# SPDX-License-Identifier: AGPL-3.0-or-later
# Copyright (C) 2026 David (Corey) Herrera
# Foundry Blackhole backend — Phase 6's MEMORY op family (EMBEDDING + KV_APPEND).
#
# TWO BINARIES, TWO BUILD TREES, AND THE FENCE BETWEEN THEM IS THE POINT.
#
#   make -f Makefile.memory all                 -> build-memory/test_memory_ops
#     The OFF-HARDWARE gate. Built WITHOUT -DFT_TT_PHASE1B_NOC_ENABLED, so every device
#     call in the objects it links returns FT_TT_EPERM before touching the NoC. That is not
#     a formality: it means this binary cannot reach the card even if it is wrong, which is
#     what makes it safe to run while somebody else owns the silicon.
#
#   make -f Makefile.memory PHASE1B=1 card      -> build-memory-card/test_phase6_memory
#     The ON-CARD acceptance. PHASE1B=1 is mandatory and deliberate (spec §7.3 — a wrong
#     NoC coordinate hangs the chip and costs a physical power cycle).
#
# THE SHIM OBJECT IS NOT AN ACCIDENT. tests/ft_tt_memory_device_shim.c compiles the
# arithmetic half of kernels/resident_memory_protocol.h for the host, so the off-hardware
# suite scores THE KERNEL'S OWN face-offset and run functions rather than a C transcription
# of them — a transcription would have agreed with itself, and face order is precisely the
# mistake that is invisible to any metric computed through the same wrong code
# (docs/tt_kernel_model.md §9). The header includes <stdint.h>/<stdbool.h> and nothing else,
# which is a property to KEEP.
#
# It was a C++ object until 2026-08-13, purely for `extern "C"` linkage over what was then a
# C++ header. Phase 6.6 made that header C99, so the shim is plain C now and this suite
# needs no C++ compiler at all.
#
# Run the off-hardware gate from this directory, so FT_TT_KERNEL_DIR's default resolves:
#   ./build-memory/test_memory_ops
#
# Run the on-card acceptance (detached, and never with these cores in use by another
# worker):
#   FT_TT_FIRMWARE_DIR=../programs/memory FT_TT_TENSIX_INDEX=30 \
#     ./build-memory-card/test_phase6_memory ../programs/memory
#
# The program directory is a RELOCATE_KERNEL_TEXT=1 build. It has FIVE images because
# build_tensix_program.sh builds five; only brisc.elf and ncrisc.elf are ever loaded, and
# noop_compute.c is there to satisfy the builder rather than to run — this op has NO
# compute thread and its enables mask is DM0|DM1 (kernels/resident_memory_protocol.h says
# why that is safe rather than merely small).
#
#   RELOCATE_KERNEL_TEXT=1 READER_CTARGS=16,0,1,2048,1,17 COMPUTE_CTARGS=0 \
#     WRITER_CTARGS=16,2048,1,17 \
#     build_tensix_program.sh kernels/resident_gather_reader.c \
#         kernels/noop_compute.c kernels/resident_place_writer.c \
#         programs/memory
#
# THE RECIPE ABOVE WAS WRONG UNTIL 2026-08-14 — it omitted the trailing CB_OUT_KV on both
# lists, giving five/three ctargs where the kernels want six/four. Derived, not guessed:
# resident_gather_reader.c:70-75 declares ft_tensix_compile_time_arg(0..5) as
# CB_OUT_EMBEDDING / CB_STAGE / CB_INDEX / TILE_BYTES / WAIT_MODE / CB_OUT_KV, and
# resident_place_writer.c:69-72 declares (0..3) as CB_OUT_EMBEDDING / TILE_BYTES /
# WAIT_MODE / CB_OUT_KV. `ft_tensix_compile_time_arg()` has no bounds check, so a short
# list does not fail to build — it reads past the array, which is the case
# resident_gather_reader.c:43 documents in its own header.
#
# No CB_TABLE: a `.cbtable` generates ft_tensix_cb_formats.h for the COMPUTE images, and this
# program has no compute kernel. The data-movement images learn their tile size from
# READER_CTARGS / WRITER_CTARGS, and those two MUST agree.

CC      ?= cc
SRC_DIR := ..

