From c4181af08d3c3282e889e30a3eb11fff9a392d2e Mon Sep 17 00:00:00 2001
From: Cosmin Apreutesei <cosmin.apreutesei@gmail.com>
Date: Fri, 30 Oct 2015 13:41:52 +0200
Subject: [PATCH] DynASM: Lua mode

---
 dynasm/dasm.lua      | 274 ++++++++++++++++++++++++++++++++++
 dynasm/dasm_extern.h |   5 +
 dynasm/dasm_mm.lua   | 144 ++++++++++++++++++
 dynasm/dasm_x64.lua  |  12 +-
 dynasm/dasm_x86.c    |  11 ++
 dynasm/dasm_x86.lua  | 183 ++++++++++++++++++-----
 dynasm/dynasm.lua    | 339 ++++++++++++++++++++++++++++++++++++-------
 7 files changed, 873 insertions(+), 95 deletions(-)
 create mode 100644 dynasm/dasm.lua
 create mode 100644 dynasm/dasm_extern.h
 create mode 100644 dynasm/dasm_mm.lua
 create mode 100644 dynasm/dasm_x86.c

diff --git a/dynasm/dasm.lua b/dynasm/dasm.lua
new file mode 100644
index 00000000..fc17500a
--- /dev/null
+++ b/dynasm/dasm.lua
@@ -0,0 +1,274 @@
+
+--Binding to the DynASM encoding engine.
+--Written by Cosmin Apreutesei. Public Domain.
+
+local ffi = require'ffi'
+local bit = require'bit'
+local arch = ffi.arch
+if arch == 'x64' then arch = 'x86' end --same linker for x64
+local C = ffi.load('dasm_'..arch)
+local M = {C = C}
+
+M._VERSION = 10400
+
+ffi.cdef[[
+enum {
+	DASM_S_OK         = 0x00000000,
+	DASM_S_NOMEM      = 0x01000000,
+	DASM_S_PHASE      = 0x02000000,
+	DASM_S_MATCH_SEC  = 0x03000000,
+	DASM_S_RANGE_I    = 0x11000000,
+	DASM_S_RANGE_SEC  = 0x12000000,
+	DASM_S_RANGE_LG   = 0x13000000,
+	DASM_S_RANGE_PC   = 0x14000000,
+	DASM_S_RANGE_VREG	= 0x15000000,
+	DASM_S_UNDEF_L    = 0x21000000,
+	DASM_S_UNDEF_PC   = 0x22000000,
+};
+
+/* Internal DynASM encoder state. */
+typedef struct dasm_State_Ref { struct dasm_State *p; } dasm_State_Ref;
+typedef dasm_State_Ref *Dst_DECL;
+
+/* Initialize and free DynASM state. */
+void dasm_init(Dst_DECL, int maxsection);
+void dasm_free(Dst_DECL);
+
+/* Setup global array. Must be called before dasm_setup(). */
+void dasm_setupglobal(Dst_DECL, void **gl, unsigned int maxgl);
+
+/* Grow PC label array. Can be called after dasm_setup(), too. */
+void dasm_growpc(Dst_DECL, unsigned int maxpc);
+
+/* Setup encoder. */
+void dasm_setup(Dst_DECL, const void *actionlist);
+
+/* Feed encoder with actions. Calls are generated by pre-processor. */
+void dasm_put(Dst_DECL, int start, ...);
+
+/* Link sections and return the resulting size. */
+int dasm_link(Dst_DECL, size_t *szp);
+
+/* Encode sections into buffer. */
+int dasm_encode(Dst_DECL, void *buffer);
+
+/* Get PC label offset. */
+int dasm_getpclabel(Dst_DECL, unsigned int pc);
+
+/* Optional sanity checker to call between isolated encoding steps. */
+int dasm_checkstep(Dst_DECL, int secmatch);
+
+typedef int (*DASM_EXTERN_TYPE) (void *ctx, unsigned char *addr, int idx, int type);
+DASM_EXTERN_TYPE DASM_EXTERN_FUNC;
+]]
+
+local function err(...)
+  io.stderr:setvbuf'no'
+  io.stderr:write('dasm error: ', ...)
+  io.stderr:write'\n'
+  os.exit(1)
+end
+
+--status check helper
+
+local status_map = {
+	[C.DASM_S_NOMEM]      = 'out of memory',
+	[C.DASM_S_PHASE]      = 'phase error',
+	[C.DASM_S_MATCH_SEC]  = 'section not found',
+	[C.DASM_S_RANGE_I]    = 'immediate value out of range',
+	[C.DASM_S_RANGE_SEC]  = 'too many sections',
+	[C.DASM_S_RANGE_LG]   = 'too many global labels',
+	[C.DASM_S_RANGE_PC]   = 'too many pclabels',
+	[C.DASM_S_RANGE_VREG] = 'variable register out of range',
+	[C.DASM_S_UNDEF_L]    = 'undefined global label',
+	[C.DASM_S_UNDEF_PC]   = 'undefined pclabel',
+}
+
+local function checkst(st)
+	if st == C.DASM_S_OK then return end
+	local status, arg = status_map[bit.band(st, 0xff000000)], bit.band(st, 0x00ffffff)
+	if status then
+		err(status, '. :', arg)
+	else
+		err(string.format('0x%08X', st))
+	end
+end
+
+--low level API
+
+M.init = C.dasm_init
+M.free = C.dasm_free
+M.setupglobal = C.dasm_setupglobal
+M.growpc = C.dasm_growpc
+M.setup = C.dasm_setup
+
+local int_ct = ffi.typeof'int'
+local function convert_arg(arg)   --dasm_put() accepts only int32 varargs.
+	if type(arg) == "number" then  --but we make it accept uint32 too by normalizing the arg.
+		arg = bit.tobit(arg)        --non-number args are converted to int32 according to ffi rules.
+	end
+	return ffi.cast(int_ct, arg)
+end
+local function convert_args(...) --not a tailcall but at least it doesn't make any garbage
+	if select('#', ...) == 0 then return end
+	return convert_arg(...), convert_args(select(2, ...))
+end
+function M.put(state, start, ...)
+	C.dasm_put(state, start, convert_args(...))
+end
+
+function M.link(state, sz)
+	sz = sz or ffi.new'size_t[1]'
+	checkst(C.dasm_link(state, sz))
+	return tonumber(sz[0])
+end
+
+function M.encode(state, buf)
+	checkst(C.dasm_encode(state, buf))
+end
+jit.off(M.encode) --calls the DASM_EXTERN_FUNC callback
+
+local voidptr_ct = ffi.typeof'void*'
+local byteptr_ct = ffi.typeof'int8_t*'
+function M.getpclabel(state, pc, buf)
+	local offset = C.dasm_getpclabel(state, pc)
+	if buf then
+		return ffi.cast(voidptr_ct, ffi.cast(byteptr_ct, buf) + offset)
+	end
+	return offset
+end
+
+function M.checkstep(state, section)
+	checkst(C.dasm_checkstep(state, section or -1))
+end
+
+--get the address of a standard symbol.
+--TODO: ask Mike to expose clib_getsym() in ffi so we can get the address
+--of symbols without having to declare them first.
+local function getsym(name)
+	return ffi.C[name]
+end
+local getsym = function(name)
+	local ok, sym = pcall(getsym, name)
+	if not ok then --not found or not defined: define it and try again
+		ffi.cdef(string.format('void %s();', name))
+		return getsym(name)
+	else
+		return sym
+	end
+end
+
+--DASM_EXTERN callback plumbing
+
+local extern_names    --t[idx] -> name
+local extern_get      --f(name) -> ptr
+
+local byteptr_ct = ffi.typeof'uint8_t*'
+
+local function DASM_EXTERN_FUNC(ctx, addr, idx, type)
+	if not extern_names or not extern_get then
+		err'extern callback not initialized.'
+	end
+	local name = extern_names[idx]
+	local ptr = extern_get(name)
+	if ptr == nil then
+	  err('extern not found: ', name, '.')
+	end
+	if type ~= 0 then
+		return ffi.cast(byteptr_ct, ptr) - addr - 4
+	else
+		return ptr
+	end
+end
+
+function M.setupextern(_, names, getter)
+	extern_names = names
+	extern_get = getter or getsym
+	if C.DASM_EXTERN_FUNC == nil then
+		C.DASM_EXTERN_FUNC = DASM_EXTERN_FUNC
+	end
+end
+
+--hi-level API
+
+function M.new(actionlist, externnames, sectioncount, globalcount, externget, globals)
+	local state = ffi.new'dasm_State_Ref'
+	M.init(state, sectioncount or 1)
+	globalcount = globalcount or 256
+	globals = globals or ffi.new('void*[?]', globalcount)
+	ffi.gc(state, function(state)
+		local _ = actionlist, externnames, globals, externget --anchor those: don't rely on the user doing so
+		M.free(state)
