/src/libpcap/gencode.c

Source
/*
 * Copyright (c) 1990, 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998
 *  The Regents of the University of California.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that: (1) source code distributions
 * retain the above copyright notice and this paragraph in its entirety, (2)
 * distributions including binary code include the above copyright notice and
 * this paragraph in its entirety in the documentation or other materials
 * provided with the distribution, and (3) all advertising materials mentioning
 * features or use of this software display the following acknowledgement:
 * ``This product includes software developed by the University of California,
 * Lawrence Berkeley Laboratory and its contributors.'' Neither the name of
 * the University nor the names of its contributors may be used to endorse
 * or promote products derived from this software without specific prior
 * written permission.
 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED
 * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
 */

#include <config.h>

#ifdef _WIN32
  #include <ws2tcpip.h>
#else
  #include <netinet/in.h>
#endif /* _WIN32 */

#include <stdlib.h>
#include <string.h>
#include <memory.h>
#include <setjmp.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>

#include "pcap-int.h"
#include "thread-local.h"

#include "extract.h"

#include "ethertype.h"
#include "llc.h"
#include "gencode.h"
#include "ieee80211.h"
#include "pflog.h"
#include "ppp.h"
#include "pcap/sll.h"
#include "pcap/ipnet.h"
#include "diag-control.h"
#include "no_sanitize.h"
#include "pcap-util.h"

#include "scanner.h"

#if defined(__linux__)
#include <linux/types.h>
#include <linux/if_packet.h>
#include <linux/filter.h>
#endif

#ifdef _WIN32
  #ifdef HAVE_NPCAP_BPF_H
    /* Defines BPF extensions for Npcap */
    #include <npcap-bpf.h>
  #endif
    #if defined(__MINGW32__) && defined(DEFINE_ADDITIONAL_IPV6_STUFF)
/* IPv6 address */
struct in6_addr
  {
    union
      {
  uint8_t   u6_addr8[16];
  uint16_t  u6_addr16[8];
  uint32_t  u6_addr32[4];
      } in6_u;
#define s6_addr     in6_u.u6_addr8
#define s6_addr16   in6_u.u6_addr16
#define s6_addr32   in6_u.u6_addr32
#define s6_addr64   in6_u.u6_addr64
  };

typedef unsigned short  sa_family_t;

#define __SOCKADDR_COMMON(sa_prefix) \
  sa_family_t sa_prefix##family

/* Ditto, for IPv6.  */
struct sockaddr_in6
  {
    __SOCKADDR_COMMON (sin6_);
    uint16_t sin6_port;   /* Transport layer port # */
    uint32_t sin6_flowinfo; /* IPv6 flow information */
    struct in6_addr sin6_addr;  /* IPv6 address */
  };

      #ifndef EAI_ADDRFAMILY
struct addrinfo {
  int ai_flags; /* AI_PASSIVE, AI_CANONNAME */
  int ai_family;  /* PF_xxx */
  int ai_socktype;  /* SOCK_xxx */
  int ai_protocol;  /* 0 or IPPROTO_xxx for IPv4 and IPv6 */
  size_t  ai_addrlen; /* length of ai_addr */
  char  *ai_canonname;  /* canonical name for hostname */
  struct sockaddr *ai_addr; /* binary address */
  struct addrinfo *ai_next; /* next structure in linked list */
};
      #endif /* EAI_ADDRFAMILY */
    #endif /* defined(__MINGW32__) && defined(DEFINE_ADDITIONAL_IPV6_STUFF) */
#else /* _WIN32 */
  #include <netdb.h>  /* for "struct addrinfo" */
#endif /* _WIN32 */
#include <pcap/namedb.h>

#include "nametoaddr.h"

#define ETHERMTU  1500

#ifndef IPPROTO_HOPOPTS
#define IPPROTO_HOPOPTS    0
#endif
#ifndef IPPROTO_IGMP
#define IPPROTO_IGMP       2
#endif
#ifndef IPPROTO_IGRP
#define IPPROTO_IGRP       9
#endif
#ifndef IPPROTO_ROUTING
#define IPPROTO_ROUTING   43
#endif
#ifndef IPPROTO_FRAGMENT
#define IPPROTO_FRAGMENT  44
#endif
#ifndef IPPROTO_ESP
#define IPPROTO_ESP       50
#endif
#ifndef IPPROTO_AH
#define IPPROTO_AH        51
#endif
#ifndef IPPROTO_ICMPV6
#define IPPROTO_ICMPV6    58
#endif
#ifndef IPPROTO_NONE
#define IPPROTO_NONE      59
#endif
#ifndef IPPROTO_DSTOPTS
#define IPPROTO_DSTOPTS   60
#endif
#ifndef IPPROTO_PIM
#define IPPROTO_PIM      103
#endif
#ifndef IPPROTO_CARP
#define IPPROTO_CARP     112
#endif
#ifndef IPPROTO_VRRP
#define IPPROTO_VRRP     112
#endif
#ifndef IPPROTO_SCTP
#define IPPROTO_SCTP     132
#endif

#define GENEVE_PORT 6081
#define VXLAN_PORT  4789


/*
 * from: NetBSD: if_arc.h,v 1.13 1999/11/19 20:41:19 thorpej Exp
 */

/* RFC 1051 */
#define ARCTYPE_IP_OLD    240 /* IP protocol */
#define ARCTYPE_ARP_OLD   241 /* address resolution protocol */

/* RFC 1201 */
#define ARCTYPE_IP    212  /* IP protocol */
#define ARCTYPE_ARP   213  /* address resolution protocol */
#define ARCTYPE_REVARP    214  /* reverse addr resolution protocol */

#define ARCTYPE_ATALK   221  /* Appletalk */
#define ARCTYPE_BANIAN    247 /* Banyan Vines */
#define ARCTYPE_IPX   250 /* Novell IPX */

#define ARCTYPE_INET6   0xc4  /* IPng */
#define ARCTYPE_DIAGNOSE  0x80  /* as per ANSI/ATA 878.1 */


/* Based on UNI3.1 standard by ATM Forum */

/* ATM traffic types based on VPI=0 and (the following VCI */
#define VCI_PPC     0x05  /* Point-to-point signal msg */
#define VCI_BCC     0x02  /* Broadcast signal msg */
#define VCI_OAMF4SC   0x03  /* Segment OAM F4 flow cell */
#define VCI_OAMF4EC   0x04  /* End-to-end OAM F4 flow cell */
#define VCI_METAC   0x01  /* Meta signal msg */
#define VCI_ILMIC   0x10  /* ILMI msg */

/* Q.2931 signalling messages */
#define CALL_PROCEED    0x02  /* call proceeding */
#define CONNECT     0x07  /* connect */
#define CONNECT_ACK   0x0f  /* connect_ack */
#define SETUP     0x05  /* setup */
#define RELEASE     0x4d  /* release */
#define RELEASE_DONE    0x5a  /* release_done */
#define RESTART     0x46  /* restart */
#define RESTART_ACK   0x4e  /* restart ack */
#define STATUS      0x7d  /* status */
#define STATUS_ENQ    0x75  /* status ack */
#define ADD_PARTY   0x80  /* add party */
#define ADD_PARTY_ACK   0x81  /* add party ack */
#define ADD_PARTY_REJ   0x82  /* add party rej */
#define DROP_PARTY    0x83  /* drop party */
#define DROP_PARTY_ACK    0x84  /* drop party ack */

/* Information Element Parameters in the signalling messages */
#define CAUSE     0x08  /* cause */
#define ENDPT_REF   0x54  /* endpoint reference */
#define AAL_PARA    0x58  /* ATM adaptation layer parameters */
#define TRAFF_DESCRIP   0x59  /* atm traffic descriptors */
#define CONNECT_ID    0x5a  /* connection identifier */
#define QOS_PARA    0x5c  /* quality of service parameters */
#define B_HIGHER    0x5d  /* broadband higher layer information */
#define B_BEARER    0x5e  /* broadband bearer capability */
#define B_LOWER     0x5f  /* broadband lower information */
#define CALLING_PARTY   0x6c  /* calling party number */
#define CALLED_PARTY    0x70  /* called party number */

#define Q2931     0x09

/* Q.2931 signalling general messages format */
#define PROTO_POS       0  /* offset of protocol discriminator */
#define CALL_REF_POS    2 /* offset of call reference value */
#define MSG_TYPE_POS    5  /* offset of message type */
#define MSG_LEN_POS     7 /* offset of message length */
#define IE_BEGIN_POS    9 /* offset of first information element */

/* format of signalling messages */
#define TYPE_POS  0
#define LEN_POS   2
#define FIELD_BEGIN_POS 4


/* SunATM header for ATM packet */
#define SUNATM_DIR_POS    0
#define SUNATM_VPI_POS    1
#define SUNATM_VCI_POS    2
#define SUNATM_PKT_BEGIN_POS  4  /* Start of ATM packet */

/* Protocol type values in the bottom for bits of the byte at SUNATM_DIR_POS. */
#define PT_LANE   0x01  /* LANE */
#define PT_LLC    0x02  /* LLC encapsulation */
#define PT_ILMI   0x05  /* ILMI */
#define PT_QSAAL  0x06  /* Q.SAAL */


/* Types missing from some systems */

/*
 * Network layer protocol identifiers
 * ITU-T Rec. X.263 (1998 E)
 * ISO/IEC TR 9577:1999(E)
 */
#ifndef ISO8473_CLNP
#define ISO8473_CLNP    0x81
#endif
#ifndef ISO9542_ESIS
#define ISO9542_ESIS    0x82
#endif
#ifndef ISO10589_ISIS
#define ISO10589_ISIS   0x83
#endif
#ifndef ISO9577_IPV6
#define ISO9577_IPV6    0x8e
#endif
#ifndef ISO9577_IPV4
#define ISO9577_IPV4    0xcc
#endif

#define ISIS_L1_LAN_IIH      15
#define ISIS_L2_LAN_IIH      16
#define ISIS_PTP_IIH         17
#define ISIS_L1_LSP          18
#define ISIS_L2_LSP          20
#define ISIS_L1_CSNP         24
#define ISIS_L2_CSNP         25
#define ISIS_L1_PSNP         26
#define ISIS_L2_PSNP         27
/*
 * The maximum possible value can also be used as a bit mask because the
 * "PDU Type" field comprises the least significant 5 bits of a particular
 * octet, see sections 9.5~9.13 of ISO/IEC 10589:2002(E).
 */
#define ISIS_PDU_TYPE_MAX 0x1FU

// Same as in tcpdump/print-sl.c.
#define SLIPDIR_IN 0
#define SLIPDIR_OUT 1

/*
 * Offsets of various fields from the beginning of their network-layer
 * header, which is the link-layer payload (OR_LINKPL).
 */
#define IPV6_PROTO_OFFSET    6
#define IPV6_SRCADDR_OFFSET  8
#define IPV6_DSTADDR_OFFSET 24
#define IPV4_PROTO_OFFSET    9
#define IPV4_SRCADDR_OFFSET 12
#define IPV4_DSTADDR_OFFSET 16
#define ARP_SRCADDR_OFFSET  14
#define ARP_DSTADDR_OFFSET  24
#define RARP_SRCADDR_OFFSET 14
#define RARP_DSTADDR_OFFSET 24

/*
 * Offsets of supported (TCP, UDP and SCTP) ports from the beginning of their
 * header, which is the network-layer payload (OR_TRAN_IPV4 and OR_TRAN_IPV6).
 */
#define TRAN_SRCPORT_OFFSET 0
#define TRAN_DSTPORT_OFFSET 2

// IPv6 mandatory outer header (Version, ..., Destination Address) length.
#define IP6_HDRLEN 40

// RFC 3032 Section 2.1, the "Label Stack Entry" 32-bit structure.
#define MPLS_STACKENTRY_LEN 4
// Ibid., the "Label" 20-bit field.
#define MPLS_LABEL_MAX 0xfffffU
#define MPLS_LABEL_SHIFT 12

#ifdef HAVE_OS_PROTO_H
#include "os-proto.h"
#endif

/*
 * Valid BPF opcodes comprise one (e.g. BPF_ST), two (e.g. BPF_RET|BPF_K) or
 * three (e.g. BPF_ALU|BPF_ADD|BPF_X) constants.  Even constants that are only
 * nominally specified because they are defined to zero (e.g. BPF_K) have
 * meaning in the source code and the ISA.  Instead of ORing the constants in
 * place use the following two macros: so long as the code uses a correct
 * macro, not specifying the correct number of constants will be a compile-time
 * error.  Groups of related BPF opcodes can wrap these with their own macros
 * to factor common parts out.
 *
 * Each of the statement-specific NEW_STMT_xxx() macros wraps either new_stmt()
 * (if the statement does not use the 'k' member) or new_stmt_k() (if it does).
 * This likewise makes an erroneous absence or presence of 'k' a compile-time
 * error.
 */
#define OPCODE_2ARY(x1, x2) ((x1) | (x2))
#define OPCODE_3ARY(x1, x2, x3) ((x1) | (x2) | (x3))

/*
 * Load instruction opcodes comprise three constants.
 */
#define OPCODE_LOAD(size, mode) OPCODE_3ARY(BPF_LD, (size), (mode))
#define NEW_STMT_LD_LEN(cstate) \
        new_stmt((cstate), OPCODE_LOAD(BPF_W, BPF_LEN));
#define NEW_STMT_LD_IMM(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_W, BPF_IMM), (k))
#define NEW_STMT_LD_M(cstate, regno) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_W, BPF_MEM), guard_reg((cstate), (regno)))
#define NEW_STMT_LDB_ABS(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_B, BPF_ABS), (k))
#define NEW_STMT_LDH_ABS(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_H, BPF_ABS), (k))
#define NEW_STMT_LD_ABS(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_W, BPF_ABS), (k))
#define NEW_STMT_LDB_IND(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_B, BPF_IND), (k))
#define NEW_STMT_LDH_IND(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOAD(BPF_H, BPF_IND), (k))

#define OPCODE_LOADX(size, mode) OPCODE_3ARY(BPF_LDX, (size), (mode))
#define NEW_STMT_LDX_IMM(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOADX(BPF_W, BPF_IMM), (k))
#define NEW_STMT_LDX_M(cstate, regno) \
        new_stmt_k((cstate), OPCODE_LOADX(BPF_W, BPF_MEM), guard_reg((cstate), (regno)))
#define NEW_STMT_LDXB_MSH(cstate, k) \
        new_stmt_k((cstate), OPCODE_LOADX(BPF_B, BPF_MSH), (k))

/*
 * Store instruction opcodes are single constants.
 */
#define NEW_STMT_ST_M(cstate, regno) \
        new_stmt_k((cstate), BPF_ST, guard_reg((cstate), (regno)))
#define NEW_STMT_STX_M(cstate, regno) \
        new_stmt_k((cstate), BPF_STX, guard_reg((cstate), (regno)))

/*
 * Branch instruction opcodes comprise two (for "ja L") or three (for
 * everything else) constants.
 */
#define NEW_STMT_JA(cstate, k) \
        new_stmt_k((cstate), OPCODE_2ARY(BPF_JMP, BPF_JA), (k))

#define OPCODE_BRANCH_X(jtype) OPCODE_3ARY(BPF_JMP, (jtype), BPF_X)

#define OPCODE_BRANCH_K(jtype) OPCODE_3ARY(BPF_JMP, (jtype), BPF_K)
#define NEW_STMT_JEQ_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_BRANCH_K(BPF_JEQ), (k))
#define NEW_STMT_JSET_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_BRANCH_K(BPF_JSET), (k))

/*
 * ALU instruction opcodes comprise two (for "neg") or three (for everything
 * else) constants.
 */
#define NEW_STMT_NEG(cstate) \
        new_stmt((cstate), OPCODE_2ARY(BPF_ALU, BPF_NEG))

#define OPCODE_ALU_X(op) OPCODE_3ARY(BPF_ALU, (op), BPF_X)
#define NEW_STMT_ADD_X(cstate) \
        new_stmt((cstate), OPCODE_ALU_X(BPF_ADD))
#define NEW_STMT_OR_X(cstate) \
        new_stmt((cstate), OPCODE_ALU_X(BPF_OR))

#define OPCODE_ALU_K(op) OPCODE_3ARY(BPF_ALU, (op), BPF_K)
#define NEW_STMT_ADD_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_ALU_K(BPF_ADD), (k))
#define NEW_STMT_MUL_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_ALU_K(BPF_MUL), (k))
#define NEW_STMT_AND_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_ALU_K(BPF_AND), (k))
#define NEW_STMT_OR_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_ALU_K(BPF_OR), (k))
#define NEW_STMT_LSH_K(cstate, k) \
        new_stmt_k((cstate), OPCODE_ALU_K(BPF_LSH), (k))

/*
 * Return instruction opcodes always comprise two constants.
 */
#define OPCODE_RET(rval) OPCODE_2ARY(BPF_RET, (rval))

/*
 * Miscellaneous instruction opcodes always comprise two constants.
 */
#define OPCODE_MISC(op) OPCODE_2ARY(BPF_MISC, (op))
#define NEW_STMT_TAX(cstate) \
        new_stmt((cstate), OPCODE_MISC(BPF_TAX))
#define NEW_STMT_TXA(cstate) \
        new_stmt((cstate), OPCODE_MISC(BPF_TXA))

/*
 * "Push" the current value of the link-layer header type and link-layer
 * header offset onto a "stack", and set a new value.  (It's not a
 * full-blown stack; we keep only the top two items.)
 */
#define PUSH_LINKHDR(cs, new_linktype, new_is_variable, new_constant_part, new_reg) \
{ \
  (cs)->prevlinktype = (cs)->linktype; \
  (cs)->off_prevlinkhdr = (cs)->off_linkhdr; \
  (cs)->linktype = (new_linktype); \
  (cs)->off_linkhdr.is_variable = (new_is_variable); \
  (cs)->off_linkhdr.constant_part = (new_constant_part); \
  (cs)->off_linkhdr.reg = (new_reg); \
  (cs)->is_encap = 0; \
}

/*
 * Offset "not set" value.
 */
#define OFFSET_NOT_SET  0xffffffffU

/*
 * Absolute offsets, which are offsets from the beginning of the raw
 * packet data, are, in the general case, the sum of a variable value
 * and a constant value; the variable value may be absent, in which
 * case the offset is only the constant value, and the constant value
 * may be zero, in which case the offset is only the variable value.
 *
 * bpf_abs_offset is a structure containing all that information:
 *
 *   is_variable is 1 if there's a variable part.
 *
 *   constant_part is the constant part of the value, possibly zero;
 *
 *   if is_variable is 1, reg is the register number for a register
 *   containing the variable value if the register has been assigned,
 *   and -1 otherwise.
 */
typedef struct {
  int is_variable;
  u_int constant_part;
  int reg;
} bpf_abs_offset;

/*
 * Value passed to gen_load_a() to indicate what the offset argument
 * is relative to the beginning of.
 */
enum e_offrel {
  OR_PACKET,    /* full packet data */
  OR_LINKHDR,   /* link-layer header */
  OR_PREVLINKHDR,   /* previous link-layer header */
  OR_LLC,     /* 802.2 LLC header */
  OR_PREVMPLSHDR,   /* previous MPLS header */
  OR_LINKTYPE,    /* link-layer type */
  OR_LINKPL,    /* link-layer payload */
  OR_LINKPL_NOSNAP, /* link-layer payload, with no SNAP header at the link layer */
  OR_TRAN_IPV4,   /* transport-layer header, with IPv4 network layer */
  OR_TRAN_IPV6    /* transport-layer header, with IPv6 network layer */
};

/*
 * Divvy out chunks of memory rather than call calloc() each time: this way
 * pcap_compile() induces orders of magnitude fewer calloc() calls, which
 * eventually require orders of magnitude fewer free() calls, which makes it
 * much easier to prevent memory leaks, which is important in a library.
 *
 * The total amount of memory that can be allocated using 16 chunks, where
 * chunk 0 size is 1KiB and each next chunk is double the size of the previous,
 * is (64MiB - 1KiB).
 */
#define NCHUNKS 16
#define CHUNKSIZE(idx) (1024U << (idx))
struct chunk {
  size_t n_left;
  void *m;
};

/*
 * A chunk can store any of:
 *  - a string (guaranteed alignment 1 but present for completeness)
 *  - a block
 *  - an slist
 *  - an arth
 * For this simple allocator every allocated chunk gets rounded up to the
 * alignment needed for any chunk.
 */
struct chunk_align {
  char dummy;
  union {
    char c;
    struct block b;
    struct slist s;
    struct arth a;
  } u;
};
#define CHUNK_ALIGN (offsetof(struct chunk_align, u))

/* Code generator state */

struct _compiler_state {
  jmp_buf top_ctx;
  pcap_t *bpf_pcap;
  int error_set;

  struct icode ic;

  int snaplen;

  int linktype;
  int prevlinktype;
  int outermostlinktype;

  bpf_u_int32 netmask;
  int no_optimize;

  /* Hack for handling VLAN and MPLS stacks. */
  u_int label_stack_depth;
  u_int vlan_stack_depth;

  /* XXX */
  u_int pcap_fddipad;

  /*
   * As errors are handled by a longjmp, anything allocated must
   * be freed in the longjmp handler, so it must be reachable
   * from that handler.
   *
   * One thing that's allocated is the result of pcap_nametoaddrinfo();
   * it must be freed with freeaddrinfo().  This variable points to
   * any addrinfo structure that would need to be freed.
   */
  struct addrinfo *ai;

  /*
   * Various code constructs need to know the layout of the packet.
   * These values give the necessary offsets from the beginning
   * of the packet data.
   */

  /*
   * Absolute offset of the beginning of the link-layer header.
   */
  bpf_abs_offset off_linkhdr;

  /*
   * If we're checking a link-layer header for a packet encapsulated
   * in another protocol layer, this is the equivalent information
   * for the previous layers' link-layer header from the beginning
   * of the raw packet data.
   */
  bpf_abs_offset off_prevlinkhdr;

  /*
   * This is the equivalent information for the outermost layers'
   * link-layer header.
   */
  bpf_abs_offset off_outermostlinkhdr;

  /*
   * Absolute offset of the beginning of the link-layer payload.
   */
  bpf_abs_offset off_linkpl;

  /*
   * "off_linktype" is the offset to information in the link-layer
   * header giving the packet type. This is an absolute offset
   * from the beginning of the packet.
   *
   * For Ethernet, it's the offset of the Ethernet type field; this
   * means that it must have a value that skips VLAN tags.
   *
   * For link-layer types that always use 802.2 headers, it's the
   * offset of the LLC header; this means that it must have a value
   * that skips VLAN tags.
   *
   * For PPP, it's the offset of the PPP type field.
   *
   * For Cisco HDLC, it's the offset of the CHDLC type field.
   *
   * For BSD loopback, it's the offset of the AF_ value.
   *
   * For Linux cooked sockets, it's the offset of the type field.
   *
   * off_linktype.constant_part is set to OFFSET_NOT_SET for no
   * encapsulation, in which case, IP is assumed.
   */
  bpf_abs_offset off_linktype;

  /*
   * TRUE if the link layer includes an ATM pseudo-header.
   */
  int is_atm;

  /* TRUE if "geneve" or "vxlan" appeared in the filter; it
   * causes us to generate code that checks for a Geneve or
   * VXLAN header respectively and assume that later filters
   * apply to the encapsulated payload.
   */
  int is_encap;

  /*
   * TRUE if we need variable length part of VLAN offset
   */
  int is_vlan_vloffset;

  /*
   * These are offsets for the ATM pseudo-header.
   */
  u_int off_vpi;
  u_int off_vci;
  u_int off_proto;

  /*
   * These are offsets for the MTP2 fields.
   */
  u_int off_li;
  u_int off_li_hsl;

  /*
   * These are offsets for the MTP3 fields.
   */
  u_int off_sio;
  u_int off_opc;
  u_int off_dpc;
  u_int off_sls;

  /*
   * This is the offset of the first byte after the ATM pseudo_header,
   * or -1 if there is no ATM pseudo-header.
   */
  u_int off_payload;

  /*
   * These are offsets to the beginning of the network-layer header.
   * They are relative to the beginning of the link-layer payload
   * (i.e., they don't include off_linkhdr.constant_part or
   * off_linkpl.constant_part).
   *
   * If the link layer never uses 802.2 LLC:
   *
   *  "off_nl" and "off_nl_nosnap" are the same.
   *
   * If the link layer always uses 802.2 LLC:
   *
   *  "off_nl" is the offset if there's a SNAP header following
   *  the 802.2 header;
   *
   *  "off_nl_nosnap" is the offset if there's no SNAP header.
   *
   * If the link layer is Ethernet:
   *
   *  "off_nl" is the offset if the packet is an Ethernet II packet
   *  (we assume no 802.3+802.2+SNAP);
   *
   *  "off_nl_nosnap" is the offset if the packet is an 802.3 packet
   *  with an 802.2 header following it.
   */
  u_int off_nl;
  u_int off_nl_nosnap;

  /*
   * Here we handle simple allocation of the scratch registers.
   * If too many registers are alloc'd, the allocator punts.
   */
  int regused[BPF_MEMWORDS];
  int curreg;

  /*
   * Memory chunks.
   */
  struct chunk chunks[NCHUNKS];
  unsigned cur_chunk;
};

/*
 * For use by routines outside this file.
 */
/* VARARGS */
void
bpf_set_error(compiler_state_t *cstate, const char *fmt, ...)
{
  va_list ap;

  /*
   * If we've already set an error, don't override it.
   * The lexical analyzer reports some errors by setting
   * the error and then returning a LEX_ERROR token, which
   * is not recognized by any grammar rule, and thus forces
   * the parse to stop.  We don't want the error reported
   * by the lexical analyzer to be overwritten by the syntax
   * error.
   */
  if (!cstate->error_set) {
    va_start(ap, fmt);
    (void)vsnprintf(cstate->bpf_pcap->errbuf, PCAP_ERRBUF_SIZE,
        fmt, ap);
    va_end(ap);
    cstate->error_set = 1;
  }
}

/*
 * For use *ONLY* in routines in this file.
 */
static void PCAP_NORETURN bpf_error(compiler_state_t *, const char *, ...)
    PCAP_PRINTFLIKE(2, 3);

/* VARARGS */
static void PCAP_NORETURN
bpf_error(compiler_state_t *cstate, const char *fmt, ...)
{
  va_list ap;

  va_start(ap, fmt);
  (void)vsnprintf(cstate->bpf_pcap->errbuf, PCAP_ERRBUF_SIZE,
      fmt, ap);
  va_end(ap);
  longjmp(cstate->top_ctx, 1);
  /*NOTREACHED*/
#ifdef _AIX
  PCAP_UNREACHABLE
#endif /* _AIX */
}

static int init_linktype(compiler_state_t *, pcap_t *);

static void init_regs(compiler_state_t *);
static int alloc_reg(compiler_state_t *);
static int guard_reg(compiler_state_t *, const int);
static void free_reg(compiler_state_t *, int);
static void free_reg_arth(compiler_state_t *, struct arth *);

static bool initchunks_ok(compiler_state_t *cstate);
static void *newchunk_nolongjmp(compiler_state_t *cstate, size_t);
static void *newchunk(compiler_state_t *cstate, size_t);
static void freechunks(compiler_state_t *cstate);
static inline struct block *new_block(compiler_state_t *cstate, int);
static inline struct slist *new_stmt(compiler_state_t *cstate, int);
static struct block *sprepend_to_block(struct slist *, struct block *);
static struct block *gen_retblk(compiler_state_t *cstate, int);
static inline void syntax(compiler_state_t *cstate);

static void backpatch(struct block *, struct block *);
static void merge(struct block *, struct block *);
static struct block *gen_cmp(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32);
static struct block *gen_cmp_gt(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32);
static struct block *gen_cmp_ge(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32);
static struct block *gen_cmp_lt(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32);
static struct block *gen_cmp_le(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32);
static struct block *gen_cmp_ne(compiler_state_t *, enum e_offrel, u_int,
    u_int size, bpf_u_int32);
static struct block *gen_mcmp(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32, bpf_u_int32);
static struct block *gen_mcmp_ne(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32, bpf_u_int32);
static struct block *gen_bcmp(compiler_state_t *, enum e_offrel, u_int,
    u_int, const u_char *);
static struct block *gen_jmp_k(compiler_state_t *, const int,
    const bpf_u_int32, struct slist *);
static struct block *gen_jmp_x(compiler_state_t *, const int, struct slist *);
static struct block *gen_set(compiler_state_t *, bpf_u_int32, struct slist *);
static struct block *gen_unset(compiler_state_t *, bpf_u_int32, struct slist *);
static struct block *gen_ncmp(compiler_state_t *, enum e_offrel, u_int,
    u_int, bpf_u_int32, int, int, bpf_u_int32);
static struct slist *gen_load_absoffsetrel(compiler_state_t *, struct slist *,
    const u_int, const u_int);
static struct slist *gen_load_absoffsetarthrel(compiler_state_t *,
    struct slist *, const bpf_u_int32, const struct arth *, const u_int);
static struct slist *gen_load_a(compiler_state_t *, const enum e_offrel, u_int,
    const u_int);
static struct slist *gen_loadx_iphdrlen(compiler_state_t *);
static struct block *gen_uncond(compiler_state_t *, const u_char);
static inline struct block *gen_true(compiler_state_t *);
static inline struct block *gen_false(compiler_state_t *);
static struct block *gen_ether_linktype(compiler_state_t *, bpf_u_int32);
static struct block *gen_ipnet_linktype(compiler_state_t *, bpf_u_int32);
static struct block *gen_linux_sll_linktype(compiler_state_t *, bpf_u_int32);
static struct slist *gen_load_pflog_llprefixlen(compiler_state_t *);
static struct slist *gen_load_prism_llprefixlen(compiler_state_t *);
static struct slist *gen_load_avs_llprefixlen(compiler_state_t *);
static struct slist *gen_load_radiotap_llprefixlen(compiler_state_t *);
static struct slist *gen_load_ppi_llprefixlen(compiler_state_t *);
static void insert_compute_vloffsets(compiler_state_t *, struct block *);
static struct slist *gen_abs_offset_varpart(compiler_state_t *,
    bpf_abs_offset *);
static uint16_t ethertype_to_ppptype(compiler_state_t *, bpf_u_int32);
static struct block *gen_linktype(compiler_state_t *, bpf_u_int32);
static struct block *gen_snap(compiler_state_t *, bpf_u_int32, bpf_u_int32);
static struct block *gen_llc_linktype(compiler_state_t *, bpf_u_int32);
static struct block *gen_hostop(compiler_state_t *, bpf_u_int32, bpf_u_int32,
    int, u_int, u_int);
static struct block *gen_hostop6(compiler_state_t *, const struct in6_addr *,
    const struct in6_addr *, const u_char);
static struct block *gen_wlanhostop(compiler_state_t *, const u_char *, int);
static unsigned char is_mac48_linktype(const int);
static struct block *gen_mac48host(compiler_state_t *, const u_char *,
    const u_char, const char *);
static struct block *gen_mac48host_byname(compiler_state_t *, const char *,
    const u_char, const char *);
static struct block *gen_mac8host(compiler_state_t *, const uint8_t,
    const u_char, const char *);
static struct block *gen_dnhostop(compiler_state_t *, bpf_u_int32, int);
static struct block *gen_mpls_linktype(compiler_state_t *, bpf_u_int32);
static struct block *gen_host(compiler_state_t *, const size_t,
    const bpf_u_int32 *, const bpf_u_int32 *, const u_char, const u_char,
    const u_char, const char *);
static struct block *gen_host6(compiler_state_t *, const size_t,
    const struct in6_addr *, const struct in6_addr *, const u_char,
    const u_char, const u_char, const char *);
static struct block *gen_host46_byname(compiler_state_t *, const char *,
    const u_char, const u_char, const u_char, const u_char);
static struct block *gen_dnhost(compiler_state_t *, const char *, bpf_u_int32,
    const struct qual);
static struct block *gen_gateway(compiler_state_t *, const char *, const u_char);
static struct block *gen_ip_proto(compiler_state_t *, const uint8_t);
static struct block *gen_ip6_proto(compiler_state_t *, const uint8_t);
static struct block *gen_ipfrag(compiler_state_t *);
static struct block *gen_portatom(compiler_state_t *, int, uint16_t);
static struct block *gen_portrangeatom(compiler_state_t *, u_int, uint16_t,
    uint16_t);
static struct block *gen_portatom6(compiler_state_t *, int, uint16_t);
static struct block *gen_portrangeatom6(compiler_state_t *, u_int, uint16_t,
    uint16_t);
static struct block *gen_port(compiler_state_t *, const uint16_t, const int,
    const u_char, const u_char);
static struct block *gen_port_common(compiler_state_t *, int, struct block *);
static struct block *gen_portrange(compiler_state_t *, uint16_t, uint16_t,
    int, int);
static struct block *gen_port6(compiler_state_t *, const uint16_t, const int,
    const u_char, const u_char);
static struct block *gen_port6_common(compiler_state_t *, int, struct block *);
static struct block *gen_portrange6(compiler_state_t *, uint16_t, uint16_t,
    int, int);
static int lookup_proto(compiler_state_t *, const char *, const struct qual);
#if !defined(NO_PROTOCHAIN)
static struct block *gen_protochain(compiler_state_t *, bpf_u_int32, int);
#endif /* !defined(NO_PROTOCHAIN) */
static struct block *gen_proto(compiler_state_t *, bpf_u_int32, int);
static struct slist *xfer_to_x(compiler_state_t *, const struct arth *);
static struct slist *xfer_to_a(compiler_state_t *, const struct arth *);
static struct block *gen_mac_multicast(compiler_state_t *, int);
static struct block *gen_len(compiler_state_t *, int, int);
static struct block *gen_encap_ll_check(compiler_state_t *cstate);

static struct block *gen_atmfield_code_internal(compiler_state_t *, int,
    bpf_u_int32, int, int);
static struct block *gen_atmtype_llc(compiler_state_t *);
static struct block *gen_msg_abbrev(compiler_state_t *, const uint8_t);
static struct block *gen_atm_prototype(compiler_state_t *, const uint8_t);
static struct block *gen_atm_vpi(compiler_state_t *, const uint8_t);
static struct block *gen_atm_vci(compiler_state_t *, const uint16_t);

#define ERRSTR_FUNC_VAR_INT "internal error in %s(): %s == %d"

static bool
initcurrentchunk_ok(compiler_state_t *cstate)
{
  if (cstate->cur_chunk >= NCHUNKS) {
    bpf_set_error(cstate, ERRSTR_FUNC_VAR_INT, __func__,
        "cur_chunk", cstate->cur_chunk);
    return false;
  }
  const size_t size = CHUNKSIZE(cstate->cur_chunk);
  cstate->chunks[cstate->cur_chunk].m = calloc(1, size);
  if (cstate->chunks[cstate->cur_chunk].m == NULL) {
    bpf_set_error(cstate, "%s: calloc() failed", __func__);
    return false;
  }
  cstate->chunks[cstate->cur_chunk].n_left = size;
  return true;
}

static bool
initchunks_ok(compiler_state_t *cstate)
{
  int i;

  for (i = 0; i < NCHUNKS; i++) {
    cstate->chunks[i].n_left = 0;
    cstate->chunks[i].m = NULL;
  }
  cstate->cur_chunk = 0;
  return initcurrentchunk_ok(cstate);
}

static void *
newchunk_nolongjmp(compiler_state_t *cstate, size_t n)
{
  /* Round up to chunk alignment. */
  n = (n + CHUNK_ALIGN - 1) & ~(CHUNK_ALIGN - 1);

  if (n > cstate->chunks[cstate->cur_chunk].n_left) {
    if (cstate->cur_chunk >= NCHUNKS - 1) {
      bpf_set_error(cstate,
          "will not allocate more than %u chunks", NCHUNKS);
      return (NULL);
    }
    if (n > CHUNKSIZE(cstate->cur_chunk + 1)) {
      bpf_set_error(cstate,
          "%zu bytes would not fit into chunk %u",
          n, cstate->cur_chunk + 1);
      return (NULL);
    }
    ++cstate->cur_chunk;
    if (! initcurrentchunk_ok(cstate))
      return (NULL); // The error buffer has been filled.
  }
  cstate->chunks[cstate->cur_chunk].n_left -= n;
  return (void *)((char *)cstate->chunks[cstate->cur_chunk].m +
      cstate->chunks[cstate->cur_chunk].n_left);
}

static void *
newchunk(compiler_state_t *cstate, size_t n)
{
  void *p;

  p = newchunk_nolongjmp(cstate, n);
  if (p == NULL) {
    longjmp(cstate->top_ctx, 1);
    /*NOTREACHED*/
  }
  return (p);
}

static void
freechunks(compiler_state_t *cstate)
{
  int i;

  for (i = 0; i < NCHUNKS; ++i)
    if (cstate->chunks[i].m != NULL)
      free(cstate->chunks[i].m);
}

/*
 * A strdup whose allocations are freed after code generation is over.
 * This is used by the lexical analyzer, so it can't longjmp; it just
 * returns NULL on an allocation error, and the callers must check
 * for it.
 */
char *
sdup(compiler_state_t *cstate, const char *s)
{
  size_t n = strlen(s) + 1;
  char *cp = newchunk_nolongjmp(cstate, n);

  if (cp == NULL)
    return (NULL);
  pcapint_strlcpy(cp, s, n);
  return (cp);
}

static inline struct block *
new_block(compiler_state_t *cstate, int code)
{
  struct block *p;

  p = (struct block *)newchunk(cstate, sizeof(*p));
  p->s.code = code;
  p->head = p;

  return p;
}

static inline struct slist *
new_stmt(compiler_state_t *cstate, int code)
{
  struct slist *p;

  p = (struct slist *)newchunk(cstate, sizeof(*p));
  p->s.code = code;

  return p;
}

static inline struct slist *
new_stmt_k(compiler_state_t *cstate, const int code, const uint32_t k)
{
  struct slist *ret = new_stmt(cstate, code);
  ret->s.k = k;
  return ret;
}

static struct block *
gen_retblk_internal(compiler_state_t *cstate, int v)
{
  struct block *b = new_block(cstate, OPCODE_RET(BPF_K));

  b->s.k = v;
  return b;
}

static struct block *
gen_retblk(compiler_state_t *cstate, int v)
{
  if (setjmp(cstate->top_ctx)) {
    /*
     * gen_retblk() only fails because a memory
     * allocation failed in newchunk(), meaning
     * that it can't return a pointer.
     *
     * Return NULL.
     */
    return NULL;
  }
  return gen_retblk_internal(cstate, v);
}

static inline PCAP_NORETURN_DEF void
syntax(compiler_state_t *cstate)
{
  bpf_error(cstate, "syntax error in filter expression");
}

/*
 * For the given integer return a string with the keyword (or the nominal
 * keyword if there is more than one).  This is a simpler version of tok2str()
 * in tcpdump because in this problem space a valid integer value is not
 * greater than 71.
 */
static const char *
qual2kw(const char *kind, const unsigned id, const char *tokens[],
    const size_t size)
{
  static thread_local char buf[4][64];
  static thread_local int idx = 0;

  if (id < size && tokens[id])
    return tokens[id];

  char *ret = buf[idx];
  idx = (idx + 1) % (sizeof(buf) / sizeof(buf[0]));
  ret[0] = '\0'; // just in case
  snprintf(ret, sizeof(buf[0]), "<invalid %s %u>", kind, id);
  return ret;
}

