protobuf/upb/pb/decoder_x64.dasc

|//
|// upb - a minimalist implementation of protocol buffers.
|//
|// Copyright (c) 2011 Google Inc.  See LICENSE for details.
|// Author: Josh Haberman <jhaberman@gmail.com>
|//
|// JIT compiler for upb_decoder on x86.  Given a upb_handlers object,
|// generates code specialized to parsing the specific message and
|// calling specific handlers.
|//
|// Since the JIT can call other functions (the JIT'ted code is not a leaf
|// function) we must respect alignment rules.  On OS X, this means aligning
|// the stack to 16 bytes.

#define UPB_NONE -1
#define UPB_MULTIPLE -2
#define UPB_TOPLEVEL_ONE -3

#include <sys/mman.h>
#include "dynasm/dasm_proto.h"
#include "dynasm/dasm_x86.h"

#ifndef MAP_ANONYMOUS
# define MAP_ANONYMOUS MAP_ANON
#endif

// We map into the low 32 bits when we can, but if this is not available
// (like on OS X) we take what we can get.  It's not required for correctness,
// it's just a performance thing that makes it more likely that our jumps
// can be rel32 (i.e. within 32-bits of our pc) instead of the longer
// sequence required for other jumps (see callp).
#ifndef MAP_32BIT
#define MAP_32BIT 0
#endif

// To debug JIT-ted code with GDB we need to tell GDB about the JIT-ted code
// at runtime.  GDB 7.x+ has defined an interface for doing this, and these
// structure/function defintions are copied out of gdb/jit.h
//
// We need to give GDB an ELF file at runtime describing the symbols we have
// generated.  To avoid implementing the ELF format, we generate an ELF file
// at compile-time and compile it in as a character string.  We can replace
// a few key constants (address of JIT-ted function and its size) by looking
// for a few magic numbers and doing a dumb string replacement.

#ifndef __APPLE__
#include "upb/pb/jit_debug_elf_file.h"

typedef enum
{
  GDB_JIT_NOACTION = 0,
  GDB_JIT_REGISTER,
  GDB_JIT_UNREGISTER
} jit_actions_t;

typedef struct gdb_jit_entry {
  struct gdb_jit_entry *next_entry;
  struct gdb_jit_entry *prev_entry;
  const char *symfile_addr;
  uint64_t symfile_size;
} gdb_jit_entry;

typedef struct {
  uint32_t version;
  uint32_t action_flag;
  gdb_jit_entry *relevant_entry;
  gdb_jit_entry *first_entry;
} gdb_jit_descriptor;

gdb_jit_descriptor __jit_debug_descriptor = {1, GDB_JIT_NOACTION, NULL, NULL};

void __attribute__((noinline)) __jit_debug_register_code() { __asm__ __volatile__(""); }

void upb_reg_jit_gdb(upb_decoder *d) {
  // Create debug info.
  size_t elf_len = upb_pb_jit_debug_elf_file_o_len;
  d->debug_info = malloc(elf_len);
  memcpy(d->debug_info, upb_pb_jit_debug_elf_file_o, elf_len);
  uint64_t *p = (void*)d->debug_info;
  for (; (void*)(p+1) <= (void*)d->debug_info + elf_len; ++p) {
    if (*p == 0x12345678) { *p = (uintptr_t)d->jit_code; }
    if (*p == 0x321) { *p = d->jit_size; }
  }

  // Register the JIT-ted code with GDB.
  gdb_jit_entry *e = malloc(sizeof(gdb_jit_entry));
  e->next_entry = __jit_debug_descriptor.first_entry;
  e->prev_entry = NULL;
  if (e->next_entry) e->next_entry->prev_entry = e;
  e->symfile_addr = d->debug_info;
  e->symfile_size = elf_len;
  __jit_debug_descriptor.first_entry = e;
  __jit_debug_descriptor.relevant_entry = e;
  __jit_debug_descriptor.action_flag = GDB_JIT_REGISTER;
  __jit_debug_register_code();
}

