// selfhost/cmd/6a/obj.ww — port of cmd/6a/obj.c.
//
// Emit a tiny ELF64 relocatable object. Layout (in file order):
//   [0] ELF header
//   [1] Section .text (program bytes)
//   [2] Section .rela.text (relocations)
//   [3] Section .symtab
//   [4] Section .strtab
//   [5] Section .shstrtab
//   [6] Section header table
//
// Symtab indices: 0 = STN_UNDEF, 1.. = our syms. Only GLOBAL symbols.

use os;
use mem;
use types;

// ---- ELF constants ----------------------------------------------------
def ELFCLASS64:   u8  = 2u8;
def ELFDATA2LSB:  u8  = 1u8;
def EV_CURRENT_W: u32 = 1u32;
def ET_REL_W:     u16 = 1u16;
def EM_X86_64_W:  u16 = 62u16;

def SHT_NULL_C:     u32 = 0u32;
def SHT_PROGBITS_C: u32 = 1u32;
def SHT_SYMTAB_C:   u32 = 2u32;
def SHT_STRTAB_C:   u32 = 3u32;
def SHT_RELA_C:     u32 = 4u32;

def SHF_ALLOC:      u64 = 2u64;
def SHF_EXECINSTR:  u64 = 4u64;
def SHF_INFO_LINK:  u64 = 64u64;	// 0x40

def STB_GLOBAL:     u8 = 1u8;
def STT_NOTYPE:     u8 = 0u8;
def STT_FUNC:       u8 = 2u8;

// Sizes of fixed structures.
def EHDR_SZ: u64 = 64u64;
def SHDR_SZ: u64 = 64u64;
def SYM_SZ:  u64 = 24u64;
def RELA_SZ: u64 = 24u64;

// ---- LE byte writers (own the bytes — write into a *u8 + offset) ----

fn wr_u8(p: *u8, off: u64, v: u8) void = { p[off] = v; };
fn wr_u16(p: *u8, off: u64, v: u16) void = {
	p[off]        = (v & 255u16): u8;
	p[off + 1u64] = ((v >> 8u16) & 255u16): u8;
};
fn wr_u32(p: *u8, off: u64, v: u32) void = {
	p[off]        = (v & 255u32): u8;
	p[off + 1u64] = ((v >> 8u32)  & 255u32): u8;
	p[off + 2u64] = ((v >> 16u32) & 255u32): u8;
	p[off + 3u64] = ((v >> 24u32) & 255u32): u8;
};
fn wr_u64(p: *u8, off: u64, v: u64) void = {
	wr_u32(p, off,        (v & 4294967295u64): u32);
	wr_u32(p, off + 4u64, ((v >> 32u64) & 4294967295u64): u32);
};

// ---- growable byte buffer ---------------------------------------------

type buf = struct {
	a:    *arena,
	p:    *u8,
	n:    u64,
	cap:  u64,
};

fn buf_init(b: *buf, a: *arena) void = {
	b.a = a;
	b.cap = 256u64;
	b.n = 0u64;
	b.p = amalloc(a, b.cap): *u8;
};

fn buf_grow(b: *buf, need: u64) void = {
	if (b.n + need <= b.cap) { return; };
	let nc: u64 = b.cap;
	for (nc < b.n + need) { nc = nc * 2u64; };
	let np: *u8 = amalloc(b.a, nc): *u8;
	let i: u64 = 0u64;
	for (i < b.n) { np[i] = b.p[i]; i += 1u64; };
	b.p = np;
	b.cap = nc;
};

fn buf_putb(b: *buf, src: *u8, n: u64) void = {
	buf_grow(b, n);
	let i: u64 = 0u64;
	for (i < n) { b.p[b.n + i] = src[i]; i += 1u64; };
	b.n += n;
};

// Write a NUL-terminated C-string copy of `s` into b. Returns offset
// where it started (suitable for st_name / sh_name fields).
fn buf_put_cstr(b: *buf, s: str) u32 = {
	let off: u32 = b.n: u32;
	buf_grow(b, s.len: u64 + 1u64);
	let i: i32 = 0;
	for (i < s.len) { b.p[b.n] = s[i]; b.n += 1u64; i += 1; };
	b.p[b.n] = 0u8;
	b.n += 1u64;
	return off;
};

// ---- emit_elf ---------------------------------------------------------

export fn a_emit_elf(a: *asm_, fd: i32) i32 = {
	let shstr: buf;  buf_init(&shstr, a.a);
	let str_:  buf;  buf_init(&str_,  a.a);
	let sym:   buf;  buf_init(&sym,   a.a);
	let rela:  buf;  buf_init(&rela,  a.a);