// protocol qualifier keywords
static const char *
pqkw(const unsigned id)
{
  const char * tokens[] = {
    [Q_LINK] = "link",
    [Q_IP] = "ip",
    [Q_ARP] = "arp",
    [Q_RARP] = "rarp",
    [Q_SCTP] = "sctp",
    [Q_TCP] = "tcp",
    [Q_UDP] = "udp",
    [Q_ICMP] = "icmp",
    [Q_IGMP] = "igmp",
    [Q_IGRP] = "igrp",
    [Q_ATALK] = "atalk",
    [Q_DECNET] = "decnet",
    [Q_LAT] = "lat",
    [Q_SCA] = "sca",
    [Q_MOPRC] = "moprc",
    [Q_MOPDL] = "mopdl",
    [Q_IPV6] = "ip6",
    [Q_ICMPV6] = "icmp6",
    [Q_AH] = "ah",
    [Q_ESP] = "esp",
    [Q_PIM] = "pim",
    [Q_VRRP] = "vrrp",
    [Q_AARP] = "aarp",
    [Q_ISO] = "iso",
    [Q_ESIS] = "esis",
    [Q_ISIS] = "isis",
    [Q_CLNP] = "clnp",
    [Q_STP] = "stp",
    [Q_IPX] = "ipx",
    [Q_NETBEUI] = "netbeui",
    [Q_ISIS_L1] = "l1",
    [Q_ISIS_L2] = "l2",
    [Q_ISIS_IIH] = "iih",
    [Q_ISIS_SNP] = "snp",
    [Q_ISIS_CSNP] = "csnp",
    [Q_ISIS_PSNP] = "psnp",
    [Q_ISIS_LSP] = "lsp",
    [Q_RADIO] = "radio",
    [Q_CARP] = "carp",
  };
  return qual2kw("proto", id, tokens, sizeof(tokens) / sizeof(tokens[0]));
}

// direction qualifier keywords
static const char *
dqkw(const unsigned id)
{
  const char * tokens[] = {
    [Q_SRC] = "src",
    [Q_DST] = "dst",
    [Q_OR] = "src or dst",
    [Q_AND] = "src and dst",
    [Q_ADDR1] = "addr1",
    [Q_ADDR2] = "addr2",
    [Q_ADDR3] = "addr3",
    [Q_ADDR4] = "addr4",
    [Q_RA] = "ra",
    [Q_TA] = "ta",
  };
  return qual2kw("dir", id, tokens, sizeof(tokens) / sizeof(tokens[0]));
}

// type (in the man page) / address (in the code) qualifier keywords
static const char *
tqkw(const unsigned id)
{
  const char * tokens[] = {
    [Q_HOST] = "host",
    [Q_NET] = "net",
    [Q_PORT] = "port",
    [Q_GATEWAY] = "gateway",
    [Q_PROTO] = "proto",
    [Q_PROTOCHAIN] = "protochain",
    [Q_PORTRANGE] = "portrange",
  };
  return qual2kw("type", id, tokens, sizeof(tokens) / sizeof(tokens[0]));
}

// ATM keywords
static const char *
atmkw(const unsigned id)
{
  const char * tokens[] = {
    [A_METAC] = "metac",
    [A_BCC] = "bcc",
    [A_OAMF4SC] = "oamf4sc",
    [A_OAMF4EC] = "oamf4ec",
    [A_SC] = "sc",
    [A_ILMIC] = "ilmic",
    [A_OAM] = "oam",
    [A_OAMF4] = "oamf4",
    [A_LANE] = "lane",
    [A_VPI] = "vpi",
    [A_VCI] = "vci",
    [A_CONNECTMSG] = "connectmsg",
    [A_METACONNECT] = "metaconnect",
  };
  return qual2kw("ATM keyword", id, tokens, sizeof(tokens) / sizeof(tokens[0]));
}

// SS7 keywords
static const char *
ss7kw(const unsigned id)
{
  const char * tokens[] = {
    [M_FISU] = "fisu",
    [M_LSSU] = "lssu",
    [M_MSU] = "msu",
    [MH_FISU] = "hfisu",
    [MH_LSSU] = "hlssu",
    [MH_MSU] = "hmsu",
    [M_SIO] = "sio",
    [M_OPC] = "opc",
    [M_DPC] = "dpc",
    [M_SLS] = "sls",
    [MH_SIO] = "hsio",
    [MH_OPC] = "hopc",
    [MH_DPC] = "hdpc",
    [MH_SLS] = "hsls",
  };
  return qual2kw("MTP keyword", id, tokens, sizeof(tokens) / sizeof(tokens[0]));
}

// Produce as descriptive an identification string of the DLT as possible.
static const char *
pcapint_datalink_val_to_string(const int dlt)
{
  static thread_local char ret[1024];
  const char *name = pcap_datalink_val_to_name(dlt);
  const char *descr = pcap_datalink_val_to_description(dlt);
  /*
   * Belt and braces: if dlt_choices[] continues to be defined the way it is
   * defined now and everything goes well, either both pointers are NULL or
   * both pointers are not NULL.  But let's not rely on that.
   */
  if (name) {
    if (descr)
      snprintf(ret, sizeof(ret), "DLT_%s (%s)", name, descr);
    else
      snprintf(ret, sizeof(ret), "DLT_%s", name);
    return ret;
  }
  // name == NULL
  if (descr) {
    snprintf(ret, sizeof(ret), "DLT %d (%s)", dlt, descr);
    return ret;
  }
  // Both are NULL, use a function that always returns a non-NULL.
  return pcap_datalink_val_to_description_or_dlt(dlt);
}

static PCAP_NORETURN_DEF void
fail_kw_on_dlt(compiler_state_t *cstate, const char *keyword)
{
  bpf_error(cstate, "'%s' not supported on %s", keyword,
      pcapint_datalink_val_to_string(cstate->linktype));
}

static void
assert_pflog(compiler_state_t *cstate, const char *kw)
{
  if (cstate->linktype != DLT_PFLOG)
    bpf_error(cstate, "'%s' supported only on PFLOG linktype", kw);
}

static void
assert_atm(compiler_state_t *cstate, const char *kw)
{
  /*
   * Belt and braces: init_linktype() sets either all of these struct
   * members (for DLT_SUNATM) or none (otherwise).
   */
  if (cstate->linktype != DLT_SUNATM ||
      ! cstate->is_atm ||
      cstate->off_vpi == OFFSET_NOT_SET ||
      cstate->off_vci == OFFSET_NOT_SET ||
      cstate->off_proto == OFFSET_NOT_SET ||
      cstate->off_payload == OFFSET_NOT_SET)
    bpf_error(cstate, "'%s' supported only on SUNATM", kw);
}

static void
assert_ss7(compiler_state_t *cstate, const char *kw)
{
  switch (cstate->linktype) {
  case DLT_MTP2:
  case DLT_ERF:
  case DLT_MTP2_WITH_PHDR:
    // Belt and braces, same as in assert_atm().
    if (cstate->off_sio != OFFSET_NOT_SET &&
        cstate->off_opc != OFFSET_NOT_SET &&
        cstate->off_dpc != OFFSET_NOT_SET &&
        cstate->off_sls != OFFSET_NOT_SET)
      return;
  }
  bpf_error(cstate, "'%s' supported only on SS7", kw);
}

static void
assert_maxval(compiler_state_t *cstate, const char *name,
    const bpf_u_int32 val, const bpf_u_int32 maxval)
{
  if (val > maxval)
    bpf_error(cstate, "%s %u greater than maximum %u",
        name, val, maxval);
}

static void
assert_nonwlan_dqual(compiler_state_t *cstate, const u_char dir)
{
  switch (dir) {
  case Q_SRC:
  case Q_DST:
  case Q_AND:
  case Q_DEFAULT:
  case Q_OR:
    break;
  default:
    bpf_error(cstate, "'%s' is valid for 802.11 syntax only", dqkw(dir));
  }
}

#define ERRSTR_INVALID_QUAL "'%s' is not a valid qualifier for '%s'"
#define ERRSTR_UNKNOWN_MAC48HOST "unknown Ethernet-like host '%s'"
#define ERRSTR_INVALID_IPV4_ADDR "invalid IPv4 address '%s'"
#define ERRSTR_FUNC_VAR_STR "internal error in %s(): %s == '%s'"

// Validate a port/portrange proto qualifier and map to an IP protocol number.
static int
port_pq_to_ipproto(compiler_state_t *cstate, const int proto, const char *kw)
{
  switch (proto) {
  case Q_UDP:
    return IPPROTO_UDP;
  case Q_TCP:
    return IPPROTO_TCP;
  case Q_SCTP:
    return IPPROTO_SCTP;
  case Q_DEFAULT:
    return PROTO_UNDEF;
  }
  bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), kw);
}

static uint8_t
pq_to_ipproto(compiler_state_t *cstate, const uint8_t pqual)
{
  static const uint8_t map[UINT8_MAX + 1] = {
    [Q_AH]     = IPPROTO_AH,
    [Q_CARP]   = IPPROTO_CARP,
    [Q_ESP]    = IPPROTO_ESP,
    [Q_ICMP]   = IPPROTO_ICMP,
    [Q_ICMPV6] = IPPROTO_ICMPV6,
    [Q_IGMP]   = IPPROTO_IGMP,
    [Q_IGRP]   = IPPROTO_IGRP,
    [Q_PIM]    = IPPROTO_PIM,
    [Q_SCTP]   = IPPROTO_SCTP,
    [Q_TCP]    = IPPROTO_TCP,
    [Q_UDP]    = IPPROTO_UDP,
    [Q_VRRP]   = IPPROTO_VRRP,
  };
  if (map[pqual])
    return map[pqual];
  bpf_error(cstate, "Proto qualifier '%s' has no IP protocol",
      pqkw(pqual));
}

static uint8_t
pq_to_llcsap(compiler_state_t *cstate, const uint8_t pqual)
{
  static const uint8_t map[UINT8_MAX + 1] = {
    [Q_IPX]     = LLCSAP_IPX,
    [Q_ISO]     = LLCSAP_ISONS,
    [Q_NETBEUI] = LLCSAP_NETBEUI,
    [Q_STP]     = LLCSAP_8021D,
  };
  if (map[pqual])
    return map[pqual];
  bpf_error(cstate, "Proto qualifier '%s' has no LLC SAP", pqkw(pqual));
}

static uint16_t
pq_to_ethertype(compiler_state_t *cstate, const uint8_t pqual)
{
  static const uint16_t map[UINT8_MAX + 1] = {
    [Q_AARP]   = ETHERTYPE_AARP,
    [Q_ARP]    = ETHERTYPE_ARP,
    [Q_ATALK]  = ETHERTYPE_ATALK,
    [Q_DECNET] = ETHERTYPE_DN,
    [Q_IP]     = ETHERTYPE_IP,
    [Q_IPV6]   = ETHERTYPE_IPV6,
    [Q_LAT]    = ETHERTYPE_LAT,
    [Q_MOPDL]  = ETHERTYPE_MOPDL,
    [Q_MOPRC]  = ETHERTYPE_MOPRC,
    [Q_RARP]   = ETHERTYPE_REVARP,
    [Q_SCA]    = ETHERTYPE_SCA,
  };
  if (map[pqual])
    return map[pqual];
  bpf_error(cstate, "Proto qualifier '%s' has no EtherType", pqkw(pqual));
}

static uint8_t
pq_to_nlpid(compiler_state_t *cstate, const uint8_t pqual)
{
  static const uint8_t map[UINT8_MAX + 1] = {
    [Q_ESIS] = ISO9542_ESIS,
    [Q_ISIS] = ISO10589_ISIS,
    [Q_CLNP] = ISO8473_CLNP,
  };
  if (map[pqual])
    return map[pqual];
  bpf_error(cstate, "Proto qualifier '%s' has no NLPID", pqkw(pqual));
}

int
pcap_compile(pcap_t *p, struct bpf_program *program,
       const char *buf, int optimize, bpf_u_int32 mask)
{
#ifdef _WIN32
  int err;
  WSADATA wsaData;
#endif
  compiler_state_t cstate;
  yyscan_t scanner = NULL;
  YY_BUFFER_STATE in_buffer = NULL;
  int rc;

  /*
   * If this pcap_t hasn't been activated, it doesn't have a
   * link-layer type, so we can't use it.
   */
  if (!p->activated) {
    (void)snprintf(p->errbuf, PCAP_ERRBUF_SIZE,
        "not-yet-activated pcap_t passed to pcap_compile");
    return (PCAP_ERROR);
  }

#ifdef _WIN32
  /*
   * Initialize Winsock, asking for the latest version (2.2),
   * as we may be calling Winsock routines to translate
   * host names to addresses.
   */
  err = WSAStartup(MAKEWORD(2, 2), &wsaData);
  if (err != 0) {
    pcapint_fmt_errmsg_for_win32_err(p->errbuf, PCAP_ERRBUF_SIZE,
        err, "Error calling WSAStartup()");
    return (PCAP_ERROR);
  }
#endif

#ifdef ENABLE_REMOTE
  /*
   * If the device on which we're capturing need to be notified
   * that a new filter is being compiled, do so.
   *
   * This allows them to save a copy of it, in case, for example,
   * they're implementing a form of remote packet capture, and
   * want the remote machine to filter out the packets in which
   * it's sending the packets it's captured.
   *
   * XXX - the fact that we happen to be compiling a filter
   * doesn't necessarily mean we'll be installing it as the
   * filter for this pcap_t; we might be running it from userland
   * on captured packets to do packet classification.  We really
   * need a better way of handling this, but this is all that
   * the WinPcap remote capture code did.
   */
  if (p->save_current_filter_op != NULL)
    (p->save_current_filter_op)(p, buf);
#endif

  cstate.no_optimize = 0;
  cstate.ai = NULL;
  cstate.ic.root = NULL;
  cstate.ic.cur_mark = 0;
  cstate.bpf_pcap = p;
  cstate.error_set = 0;
  init_regs(&cstate);

  // cstate.error_set must have been initialized first.
  if (! initchunks_ok(&cstate)) {
    // The error buffer has been filled.
    rc = PCAP_ERROR;
    goto quit;
  }

  cstate.netmask = mask;

  cstate.snaplen = pcap_snapshot(p);
  if (cstate.snaplen == 0) {
    (void)snprintf(p->errbuf, PCAP_ERRBUF_SIZE,
       "snaplen of 0 rejects all packets");
    rc = PCAP_ERROR;
    goto quit;
  }

  if (pcap_lex_init(&scanner) != 0) {
    pcapint_fmt_errmsg_for_errno(p->errbuf, PCAP_ERRBUF_SIZE,
        errno, "can't initialize scanner");
    rc = PCAP_ERROR;
    goto quit;
  }
  in_buffer = pcap__scan_string(buf ? buf : "", scanner);

  /*
   * Associate the compiler state with the lexical analyzer
   * state.
   */
  pcap_set_extra(&cstate, scanner);

  if (init_linktype(&cstate, p) == -1) {
    rc = PCAP_ERROR;
    goto quit;
  }
  if (pcap_parse(scanner, &cstate) != 0) {
    if (cstate.ai != NULL)
      freeaddrinfo(cstate.ai);
    rc = PCAP_ERROR;
    goto quit;
  }

  if (cstate.ic.root == NULL) {
    cstate.ic.root = gen_retblk(&cstate, cstate.snaplen);

    /*
     * Catch errors reported by gen_retblk().
     */
    if (cstate.ic.root== NULL) {
      rc = PCAP_ERROR;
      goto quit;
    }
  }

  if (optimize && !cstate.no_optimize) {
    if (bpf_optimize(&cstate.ic, p->errbuf) == -1) {
      /* Failure */
      rc = PCAP_ERROR;
      goto quit;
    }
    if (cstate.ic.root == NULL ||
        (cstate.ic.root->s.code == (BPF_RET|BPF_K) && cstate.ic.root->s.k == 0)) {
      (void)snprintf(p->errbuf, PCAP_ERRBUF_SIZE,
          "expression rejects all packets");
      rc = PCAP_ERROR;
      goto quit;
    }
  }
  program->bf_insns = icode_to_fcode(&cstate.ic,
      cstate.ic.root, &program->bf_len, p->errbuf);
  if (program->bf_insns == NULL) {
    /* Failure */
    rc = PCAP_ERROR;
    goto quit;
  }

  /*
   * If the code generator and the optimizer (if involved) work
   * correctly, the resulting filter program is valid.  If it is invalid,
   * fail now to make these types of bugs easier to detect and to debug.
   *
   * This sanity check is duplicate when the result is immediately used
   * with pcap_setfilter(), which validates the program too.  However,
   * pcap_offline_filter() will just quietly reject the packet if the BPF
   * interpreter runs into an invalid detail.  Also the program could be
   * used in external code and/or at a later time and/or after being
   * stored in a file or transmitted over the network.
   */
  if (pcapint_validate_filter(program->bf_insns, program->bf_len))
    rc = 0; /* We're all okay */
  else {
    snprintf(p->errbuf, PCAP_ERRBUF_SIZE,
        "%s: program validation failed", __func__);
    rc = PCAP_ERROR;
  }

quit:
  /*
   * Clean up everything for the lexical analyzer.
   */
  if (in_buffer != NULL)
    pcap__delete_buffer(in_buffer, scanner);
  if (scanner != NULL)
    pcap_lex_destroy(scanner);

  /*
   * Clean up our own allocated memory.
   */
  freechunks(&cstate);

#ifdef _WIN32
  WSACleanup();
#endif

  return (rc);
}

/*
 * entry point for using the compiler with no pcap open
 * pass in all the stuff that is needed explicitly instead.
 */
int
pcap_compile_nopcap(int snaplen_arg, int linktype_arg,
        struct bpf_program *program,
        const char *buf, int optimize, bpf_u_int32 mask)
{
  pcap_t *p;
  int ret;

  p = pcap_open_dead(linktype_arg, snaplen_arg);
  if (p == NULL)
    return (PCAP_ERROR);
  ret = pcap_compile(p, program, buf, optimize, mask);
  pcap_close(p);
  return (ret);
}

/*
 * Clean up a "struct bpf_program" by freeing all the memory allocated
 * in it.
 */
void
pcap_freecode(struct bpf_program *program)
{
  program->bf_len = 0;
  if (program->bf_insns != NULL) {
    free(program->bf_insns);
    program->bf_insns = NULL;
  }
}

/*
 * Backpatch the blocks in 'list' to 'target'.  The 'sense' field indicates
 * which of the jt and jf fields has been resolved and which is a pointer
 * back to another unresolved block (or nil).  At least one of the fields
 * in each block is already resolved.
 */
static void
backpatch(struct block *list, struct block *target)
{
  struct block *next;

  while (list) {
    if (!list->sense) {
      next = JT(list);
      JT(list) = target;
    } else {
      next = JF(list);
      JF(list) = target;
    }
    list = next;
  }
}

/*
 * Merge the lists in b0 and b1, using the 'sense' field to indicate
 * which of jt and jf is the link.
 */
static void
merge(struct block *b0, struct block *b1)
{
  struct block **p = &b0;

  /* Find end of list. */
  while (*p)
    p = !((*p)->sense) ? &JT(*p) : &JF(*p);

  /* Concatenate the lists. */
  *p = b1;
}

int
finish_parse(compiler_state_t *cstate, struct block *p_arg)
{
  /*
   * Catch errors reported by us and routines below us, and return -1
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (-1);

  struct block *p = p_arg; // "might be clobbered by longjmp()"

  /*
   * Insert before the statements of the first (root) block any
   * statements needed to load the lengths of any variable-length
   * headers into registers.
   *
   * XXX - a fancier strategy would be to insert those before the
   * statements of all blocks that use those lengths and that
   * have no predecessors that use them, so that we only compute
   * the lengths if we need them.  There might be even better
   * approaches than that.
   *
   * However, those strategies would be more complicated, and
   * as we don't generate code to compute a length if the
   * program has no tests that use the length, and as most
   * tests will probably use those lengths, we would just
   * postpone computing the lengths so that it's not done
   * for tests that fail early, and it's not clear that's
   * worth the effort.
   */
  insert_compute_vloffsets(cstate, p->head);

  /*
   * For DLT_PPI captures, generate a check of the per-packet
   * DLT value to make sure it's DLT_IEEE802_11.
   *
   * XXX - TurboCap cards use DLT_PPI for Ethernet.
   * Can we just define some DLT_ETHERNET_WITH_PHDR pseudo-header
   * with appropriate Ethernet information and use that rather
   * than using something such as DLT_PPI where you don't know
   * the link-layer header type until runtime, which, in the
   * general case, would force us to generate both Ethernet *and*
   * 802.11 code (*and* anything else for which PPI is used)
   * and choose between them early in the BPF program?
   */
  if (cstate->linktype == DLT_PPI) {
    struct block *ppi_dlt_check = gen_cmp(cstate, OR_PACKET,
      4, BPF_W, PCAP_BSWAP_32(DLT_IEEE802_11));
    p = gen_and(ppi_dlt_check, p);
  }

  backpatch(p, gen_retblk_internal(cstate, cstate->snaplen));
  p->sense = !p->sense;
  backpatch(p, gen_retblk_internal(cstate, 0));
  cstate->ic.root = p->head;
  return (0);
}

struct block *
gen_and(struct block *b0, struct block *b1)
{
  // False and X is false.
  if (b0->meaning == IS_FALSE)
    return b0;
  // X and false is false.
  if (b1->meaning == IS_FALSE)
    return b1;
  // True and X is X.
  if (b0->meaning == IS_TRUE)
    return b1;
  // X and true is X.
  if (b1->meaning == IS_TRUE)
    return b0;

  // b0->meaning == IS_UNCERTAIN && b1->meaning == IS_UNCERTAIN
  backpatch(b0, b1->head);
  b0->sense = !b0->sense;
  b1->sense = !b1->sense;
  merge(b1, b0);
  b1->sense = !b1->sense;
  b1->head = b0->head;
  return b1;
}

struct block *
gen_or(struct block *b0, struct block *b1)
{
  // False or X is X.
  if (b0->meaning == IS_FALSE)
    return b1;
  // X or false is X.
  if (b1->meaning == IS_FALSE)
    return b0;
  // True or X is true.
  if (b0->meaning == IS_TRUE)
    return b0;
  // X or true is true.
  if (b1->meaning == IS_TRUE)
    return b1;

  // b0->meaning == IS_UNCERTAIN && b1->meaning == IS_UNCERTAIN
  b0->sense = !b0->sense;
  backpatch(b0, b1->head);
  b0->sense = !b0->sense;
  merge(b1, b0);
  b1->head = b0->head;
  return b1;
}

struct block *
gen_not(struct block *b)
{
  b->sense = !b->sense;
  // A switch on an enum is a source of compiler warnings.
  if (b->meaning == IS_TRUE)
    b->meaning = IS_FALSE;
  else if (b->meaning == IS_FALSE)
    b->meaning = IS_TRUE;
  return b;
}

static struct block *
gen_cmp(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JEQ, 0, v);
}

static struct block *
gen_cmp_gt(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JGT, 0, v);
}

static struct block *
gen_cmp_ge(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JGE, 0, v);
}

static struct block *
gen_cmp_lt(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JGE, 1, v);
}

static struct block *
gen_cmp_le(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JGT, 1, v);
}

static struct block *
gen_cmp_ne(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v)
{
  return gen_ncmp(cstate, offrel, offset, size, 0xffffffff, BPF_JEQ, 1, v);
}

static struct block *
gen_mcmp(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v, bpf_u_int32 mask)
{
  /*
   * For any A: if mask == 0, it means A & mask == 0, so the result is
   * true iff v == 0.  In this case ideally the caller should have
   * skipped this invocation and have fewer statement blocks to juggle.
   * If the caller could have skipped, but has not, produce a block with
   * fewer statements.
   *
   * This could be done in gen_ncmp() in a more generic way, but this
   * function is the only code path that can have mask == 0.
   */
  if (mask == 0)
    return v ? gen_false(cstate) : gen_true(cstate);

  return gen_ncmp(cstate, offrel, offset, size, mask, BPF_JEQ, 0, v);
}

static struct block *
gen_mcmp_ne(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 v, bpf_u_int32 mask)
{
  return gen_ncmp(cstate, offrel, offset, size, mask, BPF_JEQ, 1, v);
}

static struct block *
gen_bcmp(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, const u_char *v)
{
  struct block *b, *tmp;

  b = NULL;
  /*
   * If everything everywhere always goes right, the initial value of
   * 'size' is greater than zero, this check is dead code and 'b' will
   * not remain NULL.  However, various code that calls this function
   * does not check for a NULL return value, so just in case something
   * goes wrong somewhere else fail safely here instead of causing a NULL
   * dereference upon return.
   */
  if (! size)
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "size", size);
  while (size >= 4) {
    const u_char *p = &v[size - 4];

    tmp = gen_cmp(cstate, offrel, offset + size - 4, BPF_W,
        EXTRACT_BE_U_4(p));
    b = b ? gen_and(b, tmp) : tmp;
    size -= 4;
  }
  while (size >= 2) {
    const u_char *p = &v[size - 2];

    tmp = gen_cmp(cstate, offrel, offset + size - 2, BPF_H,
        EXTRACT_BE_U_2(p));
    b = b ? gen_and(b, tmp) : tmp;
    size -= 2;
  }
  if (size > 0) {
    tmp = gen_cmp(cstate, offrel, offset, BPF_B, v[0]);
    b = b ? gen_and(b, tmp) : tmp;
  }
  return b;
}

/*
 * Generate an instruction block for one of {"jeq #k", "jgt #k", "jge #k",
 * "jset #k", "ja L"}.
 */
static struct block *
gen_jmp_k(compiler_state_t *cstate, const int jtype, const bpf_u_int32 v,
          struct slist *stmts)
{
  struct block *b = new_block(cstate, OPCODE_BRANCH_K(jtype));
  b->s.k = v;
  b->stmts = stmts;
  return b;
}

/*
 * Generate an instruction block for one of {"jeq x", "jgt x", "jge x",
 * "jset x"}.
 */
static struct block *
gen_jmp_x(compiler_state_t *cstate, const int jtype, struct slist *stmts)
{
  struct block *b = new_block(cstate, OPCODE_BRANCH_X(jtype));
  b->stmts = stmts;
  return b;
}

static struct block *
gen_set(compiler_state_t *cstate, bpf_u_int32 v, struct slist *stmts)
{
  return gen_jmp_k(cstate, BPF_JSET, v, stmts);
}

static struct block *
gen_unset(compiler_state_t *cstate, bpf_u_int32 v, struct slist *stmts)
{
  return gen_not(gen_set(cstate, v, stmts));
}

/*
 * AND the field of size "size" at offset "offset" relative to the header
 * specified by "offrel" with "mask", and compare it with the value "v"
 * with the test specified by "jtype"; if "reverse" is true, the test
 * should test the opposite of "jtype".
 */
static struct block *
gen_ncmp(compiler_state_t *cstate, enum e_offrel offrel, u_int offset,
    u_int size, bpf_u_int32 mask, int jtype, int reverse,
    bpf_u_int32 v)
{
  struct slist *s, *s2;
  struct block *b;

  s = gen_load_a(cstate, offrel, offset, size);

  if (mask != 0xffffffff) {
    s2 = NEW_STMT_AND_K(cstate, mask);
    sappend(s, s2);
  }

  b = gen_jmp_k(cstate, jtype, v, s);
  return reverse ? gen_not(b) : b;
}

static int
init_linktype(compiler_state_t *cstate, pcap_t *p)
{
  cstate->pcap_fddipad = p->fddipad;

  /*
   * We start out with only one link-layer header.
   */
  cstate->outermostlinktype = pcap_datalink(p);
  cstate->off_outermostlinkhdr.constant_part = 0;
  cstate->off_outermostlinkhdr.is_variable = 0;
  cstate->off_outermostlinkhdr.reg = -1;

  cstate->prevlinktype = cstate->outermostlinktype;
  cstate->off_prevlinkhdr.constant_part = 0;
  cstate->off_prevlinkhdr.is_variable = 0;
  cstate->off_prevlinkhdr.reg = -1;

  cstate->linktype = cstate->outermostlinktype;
  cstate->off_linkhdr.constant_part = 0;
  cstate->off_linkhdr.is_variable = 0;
  cstate->off_linkhdr.reg = -1;

  /*
   * XXX
   */
  cstate->off_linkpl.constant_part = 0;
  cstate->off_linkpl.is_variable = 0;
  cstate->off_linkpl.reg = -1;

  cstate->off_linktype.constant_part = 0;
  cstate->off_linktype.is_variable = 0;
  cstate->off_linktype.reg = -1;

  /*
   * Assume it's not raw ATM with a pseudo-header, for now.
   */
  cstate->is_atm = 0;
  cstate->off_vpi = OFFSET_NOT_SET;
  cstate->off_vci = OFFSET_NOT_SET;
  cstate->off_proto = OFFSET_NOT_SET;
  cstate->off_payload = OFFSET_NOT_SET;

  /*
   * And not encapsulated with either Geneve or VXLAN.
   */
  cstate->is_encap = 0;

  /*
   * No variable length VLAN offset by default
   */
  cstate->is_vlan_vloffset = 0;

  /*
   * And assume we're not doing SS7.
   */
  cstate->off_li = OFFSET_NOT_SET;
  cstate->off_li_hsl = OFFSET_NOT_SET;
  cstate->off_sio = OFFSET_NOT_SET;
  cstate->off_opc = OFFSET_NOT_SET;
  cstate->off_dpc = OFFSET_NOT_SET;
  cstate->off_sls = OFFSET_NOT_SET;

  cstate->label_stack_depth = 0;
  cstate->vlan_stack_depth = 0;

  switch (cstate->linktype) {

  case DLT_ARCNET:
    cstate->off_linktype.constant_part = 2;
    cstate->off_linkpl.constant_part = 6;
    cstate->off_nl = 0;   /* XXX in reality, variable! */
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_ARCNET_LINUX:
    cstate->off_linktype.constant_part = 4;
    cstate->off_linkpl.constant_part = 8;
    cstate->off_nl = 0;   /* XXX in reality, variable! */
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_EN10MB:
    cstate->off_linktype.constant_part = 12;
    cstate->off_linkpl.constant_part = 14;  /* Ethernet header length */
    cstate->off_nl = 0;   /* Ethernet II */
    cstate->off_nl_nosnap = 3;  /* 802.3+802.2 */
    break;

  case DLT_SLIP:
    /*
     * SLIP doesn't have a link level type.  The 16 byte
     * header is hacked into our SLIP driver.
     */
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 16;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_SLIP_BSDOS:
    /* XXX this may be the same as the DLT_PPP_BSDOS case */
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    /* XXX end */
    cstate->off_linkpl.constant_part = 24;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_NULL:
  case DLT_LOOP:
    cstate->off_linktype.constant_part = 0;
    cstate->off_linkpl.constant_part = 4;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_ENC:
    cstate->off_linktype.constant_part = 0;
    cstate->off_linkpl.constant_part = 12;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_PPP:
  case DLT_PPP_PPPD:
  case DLT_C_HDLC:   /* BSD/OS Cisco HDLC */
  case DLT_HDLC:     /* NetBSD (Cisco) HDLC */
  case DLT_PPP_SERIAL:   /* NetBSD sync/async serial PPP */
    cstate->off_linktype.constant_part = 2; /* skip HDLC-like framing */
    cstate->off_linkpl.constant_part = 4; /* skip HDLC-like framing and protocol field */
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_PPP_ETHER:
    /*
     * This does not include the Ethernet header, and
     * only covers session state.
     */
    cstate->off_linktype.constant_part = 6;
    cstate->off_linkpl.constant_part = 8;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_PPP_BSDOS:
    cstate->off_linktype.constant_part = 5;
    cstate->off_linkpl.constant_part = 24;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_FDDI:
    /*
     * FDDI doesn't really have a link-level type field.
     * We set "off_linktype" to the offset of the LLC header.
     *
     * To check for Ethernet types, we assume that SSAP = SNAP
     * is being used and pick out the encapsulated Ethernet type.
     * XXX - should we generate code to check for SNAP?
     */
    cstate->off_linktype.constant_part = 13;
    cstate->off_linktype.constant_part += cstate->pcap_fddipad;
    cstate->off_linkpl.constant_part = 13;  /* FDDI MAC header length */
    cstate->off_linkpl.constant_part += cstate->pcap_fddipad;
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_IEEE802:
    /*
     * Token Ring doesn't really have a link-level type field.
     * We set "off_linktype" to the offset of the LLC header.
     *
     * To check for Ethernet types, we assume that SSAP = SNAP
     * is being used and pick out the encapsulated Ethernet type.
     * XXX - should we generate code to check for SNAP?
     *
     * XXX - the header is actually variable-length.
     * Some various Linux patched versions gave 38
     * as "off_linktype" and 40 as "off_nl"; however,
     * if a token ring packet has *no* routing
     * information, i.e. is not source-routed, the correct
     * values are 20 and 22, as they are in the vanilla code.
     *
     * A packet is source-routed iff the uppermost bit
     * of the first byte of the source address, at an
     * offset of 8, has the uppermost bit set.  If the
     * packet is source-routed, the total number of bytes
     * of routing information is 2 plus bits 0x1F00 of
     * the 16-bit value at an offset of 14 (shifted right
     * 8 - figure out which byte that is).
     */
    cstate->off_linktype.constant_part = 14;
    cstate->off_linkpl.constant_part = 14;  /* Token Ring MAC header length */
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
    cstate->off_linkhdr.is_variable = 1;
    /* Fall through, 802.11 doesn't have a variable link
     * prefix but is otherwise the same. */
    /* FALLTHROUGH */

  case DLT_IEEE802_11:
    /*
     * 802.11 doesn't really have a link-level type field.
     * We set "off_linktype.constant_part" to the offset of
     * the LLC header.
     *
     * To check for Ethernet types, we assume that SSAP = SNAP
     * is being used and pick out the encapsulated Ethernet type.
     * XXX - should we generate code to check for SNAP?
     *
     * We also handle variable-length radio headers here.
     * The Prism header is in theory variable-length, but in
     * practice it's always 144 bytes long.  However, some
     * drivers on Linux use ARPHRD_IEEE80211_PRISM, but
     * sometimes or always supply an AVS header, so we
     * have to check whether the radio header is a Prism
     * header or an AVS header, so, in practice, it's
     * variable-length.
     */
    cstate->off_linktype.constant_part = 24;
    cstate->off_linkpl.constant_part = 0; /* link-layer header is variable-length */
    cstate->off_linkpl.is_variable = 1;
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_PPI:
    /*
     * At the moment we treat PPI the same way that we treat
     * normal Radiotap encoded packets. The difference is in
     * the function that generates the code at the beginning
     * to compute the header length.  Since this code generator
     * of PPI supports bare 802.11 encapsulation only (i.e.
     * the encapsulated DLT should be DLT_IEEE802_11) we
     * generate code to check for this too.
     */
    cstate->off_linktype.constant_part = 24;
    cstate->off_linkpl.constant_part = 0; /* link-layer header is variable-length */
    cstate->off_linkpl.is_variable = 1;
    cstate->off_linkhdr.is_variable = 1;
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_ATM_RFC1483:
  case DLT_ATM_CLIP: /* Linux ATM defines this */
    /*
     * assume routed, non-ISO PDUs
     * (i.e., LLC = 0xAA-AA-03, OUT = 0x00-00-00)
     *
     * XXX - what about ISO PDUs, e.g. CLNP, ISIS, ESIS,
     * or PPP with the PPP NLPID (e.g., PPPoA)?  The
     * latter would presumably be treated the way PPPoE
     * should be, so you can do "pppoe and udp port 2049"
     * or "pppoa and tcp port 80" and have it check for
     * PPPo{A,E} and a PPP protocol of IP and....
     */
    cstate->off_linktype.constant_part = 0;
    cstate->off_linkpl.constant_part = 0; /* packet begins with LLC header */
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_SUNATM:
    /*
     * Full Frontal ATM; you get AALn PDUs with an ATM
     * pseudo-header.
     */
    cstate->is_atm = 1;
    cstate->off_vpi = SUNATM_VPI_POS;
    cstate->off_vci = SUNATM_VCI_POS;
    cstate->off_proto = PROTO_POS;
    cstate->off_payload = SUNATM_PKT_BEGIN_POS;
    cstate->off_linktype.constant_part = cstate->off_payload;
    cstate->off_linkpl.constant_part = cstate->off_payload; /* if LLC-encapsulated */
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_RAW:
  case DLT_IPV4:
  case DLT_IPV6:
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 0;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_LINUX_SLL: /* fake header for Linux cooked socket v1 */
    cstate->off_linktype.constant_part = 14;
    cstate->off_linkpl.constant_part = 16;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_LINUX_SLL2: /* fake header for Linux cooked socket v2 */
    cstate->off_linktype.constant_part = 0;
    cstate->off_linkpl.constant_part = 20;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_LTALK:
    /*
     * LocalTalk does have a 1-byte type field in the LLAP header,
     * but really it just indicates whether there is a "short" or
     * "long" DDP packet following.
     */
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 0;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_IP_OVER_FC:
    /*
     * RFC 2625 IP-over-Fibre-Channel doesn't really have a
     * link-level type field.  We set "off_linktype" to the
     * offset of the LLC header.
     *
     * To check for Ethernet types, we assume that SSAP = SNAP
     * is being used and pick out the encapsulated Ethernet type.
     * XXX - should we generate code to check for SNAP? RFC
     * 2625 says SNAP should be used.
     */
    cstate->off_linktype.constant_part = 16;
    cstate->off_linkpl.constant_part = 16;
    cstate->off_nl = 8;   /* 802.2+SNAP */
    cstate->off_nl_nosnap = 3;  /* 802.2 */
    break;

  case DLT_FRELAY:
    /*
     * XXX - we should set this to handle SNAP-encapsulated
     * frames (NLPID of 0x80).
     */
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 0;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

    /*
     * the only BPF-interesting FRF.16 frames are non-control frames;
     * Frame Relay has a variable length link-layer
     * so lets start with offset 4 for now and increments later on (FIXME);
     */
  case DLT_MFR:
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 0;
    cstate->off_nl = 4;
    cstate->off_nl_nosnap = 0;  /* XXX - for now -> no 802.2 LLC */
    break;

  case DLT_APPLE_IP_OVER_IEEE1394:
    cstate->off_linktype.constant_part = 16;
    cstate->off_linkpl.constant_part = 18;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_SYMANTEC_FIREWALL:
    cstate->off_linktype.constant_part = 6;
    cstate->off_linkpl.constant_part = 44;
    cstate->off_nl = 0;   /* Ethernet II */
    cstate->off_nl_nosnap = 0;  /* XXX - what does it do with 802.3 packets? */
    break;

  case DLT_PFLOG:
    cstate->off_linktype.constant_part = 0;
    cstate->off_linkpl.constant_part = 0; /* link-layer header is variable-length */
    cstate->off_linkpl.is_variable = 1;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_MFR:
  case DLT_JUNIPER_MLFR:
  case DLT_JUNIPER_MLPPP:
  case DLT_JUNIPER_PPP:
  case DLT_JUNIPER_CHDLC:
  case DLT_JUNIPER_FRELAY:
    cstate->off_linktype.constant_part = 4;
    cstate->off_linkpl.constant_part = 4;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_ATM1:
    cstate->off_linktype.constant_part = 4;   /* in reality variable between 4-8 */
    cstate->off_linkpl.constant_part = 4; /* in reality variable between 4-8 */
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 10;
    break;

  case DLT_JUNIPER_ATM2:
    cstate->off_linktype.constant_part = 8;   /* in reality variable between 8-12 */
    cstate->off_linkpl.constant_part = 8; /* in reality variable between 8-12 */
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 10;
    break;

