"
	@echo 'C: $(PCH_CHEADERS)'
	@echo 'C++: $(PCH_CXXHEADERS)'
# Invoking target `pch` itself is not necissary, they are build automatically where they are used.

dirs:
	@mkdir -p $(addprefix {obj$(__COMMA)prof}/c{$(__COMMA)xx}/,$(shell find $(SRC)/ -type d))


$(PCH_OUTPUT_LOCATION)%.hh.gch: CXXFLAGS+=$(PCH_BUILD_CXXFLAGS)
$(PCH_OUTPUT_LOCATION)%.hh.gch: %.hh $(PCH_INCLUDES)
	@mkdir -p $(dir $@)
	$(CXX) $(CXXFLAGS) -o $@ -c $<

$(PCH_OUTPUT_LOCATION)%.h.gch: CFLAGS+=$(PCH_BUILD_CFLAGS)
$(PCH_OUTPUT_LOCATION)%.h.gch: %.h $(PCH_INCLUDES)
	@mkdir -p $(dir $@)
	$(CC) $(CFLAGS) -o $@ -c $<

obj/c/%.o: CFLAGS+= $(PCH_USE_CFLAGS)
obj/c/%.o: %.c $(PCH_OUTPUT)
	$(CC) -c $< $(CFLAGS) -o $@

obj/cxx/%.o: CXXFLAGS+= $(PCH_USE_CXXFLAGS)
obj/cxx/%.o: %.cpp $(PCH_OUTPUT)
	$(CXX) -c $< $(CXXFLAGS) -o $@

prof/c/%.o: CFLAGS+= $(PCH_USE_CFLAGS)
prof/c/%.o: %.c $(PCH_OUTPUT)
	$(CC) -c $< $(CFLAGS) -o $@ $(PROF_FLAGS) 
#$(LDFLAGS)

prof/cxx/%.o: CXXFLAGS+= $(PCH_USE_CXXFLAGS)
prof/cxx/%.o: %.cpp $(PCH_OUTPUT)
	$(CXX) -c $< $(CXXFLAGS) -o $@ $(PROF_FLAGS) 
#$(LDFLAGS)

lib$(PROJECT)-release.a: CFLAGS+= $(RELEASE_CFLAGS) $(STATIC_FLAGS) $(STATIC_RELEASE_FLAGS)
lib$(PROJECT)-release.a: CXXFLAGS += $(RELEASE_CXXFLAGS) $(STATIC_FLAGS) $(STATIC_RELEASE_FLAGS)
lib$(PROJECT)-release.a: LDFLAGS += $(RELEASE_LDFLAGS)
lib$(PROJECT)-release.a: $(OBJ)
	$(AR) rcs $@ $^
	$(RANLIB) $@

lib$(PROJECT)-debug.a: CFLAGS+= $(DEBUG_CFLAGS) $(STATIC_FLAGS) $(STATIC_DEBUG_FLAGS)
lib$(PROJECT)-debug.a: CXXFLAGS += $(DEBUG_CXXFLAGS) $(STATIC_FLAGS) $(STATIC_DEBUG_FLAGS)
lib$(PROJECT)-debug.a: LDFLAGS += $(DEBUG_LDFLAGS)
lib$(PROJECT)-debug.a: $(OBJ)
	$(AR) rcs $@ $^ 
	$(RANLIB) $@

lib$(PROJECT)-release.so: CFLAGS+= $(RELEASE_CFLAGS) $(SHARED_FLAGS) $(SHARED_RELEASE_FLAGS)
lib$(PROJECT)-release.so: CXXFLAGS += $(RELEASE_CXXFLAGS) $(SHARED_FLAGS) $(SHARED_RELEASE_FLAGS)
lib$(PROJECT)-release.so: LDFLAGS += $(RELEASE_LDFLAGS)
lib$(PROJECT)-release.so: BINFLAGS += $(RELEASE_BINFLAGS)
lib$(PROJECT)-release.so: $(OBJ)
	$(CXX) -shared $^ $(BINFLAGS) $(CXXFLAGS) -o $@ $(LDFLAGS)
	$(STRIP) $@

lib$(PROJECT)-debug.so: CFLAGS+= $(DEBUG_CFLAGS) $(SHARED_FLAGS) $(SHARED_DEBUG_FLAGS)
lib$(PROJECT)-debug.so: CXXFLAGS += $(DEBUG_CXXFLAGS) $(SHARED_FLAGS) $(SHARED_DEBUG_FLAGS)
lib$(PROJECT)-debug.so: LDFLAGS += $(DEBUG_LDFLAGS)
lib$(PROJECT)-debug.so: BINFLAGS += $(DEBUG_BINFLAGS)
lib$(PROJECT)-debug.so: $(OBJ)
	$(CXX) -shared $^ $(BINFLAGS) $(CXXFLAGS) -o $@ $(LDFLAGS)

lib$(PROJECT).a: lib$(PROJECT)-release.a
	ln -f $< $@

lib$(PROJECT).so: LDFLAGS+= -Wl,-soname,lib$(PROJECT).so.$(VERSION_MAJOR)
lib$(PROJECT).so: lib$(PROJECT)-release.so
	ln -f $< $@.$(VERSION)
	ln -sf $@.$(VERSION) $@.$(VERSION_MAJOR)
	ln -sf $@.$(VERSION_MAJOR) $@

clean-source:
	find -O2 {obj,prof}/ -type f -exec rm {} +

clean-rebuild: clean-source
	find $(INCLUDE) $(PCH_OUTPUT_LOCATION) -type f -name \*.gch -exec rm {} +


clean: clean-rebuild
	rm -f lib$(PROJECT){,-{release,debug,pgo}}.{a,so{,.*}}
	rm -f $(PROJECT)-test

clean-full: clean
	rm -rf {obj,prof}

install:
	install -d $(DESTDIR)$(PREFIX)/lib/
	install -m 644 lib$(PROJECT).a $(DESTDIR)$(PREFIX)/lib/
	install -s -m 755 lib$(PROJECT).so.$(VERSION) $(DESTDIR)$(PREFIX)/lib/
	ln -sf lib$(PROJECT).so.$(VERSION) $(DESTDIR)$(PREFIX)/lib/lib$(PROJECT).so.$(VERSION_MAJOR)
	ln -sf lib$(PROJECT).so.$(VERSION_MAJOR) $(DESTDIR)$(PREFIX)/lib/lib$(PROJECT).so
	install -d $(DESTDIR)$(PREFIX)/include/$(INCLUDE_PREFIX)
	install -m 644 $(wildcard $(INCLUDE)/*.*) $(DESTDIR)$(PREFIX)/include/$(INCLUDE_PREFIX)
uninstall:
	-rm $(DESTDIR)$(PREFIX)/lib/lib$(PROJECT).{a,so{,.*}}
	cd $(INCLUDE) && find . -type f | xargs -I {} rm "$(DESTDIR)$(PREFIX)/include/$(INCLUDE_PREFIX){}"
	[[ -d "$(DESTDIR)$(PREFIX)/include/$(INCLUDE_PREFIX)" ]] && \
		rmdir $(DESTDIR)$(PREFIX)/include/$(INCLUDE_PREFIX) || :

$(PROJECT)-test: lib$(PROJECT)-$(TARGET)
	$(CXX) $(CXXFLAGS) $(TEST_CFLAGS) $(TEST_CXXFLAGS) src/test/*.cpp -o $@ -l:$< $(LDFLAGS) $(TEST_LDFLAGS)