#else

void upb_reg_jit_gdb(upb_decoder *d) {
  (void)d;
}

#endif

|.arch x64
|.actionlist upb_jit_actionlist
|.globals UPB_JIT_GLOBAL_
|.globalnames upb_jit_globalnames
|
|// Calling conventions.  Note -- this will need to be changed for
|// Windows, which uses a different calling convention!
|.define ARG1_64,   rdi
|.define ARG2_8,    sil
|.define ARG2_32,   esi
|.define ARG2_64,   rsi
|.define ARG3_8,    dl
|.define ARG3_32,   edx
|.define ARG3_64,   rdx
|.define ARG4_64,   rcx
|.define ARG5_32,   r8d
|
|// Register allocation / type map.
|// ALL of the code in this file uses these register allocations.
|// When we "call" within this file, we do not use regular calling
|// conventions, but of course when calling to user callbacks we must.
|.define PTR,       rbx
|.define CLOSURE,   r12
|.type   FRAME,     upb_dispatcher_frame, r13
|.type   STRREF,    upb_strref, r14
|.type   DECODER,   upb_decoder, r15
|.type   STDARRAY,  upb_stdarray
|
|.macro callp, addr
|| if ((uintptr_t)addr < 0xffffffff) {
     |  call   &addr
|| } else {
     |  mov64  rax, (uintptr_t)addr
     |  call   rax
|| }
|.endmacro
|
|// Checks PTR for end-of-buffer.
|.macro check_eob, m
|  cmp   PTR, DECODER->effective_end
|| if (m->is_group) {
     |  jae  ->exit_jit
|| } else {
     |  jae  =>m->jit_endofbuf_pclabel
|| }
|.endmacro
|
|// Decodes varint from [PTR + offset] -> ARG3.
|// Saves new pointer as rax.
|.macro decode_loaded_varint, offset
|  // Check for <=2 bytes inline, otherwise jump to 2-10 byte decoder.
|  lea    rax, [PTR + offset + 1]
|  mov    ARG3_32, ecx
|  and    ARG3_32, 0x7f
|  test   cl, cl
|  jns    >9
|  lea    rax, [PTR + offset + 2]
|  movzx  esi, ch
|  and    esi, 0x7f
|  shl    esi, 7
|  or     ARG3_32, esi
|  test   cx, cx
|  jns    >9
|  mov    ARG1_64, rax
|  mov    ARG2_32, ARG3_32
|  callp  upb_vdecode_max8_fast
|  test   rax, rax
|  jz     ->exit_jit   // >10-byte varint.
|9:
|.endmacro
|
|.macro decode_varint, offset
|  mov    ecx, dword [PTR + offset]
|  decode_loaded_varint offset
|  mov    PTR, rax
|.endmacro
|
|// Decode the tag -> edx.
|// Could specialize this by avoiding the value masking: could just key the
|// table on the raw (length-masked) varint to save 3-4 cycles of latency.
|// Currently only support tables where all entries are in the array part.
|.macro dyndispatch_, m
|=>m->jit_dyndispatch_pclabel:
|  decode_loaded_varint, 0
|  mov  ecx, edx
|  shr  ecx, 3
|  and  edx, 0x7
|  cmp  ecx, m->max_field_number  // Bounds-check the field.
|  ja   ->exit_jit                // In the future; could be unknown label
|| if ((uintptr_t)m->tablearray < 0xffffffff) {
|    mov  rax, qword [rcx*8 + m->tablearray]  // TODO: support hybrid array/hash tables.
|| } else {
|    mov64  rax, (uintptr_t)m->tablearray
