/*
 * The shape we produce is the simplest valid one: PT_INTERP +
 * PT_DYNAMIC + DT_BIND_NOW so the loader resolves every PLT slot at
 * startup (no lazy binding, no PLT0 trampoline). SysV .hash, not
 * .gnu.hash. Non-PIE, fixed base.
 *
 * File layout:
 *   [0]            Ehdr
 *   [64]           Phdrs (PT_LOAD R+X, PT_LOAD R+W, PT_INTERP, PT_DYNAMIC)
 *   [interp_off]   "/lib64/ld-linux-x86-64.so.2\0"
 *   [dynstr_off]   .dynstr
 *   [dynsym_off]   .dynsym
 *   [hash_off]     .hash
 *   [relaplt_off]  .rela.plt
 *   [pad to 0x1000]
 *   [text_off]     .text
 *   [plt_off]      .plt
 *   [pad to next page]
 *   [gotplt_off]   .got.plt   (writable; mapped by PT_LOAD #2)
 *   [dynamic_off]  .dynamic   (writable; covered by PT_DYNAMIC)
 *
 * Each PLT entry is 8 bytes: `jmpq *(rip+disp)` (6 bytes) + 2 bytes
 * pad so the next entry stays naturally aligned.
 */
#include "l.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#define ET_EXEC		2
#define EM_X86_64	62
#define EV_CURRENT	1
#define ELFCLASS64	2
#define ELFDATA2LSB	1

#define PT_LOAD		1
#define PT_DYNAMIC	2
#define PT_INTERP	3
#define PF_X		1
#define PF_W		2
#define PF_R		4

#define DT_NULL		0
#define DT_NEEDED	1
#define DT_PLTRELSZ	2
#define DT_PLTGOT	3
#define DT_HASH		4
#define DT_STRTAB	5
#define DT_SYMTAB	6
#define DT_STRSZ	10
#define DT_SYMENT	11
#define DT_PLTREL	20
#define DT_RELA		7
#define DT_JMPREL	23
#define DT_BIND_NOW	24
/* GNU extensions for symbol versioning. */
#define DT_VERSYM	0x6ffffff0
#define DT_VERNEED	0x6ffffffe
#define DT_VERNEEDNUM	0x6fffffff

#define VER_NDX_LOCAL	0
#define VER_NDX_GLOBAL	1

#define R_X86_64_PC32		2
#define R_X86_64_PLT32		4
#define R_X86_64_JUMP_SLOT	7

#define STB_GLOBAL	1
#define STT_FUNC	2
#define ST_INFO(b,t)	(((b) << 4) | ((t) & 0xf))

#define INTERP	"/lib64/ld-linux-x86-64.so.2"

#pragma pack(push, 1)
typedef struct {
	u8 e_ident[16];
	u16 e_type, e_machine;
	u32 e_version;
	u64 e_entry, e_phoff, e_shoff;
	u32 e_flags;
	u16 e_ehsize, e_phentsize, e_phnum, e_shentsize, e_shnum, e_shstrndx;
} Ehdr;

typedef struct {
	u32 p_type, p_flags;
	u64 p_offset, p_vaddr, p_paddr;
	u64 p_filesz, p_memsz, p_align;
} Phdr;

typedef struct {
	u32 st_name;
	u8 st_info, st_other;
	u16 st_shndx;
	u64 st_value, st_size;
} Sym64;

typedef struct {
	u64 r_offset;
	u64 r_info;
	i64 r_addend;
} Rela64;

typedef struct {
	i64 d_tag;
	u64 d_val;
} Dyn64;
#pragma pack(pop)

#define ELF64_R_INFO(s,t)	(((u64)(s) << 32) | ((u32)(t)))

/* SysV ELF hash (the older format; .gnu.hash is faster but more code). */
static u32
elf_hash(const char *name)
{
	u32 h = 0, g;
	for (const u8 *s = (const u8 *)name; *s; s++) {
		h = (h << 4) + *s;
		g = h & 0xf0000000u;
		if (g) h ^= g >> 24;
		h &= ~g;
	}
	return h;
}

static void
poke32(u8 *buf, u64 off, u32 v)
{
	buf[off + 0] = (u8)(v);
	buf[off + 1] = (u8)(v >> 8);
	buf[off + 2] = (u8)(v >> 16);
	buf[off + 3] = (u8)(v >> 24);
}