	// Index 0 = empty.
	let zero: u8 = 0u8;
	buf_putb(&shstr, &zero, 1u64);
	buf_putb(&str_,  &zero, 1u64);

	// Section name offsets.
	let shn_text:     u32 = buf_put_cstr(&shstr, ".text");
	let shn_rela:     u32 = buf_put_cstr(&shstr, ".rela.text");
	let shn_symtab:   u32 = buf_put_cstr(&shstr, ".symtab");
	let shn_strtab:   u32 = buf_put_cstr(&shstr, ".strtab");
	let shn_shstrtab: u32 = buf_put_cstr(&shstr, ".shstrtab");

	// Symbol 0 — STN_UNDEF (24 zero bytes).
	let zsym: [24]u8;
	let zi: i32 = 0;
	for (zi < 24) { zsym[zi] = 0u8; zi += 1; };
	buf_putb(&sym, zsym.ptr, 24u64);

	let SH_TEXT: u16 = 1u16;

	// Build symbols.
	let idx: i32 = 1;
	let s: *asym = a.syms;
	for (s != nil) {
		let entry: [24]u8;
		let ei: i32 = 0;
		for (ei < 24) { entry[ei] = 0u8; ei += 1; };
		let st_name: u32 = buf_put_cstr(&str_, s.name);
		wr_u32(entry.ptr, 0u64, st_name);
		if (s.defined != 0) {
			wr_u8(entry.ptr,  4u64, ((STB_GLOBAL << 4u8) | STT_FUNC));
			wr_u16(entry.ptr, 6u64, SH_TEXT);
			wr_u64(entry.ptr, 8u64, s.addr);
		} else {
			wr_u8(entry.ptr,  4u64, ((STB_GLOBAL << 4u8) | STT_NOTYPE));
			wr_u16(entry.ptr, 6u64, 0u16);
		};
		buf_putb(&sym, entry.ptr, 24u64);
		s.idx = idx;
		idx += 1;
		s = s.snext;
	};

	// Build relocations.
	let r: *areloc = a.relocs;
	for (r != nil) {
		let entry: [24]u8;
		wr_u64(entry.ptr,  0u64, r.off);
		let r_info: u64 = (r.asy.idx: u64 << 32u64) | (r.kind: u64 & 4294967295u64);
		wr_u64(entry.ptr,  8u64, r_info);
		wr_u64(entry.ptr, 16u64, r.addend: u64);
		buf_putb(&rela, entry.ptr, 24u64);
		r = r.rnext;
	};

	// File offsets.
	let off:      u64 = EHDR_SZ;
	let off_text: u64 = off;  off = off + a.textlen;
	let off_rela: u64 = off;  off = off + rela.n;
	let off_sym:  u64 = off;  off = off + sym.n;
	let off_str:  u64 = off;  off = off + str_.n;
	let off_shstr: u64 = off; off = off + shstr.n;
	for ((off & 7u64) != 0u64) { off += 1u64; };
	let off_shdr: u64 = off;
	let NSECT: u16 = 6u16;

	// ---- Ehdr ----
	let eh: [64]u8;
	let i: i32 = 0;
	for (i < 64) { eh[i] = 0u8; i += 1; };
	eh[0] = 127u8;	// 0x7f
	eh[1] = 69u8;	// 'E'
	eh[2] = 76u8;	// 'L'
	eh[3] = 70u8;	// 'F'
	eh[4] = ELFCLASS64;
	eh[5] = ELFDATA2LSB;
	eh[6] = EV_CURRENT_W: u8;
	wr_u16(eh.ptr, 16u64, ET_REL_W);
	wr_u16(eh.ptr, 18u64, EM_X86_64_W);
	wr_u32(eh.ptr, 20u64, EV_CURRENT_W);
	wr_u64(eh.ptr, 24u64, 0u64);		// e_entry
	wr_u64(eh.ptr, 32u64, 0u64);		// e_phoff
	wr_u64(eh.ptr, 40u64, off_shdr);	// e_shoff
	wr_u32(eh.ptr, 48u64, 0u32);		// e_flags
	wr_u16(eh.ptr, 52u64, 64u16);		// e_ehsize
	wr_u16(eh.ptr, 54u64, 0u16);		// e_phentsize
	wr_u16(eh.ptr, 56u64, 0u16);		// e_phnum
	wr_u16(eh.ptr, 58u64, 64u16);		// e_shentsize
	wr_u16(eh.ptr, 60u64, NSECT);		// e_shnum
	wr_u16(eh.ptr, 62u64, 5u16);		// e_shstrndx

