Update ICU to 54.1 step 1

source/i18n/regexcmp.cpp

 //
 //  file:  regexcmp.cpp
 //
-//  Copyright (C) 2002-2013 International Business Machines Corporation and others.
+//  Copyright (C) 2002-2014 International Business Machines Corporation and others.
 //  All Rights Reserved.
 //
 //  This file contains the ICU regular expression compiler, which is responsible
 //  for processing a regular expression pattern into the compiled form that
 //  is used by the match finding engine.
 //
 
 #include "unicode/utypes.h"
 
 #if !UCONFIG_NO_REGULAR_EXPRESSIONS
 
 #include "unicode/ustring.h"
 #include "unicode/unistr.h"
 #include "unicode/uniset.h"
 #include "unicode/uchar.h"
 #include "unicode/uchriter.h"
 #include "unicode/parsepos.h"
 #include "unicode/parseerr.h"
 #include "unicode/regex.h"
 #include "unicode/utf.h"
 #include "unicode/utf16.h"
 #include "patternprops.h"
 #include "putilimp.h"
 #include "cmemory.h"
 #include "cstring.h"
 #include "uvectr32.h"
 #include "uvectr64.h"
 #include "uassert.h"
-#include "ucln_in.h"
 #include "uinvchar.h"
 
 #include "regeximp.h"
 #include "regexcst.h"   // Contains state table for the regex pattern parser.
                         //   generated by a Perl script.
 #include "regexcmp.h"
 #include "regexst.h"
 #include "regextxt.h"
 
 
@@ skipping 58 matching lines @@
 //
 //------------------------------------------------------------------------------
 void    RegexCompile::compile(
                          const UnicodeString &pat,   // Source pat to be compiled.
                          UParseError &pp,            // Error position info
                          UErrorCode &e)              // Error Code
 {
     fRXPat->fPatternString = new UnicodeString(pat);
     UText patternText = UTEXT_INITIALIZER;
     utext_openConstUnicodeString(&patternText, fRXPat->fPatternString, &e);
-    
+
     if (U_SUCCESS(e)) {
         compile(&patternText, pp, e);
         utext_close(&patternText);
     }
 }
 
 //
 //   compile, UText mode
 //     All the work is actually done here.
 //
@@ skipping 172 matching lines @@
         n *= 10;
     }
 
     //
     // The pattern's fFrameSize so far has accumulated the requirements for
     //   storage for capture parentheses, counters, etc. that are encountered
     //   in the pattern.  Add space for the two variables that are always
     //   present in the saved state:  the input string position (int64_t) and
     //   the position in the compiled pattern.
     //
-    allocateStackData(RESTACKFRAME_HDRCOUNT);
+    fRXPat->fFrameSize+=RESTACKFRAME_HDRCOUNT;
 
     //
     // Optimization pass 1: NOPs, back-references, and case-folding
     //
     stripNOPs();
 
     //
     // Get bounds for the minimum and maximum length of a string that this
     //   pattern can match.  Used to avoid looking for matches in strings that
     //   are too short.
@@ skipping 45 matching lines @@
 
     case doPatStart:
         // Start of pattern compiles to:
         //0   SAVE   2        Fall back to position of FAIL
         //1   jmp    3
         //2   FAIL            Stop if we ever reach here.
         //3   NOP             Dummy, so start of pattern looks the same as
         //                    the start of an ( grouping.
         //4   NOP             Resreved, will be replaced by a save if there are
         //                    OR | operators at the top level
-        appendOp(URX_BUILD(URX_STATE_SAVE, 2));
-        appendOp(URX_BUILD(URX_JMP,  3));
-        appendOp(URX_BUILD(URX_FAIL, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_STATE_SAVE, 2), *fStatus);
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_JMP,  3), *fStatus);
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_FAIL, 0), *fStatus);
 