#define PLT_STUB_BYTES	8	/* jmpq *disp(%rip) + 2 nop pad */

int
l_emit_dyn_elf(Lnk *l, FILE *f, u64 base, u64 entry)
{
	/* Writable globals share the dyn-path R+W segment with .got.plt
	 * and .dynamic. .data is placed after .dynamic; the segment's
	 * filesz/memsz are extended to cover all three. Relocation
	 * targeting .data uses data_va computed inside this function
	 * (l_relocate now runs from here, not main, so the dyn layout's
	 * data VA is the one that lands in patched offsets). */

	const int N = l->dyn_n;

	/* dynstr layout: [0]='\0', then DT_NEEDED soname strings, then
	 * one symbol name per dynamic Lsym. We index dyn syms by
	 * plt_idx (assigned in l_resolve). Build an array sorted by
	 * plt_idx so we can walk in slot order. */
	Lsym **dynsyms = calloc((size_t)N, sizeof *dynsyms);
	for (Lsym *s = l->syms; s; s = s->next) {
		if (s->is_dyn && s->plt_idx >= 0 && s->plt_idx < N)
			dynsyms[s->plt_idx] = s;
	}
	for (int i = 0; i < N; i++) {
		if (dynsyms[i] == NULL) {
			fprintf(stderr, "w6l: dynout: no sym for plt_idx %d\n", i);
			free(dynsyms);
			return 1;
		}
	}

	/* Count Lso's that any dyn sym references; only those need DT_NEEDED. */
	int nsos = 0;
	for (Lso *so = l->sos; so; so = so->next) {
		int used = 0;
		for (int i = 0; i < N; i++)
			if (dynsyms[i]->dyn_lib == so) { used = 1; break; }
		if (used) nsos++;
	}
	Lso **sos_used = calloc((size_t)nsos, sizeof *sos_used);
	{
		int idx = 0;
		for (Lso *so = l->sos; so; so = so->next) {
			int used = 0;
			for (int i = 0; i < N; i++)
				if (dynsyms[i]->dyn_lib == so) { used = 1; break; }
			if (used) sos_used[idx++] = so;
		}
	}

	u8 *dynstr = NULL;
	u64 dynstr_cap = 0, dynstr_len = 0;
#define DSTR_PUT(s) do { \
		size_t _n = strlen(s) + 1; \
		if (dynstr_len + _n > dynstr_cap) { \
			dynstr_cap = dynstr_cap ? dynstr_cap * 2 : 256; \
			while (dynstr_cap < dynstr_len + _n) dynstr_cap *= 2; \
			dynstr = realloc(dynstr, dynstr_cap); \
		} \
		memcpy(dynstr + dynstr_len, s, _n); \
		dynstr_len += _n; \
	} while (0)

	DSTR_PUT("");	/* leading null entry */

	u32 *soname_str = calloc((size_t)nsos, sizeof *soname_str);
	for (int i = 0; i < nsos; i++) {
		soname_str[i] = (u32)dynstr_len;
		DSTR_PUT(sos_used[i]->soname);
	}
	u32 *symname_str = calloc((size_t)N, sizeof *symname_str);
	for (int i = 0; i < N; i++) {
		symname_str[i] = (u32)dynstr_len;
		DSTR_PUT(dynsyms[i]->name);
	}

	/* For every sym whose dyn_version is non-NULL, there's a
	 * Vernaux record under that lib's Verneed. The vna_other
	 * value (assigned starting at 2; 1 is reserved for "global,
	 * unversioned") becomes that sym's .gnu.version entry.
	 * Unversioned syms get .gnu.version = 1.
	 *
	 * vlibs is parallel-indexed with sos_used so we can look
	 * up the SONAME's dynstr offset directly.
	 */
	struct vlib_ver { const char *name; u32 dynstr_off; u16 vna_other; };
	struct vlib { int sos_idx; int n_versions; struct vlib_ver *versions; };
	struct vlib *vlibs = calloc((size_t)nsos, sizeof *vlibs);
	int n_vlibs = 0;

