/[pcre]/code/trunk/pcre_jit_compile.c

Diff of /code/trunk/pcre_jit_compile.c

Parent Directory | Revision Log | View Patch Patch

-revision 850 by zherczeg,
Wed Jan  4 17:29:11 2012 UTC
+revision 1275 by zherczeg,
Sun Mar 10 05:32:10 2013 UTC
 Line 6
  and semantics are as close as possible to those of the Perl 5 language.
                         Written by Philip Hazel
-            Copyright (c) 1997-2012 University of Cambridge
+            Copyright (c) 1997-2013 University of Cambridge
    The machine code generator part (this module) was written by Zoltan Herczeg
-                       Copyright (c) 2010-2012
+                       Copyright (c) 2010-2013
  -----------------------------------------------------------------------------
  Redistribution and use in source and binary forms, with or without
 Line 46 
 POSSIBILITY OF SUCH DAMAGE.
  #include "pcre_internal.h"
- #ifdef SUPPORT_JIT
+ #if defined SUPPORT_JIT
  /* All-in-one: Since we use the JIT compiler only from here,
  we just include it. This way we don't need to touch the build
 Line 65 
 system files. */
  #error Unsupported architecture
  #endif
- /* Allocate memory on the stack. Fast, but limited size. */
+ /* Defines for debugging purposes. */
- #define LOCAL_SPACE_SIZE 32768
+ /* 1 - Use unoptimized capturing brackets.
+- Enable capture_last_ptr (includes option 1). */
+ /* #define DEBUG_FORCE_UNOPTIMIZED_CBRAS 2 */
+ /* 1 - Always have a control head. */
+ /* #define DEBUG_FORCE_CONTROL_HEAD 1 */
+ /* Allocate memory for the regex stack on the real machine stack.
+ Fast, but limited size. */
+ #define MACHINE_STACK_SIZE 32768
+ /* Growth rate for stack allocated by the OS. Should be the multiply
+ of page size. */
  #define STACK_GROWTH_RATE 8192
  /* Enable to check that the allocation could destroy temporaries. */
-Line 82 
 The code generator follows the recursive
+Line 94 
 The code generator follows the recursive
  expressions. The basic blocks of regular expressions are condition checkers
  whose execute different commands depending on the result of the condition check.
  The relationship between the operators can be horizontal (concatenation) and
- vertical (sub-expression) (See struct fallback_common for more details).
+ vertical (sub-expression) (See struct backtrack_common for more details).
    'ab' - 'a' and 'b' regexps are concatenated
    'a+' - 'a' is the sub-expression of the '+' operator
  The condition checkers are boolean (true/false) checkers. Machine code is generated
  for the checker itself and for the actions depending on the result of the checker.
- The 'true' case is called as the hot path (expected path), and the other is called as
+ The 'true' case is called as the matching path (expected path), and the other is called as
- the 'fallback' path. Branch instructions are expesive for all CPUs, so we avoid taken
+ the 'backtrack' path. Branch instructions are expesive for all CPUs, so we avoid taken
- branches on the hot path.
+ branches on the matching path.
   Greedy star operator (*) :
-    Hot path: match happens.
+    Matching path: match happens.
-    Fallback path: match failed.
+    Backtrack path: match failed.
   Non-greedy star operator (*?) :
-    Hot path: no need to perform a match.
+    Matching path: no need to perform a match.
-    Fallback path: match is required.
+    Backtrack path: match is required.
  The following example shows how the code generated for a capturing bracket
  with two alternatives. Let A, B, C, D are arbirary regular expressions, and
-Line 108 
 we have the following regular expression
+Line 120 
 we have the following regular expression
  The generated code will be the following:
-  A hot path
+  A matching path
-  '(' hot path (pushing arguments to the stack)
+  '(' matching path (pushing arguments to the stack)
-  B hot path
+  B matching path
-  ')' hot path (pushing arguments to the stack)
+  ')' matching path (pushing arguments to the stack)
-  D hot path
+  D matching path
   return with successful match
-  D fallback path
+  D backtrack path
-  ')' fallback path (If we arrived from "C" jump to the fallback of "C")
+  ')' backtrack path (If we arrived from "C" jump to the backtrack of "C")
-  B fallback path
+  B backtrack path
   C expected path
-  jump to D hot path
+  jump to D matching path
-  C fallback path
+  C backtrack path
-  A fallback path
+  A backtrack path
-  Notice, that the order of fallback code paths are the opposite of the fast
+  Notice, that the order of backtrack code paths are the opposite of the fast
   code paths. In this way the topmost value on the stack is always belong
-  to the current fallback code path. The fallback code path must check
+  to the current backtrack code path. The backtrack path must check
   whether there is a next alternative. If so, it needs to jump back to
-  the hot path eventually. Otherwise it needs to clear out its own stack
+  the matching path eventually. Otherwise it needs to clear out its own stack
-  frame and continue the execution on the fallback code paths.
+  frame and continue the execution on the backtrack code paths.
  */
  /*
  Saved stack frames:
- Atomic blocks and asserts require reloading the values of local variables
+ Atomic blocks and asserts require reloading the values of private data
- when the fallback mechanism performed. Because of OP_RECURSE, the locals
+ when the backtrack mechanism performed. Because of OP_RECURSE, the data
  are not necessarly known in compile time, thus we need a dynamic restore
  mechanism.
  The stack frames are stored in a chain list, and have the following format:
  ([ capturing bracket offset ][ start value ][ end value ])+ ... [ 0 ] [ previous head ]
- Thus we can restore the locals to a particular point in the stack.
+ Thus we can restore the private data to a particular point in the stack.
  */
  typedef struct jit_arguments {
-Line 152 
 typedef struct jit_arguments {
+Line 164 
 typedef struct jit_arguments {
    const pcre_uchar *begin;
    const pcre_uchar *end;
    int *offsets;
-   pcre_uchar *ptr;
+   pcre_uchar *uchar_ptr;
+   pcre_uchar *mark_ptr;
+   void *callout_data;
    /* Everything else after. */
-   int offsetcount;
+   int real_offset_count;
-   int calllimit;
+   int offset_count;
+   int call_limit;
    pcre_uint8 notbol;
    pcre_uint8 noteol;
    pcre_uint8 notempty;
    pcre_uint8 notempty_atstart;
  } jit_arguments;
- typedef struct executable_function {
+ typedef struct executable_functions {
-   void *executable_func;
+   void *executable_funcs[JIT_NUMBER_OF_COMPILE_MODES];
-   pcre_jit_callback callback;
+   PUBL(jit_callback) callback;
    void *userdata;
-   sljit_uw executable_size;
+   pcre_uint32 top_bracket;
- } executable_function;
+   sljit_uw executable_sizes[JIT_NUMBER_OF_COMPILE_MODES];
+ } executable_functions;
  typedef struct jump_list {
    struct sljit_jump *jump;
    struct jump_list *next;
  } jump_list;
- enum stub_types { stack_alloc };
  typedef struct stub_list {
-   enum stub_types type;
-   int data;
    struct sljit_jump *start;
-   struct sljit_label *leave;
+   struct sljit_label *quit;
    struct stub_list *next;
  } stub_list;
+ enum frame_types {
+   no_frame = -1,
+   no_stack = -2
+ };
+ enum control_types {
+   type_commit = 0,
+   type_prune = 1,
+   type_skip = 2
+ };
  typedef int (SLJIT_CALL *jit_function)(jit_arguments *args);
  /* The following structure is the key data type for the recursive
- code generator. It is allocated by compile_hotpath, and contains
+ code generator. It is allocated by compile_matchingpath, and contains
- the aguments for compile_fallbackpath. Must be the first member
+ the aguments for compile_backtrackingpath. Must be the first member
  of its descendants. */
- typedef struct fallback_common {
+ typedef struct backtrack_common {
    /* Concatenation stack. */
-   struct fallback_common *prev;
+   struct backtrack_common *prev;
-   jump_list *nextfallbacks;
+   jump_list *nextbacktracks;
    /* Internal stack (for component operators). */
-   struct fallback_common *top;
+   struct backtrack_common *top;
-   jump_list *topfallbacks;
+   jump_list *topbacktracks;
    /* Opcode pointer. */
    pcre_uchar *cc;
- } fallback_common;
+ } backtrack_common;
- typedef struct assert_fallback {
+ typedef struct assert_backtrack {
-   fallback_common common;
+   backtrack_common common;
    jump_list *condfailed;
    /* Less than 0 (-1) if a frame is not needed. */
    int framesize;
    /* Points to our private memory word on the stack. */
-   int localptr;
+   int private_data_ptr;
    /* For iterators. */
-   struct sljit_label *hotpath;
+   struct sljit_label *matchingpath;
- } assert_fallback;
+ } assert_backtrack;
- typedef struct bracket_fallback {
+ typedef struct bracket_backtrack {
-   fallback_common common;
+   backtrack_common common;
    /* Where to coninue if an alternative is successfully matched. */
-   struct sljit_label *althotpath;
+   struct sljit_label *alternative_matchingpath;
    /* For rmin and rmax iterators. */
-   struct sljit_label *recursivehotpath;
+   struct sljit_label *recursive_matchingpath;
    /* For greedy ? operator. */
-   struct sljit_label *zerohotpath;
+   struct sljit_label *zero_matchingpath;
    /* Contains the branches of a failed condition. */
    union {
      /* Both for OP_COND, OP_SCOND. */
      jump_list *condfailed;
-     assert_fallback *assert;
+     assert_backtrack *assert;
      /* For OP_ONCE. -1 if not needed. */
      int framesize;
    } u;
    /* Points to our private memory word on the stack. */
-   int localptr;
+   int private_data_ptr;
- } bracket_fallback;
+ } bracket_backtrack;
- typedef struct bracketpos_fallback {
+ typedef struct bracketpos_backtrack {
-   fallback_common common;
+   backtrack_common common;
    /* Points to our private memory word on the stack. */
-   int localptr;
+   int private_data_ptr;
    /* Reverting stack is needed. */
    int framesize;
    /* Allocated stack size. */
    int stacksize;
- } bracketpos_fallback;
+ } bracketpos_backtrack;
- typedef struct braminzero_fallback {
+ typedef struct braminzero_backtrack {
-   fallback_common common;
+   backtrack_common common;
-   struct sljit_label *hotpath;
+   struct sljit_label *matchingpath;
- } braminzero_fallback;
+ } braminzero_backtrack;
- typedef struct iterator_fallback {
+ typedef struct iterator_backtrack {
-   fallback_common common;
+   backtrack_common common;
    /* Next iteration. */
-   struct sljit_label *hotpath;
+   struct sljit_label *matchingpath;
- } iterator_fallback;
+ } iterator_backtrack;
  typedef struct recurse_entry {
    struct recurse_entry *next;
-Line 263 
 typedef struct recurse_entry {
+Line 286 
 typedef struct recurse_entry {
    int start;
  } recurse_entry;
- typedef struct recurse_fallback {
+ typedef struct recurse_backtrack {
-   fallback_common common;
+   backtrack_common common;
- } recurse_fallback;
+   BOOL inlined_pattern;
+ } recurse_backtrack;
+ #define MAX_RANGE_SIZE 6
  typedef struct compiler_common {
+   /* The sljit ceneric compiler. */
    struct sljit_compiler *compiler;
+   /* First byte code. */
    pcre_uchar *start;
-   int localsize;
+   /* Maps private data offset to each opcode. */
-   int *localptrs;
+   int *private_data_ptrs;
+   /* Tells whether the capturing bracket is optimized. */
+   pcre_uint8 *optimized_cbracket;
+   /* Starting offset of private data for capturing brackets. */
+   int cbra_ptr;
+   /* Output vector starting point. Must be divisible by 2. */
+   int ovector_start;
+   /* Last known position of the requested byte. */
+   int req_char_ptr;
+   /* Head of the last recursion. */
+   int recursive_head_ptr;
+   /* First inspected character for partial matching. */
+   int start_used_ptr;
+   /* Starting pointer for partial soft matches. */
+   int hit_start;
+   /* End pointer of the first line. */
+   int first_line_end;
+   /* Points to the marked string. */
+   int mark_ptr;
+   /* Recursive control verb management chain. */
+   int control_head_ptr;
+   /* Points to the last matched capture block index. */
+   int capture_last_ptr;
+   /* Points to the starting position of the current match. */
+   int start_ptr;
+   /* Flipped and lower case tables. */
    const pcre_uint8 *fcc;
-   sljit_w lcc;
+   sljit_sw lcc;
-   int cbraptr;
+   /* Mode can be PCRE_STUDY_JIT_COMPILE and others. */
+   int mode;
+   /* \K is in the pattern. */
+   BOOL has_set_som;
+   /* Needs to know the start position anytime. */
+   BOOL needs_start_ptr;
+   /* Currently in recurse or assert. */
+   BOOL local_exit;
+   /* Newline control. */
    int nltype;
    int newline;
    int bsr_nltype;
+   /* Dollar endonly. */
    int endonly;
-   sljit_w ctypes;
+   /* Tables. */
+   sljit_sw ctypes;
+   int digits[2 + MAX_RANGE_SIZE];
+   /* Named capturing brackets. */
    sljit_uw name_table;
-   sljit_w name_count;
+   sljit_sw name_count;
-   sljit_w name_entry_size;
+   sljit_sw name_entry_size;
-   struct sljit_label *acceptlabel;
+   /* Labels and jump lists. */
+   struct sljit_label *partialmatchlabel;
+   struct sljit_label *quit_label;
+   struct sljit_label *forced_quit_label;
+   struct sljit_label *accept_label;
    stub_list *stubs;
    recurse_entry *entries;
    recurse_entry *currententry;
+   jump_list *partialmatch;
+   jump_list *quit;
+   jump_list *forced_quit;
    jump_list *accept;
    jump_list *calllimit;
    jump_list *stackalloc;
-Line 297 
 typedef struct compiler_common {
+Line 371 
 typedef struct compiler_common {
    jump_list *vspace;
    jump_list *casefulcmp;
    jump_list *caselesscmp;
+   jump_list *reset_match;
    BOOL jscript_compat;
  #ifdef SUPPORT_UTF
    BOOL utf;
  #ifdef SUPPORT_UCP
    BOOL use_ucp;
  #endif
+ #ifndef COMPILE_PCRE32
    jump_list *utfreadchar;
+ #endif
  #ifdef COMPILE_PCRE8
    jump_list *utfreadtype8;
  #endif
-Line 321 
 typedef struct compare_context {
+Line 398 
 typedef struct compare_context {
  #if defined SLJIT_UNALIGNED && SLJIT_UNALIGNED
    int ucharptr;
    union {
-     sljit_i asint;
+     sljit_si asint;
      sljit_uh asushort;
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
      sljit_ub asbyte;
      sljit_ub asuchars[4];
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
      sljit_uh asuchars[2];
- #endif
+ #elif defined COMPILE_PCRE32
+     sljit_ui asuchars[1];
  #endif
    } c;
    union {
-     sljit_i asint;
+     sljit_si asint;
      sljit_uh asushort;
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
      sljit_ub asbyte;
      sljit_ub asuchars[4];
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
      sljit_uh asuchars[2];
- #endif
+ #elif defined COMPILE_PCRE32
+     sljit_ui asuchars[1];
  #endif
    } oc;
  #endif
  } compare_context;
- enum {