+	end)
+	M.setupglobal(state, globals, globalcount)
+	M.setupextern(state, externnames, externget)
+	M.setup(state, actionlist)
+	return state, globals
+end
+
+function M.build(state)
+	state:checkstep(-1)
+	local sz = state:link()
+	if sz == 0 then err'no code?' end
+	local mm = require'dasm_mm' --runtime dependency
+	local buf = mm.new(sz)
+	state:encode(buf)
+	mm.protect(buf, sz)
+	return buf, sz
+end
+
+function M.dump(addr, size, out)
+	local disass = require('jit.dis_'..jit.arch).disass
+	disass(ffi.string(addr, size), tonumber(ffi.cast('uintptr_t', addr)), out)
+end
+
+--given the globals array from dasm.new() and the globalnames list
+--from the `.globalnames` directive, return a map {global_name -> global_addr}.
+function M.globals(globals, globalnames)
+	local t = {}
+	for i = 0, #globalnames do
+		if globals[i] ~= nil then
+			t[globalnames[i]] = globals[i]
+		end
+	end
+	return t
+end
+
+--object interface
+
+ffi.metatype('dasm_State_Ref', {__index = {
+	--low-level API
+	init = M.init,
+	free = M.free,
+	setupglobal = M.setupglobal,
+	setupextern = M.setupextern,
+	growpc = M.growpc,
+	setup = M.setup,
+	put = M.put,
+	link = M.link,
+	encode = M.encode,
+	getpclabel = M.getpclabel,
+	checkstep = M.checkstep,
+	--hi-level API
+	build = M.build,
+}})
+
+if not ... then --demo
+	local dasm = M
+	local actions = ffi.new('const uint8_t[19]', {254,0,102,184,5,0,254,1,102,187,3,0,254,2,102,187,3,0,255})
+	local Dst, globals = dasm.new(actions, nil, 3)
+	--|.code
+	dasm.put(Dst, 0)
+	--| mov ax, 5
+	--|.sub1
+	dasm.put(Dst, 2)
+	--| mov bx, 3
+	--|.sub2
+	dasm.put(Dst, 8)
+	--| mov bx, 3
+	dasm.put(Dst, 14)
+	local addr, size = Dst:build()
+	dasm.dump(addr, size)
+end
+
+return M
diff --git a/dynasm/dasm_extern.h b/dynasm/dasm_extern.h
new file mode 100644
index 00000000..856da17f
--- /dev/null
+++ b/dynasm/dasm_extern.h
@@ -0,0 +1,5 @@
+// DASM_EXTERN_FUNC is a global function pointer to be set at runtime.
+// It will be used in place of the DASM_EXTERN() macro.
+typedef int (*DASM_EXTERN_TYPE) (void *ctx, unsigned char *addr, int idx, int type);
+DASM_EXTERN_TYPE DASM_EXTERN_FUNC;
+#define DASM_EXTERN(a,b,c,d) DASM_EXTERN_FUNC(a,b,c,d)
diff --git a/dynasm/dasm_mm.lua b/dynasm/dasm_mm.lua
new file mode 100644
index 00000000..76dedc7e
--- /dev/null
+++ b/dynasm/dasm_mm.lua
@@ -0,0 +1,144 @@
+
+--wrappers around mmap to support dynamic code exection.
+--Written by Cosmin Apreutesei. Public Domain.
+--Tested with Windows, Linux and OSX, x86 and x86-64.
+
+local ffi = require'ffi'
+local C = ffi.C
+
+local function checkh(ptr) return assert(ptr ~= nil and ptr) end
+local function checkz(ret) assert(ret == 0) end
+local function checknz(ret) assert(ret ~= 0) end
+
+local new, free, protect
+
+--Using VirtualAlloc allows memory protection, but can only allocate memory in multiple-of-64K chunks.
+local USE_VIRTUALALLOC = false
+
+if ffi.os == 'Windows' then
+
+	if USE_VIRTUALALLOC then
+
+		ffi.cdef[[
+		void* VirtualAlloc(void* lpAddress, size_t dwSize, uint32_t flAllocationType, uint32_t flProtect);
+		int VirtualFree(void* lpAddress, size_t dwSize, uint32_t dwFreeType);
+		int VirtualProtect(void* lpAddress, size_t dwSize, uint32_t flNewProtect, uint32_t* lpflOldProtect);
+		]]
+
+		local PAGE_READWRITE    = 0x04
+		local PAGE_EXECUTE_READ = 0x20
+		local MEM_COMMIT  = 0x1000
+		local MEM_RESERVE = 0x2000
+		local MEM_RELEASE = 0x8000
+
+		function new(size)
+			return checkh(C.VirtualAlloc(nil, size, bit.bor(MEM_RESERVE, MEM_COMMIT), PAGE_READWRITE))
+		end
+
+		function protect(addr, size)
+			local oldprotect = ffi.new'uint32_t[1]' --because null not accepted
+			checknz(C.VirtualProtect(addr, size, PAGE_EXECUTE_READ, oldprotect))
+		end
+
+		function free(addr, size)
+			assert(size, 'size required') --on other platforms
+			checknz(C.VirtualFree(addr, 0, MEM_RELEASE))
+		end
+
+	else
+
+		local HEAP_NO_SERIALIZE          = 0x00000001
+		local HEAP_ZERO_MEMORY           = 0x00000008
+		local HEAP_CREATE_ENABLE_EXECUTE = 0x00040000
+
+		ffi.cdef[[
+		void* HeapCreate(uint32_t flOptions, size_t dwInitialSize, size_t dwMaximumSize);
+		void* HeapAlloc(void* hHeap, uint32_t dwFlags, size_t dwBytes);
+		int HeapFree(void* hHeap, uint32_t dwFlags, void* lpMem);
+		]]
+
+		local heap
+
+		function new(size)
+			heap = heap or checkh(C.HeapCreate(bit.bor(HEAP_NO_SERIALIZE, HEAP_CREATE_ENABLE_EXECUTE), 0, 0))
+			return checkh(C.HeapAlloc(heap, HEAP_ZERO_MEMORY, size))
+		end
+
+		function protect(addr, size) end
+
+		function free(addr, size)
+			assert(size, 'size required') --on other platforms
+			checknz(C.HeapFree(heap, HEAP_NO_SERIALIZE, addr))
+		end
+
+	end
+
+elseif ffi.os == 'Linux' or ffi.os == 'OSX' then
+
+	if ffi.os == 'OSX' then
+		ffi.cdef'typedef int64_t off_t;'
+	else
+		ffi.cdef'typedef long int off_t;'
+	end
+
+	ffi.cdef[[
+	void* mmap(void *addr, size_t length, int prot, int flags, int fd, off_t offset);
+	int munmap(void *addr, size_t length);
+	int mprotect(void *addr, size_t len, int prot);
+	]]
+
+	local PROT_READ  = 1
+	local PROT_WRITE = 2
+	local PROT_EXEC  = 4
+	local MAP_PRIVATE   = 2
+	local MAP_ANON = ffi.os == 'Linux' and 0x20 or 0x1000
+
+	function new(size)
+		local ret = C.mmap(nil, size, bit.bor(PROT_READ, PROT_WRITE), bit.bor(MAP_PRIVATE, MAP_ANON), -1, 0)
+		if ffi.cast('intptr_t', ret) == ffi.cast('intptr_t', -1) then
+			error(string.format('mmap errno: %d', ffi.errno()))
+		end
+		return checkh(ret)
+	end
+
+	function protect(addr, size)
+		checkz(C.mprotect(addr, size, bit.bor(PROT_READ, PROT_EXEC)))
+	end
+
+	function free(addr, size)
+		checkz(C.munmap(addr, size))
+	end
+
+end
+
+local new = function(size) --override for hooking to gc
+	local addr = new(size)
+	ffi.gc(addr, function(addr)
+		free(addr, size)
+		ffi.gc(addr, nil)
+	end)
+	return addr
+end
+
+if not ... then
+	local function test(size)
+		local addr = new(size)
+		print(addr)
+		addr = ffi.cast('int32_t*', addr)
+		assert(addr[0] == 0)
+		addr[0] = 1234 --writable
+		assert(addr[0] == 1234)
+		protect(addr, size)
+		--addr[0] = 4321 --uncomment this to get a crash (r/o memory); TODO: test if executable
+		--addr = nil; collectgarbage() --enable this to fail the assertion below
+		return addr
+	end
+	local a1 = test(64*1024*1000) --64MB
+	local a2 = test(16) --16 bytes
+	assert(a1 ~= a2) --different pages
+	a1 = nil
+	a2 = nil
+	collectgarbage() --TODO: test if finalizer was called
+end
+
+return {new = new, free = free, protect = protect}
diff --git a/dynasm/dasm_x64.lua b/dynasm/dasm_x64.lua
index b1b62022..c22ddcfd 100644
--- a/dynasm/dasm_x64.lua
+++ b/dynasm/dasm_x64.lua
@@ -8,5 +8,15 @@
 -- All the interesting stuff is there.
 ------------------------------------------------------------------------------
 