	for (int i = 0; i < nsos; i++) {
		int has = 0;
		for (int j = 0; j < N; j++) {
			if (dynsyms[j]->dyn_lib == sos_used[i]
			    && dynsyms[j]->dyn_version != NULL) {
				has = 1; break;
			}
		}
		if (!has) continue;
		struct vlib *vl = &vlibs[n_vlibs];
		vl->sos_idx = i;
		vl->versions = calloc((size_t)N, sizeof *vl->versions);
		vl->n_versions = 0;
		for (int j = 0; j < N; j++) {
			if (dynsyms[j]->dyn_lib != sos_used[i]) continue;
			const char *vname = dynsyms[j]->dyn_version;
			if (vname == NULL) continue;
			int seen = 0;
			for (int k = 0; k < vl->n_versions; k++) {
				if (strcmp(vl->versions[k].name, vname) == 0) {
					seen = 1; break;
				}
			}
			if (!seen) {
				vl->versions[vl->n_versions].name = vname;
				vl->n_versions++;
			}
		}
		n_vlibs++;
	}

	u16 next_vna = 2;
	for (int i = 0; i < n_vlibs; i++)
		for (int k = 0; k < vlibs[i].n_versions; k++)
			vlibs[i].versions[k].vna_other = next_vna++;

	for (int i = 0; i < n_vlibs; i++) {
		for (int k = 0; k < vlibs[i].n_versions; k++) {
			vlibs[i].versions[k].dynstr_off = (u32)dynstr_len;
			DSTR_PUT(vlibs[i].versions[k].name);
		}
	}

	/* Per-dyn-sym versym index: 1 (global) for unversioned, else
	 * the matched Vernaux's vna_other. */
	u16 *versym_for = calloc((size_t)N, sizeof *versym_for);
	for (int j = 0; j < N; j++) {
		const char *vname = dynsyms[j]->dyn_version;
		if (vname == NULL) { versym_for[j] = VER_NDX_GLOBAL; continue; }
		int matched = 0;
		for (int i = 0; i < n_vlibs && !matched; i++) {
			if (sos_used[vlibs[i].sos_idx] != dynsyms[j]->dyn_lib)
				continue;
			for (int k = 0; k < vlibs[i].n_versions; k++) {
				if (strcmp(vlibs[i].versions[k].name, vname) == 0) {
					versym_for[j] = vlibs[i].versions[k].vna_other;
					matched = 1; break;
				}
			}
		}
		if (!matched) {
			fprintf(stderr, "w6l: dynout: unmatched version %s for %s\n",
			    vname, dynsyms[j]->name);
			versym_for[j] = VER_NDX_GLOBAL;
		}
	}

	const u64 ehdr_sz   = sizeof(Ehdr);
	const int n_phdrs   = 4;
	const u64 phdr_sz   = (u64)n_phdrs * sizeof(Phdr);

	const u64 interp_sz = strlen(INTERP) + 1;

	/* nsyms = 1 (undef at index 0) + N */
	const int nsyms_total = 1 + N;
	const u64 dynsym_sz   = (u64)nsyms_total * sizeof(Sym64);
	const u64 dynstr_sz   = dynstr_len;

	/* SysV hash: nbuckets + nchain + buckets[] + chain[]. We use one
	 * bucket; loader scans the whole chain. Cheap to compute, easy to
	 * validate. */
	const u32 nbuckets = 1;
	const u32 nchain   = (u32)nsyms_total;
	const u64 hash_sz  = (2 + nbuckets + nchain) * 4;

	const u64 relaplt_sz = (u64)N * sizeof(Rela64);

	const u64 plt_sz    = (u64)N * PLT_STUB_BYTES;
	const u64 gotplt_sz = (3 + (u64)N) * 8;

	/* .gnu.version: one Elf64_Half per .dynsym entry. */
	const u64 versym_sz = (u64)nsyms_total * 2;

	/* .gnu.version_r: per lib, 16-byte Verneed plus 16-byte Vernaux
	 * for each version under it. */
	u64 verneed_sz = 0;
	for (int i = 0; i < n_vlibs; i++)
		verneed_sz += 16 + 16 * (u64)vlibs[i].n_versions;

	/* dynamic entries: NEEDED*nsos, HASH, STRTAB, SYMTAB, STRSZ, SYMENT,
	 * PLTGOT, PLTRELSZ, PLTREL, JMPREL, BIND_NOW, [VERSYM, VERNEED,
	 * VERNEEDNUM], NULL. The version trio is conditional on having any
	 * versioned references. */
	const int with_ver = (n_vlibs > 0);
	const u64 ndyn = (u64)nsos + 11 + (with_ver ? 3 : 0);
	const u64 dynamic_sz = ndyn * sizeof(Dyn64);