	if (os.writefull(fd, eh.ptr, 64u64) != 64i64) { return -1; };
	if (a.textlen > 0u64) {
		if (os.writefull(fd, a.text, a.textlen) != a.textlen: i64) { return -1; };
	};
	if (rela.n > 0u64) {
		if (os.writefull(fd, rela.p, rela.n) != rela.n: i64) { return -1; };
	};
	if (sym.n > 0u64) {
		if (os.writefull(fd, sym.p, sym.n) != sym.n: i64) { return -1; };
	};
	if (str_.n > 0u64) {
		if (os.writefull(fd, str_.p, str_.n) != str_.n: i64) { return -1; };
	};
	if (shstr.n > 0u64) {
		if (os.writefull(fd, shstr.p, shstr.n) != shstr.n: i64) { return -1; };
	};

	// Pad to 8 before shdrs.
	let written: u64 = EHDR_SZ + a.textlen + rela.n + sym.n + str_.n + shstr.n;
	for ((written & 7u64) != 0u64) {
		os.writefull(fd, &zero, 1u64);
		written += 1u64;
	};

	// Section header table — 6 headers of 64 bytes each = 384 bytes.
	let shbuf: [64]u8;
	// SHT_NULL
	let sn: i32 = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	os.writefull(fd, shbuf.ptr, 64u64);
	// .text
	sn = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	wr_u32(shbuf.ptr,  0u64, shn_text);
	wr_u32(shbuf.ptr,  4u64, SHT_PROGBITS_C);
	wr_u64(shbuf.ptr,  8u64, SHF_ALLOC | SHF_EXECINSTR);
	wr_u64(shbuf.ptr, 24u64, off_text);
	wr_u64(shbuf.ptr, 32u64, a.textlen);
	wr_u64(shbuf.ptr, 48u64, 1u64);	// sh_addralign
	os.writefull(fd, shbuf.ptr, 64u64);
	// .rela.text
	sn = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	wr_u32(shbuf.ptr,  0u64, shn_rela);
	wr_u32(shbuf.ptr,  4u64, SHT_RELA_C);
	wr_u64(shbuf.ptr,  8u64, SHF_INFO_LINK);
	wr_u64(shbuf.ptr, 24u64, off_rela);
	wr_u64(shbuf.ptr, 32u64, rela.n);
	wr_u32(shbuf.ptr, 40u64, 3u32);	// sh_link = symtab idx
	wr_u32(shbuf.ptr, 44u64, 1u32);	// sh_info = .text idx
	wr_u64(shbuf.ptr, 48u64, 8u64);
	wr_u64(shbuf.ptr, 56u64, RELA_SZ);
	os.writefull(fd, shbuf.ptr, 64u64);
	// .symtab
	sn = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	wr_u32(shbuf.ptr,  0u64, shn_symtab);
	wr_u32(shbuf.ptr,  4u64, SHT_SYMTAB_C);
	wr_u64(shbuf.ptr, 24u64, off_sym);
	wr_u64(shbuf.ptr, 32u64, sym.n);
	wr_u32(shbuf.ptr, 40u64, 4u32);	// sh_link = strtab
	wr_u32(shbuf.ptr, 44u64, 1u32);	// sh_info = one local (STN_UNDEF)
	wr_u64(shbuf.ptr, 48u64, 8u64);
	wr_u64(shbuf.ptr, 56u64, SYM_SZ);
	os.writefull(fd, shbuf.ptr, 64u64);
	// .strtab
	sn = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	wr_u32(shbuf.ptr,  0u64, shn_strtab);
	wr_u32(shbuf.ptr,  4u64, SHT_STRTAB_C);
	wr_u64(shbuf.ptr, 24u64, off_str);
	wr_u64(shbuf.ptr, 32u64, str_.n);
	wr_u64(shbuf.ptr, 48u64, 1u64);
	os.writefull(fd, shbuf.ptr, 64u64);
	// .shstrtab
	sn = 0;
	for (sn < 64) { shbuf[sn] = 0u8; sn += 1; };
	wr_u32(shbuf.ptr,  0u64, shn_shstrtab);
	wr_u32(shbuf.ptr,  4u64, SHT_STRTAB_C);
	wr_u64(shbuf.ptr, 24u64, off_shstr);
	wr_u64(shbuf.ptr, 32u64, shstr.n);
	wr_u64(shbuf.ptr, 48u64, 1u64);
	os.writefull(fd, shbuf.ptr, 64u64);

	return 0;
};