(***********************************************************************)
(*                                                                     *)
(*                           Objective Caml                            *)
(*                                                                     *)
(*            Xavier Leroy, projet Cristal, INRIA Rocquencourt         *)
(*                                                                     *)
(*  Copyright 1996 Institut National de Recherche en Informatique et   *)
(*  en Automatique.  All rights reserved.  This file is distributed    *)
(*  under the terms of the Q Public License version 1.0.               *)
(*                                                                     *)
(***********************************************************************)

(* $Id$ *)

(* Emission of Intel 386 assembly code, MASM syntax. *)

module StringSet =
  Set.Make(struct type t = string let compare = compare end)

open Misc
open Cmm
open Arch
open Proc
open Reg
open Mach
open Linearize
open Emitaux

(* Tradeoff between code size and code speed *)

let fastcode_flag = ref true

(* Layout of the stack frame *)

let stack_offset = ref 0

let frame_size () =                     (* includes return address *)
  !stack_offset + 4 * num_stack_slots.(0) + 8 * num_stack_slots.(1) + 4

let slot_offset loc cl =
  match loc with
    Incoming n -> frame_size() + n
  | Local n ->
      if cl = 0
      then !stack_offset + n * 4
      else !stack_offset + num_stack_slots.(0) * 4 + n * 8
  | Outgoing n -> n

(* Record symbols used and defined - at the end generate extern for those 
   used but not defined *)

let symbols_defined = ref StringSet.empty
let symbols_used = ref StringSet.empty

let add_def_symbol s =
  symbols_defined := StringSet.add s !symbols_defined

let add_used_symbol s =
  symbols_used := StringSet.add s !symbols_used

let emit_symbol s =
  emit_string "_"; Emitaux.emit_symbol '$' s

(* Output a label *)

let emit_label lbl =
  emit_string "L"; emit_int lbl

(* Output an align directive. *)

let emit_align n = `	ALIGN	{emit_int n}\n`
  
(* Output a pseudo-register *)

let emit_reg = function
    { loc = Reg r } ->
      emit_string (register_name r)
  | { loc = Stack s; typ = Float } as r ->
      let ofs = slot_offset s (register_class r) in
      `REAL8 PTR {emit_int ofs}[esp]`
  | { loc = Stack s } as r ->
      let ofs = slot_offset s (register_class r) in
      `DWORD PTR {emit_int ofs}[esp]`
  | { loc = Unknown } ->
      fatal_error "Emit.emit_reg"

(* Output a reference to the lower 8 bits or lower 16 bits of a register *)

let reg_low_byte_name = [| "al"; "bl"; "cl"; "dl" |]
let reg_low_half_name = [| "ax"; "bx"; "cx"; "dx"; "si"; "di"; "bp" |]

let emit_reg8 r =
  match r.loc with
    Reg r when r < 4 -> emit_string (reg_low_byte_name.(r))
  | _ -> fatal_error "Emit.emit_reg8"

let emit_reg16 r =
  match r.loc with
    Reg r when r < 7 -> emit_string (reg_low_half_name.(r))
  | _ -> fatal_error "Emit.emit_reg16"

(* Check if the given register overlaps (same location) with the given
   array of registers *)

let register_overlap reg arr =
  try
    for i = 0 to Array.length arr - 1 do
      if reg.loc = arr.(i).loc then raise Exit
    done;
    false
  with Exit ->
    true

(* Output an addressing mode *)

let emit_signed_int d =
  if d > 0 then emit_char '+';
  if d <> 0 then emit_int d

let emit_addressing addr r n =
  match addr with
    Ibased(s, d) ->
      add_used_symbol s;
      `{emit_symbol s}{emit_signed_int d}`
  | Iindexed d ->
      `[{emit_reg r.(n)}{emit_signed_int d}]`
  | Iindexed2 d ->
      `[{emit_reg r.(n)}+{emit_reg r.(n+1)}{emit_signed_int d}]`
  | Iscaled(scale, d) ->
      `[{emit_reg r.(n)}*{emit_int scale}{emit_signed_int d}]`
  | Iindexed2scaled(scale, d) ->
      `[{emit_reg r.(n)}+{emit_reg r.(n+1)}*{emit_int scale}{emit_signed_int d}]`