         // Standard open nonCapture paren action emits the two NOPs and
         //   sets up the paren stack frame.
         doParseActions(doOpenNonCaptureParen);
         break;
 
     case doPatFinish:
         // We've scanned to the end of the pattern
         //  The end of pattern compiles to:
         //        URX_END
         //    which will stop the runtime match engine.
         //  Encountering end of pattern also behaves like a close paren,
         //   and forces fixups of the State Save at the beginning of the compiled pattern
         //   and of any OR operations at the top level.
         //
         handleCloseParen();
         if (fParenStack.size() > 0) {
             // Missing close paren in pattern.
             error(U_REGEX_MISMATCHED_PAREN);
         }
 
         // add the END operation to the compiled pattern.
-        appendOp(URX_BUILD(URX_END, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_END, 0), *fStatus);
 
         // Terminate the pattern compilation state machine.
         returnVal = FALSE;
         break;
 
 
 
     case doOrOperator:
         // Scanning a '|', as in (A|B)
         {
             // Generate code for any pending literals preceding the '|'
             fixLiterals(FALSE);
 
             // Insert a SAVE operation at the start of the pattern section preceding
             //   this OR at this level.  This SAVE will branch the match forward
             //   to the right hand side of the OR in the event that the left hand
             //   side fails to match and backtracks.  Locate the position for the
             //   save from the location on the top of the parentheses stack.
             int32_t savePosition = fParenStack.popi();
             int32_t op = (int32_t)fRXPat->fCompiledPat->elementAti(savePosition);
             U_ASSERT(URX_TYPE(op) == URX_NOP);  // original contents of reserved location
             op = URX_BUILD(URX_STATE_SAVE, fRXPat->fCompiledPat->size()+1);
             fRXPat->fCompiledPat->setElementAt(op, savePosition);
 
             // Append an JMP operation into the compiled pattern.  The operand for
             //  the JMP will eventually be the location following the ')' for the
             //  group.  This will be patched in later, when the ')' is encountered.
             op = URX_BUILD(URX_JMP, 0);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Push the position of the newly added JMP op onto the parentheses stack.
             // This registers if for fixup when this block's close paren is encountered.
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);
 
             // Append a NOP to the compiled pattern.  This is the slot reserved
             //   for a SAVE in the event that there is yet another '|' following
             //   this one.
-            appendOp(URX_BUILD(URX_NOP, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);
         }
         break;
 
 
     case doOpenCaptureParen:
         // Open Paren.
         //   Compile to a
         //      - NOP, which later may be replaced by a save-state if the
         //         parenthesized group gets a * quantifier, followed by
         //      - START_CAPTURE  n    where n is stack frame offset to the capture group variables.
         //      - NOP, which may later be replaced by a save-state if there
         //             is an '|' alternation within the parens.
         //
         //    Each capture group gets three slots in the save stack frame:
         //         0: Capture Group start position (in input string being matched.)
         //         1: Capture Group end position.
         //         2: Start of Match-in-progress.
         //    The first two locations are for a completed capture group, and are
         //     referred to by back references and the like.
         //    The third location stores the capture start position when an START_CAPTURE is
         //      encountered.  This will be promoted to a completed capture when (and if) the corresponding
         //      END_CAPTURE is encountered.
         {
             fixLiterals();
-            appendOp(URX_BUILD(URX_NOP, 0));
-            int32_t  varsLoc = allocateStackData(3);    // Reserve three slots in match stack frame.
-            int32_t  cop     = URX_BUILD(URX_START_CAPTURE, varsLoc);
-            appendOp(cop);
-            appendOp(URX_BUILD(URX_NOP, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
+            int32_t  varsLoc    = fRXPat->fFrameSize;    // Reserve three slots in match stack frame.
+            fRXPat->fFrameSize += 3;
+            int32_t  cop        = URX_BUILD(URX_START_CAPTURE, varsLoc);
+            fRXPat->fCompiledPat->addElement(cop, *fStatus);
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the two NOPs.  Depending on what follows in the pattern, the
             //   NOPs may be changed to SAVE_STATE or JMP ops, with a target
             //   address of the end of the parenthesized group.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(capturing, *fStatus);                        // Frame type.
             fParenStack.push(fRXPat->fCompiledPat->size()-3, *fStatus);   // The first  NOP location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP loc
 