	/* Everything from the Ehdr through .text+.plt is in the R+X
	 * load segment at base+0..text_end. .got.plt and .dynamic land
	 * in the R+W segment at the next page boundary. */

	u64 off = ehdr_sz + phdr_sz;
	const u64 interp_off  = off; off += interp_sz;
	off = (off + 7) & ~(u64)7;
	const u64 dynstr_off  = off; off += dynstr_sz;
	off = (off + 7) & ~(u64)7;
	const u64 dynsym_off  = off; off += dynsym_sz;
	const u64 hash_off    = off; off += hash_sz;
	off = (off + 1) & ~(u64)1;
	const u64 versym_off  = off; off += versym_sz;
	off = (off + 3) & ~(u64)3;
	const u64 verneed_off = off; off += verneed_sz;
	off = (off + 7) & ~(u64)7;
	const u64 relaplt_off = off; off += relaplt_sz;

	/* Pad to 0x1000 so .text is page-aligned (matters for the loader
	 * mapping our R+X PT_LOAD). */
	const u64 page = 0x1000;
	const u64 text_off = (off + page - 1) & ~(page - 1);
	const u64 plt_off  = text_off + l->textlen;
	const u64 rx_end   = plt_off + plt_sz;

	/* Page-align the writable segment. We skip a page of file bytes;
	 * the data lands at file offset gotplt_off, vaddr at base+gotplt_va.
	 * .data sits after .dynamic so the whole R+W run is one segment. */
	const u64 gotplt_off  = (rx_end + page - 1) & ~(page - 1);
	const u64 dynamic_off = gotplt_off + gotplt_sz;
	const u64 data_off    = dynamic_off + dynamic_sz;
	const u64 file_end    = data_off + l->datalen;

	/* Virtual addresses mirror file offsets within their segment.
	 * The R+W segment in particular needs vaddr = base + gotplt_off
	 * so file offset and vaddr modulo page agree (loader requirement). */
	const u64 interp_va  = base + interp_off;
	const u64 dynstr_va  = base + dynstr_off;
	const u64 dynsym_va  = base + dynsym_off;
	const u64 hash_va    = base + hash_off;
	const u64 versym_va  = base + versym_off;
	const u64 verneed_va = base + verneed_off;
	const u64 relaplt_va = base + relaplt_off;
	const u64 text_va    = base + text_off;
	const u64 plt_va     = base + plt_off;
	const u64 gotplt_va  = base + gotplt_off;
	const u64 dynamic_va = base + dynamic_off;
	const u64 data_va    = base + data_off;

	/* Now that the dyn layout pins text_va/data_va, apply
	 * relocations. main.c defers this; the static path runs it from
	 * out.c with its own VAs. */
	if (l_relocate(l, text_va, data_va) != 0) return 1;

	/* BSS optimisation — same trailing-zero scan as out.c. */
	u64 bsslen = 0;
	if (l->datalen > 0) {
		while (bsslen < l->datalen
		    && l->data[l->datalen - 1 - bsslen] == 0)
			bsslen++;
	}
	const u64 data_file_len = l->datalen - bsslen;
	const u64 file_data_end = data_off + data_file_len;

	Sym64 *dynsym = calloc((size_t)nsyms_total, sizeof *dynsym);
	for (int i = 0; i < N; i++) {
		Sym64 *e = &dynsym[1 + i];
		e->st_name  = symname_str[i];
		e->st_info  = ST_INFO(STB_GLOBAL, STT_FUNC);
		e->st_other = 0;
		e->st_shndx = 0;	/* SHN_UNDEF */
		e->st_value = 0;
		e->st_size  = 0;
		dynsyms[i]->dynsym_idx = 1 + i;
	}