(* Record live pointers at call points *)

type frame_descr =
  { fd_lbl: int;                        (* Return address *)
    fd_frame_size: int;                 (* Size of stack frame *)
    fd_live_offset: int list }          (* Offsets/regs of live addresses *)

let frame_descriptors = ref([] : frame_descr list)

let record_frame_label live =
  let lbl = new_label() in
  let live_offset = ref [] in
  Reg.Set.iter
    (function
        {typ = Addr; loc = Reg r} ->
          live_offset := ((r lsl 1) + 1) :: !live_offset
      | {typ = Addr; loc = Stack s} as reg ->
          live_offset := slot_offset s (register_class reg) :: !live_offset
      | _ -> ())
    live;
  frame_descriptors :=
    { fd_lbl = lbl;
      fd_frame_size = frame_size();
      fd_live_offset = !live_offset } :: !frame_descriptors;
  lbl

let record_frame live =
  let lbl = record_frame_label live in `{emit_label lbl}:\n`

let emit_frame fd =
  `	DWORD	{emit_label fd.fd_lbl}\n`;
  `	WORD	{emit_int fd.fd_frame_size}\n`;
  `	WORD	{emit_int (List.length fd.fd_live_offset)}\n`;
  List.iter
    (fun n ->
      `	WORD	{emit_int n}\n`)
    fd.fd_live_offset;
  emit_align 4

(* Record calls to the GC -- we've moved them out of the way *)

type gc_call =
  { gc_lbl: label;                      (* Entry label *)
    gc_return_lbl: label;               (* Where to branch after GC *)
    gc_frame: label }                   (* Label of frame descriptor *)

let call_gc_sites = ref ([] : gc_call list)

let emit_call_gc gc =
  `{emit_label gc.gc_lbl}:	call	_caml_call_gc\n`;
  `{emit_label gc.gc_frame}:	jmp	{emit_label gc.gc_return_lbl}\n`

(* Names for instructions *)

let instr_for_intop = function
    Iadd -> "add"
  | Isub -> "sub"
  | Imul -> "imul"
  | Iand -> "and"
  | Ior -> "or"
  | Ixor -> "xor"
  | Ilsl -> "sal"
  | Ilsr -> "shr"
  | Iasr -> "sar"
  | _ -> fatal_error "Emit: instr_for_intop"

let instr_for_floatop = function
    Inegf -> "fchs"
  | Iabsf -> "fabs"
  | Iaddf -> "fadd"
  | Isubf -> "fsub"
  | Imulf -> "fmul"
  | Idivf -> "fdiv"
  | Ispecific Isubfrev -> "fsubr"
  | Ispecific Idivfrev -> "fdivr"
  | _ -> fatal_error "Emit: instr_for_floatop"

let instr_for_floatop_reversed = function
    Iaddf -> "fadd"
  | Isubf -> "fsubr"
  | Imulf -> "fmul"
  | Idivf -> "fdivr"
  | Ispecific Isubfrev -> "fsub"
  | Ispecific Idivfrev -> "fdiv"
  | _ -> fatal_error "Emit: instr_for_floatop_reversed"

let instr_for_floatarithmem = function
    Ifloatadd -> "fadd"
  | Ifloatsub -> "fsub"
  | Ifloatsubrev -> "fsubr"
  | Ifloatmul -> "fmul"
  | Ifloatdiv -> "fdiv"
  | Ifloatdivrev -> "fdivr"

let name_for_cond_branch = function
    Isigned Ceq -> "e"     | Isigned Cne -> "ne"
  | Isigned Cle -> "le"    | Isigned Cgt -> "g"
  | Isigned Clt -> "l"     | Isigned Cge -> "ge"
  | Iunsigned Ceq -> "e"   | Iunsigned Cne -> "ne"
  | Iunsigned Cle -> "be"  | Iunsigned Cgt -> "a"
  | Iunsigned Clt -> "b"   | Iunsigned Cge -> "ae"
    
(* Output an = 0 or <> 0 test. *)

let output_test_zero arg =
  match arg.loc with
    Reg r -> `	test   {emit_reg arg}, {emit_reg arg}\n`
  | _     -> `	cmp    {emit_reg arg}, 0\n`

(* Deallocate the stack frame before a return or tail call *)

let output_epilogue () =
  let n = frame_size() - 4 in
  if n > 0 then `	add    esp, {emit_int n}\n`

(* Output the assembly code for an instruction *)