+ /* Undefine sljit macros. */
-   frame_end = 0,
+ #undef CMP
-   frame_setstrbegin = -1
- };
  /* Used for accessing the elements of the stack. */
- #define STACK(i)      ((-(i) - 1) * (int)sizeof(sljit_w))
+ #define STACK(i)      ((-(i) - 1) * (int)sizeof(sljit_sw))
- #define TMP1          SLJIT_TEMPORARY_REG1
+ #define TMP1          SLJIT_SCRATCH_REG1
- #define TMP2          SLJIT_TEMPORARY_REG3
+ #define TMP2          SLJIT_SCRATCH_REG3
  #define TMP3          SLJIT_TEMPORARY_EREG2
- #define STR_PTR       SLJIT_GENERAL_REG1
+ #define STR_PTR       SLJIT_SAVED_REG1
- #define STR_END       SLJIT_GENERAL_REG2
+ #define STR_END       SLJIT_SAVED_REG2
- #define STACK_TOP     SLJIT_TEMPORARY_REG2
+ #define STACK_TOP     SLJIT_SCRATCH_REG2
- #define STACK_LIMIT   SLJIT_GENERAL_REG3
+ #define STACK_LIMIT   SLJIT_SAVED_REG3
- #define ARGUMENTS     SLJIT_GENERAL_EREG1
+ #define ARGUMENTS     SLJIT_SAVED_EREG1
- #define CALL_COUNT    SLJIT_GENERAL_EREG2
+ #define CALL_COUNT    SLJIT_SAVED_EREG2
  #define RETURN_ADDR   SLJIT_TEMPORARY_EREG1
- /* Locals layout. */
+ /* Local space layout. */
  /* These two locals can be used by the current opcode. */
- #define LOCALS0          (0 * sizeof(sljit_w))
+ #define LOCALS0          (0 * sizeof(sljit_sw))
- #define LOCALS1          (1 * sizeof(sljit_w))
+ #define LOCALS1          (1 * sizeof(sljit_sw))
  /* Two local variables for possessive quantifiers (char1 cannot use them). */
- #define POSSESSIVE0      (2 * sizeof(sljit_w))
+ #define POSSESSIVE0      (2 * sizeof(sljit_sw))
- #define POSSESSIVE1      (3 * sizeof(sljit_w))
+ #define POSSESSIVE1      (3 * sizeof(sljit_sw))
- /* Head of the last recursion. */
- #define RECURSIVE_HEAD   (4 * sizeof(sljit_w))
  /* Max limit of recursions. */
- #define CALL_LIMIT       (5 * sizeof(sljit_w))
+ #define CALL_LIMIT       (4 * sizeof(sljit_sw))
- /* Last known position of the requested byte. */
- #define REQ_CHAR_PTR     (6 * sizeof(sljit_w))
- /* End pointer of the first line. */
- #define FIRSTLINE_END    (7 * sizeof(sljit_w))
  /* The output vector is stored on the stack, and contains pointers
  to characters. The vector data is divided into two groups: the first
  group contains the start / end character pointers, and the second is
  the start pointers when the end of the capturing group has not yet reached. */
- #define OVECTOR_START    (8 * sizeof(sljit_w))
+ #define OVECTOR_START    (common->ovector_start)
- #define OVECTOR(i)       (OVECTOR_START + (i) * sizeof(sljit_w))
+ #define OVECTOR(i)       (OVECTOR_START + (i) * sizeof(sljit_sw))
- #define OVECTOR_PRIV(i)  (common->cbraptr + (i) * sizeof(sljit_w))
+ #define OVECTOR_PRIV(i)  (common->cbra_ptr + (i) * sizeof(sljit_sw))
- #define PRIV_DATA(cc)    (common->localptrs[(cc) - common->start])
+ #define PRIVATE_DATA(cc) (common->private_data_ptrs[(cc) - common->start])
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
  #define MOV_UCHAR  SLJIT_MOV_UB
  #define MOVU_UCHAR SLJIT_MOVU_UB
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
  #define MOV_UCHAR  SLJIT_MOV_UH
  #define MOVU_UCHAR SLJIT_MOVU_UH
+ #elif defined COMPILE_PCRE32
+ #define MOV_UCHAR  SLJIT_MOV_UI
+ #define MOVU_UCHAR SLJIT_MOVU_UI
  #else
  #error Unsupported compiling mode
  #endif
- #endif
  /* Shortcuts. */
  #define DEFINE_COMPILER \
-Line 417 
 the start pointers when the end of the c
+Line 487 
 the start pointers when the end of the c
    sljit_set_label(sljit_emit_jump(compiler, (type)), (label))
  #define JUMPHERE(jump) \
    sljit_set_label((jump), sljit_emit_label(compiler))
+ #define SET_LABEL(jump, label) \
+   sljit_set_label((jump), (label))
  #define CMP(type, src1, src1w, src2, src2w) \
    sljit_emit_cmp(compiler, (type), (src1), (src1w), (src2), (src2w))
  #define CMPTO(type, src1, src1w, src2, src2w, label) \
    sljit_set_label(sljit_emit_cmp(compiler, (type), (src1), (src1w), (src2), (src2w)), (label))
- #define COND_VALUE(op, dst, dstw, type) \
+ #define OP_FLAGS(op, dst, dstw, src, srcw, type) \
-   sljit_emit_cond_value(compiler, (op), (dst), (dstw), (type))
+   sljit_emit_op_flags(compiler, (op), (dst), (dstw), (src), (srcw), (type))
+ #define GET_LOCAL_BASE(dst, dstw, offset) \
+   sljit_get_local_base(compiler, (dst), (dstw), (offset))
  static pcre_uchar* bracketend(pcre_uchar* cc)
  {
-Line 435 
 return cc;
+Line 509 
 return cc;
  /* Functions whose might need modification for all new supported opcodes:
   next_opcode
-  get_localspace
+  get_private_data_length
-  set_localptrs
+  set_private_data_ptrs
   get_framesize
   init_frame
-  get_localsize
+  get_private_data_length_for_copy
-  copy_locals
+  copy_private_data
-  compile_hotpath
+  compile_matchingpath
-  compile_fallbackpath
+  compile_backtrackingpath
  */
  static pcre_uchar *next_opcode(compiler_common *common, pcre_uchar *cc)
-Line 463 
 switch(*cc)
+Line 537 
 switch(*cc)
    case OP_WORDCHAR:
    case OP_ANY:
    case OP_ALLANY:
+   case OP_NOTPROP:
+   case OP_PROP:
    case OP_ANYNL:
    case OP_NOT_HSPACE:
    case OP_HSPACE:
-Line 475 
 switch(*cc)
+Line 551 
 switch(*cc)
    case OP_CIRCM:
    case OP_DOLL:
    case OP_DOLLM:
-   case OP_TYPESTAR:
-   case OP_TYPEMINSTAR:
-   case OP_TYPEPLUS:
-   case OP_TYPEMINPLUS:
-   case OP_TYPEQUERY:
-   case OP_TYPEMINQUERY:
-   case OP_TYPEPOSSTAR:
-   case OP_TYPEPOSPLUS:
-   case OP_TYPEPOSQUERY:
    case OP_CRSTAR:
    case OP_CRMINSTAR:
    case OP_CRPLUS:
    case OP_CRMINPLUS:
    case OP_CRQUERY:
    case OP_CRMINQUERY:
+   case OP_CRRANGE:
+   case OP_CRMINRANGE:
+   case OP_CLASS:
+   case OP_NCLASS:
+   case OP_REF:
+   case OP_REFI:
+   case OP_RECURSE:
+   case OP_CALLOUT:
+   case OP_ALT:
+   case OP_KET:
+   case OP_KETRMAX:
+   case OP_KETRMIN:
+   case OP_KETRPOS:
+   case OP_REVERSE:
+   case OP_ASSERT:
+   case OP_ASSERT_NOT:
+   case OP_ASSERTBACK:
+   case OP_ASSERTBACK_NOT:
+   case OP_ONCE:
+   case OP_ONCE_NC:
+   case OP_BRA:
+   case OP_BRAPOS:
+   case OP_CBRA:
+   case OP_CBRAPOS:
+   case OP_COND:
+   case OP_SBRA:
+   case OP_SBRAPOS:
+   case OP_SCBRA:
+   case OP_SCBRAPOS:
+   case OP_SCOND:
+   case OP_CREF:
+   case OP_NCREF:
+   case OP_RREF:
+   case OP_NRREF:
    case OP_DEF:
    case OP_BRAZERO:
    case OP_BRAMINZERO:
    case OP_BRAPOSZERO:
+   case OP_PRUNE:
+   case OP_SKIP:
+   case OP_COMMIT:
    case OP_FAIL:
    case OP_ACCEPT:
    case OP_ASSERT_ACCEPT:
+   case OP_CLOSE:
    case OP_SKIPZERO:
-   return cc + 1;
+   return cc + PRIV(OP_lengths)[*cc];
-   case OP_ANYBYTE:
- #ifdef SUPPORT_UTF
-   if (common->utf) return NULL;
- #endif
-   return cc + 1;
    case OP_CHAR:
    case OP_CHARI:
-Line 516 
 switch(*cc)
+Line 615 
 switch(*cc)
    case OP_MINPLUS:
    case OP_QUERY:
    case OP_MINQUERY:
+   case OP_UPTO:
+   case OP_MINUPTO:
+   case OP_EXACT:
    case OP_POSSTAR:
    case OP_POSPLUS:
    case OP_POSQUERY:
+   case OP_POSUPTO:
    case OP_STARI:
    case OP_MINSTARI:
    case OP_PLUSI:
    case OP_MINPLUSI:
    case OP_QUERYI:
    case OP_MINQUERYI:
+   case OP_UPTOI:
+   case OP_MINUPTOI:
+   case OP_EXACTI:
    case OP_POSSTARI:
    case OP_POSPLUSI:
    case OP_POSQUERYI:
+   case OP_POSUPTOI:
    case OP_NOTSTAR:
    case OP_NOTMINSTAR:
    case OP_NOTPLUS:
    case OP_NOTMINPLUS:
    case OP_NOTQUERY:
    case OP_NOTMINQUERY:
+   case OP_NOTUPTO:
+   case OP_NOTMINUPTO:
+   case OP_NOTEXACT:
    case OP_NOTPOSSTAR:
    case OP_NOTPOSPLUS:
    case OP_NOTPOSQUERY:
+   case OP_NOTPOSUPTO:
    case OP_NOTSTARI:
    case OP_NOTMINSTARI:
    case OP_NOTPLUSI:
    case OP_NOTMINPLUSI:
    case OP_NOTQUERYI:
    case OP_NOTMINQUERYI:
-   case OP_NOTPOSSTARI:
-   case OP_NOTPOSPLUSI:
-   case OP_NOTPOSQUERYI:
-   cc += 2;
- #ifdef SUPPORT_UTF
-   if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
- #endif
-   return cc;
-   case OP_UPTO:
-   case OP_MINUPTO:
-   case OP_EXACT:
-   case OP_POSUPTO:
-   case OP_UPTOI:
-   case OP_MINUPTOI:
-   case OP_EXACTI:
-   case OP_POSUPTOI:
-   case OP_NOTUPTO:
-   case OP_NOTMINUPTO:
-   case OP_NOTEXACT:
-   case OP_NOTPOSUPTO:
    case OP_NOTUPTOI:
    case OP_NOTMINUPTOI:
    case OP_NOTEXACTI:
+   case OP_NOTPOSSTARI:
+   case OP_NOTPOSPLUSI:
+   case OP_NOTPOSQUERYI:
    case OP_NOTPOSUPTOI:
-   cc += 2 + IMM2_SIZE;
+   cc += PRIV(OP_lengths)[*cc];
  #ifdef SUPPORT_UTF
    if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
  #endif
    return cc;
-   case OP_NOTPROP:
+   /* Special cases. */
-   case OP_PROP:
+   case OP_TYPESTAR:
-   return cc + 1 + 2;
+   case OP_TYPEMINSTAR:
+   case OP_TYPEPLUS:
+   case OP_TYPEMINPLUS:
+   case OP_TYPEQUERY:
+   case OP_TYPEMINQUERY:
    case OP_TYPEUPTO:
    case OP_TYPEMINUPTO:
    case OP_TYPEEXACT:
+   case OP_TYPEPOSSTAR:
+   case OP_TYPEPOSPLUS:
+   case OP_TYPEPOSQUERY:
    case OP_TYPEPOSUPTO:
-   case OP_REF:
+   return cc + PRIV(OP_lengths)[*cc] - 1;
-   case OP_REFI:
-   case OP_CREF:
-   case OP_NCREF:
-   case OP_RREF:
-   case OP_NRREF:
-   case OP_CLOSE:
-   cc += 1 + IMM2_SIZE;
-   return cc;
-   case OP_CRRANGE:
+   case OP_ANYBYTE:
-   case OP_CRMINRANGE:
+ #ifdef SUPPORT_UTF
-   return cc + 1 + 2 * IMM2_SIZE;
+   if (common->utf) return NULL;
+ #endif
-   case OP_CLASS:
+   return cc + 1;
-   case OP_NCLASS:
-   return cc + 1 + 32 / sizeof(pcre_uchar);
  #if defined SUPPORT_UTF || !defined COMPILE_PCRE8
    case OP_XCLASS:
    return cc + GET(cc, 1);
  #endif
-   case OP_RECURSE:
+   case OP_MARK:
-   case OP_ASSERT:
+   case OP_PRUNE_ARG:
-   case OP_ASSERT_NOT:
+   return cc + 1 + 2 + cc[1];
-   case OP_ASSERTBACK:
-   case OP_ASSERTBACK_NOT:
-   case OP_REVERSE:
-   case OP_ONCE:
-   case OP_ONCE_NC:
-   case OP_BRA:
-   case OP_BRAPOS:
-   case OP_COND:
-   case OP_SBRA:
-   case OP_SBRAPOS:
-   case OP_SCOND:
-   case OP_ALT:
-   case OP_KET:
-   case OP_KETRMAX:
-   case OP_KETRMIN:
-   case OP_KETRPOS:
-   return cc + 1 + LINK_SIZE;
-   case OP_CBRA:
-   case OP_CBRAPOS:
-   case OP_SCBRA:
-   case OP_SCBRAPOS:
-   return cc + 1 + LINK_SIZE + IMM2_SIZE;
    default:
    return NULL;
    }
  }
- static int get_localspace(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend)
+ #define CASE_ITERATOR_PRIVATE_DATA_1 \
+     case OP_MINSTAR: \
+     case OP_MINPLUS: \
+     case OP_QUERY: \
+     case OP_MINQUERY: \
+     case OP_MINSTARI: \
+     case OP_MINPLUSI: \
+     case OP_QUERYI: \
+     case OP_MINQUERYI: \
+     case OP_NOTMINSTAR: \
+     case OP_NOTMINPLUS: \
+     case OP_NOTQUERY: \
+     case OP_NOTMINQUERY: \
+     case OP_NOTMINSTARI: \
+     case OP_NOTMINPLUSI: \
+     case OP_NOTQUERYI: \
+     case OP_NOTMINQUERYI:
+ #define CASE_ITERATOR_PRIVATE_DATA_2A \
+     case OP_STAR: \
+     case OP_PLUS: \
+     case OP_STARI: \
+     case OP_PLUSI: \
+     case OP_NOTSTAR: \
+     case OP_NOTPLUS: \
+     case OP_NOTSTARI: \
+     case OP_NOTPLUSI:
+ #define CASE_ITERATOR_PRIVATE_DATA_2B \
+     case OP_UPTO: \
+     case OP_MINUPTO: \
+     case OP_UPTOI: \
+     case OP_MINUPTOI: \
+     case OP_NOTUPTO: \
+     case OP_NOTMINUPTO: \
+     case OP_NOTUPTOI: \
+     case OP_NOTMINUPTOI:
+ #define CASE_ITERATOR_TYPE_PRIVATE_DATA_1 \
+     case OP_TYPEMINSTAR: \
+     case OP_TYPEMINPLUS: \
+     case OP_TYPEQUERY: \
+     case OP_TYPEMINQUERY:
+ #define CASE_ITERATOR_TYPE_PRIVATE_DATA_2A \
+     case OP_TYPESTAR: \
+     case OP_TYPEPLUS:
+ #define CASE_ITERATOR_TYPE_PRIVATE_DATA_2B \
+     case OP_TYPEUPTO: \
+     case OP_TYPEMINUPTO:
+ static int get_class_iterator_size(pcre_uchar *cc)
+ {
+ switch(*cc)
+   {
+   case OP_CRSTAR:
+   case OP_CRPLUS:
+   return 2;
+   case OP_CRMINSTAR:
+   case OP_CRMINPLUS:
+   case OP_CRQUERY:
+   case OP_CRMINQUERY:
+   return 1;
+   case OP_CRRANGE:
+   case OP_CRMINRANGE:
+   if (GET2(cc, 1) == GET2(cc, 1 + IMM2_SIZE))
+     return 0;
+   return 2;
+   default:
+   return 0;
+   }
+ }
+ static int get_private_data_length(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend)
  {
- int localspace = 0;
+ int private_data_length = 0;
  pcre_uchar *alternative;
+ pcre_uchar *name;
+ pcre_uchar *end = NULL;
+ int space, size, i;
+ pcre_uint32 bracketlen;
  /* Calculate important variables (like stack size) and checks whether all opcodes are supported. */
  while (cc < ccend)
    {
+   space = 0;
+   size = 0;
+   bracketlen = 0;
    switch(*cc)
      {
+     case OP_SET_SOM:
+     common->has_set_som = TRUE;
+     cc += 1;
+     break;
+     case OP_REF:
+     case OP_REFI:
+     common->optimized_cbracket[GET2(cc, 1)] = 0;
+     cc += 1 + IMM2_SIZE;
+     break;
      case OP_ASSERT:
      case OP_ASSERT_NOT:
      case OP_ASSERTBACK:
-Line 655 
 while (cc < ccend)
+Line 817 
 while (cc < ccend)
      case OP_BRAPOS:
      case OP_SBRA:
      case OP_SBRAPOS:
-     case OP_SCOND:
+     private_data_length += sizeof(sljit_sw);
-     localspace += sizeof(sljit_w);
+     bracketlen = 1 + LINK_SIZE;
-     cc += 1 + LINK_SIZE;
      break;
      case OP_CBRAPOS:
      case OP_SCBRAPOS:
-     localspace += sizeof(sljit_w);
+     private_data_length += sizeof(sljit_sw);
-     cc += 1 + LINK_SIZE + IMM2_SIZE;
+     common->optimized_cbracket[GET2(cc, 1 + LINK_SIZE)] = 0;
+     bracketlen = 1 + LINK_SIZE + IMM2_SIZE;
      break;
      case OP_COND:
-     /* Might be a hidden SCOND. */
+     case OP_SCOND:
-     alternative = cc + GET(cc, 1);
+     /* Only AUTO_CALLOUT can insert this opcode. We do
-     if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
+        not intend to support this case. */
-       localspace += sizeof(sljit_w);
+     if (cc[1 + LINK_SIZE] == OP_CALLOUT)
+       return -1;
+     if (*cc == OP_COND)
+       {
+       /* Might be a hidden SCOND. */
+       alternative = cc + GET(cc, 1);
+       if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
+         private_data_length += sizeof(sljit_sw);
+       }
+     else
+       private_data_length += sizeof(sljit_sw);
+     bracketlen = 1 + LINK_SIZE;
+     break;
+     case OP_CREF:
+     i = GET2(cc, 1);
+     common->optimized_cbracket[i] = 0;
+     cc += 1 + IMM2_SIZE;
+     break;
+     case OP_NCREF:
+     bracketlen = GET2(cc, 1);
+     name = (pcre_uchar *)common->name_table;
+     alternative = name;
+     for (i = 0; i < common->name_count; i++)
+       {
+       if (GET2(name, 0) == bracketlen) break;
+       name += common->name_entry_size;
+       }
+     SLJIT_ASSERT(i != common->name_count);
+     for (i = 0; i < common->name_count; i++)
+       {
+       if (STRCMP_UC_UC(alternative + IMM2_SIZE, name + IMM2_SIZE) == 0)
+         common->optimized_cbracket[GET2(alternative, 0)] = 0;
+       alternative += common->name_entry_size;
+       }
+     bracketlen = 0;
+     cc += 1 + IMM2_SIZE;
+     break;
+     case OP_BRA:
+     bracketlen = 1 + LINK_SIZE;
+     break;
+     case OP_CBRA:
+     case OP_SCBRA:
+     bracketlen = 1 + LINK_SIZE + IMM2_SIZE;