|    mov  rax, qword [rax + rcx*8]
|| }
|  jmp  rax  // Dispatch: unpredictable jump.
|.endmacro
|
|.if 1
|  // Replicated dispatch: larger code, but better branch prediction.
|  .define dyndispatch, dyndispatch_
|.else
|  .macro dyndispatch, m
|    jmp =>m->jit_dyndispatch_pclabel
|  .endmacro
|.endif
|
|// Push a stack frame (not the CPU stack, the upb_decoder stack).
|.macro pushframe, f, end_offset_, is_sequence_
|  lea   rax, [FRAME + sizeof(upb_dispatcher_frame)]  // rax for shorter addressing.
|  cmp   rax, qword DECODER->dispatcher.limit
|  jae   ->exit_jit  // Frame stack overflow.
|  mov   qword FRAME:rax->f, f
|  mov   dword FRAME:rax->end_ofs, end_offset_
|  mov   byte FRAME:rax->is_sequence, is_sequence_
|  mov   DECODER->dispatcher.top, rax
|  mov   FRAME, rax
|.endmacro
|
|.macro popframe, m
|  sub   FRAME, sizeof(upb_dispatcher_frame)
|  mov   DECODER->dispatcher.top, FRAME
|  setmsgend  m
|  mov   CLOSURE, FRAME->closure
|.endmacro
|
|.macro setmsgend, m
|    mov    rsi, DECODER->jit_end
|| if (m->is_group) {
|    mov64  rax, 0xffffffffffffffff
|    mov    qword DECODER->delim_end, rax
|    mov    DECODER->effective_end, rsi
|| } else {
|    // Could store a correctly-biased version in the frame, at the cost of
|    // a larger stack.
|    mov    eax, dword FRAME->end_ofs
|    add    rax, qword DECODER->buf
|    mov    DECODER->delim_end, rax  // delim_end = d->buf + f->end_ofs
|    cmp    rax, rsi
|    jb     >8
|    mov    rax, rsi                  // effective_end = min(d->delim_end, d->jit_end)
|8:
|    mov    DECODER->effective_end, rax
|| }
|.endmacro
|
|// rax contains the tag, compare it against "tag", but since it is a varint
|// we must only compare as many bytes as actually have data.
|.macro checktag, tag
|| switch (upb_value_size(tag)) {
||    case 1:
|       cmp   cl, tag
||      break;
||    case 2:
|       cmp   cx, tag
||      break;
||    case 3:
|       and   ecx, 0xffffff  // 3 bytes
|       cmp   rcx, tag
||    case 4:
|       cmp   ecx, tag
||      break;
||    case 5:
|       mov64 rdx, 0xffffffffff  // 5 bytes
|       and   rcx, rdx
|       cmp   rcx, tag
||      break;
||    default: abort();
||  }
|.endmacro
|
|// TODO: optimize for 0 (xor) and 32-bits.
|.macro loadfval, f
||#ifndef NDEBUG
||// Since upb_value carries type information in debug mode
||// only, we need to pass the arguments slightly differently.
|    mov ARG3_32, f->fval.type
||#endif
|| if (f->fval.val.uint64 == 0) {
|    xor     ARG2_32, ARG2_32
|| } else if (f->fval.val.uint64 < 0xffffffff) {
|    mov     ARG2_32, f->fval.val.uint64
|| } else {
|    mov64   ARG2_64, f->fval.val.uint64
|| }
|.endmacro
|
|.macro sethas, reg, hasbit
|| if (hasbit >= 0) {
|    or   byte [reg + (hasbit / 8)], (1 << (hasbit % 8))
|| }
|.endmacro