+--unload dasm_x86 if it's already loaded.
+if not package then package = {loaded = {}} end --for compat. with minilua
+local dasm_x86 = package.loaded.dasm_x86
+package.loaded.dasm_x86 = nil
+
 x64 = true -- Using a global is an ugly, but effective solution.
-return require("dasm_x86")
+local dasm_x64 = require("dasm_x86")
+
+package.loaded.dasm_x86 = dasm_x86 --put it back
+
+return dasm_x64
+
diff --git a/dynasm/dasm_x86.c b/dynasm/dasm_x86.c
new file mode 100644
index 00000000..59c3bb63
--- /dev/null
+++ b/dynasm/dasm_x86.c
@@ -0,0 +1,11 @@
+/*
+  Encoding engine to use with dasm.lua.
+
+  Compile with:
+
+    gcc dasm_x86.c -DDASM_CHECKS -shared -s -o dasm_x86.so
+*/
+
+#include "dasm_extern.h"
+#include "dasm_proto.h"
+#include "dasm_x86.h"
diff --git a/dynasm/dasm_x86.lua b/dynasm/dasm_x86.lua
index 322d77f7..20e207bb 100644
--- a/dynasm/dasm_x86.lua
+++ b/dynasm/dasm_x86.lua
@@ -5,7 +5,7 @@
 -- See dynasm.lua for full copyright notice.
 ------------------------------------------------------------------------------
 