(* Name of current function *)
let function_name = ref ""
(* Entry point for tail recursive calls *)
let tailrec_entry_point = ref 0
(* Label of trap for out-of-range accesses *)
let range_check_trap = ref 0

let float_constants = ref ([] : (int * string) list)

let tos = phys_reg 100

let emit_instr i =
    match i.desc with
      Lend -> ()
    | Lop(Imove | Ispill | Ireload) ->
        let src = i.arg.(0) and dst = i.res.(0) in
        if src.loc <> dst.loc then begin
          if src.typ = Float then
            if src = tos then
              `	fstp	{emit_reg dst}\n`
            else begin
              `	fld	{emit_reg src}\n`;
              `	fstp	{emit_reg dst}\n`
            end
          else
            `	mov	{emit_reg dst}, {emit_reg src}\n`
        end
    | Lop(Iconst_int n) ->
        if n = Nativeint.zero then begin
          match i.res.(0).loc with
            Reg n -> `	xor	{emit_reg i.res.(0)}, {emit_reg i.res.(0)}\n`
          | _     -> `	mov	{emit_reg i.res.(0)}, 0\n`
        end else
          `	mov	{emit_reg i.res.(0)}, {emit_nativeint n}\n`
    | Lop(Iconst_float s) ->
        let f = float_of_string s in
        if f = 0.0 then
          `	fldz\n`
        else if f = 1.0 then
          `	fld1\n`
        else begin
          let lbl = new_label() in
          float_constants := (lbl, s) :: !float_constants;
          `	fld	{emit_label lbl}\n`
        end
    | Lop(Iconst_symbol s) ->
        add_used_symbol s;
        `	mov	{emit_reg i.res.(0)}, OFFSET {emit_symbol s}\n`
    | Lop(Icall_ind) ->
        `	call	{emit_reg i.arg.(0)}\n`;
        record_frame i.live
    | Lop(Icall_imm s) ->
        add_used_symbol s;
        `	call	{emit_symbol s}\n`;
        record_frame i.live
    | Lop(Itailcall_ind) ->
        output_epilogue();
        `	jmp	{emit_reg i.arg.(0)}\n`
    | Lop(Itailcall_imm s) ->
        if s = !function_name then
          `	jmp	{emit_label !tailrec_entry_point}\n`
        else begin
          output_epilogue();
          add_used_symbol s;
          `	jmp	{emit_symbol s}\n`
        end
    | Lop(Iextcall(s, alloc)) ->
        add_used_symbol s ;
        if alloc then begin
          `	mov	eax, OFFSET {emit_symbol s}\n`;
          `	call	_caml_c_call\n`;
          record_frame i.live
        end else begin
          `	call	{emit_symbol s}\n`
        end
    | Lop(Istackoffset n) ->
        if n >= 0
        then `	sub	esp, {emit_int n}\n`
        else `	add	esp, {emit_int(-n)}\n`;
        stack_offset := !stack_offset + n
    | Lop(Iload(chunk, addr)) ->
        let dest = i.res.(0) in
        begin match chunk with
          | Byte_unsigned ->
              `	movzx	{emit_reg dest}, BYTE PTR {emit_addressing addr i.arg 0}\n`
          | Byte_signed ->
              `	movsx	{emit_reg dest}, BYTE PTR {emit_addressing addr i.arg 0}\n`
          | Sixteen_unsigned ->
              `	movzx	{emit_reg dest}, WORD PTR {emit_addressing addr i.arg 0}\n`
          | Sixteen_signed ->
              `	movsx	{emit_reg dest}, WORD PTR {emit_addressing addr i.arg 0}\n`
          | Single ->
            `	fld	REAL4 PTR {emit_addressing addr i.arg 0}\n`
          | Double | Double_u ->
            `	fld	REAL8 PTR {emit_addressing addr i.arg 0}\n`
          | _ (* Word | Thirtytwo_signed | Thirtytwo_unsigned *) ->
            `	mov	{emit_reg dest}, DWORD PTR {emit_addressing addr i.arg 0}\n`
        end
    | Lop(Istore(chunk, addr)) ->
        begin match chunk with
          | Word | Thirtytwo_signed | Thirtytwo_unsigned ->
            `	mov	DWORD PTR {emit_addressing addr i.arg 1}, {emit_reg i.arg.(0)}\n`
          | Byte_unsigned | Byte_signed ->
            `	mov	BYTE PTR {emit_addressing addr i.arg 1}, {emit_reg8 i.arg.(0)}\n`
          | Sixteen_unsigned | Sixteen_signed ->
            `	mov	WORD PTR {emit_addressing addr i.arg 1}, {emit_reg16 i.arg.(0)}\n`
          | Single ->
              if i.arg.(0) = tos then
                `	fstp	REAL4 PTR {emit_addressing addr i.arg 1}\n`
              else begin
                `	fld	{emit_reg i.arg.(0)}\n`;
                `	fstp	REAL4 PTR {emit_addressing addr i.arg 1}\n`
              end
          | Double | Double_u ->