#include <stdlib.h>
#include "upb/pb/varint.h"
#include "upb/msg.h"

// Decodes the next val into ARG3, advances PTR.
static void upb_decoder_jit_decodefield(upb_decoder *d, upb_mhandlers *m,
                                        uint8_t type, size_t tag_size) {
  // Decode the value into arg 3 for the callback.
  switch (type) {
    case UPB_TYPE(DOUBLE):
    case UPB_TYPE(FIXED64):
    case UPB_TYPE(SFIXED64):
      |  mov  ARG3_64, qword [PTR + tag_size]
      |  add  PTR, 8 + tag_size
      break;

    case UPB_TYPE(FLOAT):
    case UPB_TYPE(FIXED32):
    case UPB_TYPE(SFIXED32):
      |  mov  ARG3_32, dword [PTR + tag_size]
      |  add  PTR, 4 + tag_size
      break;

    case UPB_TYPE(BOOL):
      // Can't assume it's one byte long, because bool must be wire-compatible
      // with all of the varint integer types.
      |  decode_varint  tag_size
      |  test  ARG3_64, ARG3_64
      |  setne ARG3_8   // Other bytes left with val, should be ok.
      break;

    case UPB_TYPE(INT64):
    case UPB_TYPE(UINT64):
    case UPB_TYPE(INT32):
    case UPB_TYPE(UINT32):
    case UPB_TYPE(ENUM):
      |  decode_varint  tag_size
      break;

    case UPB_TYPE(SINT64):
      // 64-bit zig-zag decoding.
      |  decode_varint  tag_size
      |  mov  rax, ARG3_64
      |  shr  ARG3_64, 1
      |  and  rax, 1
      |  neg  rax
      |  xor  ARG3_64, rax
      break;

    case UPB_TYPE(SINT32):
      // 32-bit zig-zag decoding.
      |  decode_varint  tag_size
      |  mov  eax, ARG3_32
      |  shr  ARG3_32, 1
      |  and  eax, 1
      |  neg  eax
      |  xor  ARG3_32, eax
      break;

    case UPB_TYPE(STRING):
    case UPB_TYPE(BYTES):
      // We only handle the case where the entire string is in our current
      // buf, which sidesteps any security problems.  The C path has more
      // robust checks.
      |  mov  ecx, dword [PTR + tag_size]
      |  decode_loaded_varint tag_size
      |  mov  rdi, rax
      |  add  rdi, ARG3_64
      |  mov  STRREF->len, ARG3_32
      |  mov  STRREF->ptr, rax
      |  sub  rax, DECODER->buf
      |  add  eax, DECODER->bufstart_ofs  // = d->ptr - d->buf + d->bufstart_ofs
      |  mov  STRREF->stream_offset, eax
      |  mov  ARG3_64, STRREF
      |  cmp  rdi, DECODER->effective_end
      |  ja   ->exit_jit   // Can't deliver, whole string not in buf.
      |  mov  PTR, rdi
      break;

    case UPB_TYPE_ENDGROUP:  // A pseudo-type.
      |  add  PTR, tag_size
      |  jmp  =>m->jit_endofmsg_pclabel
      return;

    // Will dispatch callbacks and call submessage in a second.
    case UPB_TYPE(MESSAGE):
      |  decode_varint  tag_size
      break;
    case UPB_TYPE(GROUP):
      |  add  PTR, tag_size
      break;

    default: abort();
  }
}

#if 0
// These appear not to speed things up, but keeping around for
// further experimentation.
static void upb_decoder_jit_doappend(upb_decoder *d, uint8_t size,
                                     upb_fhandlers *f) {
  |  mov   eax, STDARRAY:ARG1_64->len
  |  cmp   eax, STDARRAY:ARG1_64->size
  |  jne   >2
  // If array is full, fall back to actual function.
  |  loadfval f
  |  callp  f->value
  |  jmp   >3
  |2:
  |  mov   rcx, STDARRAY:ARG1_64->ptr
  |  mov   esi, eax
  |  add   eax, 1

  switch (size) {
    case 8:
      |  mov   [rcx + rsi * 8], ARG3_64
      break;

    case 4:
      |  mov   [rcx + rsi * 4], ARG3_32
      break;

    case 1:
      |  mov   [rcx + rsi * 4], ARG3_8
      break;
  }

  |  mov   STDARRAY:ARG1_64->len, eax
  |3:
}
#endif

static void upb_decoder_jit_callcb(upb_decoder *d, upb_fhandlers *f) {
  // Call callbacks.
  if (upb_issubmsgtype(f->type)) {
    if (f->type == UPB_TYPE(MESSAGE)) {
      |   mov   rsi, PTR
      |   sub   rsi, DECODER->buf
      |   add   esi, ARG3_32   // = (d->ptr - d->buf) + delim_len
    } else {
      assert(f->type == UPB_TYPE(GROUP));
      |   mov   esi, UPB_NONDELIMITED
    }
    |  pushframe  f, esi, false