-local x64 = x64
+local x64 = rawget(_G, "x64") --rawget so it works with strict.lua
 
 -- Module information:
 local _info = {
@@ -32,9 +32,12 @@ local bit = bit or require("bit")
 local band, bxor, shl, shr = bit.band, bit.bxor, bit.lshift, bit.rshift
 
 -- Inherited tables and callbacks.
-local g_opt, g_arch
+local g_opt, g_arch, g_map_def
 local wline, werror, wfatal, wwarn
 
+-- Global flag to generate Lua code instead of C code.
+local luamode
+
 -- Action name list.
 -- CHECK: Keep this in sync with the C code!
 local action_names = {
@@ -74,14 +77,20 @@ local map_action = {}
 local actfirst = 256-#action_names
 
 -- Action list buffer and string (only used to remove dupes).
-local actlist = {}
-local actstr = ""
+local actlist, actstr
 
--- Argument list for next dasm_put(). Start with offset 0 into action list.
-local actargs = { 0 }
+-- Argument list for next dasm_put().
+local actargs
 
 -- Current number of section buffer positions for dasm_put().
-local secpos = 1
+local secpos
+
+local function init_actionlist()
+  actlist = {}
+  actstr = ""
+  actargs = { 0 } -- Start with offset 0 into the action list.
+  secpos = 1
+end
 
 ------------------------------------------------------------------------------
 
@@ -107,7 +116,11 @@ local function writeactions(out, name)
   local last = actlist[nn] or 255
   actlist[nn] = nil -- Remove last byte.
   if nn == 0 then nn = 1 end
-  out:write("static const unsigned char ", name, "[", nn, "] = {\n")
+  if luamode then
+    out:write("local ", name, " = ffi.new('const uint8_t[", nn, "]', {\n")
+  else
+    out:write("static const unsigned char ", name, "[", nn, "] = {\n")
+  end
   local s = "  "
   for n,b in ipairs(actlist) do
     s = s..b..","
@@ -116,7 +129,11 @@ local function writeactions(out, name)
       s = "  "
     end
   end
-  out:write(s, last, "\n};\n\n") -- Add last byte back.
+  if luamode then
+    out:write(s, last, "\n})\n\n") -- Add last byte back.
+  else
+    out:write(s, last, "\n};\n\n") -- Add last byte back.
+  end
 end
 
 ------------------------------------------------------------------------------
@@ -136,7 +153,11 @@ end
 
 -- Add call to embedded DynASM C code.
 local function wcall(func, args)
-  wline(format("dasm_%s(Dst, %s);", func, concat(args, ", ")), true)
+  if luamode then
+    wline(format("dasm.%s(Dst, %s)", func, concat(args, ", ")), true)
+  else
+    wline(format("dasm_%s(Dst, %s);", func, concat(args, ", ")), true)
+  end
 end
 
 -- Delete duplicate action list chunks. A tad slow, but so what.
@@ -172,15 +193,21 @@ end
 ------------------------------------------------------------------------------
 
 -- Global label name -> global label number. With auto assignment on 1st use.
-local next_global = 10
-local map_global = setmetatable({}, { __index = function(t, name)
+local next_global, map_global
+
+local globals_meta = { __index = function(t, name)
   if not match(name, "^[%a_][%w_@]*$") then werror("bad global label") end
   local n = next_global
   if n > 246 then werror("too many global labels") end
   next_global = n + 1
   t[name] = n
   return n
-end})
+end}
+
+local function init_map_global()
+  next_global = 10
+  map_global = setmetatable({}, globals_meta)
+end
 
 -- Dump global labels.
 local function dumpglobals(out, lvl)
@@ -197,36 +224,60 @@ end
 local function writeglobals(out, prefix)
   local t = {}
   for name, n in pairs(map_global) do t[n] = name end
-  out:write("enum {\n")
-  for i=10,next_global-1 do
-    out:write("  ", prefix, gsub(t[i], "@.*", ""), ",\n")
+  if luamode then
+    local n = 0
+    for i=10,next_global-1 do
+      out:write("local ", prefix, gsub(t[i], "@.*", ""), "\t= ", n, "\n")
+      n = n + 1
+    end
+    out:write("local ", prefix, "_MAX\t= ", n, "\n") --for compatibility with the C protocol
+    out:write("local DASM_MAXGLOBAL\t= ", n, "\n")
+  else
+    out:write("enum {\n")
+    for i=10,next_global-1 do
+      out:write("  ", prefix, gsub(t[i], "@.*", ""), ",\n")
+    end
+    out:write("  ", prefix, "_MAX\n};\n")
   end
-  out:write("  ", prefix, "_MAX\n};\n")
 end
 
 -- Write global label names.
 local function writeglobalnames(out, name)
   local t = {}
   for name, n in pairs(map_global) do t[n] = name end
-  out:write("static const char *const ", name, "[] = {\n")
-  for i=10,next_global-1 do
-    out:write("  \"", t[i], "\",\n")
+  if luamode then
+    out:write("local ", name, " = {\n")
+    for i=10,next_global-1 do
+      out:write("  ", i == 10 and "[0] = " or "", "\"", t[i], "\",\n")
+    end
+    out:write("}\n")
+  else
+    out:write("static const char *const ", name, "[] = {\n")
+    for i=10,next_global-1 do
+      out:write("  \"", t[i], "\",\n")
+    end
+    out:write("  (const char *)0\n};\n")
   end
-  out:write("  (const char *)0\n};\n")
 end
 
 ------------------------------------------------------------------------------
 
 -- Extern label name -> extern label number. With auto assignment on 1st use.
-local next_extern = -1
-local map_extern = setmetatable({}, { __index = function(t, name)
+local next_extern, map_extern
+
+local extern_meta = { __index = function(t, name)
   -- No restrictions on the name for now.
   local n = next_extern
   if n < -256 then werror("too many extern labels") end
   next_extern = n - 1
   t[name] = n
   return n
-end})
+end}
+
+local function init_map_extern()
+  next_extern = -1
+  map_extern = setmetatable({}, extern_meta)
+end
 
