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

Diff of /code/trunk/pcre_jit_compile.c

Parent Directory | Revision Log | View Patch Patch

-revision 915 by zherczeg,
Tue Feb 14 13:05:39 2012 UTC
+revision 1209 by zherczeg,
Wed Nov  7 07:25:07 2012 UTC
 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. */
+ /* Allocate memory for the regex stack on the real machine stack.
- #define LOCAL_SPACE_SIZE 32768
+ 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 85 
 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 111 
 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 155 
 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;
    /* Everything else after. */
    int offsetcount;
    int calllimit;
-Line 166 
 typedef struct executable_functions {
+Line 170 
 typedef struct executable_functions {
    void *executable_funcs[JIT_NUMBER_OF_COMPILE_MODES];
    PUBL(jit_callback) callback;
    void *userdata;
+   pcre_uint32 top_bracket;
    sljit_uw executable_sizes[JIT_NUMBER_OF_COMPILE_MODES];
  } executable_functions;
-Line 180 
 typedef struct stub_list {
+Line 185 
 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;
  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 268 
 typedef struct recurse_entry {
    int start;
  } recurse_entry;
- typedef struct recurse_fallback {
+ typedef struct recurse_backtrack {
-   fallback_common common;
+   backtrack_common common;
- } recurse_fallback;
+ } recurse_backtrack;
+ #define MAX_RANGE_SIZE 6
  typedef struct compiler_common {
    struct sljit_compiler *compiler;
    pcre_uchar *start;
-   int localsize;
-   int *localptrs;
+   /* Maps private data offset to each opcode. */
-   const pcre_uint8 *fcc;
+   int *private_data_ptrs;
-   sljit_w lcc;
+   /* Tells whether the capturing bracket is optimized. */
+   pcre_uint8 *optimized_cbracket;
+   /* Starting offset of private data for capturing brackets. */
    int cbraptr;
+   /* OVector 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;
+   /* 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;
+   /* Flipped and lower case tables. */
+   const pcre_uint8 *fcc;
+   sljit_sw lcc;
+   /* Mode can be PCRE_STUDY_JIT_COMPILE and others. */
    int mode;
+   /* Newline control. */
    int nltype;
    int newline;
    int bsr_nltype;
+   /* Dollar endonly. */
    int endonly;
-   sljit_w ctypes;
+   BOOL has_set_som;
+   /* 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;
+   /* Labels and jump lists. */
    struct sljit_label *partialmatchlabel;
+   struct sljit_label *quitlabel;
    struct sljit_label *acceptlabel;
    stub_list *stubs;
    recurse_entry *entries;
    recurse_entry *currententry;
    jump_list *partialmatch;
+   jump_list *quit;
    jump_list *accept;
    jump_list *calllimit;
    jump_list *stackalloc;
-Line 306 
 typedef struct compiler_common {
+Line 344 
 typedef struct compiler_common {
  #ifdef SUPPORT_UCP
    BOOL use_ucp;
  #endif
+ #ifndef COMPILE_PCRE32
    jump_list *utfreadchar;
+ #endif
  #ifdef COMPILE_PCRE8
    jump_list *utfreadtype8;
  #endif
-Line 324 
 typedef struct compare_context {
+Line 364 
 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
-Line 352 
 typedef struct compare_context {
+Line 392 
 typedef struct compare_context {
  enum {
    frame_end = 0,
-   frame_setstrbegin = -1
+   frame_setstrbegin = -1,
+   frame_setmark = -2
  };
  /* Undefine sljit macros. */
  #undef CMP
  /* 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 TMP2          SLJIT_TEMPORARY_REG3
-Line 372 
 enum {
+Line 413 
 enum {
  #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.
- Same as START_USED_PTR. (Partial matching and req_char are exclusive) */
- #define REQ_CHAR_PTR     (6 * sizeof(sljit_w))
- /* First inspected character for partial matching.
- Same as REQ_CHAR_PTR. (Partial matching and req_char are exclusive) */
- #define START_USED_PTR   (6 * sizeof(sljit_w))
- /* Starting pointer for partial soft matches. */
- #define HIT_START        (8 * sizeof(sljit_w))
- /* End pointer of the first line. */
- #define FIRSTLINE_END    (9 * 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    (10 * 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->cbraptr + (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 433 
 the start pointers when the end of the c
+Line 463 
 the start pointers when the end of the c
    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 447 
 return cc;
+Line 479 
 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 506 
 switch(*cc)
+Line 538 
 switch(*cc)
    case OP_BRAZERO:
    case OP_BRAMINZERO:
    case OP_BRAPOSZERO:
+   case OP_COMMIT:
    case OP_FAIL:
    case OP_ACCEPT:
    case OP_ASSERT_ACCEPT:
-Line 644 
 switch(*cc)
+Line 677 
 switch(*cc)
    case OP_SCBRAPOS:
    return cc + 1 + LINK_SIZE + IMM2_SIZE;
+   case OP_MARK:
+   return cc + 1 + 2 + cc[1];
    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 667 
 while (cc < ccend)
+Line 799 
 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);
+     bracketlen = cc[1 + LINK_SIZE];
-     if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
+     if (bracketlen == OP_CREF)
-       localspace += sizeof(sljit_w);
+       {
+       bracketlen = GET2(cc, 1 + LINK_SIZE + 1);
+       common->optimized_cbracket[bracketlen] = 0;
+       }
+     else if (bracketlen == OP_NCREF)
+       {
+       bracketlen = GET2(cc, 1 + LINK_SIZE + 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;
+         }
+       }
+     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_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 == 0)
+       {
+       common->recursive_head = common->ovector_start;
+       common->ovector_start += sizeof(sljit_sw);
+       }
      cc += 1 + LINK_SIZE;
      break;
+     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;
      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 714 
 while (cc < ccend)
+Line 984 
 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->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 + IMM2_SIZE;
+     bracketlen = 1 + LINK_SIZE + IMM2_SIZE;
      break;
      case OP_COND:
-Line 731 
 while (cc < ccend)
+Line 1001 
 while (cc < ccend)
      alternative = cc + GET(cc, 1);
      if (*alternative == OP_KETRMAX || *alternative == OP_KETRMIN)
        {
-       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_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
      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;
+     }
    }
  }
-Line 751 
 static int get_framesize(compiler_common
+Line 1105 
 static int get_framesize(compiler_common
  pcre_uchar *ccend = bracketend(cc);
  int length = 0;
  BOOL possessive = FALSE;
- BOOL setsom_found = FALSE;
+ BOOL setsom_found = recursive;
+ BOOL setmark_found = recursive;
  if (!recursive && (*cc == OP_CBRAPOS || *cc == OP_SCBRAPOS))
    {
-Line 765 
 while (cc < ccend)
+Line 1120 
 while (cc < ccend)
    switch(*cc)
      {
      case OP_SET_SOM:
-     case OP_RECURSE:
+     SLJIT_ASSERT(common->has_set_som);
      if (!setsom_found)
        {
        length += 2;
        setsom_found = TRUE;
        }
-     cc += (*cc == OP_SET_SOM) ? 1 : 1 + LINK_SIZE;
+     cc += 1;
+     break;
+     case OP_MARK:
+     SLJIT_ASSERT(common->mark_ptr != 0);
