# if there is Makefile.local file then include it
ifneq (, $(wildcard Makefile.local))
include Makefile.local
endif

export CC = gcc
export LD = g++

arch = $(shell uname -m)

ifdef DEBUG
CFLAGS_COMMON = -O1 -g
else
CFLAGS_COMMON = -O3
endif
CFLAGS_COMMON += -fPIC -Wall -Wno-unused-parameter -W -Wshadow -Wpointer-arith -Wwrite-strings -Wuninitialized

CFLAGS_COMMON += -DRIB_DIR='"$(PWD)"'
CFLAGS = $(CFLAGS_COMMON) -I $(PWD) -I benchmarks/
export LDFLAGS = -lc -lm 
THIRDPARTYDIR = ../thirdparty

OBJ_GZSTREAM = $(THIRDPARTYDIR)/gzstream.o

# Optional FFTW support
ifdef WITH_FFTW
FFTWDIR = $(THIRDPARTYDIR)/fftw/dist
CFLAGS += -I $(FFTWDIR)/include -DENABLE_FFTW
LDFLAGS += $(FFTWDIR)/lib/libfftw3.a
endif

# Optional PAPI support
ifdef WITH_PAPI
PAPIDIR = $(THIRDPARTYDIR)/papi/dist
LIBPAPI = $(PAPIDIR)/lib/libpapi.a
CFLAGS += -I $(PAPIDIR)/include -DENABLE_PAPI
LDFLAGS += $(LIBPAPI)
endif

# Optional nasm-generated NOPs for cache_artificial code access
ifdef WITH_NASM
NASMOBJ = ictrash.o
CFLAGS += -DENABLE_NASM
endif

# Optional valgrind support
ifdef VALGRIND_PATH
CFLAGS += -DENABLE_VALGRIND -I$(VALGRIND_PATH)/include/
endif

CXXFLAGS = $(CFLAGS)
export CXXFLAGS_BENCHMARKS = $(CFLAGS_BENCHMARKS)

SRC_COMPRESS = $(THIRDPARTYDIR)/compress/compress42.c
OBJ_COMPRESS = $(THIRDPARTYDIR)/compress/compress42.o

SRC_FFT = $(THIRDPARTYDIR)/fft/fft.c
OBJ_FFT = $(THIRDPARTYDIR)/fft/fft.o

SRC_SFMT = $(THIRDPARTYDIR)/SFMT/SFMT.c
OBJ_SFMT = $(THIRDPARTYDIR)/SFMT/SFMT.o
CFLAGS_SFMT = -msse2 -DHAVE_SSE2 -DMEXP=607 -O3

SRCS_BENCHMARKS = $(shell find benchmarks/ -name '*.cpp')
OBJS_BENCHMARKS = $(SRCS_BENCHMARKS:.cpp=.o)

SRCS = rib_core.cpp rib.cpp benchmark.cpp params.cpp config.cpp ribplot.cpp \
	$(SRC_COMPRESS) $(SRC_SFMT) $(SRC_FFT) $(SRCS_BENCHMARKS) stack_profile.cpp va2pa.cpp

OBJS_COMMON = rib_core.o benchmark.o params.o config.o \
	$(OBJ_COMPRESS) $(OBJ_SFMT) stack_profile.o va2pa.o $(NASMOBJ)
OBJS_STANDARD = $(OBJS_COMMON) $(OBJ_FFT)
OBJS_TC = $(OBJS_COMMON) benchmarks/cache_tc.o

OBJS_RIB = $(OBJS_STANDARD) rib.o 
OBJS_RIBPLOT = $(OBJS_STANDARD) ribplot.o
OBJS = $(OBJS_STANDARD) $(OBJS_BENCHMARKS) rib.o ribplot.o

BENCHLIB = benchmarks.a

TARGETS = rib ribplot rib_tc.a dropcaches $(BENCHLIB)

all: .depend $(TARGETS) Makefile

$(BENCHLIB): $(OBJS_BENCHMARKS)
	ar rc $@ $(OBJS_BENCHMARKS)

fftw:
	cd fftw-3.1.1 && ./build.sh

papilib:
	echo "Please run ./build.sh inside papi/"

rib: $(OBJS_RIB) $(BENCHLIB)
	$(LD) -o $@ $(OBJS_RIB) -Wl,--whole-archive $(BENCHLIB) -Wl,--no-whole-archive $(LDFLAGS)
	
rib_tc.a: $(OBJS_TC)
	ar r $@ $(OBJS_TC) $(LIBPAPI)
	
ribplot: $(OBJS_RIBPLOT) $(BENCHLIB) $(OBJ_GZSTREAM)
	$(LD) -o $@ $(OBJS_RIBPLOT) $(OBJ_GZSTREAM) -Wl,--whole-archive $(BENCHLIB) -Wl,--no-whole-archive $(LDFLAGS) -lz -Wl,--defsym,_tc_filter_bench=0

ictrash.o:	ictrash.asm
ifeq ($(arch),x86_64)
	nasm -f elf64 ictrash.asm
else
	nasm -f elf ictrash.asm
endif

# development tools
#papi: tools/papi.o
#	$(LD) ${CFLAGS} -o $@ tools/papi.o $(LDFLAGS)

coltest: tools/coltest.c
	${CC} ${CFLAGS} -o $@ tools/coltest.c $(LDFLAGS)
	
coltest2: tools/coltest2.c
	${CC} ${CFLAGS} -I ./papi/dist/include -o $@ tools/coltest2.c -L./papi/dist/lib -static -lpapi
	
scaletest: tools/scaletest.o $(OBJ_FFT)
	$(LD) -o $@ tools/scaletest.o $(OBJ_FFT)
	
dropcaches: tools/dropcaches.c
	${CC} ${CFLAGS} -o $@ tools/dropcaches.c

tools/va2patest.o: tools/va2patest.cpp
	${CC} -c ${CFLAGS} -o $@ $<
	
va2patest: tools/va2patest.o va2pa.o
	$(LD) -o $@ $< va2pa.o $(LDFLAGS)
	
$(OBJ_SFMT): $(SRC_SFMT)
	${CC} -c ${CFLAGS} ${CFLAGS_SFMT} -o $@ $<

.depend:
	-makedepend -f - -- $(CFLAGS) -- $(SRCS) > .depend 2>/dev/null

clean:
	rm -f *.o $(OBJ_GZSTREAM) $(OBJ_FFT) $(OBJ_COMPRESS) *.a benchmarks/*.o benchmarks/*.a

distclean: clean
	rm -f .depend
	rm -f $(TARGETS)
	
# DO NOT DELETE
include .depend