    /* frames captured on a Juniper PPPoE service PIC
     * contain raw Ethernet frames */
  case DLT_JUNIPER_PPPOE:
  case DLT_JUNIPER_ETHER:
    cstate->off_linkpl.constant_part = 14;
    cstate->off_linktype.constant_part = 16;
    cstate->off_nl = 18;    /* Ethernet II */
    cstate->off_nl_nosnap = 21; /* 802.3+802.2 */
    break;

  case DLT_JUNIPER_PPPOE_ATM:
    cstate->off_linktype.constant_part = 4;
    cstate->off_linkpl.constant_part = 6;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_GGSN:
    cstate->off_linktype.constant_part = 6;
    cstate->off_linkpl.constant_part = 12;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_ES:
    cstate->off_linktype.constant_part = 6;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET; /* not really a network layer but raw IP addresses */
    cstate->off_nl = OFFSET_NOT_SET; /* not really a network layer but raw IP addresses */
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_MONITOR:
    cstate->off_linktype.constant_part = 12;
    cstate->off_linkpl.constant_part = 12;
    cstate->off_nl = 0;     /* raw IP/IP6 header */
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_SERVICES:
    cstate->off_linktype.constant_part = 12;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET; /* L3 proto location dep. on cookie type */
    cstate->off_nl = OFFSET_NOT_SET; /* L3 proto location dep. on cookie type */
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  case DLT_JUNIPER_VP:
    cstate->off_linktype.constant_part = 18;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_JUNIPER_ST:
    cstate->off_linktype.constant_part = 18;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_JUNIPER_ISM:
    cstate->off_linktype.constant_part = 8;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_JUNIPER_VS:
  case DLT_JUNIPER_SRX_E2E:
  case DLT_JUNIPER_FIBRECHANNEL:
  case DLT_JUNIPER_ATM_CEMIC:
    cstate->off_linktype.constant_part = 8;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_MTP2:
    cstate->off_li = 2;
    cstate->off_li_hsl = 4;
    cstate->off_sio = 3;
    cstate->off_opc = 4;
    cstate->off_dpc = 4;
    cstate->off_sls = 7;
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_MTP2_WITH_PHDR:
    cstate->off_li = 6;
    cstate->off_li_hsl = 8;
    cstate->off_sio = 7;
    cstate->off_opc = 8;
    cstate->off_dpc = 8;
    cstate->off_sls = 11;
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_ERF:
    cstate->off_li = 22;
    cstate->off_li_hsl = 24;
    cstate->off_sio = 23;
    cstate->off_opc = 24;
    cstate->off_dpc = 24;
    cstate->off_sls = 27;
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_PFSYNC:
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = 4;
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = 0;
    break;

  case DLT_IPNET:
    cstate->off_linktype.constant_part = 1;
    cstate->off_linkpl.constant_part = 24;  /* ipnet header length */
    cstate->off_nl = 0;
    cstate->off_nl_nosnap = OFFSET_NOT_SET;
    break;

  case DLT_NETANALYZER:
    cstate->off_linkhdr.constant_part = 4;  /* Ethernet header is past 4-byte pseudo-header */
    cstate->off_linktype.constant_part = cstate->off_linkhdr.constant_part + 12;
    cstate->off_linkpl.constant_part = cstate->off_linkhdr.constant_part + 14;  /* pseudo-header+Ethernet header length */
    cstate->off_nl = 0;   /* Ethernet II */
    cstate->off_nl_nosnap = 3;  /* 802.3+802.2 */
    break;

  case DLT_NETANALYZER_TRANSPARENT:
    cstate->off_linkhdr.constant_part = 12; /* MAC header is past 4-byte pseudo-header, preamble, and SFD */
    cstate->off_linktype.constant_part = cstate->off_linkhdr.constant_part + 12;
    cstate->off_linkpl.constant_part = cstate->off_linkhdr.constant_part + 14;  /* pseudo-header+preamble+SFD+Ethernet header length */
    cstate->off_nl = 0;   /* Ethernet II */
    cstate->off_nl_nosnap = 3;  /* 802.3+802.2 */
    break;

  case DLT_DSA_TAG_BRCM:
    cstate->off_linktype.constant_part = 6 + 6 + 4; // dst, src, DSA tag
    cstate->off_linkpl.constant_part = cstate->off_linktype.constant_part + 2; // idem + EtherType
    cstate->off_nl = 0; // Ethernet II
    cstate->off_nl_nosnap = 3; // 802.3+802.2
    break;

  case DLT_DSA_TAG_DSA:
    cstate->off_linktype.constant_part = 6 + 6 + 4; // dst, src, DSA tag
    cstate->off_linkpl.constant_part = cstate->off_linktype.constant_part + 2; // idem + EtherType
    cstate->off_nl = 0; // Ethernet II
    cstate->off_nl_nosnap = 3; // 802.3+802.2
    break;

  case DLT_EN3MB:
  case DLT_AX25:
  case DLT_PRONET:
  case DLT_CHAOS:
#ifdef DLT_HIPPI
  case DLT_HIPPI:
#endif
  case DLT_REDBACK_SMARTEDGE:
#ifdef DLT_HHDLC
  case DLT_HHDLC:
#endif
    /*
     * Currently, only raw "link[N:M]" filtering is supported.
     */
  case DLT_AX25_KISS:
    /*
     * Idem, plus the initial code for AX.25 KISS commented:
     *
     * - "variable, min 15, max 71 steps of 7" about off_linktype
     * - "variable, min 16, max 71 steps of 7" about off_nl
     *
     * It is not clear how that relates with the AX.25 and KISS
     * specifications, also there is a possibility of Linux kernel
     * modifying the packet type and/or structure.  So if anybody
     * would like to implement a better filtering support for this
     * DLT, it would be a good idea to verify and to document all
     * particulars of the encoding first.
     */
  case DLT_BACNET_MS_TP:
    /*
     * This DLT supports a few primitives besides "link[N:M]", but
     * "link proto", whether explicit or implicit, is not one of
     * these.
     *
     * The third octet of an MS/TP frame is Frame Type, but it is
     * the MS/TP frame type [0..7] rather than a network protocol
     * type.  It can be tested using "link[2]".  If in future it
     * becomes necessary to have a solution that matches the
     * problem space better, it would need to be a new special
     * primitive that works on MS/TP DLT(s) only and takes names
     * for the types, for example, "ms-tp type token".
     */
    cstate->off_linktype.constant_part = OFFSET_NOT_SET;
    cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
    cstate->off_nl = OFFSET_NOT_SET;
    cstate->off_nl_nosnap = OFFSET_NOT_SET; /* no 802.2 LLC */
    break;

  default:
    /*
     * For values in the range in which we've assigned new
     * DLT_ values, only raw "link[N:M]" filtering is supported.
     */
    if (cstate->linktype >= DLT_HIGH_MATCHING_MIN &&
        cstate->linktype <= DLT_HIGH_MATCHING_MAX) {
      cstate->off_linktype.constant_part = OFFSET_NOT_SET;
      cstate->off_linkpl.constant_part = OFFSET_NOT_SET;
      cstate->off_nl = OFFSET_NOT_SET;
      cstate->off_nl_nosnap = OFFSET_NOT_SET;
    } else {
      bpf_set_error(cstate, "unknown data link type %d",
          cstate->linktype);
      return (-1);
    }
    break;
  }

  cstate->off_outermostlinkhdr = cstate->off_prevlinkhdr = cstate->off_linkhdr;
  return (0);
}

/*
 * Load a value relative to the specified absolute offset.
 */
static struct slist *
gen_load_absoffsetrel(compiler_state_t *cstate, struct slist *s,
    const u_int offset, const u_int size)
{
  switch (size) {
  case BPF_B:
  case BPF_H:
  case BPF_W:
    break;
  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "size", size);
  }

  /*
   * If "s" is non-null, it has code to arrange that the X register
   * contains the variable part of the absolute offset, so we
   * generate a load relative to that, with an offset of the constant
   * part of the absolute offset:
   *   (ldb|ldh|ld) [x + k]
   *
   * Otherwise, we can do an absolute load with an offset of the
   * constant part of the absolute offset:
   *   (ldb|ldh|ld) [k]
   */
  if (s != NULL) {
    /*
     * "s" points to a list of statements that puts the
     * variable part of the absolute offset into the X register.
     * Do an indirect load, to use the X register as an offset.
     */
    struct slist *s2 = new_stmt_k(cstate,
        OPCODE_LOAD(size, BPF_IND), offset);
    sappend(s, s2);
  } else {
    /*
     * There is no variable part of the absolute offset, so
     * just do an absolute load.
     */
    s = new_stmt_k(cstate, OPCODE_LOAD(size, BPF_ABS), offset);
  }
  return s;
}

/*
 * Load a value relative to the specified absolute offset and the specified
 * arithmetic expression.
 */
static struct slist *
gen_load_absoffsetarthrel(compiler_state_t *cstate, struct slist *varpart,
    const bpf_u_int32 constpart, const struct arth *arthpart,
    const u_int bpf_size)
{
  /*
   * The required loading offset is a function of three inputs:
   *
   * - the variable part of an absolute offset (either absent or already
   *   loaded into X using the given sequence of instructions),
   * - the constant part of an absolute offset (the given integer), and
   * - the value of a given arithmetic expression (loadable into A or X
   *   from a scratch memory register).
   *
   * Converge this to "(ld|ldh|ldb) [x + k]", where 'k' holds the
   * constant part and 'x' holds the sum of the variable part (if any)
   * and the arithmetic expression value.  That is, if the variable part
   * is absent:
   *   X = <arithmetic expression value>
   * otherwise:
   *   A = <arithmetic expression value>
   *   A = A + X
   *   X = A
   * The rest is a case of a problem that already has a solution.
   */
  if (! varpart)
    varpart = xfer_to_x(cstate, arthpart);
  else {
    sappend(varpart, xfer_to_a(cstate, arthpart));
    sappend(varpart, NEW_STMT_ADD_X(cstate));
    sappend(varpart, NEW_STMT_TAX(cstate));
  }
  return gen_load_absoffsetrel(cstate, varpart, constpart, bpf_size);
}

/*
 * Load a value relative to the beginning of the specified header.
 */
static struct slist *
gen_load_a(compiler_state_t *cstate, const enum e_offrel offrel, u_int offset,
    const u_int size)
{
  struct slist *s;

  /*
   * Squelch warnings from compilers that *don't* assume that
   * offrel always has a valid enum value and therefore don't
   * assume that we'll always go through one of the case arms.
   *
   * If we have a default case, compilers that *do* assume that
   * will then complain about the default case code being
   * unreachable.
   *
   * Damned if you do, damned if you don't.
   */
  s = NULL;

  switch (offrel) {

  case OR_PACKET:
    break;

  case OR_LINKHDR:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkhdr);
    offset += cstate->off_linkhdr.constant_part;
    break;

  case OR_PREVLINKHDR:
    s = gen_abs_offset_varpart(cstate, &cstate->off_prevlinkhdr);
    offset += cstate->off_prevlinkhdr.constant_part;
    break;

  case OR_LLC:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    offset += cstate->off_linkpl.constant_part;
    break;

  case OR_PREVMPLSHDR:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    offset += cstate->off_linkpl.constant_part + cstate->off_nl -
        MPLS_STACKENTRY_LEN;
    break;

  case OR_LINKPL:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    offset += cstate->off_linkpl.constant_part + cstate->off_nl;
    break;

  case OR_LINKPL_NOSNAP:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    offset += cstate->off_linkpl.constant_part +
        cstate->off_nl_nosnap;
    break;

  case OR_LINKTYPE:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linktype);
    offset += cstate->off_linktype.constant_part;
    break;

  case OR_TRAN_IPV4:
    /*
     * Load the X register with the length of the IPv4 header
     * (plus the offset of the link-layer header, if it's
     * preceded by a variable-length header such as a radio
     * header), in bytes.
     */
    s = gen_loadx_iphdrlen(cstate);

    /*
     * Load the item at {offset of the link-layer payload} +
     * {offset, relative to the start of the link-layer
     * payload, of the IPv4 header} + {length of the IPv4 header} +
     * {specified offset}.
     *
     * If the offset of the link-layer payload is variable,
     * the variable part of that offset is included in the
     * value in the X register, and we include the constant
     * part in the offset of the load.
     */
    offset += cstate->off_linkpl.constant_part + cstate->off_nl;
    break;

  case OR_TRAN_IPV6:
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    offset += cstate->off_linkpl.constant_part + cstate->off_nl +
        IP6_HDRLEN;
    break;
  }
  return gen_load_absoffsetrel(cstate, s, offset, size);
}

/*
 * Generate code to load into the X register the sum of the length of
 * the IPv4 header and the variable part of the offset of the link-layer
 * payload.
 */
static struct slist *
gen_loadx_iphdrlen(compiler_state_t *cstate)
{
  struct slist *s, *s2;

  s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
  if (s != NULL) {
    /*
     * The offset of the link-layer payload has a variable
     * part.  "s" points to a list of statements that put
     * the variable part of that offset into the X register.
     *
     * The 4*([k]&0xf) addressing mode can't be used, as we
     * don't have a constant offset, so we have to load the
     * value in question into the A register and add to it
     * the value from the X register.
     */
    s2 = NEW_STMT_LDB_IND(cstate,
        cstate->off_linkpl.constant_part + cstate->off_nl);
    sappend(s, s2);
    s2 = NEW_STMT_AND_K(cstate, 0xf);
    sappend(s, s2);
    s2 = NEW_STMT_LSH_K(cstate, 2);
    sappend(s, s2);

    /*
     * The A register now contains the length of the IP header.
     * We need to add to it the variable part of the offset of
     * the link-layer payload, which is still in the X
     * register, and move the result into the X register.
     */
    sappend(s, NEW_STMT_ADD_X(cstate));
    sappend(s, NEW_STMT_TAX(cstate));
  } else {
    /*
     * The offset of the link-layer payload is a constant,
     * so no code was generated to load the (nonexistent)
     * variable part of that offset.
     *
     * This means we can use the 4*([k]&0xf) addressing
     * mode.  Load the length of the IPv4 header, which
     * is at an offset of cstate->off_nl from the beginning of
     * the link-layer payload, and thus at an offset of
     * cstate->off_linkpl.constant_part + cstate->off_nl from the beginning
     * of the raw packet data, using that addressing mode.
     */
    s = NEW_STMT_LDXB_MSH(cstate,
        cstate->off_linkpl.constant_part + cstate->off_nl);
  }
  return s;
}

/*
 * Produce an instruction block with a final branch statement that takes the
 * true branch iff rsense is not zero.  Since this function detects Boolean
 * constants for potential later use, the resulting block must not be modified
 * directly afterwards, instead it should be used as an argument to gen_and(),
 * gen_or(), gen_not() and sprepend_to_block().
 */
static struct block *
gen_uncond(compiler_state_t *cstate, const u_char rsense)
{
  struct slist *s;

  s = NEW_STMT_LD_IMM(cstate, !rsense);
  struct block *ret = gen_jmp_k(cstate, BPF_JEQ, 0, s);
  ret->meaning = rsense ? IS_TRUE : IS_FALSE;
  return ret;
}

static inline struct block *
gen_true(compiler_state_t *cstate)
{
  return gen_uncond(cstate, 1);
}

static inline struct block *
gen_false(compiler_state_t *cstate)
{
  return gen_uncond(cstate, 0);
}

/*
 * Generate code to match a particular packet type.
 *
 * "proto" is an Ethernet type value, if > ETHERMTU, or an LLC SAP
 * value, if <= ETHERMTU.  We use that to determine whether to
 * match the type/length field or to check the type/length field for
 * a value <= ETHERMTU to see whether it's a type field and then do
 * the appropriate test.
 */
static struct block *
gen_ether_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  struct block *b0, *b1;

  switch (ll_proto) {

  case LLCSAP_ISONS:
  case LLCSAP_IP:
  case LLCSAP_NETBEUI:
    /*
     * OSI protocols and NetBEUI always use 802.2 encapsulation,
     * so we check the DSAP and SSAP.
     *
     * LLCSAP_IP checks for IP-over-802.2, rather
     * than IP-over-Ethernet or IP-over-SNAP.
     *
     * XXX - should we check both the DSAP and the
     * SSAP, like this, or should we check just the
     * DSAP, as we do for other types <= ETHERMTU
     * (i.e., other SAP values)?
     */
    b0 = gen_cmp_le(cstate, OR_LINKTYPE, 0, BPF_H, ETHERMTU);
    b1 = gen_cmp(cstate, OR_LLC, 0, BPF_H, (ll_proto << 8) | ll_proto);
    return gen_and(b0, b1);

  case LLCSAP_IPX:
    /*
     * Check for;
     *
     *  Ethernet_II frames, which are Ethernet
     *  frames with a frame type of ETHERTYPE_IPX;
     *
     *  Ethernet_802.3 frames, which are 802.3
     *  frames (i.e., the type/length field is
     *  a length field, <= ETHERMTU, rather than
     *  a type field) with the first two bytes
     *  after the Ethernet/802.3 header being
     *  0xFFFF;
     *
     *  Ethernet_802.2 frames, which are 802.3
     *  frames with an 802.2 LLC header and
     *  with the IPX LSAP as the DSAP in the LLC
     *  header;
     *
     *  Ethernet_SNAP frames, which are 802.3
     *  frames with an LLC header and a SNAP
     *  header and with an OUI of 0x000000
     *  (encapsulated Ethernet) and a protocol
     *  ID of ETHERTYPE_IPX in the SNAP header.
     *
     * XXX - should we generate the same code both
     * for tests for LLCSAP_IPX and for ETHERTYPE_IPX?
     */

    /*
     * This generates code to check both for the
     * IPX LSAP (Ethernet_802.2) and for Ethernet_802.3.
     */
    b0 = gen_cmp(cstate, OR_LLC, 0, BPF_B, LLCSAP_IPX);
    b1 = gen_cmp(cstate, OR_LLC, 0, BPF_H, 0xFFFF);
    b1 = gen_or(b0, b1);

    /*
     * Now we add code to check for SNAP frames with
     * ETHERTYPE_IPX, i.e. Ethernet_SNAP.
     */
    b0 = gen_snap(cstate, 0x000000, ETHERTYPE_IPX);
    b1 = gen_or(b0, b1);

    /*
     * Now we generate code to check for 802.3
     * frames in general.
     */
    b0 = gen_cmp_le(cstate, OR_LINKTYPE, 0, BPF_H, ETHERMTU);

    /*
     * Now add the check for 802.3 frames before the
     * check for Ethernet_802.2 and Ethernet_802.3,
     * as those checks should only be done on 802.3
     * frames, not on Ethernet frames.
     */
    b1 = gen_and(b0, b1);

    /*
     * Now add the check for Ethernet_II frames, and
     * do that before checking for the other frame
     * types.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ETHERTYPE_IPX);
    return gen_or(b0, b1);

  case ETHERTYPE_ATALK:
  case ETHERTYPE_AARP:
    /*
     * EtherTalk (AppleTalk protocols on Ethernet link
     * layer) may use 802.2 encapsulation.
     */

    /*
     * Check for 802.2 encapsulation (EtherTalk phase 2?);
     * we check for an Ethernet type field less or equal than
     * 1500, which means it's an 802.3 length field.
     */
    b0 = gen_cmp_le(cstate, OR_LINKTYPE, 0, BPF_H, ETHERMTU);

    /*
     * 802.2-encapsulated ETHERTYPE_ATALK packets are
     * SNAP packets with an organization code of
     * 0x080007 (Apple, for Appletalk) and a protocol
     * type of ETHERTYPE_ATALK (Appletalk).
     *
     * 802.2-encapsulated ETHERTYPE_AARP packets are
     * SNAP packets with an organization code of
     * 0x000000 (encapsulated Ethernet) and a protocol
     * type of ETHERTYPE_AARP (Appletalk ARP).
     */
    if (ll_proto == ETHERTYPE_ATALK)
      b1 = gen_snap(cstate, 0x080007, ETHERTYPE_ATALK);
    else  /* ll_proto == ETHERTYPE_AARP */
      b1 = gen_snap(cstate, 0x000000, ETHERTYPE_AARP);
    b1 = gen_and(b0, b1);

    /*
     * Check for Ethernet encapsulation (Ethertalk
     * phase 1?); we just check for the Ethernet
     * protocol type.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);

    return gen_or(b0, b1);

  default:
    if (ll_proto <= ETHERMTU) {
      assert_maxval(cstate, "LLC DSAP", ll_proto, UINT8_MAX);
      /*
       * This is an LLC SAP value, so the frames
       * that match would be 802.2 frames.
       * Check that the frame is an 802.2 frame
       * (i.e., that the length/type field is
       * a length field, <= ETHERMTU) and
       * then check the DSAP.
       */
      b0 = gen_cmp_le(cstate, OR_LINKTYPE, 0, BPF_H, ETHERMTU);
      b1 = gen_cmp(cstate, OR_LINKTYPE, 2, BPF_B, ll_proto);
      return gen_and(b0, b1);
    } else {
      assert_maxval(cstate, "EtherType", ll_proto, UINT16_MAX);
      /*
       * This is an Ethernet type, so compare
       * the length/type field with it (if
       * the frame is an 802.2 frame, the length
       * field will be <= ETHERMTU, and, as
       * "ll_proto" is > ETHERMTU, this test
       * will fail and the frame won't match,
       * which is what we want).
       */
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);
    }
  }
}

/*
 * AF_INET is 2 in all the operating systems we support...
 *
 * ...except for Haiku, which defines it as 1.
 *
 * So we define BSD_AFNUM_INET as 2 (as AF_INET originated in 4.2BSD,
 * and *almost* everybody just adopted it).
 *
 * Haiku doesn't use DLT_NULL (it uses DLT_RAW for the loopback device),
 * so we don't need to check for it in DLT_NULL captures. We should,
 * however, use BSD_AFNUM_INET rathr than AF_INET when checking for
 * IPv4 in DLT_NULL, DLT_LOOP, and DLT_ENC captures.
 */
#define BSD_AFNUM_INET    2  /* Everybody but Haiku (and BeOS?) */

/*
 * The three different values we should check for when checking for an
 * IPv6 packet with DLT_NULL.
 */
#define BSD_AFNUM_INET6_BSD 24  /* NetBSD, OpenBSD, BSD/OS, Npcap */
#define BSD_AFNUM_INET6_FREEBSD 28  /* FreeBSD */
#define BSD_AFNUM_INET6_DARWIN  30  /* macOS, iOS, other Darwin-based OSes */

static struct block *
gen_endian_linktype(compiler_state_t *cstate, u_int offset, u_int size,
    bpf_u_int32 ll_proto, int swapped)
{
  return (gen_cmp(cstate, OR_LINKHDR, offset, size,
      swapped ? PCAP_BSWAP_32(ll_proto) : ll_proto));
}

static struct block *
gen_bsd_af_linktype_live(compiler_state_t *cstate, u_int offset, u_int size,
    bpf_u_int32 ll_proto, int swapped)
{
  switch (ll_proto) {

  case ETHERTYPE_IP:
    return (gen_endian_linktype(cstate, offset, size, AF_INET,
        swapped));

  case ETHERTYPE_IPV6:
    return (gen_endian_linktype(cstate, offset, size, AF_INET6,
        swapped));

  default:
    /*
     * Not a type on which we support filtering.
     * XXX - support those that have AF_ values
     * #defined on this platform, at least?
     */
    return gen_false(cstate);
  }
}

static struct block *
gen_bsd_af_linktype_offline(compiler_state_t *cstate, u_int offset, u_int size,
    bpf_u_int32 ll_proto, int swapped)
{
  struct block *b0, *b1;

  switch (ll_proto) {

  case ETHERTYPE_IP:
    /*
     * Only Haiku (and BeOS?) define AF_INET differently
     * from the value 4.2BSD used (2), and Haiku doesn't
     * have any capture type that uses AF_INET values
     * (its loopback device uses DLT_RAW), so, while
     * we must use BSD_AFNUM_INET when comparing,
     * we don't have to test for more than one value.
     */
    return (gen_endian_linktype(cstate, offset, size, BSD_AFNUM_INET,
        swapped));

  case ETHERTYPE_IPV6:
    /*
     * AF_INET6 values are, unfortunately, be platform-dependent,
     * even on platforms that use it in link-layer headers,
     * because 4.2BSD didn't have a value for it (given that
     * IPv6 didn't exist back in the early 1980's), and they
     * all picked their own values.
     *
     * This means that, if we're reading from a savefile, we
     * need to check for all the possible values.
     *
     * If we're doing a live capture, we only need to check
     * for this platform's value; however, Npcap uses 24,
     * which isn't Windows's AF_INET6 value.  (Given the
     * multiple different values, programs that read pcap
     * files shouldn't be checking for their platform's
     * AF_INET6 value anyway, they should check for all of the
     * possible values. and they might as well do that even for
     * live captures.)
     */
    b0 = gen_endian_linktype(cstate, offset, size,
        BSD_AFNUM_INET6_BSD, swapped);
    b1 = gen_endian_linktype(cstate, offset, size,
        BSD_AFNUM_INET6_FREEBSD, swapped);
    b1 = gen_or(b0, b1);
    b0 = gen_endian_linktype(cstate, offset, size,
        BSD_AFNUM_INET6_DARWIN, swapped);
    return gen_or(b0, b1);

  default:
    /*
     * Not a type on which we support filtering.
     * XXX - support those that have AF_ values
     * #defined on this platform, at least?
     */
    return gen_false(cstate);
  }
}

/*
 * Generate a test for a loopback or DLT_ENC link-layer type; the type
 * is the link-layer type.
 */
static struct block *
gen_loopback_linktype_live(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  switch (cstate->linktype) {

  case DLT_NULL:
  case DLT_ENC:
    /*
     * The value is in host byte order in the packet.
     *
     * If that's big-endian, we can just compare it
     * with the specified type.
     *
     * if that's little-endian, we have to compare it
     * with a byte-swapped version of the specified
     * type.
     *
     * htonl(the specified type) will do nothing to
     * the specified type on a big-endian machine, and
     * will byte-swap the specified type on a little-
     * endian machine, so just use that as the value
     * against which to compare.
     */
    return (gen_cmp(cstate, OR_LINKHDR, 0, BPF_W, htonl(ll_proto)));

  case DLT_LOOP:
    /*
     * The value is in network byte order in the packet,
     * so just compare it with the specified type.
     */
    return (gen_cmp(cstate, OR_LINKHDR, 0, BPF_W, ll_proto));

  default:
    /* Should not happen. */
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "linktype",
        cstate->linktype);
  }
}

/*
 * Generate a test for the DLT_PFLOG  link-layer type; the type is the
 * link-layer type.
 */
static struct block *
gen_pflog_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  if (cstate->bpf_pcap->bpf_codegen_flags & BPF_OFFLINE_AF_HANDLING) {
    return (gen_bsd_af_linktype_offline(cstate,
        offsetof(struct pfloghdr, af), BPF_B, ll_proto, 0));
  }
  return (gen_bsd_af_linktype_live(cstate, offsetof(struct pfloghdr, af),
      BPF_B, ll_proto, 0));
}

static struct block *
gen_loopback_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  struct block *b0, *b1;

  /*
   * For DLT_NULL, the link-layer header is a 32-bit word
   * containing an AF_ value in *host* byte order, and for
   * DLT_ENC, the link-layer header begins with a 32-bit
   * word containing an AF_ value in host byte order.
   *
   * In addition, if we're reading a saved capture file,
   * the host byte order in the capture may not be the
   * same as the host byte order on this machine.
   *
   * For DLT_LOOP, the link-layer header is a 32-bit
   * word containing an AF_ value in *network* byte order.
   */
  if (!(cstate->bpf_pcap->bpf_codegen_flags & BPF_OFFLINE_AF_HANDLING)) {
    /*
     * This is a live caapture, so we just check for this
     * platform's AF_ value (except when we don't).
     *
     * The AF_ value is in host byte order, but the BPF
     * interpreter will convert it to network byte order,
     * so we run it through "htonl()", and generate
     * code to compare against the result.
     */
    switch (ll_proto) {

    case ETHERTYPE_IP:
      return (gen_loopback_linktype_live(cstate, AF_INET));

    case ETHERTYPE_IPV6:
#ifdef _WIN32
      /*
       * Npcap doesn't use Windows's AF_INET6,
       * as that collides with AF_IPX on
       * some BSDs (both have the value 23).
       * Instead, it uses 24 (BSD_AFNUM_INET6_BSD).
       */
      return (gen_loopback_linktype_live(cstate, BSD_AFNUM_INET6_BSD));
#else /* _WIN32 */
      return (gen_loopback_linktype_live(cstate, AF_INET6));
#endif /* _WIN32 */

    default:
      /*
       * Not a type on which we support filtering.
       * XXX - support those that have AF_ values
       * #defined on this platform, at least?
       */
      return gen_false(cstate);
    }
  }

  /*
   * This is a savefile.
   *
   * for DLT_NULL and DLT_ENC, the endianness of the value in the
   * packets is not necessarily the endianness of the capture file,
   * as the endianness of the value in the packets is the endianness
   * of the host that did the capture, but the endianness of the file
   * is the endianness of the host that wrote the file, and this
   * file might be the result of a host with one byte order processing
   * another file from a host with a different order.
   *
   * For those types, we first test for all the types using the
   * byte order of the file, and then test again for all the types
   * with the opposite byte order of the file, under the assumption
   * that the most likely case is that the file was written as
   * a live capture.
   *
   * for DLT_LOOP, the endianness of the value in the packets
   * is always big-endian.
   *
   * For DLT_PFLOG, the field is one byte long, so it has no
   * endianness. (None of our platform are nibble-addressible. :-))
   */
  switch (cstate->linktype) {

  case DLT_NULL:
  case DLT_ENC:
    b0 = gen_bsd_af_linktype_offline(cstate, 0, BPF_W, ll_proto,
        cstate->bpf_pcap->swapped);
    b1 = gen_bsd_af_linktype_offline(cstate, 0, BPF_W, ll_proto,
        !cstate->bpf_pcap->swapped);
    return (gen_or(b0, b1));

  case DLT_LOOP:
    return (gen_bsd_af_linktype_offline(cstate, 0, BPF_W, ll_proto, 0));

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "linktype",
        cstate->linktype);
  }
}

/*
 * "proto" is an Ethernet type value and for IPNET, if it is not IPv4
 * or IPv6 then we have an error.
 */
static struct block *
gen_ipnet_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  switch (ll_proto) {

  case ETHERTYPE_IP:
    return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B, IPH_AF_INET);
    /*NOTREACHED*/

  case ETHERTYPE_IPV6:
    return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B, IPH_AF_INET6);
    /*NOTREACHED*/

  default:
    break;
  }

  return gen_false(cstate);
}

/*
 * Generate code to match a particular packet type.
 *
 * "ll_proto" is an Ethernet type value, if > ETHERMTU, or an LLC SAP
 * value, if <= ETHERMTU.  We use that to determine whether to
 * match the type field or to check the type field for the special
 * LINUX_SLL_P_802_2 value and then do the appropriate test.
 */
static struct block *
gen_linux_sll_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  struct block *b0, *b1;

  switch (ll_proto) {

  case LLCSAP_ISONS:
  case LLCSAP_IP:
  case LLCSAP_NETBEUI:
    /*
     * OSI protocols and NetBEUI always use 802.2 encapsulation,
     * so we check the DSAP and SSAP.
     *
     * LLCSAP_IP checks for IP-over-802.2, rather
     * than IP-over-Ethernet or IP-over-SNAP.
     *
     * XXX - should we check both the DSAP and the
     * SSAP, like this, or should we check just the
     * DSAP, as we do for other types <= ETHERMTU
     * (i.e., other SAP values)?
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, LINUX_SLL_P_802_2);
    b1 = gen_cmp(cstate, OR_LLC, 0, BPF_H, (ll_proto << 8) | ll_proto);
    return gen_and(b0, b1);

  case LLCSAP_IPX:
    /*
     *  Ethernet_II frames, which are Ethernet
     *  frames with a frame type of ETHERTYPE_IPX;
     *
     *  Ethernet_802.3 frames, which have a frame
     *  type of LINUX_SLL_P_802_3;
     *
     *  Ethernet_802.2 frames, which are 802.3
     *  frames with an 802.2 LLC header (i.e, have
     *  a frame type of LINUX_SLL_P_802_2) and
     *  with the IPX LSAP as the DSAP in the LLC
     *  header;
     *
     *  Ethernet_SNAP frames, which are 802.3
     *  frames with an LLC header and a SNAP
     *  header and with an OUI of 0x000000
     *  (encapsulated Ethernet) and a protocol
     *  ID of ETHERTYPE_IPX in the SNAP header.
     *
     * First, do the checks on LINUX_SLL_P_802_2
     * frames; generate the check for either
     * Ethernet_802.2 or Ethernet_SNAP frames, and
     * then put a check for LINUX_SLL_P_802_2 frames
     * before it.
     */
    b0 = gen_cmp(cstate, OR_LLC, 0, BPF_B, LLCSAP_IPX);
    b1 = gen_snap(cstate, 0x000000, ETHERTYPE_IPX);
    b1 = gen_or(b0, b1);
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, LINUX_SLL_P_802_2);
    b1 = gen_and(b0, b1);

    /*
     * Now check for 802.3 frames and OR that with
     * the previous test.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, LINUX_SLL_P_802_3);
    b1 = gen_or(b0, b1);

    /*
     * Now add the check for Ethernet_II frames, and
     * do that before checking for the other frame
     * types.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ETHERTYPE_IPX);
    return gen_or(b0, b1);

  case ETHERTYPE_ATALK:
  case ETHERTYPE_AARP:
    /*
     * EtherTalk (AppleTalk protocols on Ethernet link
     * layer) may use 802.2 encapsulation.
     */

    /*
     * Check for 802.2 encapsulation (EtherTalk phase 2?);
     * we check for the 802.2 protocol type in the
     * "Ethernet type" field.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, LINUX_SLL_P_802_2);

    /*
     * 802.2-encapsulated ETHERTYPE_ATALK packets are
     * SNAP packets with an organization code of
     * 0x080007 (Apple, for Appletalk) and a protocol
     * type of ETHERTYPE_ATALK (Appletalk).
     *
     * 802.2-encapsulated ETHERTYPE_AARP packets are
     * SNAP packets with an organization code of
     * 0x000000 (encapsulated Ethernet) and a protocol
     * type of ETHERTYPE_AARP (Appletalk ARP).
     */
    if (ll_proto == ETHERTYPE_ATALK)
      b1 = gen_snap(cstate, 0x080007, ETHERTYPE_ATALK);
    else  /* ll_proto == ETHERTYPE_AARP */
      b1 = gen_snap(cstate, 0x000000, ETHERTYPE_AARP);
    b1 = gen_and(b0, b1);

    /*
     * Check for Ethernet encapsulation (Ethertalk
     * phase 1?); we just check for the Ethernet
     * protocol type.
     */
    b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);

    return gen_or(b0, b1);

  default:
    if (ll_proto <= ETHERMTU) {
      assert_maxval(cstate, "LLC DSAP", ll_proto, UINT8_MAX);
      /*
       * This is an LLC SAP value, so the frames
       * that match would be 802.2 frames.
       * Check for the 802.2 protocol type
       * in the "Ethernet type" field, and
       * then check the DSAP.
       */
      b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, LINUX_SLL_P_802_2);
      b1 = gen_cmp(cstate, OR_LINKHDR, cstate->off_linkpl.constant_part, BPF_B,
           ll_proto);
      return gen_and(b0, b1);
    } else {
      assert_maxval(cstate, "EtherType", ll_proto, UINT16_MAX);
      /*
       * This is an Ethernet type, so compare
       * the length/type field with it (if
       * the frame is an 802.2 frame, the length
       * field will be <= ETHERMTU, and, as
       * "ll_proto" is > ETHERMTU, this test
       * will fail and the frame won't match,
       * which is what we want).
       */
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);
    }
  }
}

/*
 * Load a value relative to the beginning of the link-layer header after the
 * pflog header.
 */
static struct slist *
gen_load_pflog_llprefixlen(compiler_state_t *cstate)
{
  struct slist *s1, *s2;

  /*
   * Generate code to load the length of the pflog header into
   * the register assigned to hold that length, if one has been
   * assigned.  (If one hasn't been assigned, no code we've
   * generated uses that prefix, so we don't need to generate any
   * code to load it.)
   */
  if (cstate->off_linkpl.reg != -1) {
    /*
     * The length is in the first byte of the header.
     */
    s1 = NEW_STMT_LDB_ABS(cstate, 0);

    /*
     * Round it up to a multiple of 4.
     * Add 3, and clear the lower 2 bits.
     */
    s2 = NEW_STMT_ADD_K(cstate, 3);
    sappend(s1, s2);
    s2 = NEW_STMT_AND_K(cstate, 0xfffffffc);
    sappend(s1, s2);

    /*
     * Now allocate a register to hold that value and store
     * it.
     */
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkpl.reg);
    sappend(s1, s2);

    /*
     * Now move it into the X register.
     */
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    return (s1);
  } else
    return (NULL);
}

static struct slist *
gen_load_prism_llprefixlen(compiler_state_t *cstate)
{
  struct slist *s1, *s2;
  struct slist *sjeq_avs_cookie;
  struct slist *sjcommon;

  /*
   * This code is not compatible with the optimizer, as
   * we are generating jmp instructions within a normal
   * slist of instructions
   */
  cstate->no_optimize = 1;

  /*
   * Generate code to load the length of the radio header into
   * the register assigned to hold that length, if one has been
   * assigned.  (If one hasn't been assigned, no code we've
   * generated uses that prefix, so we don't need to generate any
   * code to load it.)
   *
   * Some Linux drivers use ARPHRD_IEEE80211_PRISM but sometimes
   * or always use the AVS header rather than the Prism header.
   * We load a 4-byte big-endian value at the beginning of the
   * raw packet data, and see whether, when masked with 0xFFFFF000,
   * it's equal to 0x80211000.  If so, that indicates that it's
   * an AVS header (the masked-out bits are the version number).
   * Otherwise, it's a Prism header.
   *
   * XXX - the Prism header is also, in theory, variable-length,
   * but no known software generates headers that aren't 144
   * bytes long.
   */
  if (cstate->off_linkhdr.reg != -1) {
    /*
     * Load the cookie.
     */
    s1 = NEW_STMT_LD_ABS(cstate, 0);

    /*
     * AND it with 0xFFFFF000.
     */
    s2 = NEW_STMT_AND_K(cstate, 0xFFFFF000);
    sappend(s1, s2);

    /*
     * Compare with 0x80211000.
     */
    sjeq_avs_cookie = NEW_STMT_JEQ_K(cstate, 0x80211000);
    sappend(s1, sjeq_avs_cookie);

    /*
     * If it's AVS:
     *
     * The 4 bytes at an offset of 4 from the beginning of
     * the AVS header are the length of the AVS header.
     * That field is big-endian.
     */
    s2 = NEW_STMT_LD_ABS(cstate, 4);
    sappend(s1, s2);
    sjeq_avs_cookie->s.jt = s2;

    /*
     * Now jump to the code to allocate a register
     * into which to save the header length and
     * store the length there.  (The "jump always"
     * instruction needs to have the k field set;
     * it's added to the PC, so, as we're jumping
     * over a single instruction, it should be 1.)
     */
    sjcommon = NEW_STMT_JA(cstate, 1);
    sappend(s1, sjcommon);

    /*
     * Now for the code that handles the Prism header.
     * Just load the length of the Prism header (144)
     * into the A register.  Have the test for an AVS
     * header branch here if we don't have an AVS header.
     */
    s2 = NEW_STMT_LD_IMM(cstate, 144);
    sappend(s1, s2);
    sjeq_avs_cookie->s.jf = s2;

    /*
     * Now allocate a register to hold that value and store
     * it.  The code for the AVS header will jump here after
     * loading the length of the AVS header.
     */
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
    sappend(s1, s2);
    sjcommon->s.jf = s2;

    /*
     * Now move it into the X register.
     */
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    return (s1);
  } else
    return (NULL);
}

static struct slist *
gen_load_avs_llprefixlen(compiler_state_t *cstate)
{
  struct slist *s1, *s2;

  /*
   * Generate code to load the length of the AVS header into
   * the register assigned to hold that length, if one has been
   * assigned.  (If one hasn't been assigned, no code we've
   * generated uses that prefix, so we don't need to generate any
   * code to load it.)
   */
  if (cstate->off_linkhdr.reg != -1) {
    /*
     * The 4 bytes at an offset of 4 from the beginning of
     * the AVS header are the length of the AVS header.
     * That field is big-endian.
     */
    s1 = NEW_STMT_LD_ABS(cstate, 4);

    /*
     * Now allocate a register to hold that value and store
     * it.
     */
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
    sappend(s1, s2);

    /*
     * Now move it into the X register.
     */
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    return (s1);
  } else
    return (NULL);
}

static struct slist *
gen_load_radiotap_llprefixlen(compiler_state_t *cstate)
{
  struct slist *s1, *s2;

  /*
   * Generate code to load the length of the radiotap header into
   * the register assigned to hold that length, if one has been
   * assigned.  (If one hasn't been assigned, no code we've
   * generated uses that prefix, so we don't need to generate any
   * code to load it.)
   */
  if (cstate->off_linkhdr.reg != -1) {
    /*
     * The 2 bytes at offsets of 2 and 3 from the beginning
     * of the radiotap header are the length of the radiotap
     * header; unfortunately, it's little-endian, so we have
     * to load it a byte at a time and construct the value.
     */