+     if (!setmark_found)
+       {
+       length += 2;
+       setmark_found = TRUE;
+       }
+     cc += 1 + 2 + cc[1];
+     break;
+     case OP_RECURSE:
+     if (common->has_set_som && !setsom_found)
+       {
+       length += 2;
+       setsom_found = TRUE;
+       }
+     if (common->mark_ptr != 0 && !setmark_found)
+       {
+       length += 2;
+       setmark_found = TRUE;
+       }
+     cc += 1 + LINK_SIZE;
      break;
      case OP_CBRA:
-Line 801 
 static void init_frame(compiler_common *
+Line 1180 
 static void init_frame(compiler_common *
  {
  DEFINE_COMPILER;
  pcre_uchar *ccend = bracketend(cc);
- BOOL setsom_found = FALSE;
+ BOOL setsom_found = recursive;
+ BOOL setmark_found = recursive;
  int offset;
  /* >= 1 + shortest item size (2) */
-Line 816 
 while (cc < ccend)
+Line 1196 
 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);
-       stackpos += (int)sizeof(sljit_w);
+       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;
+     cc += 1;
+     break;
+     case OP_MARK:
+     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, frame_setmark);
+       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, frame_setstrbegin);
+       stackpos += (int)sizeof(sljit_sw);
+       OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), stackpos, TMP1, 0);
+       stackpos += (int)sizeof(sljit_sw);
+       setsom_found = TRUE;
+       }
+     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, frame_setmark);
+       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 + LINK_SIZE;
      break;
      case OP_CBRA:
-Line 835 
 while (cc < ccend)
+Line 1251 
 while (cc < ccend)
      case OP_SCBRAPOS:
      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 856 
 OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), st
+Line 1272 
 OP1(SLJIT_MOV, SLJIT_MEM1(STACK_TOP), st
  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 = 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 875 
 while (cc < ccend)
+Line 1293 
 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 895 
 while (cc < ccend)
+Line 1314 
 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 906 
 while (cc < ccend)
+Line 1383 
 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 931 
 stacktop = STACK(stacktop - 1);
+Line 1408 
 stacktop = STACK(stacktop - 1);
  if (!save)
    {
-   stackptr += sizeof(sljit_w);
+   stackptr += 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. */
-Line 953 
 while (status != end)
+Line 1430 
 while (status != end)
    switch(status)
      {
      case start:
-     SLJIT_ASSERT(save);
+     SLJIT_ASSERT(save && common->recursive_head != 0);
      count = 1;
-     srcw[0] = RECURSIVE_HEAD;
+     srcw[0] = common->recursive_head;
      status = loop;
      break;
-Line 979 
 while (status != end)
+Line 1456 
 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 1006 
 while (status != end)
+Line 1486 
 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 1034 
 while (status != end)
+Line 1610 
 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 1045 
 while (status != end)
+Line 1621 
 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 1062 
 while (status != end)
+Line 1638 
 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 1074 
 while (status != end)
+Line 1650 
 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;
          }
-Line 1089 
 if (save)
+Line 1665 
 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 1102 
 if (save)
+Line 1678 
 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 1124 
 static SLJIT_INLINE void set_jumps(jump_
+Line 1707 
 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);
    list = list->next;
    }
-Line 1151 
 if (list_item)
+Line 1734 
 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 1171 
 while (list_item)
+Line 1754 
 while (list_item)
      add_jump(compiler, &common->stackalloc, JUMP(SLJIT_FAST_CALL));
      break;
      }
-   JUMPTO(SLJIT_JUMP, list_item->leave);
+   JUMPTO(SLJIT_JUMP, list_item->quit);
    list_item = list_item->next;
    }
  common->stubs = NULL;
-Line 1190 
 static SLJIT_INLINE void allocate_stack(
+Line 1773 
 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 1204 
 add_stub(common, stack_alloc, 0, CMP(SLJ
+Line 1787 
 add_stub(common, stack_alloc, 0, CMP(SLJ
  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 1222 
 if (length < 8)
+Line 1805 
 if (length < 8)
    }
  else
    {
-   OP2(SLJIT_ADD, SLJIT_TEMPORARY_REG2, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, OVECTOR_START - sizeof(sljit_w));
+   GET_LOCAL_BASE(SLJIT_TEMPORARY_REG2, 0, OVECTOR_START - sizeof(sljit_sw));
    OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, length);
    loop = LABEL();
-   OP1(SLJIT_MOVU, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), sizeof(sljit_w), SLJIT_TEMPORARY_REG1, 0);
+   OP1(SLJIT_MOVU, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), sizeof(sljit_sw), SLJIT_TEMPORARY_REG1, 0);
    OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_TEMPORARY_REG3, 0, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, 1);
    JUMPTO(SLJIT_C_NOT_ZERO, loop);
    }
-Line 1242 
 OP1(SLJIT_MOV, SLJIT_SAVED_REG3, 0, SLJI
+Line 1825 
 OP1(SLJIT_MOV, SLJIT_SAVED_REG3, 0, SLJI
  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), OVECTOR(1), STR_PTR, 0);
  OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG1, 0, ARGUMENTS, 0);
+ if (common->mark_ptr != 0)
+   OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr);
  OP1(SLJIT_MOV_SI, SLJIT_TEMPORARY_REG2, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, offsetcount));
+ if (common->mark_ptr != 0)
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, mark_ptr), SLJIT_TEMPORARY_REG3, 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_TEMPORARY_REG1, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), SLJIT_OFFSETOF(jit_arguments, begin));
- OP2(SLJIT_ADD, SLJIT_SAVED_REG1, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, OVECTOR_START);
+ GET_LOCAL_BASE(SLJIT_SAVED_REG1, 0, OVECTOR_START);
  /* Unlikely, but possible */
  earlyexit = CMP(SLJIT_C_EQUAL, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 0);
  loop = LABEL();
  OP2(SLJIT_SUB, SLJIT_SAVED_REG2, 0, SLJIT_MEM1(SLJIT_SAVED_REG1), 0, SLJIT_TEMPORARY_REG1, 0);
- OP2(SLJIT_ADD, SLJIT_SAVED_REG1, 0, SLJIT_SAVED_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_SAVED_REG2, 0, SLJIT_SAVED_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_SAVED_REG2, 0);
  OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_TEMPORARY_REG2, 0, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 1);
-Line 1263 
 JUMPHERE(earlyexit);
+Line 1850 
 JUMPHERE(earlyexit);
  /* 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_TEMPORARY_REG1, 0, OVECTOR_START + topbracket * 2 * sizeof(sljit_sw));
    OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, topbracket + 1);
    /* 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_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG1), -(2 * (sljit_sw)sizeof(sljit_sw)));
    OP2(SLJIT_SUB, SLJIT_TEMPORARY_REG2, 0, SLJIT_TEMPORARY_REG2, 0, SLJIT_IMM, 1);
    CMPTO(SLJIT_C_EQUAL, SLJIT_TEMPORARY_REG3, 0, SLJIT_SAVED_REG3, 0, loop);
    OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_TEMPORARY_REG2, 0);