              if i.arg.(0) = tos then
                `	fstp	REAL8 PTR {emit_addressing addr i.arg 1}\n`
              else begin
                `	fld	{emit_reg i.arg.(0)}\n`;
                `	fstp	REAL8 PTR {emit_addressing addr i.arg 1}\n`
              end
        end
    | Lop(Ialloc n) ->
        if !fastcode_flag then begin
          let lbl_redo = new_label() in
          `{emit_label lbl_redo}:	mov	eax, _young_ptr\n`;
          `	sub	eax, {emit_int n}\n`;
          `	mov	_young_ptr, eax\n`;
          `	cmp	eax, _young_limit\n`;
          let lbl_call_gc = new_label() in
          let lbl_frame = record_frame_label i.live in
          `	jb	{emit_label lbl_call_gc}\n`;
          `	lea	{emit_reg i.res.(0)}, [eax+4]\n`;
          call_gc_sites :=
            { gc_lbl = lbl_call_gc;
              gc_return_lbl = lbl_redo;
              gc_frame = lbl_frame } :: !call_gc_sites
        end else begin
          begin match n with
            8  -> `	call	_caml_alloc1\n`
          | 12 -> `	call	_caml_alloc2\n`
          | 16 -> `	call	_caml_alloc3\n`
          | _  -> `	mov	eax, {emit_int n}\n`;
                  `	call	_caml_alloc\n`
          end;
          `{record_frame i.live}        lea     {emit_reg i.res.(0)}, [eax+4]\n`
        end
    | Lop(Iintop(Icomp cmp)) ->
        `	cmp	{emit_reg i.arg.(0)},{emit_reg i.arg.(1)}\n`;
        let b = name_for_cond_branch cmp in
        `	set{emit_string b}      al\n`;
        `	movzx	{emit_reg i.res.(0)}, al\n`
    | Lop(Iintop_imm(Icomp cmp, n)) ->
        `	cmp	{emit_reg i.arg.(0)}, {emit_int n}\n`;
        let b = name_for_cond_branch cmp in
        `	set{emit_string b}      al\n`;
        `	movzx	{emit_reg i.res.(0)}, al\n`
    | Lop(Iintop Icheckbound) ->
        if !range_check_trap = 0 then range_check_trap := new_label();
        `	cmp	{emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`;
        `	jbe	{emit_label !range_check_trap}\n`
    | Lop(Iintop_imm(Icheckbound, n)) ->
        if !range_check_trap = 0 then range_check_trap := new_label();
        `	cmp	{emit_reg i.arg.(0)}, {emit_int n}\n`;
        `	jbe	{emit_label !range_check_trap}\n`
    | Lop(Iintop(Idiv | Imod)) ->
        `	cdq\n`;
        `	idiv	{emit_reg i.arg.(1)}\n`
    | Lop(Iintop(Ilsl | Ilsr | Iasr as op)) ->
        (* We have i.arg.(0) = i.res.(0) and i.arg.(1) = %ecx *)
        `	{emit_string(instr_for_intop op)}	{emit_reg i.res.(0)}, cl\n`
    | Lop(Iintop op) ->
        (* We have i.arg.(0) = i.res.(0) *)
        `	{emit_string(instr_for_intop op)}	{emit_reg i.res.(0)}, {emit_reg i.arg.(1)}\n`
    | Lop(Iintop_imm(Iadd, 1) | Iintop_imm(Isub, -1)) ->
        `	inc	{emit_reg i.res.(0)}\n`
    | Lop(Iintop_imm(Iadd, -1) | Iintop_imm(Isub, 1)) ->
        `	dec	{emit_reg i.res.(0)}\n`
    | Lop(Iintop_imm(Iadd, n)) when i.arg.(0).loc <> i.res.(0).loc ->
        `	lea	{emit_reg i.res.(0)}, [{emit_reg i.arg.(0)}+{emit_int n}]\n`
    | Lop(Iintop_imm(Idiv, n)) ->
        let l = Misc.log2 n in
        let lbl = new_label() in
        output_test_zero i.arg.(0);
        `	jge	{emit_label lbl}\n`;
        `	add	{emit_reg i.arg.(0)}, {emit_int(n-1)}\n`;
        `{emit_label lbl}:	sar	{emit_reg i.arg.(0)}, {emit_int l}\n`
    | Lop(Iintop_imm(Imod, n)) ->
        let l = Misc.log2 n in
        let lbl = new_label() in
        `	mov	eax, {emit_reg i.arg.(0)}\n`;
        `	test	eax, eax\n`;
        `	jge	{emit_label lbl}\n`;
        `	add	eax, {emit_int(n-1)}\n`;
        `{emit_label lbl}:	and	eax, {emit_int(-n)}\n`;
        `	sub	{emit_reg i.arg.(0)}, eax\n`
    | Lop(Iintop_imm(op, n)) ->
        (* We have i.arg.(0) = i.res.(0) *)
        `	{emit_string(instr_for_intop op)}	{emit_reg i.res.(0)}, {emit_int n}\n`
    | Lop(Inegf | Iabsf as floatop) ->
        if i.arg.(0) <> tos then
          `	fld	{emit_reg i.arg.(0)}\n`;
        `	{emit_string(instr_for_floatop floatop)}\n`
    | Lop(Iaddf | Isubf | Imulf | Idivf | Ispecific(Isubfrev | Idivfrev)
          as floatop) ->
        if i.arg.(0) = tos && i.arg.(1) = tos then
          (* both operands on top of FP stack *)
          `	{emit_string(instr_for_floatop_reversed floatop)}\n`
        else if i.arg.(0) = tos then