+     break;
+     CASE_ITERATOR_PRIVATE_DATA_1
+     space = 1;
+     size = -2;
+     break;
+     CASE_ITERATOR_PRIVATE_DATA_2A
+     space = 2;
+     size = -2;
+     break;
+     CASE_ITERATOR_PRIVATE_DATA_2B
+     space = 2;
+     size = -(2 + IMM2_SIZE);
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_1
+     space = 1;
+     size = 1;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2A
+     if (cc[1] != OP_ANYNL && cc[1] != OP_EXTUNI)
+       space = 2;
+     size = 1;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2B
+     if (cc[1 + IMM2_SIZE] != OP_ANYNL && cc[1 + IMM2_SIZE] != OP_EXTUNI)
+       space = 2;
+     size = 1 + IMM2_SIZE;
+     break;
+     case OP_CLASS:
+     case OP_NCLASS:
+     size += 1 + 32 / sizeof(pcre_uchar);
+     space = get_class_iterator_size(cc + size);
+     break;
+ #if defined SUPPORT_UTF || !defined COMPILE_PCRE8
+     case OP_XCLASS:
+     size = GET(cc, 1);
+     space = get_class_iterator_size(cc + size);
+     break;
+ #endif
+     case OP_RECURSE:
+     /* Set its value only once. */
+     if (common->recursive_head_ptr == 0)
+       {
+       common->recursive_head_ptr = common->ovector_start;
+       common->ovector_start += sizeof(sljit_sw);
+       }
      cc += 1 + LINK_SIZE;
      break;
+     case OP_CALLOUT:
+     if (common->capture_last_ptr == 0)
+       {
+       common->capture_last_ptr = common->ovector_start;
+       common->ovector_start += sizeof(sljit_sw);
+       }
+     cc += 2 + 2 * LINK_SIZE;
+     break;
+     case OP_PRUNE_ARG:
+     common->needs_start_ptr = TRUE;
+     common->control_head_ptr = 1;
+     /* Fall through. */
+     case OP_MARK:
+     if (common->mark_ptr == 0)
+       {
+       common->mark_ptr = common->ovector_start;
+       common->ovector_start += sizeof(sljit_sw);
+       }
+     cc += 1 + 2 + cc[1];
+     break;
+     case OP_PRUNE:
+     case OP_SKIP:
+     common->needs_start_ptr = TRUE;
+     /* Fall through. */
+     case OP_COMMIT:
+     common->control_head_ptr = 1;
+     cc += 1;
+     break;
      default:
      cc = next_opcode(common, cc);
      if (cc == NULL)
        return -1;
      break;
      }
+   if (space > 0 && cc >= end)
+     private_data_length += sizeof(sljit_sw) * space;
+   if (size != 0)
+     {
+     if (size < 0)
+       {
+       cc += -size;
+ #ifdef SUPPORT_UTF
+       if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+       }
+     else
+       cc += size;
+     }
+   if (bracketlen != 0)
+     {
+     if (cc >= end)
+       {
+       end = bracketend(cc);
+       if (end[-1 - LINK_SIZE] == OP_KET)
+         end = NULL;
+       }
+     cc += bracketlen;
+     }
    }
- return localspace;
+ return private_data_length;
  }
- static void set_localptrs(compiler_common *common, int localptr, pcre_uchar *ccend)
+ static void set_private_data_ptrs(compiler_common *common, int private_data_ptr, pcre_uchar *ccend)
  {
  pcre_uchar *cc = common->start;
  pcre_uchar *alternative;
+ pcre_uchar *end = NULL;
+ int space, size, bracketlen;
  while (cc < ccend)
    {
+   space = 0;
+   size = 0;
+   bracketlen = 0;
    switch(*cc)
      {
      case OP_ASSERT:
-Line 702 
 while (cc < ccend)
+Line 1032 
 while (cc < ccend)
      case OP_SBRA:
      case OP_SBRAPOS:
      case OP_SCOND:
-     common->localptrs[cc - common->start] = localptr;
+     common->private_data_ptrs[cc - common->start] = private_data_ptr;
-     localptr += sizeof(sljit_w);
+     private_data_ptr += sizeof(sljit_sw);
-     cc += 1 + LINK_SIZE;
+     bracketlen = 1 + LINK_SIZE;
+     break;
+     case OP_CBRAPOS:
+     case OP_SCBRAPOS:
+     common->private_data_ptrs[cc - common->start] = private_data_ptr;
+     private_data_ptr += sizeof(sljit_sw);
+     bracketlen = 1 + LINK_SIZE + IMM2_SIZE;
+     break;
+     case OP_COND:
+     /* Might be a hidden SCOND. */
+     alternative = cc + GET(cc, 1);
+     if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
+       {
+       common->private_data_ptrs[cc - common->start] = private_data_ptr;
+       private_data_ptr += sizeof(sljit_sw);
+       }
+     bracketlen = 1 + LINK_SIZE;
+     break;
+     case OP_BRA:
+     bracketlen = 1 + LINK_SIZE;
+     break;
+     case OP_CBRA:
+     case OP_SCBRA:
+     bracketlen = 1 + LINK_SIZE + IMM2_SIZE;
      break;
-     case OP_CBRAPOS:
+     CASE_ITERATOR_PRIVATE_DATA_1
-     case OP_SCBRAPOS:
+     space = 1;
-     common->localptrs[cc - common->start] = localptr;
+     size = -2;
-     localptr += sizeof(sljit_w);
-     cc += 1 + LINK_SIZE + IMM2_SIZE;
      break;
-     case OP_COND:
+     CASE_ITERATOR_PRIVATE_DATA_2A
-     /* Might be a hidden SCOND. */
+     space = 2;
-     alternative = cc + GET(cc, 1);
+     size = -2;
-     if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
+     break;
-       {
-       common->localptrs[cc - common->start] = localptr;
+     CASE_ITERATOR_PRIVATE_DATA_2B
-       localptr += sizeof(sljit_w);
+     space = 2;
-       }
+     size = -(2 + IMM2_SIZE);
-     cc += 1 + LINK_SIZE;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_1
+     space = 1;
+     size = 1;
      break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2A
+     if (cc[1] != OP_ANYNL && cc[1] != OP_EXTUNI)
+       space = 2;
+     size = 1;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2B
+     if (cc[1 + IMM2_SIZE] != OP_ANYNL && cc[1 + IMM2_SIZE] != OP_EXTUNI)
+       space = 2;
+     size = 1 + IMM2_SIZE;
+     break;
+     case OP_CLASS:
+     case OP_NCLASS:
+     size += 1 + 32 / sizeof(pcre_uchar);
+     space = get_class_iterator_size(cc + size);
+     break;
+ #if defined SUPPORT_UTF || !defined COMPILE_PCRE8
+     case OP_XCLASS:
+     size = GET(cc, 1);
+     space = get_class_iterator_size(cc + size);
+     break;
+ #endif
      default:
      cc = next_opcode(common, cc);
      SLJIT_ASSERT(cc != NULL);
      break;
      }
+   if (space > 0 && cc >= end)
+     {
+     common->private_data_ptrs[cc - common->start] = private_data_ptr;
+     private_data_ptr += sizeof(sljit_sw) * space;
+     }
+   if (size != 0)
+     {
+     if (size < 0)
+       {
+       cc += -size;
+ #ifdef SUPPORT_UTF
+       if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+       }
+     else
+       cc += size;
+     }
+   if (bracketlen > 0)
+     {
+     if (cc >= end)
+       {
+       end = bracketend(cc);
+       if (end[-1 - LINK_SIZE] == OP_KET)
+         end = NULL;
+       }
+     cc += bracketlen;
+     }
    }
  }
- /* Returns with -1 if no need for frame. */
+ /* Returns with a frame_types (always < 0) if no need for frame. */
  static int get_framesize(compiler_common *common, pcre_uchar *cc, BOOL recursive)
  {
- pcre_uchar *ccend = bracketend(cc);
+ pcre_uchar *ccend = bracketend(cc) - (1 + LINK_SIZE);
  int length = 0;
- BOOL possessive = FALSE;
+ int possessive = 0;
- BOOL setsom_found = FALSE;
+ BOOL stack_restore = FALSE;
+ BOOL setsom_found = recursive;
+ BOOL setmark_found = recursive;
+ /* The last capture is a local variable even for recursions. */
+ BOOL capture_last_found = FALSE;
  if (!recursive && (*cc == OP_CBRAPOS || *cc == OP_SCBRAPOS))
    {
-   length = 3;
+   possessive = length = (common->capture_last_ptr != 0) ? 5 : 3;
-   possessive = TRUE;
+   /* This is correct regardless of common->capture_last_ptr. */
+   capture_last_found = TRUE;
    }
  cc = next_opcode(common, cc);
-Line 753 
 while (cc < ccend)
+Line 1172 
 while (cc < ccend)
    switch(*cc)
      {
      case OP_SET_SOM:
-     case OP_RECURSE:
+     SLJIT_ASSERT(common->has_set_som);
+     stack_restore = TRUE;
      if (!setsom_found)
        {
        length += 2;
        setsom_found = TRUE;
        }
-     cc += (*cc == OP_SET_SOM) ? 1 : 1 + LINK_SIZE;
+     cc += 1;
+     break;
+     case OP_MARK:
+     case OP_PRUNE_ARG:
+     SLJIT_ASSERT(common->mark_ptr != 0);
+     stack_restore = TRUE;
+     if (!setmark_found)
+       {
+       length += 2;
+       setmark_found = TRUE;
+       }
+     cc += 1 + 2 + cc[1];
+     break;
+     case OP_RECURSE:
+     stack_restore = TRUE;
+     if (common->has_set_som && !setsom_found)
+       {
+       length += 2;
+       setsom_found = TRUE;
+       }
+     if (common->mark_ptr != 0 && !setmark_found)
+       {
+       length += 2;
+       setmark_found = TRUE;
+       }
+     if (common->capture_last_ptr != 0 && !capture_last_found)
+       {
+       length += 2;
+       capture_last_found = TRUE;
+       }
+     cc += 1 + LINK_SIZE;
      break;
      case OP_CBRA:
      case OP_CBRAPOS:
      case OP_SCBRA:
      case OP_SCBRAPOS:
+     stack_restore = TRUE;
+     if (common->capture_last_ptr != 0 && !capture_last_found)
+       {
+       length += 2;
+       capture_last_found = TRUE;
+       }
      length += 3;
      cc += 1 + LINK_SIZE + IMM2_SIZE;
      break;
      default:
+     stack_restore = TRUE;
+     /* Fall through. */
+     case OP_NOT_WORD_BOUNDARY:
+     case OP_WORD_BOUNDARY:
+     case OP_NOT_DIGIT:
+     case OP_DIGIT:
+     case OP_NOT_WHITESPACE:
+     case OP_WHITESPACE:
+     case OP_NOT_WORDCHAR:
+     case OP_WORDCHAR:
+     case OP_ANY:
+     case OP_ALLANY:
+     case OP_ANYBYTE:
+     case OP_NOTPROP:
+     case OP_PROP:
+     case OP_ANYNL:
+     case OP_NOT_HSPACE:
+     case OP_HSPACE:
+     case OP_NOT_VSPACE:
+     case OP_VSPACE:
+     case OP_EXTUNI:
+     case OP_EODN:
+     case OP_EOD:
+     case OP_CIRC:
+     case OP_CIRCM:
+     case OP_DOLL:
+     case OP_DOLLM:
+     case OP_CHAR:
+     case OP_CHARI:
+     case OP_NOT:
+     case OP_NOTI:
+     case OP_EXACT:
+     case OP_POSSTAR:
+     case OP_POSPLUS:
+     case OP_POSQUERY:
+     case OP_POSUPTO:
+     case OP_EXACTI:
+     case OP_POSSTARI:
+     case OP_POSPLUSI:
+     case OP_POSQUERYI:
+     case OP_POSUPTOI:
+     case OP_NOTEXACT:
+     case OP_NOTPOSSTAR:
+     case OP_NOTPOSPLUS:
+     case OP_NOTPOSQUERY:
+     case OP_NOTPOSUPTO:
+     case OP_NOTEXACTI:
+     case OP_NOTPOSSTARI:
+     case OP_NOTPOSPLUSI:
+     case OP_NOTPOSQUERYI:
+     case OP_NOTPOSUPTOI:
+     case OP_TYPEEXACT:
+     case OP_TYPEPOSSTAR:
+     case OP_TYPEPOSPLUS:
+     case OP_TYPEPOSQUERY:
+     case OP_TYPEPOSUPTO:
+     case OP_CLASS:
+     case OP_NCLASS:
+     case OP_XCLASS:
      cc = next_opcode(common, cc);
      SLJIT_ASSERT(cc != NULL);
      break;
      }
  /* Possessive quantifiers can use a special case. */
- if (SLJIT_UNLIKELY(possessive) && length == 3)
+ if (SLJIT_UNLIKELY(possessive == length))
-   return -1;
+   return stack_restore ? no_frame : no_stack;
  if (length > 0)
    return length + 1;
- return -1;
+ return stack_restore ? no_frame : no_stack;
  }
  static void init_frame(compiler_common *common, pcre_uchar *cc, int stackpos, int stacktop, BOOL recursive)
  {
  DEFINE_COMPILER;
- pcre_uchar *ccend = bracketend(cc);
+ pcre_uchar *ccend = bracketend(cc) - (1 + LINK_SIZE);
- BOOL setsom_found = FALSE;
+ BOOL setsom_found = recursive;
+ BOOL setmark_found = recursive;
+ /* The last capture is a local variable even for recursions. */
+ BOOL capture_last_found = FALSE;
  int offset;
  /* >= 1 + shortest item size (2) */
+ SLJIT_UNUSED_ARG(stacktop);
  SLJIT_ASSERT(stackpos >= stacktop + 2);
  stackpos = STACK(stackpos);
-Line 803 
 while (cc < ccend)
+Line 1332 
 while (cc < ccend)
    switch(*cc)
      {
      case OP_SET_SOM:
-     case OP_RECURSE:
+     SLJIT_ASSERT(common->has_set_som);
      if (!setsom_found)
        {
        OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(0));
-       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, frame_setstrbegin);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -OVECTOR(0));
-       stackpos += (int)sizeof(sljit_w);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       setsom_found = TRUE;
+       }
+     cc += 1;
+     break;
+     case OP_MARK:
+     case OP_PRUNE_ARG:
+     SLJIT_ASSERT(common->mark_ptr != 0);
+     if (!setmark_found)
+       {
+       OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -common->mark_ptr);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       setmark_found = TRUE;
+       }
+     cc += 1 + 2 + cc[1];
+     break;
+     case OP_RECURSE:
+     if (common->has_set_som && !setsom_found)
+       {
+       OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(0));
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -OVECTOR(0));
+       stackpos += (int)sizeof(sljit_sw);
        OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
-       stackpos += (int)sizeof(sljit_w);
+       stackpos += (int)sizeof(sljit_sw);
        setsom_found = TRUE;
        }
-     cc += (*cc == OP_SET_SOM) ? 1 : 1 + LINK_SIZE;
+     if (common->mark_ptr != 0 && !setmark_found)
+       {
+       OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -common->mark_ptr);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       setmark_found = TRUE;
+       }
+     if (common->capture_last_ptr != 0 && !capture_last_found)
+       {
+       OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->capture_last_ptr);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -common->capture_last_ptr);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       capture_last_found = TRUE;
+       }
+     cc += 1 + LINK_SIZE;
      break;
      case OP_CBRA:
      case OP_CBRAPOS:
      case OP_SCBRA:
      case OP_SCBRAPOS:
+     if (common->capture_last_ptr != 0 && !capture_last_found)
+       {
+       OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->capture_last_ptr);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, -common->capture_last_ptr);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       capture_last_found = TRUE;
+       }
      offset = (GET2(cc, 1 + LINK_SIZE)) << 1;
      OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, OVECTOR(offset));
-     stackpos += (int)sizeof(sljit_w);
+     stackpos += (int)sizeof(sljit_sw);
      OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(offset));
      OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(offset + 1));
      OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
-     stackpos += (int)sizeof(sljit_w);
+     stackpos += (int)sizeof(sljit_sw);
      OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP2, 0);
-     stackpos += (int)sizeof(sljit_w);
+     stackpos += (int)sizeof(sljit_sw);
      cc += 1 + LINK_SIZE + IMM2_SIZE;
      break;
-Line 839 
 while (cc < ccend)
+Line 1423 
 while (cc < ccend)
      break;
      }
- OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, frame_end);
+ OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, SLJIT_IMM, 0);
  SLJIT_ASSERT(stackpos == STACK(stacktop));
  }
- static SLJIT_INLINE int get_localsize(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend)