	/* .hash (SysV format, 1 bucket). */
	u32 *hash = calloc(2 + nbuckets + nchain, 4);
	hash[0] = nbuckets;
	hash[1] = nchain;
	/* buckets[0] = first entry that lives in this bucket; we put
	 * everything in bucket 0, so the bucket head is symbol 1. */
	hash[2] = nsyms_total > 1 ? 1 : 0;
	/* chain[i] = next sym in the bucket. Last one terminates with 0. */
	for (int i = 1; i < nsyms_total; i++) {
		u32 next = (i + 1 < nsyms_total) ? (u32)(i + 1) : 0;
		hash[2 + nbuckets + i] = next;
	}
	/* elf_hash is also used by .gnu.version_r for vna_hash below. */

	Rela64 *relaplt = calloc((size_t)N, sizeof *relaplt);
	for (int i = 0; i < N; i++) {
		relaplt[i].r_offset = gotplt_va + (3 + (u64)i) * 8;
		relaplt[i].r_info   = ELF64_R_INFO(1 + i, R_X86_64_JUMP_SLOT);
		relaplt[i].r_addend = 0;
	}

	/* .gnu.version: u16 per .dynsym entry. [0] = LOCAL, [1+i] = the
	 * versym index we computed for dyn sym i. */
	u16 *versym = calloc((size_t)nsyms_total, 2);
	versym[0] = VER_NDX_LOCAL;
	for (int i = 0; i < N; i++)
		versym[1 + i] = versym_for[i];

	/* .gnu.version_r: chain of Verneed records, one per versioned lib,
	 * each with a chain of Vernaux records, one per version under it.
	 * We write directly into a u8 buffer with little-endian poke
	 * helpers to avoid alignment concerns. */
	u8 *verneed = NULL;
	if (verneed_sz > 0) {
		verneed = calloc((size_t)verneed_sz, 1);
		u64 vnoff = 0;
		for (int i = 0; i < n_vlibs; i++) {
			struct vlib *vl = &vlibs[i];
			u64 vn_start = vnoff;
			/* Verneed header (16 bytes). */
			u8 *vn = verneed + vnoff;
			/* vn_version = 1, vn_cnt = nversions */
			vn[0] = 1; vn[1] = 0;
			vn[2] = (u8)(vl->n_versions);
			vn[3] = (u8)(vl->n_versions >> 8);
			poke32(vn, 4, soname_str[vl->sos_idx]);	/* vn_file */
			poke32(vn, 8, 16);			/* vn_aux */
			/* vn_next set after we know the vernaux count */
			vnoff += 16;
			for (int k = 0; k < vl->n_versions; k++) {
				u8 *va = verneed + vnoff;
				poke32(va, 0, elf_hash(vl->versions[k].name));
				/* vna_flags = 0 */
				va[4] = 0; va[5] = 0;
				/* vna_other (versym index) */
				va[6] = (u8)(vl->versions[k].vna_other);
				va[7] = (u8)(vl->versions[k].vna_other >> 8);
				poke32(va, 8, vl->versions[k].dynstr_off);
				poke32(va, 12,
				    (k + 1 < vl->n_versions) ? 16u : 0u);
				vnoff += 16;
			}
			/* Now patch vn_next at vn_start+12. */
			poke32(verneed, vn_start + 12,
			    (i + 1 < n_vlibs)
			    ? (u32)(vnoff - vn_start) : 0u);
		}
	}

	/* .plt — `jmpq *got.plt[3+i](%rip)` per stub.
	 * Encoding: FF 25 <disp32>. The disp is computed from the address
	 * of the *next* instruction (RIP after the 6-byte jmp) to the
	 * GOT slot. */
	u8 *plt = calloc((size_t)plt_sz, 1);
	for (int i = 0; i < N; i++) {
		u64 stub_va    = plt_va + (u64)i * PLT_STUB_BYTES;
		u64 next_ip    = stub_va + 6;
		u64 slot_va    = gotplt_va + (3 + (u64)i) * 8;
		i64 disp       = (i64)slot_va - (i64)next_ip;
		u8 *p = plt + (u64)i * PLT_STUB_BYTES;
		p[0] = 0xff;
		p[1] = 0x25;
		poke32(p, 2, (u32)(i32)disp);
		/* p[6], p[7] left as zero — pad. */
	}