             // Save the mapping from group number to stack frame variable position.
             fRXPat->fGroupMap->addElement(varsLoc, *fStatus);
         }
          break;
 
     case doOpenNonCaptureParen:
         // Open non-caputuring (grouping only) Paren.
         //   Compile to a
         //      - NOP, which later may be replaced by a save-state if the
         //         parenthesized group gets a * quantifier, followed by
         //      - NOP, which may later be replaced by a save-state if there
         //             is an '|' alternation within the parens.
         {
             fixLiterals();
-            appendOp(URX_BUILD(URX_NOP, 0));
-            appendOp(URX_BUILD(URX_NOP, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the two NOPs.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(plain,      *fStatus);                       // Begin a new frame.
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The first  NOP location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP loc
         }
          break;
 
 
     case doOpenAtomicParen:
         // Open Atomic Paren.  (?>
         //   Compile to a
         //      - NOP, which later may be replaced if the parenthesized group
         //         has a quantifier, followed by
         //      - STO_SP  save state stack position, so it can be restored at the ")"
         //      - NOP, which may later be replaced by a save-state if there
         //             is an '|' alternation within the parens.
         {
             fixLiterals();
-            appendOp(URX_BUILD(URX_NOP, 0));
-            int32_t  varLoc = allocateData(1);    // Reserve a data location for saving the state stack ptr.
-            int32_t  stoOp = URX_BUILD(URX_STO_SP, varLoc);
-            appendOp(stoOp);
-            appendOp(URX_BUILD(URX_NOP, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
+            int32_t  varLoc    = fRXPat->fDataSize;    // Reserve a data location for saving the
+            fRXPat->fDataSize += 1;                    //  state stack ptr.
+            int32_t  stoOp     = URX_BUILD(URX_STO_SP, varLoc);
+            fRXPat->fCompiledPat->addElement(stoOp, *fStatus);
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the two NOPs.  Depending on what follows in the pattern, the
             //   NOPs may be changed to SAVE_STATE or JMP ops, with a target
             //   address of the end of the parenthesized group.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(atomic, *fStatus);                           // Frame type.
             fParenStack.push(fRXPat->fCompiledPat->size()-3, *fStatus);   // The first NOP
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP
         }
@@ skipping 22 matching lines @@
         //    6.   NOP              reserved for use by quantifiers on the block.
         //                          Look-ahead can't have quantifiers, but paren stack
         //                             compile time conventions require the slot anyhow.
         //    7.   NOP              may be replaced if there is are '|' ops in the block.
         //    8.     code for parenthesized stuff.
         //    9.   LA_END
         //
         //  Two data slots are reserved, for saving the stack ptr and the input position.
         {
             fixLiterals();
-            int32_t dataLoc = allocateData(2);
+            int32_t dataLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize += 2;
             int32_t op = URX_BUILD(URX_LA_START, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_STATE_SAVE, fRXPat->fCompiledPat->size()+ 2);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_JMP, fRXPat->fCompiledPat->size()+ 3);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_LA_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_BACKTRACK, 0);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_NOP, 0);
-            appendOp(op);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the NOPs.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(lookAhead, *fStatus);                        // Frame type.
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The first  NOP location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP location
         }
         break;
 
     case doOpenLookAheadNeg:
         // Negated Lookahead.   (?! stuff )
         // Compiles to
         //    1.    START_LA    dataloc
         //    2.    SAVE_STATE  7         // Fail within look-ahead block restores to this state,
         //                                //   which continues with the match.
         //    3.    NOP                   // Std. Open Paren sequence, for possible '|'
         //    4.       code for parenthesized stuff.
         //    5.    END_LA                // Cut back stack, remove saved state from step 2.
         //    6.    BACKTRACK             // code in block succeeded, so neg. lookahead fails.
         //    7.    END_LA                // Restore match region, in case look-ahead was using
         //                                        an alternate (transparent) region.
         {
             fixLiterals();
-            int32_t dataLoc = allocateData(2);
+            int32_t dataLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize += 2;
             int32_t op = URX_BUILD(URX_LA_START, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_STATE_SAVE, 0);    // dest address will be patched later.
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             op = URX_BUILD(URX_NOP, 0);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the StateSave and NOP.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(negLookAhead, *fStatus);                    // Frame type
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The STATE_SAVE location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP location
 
             // Instructions #5 - #7 will be added when the ')' is encountered.
         }
@@ skipping 17 matching lines @@
             //          Allocate a block of matcher data, to contain (when running a match)
             //              0:    Stack ptr on entry
             //              1:    Input Index on entry
             //              2:    Start index of match current match attempt.
             //              3:    Original Input String len.
 