    // Call startsubmsg handler (if any).
    if (f->startsubmsg) {
      // upb_sflow_t startsubmsg(void *closure, upb_value fval)
      |  mov   ARG1_64, CLOSURE
      |  loadfval f
      |  callp f->startsubmsg
      |  mov  CLOSURE, rdx
    }
    |  mov   qword FRAME->closure, CLOSURE

    const upb_mhandlers *sub_m = upb_fhandlers_getsubmsg(f);
    if (sub_m->jit_parent_field_done_pclabel != UPB_MULTIPLE) {
      |  jmp   =>sub_m->jit_startmsg_pclabel;
    } else {
      |  call  =>sub_m->jit_startmsg_pclabel;
    }

    |=>f->jit_submsg_done_pclabel:

    // Call endsubmsg handler (if any).
    if (f->endsubmsg) {
      // upb_flow_t endsubmsg(void *closure, upb_value fval);
      |  mov   ARG1_64, CLOSURE
      |  loadfval  f
      |  callp f->endsubmsg
    }
    |   popframe upb_fhandlers_getmsg(f)

  } else {
    |  mov ARG1_64, CLOSURE
    // Test for callbacks we can specialize.
    // Can't switch() on function pointers.
    if (f->value == &upb_stdmsg_setint64 ||
        f->value == &upb_stdmsg_setuint64 ||
        f->value == &upb_stdmsg_setptr ||
        f->value == &upb_stdmsg_setdouble) {
      const upb_fielddef *fd = upb_value_getfielddef(f->fval);
      |  mov   [ARG1_64 + fd->offset], ARG3_64
    } else if (f->value == &upb_stdmsg_setint32 ||
               f->value == &upb_stdmsg_setuint32 ||
               f->value == &upb_stdmsg_setfloat) {
      const upb_fielddef *fd = upb_value_getfielddef(f->fval);
      |  mov   [ARG1_64 + fd->offset], ARG3_32
    } else if (f->value == &upb_stdmsg_setbool) {
      const upb_fielddef *fd = upb_value_getfielddef(f->fval);
      |  mov   [ARG1_64 + fd->offset], ARG3_8
#if 0
    // These appear not to speed things up, but keeping around for
    // further experimentation.
    } else if (f->value == &upb_stdmsg_setint64_r ||
        f->value == &upb_stdmsg_setuint64_r ||
        f->value == &upb_stdmsg_setptr_r ||
        f->value == &upb_stdmsg_setdouble_r) {
      upb_decoder_jit_doappend(d, 8, f);
    } else if (f->value == &upb_stdmsg_setint32_r ||
               f->value == &upb_stdmsg_setuint32_r ||
               f->value == &upb_stdmsg_setfloat_r) {
      upb_decoder_jit_doappend(d, 4, f);
    } else if (f->value == &upb_stdmsg_setbool_r) {
      upb_decoder_jit_doappend(d, 1, f);
#endif
    } else {
      // Load closure and fval into arg registers.
      ||#ifndef NDEBUG
      ||// Since upb_value carries type information in debug mode
      ||// only, we need to pass the arguments slightly differently.
      |    mov ARG4_64, ARG3_64
      |    mov ARG5_32, upb_types[f->type].inmemory_type
      ||#endif
      |  loadfval f
      |  callp  f->value
    }
    |  sethas CLOSURE, f->valuehasbit
  }
  // TODO: Handle UPB_SKIPSUBMSG, UPB_BREAK
}

