SHELL := $(shell which bash)
.ONESHELL:

COL := $(patsubst %.gp,%,$(wildcard *.gp))
DAT := $(COL:=.dat)
DIS := $(shell grep '^ID=.*$$' /etc/os-release | cut -d'=' -f2)
MAX := 32
OPS := $(shell echo $$(( 2**15 )))
SEQ := cas old stack std
SRC := $(wildcard *.cc)
SVG := $(COL:=.svg)
T   := $(shell echo $$(( $(shell nproc) / 2 )))
VER := $(shell grep '^VERSION_ID=.*$$' /etc/os-release | cut -d'=' -f2 | tr -d '"')

EXE := $(basename $(SRC))
LIB := ./rpmalloc/$(DIS)/$(VER)
LOG := $(foreach e,$(EXE),$(foreach o,$(OPS),$(foreach t,$(shell for ((i=1; i<=$(MAX); i*=2)); do echo $$i; done),$(e)-o$(o)-t$(t).log)))
STA := $(LOG:.log=.sta)
TXT := $(foreach e,$(EXE),$(foreach o,$(OPS),$(e)-o$(o)-t$(T).txt))
W1  := $(shell m=0; for i in $(EXE); do l=$${#i}; (( l > m )) && m=$$l; done; echo $$((m + 2)))
W2  := 10
W3  := $(shell echo $$(( $(W1) + ( $(W2) + 1 ) * $(words $(COL)) )))

CXXFLAGS := -O3 -std=c++23 -Wall # -g
LDFLAGS := -L$(LIB)
LDLIBS := -latomic -ldl -lrpmalloc
# $(filter-out stm%,$(EXE)): CXXFLAGS += -fanalyzer -fsanitize={null,thread,undefined} # address,leak
$(filter stm%,$(EXE)): CXXFLAGS += -fgnu-tm

export LC_ALL=C.UTF-8
export LD_LIBRARY_PATH=$$LD_LIBRARY_PATH:$(LIB)

all:
	@if make -qs $(TXT); then
		make -s sort
	else
		make -s stat
	fi

check:
	@printf '%$(W1)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s\n' | tr ' ' '-'
	printf '%$(W1)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s\n' program drd futex helgrind memcheck rpmalloc
	printf '%$(W1)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s\n' | tr ' ' '-'
	for i in $(EXE); do
		printf '%$(W1)s' $$i
		make --no-print-directory -s $${i} &> /dev/null
		make --no-print-directory $${i}.{drd,futex,helgrind,memcheck,rpmalloc}
		echo
	done
	printf '%$(W1)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s\n' | tr ' ' '-'
	echo "drd, helgrind & thread sanitizer produces many false positives."
	echo "Some of them can be suppressed as seen in spinlock.h"
	printf '%$(W1)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s%$(W2)s\n' | tr ' ' '-'

clean:
	-rm -fr $(DAT) $(EXE) $(SVG) *.log *.sta *.txt .*.d core* *~
	-find -maxdepth 2 -mindepth 2 -name makefile -execdir make $@ \;
	-killall -KILL $(EXE) eog loupe &> /dev/null || true

dat: $(DAT)

exe: $(EXE)

help:
	@echo "export LD_LIBRARY_PATH=$$LD_LIBRARY_PATH:$(LIB)"

key:
	@printf '%$(W3)s\n' | tr ' _' '- '
	echo " program: program name + sequential(.) or parallel(:)"
	echo "     ops: push()'s + pop()'s"
	echo " elapsed: elapsed time (ms)"
	echo "    task: task-clock time (ms)"
	echo "  ratio1: ops / elapsed (ops/ms)"
	echo "  ratio2: ops / task (ops/ms)"

lockfree:
	@IFS=''
	while read -r line; do 
		for i in asp mutex semaphore shared_ptr spinlock stm stm2; do
			if echo "$$line" | grep -q $$i; then
				line=''
				break
			fi
		done
		if [[ -n "$$line" ]]; then
			echo "$$line"
		fi
	done < <(make --no-print-directory parallel)