-Line 1277 
 else
+Line 1864 
 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 *leave)
+ static SLJIT_INLINE void return_with_partial_match(compiler_common *common, struct sljit_label *quit)
  {
  DEFINE_COMPILER;
  SLJIT_COMPILE_ASSERT(STR_END == SLJIT_SAVED_REG2, str_end_must_be_saved_reg2);
+ SLJIT_ASSERT(common->start_used_ptr != 0 && (common->mode == JIT_PARTIAL_SOFT_COMPILE ? common->hit_start != 0 : common->hit_start == 0));
  OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG2, 0, ARGUMENTS, 0);
  OP1(SLJIT_MOV, SLJIT_RETURN_REG, 0, SLJIT_IMM, PCRE_ERROR_PARTIAL);
  OP1(SLJIT_MOV_SI, SLJIT_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), SLJIT_OFFSETOF(jit_arguments, offsetcount));
- CMPTO(SLJIT_C_LESS, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, 2, leave);
+ CMPTO(SLJIT_C_LESS, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, 2, quit);
  /* Store match begin and end. */
  OP1(SLJIT_MOV, SLJIT_SAVED_REG1, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), SLJIT_OFFSETOF(jit_arguments, begin));
  OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG2, 0, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), SLJIT_OFFSETOF(jit_arguments, offsets));
- OP1(SLJIT_MOV, SLJIT_TEMPORARY_REG3, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mode == JIT_PARTIAL_HARD_COMPILE ? START_USED_PTR : HIT_START);
+ OP1(SLJIT_MOV, SLJIT_TEMPORARY_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);
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
- OP2(SLJIT_ASHR, SLJIT_SAVED_REG2, 0, SLJIT_SAVED_REG2, 0, SLJIT_IMM, 1);
+ OP2(SLJIT_ASHR, SLJIT_SAVED_REG2, 0, SLJIT_SAVED_REG2, 0, SLJIT_IMM, UCHAR_SHIFT);
  #endif
  OP1(SLJIT_MOV_SI, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), sizeof(int), SLJIT_SAVED_REG2, 0);
  OP2(SLJIT_SUB, SLJIT_TEMPORARY_REG3, 0, SLJIT_TEMPORARY_REG3, 0, SLJIT_SAVED_REG1, 0);
- #ifdef COMPILE_PCRE16
+ #if defined COMPILE_PCRE16 || defined COMPILE_PCRE32
- OP2(SLJIT_ASHR, SLJIT_TEMPORARY_REG3, 0, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, 1);
+ OP2(SLJIT_ASHR, SLJIT_TEMPORARY_REG3, 0, SLJIT_TEMPORARY_REG3, 0, SLJIT_IMM, UCHAR_SHIFT);
  #endif
  OP1(SLJIT_MOV_SI, SLJIT_MEM1(SLJIT_TEMPORARY_REG2), 0, SLJIT_TEMPORARY_REG3, 0);
- JUMPTO(SLJIT_JUMP, leave);
+ JUMPTO(SLJIT_JUMP, quit);
  }
  static SLJIT_INLINE void check_start_used_ptr(compiler_common *common)
-Line 1315 
 struct sljit_jump *jump;
+Line 1903 
 struct sljit_jump *jump;
  if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
    {
-   /* The value of -1 must be kept for START_USED_PTR! */
+   /* The value of -1 must be kept for start_used_ptr! */
-   OP2(SLJIT_ADD, TMP1, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), START_USED_PTR, SLJIT_IMM, 1);
+   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
+   /* 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. */
+   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), START_USED_PTR, 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), START_USED_PTR, STR_PTR, 0);
+   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), START_USED_PTR, STR_PTR, 0);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->start_used_ptr, STR_PTR, 0);
    JUMPHERE(jump);
    }
  }
-Line 1415 
 if (c <= 127 && bit == 0x20)
+Line 2003 
 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 1434 
 if (common->utf && c > 127)
+Line 2022 
 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 1447 
 if (common->utf && c > 65535)
+Line 2034 
 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)
+ 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;
+ struct sljit_jump *jump = NULL;
+ SLJIT_ASSERT(!force || common->mode != JIT_COMPILE);
  if (common->mode == JIT_COMPILE)
    return;
- jump = CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), START_USED_PTR, STR_PTR, 0);
+ 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), HIT_START, SLJIT_IMM, -1);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, -1);
  else
    {
    if (common->partialmatchlabel != NULL)
-Line 1470 
 else
+Line 2063 
 else
    else
      add_jump(compiler, &common->partialmatch, JUMP(SLJIT_JUMP));
    }
- JUMPHERE(jump);
+ if (jump != NULL)
+   JUMPHERE(jump);
  }
  static struct sljit_jump *check_str_end(compiler_common *common)
-Line 1487 
 if (common->mode == JIT_COMPILE)
+Line 2082 
 if (common->mode == JIT_COMPILE)
  jump = CMP(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0);
  if (common->mode == JIT_PARTIAL_SOFT_COMPILE)
    {
-   nohit = CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), START_USED_PTR, STR_PTR, 0);
+   nohit = 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), HIT_START, SLJIT_IMM, -1);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, -1);
    JUMPHERE(nohit);
    return_value = JUMP(SLJIT_JUMP);
    }
  else
    {
-   return_value = CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), START_USED_PTR, STR_PTR, 0);
+   return_value = 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
-Line 1504 
 JUMPHERE(jump);
+Line 2099 
 JUMPHERE(jump);
  return return_value;
  }
- static void fallback_at_str_end(compiler_common *common, jump_list **fallbacks)
+ static void detect_partial_match(compiler_common *common, jump_list **backtracks)
  {
  DEFINE_COMPILER;
  struct sljit_jump *jump;
  if (common->mode == JIT_COMPILE)
    {
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
+   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, fallbacks, CMP(SLJIT_C_GREATER_EQUAL, SLJIT_MEM1(SLJIT_LOCALS_REG), START_USED_PTR, STR_PTR, 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), HIT_START, SLJIT_IMM, -1);
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->hit_start, SLJIT_IMM, -1);
-   add_jump(compiler, fallbacks, JUMP(SLJIT_JUMP));
+   add_jump(compiler, backtracks, JUMP(SLJIT_JUMP));
    }
  else
    {
-Line 1538 
 static void read_char(compiler_common *c
+Line 2133 
 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 1565 
 static void peek_char(compiler_common *c
+Line 2158 
 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 1600 
 if (common->utf)
+Line 2191 
 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 1616 
 if (common->utf)
+Line 2206 
 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 1641 
 static void skip_char_back(compiler_comm
+Line 2235 
 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 1653 
 if (common->utf)
+Line 2248 
 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 1662 
 if (common->utf)
+Line 2256 
 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 1679 
 DEFINE_COMPILER;
+Line 2274 
 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 1706 
 of the character (>= 0xc0). Return char
+Line 2301 
 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 1765 
 DEFINE_COMPILER;
+Line 2360 
 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 1786 
 sljit_emit_fast_return(compiler, RETURN_
+Line 2381 
 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 1802 
 of the character (>= 0xd800). Return cha
+Line 2396 
 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 1819 
 OP1(SLJIT_MOV, TMP2, 0, SLJIT_IMM, IN_UC
+Line 2413 
 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 1839 
 DEFINE_COMPILER;
+Line 2432 