    /*
     * Load the high-order byte, at an offset of 3, shift it
     * left a byte, and put the result in the X register.
     */
    s1 = NEW_STMT_LDB_ABS(cstate, 3);
    s2 = NEW_STMT_LSH_K(cstate, 8);
    sappend(s1, s2);
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    /*
     * Load the next byte, at an offset of 2, and OR the
     * value from the X register into it.
     */
    s2 = NEW_STMT_LDB_ABS(cstate, 2);
    sappend(s1, s2);
    s2 = NEW_STMT_OR_X(cstate);
    sappend(s1, s2);

    /*
     * Now allocate a register to hold that value and store
     * it.
     */
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
    sappend(s1, s2);

    /*
     * Now move it into the X register.
     */
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    return (s1);
  } else
    return (NULL);
}

/*
 * At the moment we treat PPI as normal Radiotap encoded
 * packets. The difference is in the function that generates
 * the code at the beginning to compute the header length.
 * Since this code generator of PPI supports bare 802.11
 * encapsulation only (i.e. the encapsulated DLT should be
 * DLT_IEEE802_11) we generate code to check for this too;
 * that's done in finish_parse().
 */
static struct slist *
gen_load_ppi_llprefixlen(compiler_state_t *cstate)
{
  struct slist *s1, *s2;

  /*
   * Generate code to load the length of the radiotap header
   * into the register assigned to hold that length, if one has
   * been assigned.
   */
  if (cstate->off_linkhdr.reg != -1) {
    /*
     * The 2 bytes at offsets of 2 and 3 from the beginning
     * of the radiotap header are the length of the radiotap
     * header; unfortunately, it's little-endian, so we have
     * to load it a byte at a time and construct the value.
     */

    /*
     * Load the high-order byte, at an offset of 3, shift it
     * left a byte, and put the result in the X register.
     */
    s1 = NEW_STMT_LDB_ABS(cstate, 3);
    s2 = NEW_STMT_LSH_K(cstate, 8);
    sappend(s1, s2);
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    /*
     * Load the next byte, at an offset of 2, and OR the
     * value from the X register into it.
     */
    s2 = NEW_STMT_LDB_ABS(cstate, 2);
    sappend(s1, s2);
    s2 = NEW_STMT_OR_X(cstate);
    sappend(s1, s2);

    /*
     * Now allocate a register to hold that value and store
     * it.
     */
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
    sappend(s1, s2);

    /*
     * Now move it into the X register.
     */
    s2 = NEW_STMT_TAX(cstate);
    sappend(s1, s2);

    return (s1);
  } else
    return (NULL);
}

/*
 * Load a value relative to the beginning of the link-layer header after the 802.11
 * header, i.e. LLC_SNAP.
 * The link-layer header doesn't necessarily begin at the beginning
 * of the packet data; there might be a variable-length prefix containing
 * radio information.
 */
static struct slist *
gen_load_802_11_header_len(compiler_state_t *cstate, struct slist *s, struct slist *snext)
{
  struct slist *s2;
  struct slist *sjset_data_frame_1;
  struct slist *sjset_data_frame_2;
  struct slist *sjset_qos;
  struct slist *sjset_radiotap_flags_present;
  struct slist *sjset_radiotap_ext_present;
  struct slist *sjset_radiotap_tsft_present;
  struct slist *sjset_tsft_datapad, *sjset_notsft_datapad;
  struct slist *s_roundup;

  if (cstate->off_linkpl.reg == -1) {
    /*
     * No register has been assigned to the offset of
     * the link-layer payload, which means nobody needs
     * it; don't bother computing it - just return
     * what we already have.
     */
    return (s);
  }

  /*
   * This code is not compatible with the optimizer, as
   * we are generating jmp instructions within a normal
   * slist of instructions
   */
  cstate->no_optimize = 1;

  /*
   * If "s" is non-null, it has code to arrange that the X register
   * contains the length of the prefix preceding the link-layer
   * header.
   *
   * Otherwise, the length of the prefix preceding the link-layer
   * header is "off_outermostlinkhdr.constant_part".
   */
  if (s == NULL) {
    /*
     * There is no variable-length header preceding the
     * link-layer header.
     *
     * Load the length of the fixed-length prefix preceding
     * the link-layer header (if any) into the X register,
     * and store it in the cstate->off_linkpl.reg register.
     * That length is off_outermostlinkhdr.constant_part.
     */
    s = NEW_STMT_LDX_IMM(cstate,
        cstate->off_outermostlinkhdr.constant_part);
  }

  /*
   * The X register contains the offset of the beginning of the
   * link-layer header; add 24, which is the minimum length
   * of the MAC header for a data frame, to that, and store it
   * in cstate->off_linkpl.reg, and then load the Frame Control field,
   * which is at the offset in the X register, with an indexed load.
   */
  s2 = NEW_STMT_TXA(cstate);
  sappend(s, s2);
  s2 = NEW_STMT_ADD_K(cstate, 24);
  sappend(s, s2);
  s2 = NEW_STMT_ST_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s2);

  s2 = NEW_STMT_LDB_IND(cstate, 0);
  sappend(s, s2);

  /*
   * Check the Frame Control field to see if this is a data frame;
   * a data frame has the 0x08 bit (b3) in that field set and the
   * 0x04 bit (b2) clear.
   */
  sjset_data_frame_1 = NEW_STMT_JSET_K(cstate, IEEE80211_FC0_TYPE_DATA);
  sappend(s, sjset_data_frame_1);

  /*
   * If b3 is set, test b2, otherwise go to the first statement of
   * the rest of the program.
   */
  sjset_data_frame_1->s.jt = sjset_data_frame_2 =
      NEW_STMT_JSET_K(cstate, IEEE80211_FC0_TYPE_CTL);
  sappend(s, sjset_data_frame_2);
  sjset_data_frame_1->s.jf = snext;

  /*
   * If b2 is not set, this is a data frame; test the QoS bit.
   * Otherwise, go to the first statement of the rest of the
   * program.
   */
  sjset_data_frame_2->s.jt = snext;
  sjset_data_frame_2->s.jf = sjset_qos =
      NEW_STMT_JSET_K(cstate, IEEE80211_FC0_SUBTYPE_QOS);
  sappend(s, sjset_qos);

  /*
   * If it's set, add 2 to cstate->off_linkpl.reg, to skip the QoS
   * field.
   * Otherwise, go to the first statement of the rest of the
   * program.
   */
  sjset_qos->s.jt = s2 = NEW_STMT_LD_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s2);
  s2 = NEW_STMT_ADD_K(cstate, 2);
  sappend(s, s2);
  s2 = NEW_STMT_ST_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s2);

  /*
   * If we have a radiotap header, look at it to see whether
   * there's Atheros padding between the MAC-layer header
   * and the payload.
   *
   * Note: all of the fields in the radiotap header are
   * little-endian, so we byte-swap all of the values
   * we test against, as they will be loaded as big-endian
   * values.
   *
   * XXX - in the general case, we would have to scan through
   * *all* the presence bits, if there's more than one word of
   * presence bits.  That would require a loop, meaning that
   * we wouldn't be able to run the filter in the kernel.
   *
   * We assume here that the Atheros adapters that insert the
   * annoying padding don't have multiple antennae and therefore
   * do not generate radiotap headers with multiple presence words.
   */
  if (cstate->linktype == DLT_IEEE802_11_RADIO) {
    /*
     * Is the IEEE80211_RADIOTAP_FLAGS bit (0x0000002) set
     * in the first presence flag word?
     */
    sjset_qos->s.jf = s2 = NEW_STMT_LD_ABS(cstate, 4);
    sappend(s, s2);

    sjset_radiotap_flags_present =
        NEW_STMT_JSET_K(cstate, PCAP_BSWAP_32(0x00000002));
    sappend(s, sjset_radiotap_flags_present);

    /*
     * If not, skip all of this.
     */
    sjset_radiotap_flags_present->s.jf = snext;

    /*
     * Otherwise, is the "extension" bit set in that word?
     */
    sjset_radiotap_ext_present =
        NEW_STMT_JSET_K(cstate, PCAP_BSWAP_32(0x80000000));
    sappend(s, sjset_radiotap_ext_present);
    sjset_radiotap_flags_present->s.jt = sjset_radiotap_ext_present;

    /*
     * If so, skip all of this.
     */
    sjset_radiotap_ext_present->s.jt = snext;

    /*
     * Otherwise, is the IEEE80211_RADIOTAP_TSFT bit set?
     */
    sjset_radiotap_tsft_present =
        NEW_STMT_JSET_K(cstate, PCAP_BSWAP_32(0x00000001));
    sappend(s, sjset_radiotap_tsft_present);
    sjset_radiotap_ext_present->s.jf = sjset_radiotap_tsft_present;

    /*
     * If IEEE80211_RADIOTAP_TSFT is set, the flags field is
     * at an offset of 16 from the beginning of the raw packet
     * data (8 bytes for the radiotap header and 8 bytes for
     * the TSFT field).
     *
     * Test whether the IEEE80211_RADIOTAP_F_DATAPAD bit (0x20)
     * is set.
     */
    s2 = NEW_STMT_LDB_ABS(cstate, 16);
    sappend(s, s2);
    sjset_radiotap_tsft_present->s.jt = s2;

    sjset_tsft_datapad = NEW_STMT_JSET_K(cstate, 0x20);
    sappend(s, sjset_tsft_datapad);

    /*
     * If IEEE80211_RADIOTAP_TSFT is not set, the flags field is
     * at an offset of 8 from the beginning of the raw packet
     * data (8 bytes for the radiotap header).
     *
     * Test whether the IEEE80211_RADIOTAP_F_DATAPAD bit (0x20)
     * is set.
     */
    s2 = NEW_STMT_LDB_ABS(cstate, 8);
    sappend(s, s2);
    sjset_radiotap_tsft_present->s.jf = s2;

    sjset_notsft_datapad = NEW_STMT_JSET_K(cstate, 0x20);
    sappend(s, sjset_notsft_datapad);

    /*
     * In either case, if IEEE80211_RADIOTAP_F_DATAPAD is
     * set, round the length of the 802.11 header to
     * a multiple of 4.  Do that by adding 3 and then
     * dividing by and multiplying by 4, which we do by
     * ANDing with ~3.
     */
    s_roundup = NEW_STMT_LD_M(cstate, cstate->off_linkpl.reg);
    sappend(s, s_roundup);
    s2 = NEW_STMT_ADD_K(cstate, 3);
    sappend(s, s2);
    s2 = NEW_STMT_AND_K(cstate, (bpf_u_int32)~3);
    sappend(s, s2);
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkpl.reg);
    sappend(s, s2);

    sjset_tsft_datapad->s.jt = s_roundup;
    sjset_tsft_datapad->s.jf = snext;
    sjset_notsft_datapad->s.jt = s_roundup;
    sjset_notsft_datapad->s.jf = snext;
  } else
    sjset_qos->s.jf = snext;

  return s;
}

static void
insert_compute_vloffsets(compiler_state_t *cstate, struct block *b)
{
  struct slist *s;

  /* There is an implicit dependency between the link
   * payload and link header since the payload computation
   * includes the variable part of the header. Therefore,
   * if nobody else has allocated a register for the link
   * header and we need it, do it now. */
  if (cstate->off_linkpl.reg != -1 && cstate->off_linkhdr.is_variable &&
      cstate->off_linkhdr.reg == -1)
    cstate->off_linkhdr.reg = alloc_reg(cstate);

  /*
   * For link-layer types that have a variable-length header
   * preceding the link-layer header, generate code to load
   * the offset of the link-layer header into the register
   * assigned to that offset, if any.
   *
   * XXX - this, and the next switch statement, won't handle
   * encapsulation of 802.11 or 802.11+radio information in
   * some other protocol stack.  That's significantly more
   * complicated.
   */
  switch (cstate->outermostlinktype) {

  case DLT_PRISM_HEADER:
    s = gen_load_prism_llprefixlen(cstate);
    break;

  case DLT_IEEE802_11_RADIO_AVS:
    s = gen_load_avs_llprefixlen(cstate);
    break;

  case DLT_IEEE802_11_RADIO:
    s = gen_load_radiotap_llprefixlen(cstate);
    break;

  case DLT_PPI:
    s = gen_load_ppi_llprefixlen(cstate);
    break;

  default:
    s = NULL;
    break;
  }

  /*
   * For link-layer types that have a variable-length link-layer
   * header, generate code to load the offset of the link-layer
   * payload into the register assigned to that offset, if any.
   */
  switch (cstate->outermostlinktype) {

  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
  case DLT_PPI:
    s = gen_load_802_11_header_len(cstate, s, b->stmts);
    /*
     * After this call s may have changed, b->stmts has not
     * changed, s and b->stmts have not merged into one linked
     * list, therefore the meaning of b, whether a Boolean constant
     * or not, has not changed.
     */
    break;

  case DLT_PFLOG:
    s = gen_load_pflog_llprefixlen(cstate);
    break;
  }

  /*
   * If there is no initialization yet and we need variable
   * length offsets for VLAN, initialize them to zero
   */
  if (s == NULL && cstate->is_vlan_vloffset) {
    struct slist *s2;

    if (cstate->off_linkpl.reg == -1)
      cstate->off_linkpl.reg = alloc_reg(cstate);
    if (cstate->off_linktype.reg == -1)
      cstate->off_linktype.reg = alloc_reg(cstate);

    s = NEW_STMT_LD_IMM(cstate, 0);
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linkpl.reg);
    sappend(s, s2);
    s2 = NEW_STMT_ST_M(cstate, cstate->off_linktype.reg);
    sappend(s, s2);
  }

  /*
   * If we have any offset-loading code, append all the
   * existing statements in the block to those statements,
   * and make the resulting list the list of statements
   * for the block.
   */
  sprepend_to_block(s, b);
}

/*
 * Take an absolute offset, and:
 *
 *    if it has no variable part, return NULL;
 *
 *    if it has a variable part, generate code to load the register
 *    containing that variable part into the X register, returning
 *    a pointer to that code - if no register for that offset has
 *    been allocated, allocate it first.
 *
 * (The code to set that register will be generated later, but will
 * be placed earlier in the code sequence.)
 */
static struct slist *
gen_abs_offset_varpart(compiler_state_t *cstate, bpf_abs_offset *off)
{
  if (off->is_variable) {
    if (off->reg == -1) {
      /*
       * We haven't yet assigned a register for the
       * variable part of the offset of the link-layer
       * header; allocate one.
       */
      off->reg = alloc_reg(cstate);
    }

    /*
     * Load the register containing the variable part of the
     * offset of the link-layer header into the X register.
     */
    return NEW_STMT_LDX_M(cstate, off->reg);
  } else {
    /*
     * That offset isn't variable, there's no variable part,
     * so we don't need to generate any code.
     */
    return NULL;
  }
}

/*
 * Map an Ethernet type to the equivalent PPP type.
 */
static uint16_t
ethertype_to_ppptype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  switch (ll_proto) {

  case ETHERTYPE_IP:
    return PPP_IP;

  case ETHERTYPE_IPV6:
    return PPP_IPV6;

  case ETHERTYPE_DN:
    return PPP_DECNET;

  case ETHERTYPE_ATALK:
    return PPP_APPLE;

  case ETHERTYPE_NS:
    return PPP_NS;

  case LLCSAP_ISONS:
    return PPP_OSI;

  case LLCSAP_8021D:
    /*
     * I'm assuming the "Bridging PDU"s that go
     * over PPP are Spanning Tree Protocol
     * Bridging PDUs.
     */
    return PPP_BRPDU;

  case LLCSAP_IPX:
    return PPP_IPX;
  }
  assert_maxval(cstate, "PPP protocol", ll_proto, UINT16_MAX);
  return (uint16_t)ll_proto;
}

/*
 * Generate any tests that, for encapsulation of a link-layer packet
 * inside another protocol stack, need to be done to check for those
 * link-layer packets (and that haven't already been done by a check
 * for that encapsulation).
 */
static struct block *
gen_prevlinkhdr_check(compiler_state_t *cstate)
{
  if (cstate->is_encap)
    return gen_encap_ll_check(cstate);

  switch (cstate->prevlinktype) {

  case DLT_SUNATM:
    /*
     * This is LANE-encapsulated Ethernet; check that the LANE
     * packet doesn't begin with an LE Control marker, i.e.
     * that it's data, not a control message.
     *
     * (We've already generated a test for LANE.)
     */
    return gen_cmp_ne(cstate, OR_PREVLINKHDR, SUNATM_PKT_BEGIN_POS, BPF_H, 0xFF00);

  default:
    /*
     * No such tests are necessary.
     */
    return NULL;
  }
  /*NOTREACHED*/
}

// Match the specified version number in the Internet Protocol header.
static struct block *
gen_ip_version(compiler_state_t *cstate, const enum e_offrel offrel,
    const uint8_t ver)
{
  switch (ver) {
  case 4:
  case 6:
    return gen_mcmp(cstate, offrel, 0, BPF_B, ver << 4, 0xf0);
  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "ver", ver);
  }
}

/*
 * Match a Frame Relay (ITU-T Rec. Q.922) header with the Control field set to
 * UI (Unnumbered information, 0x03, ibid., Table 3) and the NLPID field set to
 * the given value.
 *
 * This code assumes a Frame Relay header encoding that has the Control field
 * at offset 2 and the NLPID field at offset 3, which means no flags before the
 * Address field (thus not the RFC 2427 encoding) and exactly 2 bytes for the
 * Address field (thus the default, but not the only possible address format,
 * ibid., Table 1).
 */
static struct block *
gen_frelay_nlpid(compiler_state_t *cstate, const uint8_t nlpid)
{
  return gen_cmp(cstate, OR_LINKHDR, 2, BPF_H, (0x03 << 8) | nlpid);
}

/*
 * Generate code to match a particular packet type by matching the
 * link-layer type field or fields in the 802.2 LLC header.
 *
 * "proto" is an Ethernet type value, if > ETHERMTU, or an LLC SAP
 * value, if <= ETHERMTU.
 */
static struct block *
gen_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  struct block *b0, *b1, *b2;

  /* are we checking MPLS-encapsulated packets? */
  if (cstate->label_stack_depth > 0)
    return gen_mpls_linktype(cstate, ll_proto);

  switch (cstate->linktype) {

  case DLT_EN10MB:
  case DLT_NETANALYZER:
  case DLT_NETANALYZER_TRANSPARENT:
  case DLT_DSA_TAG_BRCM:
  case DLT_DSA_TAG_DSA:
    /* Geneve has an EtherType regardless of whether there is an
     * L2 header. VXLAN always has an EtherType. */
    if (!cstate->is_encap)
      b0 = gen_prevlinkhdr_check(cstate);
    else
      b0 = NULL;

    b1 = gen_ether_linktype(cstate, ll_proto);
    return b0 ? gen_and(b0, b1) : b1;
    /*NOTREACHED*/

  case DLT_C_HDLC:
  case DLT_HDLC:
    assert_maxval(cstate, "HDLC protocol", ll_proto, UINT16_MAX);
    switch (ll_proto) {

    case LLCSAP_ISONS:
      ll_proto = (ll_proto << 8 | LLCSAP_ISONS);
      /* fall through */

    default:
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);
      /*NOTREACHED*/
    }

  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
  case DLT_PPI:
    /*
     * Check that we have a data frame.
     */
    b0 = gen_mcmp(cstate, OR_LINKHDR, 0, BPF_B,
      IEEE80211_FC0_TYPE_DATA,
      IEEE80211_FC0_TYPE_MASK);

    /*
     * Now check for the specified link-layer type.
     */
    b1 = gen_llc_linktype(cstate, ll_proto);
    return gen_and(b0, b1);
    /*NOTREACHED*/

  case DLT_FDDI:
    /*
     * XXX - check for LLC frames.
     */
    return gen_llc_linktype(cstate, ll_proto);
    /*NOTREACHED*/

  case DLT_IEEE802:
    /*
     * XXX - check for LLC PDUs, as per IEEE 802.5.
     */
    return gen_llc_linktype(cstate, ll_proto);
    /*NOTREACHED*/

  case DLT_ATM_RFC1483:
  case DLT_ATM_CLIP:
  case DLT_IP_OVER_FC:
    return gen_llc_linktype(cstate, ll_proto);
    /*NOTREACHED*/

  case DLT_SUNATM:
    /*
     * Check for an LLC-encapsulated version of this protocol;
     * if we were checking for LANE, linktype would no longer
     * be DLT_SUNATM.
     *
     * Check for LLC encapsulation and then check the protocol.
     */
    b0 = gen_atm_prototype(cstate, PT_LLC);
    b1 = gen_llc_linktype(cstate, ll_proto);
    return gen_and(b0, b1);
    /*NOTREACHED*/

  case DLT_LINUX_SLL:
    return gen_linux_sll_linktype(cstate, ll_proto);
    /*NOTREACHED*/

  case DLT_SLIP:
  case DLT_SLIP_BSDOS:
  case DLT_RAW:
    /*
     * These types don't provide any type field; packets
     * are always IPv4 or IPv6.  Hence in this context the
     * to-be-confirmed IPv4/IPv6 header begins at the link-layer
     * header.
     */
    switch (ll_proto) {

    case ETHERTYPE_IP:
      return gen_ip_version(cstate, OR_LINKHDR, 4);

    case ETHERTYPE_IPV6:
      return gen_ip_version(cstate, OR_LINKHDR, 6);

    default:
      return gen_false(cstate); /* always false */
    }
    /*NOTREACHED*/

  case DLT_IPV4:
    /*
     * Raw IPv4, so no type field.
     */
    if (ll_proto == ETHERTYPE_IP)
      return gen_true(cstate); /* always true */

    /* Checking for something other than IPv4; always false */
    return gen_false(cstate);
    /*NOTREACHED*/

  case DLT_IPV6:
    /*
     * Raw IPv6, so no type field.
     */
    if (ll_proto == ETHERTYPE_IPV6)
      return gen_true(cstate); /* always true */

    /* Checking for something other than IPv6; always false */
    return gen_false(cstate);
    /*NOTREACHED*/

  case DLT_PPP:
  case DLT_PPP_PPPD:
  case DLT_PPP_SERIAL:
  case DLT_PPP_ETHER:
    /*
     * We use Ethernet protocol types inside libpcap;
     * map them to the corresponding PPP protocol types.
     */
    return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H,
        ethertype_to_ppptype(cstate, ll_proto));
    /*NOTREACHED*/

  case DLT_PPP_BSDOS:
    /*
     * We use Ethernet protocol types inside libpcap;
     * map them to the corresponding PPP protocol types.
     */
    switch (ll_proto) {

    case ETHERTYPE_IP:
      /*
       * Also check for Van Jacobson-compressed IP.
       * XXX - do this for other forms of PPP?
       */
      b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, PPP_IP);
      b1 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, PPP_VJC);
      b1 = gen_or(b0, b1);
      b0 = gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, PPP_VJNC);
      return gen_or(b1, b0);

    default:
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H,
          ethertype_to_ppptype(cstate, ll_proto));
    }
    /*NOTREACHED*/

  case DLT_NULL:
  case DLT_LOOP:
  case DLT_ENC:
    /*
     * 4-byte AF_ value at the beginning of the packet.
     */
    return (gen_loopback_linktype(cstate, ll_proto));

  case DLT_PFLOG:
    /*
     * 1-byte AF_ value.
     */
    return (gen_pflog_linktype(cstate, ll_proto));

  case DLT_ARCNET:
  case DLT_ARCNET_LINUX:
    /*
     * In ARCnet header the 8-bit SC (System Code) field identifies
     * the higher-level protocol in the INFO (Information) part of
     * the packet, same as the 16-bit EtherType > 1500 in Ethernet.
     * RFC 1051 (March 1988) allocated ARCTYPE_IP_OLD to IPv4 and
     * ARCTYPE_ARP_OLD to ARP, RFC 1201 (February 1991) allocated
     * ARCTYPE_IP to IPv4 and ARCTYPE_ARP to ARP.  ARCnet header
     * encoding and length differ between the two specifications.
     *
     * This DLT case previously matched IPv4 and ARP by ORing, for
     * backward compatibility reasons, respective SCs from RFC 1051
     * and RFC 1201.  This worked as expected when a filter program
     * tested SC to tell whether a packet is an IPv4/ARP packet,
     * but did not access INFO (where the IPv4 or ARP header is).
     *
     * However, for filter expressions that need to access INFO the
     * C code that processes IPv4/ARP header fields generates
     * exactly one match and uses the DLT's off_linkpl, which
     * init_linktype() initializes to RFC 1201 encoding, so
     * combining that with an RFC 1051 SC match produced incorrect
     * filter programs.  This is why this DLT case in the current
     * implementation matches RFC 1201 SCs only.
     *
     * XXX should we check for first fragment if the protocol
     * uses PHDS?
     */
    switch (ll_proto) {

    default:
      return gen_false(cstate);

    case ETHERTYPE_IPV6:
      return (gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B,
        ARCTYPE_INET6));

    case ETHERTYPE_IP:
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B,
          ARCTYPE_IP);

    case ETHERTYPE_ARP:
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B,
          ARCTYPE_ARP);

    case ETHERTYPE_REVARP:
      return (gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B,
          ARCTYPE_REVARP));

    case ETHERTYPE_ATALK:
      return (gen_cmp(cstate, OR_LINKTYPE, 0, BPF_B,
          ARCTYPE_ATALK));
    }
    /*NOTREACHED*/

  case DLT_LTALK:
    switch (ll_proto) {
    case ETHERTYPE_ATALK:
      return gen_true(cstate);
    default:
      return gen_false(cstate);
    }
    /*NOTREACHED*/

  case DLT_FRELAY:
    switch (ll_proto) {

    case ETHERTYPE_IP:
      return gen_frelay_nlpid(cstate, ISO9577_IPV4);

    case ETHERTYPE_IPV6:
      return gen_frelay_nlpid(cstate, ISO9577_IPV6);

    case LLCSAP_ISONS:
      /*
       * Check for several OSI protocols.
       *
       * Frame Relay packets typically have an OSI
       * NLPID at the beginning; we check for each
       * of them.
       */
      b0 = gen_frelay_nlpid(cstate, ISO8473_CLNP);
      b1 = gen_frelay_nlpid(cstate, ISO9542_ESIS);
      b2 = gen_frelay_nlpid(cstate, ISO10589_ISIS);
      b2 = gen_or(b1, b2);
      return gen_or(b0, b2);

    default:
      return gen_false(cstate);
    }
    /*NOTREACHED*/

  case DLT_MFR:
    break; // not implemented

  case DLT_JUNIPER_MFR:
  case DLT_JUNIPER_MLFR:
  case DLT_JUNIPER_MLPPP:
  case DLT_JUNIPER_ATM1:
  case DLT_JUNIPER_ATM2:
  case DLT_JUNIPER_PPPOE:
  case DLT_JUNIPER_PPPOE_ATM:
  case DLT_JUNIPER_GGSN:
  case DLT_JUNIPER_ES:
  case DLT_JUNIPER_MONITOR:
  case DLT_JUNIPER_SERVICES:
  case DLT_JUNIPER_ETHER:
  case DLT_JUNIPER_PPP:
  case DLT_JUNIPER_FRELAY:
  case DLT_JUNIPER_CHDLC:
  case DLT_JUNIPER_VP:
  case DLT_JUNIPER_ST:
  case DLT_JUNIPER_ISM:
  case DLT_JUNIPER_VS:
  case DLT_JUNIPER_SRX_E2E:
  case DLT_JUNIPER_FIBRECHANNEL:
  case DLT_JUNIPER_ATM_CEMIC:

    /* just lets verify the magic number for now -
     * on ATM we may have up to 6 different encapsulations on the wire
     * and need a lot of heuristics to figure out that the payload
     * might be;
     *
     * FIXME encapsulation specific BPF_ filters
     */
    return gen_mcmp(cstate, OR_LINKHDR, 0, BPF_W, 0x4d474300, 0xffffff00); /* compare the magic number */

  case DLT_IPNET:
    return gen_ipnet_linktype(cstate, ll_proto);

  default:
    /*
     * Does this link-layer header type have a field
     * indicating the type of the next protocol?  If
     * so, off_linktype.constant_part will be the offset of that
     * field in the packet; if not, it will be OFFSET_NOT_SET.
     */
    if (cstate->off_linktype.constant_part != OFFSET_NOT_SET) {
      /*
       * Yes; assume it's an Ethernet type.  (If
       * it's not, it needs to be handled specially
       * above.)
       */
      assert_maxval(cstate, "EtherType", ll_proto, UINT16_MAX);
      return gen_cmp(cstate, OR_LINKTYPE, 0, BPF_H, ll_proto);
      /*NOTREACHED */
    }
  }
  /*
   * For example, using the fixed-size NFLOG header it is possible
   * to tell only the address family of the packet, other meaningful
   * data is either missing or behind TLVs.
   */
  bpf_error(cstate, "link-layer protocol filtering not implemented for %s",
      pcapint_datalink_val_to_string(cstate->linktype));
}

/*
 * Check for an LLC SNAP packet with a given organization code and
 * protocol type; we check the entire contents of the 802.2 LLC and
 * snap headers, checking for DSAP and SSAP of SNAP and a control
 * field of 0x03 in the LLC header, and for the specified organization
 * code and protocol type in the SNAP header.
 */
static struct block *
gen_snap(compiler_state_t *cstate, bpf_u_int32 orgcode, bpf_u_int32 ptype)
{
  u_char snapblock[8];

  snapblock[0] = LLCSAP_SNAP;   /* DSAP = SNAP */
  snapblock[1] = LLCSAP_SNAP;   /* SSAP = SNAP */
  snapblock[2] = 0x03;      /* control = UI */
  snapblock[3] = (u_char)(orgcode >> 16); /* upper 8 bits of organization code */
  snapblock[4] = (u_char)(orgcode >> 8);  /* middle 8 bits of organization code */
  snapblock[5] = (u_char)(orgcode >> 0);  /* lower 8 bits of organization code */
  snapblock[6] = (u_char)(ptype >> 8);  /* upper 8 bits of protocol type */
  snapblock[7] = (u_char)(ptype >> 0);  /* lower 8 bits of protocol type */
  return gen_bcmp(cstate, OR_LLC, 0, 8, snapblock);
}

/*
 * Generate code to match frames with an LLC header.
 */
static struct block *
gen_llc_internal(compiler_state_t *cstate)
{
  struct block *b0, *b1;

  switch (cstate->linktype) {

  case DLT_EN10MB:
  case DLT_DSA_TAG_BRCM:
  case DLT_DSA_TAG_DSA:
    /*
     * We check for an Ethernet type field less or equal than
     * 1500, which means it's an 802.3 length field.
     */
    b0 = gen_cmp_le(cstate, OR_LINKTYPE, 0, BPF_H, ETHERMTU);

    /*
     * Now check for the purported DSAP and SSAP not being
     * 0xFF, to rule out NetWare-over-802.3.
     */
    b1 = gen_cmp_ne(cstate, OR_LLC, 0, BPF_H, 0xFFFF);

    return gen_and(b0, b1);

  case DLT_SUNATM:
    /*
     * We check for LLC traffic.
     */
    return gen_atmtype_llc(cstate);

  case DLT_IEEE802: /* Token Ring */
    /*
     * XXX - check for LLC frames.
     */
    return gen_true(cstate);

  case DLT_FDDI:
    /*
     * XXX - check for LLC frames.
     */
    return gen_true(cstate);

  case DLT_ATM_RFC1483:
    /*
     * For LLC encapsulation, these are defined to have an
     * 802.2 LLC header.
     *
     * For VC encapsulation, they don't, but there's no
     * way to check for that; the protocol used on the VC
     * is negotiated out of band.
     */
    return gen_true(cstate);

  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_PPI:
    /*
     * Check that we have a data frame.
     */
    return gen_mcmp(cstate, OR_LINKHDR, 0, BPF_B,
      IEEE80211_FC0_TYPE_DATA,
      IEEE80211_FC0_TYPE_MASK);

  default:
    fail_kw_on_dlt(cstate, "llc");
    /*NOTREACHED*/
  }
}

struct block *
gen_llc(compiler_state_t *cstate)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_llc_internal(cstate);
}

struct block *
gen_llc_i(compiler_state_t *cstate)
{
  struct block *b0, *b1;
  struct slist *s;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Check whether this is an LLC frame.
   */
  b0 = gen_llc_internal(cstate);

  /*
   * Load the control byte and test the low-order bit; it must
   * be clear for I frames.
   */
  s = gen_load_a(cstate, OR_LLC, 2, BPF_B);
  b1 = gen_unset(cstate, 0x01, s);

  return gen_and(b0, b1);
}

struct block *
gen_llc_s(compiler_state_t *cstate)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Check whether this is an LLC frame.
   */
  b0 = gen_llc_internal(cstate);

  /*
   * Now compare the low-order 2 bit of the control byte against
   * the appropriate value for S frames.
   */
  b1 = gen_mcmp(cstate, OR_LLC, 2, BPF_B, LLC_S_FMT, 0x03);

  return gen_and(b0, b1);
}

struct block *
gen_llc_u(compiler_state_t *cstate)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Check whether this is an LLC frame.
   */
  b0 = gen_llc_internal(cstate);

  /*
   * Now compare the low-order 2 bit of the control byte against
   * the appropriate value for U frames.
   */
  b1 = gen_mcmp(cstate, OR_LLC, 2, BPF_B, LLC_U_FMT, 0x03);

  return gen_and(b0, b1);
}

struct block *
gen_llc_s_subtype(compiler_state_t *cstate, bpf_u_int32 subtype)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Check whether this is an LLC frame.
   */
  b0 = gen_llc_internal(cstate);

  /*
   * Now check for an S frame with the appropriate type.
   */
  b1 = gen_mcmp(cstate, OR_LLC, 2, BPF_B, subtype, LLC_S_CMD_MASK);

  return gen_and(b0, b1);
}

struct block *
gen_llc_u_subtype(compiler_state_t *cstate, bpf_u_int32 subtype)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Check whether this is an LLC frame.
   */
  b0 = gen_llc_internal(cstate);

  /*
   * Now check for a U frame with the appropriate type.
   */
  b1 = gen_mcmp(cstate, OR_LLC, 2, BPF_B, subtype, LLC_U_CMD_MASK);

  return gen_and(b0, b1);
}

/*
 * Generate code to match a particular packet type, for link-layer types
 * using 802.2 LLC headers.
 *
 * This is *NOT* used for Ethernet; "gen_ether_linktype()" is used
 * for that - it handles the D/I/X Ethernet vs. 802.3+802.2 issues.
 *
 * "proto" is an Ethernet type value, if > ETHERMTU, or an LLC SAP
 * value, if <= ETHERMTU.  We use that to determine whether to
 * match the DSAP or both DSAP and LSAP or to check the OUI and
 * protocol ID in a SNAP header.
 */
static struct block *
gen_llc_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  /*
   * XXX - handle token-ring variable-length header.
   */
  switch (ll_proto) {

  case LLCSAP_IP:
  case LLCSAP_ISONS:
  case LLCSAP_NETBEUI:
    /*
     * XXX - should we check both the DSAP and the
     * SSAP, like this, or should we check just the
     * DSAP, as we do for other SAP values?
     */
    return gen_cmp(cstate, OR_LLC, 0, BPF_H, (bpf_u_int32)
           ((ll_proto << 8) | ll_proto));

  case LLCSAP_IPX:
    /*
     * XXX - are there ever SNAP frames for IPX on
     * non-Ethernet 802.x networks?
     */
    return gen_cmp(cstate, OR_LLC, 0, BPF_B, LLCSAP_IPX);

  case ETHERTYPE_ATALK:
    /*
     * 802.2-encapsulated ETHERTYPE_ATALK packets are
     * SNAP packets with an organization code of
     * 0x080007 (Apple, for Appletalk) and a protocol
     * type of ETHERTYPE_ATALK (Appletalk).
     *
     * XXX - check for an organization code of
     * encapsulated Ethernet as well?
     */
    return gen_snap(cstate, 0x080007, ETHERTYPE_ATALK);

  default:
    /*
     * XXX - we don't have to check for IPX 802.3
     * here, but should we check for the IPX EtherType?
     */
    if (ll_proto <= ETHERMTU) {
      assert_maxval(cstate, "LLC DSAP", ll_proto, UINT8_MAX);
      /*
       * This is an LLC SAP value, so check
       * the DSAP.
       */
      return gen_cmp(cstate, OR_LLC, 0, BPF_B, ll_proto);
    } else {
      assert_maxval(cstate, "EtherType", ll_proto, UINT16_MAX);
      /*
       * This is an Ethernet type; we assume that it's
       * unlikely that it'll appear in the right place
       * at random, and therefore check only the
       * location that would hold the Ethernet type
       * in a SNAP frame with an organization code of
       * 0x000000 (encapsulated Ethernet).
       *
       * XXX - if we were to check for the SNAP DSAP and
       * LSAP, as per XXX, and were also to check for an
       * organization code of 0x000000 (encapsulated
       * Ethernet), we'd do
       *
       *  return gen_snap(cstate, 0x000000, ll_proto);
       *
       * here; for now, we don't, as per the above.
       * I don't know whether it's worth the extra CPU
       * time to do the right check or not.
       */
      return gen_cmp(cstate, OR_LLC, 6, BPF_H, ll_proto);
    }
  }
}

static struct block *
gen_hostop(compiler_state_t *cstate, bpf_u_int32 addr, bpf_u_int32 mask,
    int dir, u_int src_off, u_int dst_off)
{
  struct block *b0, *b1;
  u_int offset;

  switch (dir) {

  case Q_SRC:
    offset = src_off;
    break;

  case Q_DST:
    offset = dst_off;
    break;

  case Q_AND:
    b0 = gen_hostop(cstate, addr, mask, Q_SRC, src_off, dst_off);
    b1 = gen_hostop(cstate, addr, mask, Q_DST, src_off, dst_off);
    return gen_and(b0, b1);

  case Q_DEFAULT:
  case Q_OR:
    b0 = gen_hostop(cstate, addr, mask, Q_SRC, src_off, dst_off);
    b1 = gen_hostop(cstate, addr, mask, Q_DST, src_off, dst_off);
    return gen_or(b0, b1);

  default:
    // Bug: a WLAN dqual should have been rejected earlier.
    bpf_error(cstate, ERRSTR_FUNC_VAR_STR, __func__, "dir", dqkw(dir));
    /*NOTREACHED*/
  }
  return gen_mcmp(cstate, OR_LINKPL, offset, BPF_W, addr, mask);
}

static struct block *
gen_hostop6(compiler_state_t *cstate, const struct in6_addr *addr,
    const struct in6_addr *mask, const u_char dir)
{
  struct block *b0, *b1;
  u_int offset;
  /*
   * Code below needs to access four separate 32-bit parts of the 128-bit
   * IPv6 address and mask.  In some OSes this is as simple as using the
   * s6_addr32 pseudo-member of struct in6_addr, which contains a union of
   * 8-, 16- and 32-bit arrays.  In other OSes this is not the case, as
   * far as libpcap sees it.  Hence copy the data before use to avoid
   * potential unaligned memory access and the associated compiler
   * warnings (whether genuine or not).
   */
  bpf_u_int32 a[4], m[4];

  switch (dir) {

  case Q_SRC:
    offset = IPV6_SRCADDR_OFFSET;
    break;

  case Q_DST:
    offset = IPV6_DSTADDR_OFFSET;
    break;

  case Q_AND:
    b0 = gen_hostop6(cstate, addr, mask, Q_SRC);
    b1 = gen_hostop6(cstate, addr, mask, Q_DST);
    return gen_and(b0, b1);

  case Q_DEFAULT:
  case Q_OR:
    b0 = gen_hostop6(cstate, addr, mask, Q_SRC);
    b1 = gen_hostop6(cstate, addr, mask, Q_DST);
    return gen_or(b0, b1);

  default:
    // Bug: a WLAN dqual should have been rejected earlier.
    bpf_error(cstate, ERRSTR_FUNC_VAR_STR, __func__, "dir", dqkw(dir));
    /*NOTREACHED*/
  }
  /* this order is important */
  memcpy(a, addr, sizeof(a));
  memcpy(m, mask, sizeof(m));
  b1 = gen_true(cstate);
  for (int i = 3; i >= 0; i--) {
    b0 = gen_mcmp(cstate, OR_LINKPL, offset + 4 * i, BPF_W,
        ntohl(a[i]), ntohl(m[i]));
    b1 = gen_and(b0, b1);
  }
  return b1;
}