+ static SLJIT_INLINE int get_private_data_length_for_copy(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend)
  {
- int localsize = 2;
+ int private_data_length = common->control_head_ptr ? 3 : 2;
+ int size;
  pcre_uchar *alternative;
- /* Calculate the sum of the local variables. */
+ /* Calculate the sum of the private machine words. */
  while (cc < ccend)
    {
+   size = 0;
    switch(*cc)
      {
      case OP_ASSERT:
-Line 862 
 while (cc < ccend)
+Line 1448 
 while (cc < ccend)
      case OP_SBRA:
      case OP_SBRAPOS:
      case OP_SCOND:
-     localsize++;
+     private_data_length++;
      cc += 1 + LINK_SIZE;
      break;
      case OP_CBRA:
      case OP_SCBRA:
-     localsize++;
+     if (common->optimized_cbracket[GET2(cc, 1 + LINK_SIZE)] == 0)
+       private_data_length++;
      cc += 1 + LINK_SIZE + IMM2_SIZE;
      break;
      case OP_CBRAPOS:
      case OP_SCBRAPOS:
-     localsize += 2;
+     private_data_length += 2;
      cc += 1 + LINK_SIZE + IMM2_SIZE;
      break;
-Line 882 
 while (cc < ccend)
+Line 1469 
 while (cc < ccend)
      /* Might be a hidden SCOND. */
      alternative = cc + GET(cc, 1);
      if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
-       localsize++;
+       private_data_length++;
      cc += 1 + LINK_SIZE;
      break;
+     CASE_ITERATOR_PRIVATE_DATA_1
+     if (PRIVATE_DATA(cc))
+       private_data_length++;
+     cc += 2;
+ #ifdef SUPPORT_UTF
+     if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+     break;
+     CASE_ITERATOR_PRIVATE_DATA_2A
+     if (PRIVATE_DATA(cc))
+       private_data_length += 2;
+     cc += 2;
+ #ifdef SUPPORT_UTF
+     if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+     break;
+     CASE_ITERATOR_PRIVATE_DATA_2B
+     if (PRIVATE_DATA(cc))
+       private_data_length += 2;
+     cc += 2 + IMM2_SIZE;
+ #ifdef SUPPORT_UTF
+     if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_1
+     if (PRIVATE_DATA(cc))
+       private_data_length++;
+     cc += 1;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2A
+     if (PRIVATE_DATA(cc))
+       private_data_length += 2;
+     cc += 1;
+     break;
+     CASE_ITERATOR_TYPE_PRIVATE_DATA_2B
+     if (PRIVATE_DATA(cc))
+       private_data_length += 2;
+     cc += 1 + IMM2_SIZE;
+     break;
+     case OP_CLASS:
+     case OP_NCLASS:
+ #if defined SUPPORT_UTF || !defined COMPILE_PCRE8
+     case OP_XCLASS:
+     size = (*cc == OP_XCLASS) ? GET(cc, 1) : 1 + 32 / (int)sizeof(pcre_uchar);
+ #else
+     size = 1 + 32 / (int)sizeof(pcre_uchar);
+ #endif
+     if (PRIVATE_DATA(cc))
+       private_data_length += get_class_iterator_size(cc + size);
+     cc += size;
+     break;
      default:
      cc = next_opcode(common, cc);
      SLJIT_ASSERT(cc != NULL);
-Line 893 
 while (cc < ccend)
+Line 1538 
 while (cc < ccend)
      }
    }
  SLJIT_ASSERT(cc == ccend);
- return localsize;
+ return private_data_length;
  }
- static void copy_locals(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend,
+ static void copy_private_data(compiler_common *common, pcre_uchar *cc, pcre_uchar *ccend,
    BOOL save, int stackptr, int stacktop)
  {
  DEFINE_COMPILER;
  int srcw[2];
- int count;
+ int count, size;
  BOOL tmp1next = TRUE;
  BOOL tmp1empty = TRUE;
  BOOL tmp2empty = TRUE;
-Line 918 
 stacktop = STACK(stacktop - 1);
+Line 1563 
 stacktop = STACK(stacktop - 1);
  if (!save)
    {
-   stackptr += sizeof(sljit_w);
+   stackptr += (common->control_head_ptr ? 2 : 1) * sizeof(sljit_sw);
    if (stackptr < stacktop)
      {
      OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(STACK_TOP), stackptr);
-     stackptr += sizeof(sljit_w);
+     stackptr += sizeof(sljit_sw);
      tmp1empty = FALSE;
      }
    if (stackptr < stacktop)
      {
      OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(STACK_TOP), stackptr);
-     stackptr += sizeof(sljit_w);
+     stackptr += sizeof(sljit_sw);
      tmp2empty = FALSE;
      }
    /* The tmp1next must be TRUE in either way. */
    }
- while (status != end)
+ do
    {
    count = 0;
    switch(status)
      {
      case start:
-     SLJIT_ASSERT(save);
+     SLJIT_ASSERT(save && common->recursive_head_ptr != 0);
      count = 1;
-     srcw[0] = RECURSIVE_HEAD;
+     srcw[0] = common->recursive_head_ptr;
+     if (common->control_head_ptr != 0)
+       {
+       count = 2;
+       srcw[1] = common->control_head_ptr;
+       }
      status = loop;
      break;
-Line 966 
 while (status != end)
+Line 1616 
 while (status != end)
        case OP_SBRAPOS:
        case OP_SCOND:
        count = 1;
-       srcw[0] = PRIV_DATA(cc);
+       srcw[0] = PRIVATE_DATA(cc);
        SLJIT_ASSERT(srcw[0] != 0);
        cc += 1 + LINK_SIZE;
        break;
        case OP_CBRA:
        case OP_SCBRA:
-       count = 1;
+       if (common->optimized_cbracket[GET2(cc, 1 + LINK_SIZE)] == 0)
-       srcw[0] = OVECTOR_PRIV(GET2(cc, 1 + LINK_SIZE));
+         {
+         count = 1;
+         srcw[0] = OVECTOR_PRIV(GET2(cc, 1 + LINK_SIZE));
+         }
        cc += 1 + LINK_SIZE + IMM2_SIZE;
        break;
        case OP_CBRAPOS:
        case OP_SCBRAPOS:
        count = 2;
+       srcw[0] = PRIVATE_DATA(cc);
        srcw[1] = OVECTOR_PRIV(GET2(cc, 1 + LINK_SIZE));
-       srcw[0] = PRIV_DATA(cc);
+       SLJIT_ASSERT(srcw[0] != 0 && srcw[1] != 0);
-       SLJIT_ASSERT(srcw[0] != 0);
        cc += 1 + LINK_SIZE + IMM2_SIZE;
        break;
-Line 993 
 while (status != end)
+Line 1646 
 while (status != end)
        if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
          {
          count = 1;
-         srcw[0] = PRIV_DATA(cc);
+         srcw[0] = PRIVATE_DATA(cc);
          SLJIT_ASSERT(srcw[0] != 0);
          }
        cc += 1 + LINK_SIZE;
        break;
+       CASE_ITERATOR_PRIVATE_DATA_1
+       if (PRIVATE_DATA(cc))
+         {
+         count = 1;
+         srcw[0] = PRIVATE_DATA(cc);
+         }
+       cc += 2;
+ #ifdef SUPPORT_UTF
+       if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+       break;
+       CASE_ITERATOR_PRIVATE_DATA_2A
+       if (PRIVATE_DATA(cc))
+         {
+         count = 2;
+         srcw[0] = PRIVATE_DATA(cc);
+         srcw[1] = PRIVATE_DATA(cc) + sizeof(sljit_sw);
+         }
+       cc += 2;
+ #ifdef SUPPORT_UTF
+       if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+       break;
+       CASE_ITERATOR_PRIVATE_DATA_2B
+       if (PRIVATE_DATA(cc))
+         {
+         count = 2;
+         srcw[0] = PRIVATE_DATA(cc);
+         srcw[1] = PRIVATE_DATA(cc) + sizeof(sljit_sw);
+         }
+       cc += 2 + IMM2_SIZE;
+ #ifdef SUPPORT_UTF
+       if (common->utf && HAS_EXTRALEN(cc[-1])) cc += GET_EXTRALEN(cc[-1]);
+ #endif
+       break;
+       CASE_ITERATOR_TYPE_PRIVATE_DATA_1
+       if (PRIVATE_DATA(cc))
+         {
+         count = 1;
+         srcw[0] = PRIVATE_DATA(cc);
+         }
+       cc += 1;
+       break;
+       CASE_ITERATOR_TYPE_PRIVATE_DATA_2A
+       if (PRIVATE_DATA(cc))
+         {
+         count = 2;
+         srcw[0] = PRIVATE_DATA(cc);
+         srcw[1] = srcw[0] + sizeof(sljit_sw);
+         }
+       cc += 1;
+       break;
+       CASE_ITERATOR_TYPE_PRIVATE_DATA_2B
+       if (PRIVATE_DATA(cc))
+         {
+         count = 2;
+         srcw[0] = PRIVATE_DATA(cc);
+         srcw[1] = srcw[0] + sizeof(sljit_sw);
+         }
+       cc += 1 + IMM2_SIZE;
+       break;
+       case OP_CLASS:
+       case OP_NCLASS:
+ #if defined SUPPORT_UTF || !defined COMPILE_PCRE8
+       case OP_XCLASS:
+       size = (*cc == OP_XCLASS) ? GET(cc, 1) : 1 + 32 / (int)sizeof(pcre_uchar);
+ #else
+       size = 1 + 32 / (int)sizeof(pcre_uchar);
+ #endif
+       if (PRIVATE_DATA(cc))
+         switch(get_class_iterator_size(cc + size))
+           {
+           case 1:
+           count = 1;
+           srcw[0] = PRIVATE_DATA(cc);
+           break;
+           case 2:
+           count = 2;
+           srcw[0] = PRIVATE_DATA(cc);
+           srcw[1] = srcw[0] + sizeof(sljit_sw);
+           break;
+           default:
+           SLJIT_ASSERT_STOP();
+           break;
+           }
+       cc += size;
+       break;
        default:
        cc = next_opcode(common, cc);
        SLJIT_ASSERT(cc != NULL);
-Line 1021 
 while (status != end)
+Line 1770 
 while (status != end)
          if (!tmp1empty)
            {
            OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP1, 0);
-           stackptr += sizeof(sljit_w);
+           stackptr += sizeof(sljit_sw);
            }
          OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), srcw[count]);
          tmp1empty = FALSE;
-Line 1032 
 while (status != end)
+Line 1781 
 while (status != end)
          if (!tmp2empty)
            {
            OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP2, 0);
-           stackptr += sizeof(sljit_w);
+           stackptr += sizeof(sljit_sw);
            }
          OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), srcw[count]);
          tmp2empty = FALSE;
-Line 1049 
 while (status != end)
+Line 1798 
 while (status != end)
          if (!tmp1empty)
            {
            OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(STACK_TOP), stackptr);
-           stackptr += sizeof(sljit_w);
+           stackptr += sizeof(sljit_sw);
            }
          tmp1next = FALSE;
          }
-Line 1061 
 while (status != end)
+Line 1810 
 while (status != end)
          if (!tmp2empty)
            {
            OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(STACK_TOP), stackptr);
-           stackptr += sizeof(sljit_w);
+           stackptr += sizeof(sljit_sw);
            }
          tmp1next = TRUE;
          }
        }
      }
    }
+ while (status != end);
  if (save)
    {
-Line 1076 
 if (save)
+Line 1826 
 if (save)
      if (!tmp1empty)
        {
        OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP1, 0);
-       stackptr += sizeof(sljit_w);
+       stackptr += sizeof(sljit_sw);
        }
      if (!tmp2empty)
        {
        OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP2, 0);
-       stackptr += sizeof(sljit_w);
+       stackptr += sizeof(sljit_sw);
        }
      }
    else
-Line 1089 
 if (save)
+Line 1839 
 if (save)
      if (!tmp2empty)
        {
        OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP2, 0);
-       stackptr += sizeof(sljit_w);
+       stackptr += sizeof(sljit_sw);
        }
      if (!tmp1empty)
        {
        OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackptr, TMP1, 0);
-       stackptr += sizeof(sljit_w);
+       stackptr += sizeof(sljit_sw);
        }
      }
    }
  SLJIT_ASSERT(cc == ccend && stackptr == stacktop && (save || (tmp1empty && tmp2empty)));
  }
- static SLJIT_INLINE BOOL ispowerof2(unsigned int value)
+ #undef CASE_ITERATOR_PRIVATE_DATA_1
+ #undef CASE_ITERATOR_PRIVATE_DATA_2A
+ #undef CASE_ITERATOR_PRIVATE_DATA_2B
+ #undef CASE_ITERATOR_TYPE_PRIVATE_DATA_1
+ #undef CASE_ITERATOR_TYPE_PRIVATE_DATA_2A
+ #undef CASE_ITERATOR_TYPE_PRIVATE_DATA_2B
+ static SLJIT_INLINE BOOL is_powerof2(unsigned int value)
  {
  return (value & (value - 1)) == 0;
  }
-Line 1111 
 static SLJIT_INLINE void set_jumps(jump_
+Line 1868 
 static SLJIT_INLINE void set_jumps(jump_
  while (list)
    {
    /* sljit_set_label is clever enough to do nothing
-   if either the jump or the label is NULL */
+   if either the jump or the label is NULL. */
-   sljit_set_label(list->jump, label);
+   SET_LABEL(list->jump, label);
    list = list->next;
    }
  }
-Line 1128 
 if (list_item)
+Line 1885 
 if (list_item)
    }
  }
- static void add_stub(compiler_common *common, enum stub_types type, int data, struct sljit_jump *start)
+ static void add_stub(compiler_common *common, struct sljit_jump *start)
  {
  DEFINE_COMPILER;
  stub_list* list_item = sljit_alloc_memory(compiler, sizeof(stub_list));
  if (list_item)
    {
-   list_item->type = type;
-   list_item->data = data;
    list_item->start = start;
-   list_item->leave = LABEL();
+   list_item->quit = LABEL();
    list_item->next = common->stubs;
    common->stubs = list_item;
    }
-Line 1152 
 stub_list* list_item = common->stubs;
+Line 1907 
 stub_list* list_item = common->stubs;
  while (list_item)
    {
    JUMPHERE(list_item->start);
-   switch(list_item->type)
+   add_jump(compiler, &common->stackalloc, JUMP(SLJIT_FAST_CALL));
-     {
+   JUMPTO(SLJIT_JUMP, list_item->quit);
-     case stack_alloc:
-     add_jump(compiler, &common->stackalloc, JUMP(SLJIT_FAST_CALL));
-     break;
-     }
-   JUMPTO(SLJIT_JUMP, list_item->leave);
    list_item = list_item->next;
    }
  common->stubs = NULL;
-Line 1177 
 static SLJIT_INLINE void allocate_stack(
+Line 1927 
 static SLJIT_INLINE void allocate_stack(
  /* May destroy all locals and registers except TMP2. */
  DEFINE_COMPILER;
- OP2(SLJIT_ADD, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, size * sizeof(sljit_w));
+ OP2(SLJIT_ADD, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, size * sizeof(sljit_sw));
  #ifdef DESTROY_REGISTERS
  OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 12345);
  OP1(SLJIT_MOV, TMP3, 0, TMP1, 0);
-Line 1185 
 OP1(SLJIT_MOV, RETURN_ADDR, 0, TMP1, 0);
+Line 1935 
 OP1(SLJIT_MOV, RETURN_ADDR, 0, TMP1, 0);
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0, TMP1, 0);
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1, TMP1, 0);
  #endif
- add_stub(common, stack_alloc, 0, CMP(SLJIT_C_GREATER, STACK_TOP, 0, STACK_LIMIT, 0));
+ add_stub(common, CMP(SLJIT_C_GREATER, STACK_TOP, 0, STACK_LIMIT, 0));
  }
  static SLJIT_INLINE void free_stack(compiler_common *common, int size)
  {
  DEFINE_COMPILER;
- OP2(SLJIT_SUB, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, size * sizeof(sljit_w));
+ OP2(SLJIT_SUB, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, size * sizeof(sljit_sw));
  }
  static SLJIT_INLINE void reset_ovector(compiler_common *common, int length)
-Line 1199 
 static SLJIT_INLINE void reset_ovector(c
+Line 1949 
 static SLJIT_INLINE void reset_ovector(c
  DEFINE_COMPILER;
  struct sljit_label *loop;
  int i;
  /* At this point we can freely use all temporary registers. */
+ SLJIT_ASSERT(length > 1);
  /* TMP1 returns with begin - 1. */
- OP2(SLJIT_SUB, SLJIT_TEMPORARY_REG1, 0, SLJIT_MEM1(SLJIT_GENERAL_REG1), SLJIT_OFFSETOF(jit_arguments, begin), SLJIT_IMM, IN_UCHARS(1));
+ OP2(SLJIT_SUB, SLJIT_SCRATCH_REG1, 0, SLJIT_MEM1(SLJIT_SAVED_REG1), SLJIT_OFFSETOF(jit_arguments, begin), SLJIT_IMM, IN_UCHARS(1));
+ if (length < 8)
+   {
+   for (i = 1; i < length; i++)
+     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(i), SLJIT_SCRATCH_REG1, 0);
+   }
+ else
+   {
+   GET_LOCAL_BASE(SLJIT_SCRATCH_REG2, 0, OVECTOR_START);
+   OP1(SLJIT_MOV, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, length - 1);
+   loop = LABEL();
+   OP1(SLJIT_MOVU, SLJIT_MEM1(SLJIT_SCRATCH_REG2), sizeof(sljit_sw), SLJIT_SCRATCH_REG1, 0);