 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 1861 
 struct sljit_label *newlinelabel = NULL;
+Line 2454 
 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 1875 
 if (!(hascrorlf || firstline) && (common
+Line 2468 
 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 1887 
 if (firstline)
+Line 2480 
 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), FIRSTLINE_END, STR_PTR, 0);
+     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);
-     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), FIRSTLINE_END, STR_PTR, 0);
+     JUMPHERE(end);
+     OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, STR_PTR, 0);
      set_jumps(newline, LABEL());
      }
-   JUMPHERE(end);
+   OP1(SLJIT_MOV, STR_PTR, 0, TMP3, 0);
-   OP1(SLJIT_MOV, STR_PTR, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0);
    }
  start = JUMP(SLJIT_JUMP);
-Line 1915 
 if (newlinecheck)
+Line 2509 
 if (newlinecheck)
    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);
-   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);
    nl = JUMP(SLJIT_JUMP);
-Line 1929 
 mainloop = LABEL();
+Line 2523 
 mainloop = LABEL();
  #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 + IMM2_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;
- if (readuchar)
+   while (len > 0 && location < MAX_N_CHARS * 2)
-   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
+     {
+     c = *cc;
+     bit = 0;
+     if (len == (caseless & 0xff))
+       {
+       bit = caseless >> 8;
+       c |= bit;
+       }
- if (newlinecheck)
+     chars[location] = c;
-   CMPTO(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff, newlinelabel);
+     chars[location + 1] = bit;
- OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
+     len--;
- #if defined SUPPORT_UTF && defined COMPILE_PCRE8
+     location += 2;
- if (common->utf)
+     cc++;
-   {
+     }
-   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
-   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_w)PRIV(utf8_table4) - 0xc0);
+   if (location >= MAX_N_CHARS * 2 || must_stop != 0)
-   OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, TMP1, 0);
+     break;
-   JUMPHERE(singlechar);
    }
- #endif
- #if defined SUPPORT_UTF && defined COMPILE_PCRE16
+ /* At least two characters are required. */
- if (common->utf)
+ if (location < 2 * 2)
+     return FALSE;
+ if (firstline)
    {
-   singlechar = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xd800);
+   SLJIT_ASSERT(common->first_line_end != 0);
-   OP2(SLJIT_AND, TMP1, 0, TMP1, 0, SLJIT_IMM, 0xfc00);
+   OP1(SLJIT_MOV, TMP3, 0, STR_END, 0);
-   OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 0xd800);
+   OP2(SLJIT_SUB, STR_END, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), common->first_line_end, SLJIT_IMM, (location >> 1) - 1);
-   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
+ else
- JUMPHERE(start);
+   OP2(SLJIT_SUB, STR_END, 0, STR_END, 0, SLJIT_IMM, (location >> 1) - 1);
- if (newlinecheck)
+ start = LABEL();
-   {
+ quit = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
-   JUMPHERE(end);
-   JUMPHERE(nl);
+ 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 (chars[1] != 0)
+   OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, chars[1]);
+ CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, chars[0], start);
+ if (location > 2 * 2)
+   OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
+ if (chars[3] != 0)
+   OP2(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_IMM, chars[3]);
+ CMPTO(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, chars[2], start);
+ if (location > 2 * 2)
+   {
+   if (chars[5] != 0)
+     OP2(SLJIT_OR, TMP1, 0, TMP1, 0, SLJIT_IMM, chars[5]);
+   CMPTO(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, chars[4], start);
    }
+ OP2(SLJIT_SUB, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
- return mainloop;
+ JUMPHERE(quit);
+ if (firstline)
+   OP1(SLJIT_MOV, STR_END, 0, TMP3, 0);
+ else
+   OP2(SLJIT_ADD, STR_END, 0, STR_END, 0, SLJIT_IMM, (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 2002 
 if (first_char == oc)
+Line 2767 
 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 2010 
 else
+Line 2775 
 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 2051 
 DEFINE_COMPILER;
+Line 2797 
 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 2072 
 if (common->nltype == NLTYPE_FIXED && co
+Line 2819 
 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 2110 
 set_jumps(newline, loop);
+Line 2857 
 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 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;
  #ifndef COMPILE_PCRE8
  struct sljit_jump *jump;
-Line 2142 
 struct sljit_jump *jump;
+Line 2889 
 struct sljit_jump *jump;
  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)
-Line 2170 
 if (common->utf)
+Line 2918 
 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);
- JUMPHERE(leave);
+ 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 2206 
 struct sljit_jump *alreadyfound;
+Line 2955 
 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 2236 
 if (req_char == oc)
+Line 2986 
 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 2253 
 JUMPTO(SLJIT_JUMP, loop);
+Line 3003 
 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 2265 
 DEFINE_COMPILER;
+Line 3015 
 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_ADD, TMP2, 0, TMP2, 0, SLJIT_LOCALS_REG, 0);
+ OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, TMP3, 0);
- OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), 0, SLJIT_MEM1(TMP1), sizeof(sljit_w));
+ OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), 0, SLJIT_MEM1(TMP1), sizeof(sljit_sw));
- OP1(SLJIT_MOV, SLJIT_MEM1(TMP2), sizeof(sljit_w), SLJIT_MEM1(TMP1), 2 * sizeof(sljit_w));
+ 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_w));
+ OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 3 * sizeof(sljit_sw));
  JUMPTO(SLJIT_JUMP, mainloop);
  JUMPHERE(jump);
-Line 2286 
 sljit_emit_fast_return(compiler, RETURN_
+Line 3037 
 sljit_emit_fast_return(compiler, RETURN_
  JUMPHERE(jump);
  jump = CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, frame_setstrbegin);
  /* Set string begin. */
- OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(TMP1), sizeof(sljit_w));
+ OP1(SLJIT_MOV, 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);
+ if (common->mark_ptr != 0)
+   {
+   jump = CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, frame_setmark);
+   OP1(SLJIT_MOV, TMP2, 0, SLJIT_MEM1(TMP1), sizeof(sljit_sw));
+   OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 2 * sizeof(sljit_sw));
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), common->mark_ptr, TMP2, 0);
+   JUMPTO(SLJIT_JUMP, mainloop);
+   JUMPHERE(jump);
+   }
  /* Unknown command. */
- 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));
  JUMPTO(SLJIT_JUMP, mainloop);
  }
-Line 2307 
 struct sljit_jump *jump;
+Line 3069 
 struct sljit_jump *jump;
  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));
-Line 2326 
 if (common->use_ucp)
+Line 3088 
 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 2370 
 if (common->use_ucp)
+Line 3132 
 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 2405 
 OP2(SLJIT_XOR | SLJIT_SET_E, SLJIT_UNUSE