// PTR should point to the beginning of the tag.
static void upb_decoder_jit_field(upb_decoder *d, uint32_t tag, uint32_t next_tag,
                                  upb_mhandlers *m,
                                  upb_fhandlers *f, upb_fhandlers *next_f) {
  // PC-label for the dispatch table.
  // We check the wire type (which must be loaded in edx) because the
  // table is keyed on field number, not type.
  |=>f->jit_pclabel:
  |  cmp  edx, (tag & 0x7)
  |  jne  ->exit_jit     // In the future: could be an unknown field or packed.
  |=>f->jit_pclabel_notypecheck:
  if (f->repeated) {
    |  mov   esi, FRAME->end_ofs
    |  pushframe  f, esi, true
    if (f->startseq) {
      |  mov   ARG1_64, CLOSURE
      |  loadfval f
      |  callp f->startseq
      |  mov   CLOSURE, rdx
    }
    |  mov   qword FRAME->closure, CLOSURE
  }

  |1:  // Label for repeating this field.

  upb_decoder_jit_decodefield(d, m, f->type, upb_value_size(tag));
  upb_decoder_jit_callcb(d, f);

  // Epilogue: load next tag, check for repeated field.
  |  check_eob   m
  |  mov         rcx, qword [PTR]
  if (f->repeated) {
    |  checktag  tag
    |  je  <1
    if (f->endseq) {
      |  mov   ARG1_64, CLOSURE
      |  loadfval f
      |  callp f->endseq
    }
    |  popframe m
  }
  if (next_tag != 0) {
    |  checktag  next_tag
    |  je  =>next_f->jit_pclabel_notypecheck
  }

  // Fall back to dynamic dispatch.
  |  dyndispatch  m
  |1:
}

static int upb_compare_uint32(const void *a, const void *b) {
  // TODO: always put ENDGROUP at the end.
  return *(uint32_t*)a - *(uint32_t*)b;
}

static void upb_decoder_jit_msg(upb_decoder *d, upb_mhandlers *m) {
  |=>m->jit_startmsg_pclabel:

  if (m->jit_parent_field_done_pclabel == UPB_MULTIPLE) {
    // There was a call to get here, so we need to align the stack.
    |  sub  rsp, 8
  }
  // Call startmsg handler (if any):
  if (m->startmsg) {
    // upb_flow_t startmsg(void *closure);
    |  mov   ARG1_64, FRAME->closure
    |  callp m->startmsg
    // TODO: Handle UPB_SKIPSUBMSG, UPB_BREAK
  }

  |  setmsgend  m
  |  check_eob   m
  |  mov    ecx, dword [PTR]
  |  dyndispatch_ m

  // --------- New code section (does not fall through) ------------------------

  // Emit code for parsing each field (dynamic dispatch contains pointers to
  // all of these).

  // Create an ordering over the fields (inttable ordering is undefined).
  int num_keys = upb_inttable_count(&m->fieldtab);
  uint32_t *keys = malloc(num_keys * sizeof(*keys));
  int idx = 0;
  for(upb_inttable_iter i = upb_inttable_begin(&m->fieldtab); !upb_inttable_done(i);
      i = upb_inttable_next(&m->fieldtab, i)) {
    keys[idx++] = upb_inttable_iter_key(i);
  }
  qsort(keys, num_keys, sizeof(uint32_t), &upb_compare_uint32);

  upb_fhandlers *last_f = NULL;
  uint32_t last_tag = 0;
  for(int i = 0; i < num_keys; i++) {
    uint32_t key = keys[i];
    upb_fhandlers *f = upb_inttable_lookup(&m->fieldtab, key);
    uint32_t tag = upb_vencode32(key);
    if (last_f) upb_decoder_jit_field(d, last_tag, tag, m, last_f, f);
    last_tag = tag;
    last_f = f;
  }
  upb_decoder_jit_field(d, last_tag, 0, m, last_f, NULL);

  free(keys);