+   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_SCRATCH_REG3, 0, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, 1);
+   JUMPTO(SLJIT_C_NOT_ZERO, loop);
+   }
+ }
+ static SLJIT_INLINE void do_reset_match(compiler_common *common, int length)
+ {
+ DEFINE_COMPILER;
+ struct sljit_label *loop;
+ int i;
+ SLJIT_ASSERT(length > 1);
+ /* OVECTOR(1) contains the "string begin - 1" constant. */
+ if (length > 2)
+   OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(1));
  if (length < 8)
    {
-   for (i = 0; i < length; i++)
+   for (i = 2; i < length; i++)
-     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(i), SLJIT_TEMPORARY_REG1, 0);
+     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(i), TMP1, 0);
    }
  else
    {
-   OP2(SLJIT_ADD, SLJIT_TEMPORARY_REG2, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, OVECTOR_START - sizeof(sljit_w));
+   GET_LOCAL_BASE(TMP2, 0, OVECTOR_START + sizeof(sljit_sw));
-   OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, length);
+   OP1(SLJIT_MOV, STACK_TOP, 0, SLJIT_IMM, length - 2);
    loop = LABEL();
-   OP1(SLJIT_MOVU, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), sizeof(sljit_w), SLJIT_TEMPORARY_REG1, 0);
+   OP1(SLJIT_MOVU, SLJIT_MEM1(TMP2), sizeof(sljit_sw), TMP1, 0);
-   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_TEMPORARY_REG3, 0, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, 1);
+   OP2(SLJIT_SUB | SLJIT_SET_E, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, 1);
    JUMPTO(SLJIT_C_NOT_ZERO, loop);
    }
+ OP1(SLJIT_MOV, STACK_TOP, 0, ARGUMENTS, 0);
+ if (common->mark_ptr != 0)
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr, SLJIT_IMM, 0);
+ SLJIT_ASSERT(common->control_head_ptr != 0);
+ OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->control_head_ptr, SLJIT_IMM, 0);
+ OP1(SLJIT_MOV, STACK_TOP, 0, SLJIT_MEM1(STACK_TOP), SLJIT_OFFSETOF(jit_arguments, stack));
+ OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_ptr);
+ OP1(SLJIT_MOV, STACK_TOP, 0, SLJIT_MEM1(STACK_TOP), SLJIT_OFFSETOF(struct sljit_stack, base));
+ }
+ static sljit_sw do_check_control_chain(sljit_sw *current)
+ {
+ sljit_sw return_value = 0;
+ SLJIT_ASSERT(current != NULL);
+ do
+   {
+   switch (current[-2])
+     {
+     case type_commit:
+     /* Commit overwrites all. */
+     return -1;
+     case type_prune:
+     break;
+     case type_skip:
+     /* Overwrites prune, but not other skips. */
+     if (return_value == 0)
+       return_value = current[-3];
+     break;
+     default:
+     SLJIT_ASSERT_STOP();
+     break;
+     }
+   current = (sljit_sw*)current[-1];
+   }
+ while (current != NULL);
+ return return_value;
  }
  static SLJIT_INLINE void copy_ovector(compiler_common *common, int topbracket)
  {
  DEFINE_COMPILER;
  struct sljit_label *loop;
- struct sljit_jump *earlyexit;
+ struct sljit_jump *early_quit;
  /* At this point we can freely use all registers. */
- OP1(SLJIT_MOV, SLJIT_GENERAL_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(1));
+ OP1(SLJIT_MOV, SLJIT_SAVED_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(1));
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(1), STR_PTR, 0);
- OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG1, 0, ARGUMENTS, 0);
+ OP1(SLJIT_MOV, SLJIT_SCRATCH_REG1, 0, ARGUMENTS, 0);
- OP1(SLJIT_MOV_SI, SLJIT_TEMPORARY_REG2, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, offsetcount));
+ if (common->mark_ptr != 0)
- OP2(SLJIT_SUB, SLJIT_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, offsets), SLJIT_IMM, sizeof(int));
+   OP1(SLJIT_MOV, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr);
- OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG1, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, begin));
+ OP1(SLJIT_MOV_SI, SLJIT_SCRATCH_REG2, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG1), SLJIT_OFFSETOF(jit_arguments, offset_count));
- OP2(SLJIT_ADD, SLJIT_GENERAL_REG1, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, OVECTOR_START);
+ if (common->mark_ptr != 0)
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_SCRATCH_REG1), SLJIT_OFFSETOF(jit_arguments, mark_ptr), SLJIT_SCRATCH_REG3, 0);
+ OP2(SLJIT_SUB, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG1), SLJIT_OFFSETOF(jit_arguments, offsets), SLJIT_IMM, sizeof(int));
+ OP1(SLJIT_MOV, SLJIT_SCRATCH_REG1, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG1), SLJIT_OFFSETOF(jit_arguments, begin));
+ GET_LOCAL_BASE(SLJIT_SAVED_REG1, 0, OVECTOR_START);
  /* Unlikely, but possible */
- earlyexit = CMP(SLJIT_C_EQUAL, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 0);
+ early_quit = CMP(SLJIT_C_EQUAL, SLJIT_SCRATCH_REG2, 0, SLJIT_IMM, 0);
  loop = LABEL();
- OP2(SLJIT_SUB, SLJIT_GENERAL_REG2, 0, SLJIT_MEM1(SLJIT_GENERAL_REG1), 0, SLJIT_TEMPORARY_REG1, 0);
+ OP2(SLJIT_SUB, SLJIT_SAVED_REG2, 0, SLJIT_MEM1(SLJIT_SAVED_REG1), 0, SLJIT_SCRATCH_REG1, 0);
- OP2(SLJIT_ADD, SLJIT_GENERAL_REG1, 0, SLJIT_GENERAL_REG1, 0, SLJIT_IMM, sizeof(sljit_w));
+ OP2(SLJIT_ADD, SLJIT_SAVED_REG1, 0, SLJIT_SAVED_REG1, 0, SLJIT_IMM, sizeof(sljit_sw));
  /* Copy the integer value to the output buffer */
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
- OP2(SLJIT_ASHR, SLJIT_GENERAL_REG2, 0, SLJIT_GENERAL_REG2, 0, SLJIT_IMM, 1);
+ OP2(SLJIT_ASHR, SLJIT_SAVED_REG2, 0, SLJIT_SAVED_REG2, 0, SLJIT_IMM, UCHAR_SHIFT);
  #endif
- OP1(SLJIT_MOVU_SI, SLJIT_MEM1(SLJIT_TEMPORARY_REG3), sizeof(int), SLJIT_GENERAL_REG2, 0);
+ OP1(SLJIT_MOVU_SI, SLJIT_MEM1(SLJIT_SCRATCH_REG3), sizeof(int), SLJIT_SAVED_REG2, 0);
- OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_TEMPORARY_REG2, 0, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 1);
+ OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_SCRATCH_REG2, 0, SLJIT_SCRATCH_REG2, 0, SLJIT_IMM, 1);
  JUMPTO(SLJIT_C_NOT_ZERO, loop);
- JUMPHERE(earlyexit);
+ JUMPHERE(early_quit);
  /* Calculate the return value, which is the maximum ovector value. */
  if (topbracket > 1)
    {
-   OP2(SLJIT_ADD, SLJIT_TEMPORARY_REG1, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, OVECTOR_START + topbracket * 2 * sizeof(sljit_w));
+   GET_LOCAL_BASE(SLJIT_SCRATCH_REG1, 0, OVECTOR_START + topbracket * 2 * sizeof(sljit_sw));
-   OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, topbracket + 1);
+   OP1(SLJIT_MOV, SLJIT_SCRATCH_REG2, 0, SLJIT_IMM, topbracket + 1);
-   /* OVECTOR(0) is never equal to SLJIT_GENERAL_REG3. */
+   /* OVECTOR(0) is never equal to SLJIT_SAVED_REG3. */
    loop = LABEL();
-   OP1(SLJIT_MOVU, SLJIT_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), -(2 * (sljit_w)sizeof(sljit_w)));
+   OP1(SLJIT_MOVU, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG1), -(2 * (sljit_sw)sizeof(sljit_sw)));
-   OP2(SLJIT_SUB, SLJIT_TEMPORARY_REG2, 0, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 1);
+   OP2(SLJIT_SUB, SLJIT_SCRATCH_REG2, 0, SLJIT_SCRATCH_REG2, 0, SLJIT_IMM, 1);
-   CMPTO(SLJIT_C_EQUAL, SLJIT_TEMPORARY_REG3, 0, SLJIT_GENERAL_REG3, 0, loop);
+   CMPTO(SLJIT_C_EQUAL, SLJIT_SCRATCH_REG3, 0, SLJIT_SAVED_REG3, 0, loop);
-   OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_TEMPORARY_REG2, 0);
+   OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_SCRATCH_REG2, 0);
    }
  else
    OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_IMM, 1);
  }
+ static SLJIT_INLINE void return_with_partial_match(compiler_common *common, struct sljit_label *quit)
+ {
+ DEFINE_COMPILER;
+ struct sljit_jump *jump;
+ SLJIT_COMPILE_ASSERT(STR_END == SLJIT_SAVED_REG2, str_end_must_be_saved_reg2);
+ SLJIT_ASSERT(common->start_used_ptr != 0 && common->start_ptr != 0
+   && (common->mode == JIT_PARTIAL_SOFT_COMPILE ? common->hit_start != 0 : common->hit_start == 0));
+ OP1(SLJIT_MOV, SLJIT_SCRATCH_REG2, 0, ARGUMENTS, 0);
+ OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_IMM, PCRE_ERROR_PARTIAL);
+ OP1(SLJIT_MOV_SI, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG2), SLJIT_OFFSETOF(jit_arguments, real_offset_count));
+ CMPTO(SLJIT_C_SIG_LESS, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, 2, quit);
+ /* Store match begin and end. */
+ OP1(SLJIT_MOV, SLJIT_SAVED_REG1, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG2), SLJIT_OFFSETOF(jit_arguments, begin));
+ OP1(SLJIT_MOV, SLJIT_SCRATCH_REG2, 0, SLJIT_MEM1(SLJIT_SCRATCH_REG2), SLJIT_OFFSETOF(jit_arguments, offsets));
+ jump = CMP(SLJIT_C_SIG_LESS, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, 3);
+ OP2(SLJIT_SUB, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mode == JIT_PARTIAL_HARD_COMPILE ? common->start_ptr : (common->hit_start + sizeof(sljit_sw)), SLJIT_SAVED_REG1, 0);
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
+ OP2(SLJIT_ASHR, SLJIT_SCRATCH_REG3, 0, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, UCHAR_SHIFT);
+ #endif
+ OP1(SLJIT_MOV_SI, SLJIT_MEM1(SLJIT_SCRATCH_REG2), 2 * sizeof(int), SLJIT_SCRATCH_REG3, 0);
+ JUMPHERE(jump);
+ OP1(SLJIT_MOV, SLJIT_SCRATCH_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mode == JIT_PARTIAL_HARD_COMPILE ? common->start_used_ptr : common->hit_start);
+ OP2(SLJIT_SUB, SLJIT_SAVED_REG2, 0, STR_END, 0, SLJIT_SAVED_REG1, 0);
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
+ OP2(SLJIT_ASHR, SLJIT_SAVED_REG2, 0, SLJIT_SAVED_REG2, 0, SLJIT_IMM, UCHAR_SHIFT);
+ #endif
+ OP1(SLJIT_MOV_SI, SLJIT_MEM1(SLJIT_SCRATCH_REG2), sizeof(int), SLJIT_SAVED_REG2, 0);
+ OP2(SLJIT_SUB, SLJIT_SCRATCH_REG3, 0, SLJIT_SCRATCH_REG3, 0, SLJIT_SAVED_REG1, 0);
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
+ OP2(SLJIT_ASHR, SLJIT_SCRATCH_REG3, 0, SLJIT_SCRATCH_REG3, 0, SLJIT_IMM, UCHAR_SHIFT);
+ #endif
+ OP1(SLJIT_MOV_SI, SLJIT_MEM1(SLJIT_SCRATCH_REG2), 0, SLJIT_SCRATCH_REG3, 0);
+ JUMPTO(SLJIT_JUMP, quit);
+ }
+ static SLJIT_INLINE void check_start_used_ptr(compiler_common *common)
+ {
+ /* May destroy TMP1. */
+ DEFINE_COMPILER;
+ struct sljit_jump *jump;
+ if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
+   {
+   /* The value of -1 must be kept for start_used_ptr! */
+   OP2(SLJIT_ADD, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, SLJIT_IMM, 1);
+   /* Jumps if start_used_ptr < STR_PTR, or start_used_ptr == -1. Although overwriting
+   is not necessary if start_used_ptr == STR_PTR, it does not hurt as well. */
+   jump = CMP(SLJIT_C_LESS_EQUAL, TMP1, 0, STR_PTR, 0);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0);
+   JUMPHERE(jump);
+   }
+ else if (common->mode == JIT_PARTIAL_HARD_COMPILE)
+   {
+   jump = CMP(SLJIT_C_LESS_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0);
+   JUMPHERE(jump);
+   }
+ }
  static SLJIT_INLINE BOOL char_has_othercase(compiler_common *common, pcre_uchar* cc)
  {
  /* Detects if the character has an othercase. */
-Line 1348 
 if (c <= 127 && bit == 0x20)
+Line 2237 
 if (c <= 127 && bit == 0x20)
    return (0 << 8) | 0x20;
  /* Since c != oc, they must have at least 1 bit difference. */
- if (!ispowerof2(bit))
+ if (!is_powerof2(bit))
    return 0;
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
  #ifdef SUPPORT_UTF
  if (common->utf && c > 127)
-Line 1367 
 if (common->utf && c > 127)
+Line 2256 
 if (common->utf && c > 127)
  #endif /* SUPPORT_UTF */
  return (0 << 8) | bit;
- #else /* COMPILE_PCRE8 */
+ #elif defined COMPILE_PCRE16 || defined COMPILE_PCRE32
- #ifdef COMPILE_PCRE16
  #ifdef SUPPORT_UTF
  if (common->utf && c > 65535)
    {
-Line 1380 
 if (common->utf && c > 65535)
+Line 2268 
 if (common->utf && c > 65535)
    }
  #endif /* SUPPORT_UTF */
  return (bit < 256) ? ((0 << 8) | bit) : ((1 << 8) | (bit >> 8));
- #endif /* COMPILE_PCRE16 */
- #endif /* COMPILE_PCRE8 */
+ #endif /* COMPILE_PCRE[8|16|32] */
+ }
+ static void check_partial(compiler_common *common, BOOL force)
+ {
+ /* Checks whether a partial matching is occured. Does not modify registers. */
+ DEFINE_COMPILER;
+ struct sljit_jump *jump = NULL;
+ SLJIT_ASSERT(!force || common->mode != JIT_COMPILE);
+ if (common->mode == JIT_COMPILE)
+   return;
+ if (!force)
+   jump = CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0);
+ else if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
+   jump = CMP(SLJIT_C_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, SLJIT_IMM, -1);
+ if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, 0);
+ else
+   {
+   if (common->partialmatchlabel != NULL)
+     JUMPTO(SLJIT_JUMP, common->partialmatchlabel);
+   else
+     add_jump(compiler, &common->partialmatch, JUMP(SLJIT_JUMP));
+   }
+ if (jump != NULL)
+   JUMPHERE(jump);
+ }
+ static void check_str_end(compiler_common *common, jump_list **end_reached)
+ {
+ /* Does not affect registers. Usually used in a tight spot. */
+ DEFINE_COMPILER;
+ struct sljit_jump *jump;
+ if (common->mode == JIT_COMPILE)
+   {
+   add_jump(compiler, end_reached, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
+   return;
+   }
+ jump = CMP(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0);
+ if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
+   {
+   add_jump(compiler, end_reached, CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0));
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, 0);
+   add_jump(compiler, end_reached, JUMP(SLJIT_JUMP));
+   }
+ else
+   {
+   add_jump(compiler, end_reached, CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0));
+   if (common->partialmatchlabel != NULL)
+     JUMPTO(SLJIT_JUMP, common->partialmatchlabel);
+   else
+     add_jump(compiler, &common->partialmatch, JUMP(SLJIT_JUMP));
+   }
+ JUMPHERE(jump);
  }
- static SLJIT_INLINE void check_input_end(compiler_common *common, jump_list **fallbacks)
+ static void detect_partial_match(compiler_common *common, jump_list **backtracks)
  {
  DEFINE_COMPILER;
- add_jump(compiler, fallbacks, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
+ struct sljit_jump *jump;
+ if (common->mode == JIT_COMPILE)
+   {
+   add_jump(compiler, backtracks, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
+   return;
+   }
+ /* Partial matching mode. */
+ jump = CMP(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0);
+ add_jump(compiler, backtracks, CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0));
+ if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
+   {
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, 0);
+   add_jump(compiler, backtracks, JUMP(SLJIT_JUMP));
+   }
+ else
+   {
+   if (common->partialmatchlabel != NULL)
+     JUMPTO(SLJIT_JUMP, common->partialmatchlabel);
+   else
+     add_jump(compiler, &common->partialmatch, JUMP(SLJIT_JUMP));
+   }
+ JUMPHERE(jump);
  }
  static void read_char(compiler_common *common)
-Line 1396 
 static void read_char(compiler_common *c
+Line 2366 
 static void read_char(compiler_common *c
  /* Reads the character into TMP1, updates STR_PTR.