+Line 3167 
 OP2(SLJIT_XOR | SLJIT_SET_E, SLJIT_UNUSE
  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 2437 
 static void check_hspace(compiler_common
+Line 3338 
 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 2476 
 static void check_vspace(compiler_common
+Line 3377 
 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 2508 
 DEFINE_COMPILER;
+Line 3409 
 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 2537 
 DEFINE_COMPILER;
+Line 3438 
 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 2583 
 sljit_emit_fast_return(compiler, RETURN_
+Line 3484 
 sljit_emit_fast_return(compiler, RETURN_
  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)
    {
-Line 2593 
 while (src1 < end1)
+Line 3496 
 while (src1 < end1)
      return (pcre_uchar*)1;
    GETCHARINC(c1, src1);
    GETCHARINC(c2, src2);
-   if (c1 != c2 && c1 != UCD_OTHERCASE(c2)) return NULL;
+   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 2601 
 return src2;
+Line 3513 
 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 2615 
 if (caseless && char_has_othercase(commo
+Line 3527 
 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 2640 
 if (context->sourcereg == -1)
+Line 3554 
 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 2678 
 do
+Line 3592 
 do
      }
    context->ucharptr++;
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
    if (context->ucharptr >= 4 || context->length == 0 || (context->ucharptr == 2 && context->length == 1))
- #else
+ #elif defined COMPILE_PCRE16
    if (context->ucharptr >= 2 || context->length == 0)
+ #elif defined COMPILE_PCRE32
+   if (1 /* context->ucharptr >= 1 || context->length == 0 */)
  #endif
      {
+ #if defined COMPILE_PCRE8 || defined COMPILE_PCRE16
      if (context->length >= 4)
        OP1(SLJIT_MOV_SI, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #ifdef COMPILE_PCRE8
+ #if defined COMPILE_PCRE8
      else if (context->length >= 2)
        OP1(SLJIT_MOV_UH, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
      else if (context->length >= 1)
        OP1(SLJIT_MOV_UB, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #else
+ #elif defined COMPILE_PCRE16
      else if (context->length >= 2)
        OP1(SLJIT_MOV_UH, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #endif
+ #endif /* COMPILE_PCRE[8|16] */
+ #elif defined COMPILE_PCRE32
+     OP1(MOV_UCHAR, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
+ #endif /* COMPILE_PCRE[8|16|32] */
      context->sourcereg = context->sourcereg == TMP1 ? TMP2 : TMP1;
      switch(context->ucharptr)
-Line 2702 
 do
+Line 3622 
 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;
+ #if defined COMPILE_PCRE8 || defined COMPILE_PCRE16
        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
        case 1:
        if (context->oc.asbyte != 0)
          OP2(SLJIT_OR, context->sourcereg, 0, context->sourcereg, 0, SLJIT_IMM, context->oc.asbyte);
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asbyte | context->oc.asbyte));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, context->c.asbyte | context->oc.asbyte));
        break;
  #endif
+ #endif /* COMPILE_PCRE[8|16] */
        default:
        SLJIT_ASSERT_STOP();
        break;
-Line 2729 
 do
+Line 3652 
 do
  #else
    /* Unaligned read is unsupported. */
- #ifdef COMPILE_PCRE8
-   if (context->length > 0)
-     OP1(SLJIT_MOV_UB, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #else
    if (context->length > 0)
-     OP1(SLJIT_MOV_UH, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
+     OP1(MOV_UCHAR, context->sourcereg, 0, SLJIT_MEM1(STR_PTR), -context->length);
- #endif
    context->sourcereg = context->sourcereg == TMP1 ? TMP2 : TMP1;
    if (othercasebit != 0 && othercasechar == cc)
      {
      OP2(SLJIT_OR, context->sourcereg, 0, context->sourcereg, 0, SLJIT_IMM, othercasebit);
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, *cc | othercasebit));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, *cc | othercasebit));
      }
    else
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, *cc));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, context->sourcereg, 0, SLJIT_IMM, *cc));
  #endif
-Line 2780 
 return cc;
+Line 3699 
 return cc;
      } \
    charoffset = (value);
- static void compile_xclass_hotpath(compiler_common *common, pcre_uchar *cc, jump_list **fallbacks)
+ static void compile_xclass_matchingpath(compiler_common *common, pcre_uchar *cc, jump_list **backtracks)
  {
  DEFINE_COMPILER;
  jump_list *found = NULL;
- jump_list **list = (*cc & XCL_NOT) == 0 ? &found : fallbacks;
+ jump_list **list = (*cc & XCL_NOT) == 0 ? &found : backtracks;
- unsigned int c;
+ pcre_int32 c, charoffset;
- int compares;
+ const pcre_uint32 *other_cases;
  struct sljit_jump *jump = NULL;
  pcre_uchar *ccbegin;
+ int compares, invertcmp, numberofcmps;
  #ifdef SUPPORT_UCP
  BOOL needstype = FALSE, needsscript = FALSE, needschar = FALSE;
  BOOL charsaved = FALSE;
  int typereg = TMP1, scriptreg = TMP1;
- unsigned int typeoffset;
+ pcre_int32 typeoffset;
  #endif
- int invertcmp, numberofcmps;
- unsigned int charoffset;
- /* Although SUPPORT_UTF must be defined, we are not necessary in utf mode. */
+ /* Although SUPPORT_UTF must be defined, we are
- fallback_at_str_end(common, fallbacks);
+    not necessary in utf mode even in 8 bit mode. */
+ detect_partial_match(common, backtracks);
  read_char(common);
  if ((*cc++ & XCL_MAP) != 0)
-Line 2812 
 if ((*cc++ & XCL_MAP) != 0)
+Line 3731 
 if ((*cc++ & XCL_MAP) != 0)
      jump = CMP(SLJIT_C_GREATER, TMP1, 0, SLJIT_IMM, 255);
  #endif
-   OP2(SLJIT_AND, TMP2, 0, TMP1, 0, SLJIT_IMM, 0x7);
+   if (!check_class_ranges(common, (const pcre_uint8 *)cc, TRUE, list))
-   OP2(SLJIT_LSHR, TMP1, 0, TMP1, 0, SLJIT_IMM, 3);
+     {
-   OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_w)cc);
+     OP2(SLJIT_AND, TMP2, 0, TMP1, 0, SLJIT_IMM, 0x7);
-   OP2(SLJIT_SHL, TMP2, 0, SLJIT_IMM, 1, TMP2, 0);
+     OP2(SLJIT_LSHR, TMP1, 0, TMP1, 0, SLJIT_IMM, 3);
-   OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, TMP2, 0);
+     OP1(SLJIT_MOV_UB, TMP1, 0, SLJIT_MEM1(TMP1), (sljit_sw)cc);
-   add_jump(compiler, list, JUMP(SLJIT_C_NOT_ZERO));
+     OP2(SLJIT_SHL, TMP2, 0, SLJIT_IMM, 1, TMP2, 0);
+     OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, TMP2, 0);
+     add_jump(compiler, list, JUMP(SLJIT_C_NOT_ZERO));
+     }
  #ifndef COMPILE_PCRE8
    JUMPHERE(jump);
-Line 2890 
 while (*cc != XCL_END)
+Line 3812 
 while (*cc != XCL_END)
        needschar = TRUE;
        break;
+       case PT_CLIST:
+       needschar = TRUE;
+       break;
        default:
        SLJIT_ASSERT_STOP();
        break;
-Line 2926 
 if (needstype || needsscript)
+Line 3852 
 if (needstype || needsscript)
      {
      if (scriptreg == TMP1)
        {
-       OP1(SLJIT_MOV, scriptreg, 0, SLJIT_IMM, (sljit_w)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, script));
+       OP1(SLJIT_MOV, scriptreg, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, script));
        OP1(SLJIT_MOV_UB, scriptreg, 0, SLJIT_MEM2(scriptreg, TMP2), 3);
        }
      else
        {
        OP2(SLJIT_SHL, TMP2, 0, TMP2, 0, SLJIT_IMM, 3);