	/* .got.plt — first three slots are reserved.
	 * [0] = address of .dynamic (loader reads this).
	 * [1] = link_map *           (loader writes at startup).
	 * [2] = dl_runtime_resolve   (loader writes; unused with BIND_NOW). */
	u8 *gotplt = calloc((size_t)gotplt_sz, 1);
	{
		u64 v = dynamic_va;
		for (int b = 0; b < 8; b++) gotplt[b] = (u8)(v >> (b * 8));
	}
	/* [3..3+N-1] left zero; loader fills via R_X86_64_JUMP_SLOT. */

	Dyn64 *dynamic = calloc((size_t)ndyn, sizeof *dynamic);
	{
		int k = 0;
		for (int i = 0; i < nsos; i++) {
			dynamic[k].d_tag = DT_NEEDED;
			dynamic[k].d_val = soname_str[i];
			k++;
		}
		dynamic[k].d_tag = DT_HASH;     dynamic[k].d_val = hash_va;     k++;
		dynamic[k].d_tag = DT_STRTAB;   dynamic[k].d_val = dynstr_va;   k++;
		dynamic[k].d_tag = DT_SYMTAB;   dynamic[k].d_val = dynsym_va;   k++;
		dynamic[k].d_tag = DT_STRSZ;    dynamic[k].d_val = dynstr_sz;   k++;
		dynamic[k].d_tag = DT_SYMENT;   dynamic[k].d_val = sizeof(Sym64); k++;
		dynamic[k].d_tag = DT_PLTGOT;   dynamic[k].d_val = gotplt_va;   k++;
		dynamic[k].d_tag = DT_PLTRELSZ; dynamic[k].d_val = relaplt_sz;  k++;
		dynamic[k].d_tag = DT_PLTREL;   dynamic[k].d_val = DT_RELA;     k++;
		dynamic[k].d_tag = DT_JMPREL;   dynamic[k].d_val = relaplt_va;  k++;
		dynamic[k].d_tag = DT_BIND_NOW; dynamic[k].d_val = 0;           k++;
		if (with_ver) {
			dynamic[k].d_tag = DT_VERSYM;
			dynamic[k].d_val = versym_va;
			k++;
			dynamic[k].d_tag = DT_VERNEED;
			dynamic[k].d_val = verneed_va;
			k++;
			dynamic[k].d_tag = DT_VERNEEDNUM;
			dynamic[k].d_val = (u64)n_vlibs;
			k++;
		}
		dynamic[k].d_tag = DT_NULL;     dynamic[k].d_val = 0;           k++;
		if ((u64)k != ndyn) {
			fprintf(stderr, "w6l: dynamic entry count mismatch\n");
			return 1;
		}
	}

	/* Patch .text relocations targeting dynamic syms: the site is
	 * the existing PC32/PLT32 displacement field, the target the
	 * address of the symbol's PLT stub. */
	for (Lrel *r = l->rels; r; r = r->next) {
		if (r->sym == NULL || !r->sym->is_dyn) continue;
		if (r->kind != R_X86_64_PC32 && r->kind != R_X86_64_PLT32) {
			fprintf(stderr, "w6l: dynamic reloc kind %d unsupported\n",
			    r->kind);
			free(dynsyms); free(sos_used); free(soname_str);
			free(symname_str); free(dynstr); free(dynsym);
			free(hash); free(relaplt); free(plt); free(gotplt);
			free(dynamic);
			return 1;
		}
		u64 site = text_va + r->off;
		u64 stub = plt_va + (u64)r->sym->plt_idx * PLT_STUB_BYTES;
		i64 disp = (i64)stub - (i64)site + r->addend;
		poke32(l->text, r->off, (u32)(i32)disp);
	}

	Ehdr eh = {0};
	memcpy(eh.e_ident, "\x7f""ELF", 4);
	eh.e_ident[4] = ELFCLASS64;
	eh.e_ident[5] = ELFDATA2LSB;
	eh.e_ident[6] = EV_CURRENT;
	eh.e_type = ET_EXEC;
	eh.e_machine = EM_X86_64;
	eh.e_version = EV_CURRENT;
	/* #63: main computed `entry` as base + 0x1000 + val assuming .text sits
	 * at file offset 0x1000, but text_off above is recomputed and overflows
	 * 0x1000 once the dynamic headers exceed the first page (e.g. ~100 dyn
	 * syms). Rebase e_entry onto the actual text_off so it points at the
	 * real _start instead of into the headers. e_entry is a virtual address;
	 * PT_LOAD #1 maps file offset 0 at `base`, so the entry VA tracks
	 * text_off. Both stages fixed together (#263). */
	eh.e_entry = entry - 0x1000 + text_off;
	eh.e_phoff = ehdr_sz;
	eh.e_ehsize = sizeof(Ehdr);
	eh.e_phentsize = sizeof(Phdr);
	eh.e_phnum = (u16)n_phdrs;