          (* first operand on stack *)
          `	{emit_string(instr_for_floatop floatop)}	{emit_reg i.arg.(1)}\n`
        else if i.arg.(1) = tos then
          (* second operand on stack *)
          `	{emit_string(instr_for_floatop_reversed floatop)}	{emit_reg i.arg.(0)}\n`
        else begin
          (* both operands in memory *)
          `	fld	{emit_reg i.arg.(0)}\n`;
          `	{emit_string(instr_for_floatop floatop)}	{emit_reg i.arg.(1)}\n`
        end
    | Lop(Ifloatofint) ->
        begin match i.arg.(0).loc with
          Stack s ->
            `	fild	{emit_reg i.arg.(0)}\n`
        | _ ->
            `	push	{emit_reg i.arg.(0)}\n`;
            `	fild	DWORD PTR [esp]\n`;
            `	add	esp, 4\n`
        end
    | Lop(Iintoffloat) ->
        if i.arg.(0) <> tos then
          `	fld	{emit_reg i.arg.(0)}\n`;
        stack_offset := !stack_offset - 8;
        `	sub	esp, 8\n`;
        `	fnstcw	[esp+4]\n`;
        `	mov	eax, [esp+4]\n`;
        `	mov	ah, 12\n`;
        `	mov	[esp], eax\n`;
        `	fldcw	[esp]\n`;
        begin match i.res.(0).loc with
          Stack s ->
            `	fistp	{emit_reg i.res.(0)}\n`
        | _ ->
            `	fistp	DWORD PTR [esp]\n`;
            `	mov	{emit_reg i.res.(0)}, [esp]\n`
        end;
        `	fldcw	[esp+4]\n`;
        `	add	esp, 8\n`;
        stack_offset := !stack_offset + 8
    | Lop(Ispecific(Ilea addr)) ->
        `	lea	{emit_reg i.res.(0)}, DWORD PTR {emit_addressing addr i.arg 0}\n`
    | Lop(Ispecific(Istore_int(n, addr))) ->
        `	mov	DWORD PTR {emit_addressing addr i.arg 0},{emit_nativeint n}\n`
    | Lop(Ispecific(Istore_symbol(s, addr))) ->
        add_used_symbol s ;
        `	mov	DWORD PTR {emit_addressing addr i.arg 0},OFFSET {emit_symbol s}\n`
    | Lop(Ispecific(Ioffset_loc(n, addr))) ->
        `	add	DWORD PTR {emit_addressing addr i.arg 0},{emit_int n}\n`
    | Lop(Ispecific(Ipush)) ->
        (* Push arguments in reverse order *)
        for n = Array.length i.arg - 1 downto 0 do
          let r = i.arg.(n) in
          match r with
            {loc = Reg rn; typ = Float} ->
              `	sub	esp, 8\n`;
              `	fstp	REAL8 PTR 0[esp]\n`;
              stack_offset := !stack_offset + 8
          | {loc = Stack sl; typ = Float} ->
              let ofs = slot_offset sl 1 in
              `	push	DWORD PTR {emit_int (ofs + 4)}[esp]\n`;
              `	push	DWORD PTR {emit_int (ofs + 4)}[esp]\n`;
              stack_offset := !stack_offset + 8
          | _ ->
              `	push	{emit_reg r}\n`;
              stack_offset := !stack_offset + 4
        done
    | Lop(Ispecific(Ipush_int n)) ->
        `	push	{emit_nativeint n}\n`;
        stack_offset := !stack_offset + 4
    | Lop(Ispecific(Ipush_symbol s)) ->
        add_used_symbol s;
        `	push	OFFSET {emit_symbol s}\n`;
        stack_offset := !stack_offset + 4
    | Lop(Ispecific(Ipush_load addr)) ->
        `	push	DWORD PTR {emit_addressing addr i.arg 0}\n`;
        stack_offset := !stack_offset + 4
    | Lop(Ispecific(Ipush_load_float addr)) ->
        `	push	DWORD PTR {emit_addressing (offset_addressing addr 4) i.arg 0}\n`;
        `	push	DWORD PTR {emit_addressing addr i.arg 0}\n`;
        stack_offset := !stack_offset + 8
    | Lop(Ispecific(Ifloatarithmem(double, op, addr))) ->
        if i.arg.(0) <> tos then
          `	fld	{emit_reg i.arg.(0)}\n`;
        let size = if double then "REAL8" else "REAL4" in
        `	{emit_string(instr_for_floatarithmem op)}	{emit_string size} PTR {emit_addressing addr i.arg 1}\n`
    | Lreloadretaddr ->
        ()
    | Lreturn ->
        output_epilogue();
        `	ret\n`
    | Llabel lbl ->
        `{emit_label lbl}:\n`
    | Lbranch lbl ->
        `	jmp	{emit_label lbl}\n`
    | Lcondbranch(tst, lbl) ->
        begin match tst with
          Itruetest ->
            output_test_zero i.arg.(0);
            `	jne	{emit_label lbl}\n`
        | Ifalsetest ->
            output_test_zero i.arg.(0);
            `	je	{emit_label lbl}\n`
        | Iinttest cmp ->
            `	cmp	{emit_reg i.arg.(0)},{emit_reg i.arg.(1)}\n`;
            let b = name_for_cond_branch cmp in