 -- Dump extern labels.
 local function dumpexterns(out, lvl)
@@ -243,11 +294,19 @@ end
 local function writeexternnames(out, name)
   local t = {}
   for name, n in pairs(map_extern) do t[-n] = name end
-  out:write("static const char *const ", name, "[] = {\n")
-  for i=1,-next_extern-1 do
-    out:write("  \"", t[i], "\",\n")
+  if luamode then
+    out:write("local ", name, " = {\n")
+    for i=1,-next_extern-1 do
+	   out:write(i==1 and "[0] = " or "", "\"", t[i], "\",\n")
+	end
+    out:write("}\n")
+  else
+	 out:write("static const char *const ", name, "[] = {\n")
+    for i=1,-next_extern-1 do
+      out:write("  \"", t[i], "\",\n")
+    end
+    out:write("  (const char *)0\n};\n")
   end
-  out:write("  (const char *)0\n};\n")
 end
 
 ------------------------------------------------------------------------------
@@ -262,8 +321,13 @@ local map_reg_valid_index = {}	-- Int. register name -> valid index register?
 local map_reg_needrex = {}	-- Int. register name -> need rex vs. no rex.
 local reg_list = {}		-- Canonical list of int. register names.
 
-local map_type = {}		-- Type name -> { ctype, reg }
-local ctypenum = 0		-- Type number (for _PTx macros).
+local map_type			-- Type name -> { ctype, reg }
+local ctypenum			-- Type number (for _PTx macros).
+
+local function init_map_type()
+  map_type = {}
+  ctypenum = 0
+end
 
 local addrsize = x64 and "q" or "d"	-- Size for address operands.
 
@@ -631,7 +695,7 @@ end
 local function immexpr(expr)
   -- &expr (pointer)
   if sub(expr, 1, 1) == "&" then
-    return "iPJ", format("(ptrdiff_t)(%s)", sub(expr,2))
+    return "iPJ", format(luamode and "(%s)" or "(ptrdiff_t)(%s)", sub(expr,2))
   end
 
   local prefix = sub(expr, 1, 2)
@@ -837,7 +901,11 @@ local function parseoperand(param)
 	-- type[idx], type[idx].field, type->field -> [reg+offset_expr]
 	if not tp then werror("bad operand `"..param.."'") end
 	t.mode = "xm"
-	t.disp = format(tp.ctypefmt, tailr)
+	if luamode then
+	  t.disp = tp.ctypefmt(tailr)
+	else
+	  t.disp = format(tp.ctypefmt, tailr)
+	end
       else
 	t.mode, t.imm = immexpr(expr)
 	if sub(t.mode, -1) == "J" then
@@ -1202,6 +1270,8 @@ local map_op = {
   andnps_2 =	"rmo:0F55rM",
   andpd_2 =	"rmo:660F54rM",
   andps_2 =	"rmo:0F54rM",
+  fxsave_1 =	"x.:0FAE0m",
+  fxrstor_1 =	"x.:0FAE1m",
   clflush_1 =	"x.:0FAE7m",
   cmppd_3 =	"rmio:660FC2rMU",
   cmpps_3 =	"rmio:0FC2rMU",
@@ -1990,8 +2060,13 @@ if x64 then
     end
     wputop(sz, opcode, rex)
     if vreg then waction("VREG", vreg); wputxb(0) end
-    waction("IMM_D", format("(unsigned int)(%s)", op64))
-    waction("IMM_D", format("(unsigned int)((%s)>>32)", op64))
+    if luamode then
+      waction("IMM_D", format("ffi.cast(\"uintptr_t\", %s) %% 2^32", op64))
+      waction("IMM_D", format("ffi.cast(\"uintptr_t\", %s) / 2^32", op64))
+    else
+      waction("IMM_D", format("(unsigned int)(%s)", op64))
+      waction("IMM_D", format("(unsigned int)((%s)>>32)", op64))
+    end
   end
 end
 
@@ -2137,16 +2212,41 @@ map_op[".type_3"] = function(params, nparams)
   if reg and not map_reg_valid_base[reg] then
     werror("bad base register `"..(map_reg_rev[reg] or reg).."'")
   end
-  -- Add #type to defines. A bit unclean to put it in map_archdef.
-  map_archdef["#"..name] = "sizeof("..ctype..")"
+  -- Add #type to current defines table.
+  g_map_def["#"..name] = luamode and "ffi.sizeof(\""..ctype.."\")" or "sizeof("..ctype..")"
   -- Add new type and emit shortcut define.
   local num = ctypenum + 1
+  local ctypefmt
+  if luamode then
+    ctypefmt = function(tailr)
+      local index, field
+      index, field = match(tailr, "^(%b[])(.*)")
+      index = index and sub(index, 2, -2)
+      field = field or tailr
+      field = match(field, "^%->(.*)") or match(field, "^%.(.*)")
+      if not (index or field) then
+	werror("invalid syntax `"..tailr.."`")
+      end
+      local Da = index and format("Da%X(%s)", num, index)
+      local Dt = field and format("Dt%X(\"%s\")", num, field)
+      return Da and Dt and Da.."+"..Dt or Da or Dt
+    end
+  else
+    ctypefmt = format("Dt%X(%%s)", num)
+  end
   map_type[name] = {
     ctype = ctype,
-    ctypefmt = format("Dt%X(%%s)", num),
+    ctypefmt = ctypefmt,
     reg = reg,
   }
-  wline(format("#define Dt%X(_V) (int)(ptrdiff_t)&(((%s *)0)_V)", num, ctype))
+  if luamode then
+    wline(format("local Dt%X; do local ct=ffi.typeof(\"%s\"); function Dt%X(f) return ffi.offsetof(ct,f) or error(string.format(\"'struct %s' has no member named '%%s'\", f)) end; end",
+      num, ctype, num, ctype))
+    wline(format("local Da%X; do local sz=ffi.sizeof(\"%s\"); function Da%X(i) return i*sz end; end",
+      num, ctype, num))
+  else
+    wline(format("#define Dt%X(_V) (int)(ptrdiff_t)&(((%s *)0)_V)", num, ctype))
+  end
   ctypenum = num
 end
 map_op[".type_2"] = map_op[".type_3"]
@@ -2202,12 +2302,19 @@ end
 -- Setup the arch-specific module.
 function _M.setup(arch, opt)
   g_arch, g_opt = arch, opt
+  luamode = g_opt.lang == "lua"
+  init_actionlist()
+  init_map_global()
+  init_map_extern()
+  init_map_type()
 end
 