/*
 * Like gen_mac48host(), but for DLT_IEEE802_11 (802.11 wireless LAN) and
 * various 802.11 + radio headers.
 */
static struct block *
gen_wlanhostop(compiler_state_t *cstate, const u_char *eaddr, int dir)
{
  struct block *b0, *b1, *b2;
  struct slist *s;

#ifdef ENABLE_WLAN_FILTERING_PATCH
  /*
   * TODO GV 20070613
   * We need to disable the optimizer because the optimizer is buggy
   * and wipes out some LD instructions generated by the below
   * code to validate the Frame Control bits
   */
  cstate->no_optimize = 1;
#endif /* ENABLE_WLAN_FILTERING_PATCH */

  switch (dir) {
  case Q_SRC:
    /*
     * Oh, yuk.
     *
     *  For control frames, there is no SA.
     *
     *  For management frames, SA is at an
     *  offset of 10 from the beginning of
     *  the packet.
     *
     *  For data frames, SA is at an offset
     *  of 10 from the beginning of the packet
     *  if From DS is clear, at an offset of
     *  16 from the beginning of the packet
     *  if From DS is set and To DS is clear,
     *  and an offset of 24 from the beginning
     *  of the packet if From DS is set and To DS
     *  is set.
     */

    /*
     * Generate the tests to be done for data frames
     * with From DS set.
     *
     * First, check for To DS set, i.e. check "link[1] & 0x01".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC1_DIR_TODS, s);

    /*
     * If To DS is set, the SA is at 24.
     */
    b0 = gen_bcmp(cstate, OR_LINKHDR, 24, 6, eaddr);
    b0 = gen_and(b1, b0);

    /*
     * Now, check for To DS not set, i.e. check
     * "!(link[1] & 0x01)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b2 = gen_unset(cstate, IEEE80211_FC1_DIR_TODS, s);

    /*
     * If To DS is not set, the SA is at 16.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 16, 6, eaddr);
    b1 = gen_and(b2, b1);

    /*
     * Now OR together the last two checks.  That gives
     * the complete set of checks for data frames with
     * From DS set.
     */
    b0 = gen_or(b1, b0);

    /*
     * Now check for From DS being set, and AND that with
     * the ORed-together checks.
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC1_DIR_FROMDS, s);
    b0 = gen_and(b1, b0);

    /*
     * Now check for data frames with From DS not set.
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b2 = gen_unset(cstate, IEEE80211_FC1_DIR_FROMDS, s);

    /*
     * If From DS isn't set, the SA is at 10.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 10, 6, eaddr);
    b1 = gen_and(b2, b1);

    /*
     * Now OR together the checks for data frames with
     * From DS not set and for data frames with From DS
     * set; that gives the checks done for data frames.
     */
    b0 = gen_or(b1, b0);

    /*
     * Now check for a data frame.
     * I.e, check "link[0] & 0x08".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * AND that with the checks done for data frames.
     */
    b0 = gen_and(b1, b0);

    /*
     * If the high-order bit of the type value is 0, this
     * is a management frame.
     * I.e, check "!(link[0] & 0x08)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b2 = gen_unset(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * For management frames, the SA is at 10.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 10, 6, eaddr);
    b1 = gen_and(b2, b1);

    /*
     * OR that with the checks done for data frames.
     * That gives the checks done for management and
     * data frames.
     */
    b0 = gen_or(b1, b0);

    /*
     * If the low-order bit of the type value is 1,
     * this is either a control frame or a frame
     * with a reserved type, and thus not a
     * frame with an SA.
     *
     * I.e., check "!(link[0] & 0x04)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_unset(cstate, IEEE80211_FC0_TYPE_CTL, s);

    /*
     * AND that with the checks for data and management
     * frames.
     */
    return gen_and(b1, b0);

  case Q_DST:
    /*
     * Oh, yuk.
     *
     *  For control frames, there is no DA.
     *
     *  For management frames, DA is at an
     *  offset of 4 from the beginning of
     *  the packet.
     *
     *  For data frames, DA is at an offset
     *  of 4 from the beginning of the packet
     *  if To DS is clear and at an offset of
     *  16 from the beginning of the packet
     *  if To DS is set.
     */

    /*
     * Generate the tests to be done for data frames.
     *
     * First, check for To DS set, i.e. "link[1] & 0x01".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC1_DIR_TODS, s);

    /*
     * If To DS is set, the DA is at 16.
     */
    b0 = gen_bcmp(cstate, OR_LINKHDR, 16, 6, eaddr);
    b0 = gen_and(b1, b0);

    /*
     * Now, check for To DS not set, i.e. check
     * "!(link[1] & 0x01)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
    b2 = gen_unset(cstate, IEEE80211_FC1_DIR_TODS, s);

    /*
     * If To DS is not set, the DA is at 4.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 4, 6, eaddr);
    b1 = gen_and(b2, b1);

    /*
     * Now OR together the last two checks.  That gives
     * the complete set of checks for data frames.
     */
    b0 = gen_or(b1, b0);

    /*
     * Now check for a data frame.
     * I.e, check "link[0] & 0x08".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * AND that with the checks done for data frames.
     */
    b0 = gen_and(b1, b0);

    /*
     * If the high-order bit of the type value is 0, this
     * is a management frame.
     * I.e, check "!(link[0] & 0x08)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b2 = gen_unset(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * For management frames, the DA is at 4.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 4, 6, eaddr);
    b1 = gen_and(b2, b1);

    /*
     * OR that with the checks done for data frames.
     * That gives the checks done for management and
     * data frames.
     */
    b0 = gen_or(b1, b0);

    /*
     * If the low-order bit of the type value is 1,
     * this is either a control frame or a frame
     * with a reserved type, and thus not a
     * frame with an SA.
     *
     * I.e., check "!(link[0] & 0x04)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_unset(cstate, IEEE80211_FC0_TYPE_CTL, s);

    /*
     * AND that with the checks for data and management
     * frames.
     */
    return gen_and(b1, b0);

  case Q_AND:
    b0 = gen_wlanhostop(cstate, eaddr, Q_SRC);
    b1 = gen_wlanhostop(cstate, eaddr, Q_DST);
    return gen_and(b0, b1);

  case Q_DEFAULT:
  case Q_OR:
    b0 = gen_wlanhostop(cstate, eaddr, Q_SRC);
    b1 = gen_wlanhostop(cstate, eaddr, Q_DST);
    return gen_or(b0, b1);

  /*
   * XXX - add BSSID keyword?
   */
  case Q_ADDR1:
    return (gen_bcmp(cstate, OR_LINKHDR, 4, 6, eaddr));

  case Q_ADDR2:
    /*
     * Not present in CTS or ACK control frames.
     */
    b0 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_TYPE_CTL,
      IEEE80211_FC0_TYPE_MASK);
    b1 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_SUBTYPE_CTS,
      IEEE80211_FC0_SUBTYPE_MASK);
    b2 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_SUBTYPE_ACK,
      IEEE80211_FC0_SUBTYPE_MASK);
    b2 = gen_and(b1, b2);
    b2 = gen_or(b0, b2);
    b1 = gen_bcmp(cstate, OR_LINKHDR, 10, 6, eaddr);
    return gen_and(b2, b1);

  case Q_ADDR3:
    /*
     * Not present in control frames.
     */
    b0 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_TYPE_CTL,
      IEEE80211_FC0_TYPE_MASK);
    b1 = gen_bcmp(cstate, OR_LINKHDR, 16, 6, eaddr);
    return gen_and(b0, b1);

  case Q_ADDR4:
    /*
     * Present only if the direction mask has both "From DS"
     * and "To DS" set.  Neither control frames nor management
     * frames should have both of those set, so we don't
     * check the frame type.
     */
    b0 = gen_mcmp(cstate, OR_LINKHDR, 1, BPF_B,
      IEEE80211_FC1_DIR_DSTODS, IEEE80211_FC1_DIR_MASK);
    b1 = gen_bcmp(cstate, OR_LINKHDR, 24, 6, eaddr);
    return gen_and(b0, b1);

  case Q_RA:
    /*
     * Not present in management frames; addr1 in other
     * frames.
     */

    /*
     * If the high-order bit of the type value is 0, this
     * is a management frame.
     * I.e, check "(link[0] & 0x08)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * Check addr1.
     */
    b0 = gen_bcmp(cstate, OR_LINKHDR, 4, 6, eaddr);

    /*
     * AND that with the check of addr1.
     */
    return gen_and(b1, b0);

  case Q_TA:
    /*
     * Not present in management frames; addr2, if present,
     * in other frames.
     */

    /*
     * Not present in CTS or ACK control frames.
     */
    b0 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_TYPE_CTL,
      IEEE80211_FC0_TYPE_MASK);
    b1 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_SUBTYPE_CTS,
      IEEE80211_FC0_SUBTYPE_MASK);
    b2 = gen_mcmp_ne(cstate, OR_LINKHDR, 0, BPF_B, IEEE80211_FC0_SUBTYPE_ACK,
      IEEE80211_FC0_SUBTYPE_MASK);
    b2 = gen_and(b1, b2);
    b2 = gen_or(b0, b2);

    /*
     * If the high-order bit of the type value is 0, this
     * is a management frame.
     * I.e, check "(link[0] & 0x08)".
     */
    s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
    b1 = gen_set(cstate, IEEE80211_FC0_TYPE_DATA, s);

    /*
     * AND that with the check for frames other than
     * CTS and ACK frames.
     */
    b2 = gen_and(b1, b2);

    /*
     * Check addr2.
     */
    b1 = gen_bcmp(cstate, OR_LINKHDR, 10, 6, eaddr);
    return gen_and(b2, b1);
  }
  bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "dir", dir);
  /*NOTREACHED*/
}

/*
 * This is quite tricky because there may be pad bytes in front of the
 * DECNET header, and then there are two possible data packet formats that
 * carry both src and dst addresses, plus 5 packet types in a format that
 * carries only the src node, plus 2 types that use a different format and
 * also carry just the src node.
 *
 * Yuck.
 *
 * Instead of doing those all right, we just look for data packets with
 * 0 or 1 bytes of padding.  If you want to look at other packets, that
 * will require a lot more hacking.
 *
 * To add support for filtering on DECNET "areas" (network numbers)
 * one would want to add a "mask" argument to this routine.  That would
 * make the filter even more inefficient, although one could be clever
 * and not generate masking instructions if the mask is 0xFFFF.
 */
static struct block *
gen_dnhostop(compiler_state_t *cstate, bpf_u_int32 addr, int dir)
{
  struct block *b0, *b1, *b2, *tmp;
  u_int offset_lh;  /* offset if long header is received */
  u_int offset_sh;  /* offset if short header is received */

  switch (dir) {

  case Q_DST:
    offset_sh = 1;  /* follows flags */
    offset_lh = 7;  /* flgs,darea,dsubarea,HIORD */
    break;

  case Q_SRC:
    offset_sh = 3;  /* follows flags, dstnode */
    offset_lh = 15; /* flgs,darea,dsubarea,did,sarea,ssub,HIORD */
    break;

  case Q_AND:
    /* Inefficient because we do our Calvinball dance twice */
    b0 = gen_dnhostop(cstate, addr, Q_SRC);
    b1 = gen_dnhostop(cstate, addr, Q_DST);
    return gen_and(b0, b1);

  case Q_DEFAULT:
  case Q_OR:
    /* Inefficient because we do our Calvinball dance twice */
    b0 = gen_dnhostop(cstate, addr, Q_SRC);
    b1 = gen_dnhostop(cstate, addr, Q_DST);
    return gen_or(b0, b1);

  default:
    // Bug: a WLAN dqual should have been rejected earlier.
    bpf_error(cstate, ERRSTR_FUNC_VAR_STR, __func__, "dir", dqkw(dir));
    /*NOTREACHED*/
  }
  /*
   * In a DECnet message inside an Ethernet frame the first two bytes
   * immediately after EtherType are the [little-endian] DECnet message
   * length, which is irrelevant in this context.
   *
   * "pad = 1" means the third byte equals 0x81, thus it is the PLENGTH
   * 8-bit bitmap of the optional padding before the packet route header.
   * The bitmap always has bit 7 set to 1 and in this case has bits 0-6
   * (TOTAL-PAD-SEQUENCE-LENGTH) set to integer value 1.  The latter
   * means there aren't any PAD bytes after the bitmap, so the header
   * begins at the fourth byte.  "pad = 0" means bit 7 of the third byte
   * is set to 0, thus the header begins at the third byte.
   *
   * The header can be in several (as mentioned above) formats, all of
   * which begin with the FLAGS 8-bit bitmap, which always has bit 7
   * (PF, "pad field") set to 0 regardless of any padding present before
   * the header.  "Short header" means bits 0-2 of the bitmap encode the
   * integer value 2 (SFDP), and "long header" means value 6 (LFDP).
   *
   * To test PLENGTH and FLAGS, use multiple-byte constants with the
   * values and the masks, this maps to the required single bytes of
   * the message correctly on both big-endian and little-endian hosts.
   * For the DECnet address use PCAP_BSWAP_16(), which always swaps bytes,
   * because the wire encoding is little-endian and BPF multiple-byte
   * loads are big-endian.  When the destination address is near enough
   * to PLENGTH and FLAGS, generate one 32-bit comparison instead of two
   * smaller ones.
   */
  /* Check for pad = 1, long header case */
  tmp = gen_mcmp(cstate, OR_LINKPL, 2, BPF_H, 0x8106U, 0xFF07U);
  b1 = gen_cmp(cstate, OR_LINKPL, 2 + 1 + offset_lh,
      BPF_H, PCAP_BSWAP_16(addr));
  b1 = gen_and(tmp, b1);
  /* Check for pad = 0, long header case */
  tmp = gen_mcmp(cstate, OR_LINKPL, 2, BPF_B, 0x06U, 0x07U);
  b2 = gen_cmp(cstate, OR_LINKPL, 2 + offset_lh, BPF_H,
      PCAP_BSWAP_16(addr));
  b2 = gen_and(tmp, b2);
  b1 = gen_or(b2, b1);
  /* Check for pad = 1, short header case */
  if (dir == Q_DST) {
    b2 = gen_mcmp(cstate, OR_LINKPL, 2, BPF_W,
        0x81020000U | PCAP_BSWAP_16(addr),
        0xFF07FFFFU);
  } else {
    tmp = gen_mcmp(cstate, OR_LINKPL, 2, BPF_H, 0x8102U, 0xFF07U);
    b2 = gen_cmp(cstate, OR_LINKPL, 2 + 1 + offset_sh, BPF_H,
        PCAP_BSWAP_16(addr));
    b2 = gen_and(tmp, b2);
  }
  b1 = gen_or(b2, b1);
  /* Check for pad = 0, short header case */
  if (dir == Q_DST) {
    b2 = gen_mcmp(cstate, OR_LINKPL, 2, BPF_W,
        0x02000000U | PCAP_BSWAP_16(addr) << 8,
        0x07FFFF00U);
  } else {
    tmp = gen_mcmp(cstate, OR_LINKPL, 2, BPF_B, 0x02U, 0x07U);
    b2 = gen_cmp(cstate, OR_LINKPL, 2 + offset_sh, BPF_H,
        PCAP_BSWAP_16(addr));
    b2 = gen_and(tmp, b2);
  }

  return gen_or(b2, b1);
}

/*
 * Assume the link-layer payload data just before off_nl (L3) is an MPLS label
 * (L2.5) and test whether the label has Bottom of Stack bit set.
 */
static struct block *
gen_just_after_mpls_stack(compiler_state_t *cstate)
{
  return gen_set(cstate, 0x01, gen_load_a(cstate, OR_PREVMPLSHDR, 2, BPF_B));
}

/*
 * Generate a check for IPv4 or IPv6 for MPLS-encapsulated packets;
 * test the bottom-of-stack bit, and then check the version number
 * field in the IP header.
 */
static struct block *
gen_mpls_linktype(compiler_state_t *cstate, bpf_u_int32 ll_proto)
{
  struct block *b0, *b1;

  /*
   * In this context the to-be-confirmed IPv4/IPv6 header begins at the
   * link-layer payload.
   */
  switch (ll_proto) {

  case ETHERTYPE_IP:
    b0 = gen_just_after_mpls_stack(cstate);
    b1 = gen_ip_version(cstate, OR_LINKPL, 4);
    return gen_and(b0, b1);

  case ETHERTYPE_IPV6:
    b0 = gen_just_after_mpls_stack(cstate);
    b1 = gen_ip_version(cstate, OR_LINKPL, 6);
    return gen_and(b0, b1);

  default:
    /* FIXME add other L3 proto IDs */
    bpf_error(cstate, "unsupported protocol over mpls");
    /*NOTREACHED*/
  }
}

static struct block *
gen_host(compiler_state_t *cstate, const size_t n, const bpf_u_int32 *a,
    const bpf_u_int32 *m, const u_char proto, const u_char dir,
    const u_char not, const char *context)
{
  /*
   * WLAN direction qualifiers are never valid for IPv4 addresses.
   *
   * It is important to validate this now because the call to
   * gen_hostop() may be optimized out below.
   */
  assert_nonwlan_dqual(cstate, dir);

  struct block *b0, *b1;
  bpf_u_int32 llproto;
  u_int src_off, dst_off;

  switch (proto) {

  case Q_DEFAULT:
    b0 = gen_host(cstate, n, a, m, Q_IP, dir, not, context);
    /*
     * Only check for non-IPv4 addresses if we're not
     * checking MPLS-encapsulated packets.
     */
    if (cstate->label_stack_depth == 0) {
      b1 = gen_host(cstate, n, a, m, Q_ARP, dir, not, context);
      b1 = gen_or(b0, b1);
      b0 = gen_host(cstate, n, a, m, Q_RARP, dir, not, context);
      b0 = gen_or(b1, b0);
    }
    return b0;

  case Q_IP:
    llproto = ETHERTYPE_IP;
    src_off = IPV4_SRCADDR_OFFSET;
    dst_off = IPV4_DSTADDR_OFFSET;
    break;

  case Q_RARP:
    llproto = ETHERTYPE_REVARP;
    src_off = RARP_SRCADDR_OFFSET;
    dst_off = RARP_DSTADDR_OFFSET;
    break;

  case Q_ARP:
    llproto = ETHERTYPE_ARP;
    src_off = ARP_SRCADDR_OFFSET;
    dst_off = ARP_DSTADDR_OFFSET;
    break;

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), context);
  }
  b0 = gen_linktype(cstate, llproto);
  if (b0->meaning == IS_FALSE) {
    /*
     * If this DLT does not support ARP or RARP, the result of
     * gen_linktype() is a Boolean false, then the subsequent
     * gen_and() would discard the result of gen_hostop() and
     * return the Boolean false.
     *
     * However, if this DLT also uses a variable-length link-layer
     * header (which means DLT_PFLOG only at the time of this
     * writing), a side effect of the gen_hostop() invocation
     * would be registering a demand for a variable-length offset
     * preamble, which a Boolean constant never needs, so in this
     * case return early and have one fewer reasons to produce the
     * preamble in insert_compute_vloffsets().
     */
    return b0;
  }
  b1 = gen_false(cstate);
  for (size_t i = 0; i < n; i++)
    b1 = gen_or(b1,
        gen_hostop(cstate, a[i], m[i], dir, src_off, dst_off));
  return gen_and(b0, not ? gen_not(b1) : b1);
}

static struct block *
gen_host6(compiler_state_t *cstate, const size_t n,
    const struct in6_addr *a, const struct in6_addr *m,
    const u_char proto, const u_char dir, const u_char not,
    const char *context)
{
  // WLAN direction qualifiers are never valid for IPv6 addresses.
  assert_nonwlan_dqual(cstate, dir);

  if (proto != Q_DEFAULT && proto != Q_IPV6)
    bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), context);

  struct block *linkproto = gen_linktype(cstate, ETHERTYPE_IPV6);
  struct block *host = gen_false(cstate);
  for (size_t i = 0; i < n; i++)
    host = gen_or(host, gen_hostop6(cstate, a + i, m + i, dir));
  return gen_and(linkproto, not ? gen_not(host) : host);
}

static int
uint32_t_cmp(const void *a, const void *b)
{
  /*
   * Host byte order.  One potential way to do the comparison would be
   * to return "a32 - b32", but that would require to prove -- to both
   * humans and C compilers -- that for all possible [uint32_t] values
   * of a32 and b32 the difference would always map to a correct sign of
   * the [int] return value on all architectures, so let's instead do it
   * in a way that is obviously correct.
   */
  const uint32_t a32 = *((uint32_t *)a), b32 = *((uint32_t *)b);
  return a32 < b32 ? -1 :
      a32 > b32 ? 1 :
      0;
}

static int
in6_addr_cmp(const void *a, const void *b)
{
  // Network byte order is straightforward.
  return memcmp(a, b, sizeof(struct in6_addr));
}

/*
 * The maximum supported number of resolved addresses per family (IPv4/IPv6)
 * for a given Internet hostname.
 */
#define MAX_PER_AF 100

static struct block *
gen_host46_byname(compiler_state_t *cstate, const char *name,
    const u_char proto4, const u_char proto6, const u_char dir,
    const u_char not)
{
  /*
   * Both gen_host() and gen_host6() require a context argument to
   * generate an error message if the proto qualifier is invalid.  The
   * only two invocations of this function are from gen_gateway() and
   * gen_scode().  Because the former validates pqual first, the only
   * possible context here is from the latter, so there is no sense in
   * using a function argument for what effectively is a constant.
   */
  static const char *context = "host <Internet hostname>";

  if ((cstate->ai = pcap_nametoaddrinfo(name)) == NULL)
    bpf_error(cstate, "unknown host '%s'", name);
  struct block *ret = NULL;

  /*
   * For a hostname that resolves to both IPv4 and IPv6 addresses the
   * AF_INET addresses may come before or after the AF_INET6 addresses
   * depending on which getaddrinfo() implementation it is, what the
   * resolving host's network configuration is and (on Linux with glibc)
   * the contents of gai.conf(5).  This is because getaddrinfo() presumes
   * a subsequent bind(2) or connect(2) use of the addresses, which is
   * not the case here, so there is no sense in preserving the order of
   * the AFs in the resolved addresses.  However, there is sense in
   * hard-coding the order of AFs when generating a match block for more
   * than one AF because this way the result reflects fewer external
   * effects and is easier to test.
   */

  /*
   * Ignore any IPv4 addresses when resolving "ip6 host NAME", validate
   * all other proto qualifiers in gen_host().
   */
  if (proto4 != Q_IPV6) {
    uint32_t addrs[MAX_PER_AF], masks[MAX_PER_AF];
    size_t count = 0;
    for (struct addrinfo *ai = cstate->ai; ai; ai = ai->ai_next) {
      if (ai->ai_family != AF_INET)
        continue;
      if (count == MAX_PER_AF)
        bpf_error(cstate,
                  "More than %u IPv4 addresses per name",
                  MAX_PER_AF);
      struct sockaddr_in *sin4 =
          (struct sockaddr_in *)ai->ai_addr;
      addrs[count] = ntohl(sin4->sin_addr.s_addr);
      masks[count] = 0xffffffff;
      count++;
    }
    if (count > 1)
      qsort(addrs, count, sizeof(*addrs), uint32_t_cmp);
    if (count)
      ret = gen_host(cstate, count, addrs, masks, proto4,
                     dir, not, context);
  }

  /*
   * Ignore any IPv6 addresses when resolving "(arp|ip|rarp) host NAME",
   * validate all other proto qualifiers in gen_host6().
   */
  static const struct in6_addr mask128 = { .s6_addr = {
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
  }};
  if (proto6 != Q_ARP && proto6 != Q_IP && proto6 != Q_RARP) {
    struct in6_addr addrs[MAX_PER_AF], masks[MAX_PER_AF];
    size_t count = 0;
    for (struct addrinfo *ai = cstate->ai; ai; ai = ai->ai_next) {
      if (ai->ai_family != AF_INET6)
        continue;
      if (count == MAX_PER_AF)
        bpf_error(cstate,
                  "More than %u IPv6 addresses per name",
                  MAX_PER_AF);
      struct sockaddr_in6 *sin6 =
          (struct sockaddr_in6 *)ai->ai_addr;
      addrs[count] = sin6->sin6_addr;
      masks[count] = mask128;
      count++;
    }
    if (count > 1)
      qsort(addrs, count, sizeof(*addrs), in6_addr_cmp);
    if (count) {
      struct block *hosts6 =
          gen_host6(cstate, count, addrs, masks, proto6, dir,
                    not, context);
      ret = ret ? gen_or(ret, hosts6) : hosts6;
    }
  }

  freeaddrinfo(cstate->ai);
  cstate->ai = NULL;

  if (! ret)
    bpf_error(cstate, "unknown host '%s'%s", name,
        proto4 == Q_DEFAULT
        ? ""
        : " for specified address family");
  return ret;
}

#undef MAX_PER_AF

static struct block *
gen_dnhost(compiler_state_t *cstate, const char *s, bpf_u_int32 v,
    const struct qual q)
{
  // WLAN direction qualifiers are never valid for DECnet addresses.
  assert_nonwlan_dqual(cstate, q.dir);

  /*
   * libpcap defines exactly one primitive that has "decnet" as
   * the protocol qualifier: "decnet host AREANUMBER.NODENUMBER".
   */
  if (q.addr != Q_DEFAULT && q.addr != Q_HOST)
    bpf_error(cstate, ERRSTR_INVALID_QUAL, "decnet", tqkw(q.addr));

  if (s == NULL) {
    /*
     * v contains a 32-bit unsigned parsed from a string of the
     * form {N}, which could be decimal, hexadecimal or octal.
     * Although it would be possible to use the value as a raw
     * 16-bit DECnet address when the value fits into 16 bits,
     * this would be a questionable feature: DECnet address wire
     * encoding is little-endian, so this would not work as
     * intuitively as the same works for [big-endian] IPv4
     * addresses (0x01020304 means 1.2.3.4).
     */
    bpf_error(cstate, "invalid DECnet address '%u'", v);
  }

  /*
   * s points to a string of the form {N}.{N}, {N}.{N}.{N} or
   * {N}.{N}.{N}.{N}, of which only the first potentially stands
   * for a valid DECnet address.
   */
  uint16_t addr;
  if (! pcapint_atodn(s, &addr))
    bpf_error(cstate, "invalid DECnet address '%s'", s);

  struct block *b0 = gen_linktype(cstate, ETHERTYPE_DN);
  struct block *b1 = gen_dnhostop(cstate, addr, q.dir);
  return gen_and(b0, b1);
}

static unsigned char
is_mac48_linktype(const int linktype)
{
  switch (linktype) {
  case DLT_EN10MB:
  case DLT_FDDI:
  case DLT_IEEE802:
  case DLT_IEEE802_11:
  case DLT_IEEE802_11_RADIO:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IP_OVER_FC:
  case DLT_NETANALYZER:
  case DLT_NETANALYZER_TRANSPARENT:
  case DLT_DSA_TAG_BRCM:
  case DLT_DSA_TAG_DSA:
  case DLT_PPI:
  case DLT_PRISM_HEADER:
    return 1;
  default:
    return 0;
  }
}

static struct block *
gen_mac48host(compiler_state_t *cstate, const u_char *eaddr, const u_char dir,
    const char *keyword)
{
  struct block *b1 = NULL;
  u_int src_off, dst_off;

  /*
   * Do not validate dir yet and let gen_wlanhostop() handle the DLTs
   * that support WLAN direction qualifiers.
   */
  switch (cstate->linktype) {
  case DLT_EN10MB:
  case DLT_NETANALYZER:
  case DLT_NETANALYZER_TRANSPARENT:
  case DLT_DSA_TAG_BRCM:
  case DLT_DSA_TAG_DSA:
    b1 = gen_prevlinkhdr_check(cstate);
    src_off = 6;
    dst_off = 0;
    break;
  case DLT_FDDI:
    src_off = 6 + 1 + cstate->pcap_fddipad;
    dst_off = 0 + 1 + cstate->pcap_fddipad;
    break;
  case DLT_IEEE802:
    src_off = 8;
    dst_off = 2;
    break;
  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
  case DLT_PPI:
    return gen_wlanhostop(cstate, eaddr, dir);
  case DLT_IP_OVER_FC:
    /*
     * Assume that the addresses are IEEE 48-bit MAC addresses,
     * as RFC 2625 states.
     */
    src_off = 10;
    dst_off = 2;
    break;
  case DLT_SUNATM:
    /*
     * This is LLC-multiplexed traffic; if it were
     * LANE, cstate->linktype would have been set to
     * DLT_EN10MB.
     */
     /* FALLTHROUGH */
  default:
    fail_kw_on_dlt(cstate, keyword);
  }
  // Now validate.
  assert_nonwlan_dqual(cstate, dir);

  struct block *b0, *tmp;

  switch (dir) {
  case Q_SRC:
    b0 = gen_bcmp(cstate, OR_LINKHDR, src_off, 6, eaddr);
    break;
  case Q_DST:
    b0 = gen_bcmp(cstate, OR_LINKHDR, dst_off, 6, eaddr);
    break;
  case Q_AND:
    tmp = gen_bcmp(cstate, OR_LINKHDR, src_off, 6, eaddr);
    b0 = gen_bcmp(cstate, OR_LINKHDR, dst_off, 6, eaddr);
    b0 = gen_and(tmp, b0);
    break;
  case Q_DEFAULT:
  case Q_OR:
    tmp = gen_bcmp(cstate, OR_LINKHDR, src_off, 6, eaddr);
    b0 = gen_bcmp(cstate, OR_LINKHDR, dst_off, 6, eaddr);
    b0 = gen_or(tmp, b0);
    break;
  default:
    // Bug: a WLAN dqual should have been rejected earlier.
    bpf_error(cstate, ERRSTR_FUNC_VAR_STR, __func__, "dir", dqkw(dir));
  }

  return b1 ? gen_and(b1, b0) : b0;
}

static struct block *
gen_mac48host_byname(compiler_state_t *cstate, const char *name,
    const u_char dir, const char *context)
{
  if (! is_mac48_linktype(cstate->linktype))
    fail_kw_on_dlt(cstate, context);

  u_char *eaddrp = pcap_ether_hostton(name);
  if (eaddrp == NULL)
    bpf_error(cstate, ERRSTR_UNKNOWN_MAC48HOST, name);
  u_char eaddr[6];
  memcpy(eaddr, eaddrp, sizeof(eaddr));
  free(eaddrp);

  return gen_mac48host(cstate, eaddr, dir, context);
}

static struct block *
gen_mac8host(compiler_state_t *cstate, const uint8_t mac8, const u_char dir,
    const char *context)
{
  u_int src_off, dst_off;

  switch (cstate->linktype) {
  case DLT_ARCNET:
  case DLT_ARCNET_LINUX:
    /*
     * ARCnet is different from Ethernet: the source address comes
     * before the destination address, each is one byte long.
     * This holds for all three "buffer formats" in RFC 1201
     * Section 2.1, see also page 4-10 in the 1983 edition of the
     * "ARCNET Designer's Handbook" published by Datapoint
     * (document number 61610-01).
     */
    src_off = 0;
    dst_off = 1;
    break;
  case DLT_BACNET_MS_TP:
    /*
     * MS/TP resembles both Ethernet (in that the destination
     * station address precedes the source station address) and
     * ARCnet (in that a station address is one byte long).
     */
    src_off = 4;
    dst_off = 3;
    break;
  default:
    fail_kw_on_dlt(cstate, context);
  }

  struct block *src, *dst;

  switch (dir) {
  case Q_SRC:
    return gen_cmp(cstate, OR_LINKHDR, src_off, BPF_B, mac8);
  case Q_DST:
    return gen_cmp(cstate, OR_LINKHDR, dst_off, BPF_B, mac8);
  case Q_AND:
    src = gen_cmp(cstate, OR_LINKHDR, src_off, BPF_B, mac8);
    dst = gen_cmp(cstate, OR_LINKHDR, dst_off, BPF_B, mac8);
    return gen_and(src, dst);
  case Q_DEFAULT:
  case Q_OR:
    src = gen_cmp(cstate, OR_LINKHDR, src_off, BPF_B, mac8);
    dst = gen_cmp(cstate, OR_LINKHDR, dst_off, BPF_B, mac8);
    return gen_or(src, dst);
  default:
    // Bug: a WLAN dqual should have been rejected earlier.
    bpf_error(cstate, ERRSTR_FUNC_VAR_STR, __func__, "dir", dqkw(dir));
  }
}

/*
 * This primitive is non-directional by design, so the grammar does not allow
 * to qualify it with a direction.
 */
static struct block *
gen_gateway(compiler_state_t *cstate, const char *name, const u_char proto)
{
  switch (proto) {
  case Q_DEFAULT:
  case Q_IP:
  case Q_ARP:
  case Q_RARP:
    break;
  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), "gateway");
  }
  if (cstate->label_stack_depth)
    bpf_error(cstate, "'gateway' cannot be used within MPLS");
  if (cstate->is_encap)
    bpf_error(cstate, "'gateway' cannot be used within VXLAN or Geneve");

  struct block *b0 = gen_mac48host_byname(cstate, name, Q_OR, "gateway");
  /*
   * For "gateway NAME" not qualified with a protocol skip the IPv6 leg
   * of the name-to-address translation to match the documented
   * IPv4-only behaviour.
   */
  struct block *b1 = gen_host46_byname(cstate, name, proto, Q_IP, Q_OR, 1);
  return gen_and(b0, b1);
}

static struct block *
gen_proto_abbrev_internal(compiler_state_t *cstate, int proto)
{
  struct block *b0;
  struct block *b1;

  switch (proto) {

  case Q_SCTP:
  case Q_TCP:
  case Q_UDP:
  case Q_AH:
  case Q_ESP:
  case Q_PIM:
    // protocols based on IPv4/IPv6
    return gen_proto(cstate,
        pq_to_ipproto(cstate, (uint8_t)proto), Q_DEFAULT);

  case Q_ICMP:
  case Q_IGMP:
  case Q_IGRP:
  case Q_VRRP:
  case Q_CARP:
    // protocols based on IPv4 only
    return gen_proto(cstate,
        pq_to_ipproto(cstate, (uint8_t)proto), Q_IP);

  case Q_ICMPV6:
    // protocols based on IPv6 only
    return gen_proto(cstate,
        pq_to_ipproto(cstate, (uint8_t)proto), Q_IPV6);

  case Q_IP:
  case Q_ARP:
  case Q_RARP:
  case Q_ATALK:
  case Q_AARP:
  case Q_DECNET:
  case Q_SCA:
  case Q_LAT:
  case Q_MOPDL:
  case Q_MOPRC:
  case Q_IPV6:
    // link-layer protocols not based on LLC
    return gen_linktype(cstate,
        pq_to_ethertype(cstate, (uint8_t)proto));

  case Q_ISO:
  case Q_STP:
  case Q_IPX:
  case Q_NETBEUI:
    // link-layer protocols based on LLC
    return gen_linktype(cstate,
        pq_to_llcsap(cstate, (uint8_t)proto));

  case Q_ESIS:
  case Q_ISIS:
  case Q_CLNP:
    // ISO protocols
    return gen_proto(cstate,
        pq_to_nlpid(cstate, (uint8_t)proto), Q_ISO);

  case Q_ISIS_L1: /* all IS-IS Level1 PDU-Types */
    b0 = gen_proto(cstate, ISIS_L1_LAN_IIH, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_PTP_IIH, Q_ISIS); /* FIXME extract the circuit-type bits */
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L1_LSP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L1_CSNP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L1_PSNP, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_L2: /* all IS-IS Level2 PDU-Types */
    b0 = gen_proto(cstate, ISIS_L2_LAN_IIH, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_PTP_IIH, Q_ISIS); /* FIXME extract the circuit-type bits */
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L2_LSP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L2_CSNP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L2_PSNP, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_IIH: /* all IS-IS Hello PDU-Types */
    b0 = gen_proto(cstate, ISIS_L1_LAN_IIH, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_L2_LAN_IIH, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_PTP_IIH, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_LSP:
    b0 = gen_proto(cstate, ISIS_L1_LSP, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_L2_LSP, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_SNP:
    b0 = gen_proto(cstate, ISIS_L1_CSNP, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_L2_CSNP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L1_PSNP, Q_ISIS);
    b1 = gen_or(b0, b1);
    b0 = gen_proto(cstate, ISIS_L2_PSNP, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_CSNP:
    b0 = gen_proto(cstate, ISIS_L1_CSNP, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_L2_CSNP, Q_ISIS);
    return gen_or(b0, b1);

  case Q_ISIS_PSNP:
    b0 = gen_proto(cstate, ISIS_L1_PSNP, Q_ISIS);
    b1 = gen_proto(cstate, ISIS_L2_PSNP, Q_ISIS);
    return gen_or(b0, b1);
  }
  bpf_error(cstate, "'%s' cannot be used as an abbreviation", pqkw(proto));
}

struct block *
gen_proto_abbrev(compiler_state_t *cstate, int proto)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_proto_abbrev_internal(cstate, proto);
}

static struct block *
gen_ip_proto(compiler_state_t *cstate, const uint8_t proto)
{
  return gen_cmp(cstate, OR_LINKPL, IPV4_PROTO_OFFSET, BPF_B, proto);
}

static struct block *
gen_ip6_proto(compiler_state_t *cstate, const uint8_t proto)
{
  return gen_cmp(cstate, OR_LINKPL, IPV6_PROTO_OFFSET, BPF_B, proto);
}

static struct block *
gen_ipfrag(compiler_state_t *cstate)
{
  struct slist *s;

  /* not IPv4 frag other than the first frag */
  s = gen_load_a(cstate, OR_LINKPL, 6, BPF_H);
  return gen_unset(cstate, 0x1fff, s);
}

/*
 * Generate a comparison to a port value in the transport-layer header
 * at the specified offset from the beginning of that header.
 *
 * XXX - this handles a variable-length prefix preceding the link-layer
 * header, such as the radiotap or AVS radio prefix, but doesn't handle
 * variable-length link-layer headers (such as Token Ring or 802.11
 * headers).
 */
static struct block *
gen_portatom(compiler_state_t *cstate, int off, uint16_t v)
{
  return gen_cmp(cstate, OR_TRAN_IPV4, off, BPF_H, v);
}

static struct block *
gen_portatom6(compiler_state_t *cstate, int off, uint16_t v)
{
  return gen_cmp(cstate, OR_TRAN_IPV6, off, BPF_H, v);
}

static struct block *
gen_port(compiler_state_t *cstate, const uint16_t port, const int proto,
    const u_char dir, const u_char addr)
{
  struct block *b1, *tmp;

  switch (dir) {
  case Q_SRC:
    b1 = gen_portatom(cstate, TRAN_SRCPORT_OFFSET, port);
    break;

  case Q_DST:
    b1 = gen_portatom(cstate, TRAN_DSTPORT_OFFSET, port);
    break;

  case Q_AND:
    tmp = gen_portatom(cstate, TRAN_SRCPORT_OFFSET, port);
    b1 = gen_portatom(cstate, TRAN_DSTPORT_OFFSET, port);
    b1 = gen_and(tmp, b1);
    break;

  case Q_DEFAULT:
  case Q_OR:
    tmp = gen_portatom(cstate, TRAN_SRCPORT_OFFSET, port);
    b1 = gen_portatom(cstate, TRAN_DSTPORT_OFFSET, port);
    b1 = gen_or(tmp, b1);
    break;

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, dqkw(dir), tqkw(addr));
    /*NOTREACHED*/
  }

  return gen_port_common(cstate, proto, b1);
}

static struct block *
gen_port_common(compiler_state_t *cstate, int proto, struct block *b1)
{
  struct block *b0, *tmp;