# The on-card harness opens the card through the product transport seam.
include hardware_session_sources.mk
CONVERT := ../convert
BUILD   := build-memory
CARD_BUILD := build-memory-card

WARNINGS := -Wall -Wextra -Werror -pedantic \
            -Wshadow -Wconversion -Wstrict-prototypes -Wformat=2 \
            -fstack-protector-strong

CFLAGS   := -std=c99 -O1 -g $(WARNINGS) -I$(SRC_DIR)
LDLIBS   := -lm

# `card` and `all` are separate invocations on purpose: the PHASE1B flag changes what the
# backend objects DO, so one command line cannot honestly produce both.
ifeq ($(filter card,$(MAKECMDGOALS)),card)
ifneq ($(filter all,$(MAKECMDGOALS)),)
$(error build `all` and `card` in separate invocations — they are different configurations)
endif
ifndef PHASE1B
$(error the on-card acceptance drives the card: build with PHASE1B=1, or every device call \
is refused before it reaches the NoC)
endif
CFLAGS += -DFT_TT_PHASE1B_NOC_ENABLED
endif

# Header dependency tracking. Without it, editing a HEADER rebuilds nothing and the link
# silently mixes objects compiled against two different struct layouts — not a build error,
# a wrong offset.
DEPFLAGS := -MMD -MP

BACKEND := $(SRC_DIR)/ft_tt_kmd.c \
           $(SRC_DIR)/ft_tt_arch.c \
           $(SRC_DIR)/ft_tt_arch_blackhole.c \
           $(SRC_DIR)/ft_tt_device.c \
           $(SRC_DIR)/ft_tt_tlb.c \
           $(SRC_DIR)/ft_tt_arc_telemetry.c \
           $(SRC_DIR)/ft_tt_coord.c \
           $(SRC_DIR)/ft_tt_noc.c \
           $(SRC_DIR)/ft_tt_buffer.c \
           $(SRC_DIR)/ft_tt_elf.c \
           $(SRC_DIR)/ft_tt_kernel_blob.c \
           $(SRC_DIR)/ft_tt_udrv_claim.c \
           $(SRC_DIR)/ft_tt_tensix_claim.c \
           $(SRC_DIR)/ft_tt_launch.c \
           $(SRC_DIR)/ft_tt_tilize.c \
           $(SRC_DIR)/ft_tt_op_dispatch.c \
           $(SRC_DIR)/ft_tt_op_memory.c \
           $(SRC_DIR)/ft_tt_queue.c \
           $(SRC_DIR)/ft_tt_group.c \
           $(SRC_DIR)/ft_tt_watchdog.c

SUPPORT      := ft_tt_test_report.c
CARD_SUPPORT := $(HARDWARE_SESSION_SOURCES) ft_tt_program_directory.c

TARGET      := $(BUILD)/test_memory_ops
CARD_TARGET := $(CARD_BUILD)/test_phase6_memory
SHIM        := $(BUILD)/ft_tt_memory_device_shim.o
CARD_SHIM   := $(CARD_BUILD)/ft_tt_memory_device_shim.o

.PHONY: all card red clean

all: $(TARGET)

card: $(CARD_TARGET)

$(BUILD) $(CARD_BUILD):
	mkdir -p $@

$(SHIM): ft_tt_memory_device_shim.c ../kernels/resident_memory_protocol.h | $(BUILD)
	$(CC) $(CFLAGS) -c -o $@ $<

$(CARD_SHIM): ft_tt_memory_device_shim.c ../kernels/resident_memory_protocol.h | $(CARD_BUILD)
	$(CC) $(CFLAGS) -c -o $@ $<

# Whole-program links, not incremental objects. `red` deletes the tree first, because
# Apple's GNU make 3.81 compares mtimes at one-second granularity and will silently skip a
# rebuild of a source edited in the same second as its object — which has already produced
# three "runs" of the same stale binary here (best-practices §9, dev notes §3).
$(TARGET): test_memory_ops.c $(BACKEND) $(SUPPORT) $(SHIM) | $(BUILD)
	$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(BACKEND) $(SUPPORT) $(SHIM) $(LDLIBS)

$(CARD_TARGET): test_phase6_memory.c $(BACKEND) $(SUPPORT) $(CARD_SUPPORT) $(CARD_SHIM) \
                | $(CARD_BUILD)
	$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(BACKEND) $(SUPPORT) $(CARD_SUPPORT) $(CARD_SHIM) \
	    $(LDLIBS)

red: clean all

clean:
	rm -rf $(BUILD) $(CARD_BUILD)

-include $(wildcard $(BUILD)/*.d) $(wildcard $(CARD_BUILD)/*.d)