-       OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, SLJIT_IMM, (sljit_w)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, script));
+       OP2(SLJIT_ADD, TMP2, 0, TMP2, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, script));
        OP1(SLJIT_MOV_UB, scriptreg, 0, SLJIT_MEM1(TMP2), 0);
        }
      }
-Line 2950 
 typeoffset = 0;
+Line 3876 
 typeoffset = 0;
  while (*cc != XCL_END)
    {
    compares--;
-   invertcmp = (compares == 0 && list != fallbacks);
+   invertcmp = (compares == 0 && list != backtracks);
    jump = NULL;
    if (*cc == XCL_SINGLE)
-Line 2968 
 while (*cc != XCL_END)
+Line 3894 
 while (*cc != XCL_END)
      if (numberofcmps < 3 && (*cc == XCL_SINGLE || *cc == XCL_RANGE))
        {
        OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, c - charoffset);
-       COND_VALUE(numberofcmps == 0 ? SLJIT_MOV : SLJIT_OR, TMP2, 0, SLJIT_C_EQUAL);
+       OP_FLAGS(numberofcmps == 0 ? SLJIT_MOV : SLJIT_OR, TMP2, 0, numberofcmps == 0 ? SLJIT_UNUSED : TMP2, 0, SLJIT_C_EQUAL);
        numberofcmps++;
        }
      else if (numberofcmps > 0)
        {
        OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, c - charoffset);
-       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);
        jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
        numberofcmps = 0;
        }
-Line 3007 
 while (*cc != XCL_END)
+Line 3933 
 while (*cc != XCL_END)
      if (numberofcmps < 3 && (*cc == XCL_SINGLE || *cc == XCL_RANGE))
        {
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, c - charoffset);
-       COND_VALUE(numberofcmps == 0 ? SLJIT_MOV : SLJIT_OR, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS(numberofcmps == 0 ? SLJIT_MOV : SLJIT_OR, TMP2, 0, numberofcmps == 0 ? SLJIT_UNUSED : TMP2, 0, SLJIT_C_LESS_EQUAL);
        numberofcmps++;
        }
      else if (numberofcmps > 0)
        {
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, c - charoffset);
-       COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
        jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
        numberofcmps = 0;
        }
-Line 3032 
 while (*cc != XCL_END)
+Line 3958 
 while (*cc != XCL_END)
      switch(*cc)
        {
        case PT_ANY:
-       if (list != fallbacks)
+       if (list != backtracks)
          {
          if ((cc[-1] == XCL_NOTPROP && compares > 0) || (cc[-1] == XCL_PROP && compares == 0))
            continue;
-Line 3044 
 while (*cc != XCL_END)
+Line 3970 
 while (*cc != XCL_END)
        case PT_LAMP:
        OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, typereg, 0, SLJIT_IMM, ucp_Lu - typeoffset);
-       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, typereg, 0, SLJIT_IMM, ucp_Ll - typeoffset);
-       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, typereg, 0, SLJIT_IMM, ucp_Lt - typeoffset);
-       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);
        jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
        break;
-Line 3075 
 while (*cc != XCL_END)
+Line 4001 
 while (*cc != XCL_END)
          }
        SET_CHAR_OFFSET(9);
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, 13 - 9);
-       COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS_EQUAL);
        if (*cc == PT_SPACE)
          JUMPHERE(jump);
        SET_TYPE_OFFSET(ucp_Zl);
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, typereg, 0, SLJIT_IMM, ucp_Zs - ucp_Zl);
-       COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
        jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
        break;
        case PT_WORD:
        OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, CHAR_UNDERSCORE - charoffset);
-       COND_VALUE(SLJIT_MOV, TMP2, 0, SLJIT_C_EQUAL);
+       OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
        /* ... fall through */
        case PT_ALNUM:
        SET_TYPE_OFFSET(ucp_Ll);
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, typereg, 0, SLJIT_IMM, ucp_Lu - ucp_Ll);
-       COND_VALUE((*cc == PT_ALNUM) ? SLJIT_MOV : SLJIT_OR, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS((*cc == PT_ALNUM) ? SLJIT_MOV : SLJIT_OR, TMP2, 0, (*cc == PT_ALNUM) ? SLJIT_UNUSED : TMP2, 0, SLJIT_C_LESS_EQUAL);
        SET_TYPE_OFFSET(ucp_Nd);
        OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, typereg, 0, SLJIT_IMM, ucp_No - ucp_Nd);
-       COND_VALUE(SLJIT_OR | SLJIT_SET_E, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_LESS_EQUAL);
+       jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
+       break;
+       case PT_CLIST:
+       other_cases = PRIV(ucd_caseless_sets) + cc[1];
+       /* At least three characters are required.
+          Otherwise this case would be handled by the normal code path. */
+       SLJIT_ASSERT(other_cases[0] != NOTACHAR && other_cases[1] != NOTACHAR && other_cases[2] != NOTACHAR);
+       SLJIT_ASSERT(other_cases[0] < other_cases[1] && other_cases[1] < other_cases[2]);
+       /* Optimizing character pairs, if their difference is power of 2. */
+       if (is_powerof2(other_cases[1] ^ other_cases[0]))
+         {
+         if (charoffset == 0)
+           OP2(SLJIT_OR, TMP2, 0, TMP1, 0, SLJIT_IMM, other_cases[1] ^ other_cases[0]);
+         else
+           {
+           OP2(SLJIT_ADD, TMP2, 0, TMP1, 0, SLJIT_IMM, (sljit_sw)charoffset);
+           OP2(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_IMM, other_cases[1] ^ other_cases[0]);
+           }
+         OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_IMM, other_cases[1]);
+         OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
+         other_cases += 2;
+         }
+       else if (is_powerof2(other_cases[2] ^ other_cases[1]))
+         {
+         if (charoffset == 0)
+           OP2(SLJIT_OR, TMP2, 0, TMP1, 0, SLJIT_IMM, other_cases[2] ^ other_cases[1]);
+         else
+           {
+           OP2(SLJIT_ADD, TMP2, 0, TMP1, 0, SLJIT_IMM, (sljit_sw)charoffset);
+           OP2(SLJIT_OR, TMP2, 0, TMP2, 0, SLJIT_IMM, other_cases[1] ^ other_cases[0]);
+           }
+         OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP2, 0, SLJIT_IMM, other_cases[2]);
+         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, other_cases[0] - charoffset);
+         OP_FLAGS(SLJIT_OR | ((other_cases[3] == NOTACHAR) ? SLJIT_SET_E : 0), TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
+         other_cases += 3;
+         }
+       else
+         {
+         OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, *other_cases++ - charoffset);
+         OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_EQUAL);
+         }
+       while (*other_cases != NOTACHAR)
+         {
+         OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, *other_cases++ - charoffset);
+         OP_FLAGS(SLJIT_OR | ((*other_cases == NOTACHAR) ? SLJIT_SET_E : 0), TMP2, 0, TMP2, 0, SLJIT_C_EQUAL);
+         }
        jump = JUMP(SLJIT_C_NOT_ZERO ^ invertcmp);
        break;
        }
-Line 3105 
 while (*cc != XCL_END)
+Line 4084 
 while (*cc != XCL_END)
  #endif
    if (jump != NULL)
-     add_jump(compiler, compares > 0 ? list : fallbacks, jump);
+     add_jump(compiler, compares > 0 ? list : backtracks, jump);
    }
  if (found != NULL)
-Line 3117 
 if (found != NULL)
+Line 4096 
 if (found != NULL)
  #endif