  Does not check STR_END. TMP2 Destroyed. */
  DEFINE_COMPILER;
- #ifdef SUPPORT_UTF
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
  struct sljit_jump *jump;
  #endif
  OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
- #ifdef SUPPORT_UTF
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
  if (common->utf)
    {
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
    jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800);
- #endif
+ #endif /* COMPILE_PCRE[8|16] */
- #endif /* COMPILE_PCRE8 */
    add_jump(compiler, &common->utfreadchar, JUMP(SLJIT_FAST_CALL));
    JUMPHERE(jump);
    }
- #endif
+ #endif /* SUPPORT_UTF && !COMPILE_PCRE32 */
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
  }
-Line 1423 
 static void peek_char(compiler_common *c
+Line 2391 
 static void peek_char(compiler_common *c
  /* Reads the character into TMP1, keeps STR_PTR.
  Does not check STR_END. TMP2 Destroyed. */
  DEFINE_COMPILER;
- #ifdef SUPPORT_UTF
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
  struct sljit_jump *jump;
  #endif
  OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
- #ifdef SUPPORT_UTF
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
  if (common->utf)
    {
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
    jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800);
- #endif
+ #endif /* COMPILE_PCRE[8|16] */
- #endif /* COMPILE_PCRE8 */
    add_jump(compiler, &common->utfreadchar, JUMP(SLJIT_FAST_CALL));
    OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
    JUMPHERE(jump);
    }
- #endif
+ #endif /* SUPPORT_UTF && !COMPILE_PCRE32 */
  }
  static void read_char8_type(compiler_common *common)
  {
  /* Reads the character type into TMP1, updates STR_PTR. Does not check STR_END. */
  DEFINE_COMPILER;
- #if defined SUPPORT_UTF || defined COMPILE_PCRE16
+ #if defined SUPPORT_UTF || defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  struct sljit_jump *jump;
  #endif
-Line 1458 
 if (common->utf)
+Line 2424 
 if (common->utf)
    {
    OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), 0);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    /* This can be an extra read in some situations, but hopefully
    it is needed in most cases. */
    OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), common->ctypes);
    jump = CMP(SLJIT_C_LESS, TMP2, 0, SLJIT_IMM, 0xc0);
    add_jump(compiler, &common->utfreadtype8, JUMP(SLJIT_FAST_CALL));
    JUMPHERE(jump);
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
    OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
    jump = CMP(SLJIT_C_GREATER, TMP2, 0, SLJIT_IMM, 255);
    OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), common->ctypes);
-Line 1474 
 if (common->utf)
+Line 2439 
 if (common->utf)
    /* Skip low surrogate if necessary. */
    OP2(SLJIT_AND, TMP2, 0, TMP2, 0, SLJIT_IMM, 0xfc00);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_IMM, 0xd800);
-   COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, 1);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
- #endif
+ #elif defined COMPILE_PCRE32
- #endif /* COMPILE_PCRE8 */
+   OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
+   jump = CMP(SLJIT_C_GREATER, TMP2, 0, SLJIT_IMM, 255);
+   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), common->ctypes);
+   JUMPHERE(jump);
+ #endif /* COMPILE_PCRE[8|16|32] */
    return;
    }
- #endif
+ #endif /* SUPPORT_UTF */
  OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), 0);
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  /* The ctypes array contains only 256 values. */
  OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
  jump = CMP(SLJIT_C_GREATER, TMP2, 0, SLJIT_IMM, 255);
  #endif
  OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), common->ctypes);
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  JUMPHERE(jump);
  #endif
  }
-Line 1499 
 static void skip_char_back(compiler_comm
+Line 2468 
 static void skip_char_back(compiler_comm
  {
  /* Goes one character back. Affects STR_PTR and TMP1. Does not check begin. */
  DEFINE_COMPILER;
- #if defined SUPPORT_UTF && defined COMPILE_PCRE8
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
+ #if defined COMPILE_PCRE8
  struct sljit_label *label;
  if (common->utf)
-Line 1511 
 if (common->utf)
+Line 2481 
 if (common->utf)
    CMPTO(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, 0x80, label);
    return;
    }
- #endif
+ #elif defined COMPILE_PCRE16
- #if defined SUPPORT_UTF && defined COMPILE_PCRE16
  if (common->utf)
    {
    OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), -IN_UCHARS(1));
-Line 1520 
 if (common->utf)
+Line 2489 
 if (common->utf)
    /* Skip low surrogate if necessary. */
    OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xdc00);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP1, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
    OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
    return;
    }
- #endif
+ #endif /* COMPILE_PCRE[8|16] */
+ #endif /* SUPPORT_UTF && !COMPILE_PCRE32 */
  OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
  }
- static void check_newlinechar(compiler_common *common, int nltype, jump_list **fallbacks, BOOL jumpiftrue)
+ static void check_newlinechar(compiler_common *common, int nltype, jump_list **backtracks, BOOL jumpiftrue)
  {
  /* Character comes in TMP1. Checks if it is a newline. TMP2 may be destroyed. */
  DEFINE_COMPILER;
-Line 1537 
 DEFINE_COMPILER;
+Line 2507 
 DEFINE_COMPILER;
  if (nltype == NLTYPE_ANY)
    {
    add_jump(compiler, &common->anynewline, JUMP(SLJIT_FAST_CALL));
-   add_jump(compiler, fallbacks, JUMP(jumpiftrue ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
+   add_jump(compiler, backtracks, JUMP(jumpiftrue ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
    }
  else if (nltype == NLTYPE_ANYCRLF)
    {
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, CHAR_CR);
-   COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, CHAR_NL);
-   COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
-   add_jump(compiler, fallbacks, JUMP(jumpiftrue ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
+   add_jump(compiler, backtracks, JUMP(jumpiftrue ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
    }
  else
    {
    SLJIT_ASSERT(nltype == NLTYPE_FIXED && common->newline < 256);
-   add_jump(compiler, fallbacks, CMP(jumpiftrue ? SLJIT_C_EQUAL : SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, common->newline));
+   add_jump(compiler, backtracks, CMP(jumpiftrue ? SLJIT_C_EQUAL : SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, common->newline));
    }
  }
  #ifdef SUPPORT_UTF
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
  static void do_utfreadchar(compiler_common *common)
  {
  /* Fast decoding a UTF-8 character. TMP1 contains the first byte
-Line 1564 
 of the character (>= 0xc0). Return char
+Line 2534 
 of the character (>= 0xc0). Return char
  DEFINE_COMPILER;
  struct sljit_jump *jump;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  /* Searching for the first zero. */
  OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x20);
  jump = JUMP(SLJIT_C_NOT_ZERO);
-Line 1623 
 DEFINE_COMPILER;
+Line 2593 
 DEFINE_COMPILER;
  struct sljit_jump *jump;
  struct sljit_jump *compare;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_IMM, 0x20);
  jump = JUMP(SLJIT_C_NOT_ZERO);
-Line 1644 
 sljit_emit_fast_return(compiler, RETURN_
+Line 2614 
 sljit_emit_fast_return(compiler, RETURN_
  JUMPHERE(jump);
  /* We only have types for characters less than 256. */
- OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), (sljit_w)PRIV(utf8_table4) - 0xc0);
+ OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP2), (sljit_sw)PRIV(utf8_table4) - 0xc0);
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
  OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 0);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
- #else /* COMPILE_PCRE8 */
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
  static void do_utfreadchar(compiler_common *common)
  {
  /* Fast decoding a UTF-16 character. TMP1 contains the first 16 bit char
-Line 1660 
 of the character (>= 0xd800). Return cha
+Line 2629 
 of the character (>= 0xd800). Return cha
  DEFINE_COMPILER;
  struct sljit_jump *jump;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xdc00);
  /* Do nothing, only return. */
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
-Line 1677 
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UC
+Line 2646 
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UC
  OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x10000);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
- #endif /* COMPILE_PCRE16 */
- #endif /* COMPILE_PCRE8 */
+ #endif /* COMPILE_PCRE[8|16] */
  #endif /* SUPPORT_UTF */
-Line 1697 
 DEFINE_COMPILER;
+Line 2665 
 DEFINE_COMPILER;
  SLJIT_ASSERT(UCD_BLOCK_SIZE == 128 && sizeof(ucd_record) == 8);
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_LSHR, TMP2, 0, TMP1, 0, SLJIT_IMM, UCD_BLOCK_SHIFT);
- OP1(SLJIT_MOV_UB, TMP2, 0, SLJIT_MEM1(TMP2), (sljit_w)PRIV(ucd_stage1));
+ OP1(SLJIT_MOV_UB, TMP2, 0, SLJIT_MEM1(TMP2), (sljit_sw)PRIV(ucd_stage1));
  OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, UCD_BLOCK_MASK);
  OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, UCD_BLOCK_SHIFT);
  OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, TMP2, 0);
- OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, (sljit_w)PRIV(ucd_stage2));
+ OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_stage2));
  OP1(SLJIT_MOV_UH, TMP2, 0, SLJIT_MEM2(TMP2, TMP1), 1);
- OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, (sljit_w)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, chartype));
+ OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, chartype));
  OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM2(TMP1, TMP2), 3);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
-Line 1719 
 struct sljit_label *newlinelabel = NULL;
+Line 2687 
 struct sljit_label *newlinelabel = NULL;
  struct sljit_jump *start;
  struct sljit_jump *end = NULL;
  struct sljit_jump *nl = NULL;
- #ifdef SUPPORT_UTF
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
  struct sljit_jump *singlechar;
  #endif
  jump_list *newline = NULL;
-Line 1733 
 if (!(hascrorlf || firstline) && (common
+Line 2701 
 if (!(hascrorlf || firstline) && (common
  if (firstline)
    {
    /* Search for the end of the first line. */
-   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0, STR_PTR, 0);
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END, STR_END, 0);
+   OP1(SLJIT_MOV, TMP3, 0, STR_PTR, 0);
    if (common->nltype == NLTYPE_FIXED && common->newline > 255)
      {
-Line 1745 
 if (firstline)
+Line 2713 
 if (firstline)
      OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(0));
      CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff, mainloop);
      CMPTO(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, common->newline & 0xff, mainloop);
-     OP2(SLJIT_SUB, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+     JUMPHERE(end);
+     OP2(SLJIT_SUB, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+     }
+   else
+     {
+     end = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+     mainloop = LABEL();
+     /* Continual stores does not cause data dependency. */
+     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, STR_PTR, 0);
+     read_char(common);
+     check_newlinechar(common, common->nltype, &newline, TRUE);
+     CMPTO(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0, mainloop);
+     JUMPHERE(end);
+     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, STR_PTR, 0);
+     set_jumps(newline, LABEL());
+     }
+   OP1(SLJIT_MOV, STR_PTR, 0, TMP3, 0);
+   }
+ start = JUMP(SLJIT_JUMP);
+ if (newlinecheck)
+   {
+   newlinelabel = LABEL();
+   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+   end = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
+   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, common->newline & 0xff);
+   OP_FLAGS(SLJIT_MOV, TMP1, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
+   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, UCHAR_SHIFT);
+ #endif
+   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
+   nl = JUMP(SLJIT_JUMP);
+   }
+ mainloop = LABEL();
+ /* Increasing the STR_PTR here requires one less jump in the most common case. */
+ #ifdef SUPPORT_UTF
+ if (common->utf) readuchar = TRUE;
+ #endif
+ if (newlinecheck) readuchar = TRUE;
+ if (readuchar)
+   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
+ if (newlinecheck)
+   CMPTO(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff, newlinelabel);
+ OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+ #if defined SUPPORT_UTF && !defined COMPILE_PCRE32
+ #if defined COMPILE_PCRE8
+ if (common->utf)
+   {
+   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
+   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_sw)PRIV(utf8_table4) - 0xc0);
+   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
+   JUMPHERE(singlechar);
+   }
+ #elif defined COMPILE_PCRE16
+ if (common->utf)
+   {
+   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800);
+   OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
+   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xd800);
+   OP_FLAGS(SLJIT_MOV, TMP1, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
+   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
+   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
+   JUMPHERE(singlechar);
+   }
+ #endif /* COMPILE_PCRE[8|16] */
+ #endif /* SUPPORT_UTF && !COMPILE_PCRE32 */
+ JUMPHERE(start);
+ if (newlinecheck)
+   {
+   JUMPHERE(end);
+   JUMPHERE(nl);
+   }
+ return mainloop;
+ }
+ #define MAX_N_CHARS 3
+ static SLJIT_INLINE BOOL fast_forward_first_n_chars(compiler_common *common, BOOL firstline)
+ {
+ DEFINE_COMPILER;
+ struct sljit_label *start;
+ struct sljit_jump *quit;
+ pcre_uint32 chars[MAX_N_CHARS * 2];
+ pcre_uchar *cc = common->start + 1 + LINK_SIZE;
+ int location = 0;
+ pcre_int32 len, c, bit, caseless;
+ int must_stop;
+ /* We do not support alternatives now. */
+ if (*(common->start + GET(common->start, 1)) == OP_ALT)
+   return FALSE;
+ while (TRUE)
+   {
+   caseless = 0;
+   must_stop = 1;
+   switch(*cc)
+     {
+     case OP_CHAR:
+     must_stop = 0;
+     cc++;
+     break;
+     case OP_CHARI:
+     caseless = 1;
+     must_stop = 0;
+     cc++;
+     break;
+     case OP_SOD:
+     case OP_SOM:
+     case OP_SET_SOM:
+     case OP_NOT_WORD_BOUNDARY:
+     case OP_WORD_BOUNDARY:
+     case OP_EODN:
+     case OP_EOD:
+     case OP_CIRC:
+     case OP_CIRCM:
+     case OP_DOLL:
+     case OP_DOLLM:
+     /* Zero width assertions. */
+     cc++;
+     continue;
+     case OP_PLUS:
+     case OP_MINPLUS:
+     case OP_POSPLUS:
+     cc++;
+     break;
+     case OP_EXACT:
+     cc += 1 + IMM2_SIZE;
+     break;
+     case OP_PLUSI:
+     case OP_MINPLUSI:
+     case OP_POSPLUSI:
+     caseless = 1;
+     cc++;
+     break;
+     case OP_EXACTI:
+     caseless = 1;
+     cc += 1 + IMM2_SIZE;
+     break;
+     default:
+     must_stop = 2;
+     break;
+     }
+   if (must_stop == 2)
+       break;
+   len = 1;
+ #ifdef SUPPORT_UTF
+   if (common->utf && HAS_EXTRALEN(cc[0])) len += GET_EXTRALEN(cc[0]);
+ #endif
+   if (caseless && char_has_othercase(common, cc))
+     {
+     caseless = char_get_othercase_bit(common, cc);
+     if (caseless == 0)
+       return FALSE;
+ #ifdef COMPILE_PCRE8
+     caseless = ((caseless & 0xff) << 8) | (len - (caseless >> 8));
+ #else
+     if ((caseless & 0x100) != 0)
+       caseless = ((caseless & 0xff) << 16) | (len - (caseless >> 9));
+     else
+       caseless = ((caseless & 0xff) << 8) | (len - (caseless >> 9));
+ #endif
      }
    else
+     caseless = 0;
+   while (len > 0 && location < MAX_N_CHARS * 2)
      {
-     end = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+     c = *cc;
-     mainloop = LABEL();
+     bit = 0;
-     /* Continual stores does not cause data dependency. */
+     if (len == (caseless & 0xff))
-     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END, STR_PTR, 0);
+       {
-     read_char(common);
+       bit = caseless >> 8;
-     check_newlinechar(common, common->nltype, &newline, TRUE);
+       c |= bit;
-     CMPTO(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0, mainloop);
+       }
-     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END, STR_PTR, 0);
-     set_jumps(newline, LABEL());
+     chars[location] = c;
+     chars[location + 1] = bit;
+     len--;
+     location += 2;
+     cc++;
      }
-   JUMPHERE(end);
+   if (location >= MAX_N_CHARS * 2 || must_stop != 0)
-   OP1(SLJIT_MOV, STR_PTR, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0);
+     break;
    }
- start = JUMP(SLJIT_JUMP);
+ /* At least two characters are required. */
+ if (location < 2 * 2)
+     return FALSE;
- if (newlinecheck)
+ if (firstline)
    {
-   newlinelabel = LABEL();
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+   OP1(SLJIT_MOV, TMP3, 0, STR_END, 0);
-   end = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+   OP2(SLJIT_SUB, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, SLJIT_IMM, IN_UCHARS((location >> 1) - 1));
-   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
-   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, common->newline & 0xff);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
- #ifdef COMPILE_PCRE16
-   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
- #endif
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
-   nl = JUMP(SLJIT_JUMP);
    }
+ else
+   OP2(SLJIT_SUB, STR_END, 0, STR_END, 0, SLJIT_IMM, IN_UCHARS((location >> 1) - 1));
- mainloop = LABEL();
+ start = LABEL();
+ quit = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
- /* Increasing the STR_PTR here requires one less jump in the most common case. */
- #ifdef SUPPORT_UTF
- if (common->utf) readuchar = TRUE;
- #endif
- if (newlinecheck) readuchar = TRUE;
- if (readuchar)