  // --------- New code section (does not fall through) ------------------------

  // End-of-buf / end-of-message.
  if (!m->is_group) {
    // This case doesn't exist for groups, because there eob really means
    // eob, so that case just exits the jit directly.
    |=>m->jit_endofbuf_pclabel:
    |  cmp  PTR, DECODER->delim_end
    |  jb   ->exit_jit    // We are at eob, but not end-of-submsg.
  }

  |=>m->jit_endofmsg_pclabel:
  // We are at end-of-submsg: call endmsg handler (if any):
  if (m->endmsg) {
    // void endmsg(void *closure, upb_status *status) {
    |  mov   ARG1_64, FRAME->closure
    |  lea   ARG2_64, DECODER->dispatcher.status
    |  callp m->endmsg
  }

  if (m->jit_parent_field_done_pclabel == UPB_MULTIPLE) {
    // Counter previous alignment.
    |  add  rsp, 8
    |  ret
  } else if (m->jit_parent_field_done_pclabel == UPB_TOPLEVEL_ONE) {
    |  jmp  ->exit_jit
  } else {
    |  jmp  =>m->jit_parent_field_done_pclabel
  }

}

static const char *dbgfmt =
    "JIT encountered unknown field!  wt=%d, fn=%d\n";

static void upb_decoder_jit(upb_decoder *d) {
  |  push  rbp
  |  mov   rbp, rsp
  |  push  r15
  |  push  r14
  |  push  r13
  |  push  r12
  |  push  rbx
  // Align stack.
  |  sub   rsp, 8
  |  mov   DECODER, ARG1_64
  |  mov   FRAME, DECODER:ARG1_64->dispatcher.top
  |  lea   STRREF, DECODER:ARG1_64->strref
  |  mov   CLOSURE, FRAME->closure
  |  mov   PTR, DECODER->ptr

  upb_handlers *h = d->dispatcher.handlers;
  if (h->msgs[0]->jit_parent_field_done_pclabel == UPB_MULTIPLE) {
    |  call  =>h->msgs[0]->jit_startmsg_pclabel
    |  jmp   ->exit_jit
  }

  // TODO: push return addresses for re-entry (will be necessary for multiple
  // buffer support).
  for (int i = 0; i < h->msgs_len; i++) upb_decoder_jit_msg(d, h->msgs[i]);

  |->exit_jit:
  |  mov   DECODER->ptr, PTR
  // Counter previous alignment.
  |  add   rsp, 8
  |  pop   rbx
  |  pop   r12
  |  pop   r13
  |  pop   r14
  |  pop   r15
  |  leave
  |  ret
  |=>0:
  |  mov rdi, stderr
  |  mov rsi, dbgfmt
  |  callp  fprintf
  |  callp  abort
}

void upb_decoder_jit_assignfieldlabs(upb_fhandlers *f,
                                     uint32_t *pclabel_count) {
  f->jit_pclabel = (*pclabel_count)++;
  f->jit_pclabel_notypecheck = (*pclabel_count)++;
  f->jit_submsg_done_pclabel = (*pclabel_count)++;
}

void upb_decoder_jit_assignmsglabs(upb_mhandlers *m, uint32_t *pclabel_count) {
  m->jit_startmsg_pclabel = (*pclabel_count)++;
  m->jit_endofbuf_pclabel = (*pclabel_count)++;
  m->jit_endofmsg_pclabel = (*pclabel_count)++;
  m->jit_dyndispatch_pclabel = (*pclabel_count)++;
  m->jit_unknownfield_pclabel = (*pclabel_count)++;
  m->jit_parent_field_done_pclabel = UPB_NONE;
  m->max_field_number = 0;
  upb_inttable_iter i;
  for(i = upb_inttable_begin(&m->fieldtab); !upb_inttable_done(i);
      i = upb_inttable_next(&m->fieldtab, i)) {
    uint32_t key = upb_inttable_iter_key(i);
    m->max_field_number = UPB_MAX(m->max_field_number, key);
    upb_fhandlers *f = upb_inttable_iter_value(i);
    upb_decoder_jit_assignfieldlabs(f, pclabel_count);
  }
  // XXX: Won't work for large field numbers; will need to use a upb_table.
  m->tablearray = malloc((m->max_field_number + 1) * sizeof(void*));
}