  /*
   * ether proto ip
   *
   * For FDDI, RFC 1188 says that SNAP encapsulation is used,
   * not LLC encapsulation with LLCSAP_IP.
   *
   * For IEEE 802 networks - which includes 802.5 token ring
   * (which is what DLT_IEEE802 means) and 802.11 - RFC 1042
   * says that SNAP encapsulation is used, not LLC encapsulation
   * with LLCSAP_IP.
   *
   * For LLC-encapsulated ATM/"Classical IP", RFC 1483 and
   * RFC 2225 say that SNAP encapsulation is used, not LLC
   * encapsulation with LLCSAP_IP.
   *
   * So we always check for ETHERTYPE_IP.
   *
   * At the time of this writing all three L4 protocols the "port" and
   * "portrange" primitives support (TCP, UDP and SCTP) have the source
   * and the destination ports identically encoded in the transport
   * protocol header.  So without a proto qualifier the only difference
   * between the implemented cases is the protocol number and all other
   * checks need to be made exactly once.
   *
   * If the expression syntax in future starts to support ports for
   * another L4 protocol that has unsigned integer ports encoded using a
   * different size and/or offset, this will require a different code.
   */
  switch (proto) {
  case IPPROTO_UDP:
  case IPPROTO_TCP:
  case IPPROTO_SCTP:
    tmp = gen_ip_proto(cstate, (uint8_t)proto);
    break;

  case PROTO_UNDEF:
    tmp = gen_ip_proto(cstate, IPPROTO_SCTP);
    tmp = gen_or(gen_ip_proto(cstate, IPPROTO_UDP), tmp);
    tmp = gen_or(gen_ip_proto(cstate, IPPROTO_TCP), tmp);
    break;

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "proto", proto);
  }
  // Not a fragment other than the first fragment.
  b0 = gen_ipfrag(cstate);
  b0 = gen_and(tmp, b0);
  b1 = gen_and(b0, b1);
  // "link proto \ip"
  return gen_and(gen_linktype(cstate, ETHERTYPE_IP), b1);
}

static struct block *
gen_port6(compiler_state_t *cstate, const uint16_t port, const int proto,
    const u_char dir, const u_char addr)
{
  struct block *b1, *tmp;

  switch (dir) {
  case Q_SRC:
    b1 = gen_portatom6(cstate, TRAN_SRCPORT_OFFSET, port);
    break;

  case Q_DST:
    b1 = gen_portatom6(cstate, TRAN_DSTPORT_OFFSET, port);
    break;

  case Q_AND:
    tmp = gen_portatom6(cstate, TRAN_SRCPORT_OFFSET, port);
    b1 = gen_portatom6(cstate, TRAN_DSTPORT_OFFSET, port);
    b1 = gen_and(tmp, b1);
    break;

  case Q_DEFAULT:
  case Q_OR:
    tmp = gen_portatom6(cstate, TRAN_SRCPORT_OFFSET, port);
    b1 = gen_portatom6(cstate, TRAN_DSTPORT_OFFSET, port);
    b1 = gen_or(tmp, b1);
    break;

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, dqkw(dir), tqkw(addr));
    /*NOTREACHED*/
  }

  return gen_port6_common(cstate, proto, b1);
}

static struct block *
gen_port6_common(compiler_state_t *cstate, int proto, struct block *b1)
{
  struct block *tmp;

  // "ip6 proto 'ip_proto'"
  switch (proto) {
  case IPPROTO_UDP:
  case IPPROTO_TCP:
  case IPPROTO_SCTP:
    tmp = gen_ip6_proto(cstate, (uint8_t)proto);
    break;

  case PROTO_UNDEF:
    // Same as in gen_port_common().
    tmp = gen_ip6_proto(cstate, IPPROTO_SCTP);
    tmp = gen_or(gen_ip6_proto(cstate, IPPROTO_UDP), tmp);
    tmp = gen_or(gen_ip6_proto(cstate, IPPROTO_TCP), tmp);
    break;

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "proto", proto);
  }
  // XXX - catch the first fragment of a fragmented packet?
  b1 = gen_and(tmp, b1);
  // "link proto \ip6"
  return gen_and(gen_linktype(cstate, ETHERTYPE_IPV6), b1);
}

/* gen_portrange code */
static struct block *
gen_portrangeatom(compiler_state_t *cstate, u_int off, uint16_t v1,
    uint16_t v2)
{
  if (v1 == v2)
    return gen_portatom(cstate, off, v1);

  struct block *b1, *b2;

  b1 = gen_cmp_ge(cstate, OR_TRAN_IPV4, off, BPF_H, min(v1, v2));
  b2 = gen_cmp_le(cstate, OR_TRAN_IPV4, off, BPF_H, max(v1, v2));

  return gen_and(b1, b2);
}

static struct block *
gen_portrange(compiler_state_t *cstate, uint16_t port1, uint16_t port2,
    int proto, int dir)
{
  struct block *b1, *tmp;

  switch (dir) {
  case Q_SRC:
    b1 = gen_portrangeatom(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    break;

  case Q_DST:
    b1 = gen_portrangeatom(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    break;

  case Q_AND:
    tmp = gen_portrangeatom(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    b1 = gen_portrangeatom(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    b1 = gen_and(tmp, b1);
    break;

  case Q_DEFAULT:
  case Q_OR:
    tmp = gen_portrangeatom(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    b1 = gen_portrangeatom(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    b1 = gen_or(tmp, b1);
    break;

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, dqkw(dir), "portrange");
    /*NOTREACHED*/
  }

  return gen_port_common(cstate, proto, b1);
}

static struct block *
gen_portrangeatom6(compiler_state_t *cstate, u_int off, uint16_t v1,
    uint16_t v2)
{
  if (v1 == v2)
    return gen_portatom6(cstate, off, v1);

  struct block *b1, *b2;

  b1 = gen_cmp_ge(cstate, OR_TRAN_IPV6, off, BPF_H, min(v1, v2));
  b2 = gen_cmp_le(cstate, OR_TRAN_IPV6, off, BPF_H, max(v1, v2));

  return gen_and(b1, b2);
}

static struct block *
gen_portrange6(compiler_state_t *cstate, uint16_t port1, uint16_t port2,
    int proto, int dir)
{
  struct block *b1, *tmp;

  switch (dir) {
  case Q_SRC:
    b1 = gen_portrangeatom6(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    break;

  case Q_DST:
    b1 = gen_portrangeatom6(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    break;

  case Q_AND:
    tmp = gen_portrangeatom6(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    b1 = gen_portrangeatom6(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    b1 = gen_and(tmp, b1);
    break;

  case Q_DEFAULT:
  case Q_OR:
    tmp = gen_portrangeatom6(cstate, TRAN_SRCPORT_OFFSET, port1, port2);
    b1 = gen_portrangeatom6(cstate, TRAN_DSTPORT_OFFSET, port1, port2);
    b1 = gen_or(tmp, b1);
    break;

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, dqkw(dir), "portrange");
    /*NOTREACHED*/
  }

  return gen_port6_common(cstate, proto, b1);
}

static int
lookup_proto(compiler_state_t *cstate, const char *name, const struct qual q)
{
  /*
   * Do not check here whether q.proto is valid (e.g. in "udp proto abc"
   * fail the "abc", but not the "udp proto").  Likewise, do not check
   * here whether the combination of q.proto and q.addr is valid (e.g.
   * in "(link|iso|isis) protochain abc" fail the "abc", but not the
   * "(link|iso|isis) protochain").
   *
   * On the one hand, this avoids a layering violation: gen_proto() and
   * gen_protochain() implement the semantic checks.  On the other hand,
   * the protocol name lookup error arguably is a problem smaller than
   * the semantic error, hence the latter ought to be the reported cause
   * of failure in both cases.  In future this potentially could be made
   * more consistent by attempting the lookup after the semantic checks.
   */

  int v = PROTO_UNDEF;
  switch (q.proto) {

  case Q_DEFAULT:
  case Q_IP:
  case Q_IPV6:
    v = pcap_nametoproto(name);
    break;

  case Q_LINK:
    /* XXX should look up h/w protocol type based on cstate->linktype */
    v = pcap_nametoeproto(name);
    if (v == PROTO_UNDEF)
      v = pcap_nametollc(name);
    break;

  case Q_ISO:
    if (strcmp(name, "esis") == 0)
      v = ISO9542_ESIS;
    else if (strcmp(name, "isis") == 0)
      v = ISO10589_ISIS;
    else if (strcmp(name, "clnp") == 0)
      v = ISO8473_CLNP;
    break;

  // "isis proto" is a valid syntax, but it takes only numeric IDs.
  }
  // In theory, the only possible negative value of v is PROTO_UNDEF.
  if (v >= 0)
    return v;

  if (q.proto == Q_DEFAULT)
    bpf_error(cstate, "unknown '%s' value '%s'",
        tqkw(q.addr), name);
  bpf_error(cstate, "unknown '%s %s' value '%s'",
      pqkw(q.proto), tqkw(q.addr), name);
}

#if !defined(NO_PROTOCHAIN)
/*
 * This primitive is non-directional by design, so the grammar does not allow
 * to qualify it with a direction.
 */
static struct block *
gen_protochain(compiler_state_t *cstate, bpf_u_int32 v, int proto)
{
  struct block *b0, *b;
  struct slist *s[100];

  memset(s, 0, sizeof(s));

  switch (proto) {
  case Q_IP:
  case Q_IPV6:
    assert_maxval(cstate, "protocol number", v, UINT8_MAX);
    break;
  case Q_DEFAULT:
    b0 = gen_protochain(cstate, v, Q_IP);
    b = gen_protochain(cstate, v, Q_IPV6);
    return gen_or(b0, b);
  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), "protochain");
    /*NOTREACHED*/
  }

  /*
   * We don't handle variable-length prefixes before the link-layer
   * header, or variable-length link-layer headers, here yet.
   * We might want to add BPF instructions to do the protochain
   * work, to simplify that and, on platforms that have a BPF
   * interpreter with the new instructions, let the filtering
   * be done in the kernel.  (We already require a modified BPF
   * engine to do the protochain stuff, to support backward
   * branches, and backward branch support is unlikely to appear
   * in kernel BPF engines.)
   *
   * Hence in the current implementation the gen_abs_offset_varpart()
   * invocations incurred from gen_load_a() and gen_loadx_iphdrlen()
   * below do not affect the offset because off_linkpl.is_variable == 0.
   */
  if (cstate->off_linkpl.is_variable)
    bpf_error(cstate, "'protochain' not supported with variable length headers");

  /*
   * To quote a comment in optimize.c:
   *
   * "These data structures are used in a Cocke and Schwartz style
   * value numbering scheme.  Since the flowgraph is acyclic,
   * exit values can be propagated from a node's predecessors
   * provided it is uniquely defined."
   *
   * "Acyclic" means "no backward branches", which means "no
   * loops", so we have to turn the optimizer off.
   */
  cstate->no_optimize = 1;

  /*
   * s[0] is a dummy entry to protect other BPF insn from damage
   * by s[fix] = foo with uninitialized variable "fix".  It is somewhat
   * hard to find interdependency made by jump table fixup.
   */
  unsigned i = 0;
  s[i] = NEW_STMT_LD_IMM(cstate, 0);
  i++;

  if (proto == Q_IP) {
    b0 = gen_linktype(cstate, ETHERTYPE_IP);

    /* A = ip->ip_p */
    s[i] = gen_load_a(cstate, OR_LINKPL, IPV4_PROTO_OFFSET, BPF_B);
    i++;
    /* X = ip->ip_hl << 2 */
    s[i] = gen_loadx_iphdrlen(cstate);
    i++;
  } else {
    b0 = gen_linktype(cstate, ETHERTYPE_IPV6);

    /* A = ip6->ip_nxt */
    s[i] = gen_load_a(cstate, OR_LINKPL, IPV6_PROTO_OFFSET, BPF_B);
    i++;
    /* X = sizeof(struct ip6_hdr) */
    s[i] = NEW_STMT_LDX_IMM(cstate, IP6_HDRLEN);
    i++;
  }

  /* again: if (A == v) goto end; else fall through; */
  unsigned again = i;
  s[i] = NEW_STMT_JEQ_K(cstate, v);
  s[i]->s.jt = NULL;   /*later*/
  s[i]->s.jf = NULL;   /*update in next stmt*/
  unsigned fix5 = i;
  i++;

  /* if (A == IPPROTO_NONE) goto end */
  s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_NONE);
  s[i]->s.jt = NULL; /*later*/
  s[i]->s.jf = NULL; /*update in next stmt*/
  s[fix5]->s.jf = s[i];
  unsigned fix2 = i;
  i++;

  int reg2 = alloc_reg(cstate);
  unsigned fix3 = 0;
  if (proto == Q_IPV6) {
    unsigned v6start = i;

    /* if (A == IPPROTO_HOPOPTS) goto v6advance */
    s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_HOPOPTS);
    s[i]->s.jt = NULL; /*later*/
    s[i]->s.jf = NULL; /*update in next stmt*/
    s[fix2]->s.jf = s[i];
    i++;
    /* if (A == IPPROTO_DSTOPTS) goto v6advance */
    s[i - 1]->s.jf = s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_DSTOPTS);
    s[i]->s.jt = NULL; /*later*/
    s[i]->s.jf = NULL; /*update in next stmt*/
    i++;
    /* if (A == IPPROTO_ROUTING) goto v6advance */
    s[i - 1]->s.jf = s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_ROUTING);
    s[i]->s.jt = NULL; /*later*/
    s[i]->s.jf = NULL; /*update in next stmt*/
    i++;
    /* if (A == IPPROTO_FRAGMENT) goto v6advance; else goto ahcheck; */
    s[i - 1]->s.jf = s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_FRAGMENT);
    s[i]->s.jt = NULL; /*later*/
    s[i]->s.jf = NULL; /*later*/
    fix3 = i;
    unsigned v6end = i;
    i++;

    /* v6advance: */
    unsigned v6advance = i;

    /*
     * in short,
     * A = P[X + packet head];
     * X = X + (P[X + packet head + 1] + 1) * 8;
     */
    /* A = P[X + packet head] */
    s[i] = NEW_STMT_LDB_IND(cstate,
        cstate->off_linkpl.constant_part + cstate->off_nl);
    i++;
    /* MEM[reg2] = A */
    s[i] = NEW_STMT_ST_M(cstate, reg2);
    i++;
    /* A = P[X + packet head + 1]; */
    s[i] = NEW_STMT_LDB_IND(cstate,
        cstate->off_linkpl.constant_part + cstate->off_nl + 1);
    i++;
    /* A += 1 */
    s[i] = NEW_STMT_ADD_K(cstate, 1);
    i++;
    /* A *= 8 */
    s[i] = NEW_STMT_MUL_K(cstate, 8);
    i++;
    /* A += X */
    s[i] = NEW_STMT_ADD_X(cstate);
    i++;
    /* X = A; */
    s[i] = NEW_STMT_TAX(cstate);
    i++;
    /* A = MEM[reg2] */
    s[i] = NEW_STMT_LD_M(cstate, reg2);
    i++;

    /* goto again; (must use BPF_JA for backward jump) */
    s[i] = NEW_STMT_JA(cstate, again - i - 1);
    s[i - 1]->s.jf = s[i];
    i++;

    /* fixup */
    for (unsigned j = v6start; j <= v6end; j++)
      s[j]->s.jt = s[v6advance];
  } else {
    /* nop */
    s[i] = NEW_STMT_ADD_K(cstate, 0);
    s[fix2]->s.jf = s[i];
    i++;
  }

  /* ahcheck: */
  unsigned ahcheck = i;
  /* if (A == IPPROTO_AH) then fall through; else goto end; */
  s[i] = NEW_STMT_JEQ_K(cstate, IPPROTO_AH);
  s[i]->s.jt = NULL; /*later*/
  s[i]->s.jf = NULL; /*later*/
  if (fix3)
    s[fix3]->s.jf = s[ahcheck];
  unsigned fix4 = i;
  i++;

  /*
   * in short,
   * A = P[X];
   * X = X + (P[X + 1] + 2) * 4;
   */
  /* A = P[X + packet head]; */
  s[i] = NEW_STMT_LDB_IND(cstate,
      cstate->off_linkpl.constant_part + cstate->off_nl);
  s[i - 1]->s.jt = s[i];
  i++;
  /* MEM[reg2] = A */
  s[i] = NEW_STMT_ST_M(cstate, reg2);
  i++;
  /* A = X */
  s[i - 1]->s.jt = s[i] = NEW_STMT_TXA(cstate);
  i++;
  /* A += 1 */
  s[i] = NEW_STMT_ADD_K(cstate, 1);
  i++;
  /* X = A */
  s[i] = NEW_STMT_TAX(cstate);
  i++;
  /* A = P[X + packet head] */
  s[i] = NEW_STMT_LDB_IND(cstate,
      cstate->off_linkpl.constant_part + cstate->off_nl);
  i++;
  /* A += 2 */
  s[i] = NEW_STMT_ADD_K(cstate, 2);
  i++;
  /* A *= 4 */
  s[i] = NEW_STMT_MUL_K(cstate, 4);
  i++;
  /* X = A; */
  s[i] = NEW_STMT_TAX(cstate);
  i++;
  /* A = MEM[reg2] */
  s[i] = NEW_STMT_LD_M(cstate, reg2);
  i++;

  /* goto again; (must use BPF_JA for backward jump) */
  s[i] = NEW_STMT_JA(cstate, again - i - 1);
  i++;

  /* end: nop */
  s[i] = NEW_STMT_ADD_K(cstate, 0);
  s[fix2]->s.jt = s[i];
  s[fix4]->s.jf = s[i];
  s[fix5]->s.jt = s[i];
  i++;

  /*
   * make slist chain
   */
  for (unsigned j = 0; j < i - 1; j++)
    s[j]->next = s[j + 1];
  s[i - 1]->next = NULL;

  /*
   * emit final check
   * Remember, s[0] is dummy.
   */
  b = gen_jmp_k(cstate, BPF_JEQ, v, s[1]);

  free_reg(cstate, reg2);

  return gen_and(b0, b);
}
#endif /* !defined(NO_PROTOCHAIN) */

/*
 * Generate code that checks whether the packet is a packet for protocol
 * <proto> and whether the type field in that protocol's header has
 * the value <v>, e.g. if <proto> is Q_IP, it checks whether it's an
 * IP packet and checks the protocol number in the IP header against <v>.
 *
 * If <proto> is Q_DEFAULT, i.e. just "proto" was specified, it checks
 * against Q_IP and Q_IPV6.
 *
 * This primitive is non-directional by design, so the grammar does not allow
 * to qualify it with a direction.
 */
static struct block *
gen_proto(compiler_state_t *cstate, bpf_u_int32 v, int proto)
{
  struct block *b0, *b1;
  struct block *b2;

  switch (proto) {
  case Q_DEFAULT:
    b0 = gen_proto(cstate, v, Q_IP);
    b1 = gen_proto(cstate, v, Q_IPV6);
    return gen_or(b0, b1);

  case Q_LINK:
    return gen_linktype(cstate, v);

  case Q_IP:
    assert_maxval(cstate, "protocol number", v, UINT8_MAX);
    /*
     * For FDDI, RFC 1188 says that SNAP encapsulation is used,
     * not LLC encapsulation with LLCSAP_IP.
     *
     * For IEEE 802 networks - which includes 802.5 token ring
     * (which is what DLT_IEEE802 means) and 802.11 - RFC 1042
     * says that SNAP encapsulation is used, not LLC encapsulation
     * with LLCSAP_IP.
     *
     * For LLC-encapsulated ATM/"Classical IP", RFC 1483 and
     * RFC 2225 say that SNAP encapsulation is used, not LLC
     * encapsulation with LLCSAP_IP.
     *
     * So we always check for ETHERTYPE_IP.
     */
    b0 = gen_linktype(cstate, ETHERTYPE_IP);
    // 0 <= v <= UINT8_MAX
    b1 = gen_ip_proto(cstate, (uint8_t)v);
    return gen_and(b0, b1);

  case Q_IPV6:
    assert_maxval(cstate, "protocol number", v, UINT8_MAX);
    b0 = gen_linktype(cstate, ETHERTYPE_IPV6);
    /*
     * Also check for a fragment header before the final
     * header.
     */
    b2 = gen_ip6_proto(cstate, IPPROTO_FRAGMENT);
    b1 = gen_cmp(cstate, OR_LINKPL, IP6_HDRLEN, BPF_B, v);
    b1 = gen_and(b2, b1);
    // 0 <= v <= UINT8_MAX
    b2 = gen_ip6_proto(cstate, (uint8_t)v);
    b1 = gen_or(b2, b1);
    return gen_and(b0, b1);

  case Q_ISO:
    assert_maxval(cstate, "ISO protocol", v, UINT8_MAX);
    switch (cstate->linktype) {

    case DLT_FRELAY:
      /*
       * Frame Relay packets typically have an OSI
       * NLPID at the beginning; "gen_linktype(cstate, LLCSAP_ISONS)"
       * generates code to check for all the OSI
       * NLPIDs, so calling it and then adding a check
       * for the particular NLPID for which we're
       * looking is bogus, as we can just check for
       * the NLPID.
       *
       * XXX - what about SNAP-encapsulated frames?
       */
      return gen_frelay_nlpid(cstate, (uint8_t)v);
      /*NOTREACHED*/

    case DLT_C_HDLC:
    case DLT_HDLC:
      /*
       * Cisco uses an EtherType lookalike - for OSI,
       * it's 0xfefe.
       */
      b0 = gen_linktype(cstate, LLCSAP_ISONS<<8 | LLCSAP_ISONS);
      /* OSI in C-HDLC is stuffed with a fudge byte */
      b1 = gen_cmp(cstate, OR_LINKPL_NOSNAP, 1, BPF_B, v);
      return gen_and(b0, b1);

    default:
      b0 = gen_linktype(cstate, LLCSAP_ISONS);
      b1 = gen_cmp(cstate, OR_LINKPL_NOSNAP, 0, BPF_B, v);
      return gen_and(b0, b1);
    }

  case Q_ISIS:
    assert_maxval(cstate, "IS-IS PDU type", v, ISIS_PDU_TYPE_MAX);
    b0 = gen_proto(cstate, ISO10589_ISIS, Q_ISO);
    /*
     * 4 is the offset of the PDU type relative to the IS-IS
     * header.
     * Except when it is not, see above.
     */
    unsigned pdu_type_offset;
    switch (cstate->linktype) {
    case DLT_C_HDLC:
    case DLT_HDLC:
      pdu_type_offset = 5;
      break;
    default:
      pdu_type_offset = 4;
    }
    b1 = gen_mcmp(cstate, OR_LINKPL_NOSNAP, pdu_type_offset, BPF_B,
        v, ISIS_PDU_TYPE_MAX);
    return gen_and(b0, b1);
  }
  bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), "proto");
  /*NOTREACHED*/
}

/*
 * Convert a non-numeric name to a port number.
 */
static int
nametoport(compiler_state_t *cstate, const char *name, int ipproto)
{
  struct addrinfo hints, *res, *ai;
  int error;
  struct sockaddr_in *in4;
  struct sockaddr_in6 *in6;
  int port = -1;

  /*
   * We check for both TCP and UDP in case there are
   * ambiguous entries.
   */
  memset(&hints, 0, sizeof(hints));
  hints.ai_family = PF_UNSPEC;
  hints.ai_socktype = (ipproto == IPPROTO_TCP) ? SOCK_STREAM : SOCK_DGRAM;
  hints.ai_protocol = ipproto;
  error = getaddrinfo(NULL, name, &hints, &res);
  if (error != 0) {
    switch (error) {

    case EAI_NONAME:
    case EAI_SERVICE:
      /*
       * No such port.  Just return -1.
       */
      break;

#ifdef EAI_SYSTEM
    case EAI_SYSTEM:
      /*
       * We don't use strerror() because it's not
       * guaranteed to be thread-safe on all platforms
       * (probably because it might use a non-thread-local
       * buffer into which to format an error message
       * if the error code isn't one for which it has
       * a canned string; three cheers for C string
       * handling).
       */
      bpf_set_error(cstate, "getaddrinfo(\"%s\" fails with system error: %d",
          name, errno);
      port = -2;  /* a real error */
      break;
#endif

    default:
      /*
       * This is a real error, not just "there's
       * no such service name".
       *
       * We don't use gai_strerror() because it's not
       * guaranteed to be thread-safe on all platforms
       * (probably because it might use a non-thread-local
       * buffer into which to format an error message
       * if the error code isn't one for which it has
       * a canned string; three cheers for C string
       * handling).
       */
      bpf_set_error(cstate, "getaddrinfo(\"%s\") fails with error: %d",
          name, error);
      port = -2;  /* a real error */
      break;
    }
  } else {
    /*
     * OK, we found it.  Did it find anything?
     */
    for (ai = res; ai != NULL; ai = ai->ai_next) {
      /*
       * Does it have an address?
       */
      if (ai->ai_addr != NULL) {
        /*
         * Yes.  Get a port number; we're done.
         */
        if (ai->ai_addr->sa_family == AF_INET) {
          in4 = (struct sockaddr_in *)ai->ai_addr;
          port = ntohs(in4->sin_port);
          break;
        }
        if (ai->ai_addr->sa_family == AF_INET6) {
          in6 = (struct sockaddr_in6 *)ai->ai_addr;
          port = ntohs(in6->sin6_port);
          break;
        }
      }
    }
    freeaddrinfo(res);
  }
  return port;
}

/*
 * Convert a string to a port number.
 */
static bpf_u_int32
stringtoport(compiler_state_t *cstate, const char *string, size_t string_size,
    int *proto)
{
  stoulen_ret ret;
  char *cpy;
  bpf_u_int32 val;
  int tcp_port = -1;
  int udp_port = -1;

  /*
   * See if it's a number.
   */
  ret = stoulen(string, string_size, &val, cstate);
  switch (ret) {

  case STOULEN_OK:
    /* Unknown port type - it's just a number. */
    *proto = PROTO_UNDEF;
    break;

  case STOULEN_NOT_OCTAL_NUMBER:
  case STOULEN_NOT_HEX_NUMBER:
  case STOULEN_NOT_DECIMAL_NUMBER:
    /*
     * Not a valid number; try looking it up as a port.
     */
    cpy = malloc(string_size + 1);  /* +1 for terminating '\0' */
    if (cpy == NULL) {
      bpf_set_error(cstate, "%s: out of memory", __func__);
      longjmp(cstate->top_ctx, 1);
      /*NOTREACHED*/
    }
    memcpy(cpy, string, string_size);
    cpy[string_size] = '\0';
    tcp_port = nametoport(cstate, cpy, IPPROTO_TCP);
    if (tcp_port == -2) {
      /*
       * We got a hard error; the error string has
       * already been set.
       */
      free(cpy);
      longjmp(cstate->top_ctx, 1);
      /*NOTREACHED*/
    }
    udp_port = nametoport(cstate, cpy, IPPROTO_UDP);
    if (udp_port == -2) {
      /*
       * We got a hard error; the error string has
       * already been set.
       */
      free(cpy);
      longjmp(cstate->top_ctx, 1);
      /*NOTREACHED*/
    }

    /*
     * We need to check /etc/services for ambiguous entries.
     * If we find an ambiguous entry, and it has the
     * same port number, change the proto to PROTO_UNDEF
     * so both TCP and UDP will be checked.
     */
    if (tcp_port >= 0) {
      val = (bpf_u_int32)tcp_port;
      *proto = IPPROTO_TCP;
      if (udp_port >= 0) {
        if (udp_port == tcp_port)
          *proto = PROTO_UNDEF;
#ifdef notdef
        else
          /* Can't handle ambiguous names that refer
             to different port numbers. */
          warning("ambiguous port %s in /etc/services",
            cpy);
#endif
      }
      free(cpy);
      break;
    }
    if (udp_port >= 0) {
      val = (bpf_u_int32)udp_port;
      *proto = IPPROTO_UDP;
      free(cpy);
      break;
    }
    bpf_set_error(cstate, "'%s' is not a valid port", cpy);
    free(cpy);
    longjmp(cstate->top_ctx, 1);
    /*NOTREACHED*/
#ifdef _AIX
    PCAP_UNREACHABLE
#endif /* _AIX */

  case STOULEN_ERROR:
    /* Error already set. */
    longjmp(cstate->top_ctx, 1);
    /*NOTREACHED*/
#ifdef _AIX
    PCAP_UNREACHABLE
#endif /* _AIX */

  default:
    /* Should not happen */
    bpf_set_error(cstate, "stoulen returned %d - this should not happen", ret);
    longjmp(cstate->top_ctx, 1);
    /*NOTREACHED*/
  }
  return (val);
}

/*
 * Convert a string in the form PPP-PPP, which correspond to ports, to
 * a starting and ending port in a port range.
 */
static void
stringtoportrange(compiler_state_t *cstate, const char *string,
    bpf_u_int32 *port1, bpf_u_int32 *port2, int *proto)
{
  const char *hyphen_off;
  const char *first, *second;
  size_t first_size, second_size;
  int save_proto;

  if ((hyphen_off = strchr(string, '-')) == NULL)
    bpf_error(cstate, "port range '%s' contains no hyphen", string);

  /*
   * Make sure there are no other hyphens.
   *
   * XXX - we support named ports, but there are some port names
   * in /etc/services that include hyphens, so this would rule
   * that out.
   */
  if (strchr(hyphen_off + 1, '-') != NULL)
    bpf_error(cstate, "port range '%s' contains more than one hyphen",
        string);

  /*
   * Get the length of the first port.
   */
  first = string;
  first_size = hyphen_off - string;
  if (first_size == 0) {
    /* Range of "-port", which we don't support. */
    bpf_error(cstate, "port range '%s' has no starting port", string);
  }

  /*
   * Try to convert it to a port.
   */
  *port1 = stringtoport(cstate, first, first_size, proto);
  save_proto = *proto;

  /*
   * Get the length of the second port.
   */
  second = hyphen_off + 1;
  second_size = strlen(second);
  if (second_size == 0) {
    /* Range of "port-", which we don't support. */
    bpf_error(cstate, "port range '%s' has no ending port", string);
  }

  /*
   * Try to convert it to a port.
   */
  *port2 = stringtoport(cstate, second, second_size, proto);
  if (*proto != save_proto)
    *proto = PROTO_UNDEF;
}

struct block *
gen_scode(compiler_state_t *cstate, const char *name, struct qual q)
{
  int proto = q.proto;
  int dir = q.dir;
  bpf_u_int32 mask, addr;
  int port, real_proto;
  bpf_u_int32 port1, port2;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  if (q.proto == Q_DECNET) {
    /*
     * A long time ago on Ultrix libpcap supported translation of
     * DECnet host names into DECnet addresses, but this feature
     * is history now.  The current implementation does not define
     * any primitives that have "decnet" as the protocol qualifier
     * and a name as the ID.
     */
    bpf_error(cstate, ERRSTR_INVALID_QUAL, "decnet",
              tqkw(q.addr == Q_DEFAULT ? Q_HOST : q.addr));
  }

  struct block *b, *b6;
  switch (q.addr) {

  case Q_NET:
    addr = pcap_nametonetaddr(name);
    if (addr == 0)
      bpf_error(cstate, "unknown network '%s'", name);
    /* Left justify network addr and calculate its network mask */
    mask = 0xffffffff;
    while (addr && (addr & 0xff000000) == 0) {
      addr <<= 8;
      mask <<= 8;
    }
    return gen_host(cstate, 1, &addr, &mask, q.proto, q.dir, 0,
                    "net <IPv4 network name>");

  case Q_DEFAULT:
  case Q_HOST:
    if (proto == Q_LINK) {
      return gen_mac48host_byname(cstate, name, q.dir, "link host NAME");
    } else {
      return gen_host46_byname(cstate, name, q.proto,
          q.proto, q.dir, 0);
    }

  case Q_PORT:
    (void)port_pq_to_ipproto(cstate, proto, "port"); // validate only
    if (pcap_nametoport(name, &port, &real_proto) == 0)
      bpf_error(cstate, "unknown port '%s'", name);
    if (proto == Q_UDP) {
      if (real_proto == IPPROTO_TCP)
        bpf_error(cstate, "port '%s' is tcp", name);
      else if (real_proto == IPPROTO_SCTP)
        bpf_error(cstate, "port '%s' is sctp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_UDP;
    }
    if (proto == Q_TCP) {
      if (real_proto == IPPROTO_UDP)
        bpf_error(cstate, "port '%s' is udp", name);

      else if (real_proto == IPPROTO_SCTP)
        bpf_error(cstate, "port '%s' is sctp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_TCP;
    }
    if (proto == Q_SCTP) {
      if (real_proto == IPPROTO_UDP)
        bpf_error(cstate, "port '%s' is udp", name);

      else if (real_proto == IPPROTO_TCP)
        bpf_error(cstate, "port '%s' is tcp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_SCTP;
    }

    /*
     * These two checks are redundant at this point: here name is
     * a string that the lexer does not recognize as a number
     * hence did not attempt stoulen(), pcap_nametoport() does not
     * use stoulen() and has successfully translated the string to
     * an uint16_t value using getaddrinfo().
     */
    if (port < 0)
      bpf_error(cstate, "illegal port number %d < 0", port);
    if (port > 65535)
      bpf_error(cstate, "illegal port number %d > 65535", port);

    // real_proto can be PROTO_UNDEF
    b = gen_port(cstate, (uint16_t)port, real_proto, q.dir, q.addr);
    b6 = gen_port6(cstate, (uint16_t)port, real_proto, q.dir, q.addr);
    return gen_or(b6, b);

  case Q_PORTRANGE:
    (void)port_pq_to_ipproto(cstate, proto, "portrange"); // validate only
    stringtoportrange(cstate, name, &port1, &port2, &real_proto);
    if (proto == Q_UDP) {
      if (real_proto == IPPROTO_TCP)
        bpf_error(cstate, "port in range '%s' is tcp", name);
      else if (real_proto == IPPROTO_SCTP)
        bpf_error(cstate, "port in range '%s' is sctp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_UDP;
    }
    if (proto == Q_TCP) {
      if (real_proto == IPPROTO_UDP)
        bpf_error(cstate, "port in range '%s' is udp", name);
      else if (real_proto == IPPROTO_SCTP)
        bpf_error(cstate, "port in range '%s' is sctp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_TCP;
    }
    if (proto == Q_SCTP) {
      if (real_proto == IPPROTO_UDP)
        bpf_error(cstate, "port in range '%s' is udp", name);
      else if (real_proto == IPPROTO_TCP)
        bpf_error(cstate, "port in range '%s' is tcp", name);
      else
        /* override PROTO_UNDEF */
        real_proto = IPPROTO_SCTP;
    }

    /*
     * When name is a string of the form "str1-str2", these two
     * checks are redundant at this point: in both stringtoport()
     * invocations stoulen() has rejected the argument and
     * getaddrinfo() has successfully translated it to an uint16_t
     * value.
     *
     * When name is a string of the form "num1-num2", "num-str" or
     * "str-num", these two checks are necessary: in at least one
     * stringtoport() invocation stoulen() can return any uint32_t
     * value if it has accepted the argument.
     */
    assert_maxval(cstate, "port number", port1, UINT16_MAX);
    assert_maxval(cstate, "port number", port2, UINT16_MAX);

    // real_proto can be PROTO_UNDEF
    b = gen_portrange(cstate, (uint16_t)port1, (uint16_t)port2,
        real_proto, dir);
    b6 = gen_portrange6(cstate, (uint16_t)port1, (uint16_t)port2,
        real_proto, dir);
    return gen_or(b6, b);

  case Q_GATEWAY:
    return gen_gateway(cstate, name, q.proto);

  case Q_PROTO:
    return gen_proto(cstate, lookup_proto(cstate, name, q), proto);

#if !defined(NO_PROTOCHAIN)
  case Q_PROTOCHAIN:
    return gen_protochain(cstate, lookup_proto(cstate, name, q), proto);
#endif /* !defined(NO_PROTOCHAIN) */

  case Q_UNDEF:
    syntax(cstate);
    /*NOTREACHED*/
  }
  bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), name);
  /*NOTREACHED*/
}

struct block *
gen_mcode(compiler_state_t *cstate, const char *s1, const char *s2,
    bpf_u_int32 masklen, struct qual q)
{
  int nlen, mlen;
  bpf_u_int32 n, m;
  uint64_t m64;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  if (q.proto == Q_DECNET) {
    /*
     * libpcap has never defined any primitives that have "decnet"
     * as the protocol qualifier and an IPv4 network with a
     * netmask as the ID.
     */
    bpf_error(cstate, ERRSTR_INVALID_QUAL, "decnet",
              tqkw(q.addr == Q_DEFAULT ? Q_HOST : q.addr));
  }

  nlen = pcapint_atoin(s1, &n);
  if (nlen < 0)
    bpf_error(cstate, ERRSTR_INVALID_IPV4_ADDR, s1);
  /* Promote short ipaddr */
  n <<= 32 - nlen;

  char idstr[PCAP_BUF_SIZE];
  if (s2 != NULL) {
    mlen = pcapint_atoin(s2, &m);
    if (mlen < 0)
      bpf_error(cstate, ERRSTR_INVALID_IPV4_ADDR, s2);
    /* Promote short ipaddr */
    m <<= 32 - mlen;
    snprintf(idstr, sizeof(idstr), "%s mask %s", s1, s2);
  } else {
    /* Convert mask len to mask */
    assert_maxval(cstate, "netmask length", masklen, 32);
    m64 = UINT64_C(0xffffffff) << (32 - masklen);
    m = (bpf_u_int32)m64;
    snprintf(idstr, sizeof(idstr), "%s/%u", s1, masklen);
  }
  if ((n & ~m) != 0)
    bpf_error(cstate, "non-network bits set in \"%s\"", idstr);

  switch (q.addr) {

  case Q_NET:
    return gen_host(cstate, 1, &n, &m, q.proto, q.dir, 0,
                    "net <IPv4 prefix>");

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), idstr);
    /*NOTREACHED*/
  }
  /*NOTREACHED*/
}

UNSIGNED_SHIFT_OK struct block *
gen_ncode(compiler_state_t *cstate, const char *s, bpf_u_int32 v, struct qual q)
{
  bpf_u_int32 mask;
  int proto;
  int vlen;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  if (q.proto == Q_DECNET)
    return gen_dnhost(cstate, s, v, q);

  proto = q.proto;
  char idstr[PCAP_BUF_SIZE];
  if (s == NULL) {
    /*
     * v contains a 32-bit unsigned parsed from a string of the
     * form {N}, which could be decimal, hexadecimal or octal.
     * This is a valid IPv4 address, in the sense of inet_aton(3).
     */
    vlen = 32;
    snprintf(idstr, sizeof(idstr), "%u", v);
  } else {
    /*
     * s points to a string of the form {N}.{N}, {N}.{N}.{N} or
     * {N}.{N}.{N}.{N}, all of which potentially stand for a valid
     * IPv4 address, in the sense of inet_aton(3).
     */
    vlen = pcapint_atoin(s, &v);
    if (vlen < 0)
      bpf_error(cstate, ERRSTR_INVALID_IPV4_ADDR, s);
    snprintf(idstr, sizeof(idstr), "%s", s);
  }

  struct block *b, *b6;
  switch (q.addr) {

  case Q_DEFAULT:
  case Q_HOST:
  case Q_NET:
    if (proto == Q_LINK) {
      bpf_error(cstate, "illegal link-layer address '%s'", idstr);
    } else {
      mask = 0xffffffff;
      if (s == NULL && q.addr == Q_NET) {
        /* Promote short net number */
        while (v && (v & 0xff000000) == 0) {
          v <<= 8;
          mask <<= 8;
        }
      } else {
        /* Promote short ipaddr */
        v <<= 32 - vlen;
        mask <<= 32 - vlen ;
      }
      return gen_host(cstate, 1, &v, &mask, q.proto, q.dir, 0,
                      q.addr == Q_NET ? "net <IPv4 address>" :
                      "host <IPv4 address>");
    }

  case Q_PORTRANGE: // "portrange <n>" means the same as "port <n>".
  case Q_PORT:
    proto = port_pq_to_ipproto(cstate, proto, tqkw(q.addr));

    // This check is necessary: v can hold any uint32_t value.
    assert_maxval(cstate, "port number", v, UINT16_MAX);

    // proto can be PROTO_UNDEF
    b = gen_port(cstate, (uint16_t)v, proto, q.dir, q.addr);
    b6 = gen_port6(cstate, (uint16_t)v, proto, q.dir, q.addr);
    return gen_or(b6, b);

  case Q_PROTO:
    return gen_proto(cstate, v, proto);

#if !defined(NO_PROTOCHAIN)
  case Q_PROTOCHAIN:
    return gen_protochain(cstate, v, proto);
#endif

  case Q_UNDEF:
    syntax(cstate);
    /*NOTREACHED*/

  default:
    bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), idstr);
    /*NOTREACHED*/
  }
  /*NOTREACHED*/
}

struct block *
gen_mcode6(compiler_state_t *cstate, const char *s, bpf_u_int32 masklen,
    struct qual q)
{
  struct in6_addr addr;
  struct in6_addr mask;
  bpf_u_int32 a[4], m[4]; /* Same as in gen_hostop6(). */