parallel:
	@make --no-print-directory sort | grep -v '[[:alnum:]]\. '

sequential:
	@make --no-print-directory sort | grep -v '[[:alnum:]]: '

sort: $(filter %-t$(T).txt,$(TXT)) | key
	@readarray -t F < <(head -n1 $< | tr ' ' '\n' | sed '/^$$/d')
	declare -a S=('' 'n' 'n' 'n' 'n' 'n')
	declare -a R=(''  ''  ''  '' 'r' 'r')
	for ((i=0; i<$${#F[@]}; ++i)); do
		printf '%$(W3)s\n' | tr ' ' '-'
		if [[ "$$i" == "0" ]]; then
			w=$(W1)
		else
			w=$(W2)
		fi
		in=$$(printf "%$${w}s" "$${F[$$i]}")
		out=$$(printf "%$${w}s" "[$${F[$$i]:0:$$((w - 2))}]")
		head -n1 -q $< | sed "s/$${in}/$${out}/"
		printf '%$(W3)s\n' | tr ' ' '-'
		for t in $^; do
			tail -n1 $$t
		done | sort -bf -k$$((i+1)),$$((i+1))$${S[$$i]}$${R[$$i]} -k$${#F[@]},$${#F[@]}$${S[$${#F[@]}]}$${R[$${#F[@]}]}
	done
	printf '%$(W3)s\n' | tr ' ' '-'

stat: | key
	@for i in $(filter %-t$(T).txt,$(TXT)); do
		make -s $$i &> /dev/null
		if echo $$i | grep $(firstword $(EXE)) &> /dev/null; then
			printf '%$(W3)s\n' | tr ' ' '-'
			head -n 1 $$i
			printf '%$(W3)s\n' | tr ' ' '-'
		fi
		tail -n 1 $$i
	done
	printf '%$(W3)s\n' | tr ' ' '-'

svg: $(SVG)
	@for i in $^; do
		xdg-open $$i &> /dev/null &
	done

%: %.cc
	$(CXX) $(CPPFLAGS) $(CXXFLAGS) -MMD -MP -MF .$*.d ./$< $(LDFLAGS) $(LOADLIBES) $(LDLIBS) -o $@

%.dat:
	@printf '%80s\n _$@\n%80s\n' | tr ' _' '# '
	echo "program $(EXE)" | tee $@
	for (( i = 1; i <= $(MAX); i *= 2 )); do 
		echo -n "$$i "
		for e in $(EXE); do
			make -s $${e}-o$(OPS)-t$${i}.txt &> /dev/null
			for (( j = 0; j <= $(words $(COL)); ++j )); do 
				c=$$(cat $${e}-o$(OPS)-t$${i}.txt | tr -s ' ' | cut -d' ' -f$$((j+2)))
				if echo "$$c" | grep "$*" &> /dev/null; then
					echo -en "$$c " | tail -n 1
				fi
			done
		done
		echo
	done | tee -a $@

%.drd: %
	@valgrind --tool=drd --error-exitcode=1 --exit-on-first-error=yes -- ./$< -o 1000 -t $(T) &> /dev/null
	if [ $$? -eq 0 ]; then
		echo -en "\e[1;32m$$(printf '%$(W2)s' PASSED)\e[0m"
	else
		echo -en "\e[1;31m$$(printf '%$(W2)s' FAILED)\e[0m"
	fi

%.futex: %
	@[[ " $(SEQ) " =~ " $< " ]] && threads=1 || threads=$(T)
	read -r time seconds usecs calls errors syscall < <(strace -c -e futex -f ./$< -o $(OPS) -t $$threads |& grep futex) &> /dev/null
	if (( calls == 0 )); then
		echo -en "\e[1;32m$$(printf '%$(W2)s' $$calls)\e[0m"
	else
		echo -en "\e[1;31m$$(printf '%$(W2)s' $$calls)\e[0m"
	fi