// Second pass: for messages that have only one parent, link them to the field
// from which they are called.
void upb_decoder_jit_assignmsglabs2(upb_mhandlers *m) {
  upb_inttable_iter i;
  for(i = upb_inttable_begin(&m->fieldtab); !upb_inttable_done(i);
      i = upb_inttable_next(&m->fieldtab, i)) {
    upb_fhandlers *f = upb_inttable_iter_value(i);
    if (upb_issubmsgtype(f->type)) {
      upb_mhandlers *sub_m = upb_fhandlers_getsubmsg(f);
      if (sub_m->jit_parent_field_done_pclabel == UPB_NONE) {
        sub_m->jit_parent_field_done_pclabel = f->jit_submsg_done_pclabel;
      } else {
        sub_m->jit_parent_field_done_pclabel = UPB_MULTIPLE;
      }
    }
  }
}

void upb_decoder_makejit(upb_decoder *d) {
  d->debug_info = NULL;

  // Assign pclabels.
  uint32_t pclabel_count = 1;
  upb_handlers *h = d->dispatcher.handlers;
  for (int i = 0; i < h->msgs_len; i++)
    upb_decoder_jit_assignmsglabs(h->msgs[i], &pclabel_count);
  for (int i = 0; i < h->msgs_len; i++)
    upb_decoder_jit_assignmsglabs2(h->msgs[i]);

  if (h->msgs[0]->jit_parent_field_done_pclabel == UPB_NONE) {
    h->msgs[0]->jit_parent_field_done_pclabel = UPB_TOPLEVEL_ONE;
  }

  void **globals = malloc(UPB_JIT_GLOBAL__MAX * sizeof(*globals));
  dasm_init(d, 1);
  dasm_setupglobal(d, globals, UPB_JIT_GLOBAL__MAX);
  dasm_growpc(d, pclabel_count);
  dasm_setup(d, upb_jit_actionlist);

  upb_decoder_jit(d);

  dasm_link(d, &d->jit_size);

  d->jit_code = mmap(NULL, d->jit_size, PROT_READ | PROT_WRITE,
                     MAP_32BIT | MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);

  upb_reg_jit_gdb(d);

  dasm_encode(d, d->jit_code);

  // Create dispatch tables.
  for (int i = 0; i < h->msgs_len; i++) {
    upb_mhandlers *m = h->msgs[i];
    for (uint32_t j = 0; j <= m->max_field_number; j++) {
      upb_fhandlers *f = NULL;
      for (int k = 0; k < 8; k++) {
        f = upb_inttable_lookup(&m->fieldtab, (j << 3) | k);
        if (f) break;
      }
      if (f) {
        m->tablearray[j] = d->jit_code + dasm_getpclabel(d, f->jit_pclabel);
      } else {
        // Don't handle unknown fields yet.
        m->tablearray[j] = d->jit_code + dasm_getpclabel(d, 0);
      }
    }
  }

  dasm_free(d);
  free(globals);

  mprotect(d->jit_code, d->jit_size, PROT_EXEC | PROT_READ);

  // View with: objdump -M intel -D -b binary -mi386 -Mx86-64 /tmp/machine-code
  // Or: ndisasm -b 64 /tmp/machine-code
  FILE *f = fopen("/tmp/machine-code", "wb");
  fwrite(d->jit_code, d->jit_size, 1, f);
  fclose(f);
}

void upb_decoder_freejit(upb_decoder *d) {
  munmap(d->jit_code, d->jit_size);
  free(d->debug_info);
  // TODO: unregister
}