 -- Merge the core maps and the arch-specific maps.
 function _M.mergemaps(map_coreop, map_def)
   setmetatable(map_op, { __index = map_coreop })
   setmetatable(map_def, { __index = map_archdef })
+  -- Hold a ref. to map_def to store `#type` defines in.
+  g_map_def = map_def
   return map_op, map_def
 end
 
diff --git a/dynasm/dynasm.lua b/dynasm/dynasm.lua
index 145fb0cc..bdade87e 100644
--- a/dynasm/dynasm.lua
+++ b/dynasm/dynasm.lua
@@ -45,6 +45,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 -- Cache library functions.
 local type, pairs, ipairs = type, pairs, ipairs
 local pcall, error, assert = pcall, error, assert
+local select, tostring = select, tostring
 local _s = string
 local sub, match, gmatch, gsub = _s.sub, _s.match, _s.gmatch, _s.gsub
 local format, rep, upper = _s.format, _s.rep, _s.upper
@@ -54,18 +55,70 @@ local exit = os.exit
 local io = io
 local stdin, stdout, stderr = io.stdin, io.stdout, io.stderr
 
+-- Helper to update a table with the elements another table.
+local function update(dst, src)
+  if src then
+    for k,v in pairs(src) do
+      dst[k] = v
+    end
+  end
+  return dst
+end
+
 ------------------------------------------------------------------------------
 
 -- Program options.
-local g_opt = {}
+local g_opt
 
 -- Global state for current file.
 local g_fname, g_curline, g_indent, g_lineno, g_synclineno, g_arch
-local g_errcount = 0
+local g_errcount
 
 -- Write buffer for output file.
 local g_wbuffer, g_capbuffer
 
+-- Map for defines (initially empty, chains to arch-specific map).
+local map_def
+
+-- Sections names.
+local map_sections
+
+-- The opcode map. Begins as an empty map set to inherit from map_initop,
+-- It's later changed to inherit from the arch-specific map, which itself is set
+-- to inherit from map_coreop.
+local map_op
+
+-- The initial opcode map to initialize map_op with on each global reset.
+local map_initop = {}
+
+-- Dummy action flush function. Replaced with arch-specific function later.
+local function dummy_wflush(term) end
+local wflush
+
+-- Init/reset the global state for processing a new file.
+local function reset()
+  g_opt = {}
+  g_opt.dumpdef = 0
+  g_opt.include = { "" }
+  g_opt.lang = nil
+  g_opt.comment = "//|"
+  g_opt.endcomment = ""
+  g_opt.cpp = true -- Write `#line` directives
+  g_fname = nil
+  g_curline = nil
+  g_indent = ""
+  g_lineno = 0
+  g_synclineno = -1
+  g_errcount = 0
+  g_arch = nil
+  g_wbuffer = {}
+  g_capbuffer = nil
+  map_def = {}
+  map_sections = {}
+  map_op = setmetatable({}, { __index = map_initop })
+  wflush = dummy_wflush
+end
+
 ------------------------------------------------------------------------------
 
 -- Write an output line (or callback function) to the buffer.
@@ -85,15 +138,13 @@ end
 -- Resync CPP line numbers.
 local function wsync()
   if g_synclineno ~= g_lineno and g_opt.cpp then
-    wline("#line "..g_lineno..' "'..g_fname..'"')
+    if not g_opt.lang == "lua" then
+      wline("#line "..g_lineno..' "'..g_fname..'"')
+    end
     g_synclineno = g_lineno
   end
 end
 
--- Dummy action flush function. Replaced with arch-specific function later.
-local function wflush(term)
-end
-
 -- Dump all buffered output lines.
 local function wdumplines(out, buf)
   for _,line in ipairs(buf) do
@@ -171,8 +222,6 @@ end
 
 -- Core pseudo-opcodes.
 local map_coreop = {}
--- Dummy opcode map. Replaced by arch-specific map.
-local map_op = {}
 
 -- Forward declarations.
 local dostmt
@@ -180,9 +229,6 @@ local readfile
 
 ------------------------------------------------------------------------------
 
--- Map for defines (initially empty, chains to arch-specific map).
-local map_def = {}
-
 -- Pseudo-opcode to define a substitution.
 map_coreop[".define_2"] = function(params, nparams)
   if not params then return nparams == 1 and "name" or "name, subst" end
@@ -379,11 +425,11 @@ map_coreop[".include_1"] = function(params)
 end
 
 -- Make .include and conditionals initially available, too.
-map_op[".include_1"] = map_coreop[".include_1"]
-map_op[".if_1"] = map_coreop[".if_1"]
-map_op[".elif_1"] = map_coreop[".elif_1"]
-map_op[".else_0"] = map_coreop[".else_0"]
-map_op[".endif_0"] = map_coreop[".endif_0"]
+map_initop[".include_1"] = map_coreop[".include_1"]
+map_initop[".if_1"] = map_coreop[".if_1"]
+map_initop[".elif_1"] = map_coreop[".elif_1"]
+map_initop[".else_0"] = map_coreop[".else_0"]
+map_initop[".endif_0"] = map_coreop[".endif_0"]
 
 ------------------------------------------------------------------------------
 
@@ -583,9 +629,6 @@ end
 
 ------------------------------------------------------------------------------
 
--- Sections names.
-local map_sections = {}
-
 -- Pseudo-opcode to define code sections.
 -- TODO: Data sections, BSS sections. Needs extra C code and API.
 map_coreop[".section_*"] = function(params)
@@ -599,10 +642,18 @@ map_coreop[".section_*"] = function(params)
       werror("bad section name `"..name.."'")
     end
     map_sections[#map_sections+1] = name
-    wline(format("#define DASM_SECTION_%s\t%d", upper(name), sn-1))
+    if g_opt.lang == "lua" then
+      wline(format("local DASM_SECTION_%s\t= %d", upper(name), sn-1))
+    else
+      wline(format("#define DASM_SECTION_%s\t%d", upper(name), sn-1))
+    end
     map_op[opname] = function(params) g_arch.section(sn-1) end
   end
-  wline(format("#define DASM_MAXSECTION\t\t%d", #map_sections))
+  if g_opt.lang == "lua" then
+    wline(format("local DASM_MAXSECTION\t= %d", #map_sections))
+  else
+    wline(format("#define DASM_MAXSECTION\t\t%d", #map_sections))
+  end
 end
 