-   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
- if (newlinecheck)
-   CMPTO(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff, newlinelabel);
+ OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(0));
+ OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- #if defined SUPPORT_UTF && defined COMPILE_PCRE8
+ if (chars[1] != 0)
- if (common->utf)
+   OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, chars[1]);
-   {
+ CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, chars[0], start);
-   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
+ if (location > 2 * 2)
-   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_w)PRIV(utf8_table4) - 0xc0);
+   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
+ if (chars[3] != 0)
-   JUMPHERE(singlechar);
+   OP2(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_IMM, chars[3]);
-   }
+ CMPTO(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, chars[2], start);
- #endif
+ if (location > 2 * 2)
- #if defined SUPPORT_UTF && defined COMPILE_PCRE16
+   {
- if (common->utf)
+   if (chars[5] != 0)
-   {
+     OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, chars[5]);
-   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800);
+   CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, chars[4], start);
-   OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
-   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xd800);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
-   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
-   JUMPHERE(singlechar);
    }
- #endif
+ OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- JUMPHERE(start);
- if (newlinecheck)
+ JUMPHERE(quit);
-   {
-   JUMPHERE(end);
-   JUMPHERE(nl);
-   }
- return mainloop;
+ if (firstline)
+   OP1(SLJIT_MOV, STR_END, 0, TMP3, 0);
+ else
+   OP2(SLJIT_ADD, STR_END, 0, STR_END, 0, SLJIT_IMM, IN_UCHARS((location >> 1) - 1));
+ return TRUE;
  }
+ #undef MAX_N_CHARS
  static SLJIT_INLINE void fast_forward_first_char(compiler_common *common, pcre_uchar first_char, BOOL caseless, BOOL firstline)
  {
  DEFINE_COMPILER;
  struct sljit_label *start;
- struct sljit_jump *leave;
+ struct sljit_jump *quit;
  struct sljit_jump *found;
  pcre_uchar oc, bit;
  if (firstline)
    {
-   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0, STR_END, 0);
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END);
+   OP1(SLJIT_MOV, TMP3, 0, STR_END, 0);
+   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end);
    }
  start = LABEL();
- leave = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+ quit = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
  OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
  oc = first_char;
-Line 1860 
 if (first_char == oc)
+Line 3000 
 if (first_char == oc)
  else
    {
    bit = first_char ^ oc;
-   if (ispowerof2(bit))
+   if (is_powerof2(bit))
      {
      OP2(SLJIT_OR, TMP2, 0, TMP1, 0, SLJIT_IMM, bit);
      found = CMP(SLJIT_C_EQUAL, TMP2, 0, SLJIT_IMM, first_char | bit);
-Line 1868 
 else
+Line 3008 
 else
    else
      {
      OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, first_char);
-     COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_EQUAL);
+     OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
      OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, oc);
-     COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+     OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
      found = JUMP(SLJIT_C_NOT_ZERO);
      }
    }
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- #if defined SUPPORT_UTF && defined COMPILE_PCRE8
- if (common->utf)
-   {
-   CMPTO(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0, start);
-   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_w)PRIV(utf8_table4) - 0xc0);
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
-   }
- #endif
- #if defined SUPPORT_UTF && defined COMPILE_PCRE16
- if (common->utf)
-   {
-   CMPTO(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800, start);
-   OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
-   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xd800);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
-   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
-   }
- #endif
  JUMPTO(SLJIT_JUMP, start);
  JUMPHERE(found);
- JUMPHERE(leave);
+ JUMPHERE(quit);
  if (firstline)
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0);
+   OP1(SLJIT_MOV, STR_END, 0, TMP3, 0);
  }
  static SLJIT_INLINE void fast_forward_newline(compiler_common *common, BOOL firstline)
-Line 1909 
 DEFINE_COMPILER;
+Line 3030 
 DEFINE_COMPILER;
  struct sljit_label *loop;
  struct sljit_jump *lastchar;
  struct sljit_jump *firstchar;
- struct sljit_jump *leave;
+ struct sljit_jump *quit;
  struct sljit_jump *foundcr = NULL;
  struct sljit_jump *notfoundnl;
  jump_list *newline = NULL;
  if (firstline)
    {
-   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0, STR_END, 0);
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END);
+   OP1(SLJIT_MOV, TMP3, 0, STR_END, 0);
+   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end);
    }
  if (common->nltype == NLTYPE_FIXED && common->newline > 255)
-Line 1930 
 if (common->nltype == NLTYPE_FIXED && co
+Line 3052 
 if (common->nltype == NLTYPE_FIXED && co
    OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, IN_UCHARS(2));
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, STR_PTR, 0, TMP1, 0);
-   COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_GREATER_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_GREATER_EQUAL);
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
-   OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, 1);
+   OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, UCHAR_SHIFT);
  #endif
    OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
    loop = LABEL();
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
-   leave = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+   quit = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
    OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(-2));
    OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(-1));
    CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff, loop);
    CMPTO(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, common->newline & 0xff, loop);
-   JUMPHERE(leave);
+   JUMPHERE(quit);
    JUMPHERE(firstchar);
    JUMPHERE(lastchar);
-Line 1968 
 set_jumps(newline, loop);
+Line 3090 
 set_jumps(newline, loop);
  if (common->nltype == NLTYPE_ANY || common->nltype == NLTYPE_ANYCRLF)
    {
-   leave = JUMP(SLJIT_JUMP);
+   quit = JUMP(SLJIT_JUMP);
    JUMPHERE(foundcr);
    notfoundnl = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
    OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, CHAR_NL);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP1, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
-   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
+   OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, UCHAR_SHIFT);
  #endif
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
    JUMPHERE(notfoundnl);
-   JUMPHERE(leave);
+   JUMPHERE(quit);
    }
  JUMPHERE(lastchar);
  JUMPHERE(firstchar);
  if (firstline)
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0);
+   OP1(SLJIT_MOV, STR_END, 0, TMP3, 0);
  }
+ static BOOL check_class_ranges(compiler_common *common, const pcre_uint8 *bits, BOOL nclass, jump_list **backtracks);
  static SLJIT_INLINE void fast_forward_start_bits(compiler_common *common, sljit_uw start_bits, BOOL firstline)
  {
  DEFINE_COMPILER;
  struct sljit_label *start;
- struct sljit_jump *leave;
+ struct sljit_jump *quit;
- struct sljit_jump *found;
+ struct sljit_jump *found = NULL;
+ jump_list *matches = NULL;
+ pcre_uint8 inverted_start_bits[32];
+ int i;
  #ifndef COMPILE_PCRE8
  struct sljit_jump *jump;
  #endif
+ for (i = 0; i < 32; ++i)
+   inverted_start_bits[i] = ~(((pcre_uint8*)start_bits)[i]);
  if (firstline)
    {
-   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0, STR_END, 0);
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END);
+   OP1(SLJIT_MOV, RETURN_ADDR, 0, STR_END, 0);
+   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end);
    }
  start = LABEL();
- leave = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+ quit = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
  OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
  #ifdef SUPPORT_UTF
  if (common->utf)
    OP1(SLJIT_MOV, TMP3, 0, TMP1, 0);
  #endif
+ if (!check_class_ranges(common, inverted_start_bits, (inverted_start_bits[31] & 0x80) != 0, &matches))
+   {
  #ifndef COMPILE_PCRE8
- jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 255);
+   jump = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 255);
- OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 255);
+   OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, 255);
- JUMPHERE(jump);
+   JUMPHERE(jump);
  #endif
- OP2(SLJIT_AND, TMP2, 0, TMP1, 0, SLJIT_IMM, 0x7);
+   OP2(SLJIT_AND, TMP2, 0, TMP1, 0, SLJIT_IMM, 0x7);
- OP2(SLJIT_LSHR, TMP1, 0, TMP1, 0, SLJIT_IMM, 3);
+   OP2(SLJIT_LSHR, TMP1, 0, TMP1, 0, SLJIT_IMM, 3);
- OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), start_bits);
+   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), start_bits);
- OP2(SLJIT_SHL, TMP2, 0, SLJIT_IMM, 1, TMP2, 0);
+   OP2(SLJIT_SHL, TMP2, 0, SLJIT_IMM, 1, TMP2, 0);
- OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, TMP2, 0);
+   OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, TMP2, 0);
- found = JUMP(SLJIT_C_NOT_ZERO);
+   found = JUMP(SLJIT_C_NOT_ZERO);
+   }
  #ifdef SUPPORT_UTF
  if (common->utf)
    OP1(SLJIT_MOV, TMP1, 0, TMP3, 0);
  #endif
  OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- #if defined SUPPORT_UTF && defined COMPILE_PCRE8
+ #ifdef SUPPORT_UTF
+ #if defined COMPILE_PCRE8
  if (common->utf)
    {
    CMPTO(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0, start);
-   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_w)PRIV(utf8_table4) - 0xc0);
+   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_sw)PRIV(utf8_table4) - 0xc0);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
    }
- #endif
+ #elif defined COMPILE_PCRE16
- #if defined SUPPORT_UTF && defined COMPILE_PCRE16
  if (common->utf)
    {
    CMPTO(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800, start);
    OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xd800);
-   COND_VALUE(SLJIT_MOV, TMP1, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP1, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SHL, TMP1, 0, TMP1, 0, SLJIT_IMM, 1);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
    }
- #endif
+ #endif /* COMPILE_PCRE[8|16] */
+ #endif /* SUPPORT_UTF */
  JUMPTO(SLJIT_JUMP, start);
- JUMPHERE(found);
+ if (found != NULL)
- JUMPHERE(leave);
+   JUMPHERE(found);
+ if (matches != NULL)
+   set_jumps(matches, LABEL());
+ JUMPHERE(quit);
  if (firstline)
-   OP1(SLJIT_MOV, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), POSSESSIVE0);
+   OP1(SLJIT_MOV, STR_END, 0, RETURN_ADDR, 0);
  }
  static SLJIT_INLINE struct sljit_jump *search_requested_char(compiler_common *common, pcre_uchar req_char, BOOL caseless, BOOL has_firstchar)
-Line 2064 
 struct sljit_jump *alreadyfound;
+Line 3203 
 struct sljit_jump *alreadyfound;
  struct sljit_jump *found;
  struct sljit_jump *foundoc = NULL;
  struct sljit_jump *notfound;
- pcre_uchar oc, bit;
+ pcre_uint32 oc, bit;
- OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), REQ_CHAR_PTR);
+ SLJIT_ASSERT(common->req_char_ptr != 0);
+ OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->req_char_ptr);
  OP2(SLJIT_ADD, TMP1, 0, STR_PTR, 0, SLJIT_IMM, REQ_BYTE_MAX);
  toolong = CMP(SLJIT_C_LESS, TMP1, 0, STR_END, 0);
  alreadyfound = CMP(SLJIT_C_LESS, STR_PTR, 0, TMP2, 0);
-Line 2094 
 if (req_char == oc)
+Line 3234 
 if (req_char == oc)
  else
    {
    bit = req_char ^ oc;
-   if (ispowerof2(bit))
+   if (is_powerof2(bit))
      {
      OP2(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_IMM, bit);
      found = CMP(SLJIT_C_EQUAL, TMP2, 0, SLJIT_IMM, req_char | bit);
-Line 2111 
 JUMPTO(SLJIT_JUMP, loop);
+Line 3251 
 JUMPTO(SLJIT_JUMP, loop);
  JUMPHERE(found);
  if (foundoc)
    JUMPHERE(foundoc);
- OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), REQ_CHAR_PTR, TMP1, 0);
+ OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->req_char_ptr, TMP1, 0);
  JUMPHERE(alreadyfound);
  JUMPHERE(toolong);
  return notfound;
-Line 2123 
 DEFINE_COMPILER;
+Line 3263 
 DEFINE_COMPILER;
  struct sljit_jump *jump;
  struct sljit_label *mainloop;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP1(SLJIT_MOV, TMP1, 0, STACK_TOP, 0);
+ GET_LOCAL_BASE(TMP3, 0, 0);
  /* Drop frames until we reach STACK_TOP. */
  mainloop = LABEL();
  OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(TMP1), 0);
- jump = CMP(SLJIT_C_SIG_LESS_EQUAL, TMP2, 0, SLJIT_IMM, frame_end);
+ OP2(SLJIT_SUB | SLJIT_SET_S, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_IMM, 0);
- OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, SLJIT_LOCALS_REG, 0);
+ jump = JUMP(SLJIT_C_SIG_LESS_EQUAL);
- OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), 0, SLJIT_MEM1(TMP1), sizeof(sljit_w));
- OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), sizeof(sljit_w), SLJIT_MEM1(TMP1), 2 * sizeof(sljit_w));
+ OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, TMP3, 0);
- OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 3 * sizeof(sljit_w));
+ OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), 0, SLJIT_MEM1(TMP1), sizeof(sljit_sw));
+ OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), sizeof(sljit_sw), SLJIT_MEM1(TMP1), 2 * sizeof(sljit_sw));
+ OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 3 * sizeof(sljit_sw));
  JUMPTO(SLJIT_JUMP, mainloop);
  JUMPHERE(jump);
- jump = CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, frame_end);
+ jump = JUMP(SLJIT_C_SIG_LESS);
  /* End of dropping frames. */
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  JUMPHERE(jump);
- jump = CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, frame_setstrbegin);
+ OP1(SLJIT_NEG, TMP2, 0, TMP2, 0);
- /* Set string begin. */
+ OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, TMP3, 0);
- OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(TMP1), sizeof(sljit_w));
+ OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), 0, SLJIT_MEM1(TMP1), sizeof(sljit_sw));
- OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 2 * sizeof(sljit_w));
+ OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 2 * sizeof(sljit_sw));
- OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(0), TMP2, 0);
- JUMPTO(SLJIT_JUMP, mainloop);
- JUMPHERE(jump);
- /* Unknown command. */
- OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 2 * sizeof(sljit_w));
  JUMPTO(SLJIT_JUMP, mainloop);
  }
  static void check_wordboundary(compiler_common *common)
  {
  DEFINE_COMPILER;
- struct sljit_jump *beginend;
+ struct sljit_jump *skipread;
+ jump_list *skipread_list = NULL;
  #if !(defined COMPILE_PCRE8) || defined SUPPORT_UTF
  struct sljit_jump *jump;
  #endif
  SLJIT_COMPILE_ASSERT(ctype_word == 0x10, ctype_word_must_be_16);
- sljit_emit_fast_enter(compiler, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0);
  /* Get type of the previous char, and put it to LOCALS1. */
  OP1(SLJIT_MOV, TMP1, 0, ARGUMENTS, 0);
  OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(TMP1), SLJIT_OFFSETOF(jit_arguments, begin));
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1, SLJIT_IMM, 0);
- beginend = CMP(SLJIT_C_LESS_EQUAL, STR_PTR, 0, TMP1, 0);
+ skipread = CMP(SLJIT_C_LESS_EQUAL, STR_PTR, 0, TMP1, 0);
  skip_char_back(common);
+ check_start_used_ptr(common);
  read_char(common);
  /* Testing char type. */
-Line 2183 
 if (common->use_ucp)
+Line 3322 
 if (common->use_ucp)
    add_jump(compiler, &common->getucd, JUMP(SLJIT_FAST_CALL));
    OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Ll);
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ucp_Lu - ucp_Ll);
-   COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_LESS_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS_EQUAL);
    OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Nd - ucp_Ll);
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ucp_No - ucp_Nd);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_LESS_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
    JUMPHERE(jump);
    OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1, TMP2, 0);
    }
-Line 2212 
 else
+Line 3351 
 else
      JUMPHERE(jump);
  #endif /* COMPILE_PCRE8 */
    }
- JUMPHERE(beginend);
+ JUMPHERE(skipread);
  OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, 0);
- beginend = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+ check_str_end(common, &skipread_list);
  peek_char(common);
  /* Testing char type. This is a code duplication. */
-Line 2227 
 if (common->use_ucp)
+Line 3366 
 if (common->use_ucp)
    add_jump(compiler, &common->getucd, JUMP(SLJIT_FAST_CALL));
    OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Ll);
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ucp_Lu - ucp_Ll);
-   COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_LESS_EQUAL);
+   OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS_EQUAL);
    OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Nd - ucp_Ll);
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ucp_No - ucp_Nd);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_LESS_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
    JUMPHERE(jump);
    }
  else
-Line 2256 
 else
+Line 3395 
 else
      JUMPHERE(jump);
  #endif /* COMPILE_PCRE8 */
    }
- JUMPHERE(beginend);
+ set_jumps(skipread_list, LABEL());
  OP2(SLJIT_XOR | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1);
  sljit_emit_fast_return(compiler, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0);
  }
+ /*
+   range format:
+   ranges[0] = length of the range (max MAX_RANGE_SIZE, -1 means invalid range).