             // Generate match code for any pending literals.
             fixLiterals();
 
             // Allocate data space
-            int32_t dataLoc = allocateData(4);
+            int32_t dataLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize += 4;
 
             // Emit URX_LB_START
             int32_t op = URX_BUILD(URX_LB_START, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Emit URX_LB_CONT
             op = URX_BUILD(URX_LB_CONT, dataLoc);
-            appendOp(op);
-            appendOp(0);    // MinMatchLength.  To be filled later.
-            appendOp(0);    // MaxMatchLength.  To be filled later.
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+            fRXPat->fCompiledPat->addElement(0,  *fStatus);    // MinMatchLength.  To be filled later.
+            fRXPat->fCompiledPat->addElement(0,  *fStatus);    // MaxMatchLength.  To be filled later.
 
             // Emit the NOP
             op = URX_BUILD(URX_NOP, 0);
-            appendOp(op);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the URX_LB_CONT and the NOP.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(lookBehind, *fStatus);                       // Frame type
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The first NOP location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The 2nd   NOP location
 
             // The final two instructions will be added when the ')' is encountered.
         }
@@ skipping 19 matching lines @@
             //          Allocate a block of matcher data, to contain (when running a match)
             //              0:    Stack ptr on entry
             //              1:    Input Index on entry
             //              2:    Start index of match current match attempt.
             //              3:    Original Input String len.
 
             // Generate match code for any pending literals.
             fixLiterals();
 
             // Allocate data space
-            int32_t dataLoc = allocateData(4);
+            int32_t dataLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize += 4;
 
             // Emit URX_LB_START
             int32_t op = URX_BUILD(URX_LB_START, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Emit URX_LBN_CONT
             op = URX_BUILD(URX_LBN_CONT, dataLoc);
-            appendOp(op);
-            appendOp(0);    // MinMatchLength.  To be filled later.
-            appendOp(0);    // MaxMatchLength.  To be filled later.
-            appendOp(0);    // Continue Loc.    To be filled later.
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+            fRXPat->fCompiledPat->addElement(0,  *fStatus);    // MinMatchLength.  To be filled later.
+            fRXPat->fCompiledPat->addElement(0,  *fStatus);    // MaxMatchLength.  To be filled later.
+            fRXPat->fCompiledPat->addElement(0,  *fStatus);    // Continue Loc.    To be filled later.
 
             // Emit the NOP
             op = URX_BUILD(URX_NOP, 0);
-            appendOp(op);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the URX_LB_CONT and the NOP.
             fParenStack.push(fModeFlags, *fStatus);                       // Match mode state
             fParenStack.push(lookBehindN, *fStatus);                      // Frame type
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The first NOP location
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The 2nd   NOP location
 
             // The final two instructions will be added when the ')' is encountered.
         }
@@ skipping 50 matching lines @@
             int32_t  topLoc = blockTopLoc(FALSE);        // location of item #1
             int32_t  frameLoc;
 
             // Check for simple constructs, which may get special optimized code.
             if (topLoc == fRXPat->fCompiledPat->size() - 1) {
                 int32_t repeatedOp = (int32_t)fRXPat->fCompiledPat->elementAti(topLoc);
 
                 if (URX_TYPE(repeatedOp) == URX_SETREF) {
                     // Emit optimized code for [char set]+
                     int32_t loopOpI = URX_BUILD(URX_LOOP_SR_I, URX_VAL(repeatedOp));
-                    appendOp(loopOpI);
-                    frameLoc = allocateStackData(1);
+                    fRXPat->fCompiledPat->addElement(loopOpI, *fStatus);
+                    frameLoc = fRXPat->fFrameSize;
+                    fRXPat->fFrameSize++;
                     int32_t loopOpC = URX_BUILD(URX_LOOP_C, frameLoc);
-                    appendOp(loopOpC);
+                    fRXPat->fCompiledPat->addElement(loopOpC, *fStatus);
                     break;
                 }
 
