glazer

CURDIR   := $(shell pwd)
BASEDIR	 := $(abspath $(dir $(CURDIR)))
CXX      ?= g++
PRIV_DIR ?= $(if $(REBAR_BARE_COMPILER_OUTPUT_DIR),$(REBAR_BARE_COMPILER_OUTPUT_DIR),$(BASEDIR))/priv
OBJ_DIR  ?= $(BASEDIR)/obj
DEBUG    ?= 0

SRC      := glazer_nif.cpp
OBJ      := $(OBJ_DIR)/glazer_nif.o
TARGET   := $(PRIV_DIR)/glazer.so

ERL_ERTS_INCLUDE := $(shell erl -noshell -noinput \
  -eval "io:format(\"~ts/erts-~ts/include\",[code:root_dir(),erlang:system_info(version)]),halt(0).")
ERL_EI_INCLUDE   := $(shell erl -noshell -noinput \
  -eval "io:format(\"~ts\",[code:lib_dir(erl_interface,include)]),halt(0).")
ERL_EI_LIB       := $(shell erl -noshell -noinput \
  -eval "io:format(\"~ts\",[code:lib_dir(erl_interface,lib)]),halt(0).")

CXXFLAGS := -std=c++23 -fPIC -Wall -finline-functions -fvisibility=hidden \
            -I$(ERL_ERTS_INCLUDE) -I$(ERL_EI_INCLUDE) -I.

# Detect compiler family for LTO flags.
# AppleClang uses ld64 which handles thin-LTO natively — no -fuse-ld=lld.
# Upstream Clang on Linux emits LLVM bitcode that requires lld.
# GCC uses -flto=auto with GNU ld/gold.
CXX_ID := $(shell $(CXX) --version 2>&1 | head -1)
ifneq ($(findstring Apple,$(CXX_ID)),)
  LTO_FLAGS  := -flto=thin
  LINK_FLAGS :=
else ifneq ($(findstring clang,$(CXX_ID)),)
  LTO_FLAGS  := -flto=thin
  LINK_FLAGS := -fuse-ld=lld
else
  LTO_FLAGS  := -flto=auto -fno-fat-lto-objects
  LINK_FLAGS :=
endif

# Profile-Guided Optimisation (PGO) — lets the compiler use real branch
# frequencies and call counts to lay out hot paths, inline more aggressively,
# and reorder basic blocks for better instruction-cache usage.  Typical gain
# on a JSON-heavy workload is 5–15 % over plain -O3.
#
# Three-step workflow (or just run `make pgo` from the project root):
#
#   1. make PGO=generate   — build an instrumented .so that writes profile
#                            data (*.gcda) into obj/pgo/ at runtime
#   2. run a representative workload (e.g. `rebar3 eunit`, a benchmark)
#   3. make PGO=use        — rebuild optimised using the collected profile
#
# Profile data lives in $(OBJ_DIR)/pgo/ and is removed by `make clean`.
# -fprofile-correction tolerates minor counter mismatches (e.g. if the
# workload was interrupted) instead of aborting the build.
PGO      ?= 0
PGO_DIR  := $(OBJ_DIR)/pgo
ASAN     ?= 0

ifeq ($(DEBUG),1)
  CXXFLAGS += -O0 -g
else ifeq ($(ASAN),1)
  # Compile with ASan instrumentation but do NOT link -fsanitize=address.
  # Linking it would embed or reference the runtime inside the NIF .so; when
  # the Erlang VM dlopen's that .so the runtime loads too late and ASan aborts.
  # Instead all __asan_* symbols stay undefined in the .so and are resolved
  # from the runtime preloaded via LD_PRELOAD (Linux only — see Makefile).
  CXXFLAGS += -O1 -g -fsanitize=address -fno-omit-frame-pointer
else
  CXXFLAGS += -O3 -DNDEBUG -march=native -mtune=native $(LTO_FLAGS)
  LDFLAGS  += $(LINK_FLAGS)
  ifeq ($(PGO),generate)
    CXXFLAGS += -fprofile-generate=$(PGO_DIR)
    LDFLAGS  += -fprofile-generate=$(PGO_DIR)
  else ifeq ($(PGO),use)
    CXXFLAGS += -fprofile-use=$(PGO_DIR) -fprofile-correction -Wno-missing-profile
    LDFLAGS  += -fprofile-use=$(PGO_DIR)
  endif
endif

LDFLAGS += -shared -L$(ERL_EI_LIB) -lei
ifeq ($(shell uname -s),Darwin)
  LDFLAGS += -undefined dynamic_lookup
endif

.PHONY: all clean force

all: $(TARGET)

$(TARGET): $(OBJ) | $(PRIV_DIR)
	$(CXX) $(OBJ) $(LDFLAGS) -o $@

# When PGO=use the .o is stale by timestamp (same source, same flags string)
# but must be recompiled against the newly collected profile data — force it.
# Note: GCC writes .gcda files under a path derived from the source file's
# absolute path, so the layout inside $(PGO_DIR) is nested.  This is harmless:
# GCC searches $(PGO_DIR) recursively on -fprofile-use, and clean wipes the tree.
$(OBJ): $(SRC) $(if $(filter use,$(PGO)),force) \
        | $(OBJ_DIR) $(if $(filter generate use,$(PGO)),$(PGO_DIR))
	$(CXX) $(CXXFLAGS) -c $< -o $@

$(PRIV_DIR) $(OBJ_DIR) $(PGO_DIR):
	@mkdir -p $@

clean:
	rm -f $(OBJ) $(TARGET)
	rm -rf $(PGO_DIR)