- static pcre_uchar *compile_char1_hotpath(compiler_common *common, pcre_uchar type, pcre_uchar *cc, jump_list **fallbacks)
+ static pcre_uchar *compile_char1_matchingpath(compiler_common *common, pcre_uchar type, pcre_uchar *cc, jump_list **backtracks)
  {
  DEFINE_COMPILER;
  int length;
-Line 3136 
 switch(type)
+Line 4115 
 switch(type)
    case OP_SOD:
    OP1(SLJIT_MOV, TMP1, 0, ARGUMENTS, 0);
    OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(TMP1), SLJIT_OFFSETOF(jit_arguments, begin));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, TMP1, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, TMP1, 0));
    return cc;
    case OP_SOM:
    OP1(SLJIT_MOV, TMP1, 0, ARGUMENTS, 0);
    OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(TMP1), SLJIT_OFFSETOF(jit_arguments, str));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, TMP1, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, TMP1, 0));
    return cc;
    case OP_NOT_WORD_BOUNDARY:
    case OP_WORD_BOUNDARY:
    add_jump(compiler, &common->wordboundary, JUMP(SLJIT_FAST_CALL));
-   add_jump(compiler, fallbacks, JUMP(type == OP_NOT_WORD_BOUNDARY ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
+   add_jump(compiler, backtracks, JUMP(type == OP_NOT_WORD_BOUNDARY ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
    return cc;
    case OP_NOT_DIGIT:
    case OP_DIGIT:
-   fallback_at_str_end(common, fallbacks);
+   /* Digits are usually 0-9, so it is worth to optimize them. */
-   read_char8_type(common);
+   if (common->digits[0] == -2)
-   OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ctype_digit);
+     get_ctype_ranges(common, ctype_digit, common->digits);
-   add_jump(compiler, fallbacks, JUMP(type == OP_DIGIT ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
+   detect_partial_match(common, backtracks);
+   /* Flip the starting bit in the negative case. */
+   if (type == OP_NOT_DIGIT)
+     common->digits[1] ^= 1;
+   if (!check_ranges(common, common->digits, backtracks, TRUE))
+     {
+     read_char8_type(common);
+     OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ctype_digit);
+     add_jump(compiler, backtracks, JUMP(type == OP_DIGIT ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
+     }
+   if (type == OP_NOT_DIGIT)
+     common->digits[1] ^= 1;
    return cc;
    case OP_NOT_WHITESPACE:
    case OP_WHITESPACE:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char8_type(common);
    OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ctype_space);
-   add_jump(compiler, fallbacks, JUMP(type == OP_WHITESPACE ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
+   add_jump(compiler, backtracks, JUMP(type == OP_WHITESPACE ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
    return cc;
    case OP_NOT_WORDCHAR:
    case OP_WORDCHAR:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char8_type(common);
    OP2(SLJIT_AND | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, ctype_word);
-   add_jump(compiler, fallbacks, JUMP(type == OP_WORDCHAR ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
+   add_jump(compiler, backtracks, JUMP(type == OP_WORDCHAR ? SLJIT_C_ZERO : SLJIT_C_NOT_ZERO));
    return cc;
    case OP_ANY:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char(common);
    if (common->nltype == NLTYPE_FIXED && common->newline > 255)
      {
      jump[0] = CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff);
-     jump[1] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+     if (common->mode != JIT_PARTIAL_HARD_COMPILE)
+       jump[1] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+     else
+       jump[1] = check_str_end(common);
      OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, common->newline & 0xff));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, common->newline & 0xff));
-     JUMPHERE(jump[1]);
+     if (jump[1] != NULL)
+       JUMPHERE(jump[1]);
      JUMPHERE(jump[0]);
      }
    else
-     check_newlinechar(common, common->nltype, fallbacks, TRUE);
+     check_newlinechar(common, common->nltype, backtracks, TRUE);
    return cc;
    case OP_ALLANY:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
  #ifdef SUPPORT_UTF
    if (common->utf)
      {
      OP1(MOV_UCHAR, TMP1, 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 || defined COMPILE_PCRE16
+ #if defined COMPILE_PCRE8
      jump[0] = CMP(SLJIT_C_LESS, TMP1, 0, SLJIT_IMM, 0xc0);
-     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);
- #else /* COMPILE_PCRE8 */
+ #elif defined COMPILE_PCRE16
- #ifdef COMPILE_PCRE16
      jump[0] = 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);
-     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 /* COMPILE_PCRE16 */
+ #endif
- #endif /* COMPILE_PCRE8 */
      JUMPHERE(jump[0]);
+ #endif /* COMPILE_PCRE[8|16] */
      return cc;
      }
  #endif
-Line 3220 
 switch(type)
+Line 4215 
 switch(type)
    return cc;
    case OP_ANYBYTE:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
    return cc;
-Line 3233 
 switch(type)
+Line 4228 
 switch(type)
    propdata[2] = cc[0];
    propdata[3] = cc[1];
    propdata[4] = XCL_END;
-   compile_xclass_hotpath(common, propdata, fallbacks);
+   compile_xclass_matchingpath(common, propdata, backtracks);
    return cc + 2;
  #endif
  #endif
    case OP_ANYNL:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char(common);
    jump[0] = CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, CHAR_CR);
-   jump[1] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+   /* We don't need to handle soft partial matching case. */
+   if (common->mode != JIT_PARTIAL_HARD_COMPILE)
+     jump[1] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
+   else
+     jump[1] = check_str_end(common);
    OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), 0);
    jump[2] = CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, CHAR_NL);
    OP2(SLJIT_ADD, STR_PTR, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
    jump[3] = JUMP(SLJIT_JUMP);
    JUMPHERE(jump[0]);
-   check_newlinechar(common, common->bsr_nltype, fallbacks, FALSE);
+   check_newlinechar(common, common->bsr_nltype, backtracks, FALSE);
    JUMPHERE(jump[1]);
    JUMPHERE(jump[2]);
    JUMPHERE(jump[3]);
-Line 3256 
 switch(type)
+Line 4255 
 switch(type)
    case OP_NOT_HSPACE:
    case OP_HSPACE:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char(common);
    add_jump(compiler, &common->hspace, JUMP(SLJIT_FAST_CALL));
-   add_jump(compiler, fallbacks, JUMP(type == OP_NOT_HSPACE ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
+   add_jump(compiler, backtracks, JUMP(type == OP_NOT_HSPACE ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
    return cc;
    case OP_NOT_VSPACE:
    case OP_VSPACE:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char(common);
    add_jump(compiler, &common->vspace, JUMP(SLJIT_FAST_CALL));
-   add_jump(compiler, fallbacks, JUMP(type == OP_NOT_VSPACE ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
+   add_jump(compiler, backtracks, JUMP(type == OP_NOT_VSPACE ? SLJIT_C_NOT_ZERO : SLJIT_C_ZERO));
    return cc;
  #ifdef SUPPORT_UCP
    case OP_EXTUNI:
-   fallback_at_str_end(common, fallbacks);
+   detect_partial_match(common, backtracks);
    read_char(common);