            `	j{emit_string b}        {emit_label lbl}\n`
        | Iinttest_imm((Isigned Ceq | Isigned Cne | 
                        Iunsigned Ceq | Iunsigned Cne) as cmp, 0) ->
            output_test_zero i.arg.(0);
            let b = name_for_cond_branch cmp in
            `	j{emit_string b}	{emit_label lbl}\n`
        | Iinttest_imm(cmp, n) ->
            `	cmp	{emit_reg i.arg.(0)}, {emit_int n}\n`;
            let b = name_for_cond_branch cmp in
            `	j{emit_string b}	{emit_label lbl}\n`
        | Ifloattest((Ceq | Cne as cmp), neg) ->
            if i.arg.(1) <> tos then
              `	fld	{emit_reg i.arg.(1)}\n`;
            if i.arg.(0) <> tos then
              `	fld	{emit_reg i.arg.(0)}\n`;
            `	fucompp\n`;
            `	fnstsw	ax\n`;
            let neg1 = if cmp = Ceq then neg else not neg in
            if neg1 then begin          (* branch if different *)
              `	and	ah, 68\n`;
              `	xor	ah, 64\n`;
              `	jne	{emit_label lbl}\n`
            end else begin              (* branch if equal *)
              `	and	ah, 69\n`;
              `	cmp	ah, 64\n`;
              `	je	{emit_label lbl}\n`
            end
        | Ifloattest(cmp, neg) ->
            let actual_cmp =
              if i.arg.(0) = tos && i.arg.(1) = tos then begin
                (* both args on top of FP stack *)
                `	fcompp\n`;
                cmp
              end else if i.arg.(0) = tos then begin
                (* first arg on top of FP stack *)
                `	fcomp	{emit_reg i.arg.(1)}\n`;
                cmp
              end else if i.arg.(1) = tos then begin
                (* second arg on top of FP stack *)
                `	fcomp	{emit_reg i.arg.(0)}\n`;
                Cmm.swap_comparison cmp
              end else begin
                `	fld	{emit_reg i.arg.(0)}\n`;
                `	fcomp	{emit_reg i.arg.(1)}\n`;
                cmp
              end in
            `	fnstsw	ax\n`;
            begin match actual_cmp with
              Cle ->
                `	and	ah, 69\n`;
                `	dec	ah\n`;
                `	cmp	ah, 64\n`;
                if neg
                then `	jae	`
                else `	jb	`
            | Cge ->
                `	and	ah, 5\n`;
                if neg
                then `	jne	`
                else `	je	`
            | Clt ->
                `	and	ah, 69\n`;
                `	cmp	ah, 1\n`;
                if neg
                then `	jne	`
                else `	je	`
            | Cgt ->
                `	and	ah, 69\n`;
                if neg
                then `	jne	`
                else `	je	`
            | _ -> fatal_error "Emit_i386: floattest"
            end;
            `{emit_label lbl}\n`
        | Ioddtest ->
            `	test	{emit_reg i.arg.(0)}, 1\n`;
            `	jne	{emit_label lbl}\n`
        | Ieventest ->
            `	test	{emit_reg i.arg.(0)}, 1\n`;
            `	je	{emit_label lbl}\n`
        end
    | Lcondbranch3(lbl0, lbl1, lbl2) ->
            `	cmp	{emit_reg i.arg.(0)}, 1\n`;
            begin match lbl0 with
              None -> ()
            | Some lbl -> `	jb	{emit_label lbl}\n`
            end;
            begin match lbl1 with
              None -> ()
            | Some lbl -> `	je	{emit_label lbl}\n`
            end;
            begin match lbl2 with
              None -> ()
            | Some lbl -> `	jg	{emit_label lbl}\n`
            end
    | Lswitch jumptbl ->
        let lbl = new_label() in
        `	jmp	[{emit_reg i.arg.(0)} * 4 + {emit_label lbl}]\n`;
        `	.DATA\n`;
        `{emit_label lbl}`;
        for i = 0 to Array.length jumptbl - 1 do
          `	DWORD	{emit_label jumptbl.(i)}\n`
        done;
        `	.CODE\n`
    | Lsetuptrap lbl ->
        `	call	{emit_label lbl}\n`
    | Lpushtrap ->
        `	push	_caml_exception_pointer\n`;
        `	mov	_caml_exception_pointer, esp\n`;
        stack_offset := !stack_offset + 8
    | Lpoptrap ->
        `	pop	_caml_exception_pointer\n`;
        `	add	esp, 4\n`;
        stack_offset := !stack_offset - 8
    | Lraise ->
        `	mov	esp, _caml_exception_pointer\n`;
        `	pop	_caml_exception_pointer\n`;
        `	ret\n`

