// selfhost/cmd/6l/out.ww — port of cmd/6l/out.c.
//
// Emit a static ELF64 executable. File layout (per the C original):
//   [0..64)       Ehdr
//   [64..120)     Phdr (one PT_LOAD)
//   [120..0x1000) zero pad
//   [0x1000..)    .text bytes
// Single PT_LOAD covers the whole file, R+X. No interpreter, no .bss.

use os;
use sym;
use dynout;

def ET_EXEC:    u16 = 2u16;
def EM_X86_64_W: u16 = 62u16;
def EV_CURRENT: u32 = 1u32;
def ELFCLASS64: u8 = 2u8;
def ELFDATA2LSB: u8 = 1u8;
def PT_LOAD:    u32 = 1u32;
def PF_X:       u32 = 1u32;
def PF_R:       u32 = 4u32;

def TEXT_OFF: u64 = 4096u64;	// 0x1000

// ---- little-endian byte writers ----------------------------------------

fn wr_u16(buf: *u8, off: u64, v: u16) void = {
	buf[off]         = (v & 255u16): u8;
	buf[off + 1u64]  = ((v >> 8u16) & 255u16): u8;
};

fn wr_u32(buf: *u8, off: u64, v: u32) void = {
	buf[off]         = (v & 255u32): u8;
	buf[off + 1u64]  = ((v >> 8u32)  & 255u32): u8;
	buf[off + 2u64]  = ((v >> 16u32) & 255u32): u8;
	buf[off + 3u64]  = ((v >> 24u32) & 255u32): u8;
};

fn wr_u64(buf: *u8, off: u64, v: u64) void = {
	wr_u32(buf, off,        (v & 4294967295u64): u32);
	wr_u32(buf, off + 4u64, ((v >> 32u64) & 4294967295u64): u32);
};

// ---- emit ---------------------------------------------------------------

export fn l_emit_elf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
	// Dispatch: any loaded shared object plus any dynamic ref means
	// we owe the loader a real PT_INTERP/PT_DYNAMIC binary.
	if (l.sos != nil) {
		if (l.dyn_n > 0) {
			return l_emit_dyn_elf(l, fd, base, entry);
		};
	};

	let filesz: u64 = TEXT_OFF + l.textlen;

	// One contiguous header buffer covering [0..0x1000), then .text.
	let hdr: *u8 = os.alloc(TEXT_OFF): *u8;	// zero-initialised by mmap

	// --- Ehdr (64 bytes) ---
	hdr[0u64] = 127u8;	// 0x7f
	hdr[1u64] = 69u8;	// 'E'
	hdr[2u64] = 76u8;	// 'L'
	hdr[3u64] = 70u8;	// 'F'
	hdr[4u64] = ELFCLASS64;
	hdr[5u64] = ELFDATA2LSB;
	hdr[6u64] = EV_CURRENT: u8;
	wr_u16(hdr, 16u64, ET_EXEC);		// e_type
	wr_u16(hdr, 18u64, EM_X86_64_W);	// e_machine
	wr_u32(hdr, 20u64, EV_CURRENT);		// e_version
	wr_u64(hdr, 24u64, entry);		// e_entry
	wr_u64(hdr, 32u64, 64u64);		// e_phoff = sizeof(Ehdr)
	wr_u64(hdr, 40u64, 0u64);		// e_shoff
	wr_u32(hdr, 48u64, 0u32);		// e_flags
	wr_u16(hdr, 52u64, 64u16);		// e_ehsize
	wr_u16(hdr, 54u64, 56u16);		// e_phentsize
	wr_u16(hdr, 56u64, 1u16);		// e_phnum
	wr_u16(hdr, 58u64, 0u16);		// e_shentsize
	wr_u16(hdr, 60u64, 0u16);		// e_shnum
	wr_u16(hdr, 62u64, 0u16);		// e_shstrndx

	// --- Phdr (56 bytes) at offset 64 ---
	wr_u32(hdr, 64u64, PT_LOAD);		// p_type
	wr_u32(hdr, 68u64, PF_R | PF_X);	// p_flags
	wr_u64(hdr, 72u64, 0u64);		// p_offset
	wr_u64(hdr, 80u64, base);		// p_vaddr
	wr_u64(hdr, 88u64, base);		// p_paddr
	wr_u64(hdr, 96u64, filesz);		// p_filesz
	wr_u64(hdr, 104u64, filesz);		// p_memsz
	wr_u64(hdr, 112u64, TEXT_OFF);		// p_align

	// Write [0..0x1000) then .text.
	let n1: i64 = os.writefull(fd, hdr, TEXT_OFF);
	if (n1 != TEXT_OFF: i64) { return -1; };
	if (l.textlen > 0u64) {
		let n2: i64 = os.writefull(fd, l.text, l.textlen);
		if (n2 != l.textlen: i64) { return -1; };
	};
	return 0;
};