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * If everything works correctly, this call never fails: a string that
   * is valid for HID6 and the associated validating inet_pton() in the
   * lexer is valid for inet_pton() here.
   */
  if (1 != inet_pton(AF_INET6, s, &addr))
    bpf_error(cstate, "'%s' is not a valid IPv6 address", s);

  if (masklen > sizeof(mask.s6_addr) * 8)
    bpf_error(cstate, "mask length must be <= %zu", sizeof(mask.s6_addr) * 8);
  memset(&mask, 0, sizeof(mask));
  memset(&mask.s6_addr, 0xff, masklen / 8);
  if (masklen % 8) {
    mask.s6_addr[masklen / 8] =
      (0xff << (8 - masklen % 8)) & 0xff;
  }

  memcpy(a, &addr, sizeof(a));
  memcpy(m, &mask, sizeof(m));
  if ((a[0] & ~m[0]) || (a[1] & ~m[1])
   || (a[2] & ~m[2]) || (a[3] & ~m[3])) {
    bpf_error(cstate, "non-network bits set in \"%s/%d\"", s, masklen);
  }

  char buf[INET6_ADDRSTRLEN + sizeof("/128")];
  switch (q.addr) {

  case Q_DEFAULT:
  case Q_HOST:
    if (masklen != 128) {
      snprintf(buf, sizeof(buf), "%s/%u", s, masklen);
      bpf_error(cstate, ERRSTR_INVALID_QUAL, "host", buf);
    }
    /* FALLTHROUGH */

  case Q_NET:
    return gen_host6(cstate, 1, &addr, &mask, q.proto, q.dir, 0,
                     q.addr == Q_HOST ? "host <IPv6 address>" :
                     "net <IPv6 prefix>");

  default:
    if (masklen == 128)
      bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), s);
    else {
      snprintf(buf, sizeof(buf), "%s/%u", s, masklen);
      bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), buf);
    }
    /*NOTREACHED*/
  }
}

struct block *
gen_ecode(compiler_state_t *cstate, const char *s, struct qual q)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  const char *context = "link host XX:XX:XX:XX:XX:XX";

  if (! ((q.addr == Q_HOST || q.addr == Q_DEFAULT) && q.proto == Q_LINK))
    bpf_error(cstate, "Ethernet address used in non-ether expression");
  if (! is_mac48_linktype(cstate->linktype))
    fail_kw_on_dlt(cstate, context);

  u_char eaddr[6];
  /*
   * Belt and braces: so long as the lexer regexp guards MAC-48 syntax,
   * here the attempt to parse it will always succeed.
   */
  if (! pcapint_atomac48(s, eaddr))
    bpf_error(cstate, "invalid MAC-48 address '%s'", s);

  return gen_mac48host(cstate, eaddr, q.dir, context);
}

// Process a regular primitive, the ID is a MAC-8 address string.
struct block *
gen_acode(compiler_state_t *cstate, const char *s, struct qual q)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  // WLAN direction qualifiers are never valid for MAC-8 addresses.
  assert_nonwlan_dqual(cstate, q.dir);

  if (q.addr != Q_HOST && q.addr != Q_DEFAULT)
    bpf_error(cstate, ERRSTR_INVALID_QUAL, tqkw(q.addr), "$XX");
  if (q.proto != Q_LINK)
    bpf_error(cstate, "'link' is the only valid proto qualifier for 'host $XX'");

  uint8_t addr;
  /*
   * The lexer currently defines the address format in a way that makes
   * this error condition never true.  Let's check it anyway in case this
   * part of the lexer changes in future.
   */
  if (! pcapint_atoan(s, &addr))
      bpf_error(cstate, "invalid MAC-8 address '%s'", s);

  return gen_mac8host(cstate, addr, q.dir, "link host $XX");
}

void
sappend(struct slist *s0, struct slist *s1)
{
  /*
   * This is definitely not the best way to do this, but the
   * lists will rarely get long.
   */
  while (s0->next)
    s0 = s0->next;
  s0->next = s1;
}

/*
 * Prepend the given list of statements to the list of side effect statements
 * of the block.  Either of the lists may be NULL to mean the valid edge case
 * of an empty list.
 */
static struct block *
sprepend_to_block(struct slist *s, struct block *b)
{
  if (s) {
    if (b->stmts)
      sappend(s, b->stmts);
    b->stmts = s;
    /*
     * The block has changed.  It could have been a Boolean
     * constant before.
     */
    b->meaning = IS_UNCERTAIN;
  }
  return b;
}

static struct slist *
xfer_to_x(compiler_state_t *cstate, const struct arth *a)
{
  return NEW_STMT_LDX_M(cstate, a->regno);
}

static struct slist *
xfer_to_a(compiler_state_t *cstate, const struct arth *a)
{
  return NEW_STMT_LD_M(cstate, a->regno);
}

/*
 * Modify "inst" to use the value stored into its register as an
 * offset relative to the beginning of the header for the protocol
 * "proto", and allocate a register and put an item "size" bytes long
 * (1, 2, or 4) at that offset into that register, making it the register
 * for "inst".
 */
static struct arth *
gen_load_internal(compiler_state_t *cstate, int proto, struct arth *inst,
    bpf_u_int32 size)
{
  int size_code;
  int regno = alloc_reg(cstate);

  switch (size) {

  default:
    bpf_error(cstate, "data size must be 1, 2, or 4");
    /*NOTREACHED*/

  case 1:
    size_code = BPF_B;
    break;

  case 2:
    size_code = BPF_H;
    break;

  case 4:
    size_code = BPF_W;
    break;
  }
  struct block *b = NULL; // protocol checks
  struct slist *s = NULL; // the variable part of an absolute offset
  u_int constpart = 0;    // the constant part of an absolute offset
  switch (proto) {
  default:
    bpf_error(cstate, "'%s' does not support the index operation", pqkw(proto));

  case Q_RADIO:
    /*
     * This corresponds to OR_PACKET in gen_load_a().
     *
     * The offset is relative to the beginning of the packet
     * data, if we have a radio header.  (If we don't, this
     * is an error.)
     */
    if (cstate->linktype != DLT_IEEE802_11_RADIO_AVS &&
        cstate->linktype != DLT_IEEE802_11_RADIO &&
        cstate->linktype != DLT_PRISM_HEADER)
      bpf_error(cstate, "radio information not present in capture");

    /*
     * Load into the X register the offset computed into the
     * register specified by "inst".
     *
     * Load the item at that offset.
     *
     * In other words, the variable part is not present, the
     * constant part is zero and there are no protocol checks, so
     * just break out to proceed with "inst" only.
     */
    break;

  case Q_LINK:
    /*
     * This corresponds to OR_LINKHDR in gen_load_a().
     *
     * The offset is relative to the beginning of
     * the link-layer header.
     *
     * XXX - what about ATM LANE?  Should "inst" be
     * relative to the beginning of the AAL5 frame, so
     * that 0 refers to the beginning of the LE Control
     * field, or relative to the beginning of the LAN
     * frame, so that 0 refers, for Ethernet LANE, to
     * the beginning of the destination address?
     */
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkhdr);

    /*
     * If "s" is non-null, it has code to arrange that the
     * X register contains the length of the prefix preceding
     * the link-layer header.  Add to it the offset computed
     * into the register specified by "inst", and move that
     * into the X register.  Otherwise, just load into the X
     * register the offset computed into the register specified
     * by "inst".
     *
     * Load the item at the sum of the offset we've put in the
     * X register and the offset of the start of the link
     * layer header (which is 0 if the radio header is
     * variable-length; that header length is what we put
     * into the X register and then added to "inst").
     */
    constpart = cstate->off_linkhdr.constant_part;
    // There are no protocol checks.
    break;

  case Q_IP:
  case Q_ARP:
  case Q_RARP:
  case Q_ATALK:
  case Q_DECNET:
  case Q_SCA:
  case Q_LAT:
  case Q_MOPRC:
  case Q_MOPDL:
  case Q_IPV6:
    /*
     * This corresponds to OR_LINKPL in gen_load_a().
     *
     * The offset is relative to the beginning of
     * the network-layer header.
     * XXX - are there any cases where we want
     * cstate->off_nl_nosnap?
     */
    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);

    /*
     * If "s" is non-null, it has code to arrange that the
     * X register contains the variable part of the offset
     * of the link-layer payload.  Add to it the offset
     * computed into the register specified by "inst",
     * and move that into the X register.  Otherwise, just
     * load into the X register the offset computed into
     * the register specified by "inst".
     *
     * Load the item at the sum of the offset we've put in the
     * X register, the offset of the start of the network
     * layer header from the beginning of the link-layer
     * payload, and the constant part of the offset of the
     * start of the link-layer payload.
     */
    constpart = cstate->off_linkpl.constant_part + cstate->off_nl;

    /*
     * Do the computation only if the packet contains
     * the protocol in question.
     */
    b = gen_proto_abbrev_internal(cstate, proto);
    break;

  case Q_SCTP:
  case Q_TCP:
  case Q_UDP:
  case Q_ICMP:
  case Q_IGMP:
  case Q_IGRP:
  case Q_PIM:
  case Q_VRRP:
  case Q_CARP:
    /*
     * This corresponds to OR_TRAN_IPV4 in gen_load_a().
     *
     * The offset is relative to the beginning of
     * the transport-layer header.
     *
     * Load the X register with the length of the IPv4 header
     * (plus the offset of the link-layer header, if it's
     * a variable-length header), in bytes.
     *
     * XXX - are there any cases where we want
     * cstate->off_nl_nosnap?
     * XXX - we should, if we're built with
     * IPv6 support, generate code to load either
     * IPv4, IPv6, or both, as appropriate.
     */
    s = gen_loadx_iphdrlen(cstate);

    /*
     * The X register now contains the sum of the variable
     * part of the offset of the link-layer payload and the
     * length of the network-layer header.
     *
     * Load into the A register the offset relative to
     * the beginning of the transport layer header,
     * add the X register to that, move that to the
     * X register, and load with an offset from the
     * X register equal to the sum of the constant part of
     * the offset of the link-layer payload and the offset,
     * relative to the beginning of the link-layer payload,
     * of the network-layer header.
     */
    constpart = cstate->off_linkpl.constant_part + cstate->off_nl;

    /*
     * Do the computation only if the packet contains
     * the protocol in question - which is true only
     * if this is an IP datagram and is the first or
     * only fragment of that datagram.
     *
     * Do not use gen_proto_abbrev_internal(cstate, proto): if it
     * matches the given proto qualifier using Q_DEFAULT, this
     * would produce an unreachable IPv6 branch.
     */
    b = gen_proto_abbrev_internal(cstate, Q_IP);
    b = gen_and(b, gen_ip_proto(cstate, pq_to_ipproto(cstate,
        (u_char)proto)));
    b = gen_and(b, gen_ipfrag(cstate));
    break;
  case Q_ICMPV6:
    /*
     * This corresponds to OR_TRAN_IPV6 in gen_load_a().
     *
     * Do the computation only if the packet contains
     * the protocol in question.
     *
     * Do not use gen_proto(..., Q_IPV6): this would also match
     * IPPROTO_FRAGMENT and the side effect statements would
     * quietly load incorrect data.
     */
    b = gen_proto_abbrev_internal(cstate, Q_IPV6);

    /*
     * Check if we have an icmp6 next header
     */
    b = gen_and(b, gen_ip6_proto(cstate, IPPROTO_ICMPV6));

    s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
    /*
     * If "s" is non-null, it has code to arrange that the
     * X register contains the variable part of the offset
     * of the link-layer payload.  Add to it the offset
     * computed into the register specified by "inst",
     * and move that into the X register.  Otherwise, just
     * load into the X register the offset computed into
     * the register specified by "inst".
     *
     * Load the item at the sum of the offset we've put in the
     * X register, the offset of the start of the network
     * layer header from the beginning of the link-layer
     * payload, and the constant part of the offset of the
     * start of the link-layer payload.
     */
    constpart = cstate->off_linkpl.constant_part + cstate->off_nl +
        IP6_HDRLEN;
    break;
  }

  if (b)
    inst->b = inst->b ? gen_and(inst->b, b) : b;
  // NULL is a valid value for 's'.
  sappend(inst->s, gen_load_absoffsetarthrel(cstate, s, constpart, inst,
      size_code));

  /*
   * Only now it is correct to deallocate the input register because
   * gen_load_absoffsetarthrel() was using it just before.
   */
  free_reg_arth(cstate, inst);

  inst->regno = regno;
  s = NEW_STMT_ST_M(cstate, regno);
  sappend(inst->s, s);

  return inst;
}

struct arth *
gen_load(compiler_state_t *cstate, int proto, struct arth *inst,
    bpf_u_int32 size)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_load_internal(cstate, proto, inst, size);
}

static struct block *
gen_relation_internal(compiler_state_t *cstate, int code, struct arth *a0,
    struct arth *a1, int reversed)
{
  struct slist *s0, *s1;
  struct block *b;

  s0 = xfer_to_x(cstate, a1);
  s1 = xfer_to_a(cstate, a0);
  sappend(s0, s1);
  sappend(a1->s, s0);
  sappend(a0->s, a1->s);

  b = gen_jmp_x(cstate, code, a0->s);
  if (reversed)
    gen_not(b);

  free_reg_arth(cstate, a0);
  free_reg_arth(cstate, a1);

  /* 'and' together protocol checks */
  if (a0->b)
    b = gen_and(a0->b, b);
  if (a1->b)
    b = gen_and(a1->b, b);
  return b;
}

struct block *
gen_relation(compiler_state_t *cstate, int code, struct arth *a0,
    struct arth *a1, int reversed)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_relation_internal(cstate, code, a0, a1, reversed);
}

struct arth *
gen_loadlen(compiler_state_t *cstate)
{
  int regno;
  struct arth *a;
  struct slist *s;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  regno = alloc_reg(cstate);
  a = (struct arth *)newchunk(cstate, sizeof(*a));
  s = NEW_STMT_LD_LEN(cstate);
  s->next = NEW_STMT_ST_M(cstate, regno);
  a->s = s;
  a->regno = regno;

  return a;
}

static struct arth *
gen_loadi_internal(compiler_state_t *cstate, bpf_u_int32 val)
{
  struct arth *a;
  struct slist *s;
  int reg;

  a = (struct arth *)newchunk(cstate, sizeof(*a));

  reg = alloc_reg(cstate);

  s = NEW_STMT_LD_IMM(cstate, val);
  s->next = NEW_STMT_ST_M(cstate, reg);
  a->s = s;
  a->regno = reg;

  return a;
}

struct arth *
gen_loadi(compiler_state_t *cstate, bpf_u_int32 val)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_loadi_internal(cstate, val);
}

/*
 * Return true iff the given arithmetic expression is a result of gen_loadi(),
 * which in the current implementation means it uses exactly two BPF
 * statements: an "ld #k" followed by an "st M[k]".
 */
static inline bool
is_loadi(const struct arth *a)
{
  return a->s->s.code == (BPF_LD|BPF_W|BPF_IMM) &&
      a->s->next != NULL && a->s->next->s.code == BPF_ST &&
      a->s->next->next == NULL;
}

/*
 * The a_arg dance is to avoid annoying whining by compilers that
 * a might be clobbered by longjmp - yeah, it might, but *WHO CARES*?
 * It's not *used* after setjmp returns.
 */
struct arth *
gen_neg(compiler_state_t *cstate, struct arth *a_arg)
{
  struct arth *a = a_arg;
  struct slist *s;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  s = xfer_to_a(cstate, a);
  sappend(a->s, s);
  s = NEW_STMT_NEG(cstate);
  sappend(a->s, s);
  s = NEW_STMT_ST_M(cstate, a->regno);
  sappend(a->s, s);

  return a;
}

/*
 * The a0_arg dance is to avoid annoying whining by compilers that
 * a0 might be clobbered by longjmp - yeah, it might, but *WHO CARES*?
 * It's not *used* after setjmp returns.
 */
struct arth *
gen_arth(compiler_state_t *cstate, int code, struct arth *a0_arg,
    struct arth *a1)
{
  struct arth *a0 = a0_arg;
  struct slist *s0, *s1, *s2;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Disallow division by, or modulo by, zero; we do this here
   * so that it gets done even if the optimizer is disabled.
   *
   * Also disallow shifts by a value greater than 31; we do this
   * here, for the same reason.
   *
   * These checks apply only to the simplest case of the 2nd operand of
   * a binary operation -- an immediate value.  Anything more
   * sophisticated (even a negation of an immediate value) will have to be
   * handled in the optimizer and/or the interpreter.
   */
  if (is_loadi(a1))
    switch (code) {
    case BPF_DIV:
      if (a1->s->s.k == 0)
        bpf_error(cstate, ERRSTR_DIV_BY_ZERO);
      break;
    case BPF_MOD:
      if (a1->s->s.k == 0)
        bpf_error(cstate, ERRSTR_MOD_BY_ZERO);
      break;
    case BPF_LSH:
    case BPF_RSH:
      if (a1->s->s.k > 31)
        bpf_error(cstate, ERRSTR_SHIFT_BY_MORE);
      break;
    }

  s0 = xfer_to_x(cstate, a1);
  s1 = xfer_to_a(cstate, a0);
  s2 = new_stmt(cstate, OPCODE_ALU_X(code));

  sappend(s1, s2);
  sappend(s0, s1);
  sappend(a1->s, s0);
  sappend(a0->s, a1->s);

  free_reg_arth(cstate, a0);
  free_reg_arth(cstate, a1);

  a0->regno = alloc_reg(cstate);
  s0 = NEW_STMT_ST_M(cstate, a0->regno);
  sappend(a0->s, s0);

  return a0;
}

/*
 * Initialize the table of used registers and the current register.
 */
static void
init_regs(compiler_state_t *cstate)
{
  cstate->curreg = 0;
  memset(cstate->regused, 0, sizeof cstate->regused);
}

/*
 * Return the next free register.
 */
static int
alloc_reg(compiler_state_t *cstate)
{
  int n = BPF_MEMWORDS;

  while (--n >= 0) {
    if (cstate->regused[cstate->curreg])
      cstate->curreg = (cstate->curreg + 1) % BPF_MEMWORDS;
    else {
      cstate->regused[cstate->curreg] = 1;
      return cstate->curreg;
    }
  }
  bpf_error(cstate, "too many registers needed to evaluate expression");
  /*NOTREACHED*/
}

/*
 * compiler_state_t.regused is a fixed-size array.  If the array index points
 * outside of the array, this means a bug that is about to cause OOBR or OOBW.
 * If the register index is valid, but the register has already been freed,
 * trying to use it or to free it again means another bug.  In either case,
 * fail before such bugs cascade.
 */
static int
guard_reg(compiler_state_t *cstate, const int regno)
{
  if (regno < 0 || regno >= BPF_MEMWORDS || ! cstate->regused[regno])
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "regno", regno);
  return regno;
}

/*
 * Return a register to the table so it can
 * be used later.
 */
static void
free_reg(compiler_state_t *cstate, int n)
{
  cstate->regused[guard_reg(cstate, n)] = 0;
}

/*
 * If the code deallocates an arithmetic expression register, then allocates
 * the same register for a different purpose and erroneously uses the
 * expression again, the register index will look valid (in bounds and marked
 * as allocated), but will be logically invalid.  Invalidate the index on
 * deallocation to avoid this.
 */
static void
free_reg_arth(compiler_state_t *cstate, struct arth *a)
{
  free_reg(cstate, a->regno);
  a->regno = -1;
}

static struct block *
gen_len(compiler_state_t *cstate, int jmp, int n)
{
  struct slist *s;

  s = NEW_STMT_LD_LEN(cstate);
  return gen_jmp_k(cstate, jmp, n, s);
}

struct block *
gen_greater(compiler_state_t *cstate, int n)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_len(cstate, BPF_JGE, n);
}

/*
 * Actually, this is less than or equal.
 */
struct block *
gen_less(compiler_state_t *cstate, int n)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_not(gen_len(cstate, BPF_JGT, n));
}

/*
 * This is for "byte {idx} {op} {val}"; "idx" is treated as relative to
 * the beginning of the link-layer header.
 */
struct block *
gen_byteop(compiler_state_t *cstate, int op, int idx, bpf_u_int32 val)
{
  struct block *b;
  struct slist *s;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_maxval(cstate, "byte argument", val, UINT8_MAX);

  switch (op) {
  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "op", op);

  case '=':
    return gen_cmp(cstate, OR_LINKHDR, (u_int)idx, BPF_B, val);

  case '<':
    return gen_cmp_lt(cstate, OR_LINKHDR, (u_int)idx, BPF_B, val);

  case '>':
    return gen_cmp_gt(cstate, OR_LINKHDR, (u_int)idx, BPF_B, val);

  case '|':
    s = NEW_STMT_OR_K(cstate, val);
    break;

  case '&':
    s = NEW_STMT_AND_K(cstate, val);
    break;
  }
  // Load the required byte first.
  struct slist *s0 = gen_load_a(cstate, OR_LINKHDR, idx, BPF_B);
  sappend(s0, s);
  b = gen_jmp_k(cstate, BPF_JEQ, 0, s0);

  return gen_not(b);
}

struct block *
gen_broadcast(compiler_state_t *cstate, int proto)
{
  bpf_u_int32 hostmask;
  struct block *b0, *b1, *b2;
  static const u_char ebroadcast[] = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff };

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  switch (proto) {

  case Q_DEFAULT:
  case Q_LINK:
    switch (cstate->linktype) {
    case DLT_ARCNET:
    case DLT_ARCNET_LINUX:
      // ARCnet broadcast is [8-bit] destination address 0.
      return gen_mac8host(cstate, 0, Q_DST, "broadcast");
    case DLT_BACNET_MS_TP:
      // MS/TP broadcast is [8-bit] destination address 0xFF.
      return gen_mac8host(cstate, 0xFF, Q_DST, "broadcast");
    }
    return gen_mac48host(cstate, ebroadcast, Q_DST, "broadcast");
    /*NOTREACHED*/

  case Q_IP:
    /*
     * We treat a netmask of PCAP_NETMASK_UNKNOWN (0xffffffff)
     * as an indication that we don't know the netmask, and fail
     * in that case.
     */
    if (cstate->netmask == PCAP_NETMASK_UNKNOWN)
      bpf_error(cstate, "netmask not known, so 'ip broadcast' not supported");
    b0 = gen_linktype(cstate, ETHERTYPE_IP);
    hostmask = ~cstate->netmask;
    b1 = gen_mcmp(cstate, OR_LINKPL, IPV4_DSTADDR_OFFSET, BPF_W,
        0, hostmask);
    b2 = gen_mcmp(cstate, OR_LINKPL, IPV4_DSTADDR_OFFSET, BPF_W,
        hostmask, hostmask);
    return gen_and(b0, gen_or(b1, b2));
  }
  bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), "broadcast");
  /*NOTREACHED*/
}

/*
 * Generate code to test the low-order bit of a MAC address (that's
 * the bottom bit of the *first* byte).
 */
static struct block *
gen_mac_multicast(compiler_state_t *cstate, int offset)
{
  struct slist *s;

  /* link[offset] & 1 != 0 */
  s = gen_load_a(cstate, OR_LINKHDR, offset, BPF_B);
  return gen_set(cstate, 1, s);
}

struct block *
gen_multicast(compiler_state_t *cstate, int proto)
{
  struct block *b0, *b1, *b2;
  struct slist *s;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  switch (proto) {

  case Q_DEFAULT:
  case Q_LINK:
    switch (cstate->linktype) {
    case DLT_ARCNET:
    case DLT_ARCNET_LINUX:
      // ARCnet multicast is the same as broadcast.
      return gen_mac8host(cstate, 0, Q_DST, "multicast");
    case DLT_EN10MB:
    case DLT_NETANALYZER:
    case DLT_NETANALYZER_TRANSPARENT:
    case DLT_DSA_TAG_BRCM:
    case DLT_DSA_TAG_DSA:
      b1 = gen_prevlinkhdr_check(cstate);
      /* ether[0] & 1 != 0 */
      b0 = gen_mac_multicast(cstate, 0);
      return b1 ? gen_and(b1, b0) : b0;
    case DLT_FDDI:
      /*
       * XXX TEST THIS: MIGHT NOT PORT PROPERLY XXX
       *
       * XXX - was that referring to bit-order issues?
       */
      /* fddi[1] & 1 != 0 */
      return gen_mac_multicast(cstate, 1);
    case DLT_IEEE802:
      /* tr[2] & 1 != 0 */
      return gen_mac_multicast(cstate, 2);
    case DLT_IEEE802_11:
    case DLT_PRISM_HEADER:
    case DLT_IEEE802_11_RADIO_AVS:
    case DLT_IEEE802_11_RADIO:
    case DLT_PPI:
      /*
       * Oh, yuk.
       *
       *  For control frames, there is no DA.
       *
       *  For management frames, DA is at an
       *  offset of 4 from the beginning of
       *  the packet.
       *
       *  For data frames, DA is at an offset
       *  of 4 from the beginning of the packet
       *  if To DS is clear and at an offset of
       *  16 from the beginning of the packet
       *  if To DS is set.
       */

      /*
       * Generate the tests to be done for data frames.
       *
       * First, check for To DS set, i.e. "link[1] & 0x01".
       */
      s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
      b1 = gen_set(cstate, IEEE80211_FC1_DIR_TODS, s);

      /*
       * If To DS is set, the DA is at 16.
       */
      b0 = gen_mac_multicast(cstate, 16);
      b0 = gen_and(b1, b0);

      /*
       * Now, check for To DS not set, i.e. check
       * "!(link[1] & 0x01)".
       */
      s = gen_load_a(cstate, OR_LINKHDR, 1, BPF_B);
      b2 = gen_unset(cstate, IEEE80211_FC1_DIR_TODS, s);

      /*
       * If To DS is not set, the DA is at 4.
       */
      b1 = gen_mac_multicast(cstate, 4);
      b1 = gen_and(b2, b1);

      /*
       * Now OR together the last two checks.  That gives
       * the complete set of checks for data frames.
       */
      b0 = gen_or(b1, b0);

      /*
       * Now check for a data frame.
       * I.e, check "link[0] & 0x08".
       */
      s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
      b1 = gen_set(cstate, IEEE80211_FC0_TYPE_DATA, s);

      /*
       * AND that with the checks done for data frames.
       */
      b0 = gen_and(b1, b0);

      /*
       * If the high-order bit of the type value is 0, this
       * is a management frame.
       * I.e, check "!(link[0] & 0x08)".
       */
      s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
      b2 = gen_unset(cstate, IEEE80211_FC0_TYPE_DATA, s);

      /*
       * For management frames, the DA is at 4.
       */
      b1 = gen_mac_multicast(cstate, 4);
      b1 = gen_and(b2, b1);

      /*
       * OR that with the checks done for data frames.
       * That gives the checks done for management and
       * data frames.
       */
      b0 = gen_or(b1, b0);

      /*
       * If the low-order bit of the type value is 1,
       * this is either a control frame or a frame
       * with a reserved type, and thus not a
       * frame with an SA.
       *
       * I.e., check "!(link[0] & 0x04)".
       */
      s = gen_load_a(cstate, OR_LINKHDR, 0, BPF_B);
      b1 = gen_unset(cstate, IEEE80211_FC0_TYPE_CTL, s);

      /*
       * AND that with the checks for data and management
       * frames.
       */
      return gen_and(b1, b0);
    case DLT_IP_OVER_FC:
      return gen_mac_multicast(cstate, 2);
    default:
      break;
    }
    fail_kw_on_dlt(cstate, "multicast");
    /*NOTREACHED*/

  case Q_IP:
    b0 = gen_linktype(cstate, ETHERTYPE_IP);

    /*
     * Compare address with 224.0.0.0/4
     */
    b1 = gen_mcmp(cstate, OR_LINKPL, IPV4_DSTADDR_OFFSET, BPF_B,
        0xe0, 0xf0);

    return gen_and(b0, b1);

  case Q_IPV6:
    b0 = gen_linktype(cstate, ETHERTYPE_IPV6);
    b1 = gen_cmp(cstate, OR_LINKPL, IPV6_DSTADDR_OFFSET, BPF_B, 255);
    return gen_and(b0, b1);
  }
  bpf_error(cstate, ERRSTR_INVALID_QUAL, pqkw(proto), "multicast");
  /*NOTREACHED*/
}

#ifdef __linux__
/*
 * This is Linux; we require PF_PACKET support.  If this is a *live* capture,
 * we can look at special meta-data in the filter expression; otherwise we
 * can't because it is either a savefile (rfile != NULL) or a pcap_t created
 * using pcap_open_dead() (rfile == NULL).  Thus check for a flag that
 * pcap_activate() conditionally sets.
 */
static void
require_basic_bpf_extensions(compiler_state_t *cstate, const char *keyword)
{
  if (cstate->bpf_pcap->bpf_codegen_flags & BPF_SPECIAL_BASIC_HANDLING)
    return;
  bpf_error(cstate, "not a live capture, '%s' not supported on %s",
      keyword,
      pcapint_datalink_val_to_string(cstate->linktype));
}
#endif // __linux__

struct block *
gen_ifindex(compiler_state_t *cstate, int ifindex)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Only some data link types support ifindex qualifiers.
   */
  switch (cstate->linktype) {
  case DLT_LINUX_SLL2:
    /* match packets on this interface */
    return gen_cmp(cstate, OR_LINKHDR, 4, BPF_W, ifindex);
  default:
#if defined(__linux__)
    require_basic_bpf_extensions(cstate, "ifindex");
    /* match ifindex */
    return gen_cmp(cstate, OR_LINKHDR, SKF_AD_OFF + SKF_AD_IFINDEX, BPF_W,
                 ifindex);
#else /* defined(__linux__) */
    fail_kw_on_dlt(cstate, "ifindex");
    /*NOTREACHED*/
#endif /* defined(__linux__) */
  }
}

/*
 * Filter on inbound (outbound == 0) or outbound (outbound == 1) traffic.
 * Outbound traffic is sent by this machine, while inbound traffic is
 * sent by a remote machine (and may include packets destined for a
 * unicast or multicast link-layer address we are not subscribing to).
 * These are the same definitions implemented by pcap_setdirection().
 * Capturing only unicast traffic destined for this host is probably
 * better accomplished using a higher-layer filter.
 */
struct block *
gen_inbound_outbound(compiler_state_t *cstate, const int outbound)
{
  struct block *b0;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Only some data link types support inbound/outbound qualifiers.
   */
  switch (cstate->linktype) {
  case DLT_SLIP:
    return gen_cmp(cstate, OR_LINKHDR, 0, BPF_B,
        outbound ? SLIPDIR_OUT : SLIPDIR_IN);

  case DLT_IPNET:
    return gen_cmp(cstate, OR_LINKHDR, 2, BPF_H,
        outbound ? IPNET_OUTBOUND : IPNET_INBOUND);

  case DLT_LINUX_SLL:
    /* match outgoing packets */
    b0 = gen_cmp(cstate, OR_LINKHDR, 0, BPF_H, LINUX_SLL_OUTGOING);
    // To filter on inbound traffic, invert the match.
    return outbound ? b0 : gen_not(b0);

  case DLT_LINUX_SLL2:
    /* match outgoing packets */
    b0 = gen_cmp(cstate, OR_LINKHDR, 10, BPF_B, LINUX_SLL_OUTGOING);
    // To filter on inbound traffic, invert the match.
    return outbound ? b0 : gen_not(b0);

  case DLT_PFLOG:
    return gen_cmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, dir), BPF_B,
        outbound ? PF_OUT : PF_IN);

  case DLT_PPP_PPPD:
    return gen_cmp(cstate, OR_LINKHDR, 0, BPF_B, outbound ? PPP_PPPD_OUT : PPP_PPPD_IN);

  case DLT_JUNIPER_MFR:
  case DLT_JUNIPER_MLFR:
  case DLT_JUNIPER_MLPPP:
  case DLT_JUNIPER_ATM1:
  case DLT_JUNIPER_ATM2:
  case DLT_JUNIPER_PPPOE:
  case DLT_JUNIPER_PPPOE_ATM:
  case DLT_JUNIPER_GGSN:
  case DLT_JUNIPER_ES:
  case DLT_JUNIPER_MONITOR:
  case DLT_JUNIPER_SERVICES:
  case DLT_JUNIPER_ETHER:
  case DLT_JUNIPER_PPP:
  case DLT_JUNIPER_FRELAY:
  case DLT_JUNIPER_CHDLC:
  case DLT_JUNIPER_VP:
  case DLT_JUNIPER_ST:
  case DLT_JUNIPER_ISM:
  case DLT_JUNIPER_VS:
  case DLT_JUNIPER_SRX_E2E:
  case DLT_JUNIPER_FIBRECHANNEL:
  case DLT_JUNIPER_ATM_CEMIC:
    /* juniper flags (including direction) are stored
     * the byte after the 3-byte magic number */
    return gen_mcmp(cstate, OR_LINKHDR, 3, BPF_B, outbound ? 0 : 1, 0x01);

  case DLT_DSA_TAG_BRCM:
    /*
     * This DSA tag encodes the frame direction in the three most
     * significant bits of its first octet: 0b000***** ("egress",
     * switch -> CPU) means "inbound" in libpcap terms and
     * 0b001***** ("ingress", CPU -> switch) means "outbound".
     */
    return gen_mcmp(cstate, OR_LINKHDR, 6 + 6, BPF_B,
                    outbound ? 0x20 : 0x00, 0xe0);

  case DLT_DSA_TAG_DSA:
    /*
     * This DSA tag does not encode the frame direction, but it
     * encodes the frame mode, and some modes imply exactly one
     * direction.  The mode is the two most significant bits of the
     * first octet.  0b00****** ("To_CPU ingress") and 0b10******
     * ("To_Sniffer ingress") mean "inbound" in libpcap terms and
     * 0b01****** ("From_CPU egress") means "outbound".  0x11******
     * ("Forward") can mean either direction, so cannot be used for
     * this purpose.
     *
     * So match 0b01****** for outbound and 0b*0****** otherwise.
     */
    return gen_mcmp(cstate, OR_LINKHDR, 6 + 6, BPF_B,
                    outbound ? 0x40 : 0x00,
                    outbound ? 0xc0 : 0x40);

  default:
    /*
     * If we have packet meta-data indicating a direction,
     * and that metadata can be checked by BPF code, check
     * it.  Otherwise, give up, as this link-layer type has
     * nothing in the packet data.
     *
     * Currently, the only platform where a BPF filter can
     * check that metadata is Linux with the in-kernel
     * BPF interpreter.  If other packet capture mechanisms
     * and BPF filters also supported this, it would be
     * nice.  It would be even better if they made that
     * metadata available so that we could provide it
     * with newer capture APIs, allowing it to be saved
     * in pcapng files.
     */
#if defined(__linux__)
    require_basic_bpf_extensions(cstate, outbound ? "outbound" : "inbound");
    /* match outgoing packets */
    b0 = gen_cmp(cstate, OR_LINKHDR, SKF_AD_OFF + SKF_AD_PKTTYPE, BPF_H,
                 PACKET_OUTGOING);
    // To filter on inbound traffic, invert the match.
    return outbound ? b0 : gen_not(b0);
#else /* defined(__linux__) */
    fail_kw_on_dlt(cstate, outbound ? "outbound" : "inbound");
    /*NOTREACHED*/
#endif /* defined(__linux__) */
  }
}

/* PF firewall log matched interface */
struct block *
gen_pf_ifname(compiler_state_t *cstate, const char *ifname)
{
  u_int len, off;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "ifname");

  len = sizeof(((struct pfloghdr *)0)->ifname);
  off = offsetof(struct pfloghdr, ifname);
  if (strlen(ifname) >= len) {
    bpf_error(cstate, "ifname interface names can only be %d characters",
        len-1);
    /*NOTREACHED*/
  }
  return gen_bcmp(cstate, OR_LINKHDR, off, (u_int)strlen(ifname),
      (const u_char *)ifname);
}

/* PF firewall log ruleset name */
struct block *
gen_pf_ruleset(compiler_state_t *cstate, char *ruleset)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "ruleset");

  if (strlen(ruleset) >= sizeof(((struct pfloghdr *)0)->ruleset)) {
    bpf_error(cstate, "ruleset names can only be %ld characters",
        (long)(sizeof(((struct pfloghdr *)0)->ruleset) - 1));
    /*NOTREACHED*/
  }

  return gen_bcmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, ruleset),
      (u_int)strlen(ruleset), (const u_char *)ruleset);
}

/* PF firewall log rule number */
struct block *
gen_pf_rnr(compiler_state_t *cstate, int rnr)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "rnr");

  return gen_cmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, rulenr), BPF_W,
     (bpf_u_int32)rnr);
}

/* PF firewall log sub-rule number */
struct block *
gen_pf_srnr(compiler_state_t *cstate, int srnr)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "srnr");

  return gen_cmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, subrulenr), BPF_W,
      (bpf_u_int32)srnr);
}

/* PF firewall log reason code */
struct block *
gen_pf_reason(compiler_state_t *cstate, int reason)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "reason");

  return gen_cmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, reason), BPF_B,
      (bpf_u_int32)reason);
}

/* PF firewall log action */
struct block *
gen_pf_action(compiler_state_t *cstate, int action)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_pflog(cstate, "action");

  return gen_cmp(cstate, OR_LINKHDR, offsetof(struct pfloghdr, action), BPF_B,
      (bpf_u_int32)action);
}

/* IEEE 802.11 wireless header */
struct block *
gen_p80211_type(compiler_state_t *cstate, bpf_u_int32 type, bpf_u_int32 mask)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  switch (cstate->linktype) {

  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
  case DLT_PPI:
    return gen_mcmp(cstate, OR_LINKHDR, 0, BPF_B, type, mask);

  default:
    fail_kw_on_dlt(cstate, "type/subtype");
    /*NOTREACHED*/
  }
}

struct block *
gen_p80211_fcdir(compiler_state_t *cstate, bpf_u_int32 fcdir)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  switch (cstate->linktype) {

  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
  case DLT_PPI:
    return gen_mcmp(cstate, OR_LINKHDR, 1, BPF_B, fcdir,
        IEEE80211_FC1_DIR_MASK);

  default:
    fail_kw_on_dlt(cstate, "dir");
    /*NOTREACHED*/
  }
}

static struct block *
gen_vlan_tpid_test(compiler_state_t *cstate)
{
  struct block *b0, *b1;

  /* check for VLAN, including 802.1ad and QinQ */
  b0 = gen_linktype(cstate, ETHERTYPE_8021Q);
  b1 = gen_linktype(cstate, ETHERTYPE_8021AD);
  b0 = gen_or(b0, b1);
  b1 = gen_linktype(cstate, ETHERTYPE_8021QINQ);

  return gen_or(b0, b1);
}

static struct block *
gen_vlan_vid_test(compiler_state_t *cstate, bpf_u_int32 vlan_num)
{
  assert_maxval(cstate, "VLAN tag", vlan_num, 0x0fff);
  return gen_mcmp(cstate, OR_LINKPL, 0, BPF_H, vlan_num, 0x0fff);
}

static struct block *
gen_vlan_no_bpf_extensions(compiler_state_t *cstate, bpf_u_int32 vlan_num,
    int has_vlan_tag)
{
  struct block *b0, *b1;

  b0 = gen_vlan_tpid_test(cstate);

  if (has_vlan_tag) {
    b1 = gen_vlan_vid_test(cstate, vlan_num);
    b0 = gen_and(b0, b1);
  }

  /*
   * Both payload and link header type follow the VLAN tags so that
   * both need to be updated.
   */
  cstate->off_linkpl.constant_part += 4;
  cstate->off_linktype.constant_part += 4;

  return b0;
}

#if defined(SKF_AD_VLAN_TAG_PRESENT)
/* Add v to variable part of off. */
static void
gen_vlan_vloffset_add(compiler_state_t *cstate, bpf_abs_offset *off,
    bpf_u_int32 v, struct slist *s)
{
  struct slist *s2;

  if (!off->is_variable)
    off->is_variable = 1;
  if (off->reg == -1)
    off->reg = alloc_reg(cstate);

  s2 = NEW_STMT_LD_M(cstate, off->reg);
  sappend(s, s2);
  s2 = NEW_STMT_ADD_K(cstate, v);
  sappend(s, s2);
  s2 = NEW_STMT_ST_M(cstate, off->reg);
  sappend(s, s2);
}