let rec emit_all i =
  match i.desc with Lend -> () | _ -> emit_instr i; emit_all i.next

(* Emission of the floating-point constants *)

let emit_float s =
  (* MASM doesn't like floating-point constants such as 2e9.
     Turn them into 2.0e9. *)
  let pos_e = ref (-1) and pos_dot = ref (-1) in
  for i = 0 to String.length s - 1 do
    match s.[i] with
      'e'|'E' -> pos_e := i
    | '.'     -> pos_dot := i
    | _       -> ()
  done;
  if !pos_dot < 0 && !pos_e >= 0 then begin
    emit_string (String.sub s 0 !pos_e);
    emit_string ".0";
    emit_string (String.sub s !pos_e (String.length s - !pos_e))
  end else
    emit_string s

let emit_float_constant (lbl, cst) =
  `{emit_label lbl}     REAL8   {emit_float cst}\n`

(* Emission of a function declaration *)

let fundecl fundecl =
  function_name := fundecl.fun_name;
  fastcode_flag := fundecl.fun_fast;
  tailrec_entry_point := new_label();
  stack_offset := 0;
  float_constants := [];
  call_gc_sites := [];
  range_check_trap := 0;
  `	.CODE\n`;
  add_def_symbol fundecl.fun_name;
  emit_align 4;
  `	PUBLIC	{emit_symbol fundecl.fun_name}\n`;
  `{emit_symbol fundecl.fun_name}:\n`;
  let n = frame_size() - 4 in
  if n > 0 then
    `	sub	esp, {emit_int n}\n`;
  `{emit_label !tailrec_entry_point}:\n`;
  emit_all fundecl.fun_body;
  List.iter emit_call_gc !call_gc_sites;
  if !range_check_trap > 0 then
    `{emit_label !range_check_trap}:	jmp	_array_bound_error\n`;
  begin match !float_constants with
    [] -> ()
  | _  ->
      `	.DATA\n`;
      List.iter emit_float_constant !float_constants;
      float_constants := []
  end