    add_jump(compiler, &common->getucd, JUMP(SLJIT_FAST_CALL));
-   OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Mc);
+   OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, gbprop));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_LESS_EQUAL, TMP1, 0, SLJIT_IMM, ucp_Mn - ucp_Mc));
+   /* Optimize register allocation: use a real register. */
+   OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0, STACK_TOP, 0);
+   OP1(SLJIT_MOV_UB, STACK_TOP, 0, SLJIT_MEM2(TMP1, TMP2), 3);
    label = LABEL();
    jump[0] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
    OP1(SLJIT_MOV, TMP3, 0, STR_PTR, 0);
    read_char(common);
    add_jump(compiler, &common->getucd, JUMP(SLJIT_FAST_CALL));
-   OP2(SLJIT_SUB, TMP1, 0, TMP1, 0, SLJIT_IMM, ucp_Mc);
+   OP1(SLJIT_MOV, TMP1, 0, SLJIT_IMM, (sljit_sw)PRIV(ucd_records) + SLJIT_OFFSETOF(ucd_record, gbprop));
-   CMPTO(SLJIT_C_LESS_EQUAL, TMP1, 0, SLJIT_IMM, ucp_Mn - ucp_Mc, label);
+   OP1(SLJIT_MOV_UB, TMP2, 0, SLJIT_MEM2(TMP1, TMP2), 3);
+   OP2(SLJIT_SHL, STACK_TOP, 0, STACK_TOP, 0, SLJIT_IMM, 2);
+   OP1(SLJIT_MOV_UI, TMP1, 0, SLJIT_MEM1(STACK_TOP), (sljit_sw)PRIV(ucp_gbtable));
+   OP1(SLJIT_MOV, STACK_TOP, 0, 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);
+   JUMPTO(SLJIT_C_NOT_ZERO, label);
    OP1(SLJIT_MOV, STR_PTR, 0, TMP3, 0);
    JUMPHERE(jump[0]);
+   OP1(SLJIT_MOV, STACK_TOP, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS0);
    if (common->mode == JIT_PARTIAL_HARD_COMPILE)
      {
      jump[0] = CMP(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0);
-     check_partial(common);
+     /* Since we successfully read a char above, partial matching must occure. */
+     check_partial(common, TRUE);
      JUMPHERE(jump[0]);
      }
    return cc;
  #endif
    case OP_EODN:
+   /* Requires rather complex checks. */
    jump[0] = CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0);
    if (common->nltype == NLTYPE_FIXED && common->newline > 255)
      {
      OP2(SLJIT_ADD, TMP2, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(2));
      OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(0));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, STR_END, 0));
+     if (common->mode == JIT_COMPILE)
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, STR_END, 0));
+     else
+       {
+       jump[1] = CMP(SLJIT_C_EQUAL, TMP2, 0, STR_END, 0);
+       OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP2, 0, STR_END, 0);
+       OP_FLAGS(SLJIT_MOV, TMP2, 0, SLJIT_UNUSED, 0, SLJIT_C_LESS);
+       OP2(SLJIT_SUB | SLJIT_SET_E, SLJIT_UNUSED, 0, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff);
+       OP_FLAGS(SLJIT_OR | SLJIT_SET_E, TMP2, 0, TMP2, 0, SLJIT_C_NOT_EQUAL);
+       add_jump(compiler, backtracks, JUMP(SLJIT_C_NOT_EQUAL));
+       check_partial(common, TRUE);
+       add_jump(compiler, backtracks, JUMP(SLJIT_JUMP));
+       JUMPHERE(jump[1]);
+       }
      OP1(MOV_UCHAR, TMP2, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, (common->newline >> 8) & 0xff));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, common->newline & 0xff));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, common->newline & 0xff));
      }
    else if (common->nltype == NLTYPE_FIXED)
      {
      OP2(SLJIT_ADD, TMP2, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
      OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(0));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, STR_END, 0));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, STR_END, 0));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, common->newline));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, common->newline));
      }
    else
      {
-Line 3322 
 switch(type)
+Line 4347 
 switch(type)
      OP2(SLJIT_ADD, TMP2, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(2));
      OP2(SLJIT_SUB | SLJIT_SET_U, SLJIT_UNUSED, 0, TMP2, 0, STR_END, 0);
      jump[2] = JUMP(SLJIT_C_GREATER);
-     add_jump(compiler, fallbacks, JUMP(SLJIT_C_LESS));
+     add_jump(compiler, backtracks, JUMP(SLJIT_C_LESS));
      /* Equal. */
      OP1(MOV_UCHAR, TMP1, 0, SLJIT_MEM1(STR_PTR), IN_UCHARS(1));
      jump[3] = CMP(SLJIT_C_EQUAL, TMP1, 0, SLJIT_IMM, CHAR_NL);
-     add_jump(compiler, fallbacks, JUMP(SLJIT_JUMP));
+     add_jump(compiler, backtracks, JUMP(SLJIT_JUMP));
      JUMPHERE(jump[1]);
      if (common->nltype == NLTYPE_ANYCRLF)
        {
        OP2(SLJIT_ADD, TMP2, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_LESS, TMP2, 0, STR_END, 0));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_LESS, TMP2, 0, STR_END, 0));
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, CHAR_NL));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP1, 0, SLJIT_IMM, CHAR_NL));
        }
      else
        {
        OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1, STR_PTR, 0);
        read_char(common);
-       add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, STR_END, 0));
+       add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, STR_END, 0));
        add_jump(compiler, &common->anynewline, JUMP(SLJIT_FAST_CALL));
-       add_jump(compiler, fallbacks, JUMP(SLJIT_C_ZERO));
+       add_jump(compiler, backtracks, JUMP(SLJIT_C_ZERO));
        OP1(SLJIT_MOV, STR_PTR, 0, SLJIT_MEM1(SLJIT_LOCALS_REG), LOCALS1);
        }
      JUMPHERE(jump[2]);
      JUMPHERE(jump[3]);
      }
    JUMPHERE(jump[0]);
-   check_partial(common);
+   check_partial(common, FALSE);
    return cc;
    case OP_EOD:
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, STR_PTR, 0, STR_END, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_LESS, STR_PTR, 0, STR_END, 0));
-   check_partial(common);
+   check_partial(common, FALSE);
    return cc;
    case OP_CIRC:
    OP1(SLJIT_MOV, TMP2, 0, ARGUMENTS, 0);
    OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(TMP2), SLJIT_OFFSETOF(jit_arguments, begin));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_GREATER, STR_PTR, 0, TMP1, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_GREATER, STR_PTR, 0, TMP1, 0));
    OP1(SLJIT_MOV_UB, TMP2, 0, SLJIT_MEM1(TMP2), SLJIT_OFFSETOF(jit_arguments, notbol));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, 0));
    return cc;
    case OP_CIRCM:
-Line 3369 
 switch(type)
+Line 4394 
 switch(type)
    OP1(SLJIT_MOV, TMP1, 0, SLJIT_MEM1(TMP2), SLJIT_OFFSETOF(jit_arguments, begin));
    jump[1] = CMP(SLJIT_C_GREATER, STR_PTR, 0, TMP1, 0);
    OP1(SLJIT_MOV_UB, TMP2, 0, SLJIT_MEM1(TMP2), SLJIT_OFFSETOF(jit_arguments, notbol));
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_NOT_EQUAL, TMP2, 0, SLJIT_IMM, 0));
    jump[0] = JUMP(SLJIT_JUMP);
    JUMPHERE(jump[1]);
-   add_jump(compiler, fallbacks, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
+   add_jump(compiler, backtracks, CMP(SLJIT_C_GREATER_EQUAL, STR_PTR, 0, STR_END, 0));
    if (common->nltype == NLTYPE_FIXED && common->newline > 255)
      {
      OP2(SLJIT_SUB, TMP2, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(2));
-     add_jump(compiler, fallbacks, CMP(SLJIT_C_LESS, TMP2, 0, TMP1, 0));
+     add_jump(compiler, backtracks, CMP(SLJIT_C_LESS, TMP2, 0, TMP1, 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));
      add_jump(compiler, fallbacks, CMP(SLJIT_C_NOT_EQUA