 -- Dump all sections.
@@ -635,7 +686,9 @@ local function loadarch(arch)
   g_arch = m_arch
   wflush = m_arch.passcb(wline, werror, wfatal, wwarn)
   m_arch.setup(arch, g_opt)
-  map_op, map_def = m_arch.mergemaps(map_coreop, map_def)
+  local arch_map_op
+  arch_map_op, map_def = m_arch.mergemaps(map_coreop, map_def)
+  map_op = setmetatable(map_op, { __index = arch_map_op })
 end
 
 -- Dump architecture description.
@@ -691,19 +744,27 @@ end
 
 -- Pseudo-opcode to set the architecture.
 -- Only initially available (map_op is replaced when called).
-map_op[".arch_1"] = function(params)
+map_initop[".arch_1"] = function(params)
   if not params then return "name" end
   local err = loadarch(params[1])
   if err then wfatal(err) end
-  wline(format("#if DASM_VERSION != %d", _info.vernum))
-  wline('#error "Version mismatch between DynASM and included encoding engine"')
-  wline("#endif")
+  if g_opt.lang == "lua" then
+    wline(format("if dasm._VERSION ~= %d then", _info.vernum))
+    wline('  error("Version mismatch between DynASM and included encoding engine")')
+    wline("end")
+  else
+    wline(format("#if DASM_VERSION != %d", _info.vernum))
+    wline('#error "Version mismatch between DynASM and included encoding engine"')
+    wline("#endif")
+  end
 end
 
 -- Dummy .arch pseudo-opcode to improve the error report.
 map_coreop[".arch_1"] = function(params)
   if not params then return "name" end
-  wfatal("duplicate .arch statement")
+  if g_arch._info.arch ~= params[1] then
+    wfatal("invalid .arch statement. arch already loaded: `", g_arch._info.arch, "`.")
+  end
 end
 
 ------------------------------------------------------------------------------
@@ -859,6 +920,9 @@ local function doline(line)
 
   -- Strip assembler comments.
   aline = gsub(aline, "//.*$", "")
+  if g_opt.lang == "lua" then
+    aline = gsub(aline, "%-%-.*$", "")
+  end
 
   -- Split line into statements at semicolons.
   if match(aline, ";") then
@@ -872,7 +936,21 @@ end
 
 -- Write DynASM header.
 local function dasmhead(out)
-  out:write(format([[
+  if not g_opt.comment then return end
+  if g_opt.lang == "lua" then
+    out:write(format([[
+--
+-- This file has been pre-processed with DynASM.
+-- %s
+-- DynASM version %s, DynASM %s version %s
+-- DO NOT EDIT! The original file is in "%s".
+--
+
+]], _info.url,
+    _info.version, g_arch._info.arch, g_arch._info.version,
+    g_fname))
+  else
+    out:write(format([[
 /*
 ** This file has been pre-processed with DynASM.
 ** %s
@@ -883,14 +961,11 @@ local function dasmhead(out)
 ]], _info.url,
     _info.version, g_arch._info.arch, g_arch._info.version,
     g_fname))
+  end
 end
 
 -- Read input file.
 readfile = function(fin)
-  g_indent = ""
-  g_lineno = 0
-  g_synclineno = -1
-
   -- Process all lines.
   for line in fin:lines() do
     g_lineno = g_lineno + 1
@@ -911,8 +986,10 @@ local function writefile(outfile)
   -- Open output file.
   if outfile == nil or outfile == "-" then
     fout = stdout
-  else
+  elseif type(outfile) == "string" then
     fout = assert(io.open(outfile, "w"))
+  else
+    fout = outfile
   end
 
   -- Write all buffered lines
@@ -927,10 +1004,6 @@ end
 
 -- Translate an input file to an output file.
 local function translate(infile, outfile)
-  g_wbuffer = {}
-  g_indent = ""
-  g_lineno = 0
-  g_synclineno = -1
 
   -- Put header.
   wline(dasmhead)
@@ -940,9 +1013,12 @@ local function translate(infile, outfile)
   if infile == "-" then
     g_fname = "(stdin)"
     fin = stdin
-  else
+  elseif type(infile) == "string" then
     g_fname = infile
     fin = assert(io.open(infile, "r"))
+  else
+    g_fname = "=(translate)"
+    fin = infile
   end
   readfile(fin)
 
@@ -974,7 +1050,7 @@ function opt_map.help()
   stdout:write("DynASM ", _info.version, " ", _info.release, "  ", _info.url, "\n")
   stdout:write[[
 
-Usage: dynasm [OPTION]... INFILE.dasc|-
+Usage: dynasm [OPTION]... INFILE.dasc|INFILE.dasl|-
 
   -h, --help           Display this help text.
   -V, --version        Display version and copyright information.
@@ -982,6 +1058,8 @@ Usage: dynasm [OPTION]... INFILE.dasc|-
   -o, --outfile FILE   Output file name (default is stdout).
   -I, --include DIR    Add directory to the include search path.
 
+  -l, --lang C|Lua     Generate C or Lua code (default C for dasc, Lua for dasl).
+
   -c, --ccomment       Use /* */ comments for assembler lines.
   -C, --cppcomment     Use // comments for assembler lines (default).
   -N, --nocomment      Suppress assembler lines in output.
@@ -1007,6 +1085,7 @@ function opt_map.version()
 end
 
 -- Misc. options.
+function opt_map.lang(args) g_opt.lang = optparam(args):lower() end
 function opt_map.outfile(args) g_opt.outfile = optparam(args) end
 function opt_map.include(args) insert(g_opt.include, 1, optparam(args)) end
 function opt_map.ccomment() g_opt.comment = "/*|"; g_opt.endcomment = " */" end
@@ -1023,6 +1102,7 @@ function opt_map.dumpdef() g_opt.dumpdef = g_opt.dumpdef + 1 end
 local opt_alias = {
   h = "help", ["?"] = "help", V = "version",
   o = "outfile", I = "include",
+  l = "lang",
   c = "ccomment", C = "cppcomment", N = "nocomment", M = "maccomment",
   L = "nolineno", F = "flushline",
   P = "dumpdef", A = "dumparch",
@@ -1038,14 +1118,43 @@ local function parseopt(opt, args)
   f(args)
 end
 