%.helgrind: %
	@valgrind --tool=helgrind --error-exitcode=1 --exit-on-first-error=yes -- ./$< -o 1000 -t $(T) &> /dev/null
	if [ $$? -eq 0 ]; then
		echo -en "\e[1;32m$$(printf '%$(W2)s' PASSED)\e[0m"
	else
		echo -en "\e[1;31m$$(printf '%$(W2)s' FAILED)\e[0m"
	fi

%.memcheck: %
	@[[ " $(SEQ) " =~ " $< " ]] && threads=1 || threads=2
	[[ " stm stm2 " =~ " $< " ]] && undef=no || undef=yes
	LD_PRELOAD=$$(gcc -print-file-name=libc.so.6) valgrind --tool=memcheck --run-libc-freeres=no --run-cxx-freeres=no --error-exitcode=1 --exit-on-first-error=yes --undef-value-errors=$$undef -- ./$< -o 1000 -t $$threads &> /dev/null
	if [ $$? -eq 0 ]; then
		echo -en "\e[1;32m$$(printf '%$(W2)s' PASSED)\e[0m"
	else
		echo -en "\e[1;31m$$(printf '%$(W2)s' FAILED)\e[0m"
	fi

%.rpmalloc: %
	@if ldd ./$< | grep -q librpmalloc.so; then
		echo -en "\e[1;32m$$(printf '%$(W2)s' FOUND)\e[0m"
	else
		echo -en "\e[1;31m$$(printf '%$(W2)s' MISSING)\e[0m"
	fi

%.txt: %.log %.sta
	@log="[$$(cat $*.log | tr '\n' ',')]"
	ops=$$(python -c "import statistics; print(int(statistics.mean($$log)))")
	elapsed=$$(grep -m1 duration_time $*.sta | cut -d';' -f1)
	elapsed=$$(python -c "print('%g' % ($$elapsed / 1000000))") # ns -> ms
	ratio1=$$(python -c "print('%g' % ($$ops / $$elapsed))")
	task=$$(grep -m1 task-clock $*.sta | cut -d';' -f1)
	ratio2=$$(python -c "print('%g' % ($$ops / $$task))")
	exe=$$(echo $* | sed -n 's/\([^-]*\)-.*/\1/p')
	if [[ " $(SEQ) " =~ " $$exe " ]]; then
		exe+='.'
	else
		exe+=':'
	fi
	printf '%$(W1).$(W1)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s\n' program ops elapsed task ratio1 ratio2 > $@
	printf '%$(W1).$(W1)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s %$(W2).$(W2)s\n' $$exe $$(printf '%.0f' $$ops) "$$(printf '%.1f' $$elapsed)" "$$(printf '%.1f' $$task)" "$$(printf '%.0f' $$ratio1)" "$$(printf '%.0f' $$ratio2)" >> $@

%.svg: %.gp | %.dat
	@gnuplot -e "basename='$*'" $< > $@

-include .*.d

.PHONY: all check clean dat exe help key lockfree parallel sequential sort stat svg
.PRECIOUS: $(DAT) $(LOG) $(STA) $(TXT) $(SVG)

.SECONDEXPANSION:

%.log %.sta: $$(firstword $$(subst -, ,$$*))
	@o=$$(echo $* | sed -n 's/.*-o\([[:digit:]]*\).*/\1/p')
	t=$$(echo $* | sed -n 's/.*-t\([[:digit:]]*\).*/\1/p')
	if [[ -z "$$o" || -z "$$t" ]]; then
		msg="Error: invalid parameters for ./$< -o$$o -t$$t"
		notify-send "$$msg"
		echo "$$msg" >&2
		exit 1
	fi
	if [[ $< -nt $*.sta ]]; then
		[[ " $(SEQ) " =~ " $< " ]] && (( o = o * t )) && t=1
		perf stat -e 'duration_time,task-clock' -r 10 -x ';' -- ./$< -o $$o -t $$t 1> $*.log 2> $*.sta
	fi