/*
 * Patch block b_tpid (VLAN TPID test) to update variable parts of link payload
 * and link type offsets first.
 */
static void
gen_vlan_patch_tpid_test(compiler_state_t *cstate, struct block *b_tpid)
{
  struct slist s;

  /* offset determined at run time, shift variable part */
  s.next = NULL;
  cstate->is_vlan_vloffset = 1;
  gen_vlan_vloffset_add(cstate, &cstate->off_linkpl, 4, &s);
  gen_vlan_vloffset_add(cstate, &cstate->off_linktype, 4, &s);

  /* we get a pointer to a chain of or-ed blocks, patch first of them */
  sprepend_to_block(s.next, b_tpid->head);
}

/*
 * Patch block b_vid (VLAN ID test) to load VID value either from packet
 * metadata (using BPF extensions) if SKF_AD_VLAN_TAG_PRESENT is true.
 */
static void
gen_vlan_patch_vid_test(compiler_state_t *cstate, struct block *b_vid)
{
  struct slist *s, *s2, *sjeq;
  unsigned cnt;

  s = NEW_STMT_LDB_ABS(cstate,
      (bpf_u_int32)(SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT));

  /* true -> next instructions, false -> beginning of b_vid */
  sjeq = NEW_STMT_JEQ_K(cstate, 1);
  sjeq->s.jf = b_vid->stmts;
  sappend(s, sjeq);

  s2 = NEW_STMT_LDH_ABS(cstate,
      (bpf_u_int32)(SKF_AD_OFF + SKF_AD_VLAN_TAG));
  sappend(s, s2);
  sjeq->s.jt = s2;

  /* Jump to the test in b_vid. We need to jump one instruction before
   * the end of the b_vid block so that we only skip loading the TCI
   * from packet data and not the 'and' instruction extracting VID.
   */
  cnt = 0;
  for (s2 = b_vid->stmts; s2; s2 = s2->next)
    cnt++;
  s2 = NEW_STMT_JA(cstate, cnt - 1);
  sappend(s, s2);

  /* insert our statements at the beginning of b_vid */
  sprepend_to_block(s, b_vid);
}

/*
 * Generate check for "vlan" or "vlan <id>" on systems with support for BPF
 * extensions.  Even if kernel supports VLAN BPF extensions, (outermost) VLAN
 * tag can be either in metadata or in packet data; therefore if the
 * SKF_AD_VLAN_TAG_PRESENT test is negative, we need to check link
 * header for VLAN tag.  As the decision is done at run time, we need to
 * update variable part of the offsets.
 */
static struct block *
gen_vlan_bpf_extensions(compiler_state_t *cstate, bpf_u_int32 vlan_num,
    int has_vlan_tag)
{
  struct block *b0, *b_tpid, *b_vid = NULL;
  struct slist *s;

  /* generate new filter code based on extracting packet
   * metadata */
  s = NEW_STMT_LDB_ABS(cstate,
      (bpf_u_int32)(SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT));

  b0 = gen_jmp_k(cstate, BPF_JEQ, 1, s);

  /*
   * This is tricky. We need to insert the statements updating variable
   * parts of offsets before the traditional TPID and VID tests so
   * that they are called whenever SKF_AD_VLAN_TAG_PRESENT fails but
   * we do not want this update to affect those checks. That's why we
   * generate both test blocks first and insert the statements updating
   * variable parts of both offsets after that. This wouldn't work if
   * there already were variable length link header when entering this
   * function but gen_vlan_bpf_extensions() isn't called in that case.
   */
  b_tpid = gen_vlan_tpid_test(cstate);
  if (has_vlan_tag)
    b_vid = gen_vlan_vid_test(cstate, vlan_num);

  gen_vlan_patch_tpid_test(cstate, b_tpid);
  b0 = gen_or(b0, b_tpid);

  if (has_vlan_tag) {
    gen_vlan_patch_vid_test(cstate, b_vid);
    b0 = gen_and(b0, b_vid);
  }

  return b0;
}
#endif

/*
 * Support IEEE 802.1Q VLAN trunk over Ethernet.
 */
struct block *
gen_vlan(compiler_state_t *cstate, bpf_u_int32 vlan_num, int has_vlan_tag)
{
  struct  block *b0;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /* can't check for VLAN-encapsulated packets inside MPLS */
  if (cstate->label_stack_depth > 0)
    bpf_error(cstate, "no VLAN match after MPLS");

  /*
   * Check for a VLAN packet, and then change the offsets to point
   * to the type and data fields within the VLAN packet.  Just
   * increment the offsets, so that we can support a hierarchy, e.g.
   * "vlan 100 && vlan 200" to capture VLAN 200 encapsulated within
   * VLAN 100.
   *
   * XXX - this is a bit of a kludge.  If we were to split the
   * compiler into a parser that parses an expression and
   * generates an expression tree, and a code generator that
   * takes an expression tree (which could come from our
   * parser or from some other parser) and generates BPF code,
   * we could perhaps make the offsets parameters of routines
   * and, in the handler for an "AND" node, pass to subnodes
   * other than the VLAN node the adjusted offsets.
   *
   * This would mean that "vlan" would, instead of changing the
   * behavior of *all* tests after it, change only the behavior
   * of tests ANDed with it.  That would change the documented
   * semantics of "vlan", which might break some expressions.
   * However, it would mean that "(vlan and ip) or ip" would check
   * both for VLAN-encapsulated IP and IP-over-Ethernet, rather than
   * checking only for VLAN-encapsulated IP, so that could still
   * be considered worth doing; it wouldn't break expressions
   * that are of the form "vlan and ..." or "vlan N and ...",
   * which I suspect are the most common expressions involving
   * "vlan".  "vlan or ..." doesn't necessarily do what the user
   * would really want, now, as all the "or ..." tests would
   * be done assuming a VLAN, even though the "or" could be viewed
   * as meaning "or, if this isn't a VLAN packet...".
   */
  switch (cstate->linktype) {

  case DLT_EN10MB:
    /*
     * Newer version of the Linux kernel pass around
     * packets in which the VLAN tag has been removed
     * from the packet data and put into metadata.
     *
     * This requires special treatment.
     */
#if defined(SKF_AD_VLAN_TAG_PRESENT)
    /* Verify that this is the outer part of the packet and
     * not encapsulated somehow. */
    if (cstate->vlan_stack_depth == 0 && !cstate->off_linkhdr.is_variable &&
        cstate->off_linkhdr.constant_part ==
        cstate->off_outermostlinkhdr.constant_part) {
      /*
       * Do we need special VLAN handling?
       */
      if (cstate->bpf_pcap->bpf_codegen_flags & BPF_SPECIAL_VLAN_HANDLING)
        b0 = gen_vlan_bpf_extensions(cstate, vlan_num,
            has_vlan_tag);
      else
        b0 = gen_vlan_no_bpf_extensions(cstate,
            vlan_num, has_vlan_tag);
    } else
#endif
      b0 = gen_vlan_no_bpf_extensions(cstate, vlan_num,
          has_vlan_tag);
    break;

  case DLT_NETANALYZER:
  case DLT_NETANALYZER_TRANSPARENT:
  case DLT_DSA_TAG_BRCM:
  case DLT_DSA_TAG_DSA:
  case DLT_IEEE802_11:
  case DLT_PRISM_HEADER:
  case DLT_IEEE802_11_RADIO_AVS:
  case DLT_IEEE802_11_RADIO:
    /*
     * These are either Ethernet packets with an additional
     * metadata header (the NetAnalyzer types), or 802.11
     * packets, possibly with an additional metadata header.
     *
     * For the first of those, the VLAN tag is in the normal
     * place, so the special-case handling above isn't
     * necessary.
     *
     * For the second of those, we don't do the special-case
     * handling for now.
     */
    b0 = gen_vlan_no_bpf_extensions(cstate, vlan_num, has_vlan_tag);
    break;

  default:
    fail_kw_on_dlt(cstate, "vlan");
    /*NOTREACHED*/
  }

  cstate->vlan_stack_depth++;

  return (b0);
}

/*
 * support for MPLS
 *
 * The label_num_arg dance is to avoid annoying whining by compilers that
 * label_num might be clobbered by longjmp - yeah, it might, but *WHO CARES*?
 * It's not *used* after setjmp returns.
 */
static struct block *
gen_mpls_internal(compiler_state_t *cstate, bpf_u_int32 label_num,
    int has_label_num)
{
  struct  block *b0, *b1;

  if (cstate->label_stack_depth > 0) {
    b0 = gen_not(gen_just_after_mpls_stack(cstate));
  } else {
    /*
     * We're not in an MPLS stack yet, so check the link-layer
     * type against MPLS.
     */
    switch (cstate->linktype) {

    case DLT_C_HDLC: /* fall through */
    case DLT_HDLC:
    case DLT_EN10MB:
    case DLT_NETANALYZER:
    case DLT_NETANALYZER_TRANSPARENT:
    case DLT_DSA_TAG_BRCM:
    case DLT_DSA_TAG_DSA:
      b0 = gen_linktype(cstate, ETHERTYPE_MPLS);
      break;

    case DLT_PPP:
      b0 = gen_linktype(cstate, PPP_MPLS_UCAST);
      break;

      /* FIXME add other DLT_s ...
       * for Frame-Relay/and ATM this may get messy due to SNAP headers
       * leave it for now */

    default:
      fail_kw_on_dlt(cstate, "mpls");
      /*NOTREACHED*/
    }
  }

  /* If a specific MPLS label is requested, check it */
  if (has_label_num) {
    assert_maxval(cstate, "MPLS label", label_num, MPLS_LABEL_MAX);
    b1 = gen_mcmp(cstate, OR_LINKPL, 0, BPF_W,
        label_num << MPLS_LABEL_SHIFT,
        MPLS_LABEL_MAX << MPLS_LABEL_SHIFT);
    b0 = gen_and(b0, b1);
  }

  /*
   * Change the offsets to point to the type and data fields within
   * the MPLS packet.  Just increment the offsets, so that we
   * can support a hierarchy, e.g. "mpls 100000 && mpls 1024" to
   * capture packets with an outer label of 100000 and an inner
   * label of 1024.
   *
   * Increment the MPLS stack depth as well; this indicates that
   * we're checking MPLS-encapsulated headers, to make sure higher
   * level code generators don't try to match against IP-related
   * protocols such as Q_ARP, Q_RARP etc.
   *
   * XXX - this is a bit of a kludge.  See comments in gen_vlan().
   */
  cstate->off_nl_nosnap += MPLS_STACKENTRY_LEN;
  cstate->off_nl += MPLS_STACKENTRY_LEN;
  cstate->label_stack_depth++;
  return (b0);
}

struct block *
gen_mpls(compiler_state_t *cstate, bpf_u_int32 label_num, int has_label_num)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_mpls_internal(cstate, label_num, has_label_num);
}

/*
 * Support PPPoE discovery and session.
 */
struct block *
gen_pppoed(compiler_state_t *cstate)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /* check for PPPoE discovery */
  return gen_linktype(cstate, ETHERTYPE_PPPOED);
}

/*
 * RFC 2516 Section 4:
 *
 * The Ethernet payload for PPPoE is as follows:
 *
 *                      1                   2                   3
 *  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 * |  VER  | TYPE  |      CODE     |          SESSION_ID           |
 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 * |            LENGTH             |           payload             ~
 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 */
struct block *
gen_pppoes(compiler_state_t *cstate, bpf_u_int32 sess_num, int has_sess_num)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * Test against the PPPoE session link-layer type.
   */
  b0 = gen_linktype(cstate, ETHERTYPE_PPPOES);

  /* If a specific session is requested, check PPPoE session id */
  if (has_sess_num) {
    assert_maxval(cstate, "PPPoE session number", sess_num, UINT16_MAX);
    b1 = gen_cmp(cstate, OR_LINKPL, 2, BPF_H, sess_num);
    b0 = gen_and(b0, b1);
  }

  /*
   * Change the offsets to point to the type and data fields within
   * the PPP packet, and note that this is PPPoE rather than
   * raw PPP.
   *
   * XXX - this is a bit of a kludge.  See the comments in
   * gen_vlan().
   *
   * The "network-layer" protocol is PPPoE, which has a 6-byte
   * PPPoE header, followed by a PPP packet.
   *
   * There is no HDLC encapsulation for the PPP packet (it's
   * encapsulated in PPPoES instead), so the link-layer type
   * starts at the first byte of the PPP packet.  For PPPoE,
   * that offset is relative to the beginning of the total
   * link-layer payload, including any 802.2 LLC header, so
   * it's 6 bytes past cstate->off_nl.
   */
  PUSH_LINKHDR(cstate, DLT_PPP, cstate->off_linkpl.is_variable,
      cstate->off_linkpl.constant_part + cstate->off_nl + 6, /* 6 bytes past the PPPoE header */
      cstate->off_linkpl.reg);

  cstate->off_linktype = cstate->off_linkhdr;
  cstate->off_linkpl.constant_part = cstate->off_linkhdr.constant_part + 2;

  cstate->off_nl = 0;
  cstate->off_nl_nosnap = 0;  /* no 802.2 LLC */

  return b0;
}

/* Check that this is Geneve and the VNI is correct if
 * specified. Parameterized to handle both IPv4 and IPv6. */
static struct block *
gen_geneve_check(compiler_state_t *cstate,
    struct block *(*gen_portfn)(compiler_state_t *, const uint16_t, const int, const u_char, const u_char),
    enum e_offrel offrel, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;

  b0 = gen_portfn(cstate, GENEVE_PORT, IPPROTO_UDP, Q_DST, Q_PORT);

  /* Check that we are operating on version 0. Otherwise, we
   * can't decode the rest of the fields. The version is 2 bits
   * in the first byte of the Geneve header. */
  b1 = gen_mcmp(cstate, offrel, 8, BPF_B, 0, 0xc0);
  b0 = gen_and(b0, b1);

  if (has_vni) {
    assert_maxval(cstate, "Geneve VNI", vni, 0xffffff);
    vni <<= 8; /* VNI is in the upper 3 bytes */
    b1 = gen_mcmp(cstate, offrel, 12, BPF_W, vni, 0xffffff00);
    b0 = gen_and(b0, b1);
  }

  return b0;
}

/* The IPv4 and IPv6 Geneve checks need to do two things:
 * - Verify that this actually is Geneve with the right VNI.
 * - Place the IP header length (plus variable link prefix if
 *   needed) into register A to be used later to compute
 *   the inner packet offsets. */
static struct block *
gen_geneve4(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;
  struct slist *s, *s1;

  b0 = gen_geneve_check(cstate, gen_port, OR_TRAN_IPV4, vni, has_vni);

  /* Load the IP header length into A. */
  s = gen_loadx_iphdrlen(cstate);

  s1 = NEW_STMT_TXA(cstate);
  sappend(s, s1);

  /* Forcibly append these statements to the true condition
   * of the protocol check by creating a new block that is
   * always true and ANDing them. */
  b1 = gen_jmp_x(cstate, BPF_JEQ, s);

  return gen_and(b0, b1);
}

static struct block *
gen_geneve6(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;
  struct slist *s, *s1;

  b0 = gen_geneve_check(cstate, gen_port6, OR_TRAN_IPV6, vni, has_vni);

  /* Load the IP header length. We need to account for a
   * variable length link prefix if there is one. */
  s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
  if (s) {
    s1 = NEW_STMT_LD_IMM(cstate, IP6_HDRLEN);
    sappend(s, s1);

    s1 = NEW_STMT_ADD_X(cstate);
    sappend(s, s1);
  } else {
    s = NEW_STMT_LD_IMM(cstate, IP6_HDRLEN);
  }

  /* Forcibly append these statements to the true condition
   * of the protocol check by creating a new block that is
   * always true and ANDing them. */
  s1 = NEW_STMT_TAX(cstate);
  sappend(s, s1);

  b1 = gen_jmp_x(cstate, BPF_JEQ, s);

  return gen_and(b0, b1);
}

/* We need to store three values based on the Geneve header::
 * - The offset of the linktype.
 * - The offset of the end of the Geneve header.
 * - The offset of the end of the encapsulated MAC header. */
static struct slist *
gen_geneve_offsets(compiler_state_t *cstate)
{
  struct slist *s, *s1, *s_proto;

  /* First we need to calculate the offset of the Geneve header
   * itself. This is composed of the IP header previously calculated
   * (include any variable link prefix) and stored in A plus the
   * fixed sized headers (fixed link prefix, MAC length, and UDP
   * header). */
  s = NEW_STMT_ADD_K(cstate,
      cstate->off_linkpl.constant_part + cstate->off_nl + 8);

  /* Stash this in X since we'll need it later. */
  s1 = NEW_STMT_TAX(cstate);
  sappend(s, s1);

  /* The EtherType in Geneve is 2 bytes in. Calculate this and
   * store it. */
  s1 = NEW_STMT_ADD_K(cstate, 2);
  sappend(s, s1);

  cstate->off_linktype.reg = alloc_reg(cstate);
  cstate->off_linktype.is_variable = 1;
  cstate->off_linktype.constant_part = 0;

  s1 = NEW_STMT_ST_M(cstate, cstate->off_linktype.reg);
  sappend(s, s1);

  /* Load the Geneve option length and mask and shift to get the
   * number of bytes. It is stored in the first byte of the Geneve
   * header. */
  s1 = NEW_STMT_LDB_IND(cstate, 0);
  sappend(s, s1);

  s1 = NEW_STMT_AND_K(cstate, 0x3f);
  sappend(s, s1);

  s1 = NEW_STMT_MUL_K(cstate, 4);
  sappend(s, s1);

  /* Add in the rest of the Geneve base header. */
  s1 = NEW_STMT_ADD_K(cstate, 8);
  sappend(s, s1);

  /* Add the Geneve header length to its offset and store. */
  s1 = NEW_STMT_ADD_X(cstate);
  sappend(s, s1);

  /* Set the encapsulated type as Ethernet. Even though we may
   * not actually have Ethernet inside there are two reasons this
   * is useful:
   * - The linktype field is always in EtherType format regardless
   *   of whether it is in Geneve or an inner Ethernet frame.
   * - The only link layer that we have specific support for is
   *   Ethernet. We will confirm that the packet actually is
   *   Ethernet at runtime before executing these checks. */
  PUSH_LINKHDR(cstate, DLT_EN10MB, 1, 0, alloc_reg(cstate));

  s1 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
  sappend(s, s1);

  /* Calculate whether we have an Ethernet header or just raw IP/
   * MPLS/etc. If we have Ethernet, advance the end of the MAC offset
   * and linktype by 14 bytes so that the network header can be found
   * seamlessly. Otherwise, keep what we've calculated already. */

  /* We have a bare jmp so we can't use the optimizer. */
  cstate->no_optimize = 1;

  /* Load the EtherType in the Geneve header, 2 bytes in. */
  s1 = NEW_STMT_LDH_IND(cstate, 2);
  sappend(s, s1);

  /* Load X with the end of the Geneve header. */
  s1 = NEW_STMT_LDX_M(cstate, cstate->off_linkhdr.reg);
  sappend(s, s1);

  /* Check if the EtherType is Transparent Ethernet Bridging. At the
   * end of this check, we should have the total length in X. In
   * the non-Ethernet case, it's already there. */
  s_proto = NEW_STMT_JEQ_K(cstate, ETHERTYPE_TEB);
  sappend(s, s_proto);

  s1 = NEW_STMT_TXA(cstate);
  sappend(s, s1);
  s_proto->s.jt = s1;

  /* Since this is Ethernet, use the EtherType of the payload
   * directly as the linktype. Overwrite what we already have. */
  s1 = NEW_STMT_ADD_K(cstate, 12);
  sappend(s, s1);

  s1 = NEW_STMT_ST_M(cstate, cstate->off_linktype.reg);
  sappend(s, s1);

  /* Advance two bytes further to get the end of the Ethernet
   * header. */
  s1 = NEW_STMT_ADD_K(cstate, 2);
  sappend(s, s1);

  /* Move the result to X. */
  s1 = NEW_STMT_TAX(cstate);
  sappend(s, s1);

  /* Store the final result of our linkpl calculation. */
  cstate->off_linkpl.reg = alloc_reg(cstate);
  cstate->off_linkpl.is_variable = 1;
  cstate->off_linkpl.constant_part = 0;

  s1 = NEW_STMT_STX_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s1);
  s_proto->s.jf = s1;

  cstate->off_nl = 0;

  return s;
}

/* Check to see if this is a Geneve packet. */
struct block *
gen_geneve(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  b0 = gen_geneve4(cstate, vni, has_vni);
  b1 = gen_geneve6(cstate, vni, has_vni);

  /* Later filters should act on the payload of the Geneve frame,
   * update all of the header pointers. Attach this code so that
   * it gets executed in the event that the Geneve filter matches. */
  struct block *offsets =
    sprepend_to_block(gen_geneve_offsets(cstate), gen_true(cstate));

  cstate->is_encap = 1;

  return gen_and(gen_or(b0, b1), offsets);
}

/* Check that this is VXLAN and the VNI is correct if
 * specified. Parameterized to handle both IPv4 and IPv6. */
static struct block *
gen_vxlan_check(compiler_state_t *cstate,
    struct block *(*gen_portfn)(compiler_state_t *, const uint16_t, const int, const u_char, const u_char),
    enum e_offrel offrel, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;

  b0 = gen_portfn(cstate, VXLAN_PORT, IPPROTO_UDP, Q_DST, Q_PORT);

  /* Check that the VXLAN header has the flag bits set
   * correctly. */
  b1 = gen_cmp(cstate, offrel, 8, BPF_B, 0x08);
  b0 = gen_and(b0, b1);

  if (has_vni) {
    assert_maxval(cstate, "VXLAN VNI", vni, 0xffffff);
    vni <<= 8; /* VNI is in the upper 3 bytes */
    b1 = gen_mcmp(cstate, offrel, 12, BPF_W, vni, 0xffffff00);
    b0 = gen_and(b0, b1);
  }

  return b0;
}

/* The IPv4 and IPv6 VXLAN checks need to do two things:
 * - Verify that this actually is VXLAN with the right VNI.
 * - Place the IP header length (plus variable link prefix if
 *   needed) into register A to be used later to compute
 *   the inner packet offsets. */
static struct block *
gen_vxlan4(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;
  struct slist *s, *s1;

  b0 = gen_vxlan_check(cstate, gen_port, OR_TRAN_IPV4, vni, has_vni);

  /* Load the IP header length into A. */
  s = gen_loadx_iphdrlen(cstate);

  s1 = NEW_STMT_TXA(cstate);
  sappend(s, s1);

  /* Forcibly append these statements to the true condition
   * of the protocol check by creating a new block that is
   * always true and ANDing them. */
  b1 = gen_jmp_x(cstate, BPF_JEQ, s);

  return gen_and(b0, b1);
}

static struct block *
gen_vxlan6(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;
  struct slist *s, *s1;

  b0 = gen_vxlan_check(cstate, gen_port6, OR_TRAN_IPV6, vni, has_vni);

  /* Load the IP header length. We need to account for a
   * variable length link prefix if there is one. */
  s = gen_abs_offset_varpart(cstate, &cstate->off_linkpl);
  if (s) {
    s1 = NEW_STMT_LD_IMM(cstate, IP6_HDRLEN);
    sappend(s, s1);

    s1 = NEW_STMT_ADD_X(cstate);
    sappend(s, s1);
  } else {
    s = NEW_STMT_LD_IMM(cstate, IP6_HDRLEN);
  }

  /* Forcibly append these statements to the true condition
   * of the protocol check by creating a new block that is
   * always true and ANDing them. */
  s1 = NEW_STMT_TAX(cstate);
  sappend(s, s1);

  b1 = gen_jmp_x(cstate, BPF_JEQ, s);

  return gen_and(b0, b1);
}

/* We need to store three values based on the VXLAN header:
 * - The offset of the linktype.
 * - The offset of the end of the VXLAN header.
 * - The offset of the end of the encapsulated MAC header. */
static struct slist *
gen_vxlan_offsets(compiler_state_t *cstate)
{
  struct slist *s, *s1;

  /* Calculate the offset of the VXLAN header itself. This
   * includes the IP header computed previously (including any
   * variable link prefix) and stored in A plus the fixed size
   * headers (fixed link prefix, MAC length, UDP header). */
  s = NEW_STMT_ADD_K(cstate,
      cstate->off_linkpl.constant_part + cstate->off_nl + 8);

  /* Add the VXLAN header length to its offset and store */
  s1 = NEW_STMT_ADD_K(cstate, 8);
  sappend(s, s1);

  /* Push the link header. VXLAN packets always contain Ethernet
   * frames. */
  PUSH_LINKHDR(cstate, DLT_EN10MB, 1, 0, alloc_reg(cstate));

  s1 = NEW_STMT_ST_M(cstate, cstate->off_linkhdr.reg);
  sappend(s, s1);

  /* As the payload is an Ethernet packet, we can use the
   * EtherType of the payload directly as the linktype. */
  s1 = NEW_STMT_ADD_K(cstate, 12);
  sappend(s, s1);

  cstate->off_linktype.reg = alloc_reg(cstate);
  cstate->off_linktype.is_variable = 1;
  cstate->off_linktype.constant_part = 0;

  s1 = NEW_STMT_ST_M(cstate, cstate->off_linktype.reg);
  sappend(s, s1);

  /* Two bytes further is the end of the Ethernet header and the
   * start of the payload. */
  s1 = NEW_STMT_ADD_K(cstate, 2);
  sappend(s, s1);

  /* Move the result to X. */
  s1 = NEW_STMT_TAX(cstate);
  sappend(s, s1);

  /* Store the final result of our linkpl calculation. */
  cstate->off_linkpl.reg = alloc_reg(cstate);
  cstate->off_linkpl.is_variable = 1;
  cstate->off_linkpl.constant_part = 0;

  s1 = NEW_STMT_STX_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s1);

  cstate->off_nl = 0;

  return s;
}

/* Check to see if this is a VXLAN packet. */
struct block *
gen_vxlan(compiler_state_t *cstate, bpf_u_int32 vni, int has_vni)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  /*
   * This code generates blocks that are known to trigger a bug in the
   * optimizer (see GitHub 1670), so avoid the bug until it gets fixed.
   */
  cstate->no_optimize = 1;

  b0 = gen_vxlan4(cstate, vni, has_vni);
  b1 = gen_vxlan6(cstate, vni, has_vni);

  /* Later filters should act on the payload of the VXLAN frame,
   * update all of the header pointers. Attach this code so that
   * it gets executed in the event that the VXLAN filter matches. */
  struct block *offsets =
    sprepend_to_block(gen_vxlan_offsets(cstate), gen_true(cstate));

  cstate->is_encap = 1;

  return gen_and(gen_or(b0, b1), offsets);
}

/* Check that the encapsulated frame has a link layer header
 * for Ethernet filters. */
static struct block *
gen_encap_ll_check(compiler_state_t *cstate)
{
  struct block *b0;
  struct slist *s, *s1;

  /* The easiest way to see if there is a link layer present
   * is to check if the link layer header and payload are not
   * the same. */

  /* Geneve always generates pure variable offsets so we can
   * compare only the registers. */
  s = NEW_STMT_LD_M(cstate, cstate->off_linkhdr.reg);
  s1 = NEW_STMT_LDX_M(cstate, cstate->off_linkpl.reg);
  sappend(s, s1);

  b0 = gen_jmp_x(cstate, BPF_JEQ, s);

  return gen_not(b0);
}

static struct block *
gen_atmfield_code_internal(compiler_state_t *cstate, int atmfield,
    bpf_u_int32 jvalue, int jtype, int reverse)
{
  assert_atm(cstate, atmkw(atmfield));

  switch (atmfield) {

  case A_VPI:
    assert_maxval(cstate, "VPI", jvalue, UINT8_MAX);
    return gen_ncmp(cstate, OR_LINKHDR, cstate->off_vpi, BPF_B,
        0xffffffffU, jtype, reverse, jvalue);

  case A_VCI:
    assert_maxval(cstate, "VCI", jvalue, UINT16_MAX);
    return gen_ncmp(cstate, OR_LINKHDR, cstate->off_vci, BPF_H,
        0xffffffffU, jtype, reverse, jvalue);

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "atmfield", atmfield);
  }
}

static struct block *
gen_atm_vpi(compiler_state_t *cstate, const uint8_t v)
{
  return gen_atmfield_code_internal(cstate, A_VPI, v, BPF_JEQ, 0);
}

static struct block *
gen_atm_vci(compiler_state_t *cstate, const uint16_t v)
{
  return gen_atmfield_code_internal(cstate, A_VCI, v, BPF_JEQ, 0);
}

static struct block *
gen_atm_prototype(compiler_state_t *cstate, const uint8_t v)
{
  return gen_mcmp(cstate, OR_LINKHDR, cstate->off_proto, BPF_B, v, 0x0fU);
}

static struct block *
gen_atmtype_llc(compiler_state_t *cstate)
{
  struct block *b0;

  b0 = gen_atm_prototype(cstate, PT_LLC);
  cstate->linktype = cstate->prevlinktype;
  return b0;
}

struct block *
gen_atmfield_code(compiler_state_t *cstate, int atmfield,
    bpf_u_int32 jvalue, int jtype, int reverse)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_atmfield_code_internal(cstate, atmfield, jvalue, jtype,
      reverse);
}

struct block *
gen_atmtype_abbrev(compiler_state_t *cstate, int type)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_atm(cstate, atmkw(type));

  switch (type) {

  case A_METAC:
    /* Get all packets in Meta signalling Circuit */
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 1);
    return gen_and(b0, b1);

  case A_BCC:
    /* Get all packets in Broadcast Circuit*/
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 2);
    return gen_and(b0, b1);

  case A_OAMF4SC:
    /* Get all cells in Segment OAM F4 circuit*/
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 3);
    return gen_and(b0, b1);

  case A_OAMF4EC:
    /* Get all cells in End-to-End OAM F4 Circuit*/
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 4);
    return gen_and(b0, b1);

  case A_SC:
    /*  Get all packets in connection Signalling Circuit */
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 5);
    return gen_and(b0, b1);

  case A_ILMIC:
    /* Get all packets in ILMI Circuit */
    b0 = gen_atm_vpi(cstate, 0);
    b1 = gen_atm_vci(cstate, 16);
    return gen_and(b0, b1);

  case A_LANE:
    /* Get all LANE packets */
    b1 = gen_atm_prototype(cstate, PT_LANE);

    /*
     * Arrange that all subsequent tests assume LANE
     * rather than LLC-encapsulated packets, and set
     * the offsets appropriately for LANE-encapsulated
     * Ethernet.
     *
     * We assume LANE means Ethernet, not Token Ring.
     */
    PUSH_LINKHDR(cstate, DLT_EN10MB, 0,
        cstate->off_payload + 2,  /* Ethernet header */
        -1);
    cstate->off_linktype.constant_part = cstate->off_linkhdr.constant_part + 12;
    cstate->off_linkpl.constant_part = cstate->off_linkhdr.constant_part + 14;  /* Ethernet */
    cstate->off_nl = 0;     /* Ethernet II */
    cstate->off_nl_nosnap = 3;    /* 802.3+802.2 */
    return b1;

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "type", type);
  }
}

/*
 * Filtering for MTP2 messages based on li value
 * FISU, length is null
 * LSSU, length is 1 or 2
 * MSU, length is 3 or more
 * For MTP2_HSL, sequences are on 2 bytes, and length on 9 bits
 */
struct block *
gen_mtp2type_abbrev(compiler_state_t *cstate, int type)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_ss7(cstate, ss7kw(type));

  switch (type) {

  case M_FISU:
    return gen_ncmp(cstate, OR_PACKET, cstate->off_li, BPF_B,
        0x3fU, BPF_JEQ, 0, 0U);

  case M_LSSU:
    b0 = gen_ncmp(cstate, OR_PACKET, cstate->off_li, BPF_B,
        0x3fU, BPF_JGT, 1, 2U);
    b1 = gen_ncmp(cstate, OR_PACKET, cstate->off_li, BPF_B,
        0x3fU, BPF_JGT, 0, 0U);
    return gen_and(b1, b0);

  case M_MSU:
    return gen_ncmp(cstate, OR_PACKET, cstate->off_li, BPF_B,
        0x3fU, BPF_JGT, 0, 2U);

  case MH_FISU:
    return gen_ncmp(cstate, OR_PACKET, cstate->off_li_hsl, BPF_H,
        0xff80U, BPF_JEQ, 0, 0U);

  case MH_LSSU:
    b0 = gen_ncmp(cstate, OR_PACKET, cstate->off_li_hsl, BPF_H,
        0xff80U, BPF_JGT, 1, 0x0100U);
    b1 = gen_ncmp(cstate, OR_PACKET, cstate->off_li_hsl, BPF_H,
        0xff80U, BPF_JGT, 0, 0U);
    return gen_and(b1, b0);

  case MH_MSU:
    return gen_ncmp(cstate, OR_PACKET, cstate->off_li_hsl, BPF_H,
        0xff80U, BPF_JGT, 0, 0x0100U);

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "type", type);
  }
}

/*
 * These maximum valid values are all-ones, so they double as the bitmasks
 * before any bitwise shifting.
 */
#define MTP2_SIO_MAXVAL UINT8_MAX
#define MTP3_PC_MAXVAL 0x3fffU
#define MTP3_SLS_MAXVAL 0xfU

static struct block *
gen_mtp3field_code_internal(compiler_state_t *cstate, int mtp3field,
    bpf_u_int32 jvalue, int jtype, int reverse)
{
  u_int newoff_sio;
  u_int newoff_opc;
  u_int newoff_dpc;
  u_int newoff_sls;

  newoff_sio = cstate->off_sio;
  newoff_opc = cstate->off_opc;
  newoff_dpc = cstate->off_dpc;
  newoff_sls = cstate->off_sls;

  assert_ss7(cstate, ss7kw(mtp3field));

  switch (mtp3field) {

  /*
   * See UTU-T Rec. Q.703, Section 2.2, Figure 3/Q.703.
   *
   * SIO is the simplest field: the size is one byte and the offset is a
   * multiple of bytes, so the only detail to get right is the value of
   * the [right-to-left] field offset.
   */
  case MH_SIO:
    newoff_sio += 3; /* offset for MTP2_HSL */
    /* FALLTHROUGH */

  case M_SIO:
    assert_maxval(cstate, ss7kw(mtp3field), jvalue, MTP2_SIO_MAXVAL);
    // Here the bitmask means "do not apply a bitmask".
    return gen_ncmp(cstate, OR_PACKET, newoff_sio, BPF_B, UINT32_MAX,
        jtype, reverse, jvalue);

  /*
   * See UTU-T Rec. Q.704, Section 2.2, Figure 3/Q.704.
   *
   * SLS, OPC and DPC are more complicated: none of these is sized in a
   * multiple of 8 bits, MTP3 encoding is little-endian and MTP packet
   * diagrams are meant to be read right-to-left.  This means in the
   * diagrams within individual fields and concatenations thereof
   * bitwise shifts and masks can be noted in the common left-to-right
   * manner until each final value is ready to be byte-swapped and
   * handed to gen_ncmp().  See also gen_dnhostop(), which solves a
   * similar problem in a similar way.
   *
   * Offsets of fields within the packet header always have the
   * right-to-left meaning.  Note that in DLT_MTP2 and possibly other
   * DLTs the offset does not include the F (Flag) field at the
   * beginning of each message.
   *
   * For example, if the 8-bit SIO field has a 3 byte [RTL] offset, the
   * 32-bit standard routing header has a 4 byte [RTL] offset and could
   * be tested entirely using a single BPF_W comparison.  In this case
   * the 14-bit DPC field [LTR] bitmask would be 0x3FFF, the 14-bit OPC
   * field [LTR] bitmask would be (0x3FFF << 14) and the 4-bit SLS field
   * [LTR] bitmask would be (0xF << 28), all of which conveniently
   * correlates with the [RTL] packet diagram until the byte-swapping is
   * done before use.
   *
   * The code below uses this approach for OPC, which spans 3 bytes.
   * DPC and SLS use shorter loads, SLS also uses a different offset.
   */
  case MH_OPC:
    newoff_opc += 3;

    /* FALLTHROUGH */
  case M_OPC:
    assert_maxval(cstate, ss7kw(mtp3field), jvalue, MTP3_PC_MAXVAL);
    return gen_ncmp(cstate, OR_PACKET, newoff_opc, BPF_W,
        PCAP_BSWAP_32(MTP3_PC_MAXVAL << 14), jtype, reverse,
        PCAP_BSWAP_32(jvalue << 14));

  case MH_DPC:
    newoff_dpc += 3;
    /* FALLTHROUGH */

  case M_DPC:
    assert_maxval(cstate, ss7kw(mtp3field), jvalue, MTP3_PC_MAXVAL);
    return gen_ncmp(cstate, OR_PACKET, newoff_dpc, BPF_H,
        PCAP_BSWAP_16(MTP3_PC_MAXVAL), jtype, reverse,
        PCAP_BSWAP_16(jvalue));

  case MH_SLS:
    newoff_sls += 3;
    /* FALLTHROUGH */

  case M_SLS:
    assert_maxval(cstate, ss7kw(mtp3field), jvalue, MTP3_SLS_MAXVAL);
    return gen_ncmp(cstate, OR_PACKET, newoff_sls, BPF_B,
        MTP3_SLS_MAXVAL << 4, jtype, reverse,
        jvalue << 4);

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "mtp3field", mtp3field);
  }
}

struct block *
gen_mtp3field_code(compiler_state_t *cstate, int mtp3field,
    bpf_u_int32 jvalue, int jtype, int reverse)
{
  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  return gen_mtp3field_code_internal(cstate, mtp3field, jvalue, jtype,
      reverse);
}

static struct block *
gen_msg_abbrev(compiler_state_t *cstate, const uint8_t type)
{
  /*
   * Q.2931 signalling protocol messages for handling virtual circuits
   * establishment and teardown
   */
  return gen_cmp(cstate, OR_LINKHDR, cstate->off_payload + MSG_TYPE_POS,
      BPF_B, type);
}

struct block *
gen_atmmulti_abbrev(compiler_state_t *cstate, int type)
{
  struct block *b0, *b1;

  /*
   * Catch errors reported by us and routines below us, and return NULL
   * on an error.
   */
  if (setjmp(cstate->top_ctx))
    return (NULL);

  assert_atm(cstate, atmkw(type));

  switch (type) {

  case A_OAM:
    /* OAM F4 type */
    b0 = gen_atm_vci(cstate, 3);
    b1 = gen_atm_vci(cstate, 4);
    b1 = gen_or(b0, b1);
    b0 = gen_atm_vpi(cstate, 0);
    return gen_and(b0, b1);

  case A_OAMF4:
    /* OAM F4 type */
    b0 = gen_atm_vci(cstate, 3);
    b1 = gen_atm_vci(cstate, 4);
    b1 = gen_or(b0, b1);
    b0 = gen_atm_vpi(cstate, 0);
    return gen_and(b0, b1);

  case A_CONNECTMSG:
    /*
     * Get Q.2931 signalling messages for switched
     * virtual connection
     */
    b0 = gen_msg_abbrev(cstate, SETUP);
    b1 = gen_msg_abbrev(cstate, CALL_PROCEED);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, CONNECT);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, CONNECT_ACK);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, RELEASE);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, RELEASE_DONE);
    b1 = gen_or(b0, b1);
    b0 = gen_atmtype_abbrev(cstate, A_SC);
    return gen_and(b0, b1);

  case A_METACONNECT:
    b0 = gen_msg_abbrev(cstate, SETUP);
    b1 = gen_msg_abbrev(cstate, CALL_PROCEED);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, CONNECT);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, RELEASE);
    b1 = gen_or(b0, b1);
    b0 = gen_msg_abbrev(cstate, RELEASE_DONE);
    b1 = gen_or(b0, b1);
    b0 = gen_atmtype_abbrev(cstate, A_METAC);
    return gen_and(b0, b1);

  default:
    bpf_error(cstate, ERRSTR_FUNC_VAR_INT, __func__, "type", type);
  }
}

Coverage Report

Created: 2026-10-07 07:40