+   ranges[1] = first bit (0 or 1)
+   ranges[2-length] = position of the bit change (when the current bit is not equal to the previous)
+ */
+ static BOOL check_ranges(compiler_common *common, int *ranges, jump_list **backtracks, BOOL readch)
+ {
+ DEFINE_COMPILER;
+ struct sljit_jump *jump;
+ if (ranges[0] < 0)
+   return FALSE;
+ switch(ranges[0])
+   {
+   case 1:
+   if (readch)
+     read_char(common);
+   add_jump(compiler, backtracks, CMP(ranges[1] == 0 ? SLJIT_C_LESS : SLJIT_C_GREATER_EQUAL, TMP1, 0, SLJIT_IMM, ranges[2]));
+   return TRUE;
+   case 2:
+   if (readch)
+     read_char(common);
+   OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ranges[2]);
+   add_jump(compiler, backtracks, CMP(ranges[1] != 0 ? SLJIT_C_LESS : SLJIT_C_GREATER_EQUAL, TMP1, 0, SLJIT_IMM, ranges[3] - ranges[2]));
+   return TRUE;
+   case 4:
+   if (ranges[2] + 1 == ranges[3] && ranges[4] + 1 == ranges[5])
+     {
+     if (readch)
+       read_char(common);
+     if (ranges[1] != 0)
+       {
+       add_jump(compiler, backtracks, CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, ranges[2]));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, ranges[4]));
+       }
+     else
+       {
+       jump = CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, ranges[2]);
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, ranges[4]));
+       JUMPHERE(jump);
+       }
+     return TRUE;
+     }
+   if ((ranges[3] - ranges[2]) == (ranges[5] - ranges[4]) && is_powerof2(ranges[4] - ranges[2]))
+     {
+     if (readch)
+       read_char(common);
+     OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, ranges[4] - ranges[2]);
+     OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ranges[4]);
+     add_jump(compiler, backtracks, CMP(ranges[1] != 0 ? SLJIT_C_LESS : SLJIT_C_GREATER_EQUAL, TMP1, 0, SLJIT_IMM, ranges[5] - ranges[4]));
+     return TRUE;
+     }
+   return FALSE;
+   default:
+   return FALSE;
+   }
+ }
+ static void get_ctype_ranges(compiler_common *common, int flag, int *ranges)
+ {
+ int i, bit, length;
+ const pcre_uint8 *ctypes = (const pcre_uint8*)common->ctypes;
+ bit = ctypes[0] & flag;
+ ranges[0] = -1;
+ ranges[1] = bit != 0 ? 1 : 0;
+ length = 0;
+ for (i = 1; i < 256; i++)
+   if ((ctypes[i] & flag) != bit)
+     {
+     if (length >= MAX_RANGE_SIZE)
+       return;
+     ranges[2 + length] = i;
+     length++;
+     bit ^= flag;
+     }
+ if (bit != 0)
+   {
+   if (length >= MAX_RANGE_SIZE)
+     return;
+   ranges[2 + length] = 256;
+   length++;
+   }
+ ranges[0] = length;
+ }
+ static BOOL check_class_ranges(compiler_common *common, const pcre_uint8 *bits, BOOL nclass, jump_list **backtracks)
+ {
+ int ranges[2 + MAX_RANGE_SIZE];
+ pcre_uint8 bit, cbit, all;
+ int i, byte, length = 0;
+ bit = bits[0] & 0x1;
+ ranges[1] = bit;
+ /* Can be 0 or 255. */
+ all = -bit;
+ for (i = 0; i < 256; )
+   {
+   byte = i >> 3;
+   if ((i & 0x7) == 0 && bits[byte] == all)
+     i += 8;
+   else
+     {
+     cbit = (bits[byte] >> (i & 0x7)) & 0x1;
+     if (cbit != bit)
+       {
+       if (length >= MAX_RANGE_SIZE)
+         return FALSE;
+       ranges[2 + length] = i;
+       length++;
+       bit = cbit;
+       all = -cbit;
+       }
+     i++;
+     }
+   }
+ if (((bit == 0) && nclass) || ((bit == 1) && !nclass))
+   {
+   if (length >= MAX_RANGE_SIZE)
+     return FALSE;
+   ranges[2 + length] = 256;
+   length++;
+   }
+ ranges[0] = length;
+ return check_ranges(common, ranges, backtracks, FALSE);
+ }
  static void check_anynewline(compiler_common *common)
  {
  /* Check whether TMP1 contains a newline character. TMP2 destroyed. */
  DEFINE_COMPILER;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x0a);
  OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x0d - 0x0a);
- COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_LESS_EQUAL);
+ OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS_EQUAL);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x85 - 0x0a);
- #if defined SUPPORT_UTF || defined COMPILE_PCRE16
+ #if defined SUPPORT_UTF || defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  #ifdef COMPILE_PCRE8
  if (common->utf)
    {
  #endif
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x1);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x2029 - 0x0a);
  #ifdef COMPILE_PCRE8
    }
  #endif
- #endif /* SUPPORT_UTF || COMPILE_PCRE16 */
+ #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
- COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+ OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
-Line 2294 
 static void check_hspace(compiler_common
+Line 3572 
 static void check_hspace(compiler_common
  /* Check whether TMP1 contains a newline character. TMP2 destroyed. */
  DEFINE_COMPILER;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x09);
- COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_EQUAL);
+ OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x20);
- COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+ OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xa0);
- #if defined SUPPORT_UTF || defined COMPILE_PCRE16
+ #if defined SUPPORT_UTF || defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  #ifdef COMPILE_PCRE8
  if (common->utf)
    {
  #endif
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x1680);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x180e);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x2000);
    OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x200A - 0x2000);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_LESS_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x202f - 0x2000);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x205f - 0x2000);
-   COND_VALUE(SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x3000 - 0x2000);
  #ifdef COMPILE_PCRE8
    }
  #endif
- #endif /* SUPPORT_UTF || COMPILE_PCRE16 */
+ #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
- COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+ OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
-Line 2333 
 static void check_vspace(compiler_common
+Line 3611 
 static void check_vspace(compiler_common
  /* Check whether TMP1 contains a newline character. TMP2 destroyed. */
  DEFINE_COMPILER;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x0a);
  OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x0d - 0x0a);
- COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_LESS_EQUAL);
+ OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS_EQUAL);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x85 - 0x0a);
- #if defined SUPPORT_UTF || defined COMPILE_PCRE16
+ #if defined SUPPORT_UTF || defined COMPILE_PCRE16 || defined COMPILE_PCRE32
  #ifdef COMPILE_PCRE8
  if (common->utf)
    {
  #endif
-   COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+   OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
    OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, 0x1);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0x2029 - 0x0a);
  #ifdef COMPILE_PCRE8
    }
  #endif
- #endif /* SUPPORT_UTF || COMPILE_PCRE16 */
+ #endif /* SUPPORT_UTF || COMPILE_PCRE16 || COMPILE_PCRE32 */
- COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_EQUAL);
+ OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
  sljit_emit_fast_return(compiler, RETURN_ADDR, 0);
  }
-Line 2365 
 DEFINE_COMPILER;
+Line 3643 
 DEFINE_COMPILER;
  struct sljit_jump *jump;
  struct sljit_label *label;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
  OP1(SLJIT_MOV, TMP3, 0, CHAR1, 0);
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0, CHAR2, 0);
-Line 2394 
 DEFINE_COMPILER;
+Line 3672 
 DEFINE_COMPILER;
  struct sljit_jump *jump;
  struct sljit_label *label;
- sljit_emit_fast_enter(compiler, RETURN_ADDR, 0, 1, 5, 5, common->localsize);
+ sljit_emit_fast_enter(compiler, RETURN_ADDR, 0);
  OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, TMP2, 0);
  OP1(SLJIT_MOV, TMP3, 0, LCC_TABLE, 0);
-Line 2437 
 sljit_emit_fast_return(compiler, RETURN_
+Line 3715 
 sljit_emit_fast_return(compiler, RETURN_
  #if defined SUPPORT_UTF && defined SUPPORT_UCP
- static const pcre_uchar *SLJIT_CALL do_utf_caselesscmp(pcre_uchar *src1, jit_arguments *args, pcre_uchar *end1)
+ static const pcre_uchar * SLJIT_CALL do_utf_caselesscmp(pcre_uchar *src1, jit_arguments *args, pcre_uchar *end1)
  {
  /* This function would be ineffective to do in JIT level. */
- int c1, c2;
+ pcre_uint32 c1, c2;
- const pcre_uchar *src2 = args->ptr;
+ const pcre_uchar *src2 = args->uchar_ptr;
  const pcre_uchar *end2 = args->end;
+ const ucd_record *ur;
+ const pcre_uint32 *pp;
  while (src1 < end1)
    {
    if (src2 >= end2)
-     return 0;
+     return (pcre_uchar*)1;
    GETCHARINC(c1, src1);
    GETCHARINC(c2, src2);
-   if (c1 != c2 && c1 != UCD_OTHERCASE(c2)) return 0;
+   ur = GET_UCD(c2);
+   if (c1 != c2 && c1 != c2 + ur->other_case)
+     {
+     pp = PRIV(ucd_caseless_sets) + ur->caseset;
+     for (;;)
+       {
+       if (c1 < *pp) return NULL;
+       if (c1 == *pp++) break;
+       }
+     }
    }
  return src2;
  }
-Line 2458 
 return src2;
+Line 3747 
 return src2;
  #endif /* SUPPORT_UTF && SUPPORT_UCP */
  static pcre_uchar *byte_sequence_compare(compiler_common *common, BOOL caseless, pcre_uchar *cc,
-     compare_context* context, jump_list **fallbacks)
+     compare_context* context, jump_list **backtracks)
  {
  DEFINE_COMPILER;
  unsigned int othercasebit = 0;
-Line 2472 
 if (caseless && char_has_othercase(commo
+Line 3761 
 if (caseless && char_has_othercase(commo
    othercasebit = char_get_othercase_bit(common, cc);
    SLJIT_ASSERT(othercasebit);
    /* Extracting bit difference info. */
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    othercasechar = cc + (othercasebit >> 8);
    othercasebit &= 0xff;
- #else
+ #elif defined COMPILE_PCRE16 || defined COMPILE_PCRE32
- #ifdef COMPILE_PCRE16
+   /* Note that this code only handles characters in the BMP. If there
+   ever are characters outside the BMP whose othercase differs in only one
+   bit from itself (there currently are none), this code will need to be
+   revised for COMPILE_PCRE32. */
    othercasechar = cc + (othercasebit >> 9);
    if ((othercasebit & 0x100) != 0)
      othercasebit = (othercasebit & 0xff) << 8;
    else
      othercasebit &= 0xff;
- #endif
+ #endif /* COMPILE_PCRE[8|16|32] */
- #endif
    }
  if (context->sourcereg == -1)
    {
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
  #if defined SLJIT_UNALIGNED && SLJIT_UNALIGNED
    if (context->length >= 4)
      OP1(SLJIT_MOV_SI, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
-Line 2497 
 if (context->sourcereg == -1)
+Line 3788 
 if (context->sourcereg == -1)
    else
  #endif
      OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #else
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
  #if defined SLJIT_UNALIGNED && SLJIT_UNALIGNED
    if (context->length >= 4)
      OP1(SLJIT_MOV_SI, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
    else
  #endif
-     OP1(SLJIT_MOV_UH, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
+     OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #endif
+ #elif defined COMPILE_PCRE32
- #endif /* COMPILE_PCRE8 */
+   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), -context->length);
+ #endif /* COMPILE_PCRE[8|16|32] */
    context->sourcereg = TMP2;
    }
-Line 2520 
 do
+Line 3811 
 do
  #endif
    context->length -= IN_UCHARS(1);
- #if defined SLJIT_UNALIGNED && SLJIT_UNALIGNED
+ #if (defined SLJIT_UNALIGNED && SLJIT_UNALIGNED) && (defined COMPILE_PCRE8 || defined COMPILE_PCRE16)
    /* Unaligned read is supported. */
    if (othercasebit != 0 && othercasechar == cc)
-Line 2535 
 do
+Line 3826 
 do
      }
    context->ucharptr++;
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    if (context->ucharptr >= 4 || context->length == 0 || (context->ucharptr == 2 && context->length == 1))
  #else
    if (context->ucharptr >= 2 || context->length == 0)
-Line 2543 
 do
+Line 3834 
 do
      {
      if (context->length >= 4)
        OP1(SLJIT_MOV_SI, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #ifdef COMPILE_PCRE8
      else if (context->length >= 2)
        OP1(SLJIT_MOV_UH, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
+ #if defined COMPILE_PCRE8
      else if (context->length >= 1)
        OP1(SLJIT_MOV_UB, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #else
+ #endif /* COMPILE_PCRE8 */
-     else if (context->length >= 2)
-       OP1(SLJIT_MOV_UH, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #endif
      context->sourcereg = context->sourcereg == TMP1 ? TMP2 : TMP1;
      switch(context->ucharptr)
-Line 2559 
 do
+Line 3847 
 do
        case 4 / sizeof(pcre_uchar):
        if (context->oc.asint != 0)
          OP2(SLJIT_OR, context->sourcereg, 0, context->sourcereg, 0, SLJIT_IMM, context->oc.asint);
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asint | context->oc.asint));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asint | context->oc.asint));
        break;
        case 2 / sizeof(pcre_uchar):
        if (context->oc.asushort != 0)
          OP2(SLJIT_OR, context->sourcereg, 0, context->sourcereg, 0, SLJIT_IMM, context->oc.asushort);
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asushort | context->oc.asushort));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asushort | context->oc.asushort));
        break;
  #ifdef COMPILE_PCRE8