(* Emission of data *)

let emit_item = function
    Cdefine_symbol s ->
      add_def_symbol s ;
      `	PUBLIC	{emit_symbol s}\n`;
      `{emit_symbol s} LABEL DWORD\n`
  | Cdefine_label lbl ->
      `{emit_label (100000 + lbl)} `
  | Cint8 n ->
      `	BYTE	{emit_int n}\n`
  | Cint16 n ->
      `	WORD	{emit_int n}\n`
  | Cint n ->
      `	DWORD	{emit_nativeint n}\n`
  | Cint32 n ->
      `	DWORD	{emit_nativeint n}\n`
  | Csingle f ->
      `	REAL4	{emit_float f}\n`
  | Cdouble f ->
      `	REAL8	{emit_float f}\n`
  | Csymbol_address s ->
      add_used_symbol s ;
      `	DWORD	{emit_symbol s}\n`
  | Clabel_address lbl ->
      `	DWORD	{emit_label (100000 + lbl)}\n`
  | Cstring s ->
      emit_bytes_directive "	BYTE	" s
  | Cskip n ->
      if n > 0 then `	BYTE	{emit_int n} DUP (?)\n`
  | Calign n ->
      emit_align n

let data l =
  `	.DATA\n`;
  List.iter emit_item l

(* Beginning / end of an assembly file *)

let begin_assembly() =
  `.386\n`;
  `	.MODEL	FLAT\n\n`;
  `	EXTERN _young_ptr: DWORD\n`;
  `	EXTERN _young_limit: DWORD\n`;
  `	EXTERN _caml_exception_pointer: DWORD\n`;
  `	EXTERN _caml_call_gc: PROC\n`;
  `	EXTERN _caml_c_call: PROC\n`;
  `	EXTERN _caml_alloc: PROC\n`;
  `	EXTERN _caml_alloc1: PROC\n`;
  `	EXTERN _caml_alloc2: PROC\n`;
  `	EXTERN _caml_alloc3: PROC\n`;
  `	EXTERN _array_bound_error: PROC\n`;
  `	.DATA\n`;
  let lbl_begin = Compilenv.current_unit_name() ^ "_data_begin" in
  add_def_symbol lbl_begin;
  `	PUBLIC	{emit_symbol lbl_begin}\n`;
  `{emit_symbol lbl_begin}	LABEL	DWORD\n`;
  `	.CODE\n`;
  let lbl_begin = Compilenv.current_unit_name() ^ "_code_begin" in
  add_def_symbol lbl_begin;
  `	PUBLIC	{emit_symbol lbl_begin}\n`;
  `{emit_symbol lbl_begin}	LABEL	DWORD\n`

let end_assembly() =
  `	.CODE\n`;
  let lbl_end = Compilenv.current_unit_name() ^ "_code_end" in
  add_def_symbol lbl_end;
  `	PUBLIC	{emit_symbol lbl_end}\n`;
  `{emit_symbol lbl_end}	LABEL	DWORD\n`;
  `	.DATA\n`;
  let lbl_end = Compilenv.current_unit_name() ^ "_data_end" in
  add_def_symbol lbl_end;
  `	PUBLIC	{emit_symbol lbl_end}\n`;
  `{emit_symbol lbl_end}	LABEL	DWORD\n`;
  let lbl = Compilenv.current_unit_name() ^ "_frametable" in
  add_def_symbol lbl;
  `	PUBLIC	{emit_symbol lbl}\n`;
  `{emit_symbol lbl}	DWORD	{emit_int (List.length !frame_descriptors)}\n`;
  List.iter emit_frame !frame_descriptors;
  frame_descriptors := [];
  `\n;External functions\n\n`;
  StringSet.iter
    (fun s ->
      if not (StringSet.mem s !symbols_defined) then
        `	EXTERN	{emit_symbol s}: PROC\n`)
    !symbols_used;
  symbols_used := StringSet.empty;
  symbols_defined := StringSet.empty;
  `END\n`