+local languages = {c = true, lua = true}
+local langext = {dasc = "c", dasl = "lua"}
+
+--Set language options (Lua or C code gen) based on file extension.
+local function setlang(infile)
+
+  -- Infer language from file extension, if `lang` not set.
+  if not g_opt.lang and type(infile) == "string" then
+    g_opt.lang = langext[match(infile, "%.([^%.]+)$")] or "c"
+  end
+
+  -- Check that the `lang` option is valid.
+  if not languages[g_opt.lang] then
+    opterror("invalid language `", tostring(g_opt.lang), "`.")
+  end
+
+  -- Adjust comment options for Lua mode.
+  if g_opt.lang == "lua" then
+    if g_opt.comment then
+      g_opt.cpp = false
+      g_opt.comment = "--|"
+      g_opt.endcomment = ""
+    end
+    -- Set initial defines only available in Lua mode.
+    local ffi = require("ffi")
+    map_def.ARCH = ffi.arch          --for `.arch ARCH`
+    map_def[upper(ffi.arch)] = 1     --for `.if X86 ...`
+    map_def.OS = ffi.os              --for `.if OS == 'Windows'`
+    map_def[upper(ffi.os)] = 1       --for `.if WINDOWS ...`
+  end
+end
+
 -- Parse arguments.
 local function parseargs(args)
-  -- Default options.
-  g_opt.comment = "//|"
-  g_opt.endcomment = ""
-  g_opt.cpp = true
-  g_opt.dumpdef = 0
-  g_opt.include = { "" }
+
+  --Reset globals.
+  reset()
 
   -- Process all option arguments.
   args.argn = 1
@@ -1073,22 +1182,140 @@ local function parseargs(args)
     opt_map.help()
   end
 
+  local infile = args[args.argn]
+
+  -- Set language options.
+  setlang(infile)
+
   -- Translate a single input file to a single output file
   -- TODO: Handle multiple files?
-  translate(args[args.argn], g_opt.outfile)
+  translate(infile, g_opt.outfile)
 end
 
 ------------------------------------------------------------------------------
 
--- Add the directory dynasm.lua resides in to the Lua module search path.
-local arg = arg
-if arg and arg[0] then
-  prefix = match(arg[0], "^(.*[/\\])")
-  if package and prefix then package.path = prefix.."?.lua;"..package.path end
+if ... == "dynasm" then -- use as module
+
+  -- Make a reusable translate() function with support for setting options.
+  local translate = function(infile, outfile, opt)
+    reset()
+    update(g_opt, opt)
+    setlang(infile)
+    if g_opt.subst then
+      for name, subst in pairs(g_opt.subst) do
+	map_def[name] = tostring(subst)
+      end
+    end
+    translate(infile, outfile)
+  end
+
+  -- Dummy file:close() method.
+  local function dummyclose()
+    return true
+  end
+
+  -- Create a pseudo-file object that implements the file:lines() method
+  -- which reads data from a string.
+  local function string_infile(s)
+    local lines = function()
+      local term =
+	match(s, "\r\n") and "\r\n" or
+	match(s, "\r") and "\r" or
+	match(s, "\n") and "\n" or ""
+      return gmatch(s, "([^\n\r]*)"..term)
+    end
+    return {lines = lines, close = dummyclose}
+  end
+
+  -- Create a pseudo-file object that implements the file:write() method
+  -- which forwards each non-empty value to a function.
+  local function func_outfile(func)
+    local function write(_, ...)
+      for i = 1, select('#', ...) do
+	local v = select(i, ...)
+	assert(type(v) == "string" or type(v) == "number", "invalid value")
+	local s = tostring(v)
+	if #s > 0 then
+	  func(s)
+	end
+      end
+      return true
+    end
+    return {write = write, close = dummyclose}
+  end
+
+  -- Create a pseudo-file object that accumulates writes to a table.
+  local function table_outfile(t)
+    return func_outfile(function(s)
+      t[#t+1] = s
+    end)
+  end
+
+  -- Translate an input file to a string.
+  local function translate_tostring(infile, opt)
+    local t = {}
+    translate(infile, table_outfile(t), opt)
+    return table.concat(t)
+  end
+
+  -- Create an iterator that translates an input file
+  -- and returns the translated file in chunks.
+  local function translate_toiter(infile, opt)
+    return coroutine.wrap(function()
+      translate(infile, func_outfile(coroutine.yield), opt)
+    end)
+  end
+
+  -- Load a dasl file and return it as a Lua chunk.
+  local function dasl_loadfile(infile, opt)
+    local opt = update({lang = "lua"}, opt)
+    local read = translate_toiter(infile, opt)
+    local filename = type(infile) == "string" and infile or "=(load)"
+    return load(read, filename)
+  end
+
+  -- Load a dasl string and return it as a Lua chunk.
+  local function dasl_loadstring(s, opt)
+    return dasl_loadfile(string_infile(s), opt)
+  end
+
+  -- Register a module loader for *.dasl files.
+  insert(package.loaders, function(modname)
+    local daslpath = gsub(gsub(package.path, "%.lua;", ".dasl;"), "%.lua$", ".dasl")
+    local path, reason = package.searchpath(modname, daslpath)
+    if not path then return reason end
+    return function()
+      local chunk = assert(dasl_loadfile(path, {comment = false}))
+      return chunk(modname)
+    end
+  end)
+
+  -- Make and return the DynASM API.
+  return {
+    --low-level intf.
+    translate = translate,
+    string_infile = string_infile,
+    func_outfile = func_outfile,
+    table_outfile = table_outfile,
+    translate_tostring = translate_tostring,
+    translate_toiter = translate_toiter,
+    --hi-level intf.
+    loadfile = dasl_loadfile,
+    loadstring = dasl_loadstring,
+  }
+
+else -- use as standalone script
+
+  -- Add the directory dynasm.lua resides in to the Lua module search path.
+  local arg = arg
+  if arg and arg[0] then
+    prefix = match(arg[0], "^(.*[/\\])")
+    if package and prefix then package.path = prefix.."?.lua;"..package.path end
+  end
+
+  -- Start DynASM.
+  parseargs{...}
+
 end
 
--- Start DynASM.
-parseargs{...}
-
 ------------------------------------------------------------------------------
-