	Phdr ph[4] = {0};
	/* PT_LOAD #1 — R+X covering everything from Ehdr through .plt. */
	ph[0].p_type = PT_LOAD;
	ph[0].p_flags = PF_R | PF_X;
	ph[0].p_offset = 0;
	ph[0].p_vaddr  = base;
	ph[0].p_paddr  = base;
	ph[0].p_filesz = rx_end;
	/* The loader page-rounds mappings; gotplt_off independently starts
	 * the next page, so padding is not part of the R+X extent. */
	ph[0].p_memsz  = rx_end;
	ph[0].p_align  = page;

	/* PT_LOAD #2 — R+W covering .got.plt, .dynamic, and .data.
	 * filesz drops the trailing-zero suffix (BSS); memsz keeps the
	 * full extent so the loader zero-fills the gap. */
	ph[1].p_type = PT_LOAD;
	ph[1].p_flags = PF_R | PF_W;
	ph[1].p_offset = gotplt_off;
	ph[1].p_vaddr  = gotplt_va;
	ph[1].p_paddr  = gotplt_va;
	ph[1].p_filesz = file_data_end - gotplt_off;
	ph[1].p_memsz  = file_end - gotplt_off;
	ph[1].p_align  = page;

	ph[2].p_type = PT_INTERP;
	ph[2].p_flags = PF_R;
	ph[2].p_offset = interp_off;
	ph[2].p_vaddr  = interp_va;
	ph[2].p_paddr  = interp_va;
	ph[2].p_filesz = interp_sz;
	ph[2].p_memsz  = interp_sz;
	ph[2].p_align  = 1;

	ph[3].p_type = PT_DYNAMIC;
	ph[3].p_flags = PF_R | PF_W;
	ph[3].p_offset = dynamic_off;
	ph[3].p_vaddr  = dynamic_va;
	ph[3].p_paddr  = dynamic_va;
	ph[3].p_filesz = dynamic_sz;
	ph[3].p_memsz  = dynamic_sz;
	ph[3].p_align  = 8;

	fwrite(&eh, 1, sizeof eh, f);
	fwrite(ph,  1, sizeof ph, f);

#define PAD_TO(to) do { \
		long _here = ftell(f); \
		for (long _i = _here; _i < (long)(to); _i++) fputc(0, f); \
	} while (0)

	PAD_TO(interp_off);  fwrite(INTERP, 1, interp_sz, f);
	PAD_TO(dynstr_off);  fwrite(dynstr, 1, dynstr_sz, f);
	PAD_TO(dynsym_off);  fwrite(dynsym, 1, dynsym_sz, f);
	PAD_TO(hash_off);    fwrite(hash, 4, 2 + nbuckets + nchain, f);
	PAD_TO(versym_off);  fwrite(versym, 2, (size_t)nsyms_total, f);
	if (verneed_sz > 0) {
		PAD_TO(verneed_off); fwrite(verneed, 1, verneed_sz, f);
	}
	PAD_TO(relaplt_off); fwrite(relaplt, 1, relaplt_sz, f);
	PAD_TO(text_off);    fwrite(l->text, 1, l->textlen, f);
	PAD_TO(plt_off);     fwrite(plt, 1, plt_sz, f);
	PAD_TO(gotplt_off);  fwrite(gotplt, 1, gotplt_sz, f);
	PAD_TO(dynamic_off); fwrite(dynamic, 1, dynamic_sz, f);
	if (data_file_len > 0) {
		PAD_TO(data_off); fwrite(l->data, 1, data_file_len, f);
	}

	for (int i = 0; i < n_vlibs; i++) free(vlibs[i].versions);
	free(vlibs); free(versym_for); free(versym);
	if (verneed) free(verneed);
	free(dynsyms); free(sos_used); free(soname_str);
	free(symname_str); free(dynstr); free(dynsym);
	free(hash); free(relaplt); free(plt); free(gotplt); free(dynamic);
	return 0;
}