                 if (URX_TYPE(repeatedOp) == URX_DOTANY ||
                     URX_TYPE(repeatedOp) == URX_DOTANY_ALL ||
                     URX_TYPE(repeatedOp) == URX_DOTANY_UNIX) {
                     // Emit Optimized code for .+ operations.
                     int32_t loopOpI = URX_BUILD(URX_LOOP_DOT_I, 0);
                     if (URX_TYPE(repeatedOp) == URX_DOTANY_ALL) {
                         // URX_LOOP_DOT_I operand is a flag indicating ". matches any" mode.
                         loopOpI |= 1;
                     }
                     if (fModeFlags & UREGEX_UNIX_LINES) {
                         loopOpI |= 2;
                     }
-                    appendOp(loopOpI);
-                    frameLoc = allocateStackData(1);
+                    fRXPat->fCompiledPat->addElement(loopOpI, *fStatus);
+                    frameLoc = fRXPat->fFrameSize;
+                    fRXPat->fFrameSize++;
                     int32_t loopOpC = URX_BUILD(URX_LOOP_C, frameLoc);
-                    appendOp(loopOpC);
+                    fRXPat->fCompiledPat->addElement(loopOpC, *fStatus);
                     break;
                 }
 
             }
 
             // General case.
 
             // Check for minimum match length of zero, which requires
             //    extra loop-breaking code.
             if (minMatchLength(topLoc, fRXPat->fCompiledPat->size()-1) == 0) {
                 // Zero length match is possible.
                 // Emit the code sequence that can handle it.
                 insertOp(topLoc);
-                frameLoc = allocateStackData(1);
+                frameLoc =  fRXPat->fFrameSize;
+                fRXPat->fFrameSize++;
 
                 int32_t op = URX_BUILD(URX_STO_INP_LOC, frameLoc);
                 fRXPat->fCompiledPat->setElementAt(op, topLoc);
 
                 op = URX_BUILD(URX_JMP_SAV_X, topLoc+1);
-                appendOp(op);
+                fRXPat->fCompiledPat->addElement(op, *fStatus);
             } else {
                 // Simpler code when the repeated body must match something non-empty
                 int32_t  jmpOp  = URX_BUILD(URX_JMP_SAV, topLoc);
-                appendOp(jmpOp);
+                fRXPat->fCompiledPat->addElement(jmpOp, *fStatus);
             }
         }
         break;
 
     case doNGPlus:
         //  Non-greedy '+?'  compiles to
         //     1.   stuff to be repeated  (already built)
         //     2.   state-save  1
         //     3.   ...
         {
             int32_t topLoc      = blockTopLoc(FALSE);
             int32_t saveStateOp = URX_BUILD(URX_STATE_SAVE, topLoc);
-            appendOp(saveStateOp);
+            fRXPat->fCompiledPat->addElement(saveStateOp, *fStatus);
         }
         break;
 
 
     case doOpt:
         // Normal (greedy) ? quantifier.
         //  Compiles to
         //     1. state save 3
         //     2.    body of optional block
         //     3. ...
@@ skipping 16 matching lines @@
         //  This code is less than ideal, with two jmps instead of one, because we can only
         //  insert one instruction at the top of the block being iterated.
         {
             int32_t  jmp1_loc = blockTopLoc(TRUE);
             int32_t  jmp2_loc = fRXPat->fCompiledPat->size();
 
             int32_t  jmp1_op  = URX_BUILD(URX_JMP, jmp2_loc+1);
             fRXPat->fCompiledPat->setElementAt(jmp1_op, jmp1_loc);
 
             int32_t  jmp2_op  = URX_BUILD(URX_JMP, jmp2_loc+2);
-            appendOp(jmp2_op);
+            fRXPat->fCompiledPat->addElement(jmp2_op, *fStatus);
 
             int32_t  save_op  = URX_BUILD(URX_STATE_SAVE, jmp1_loc+1);
-            appendOp(save_op);
+            fRXPat->fCompiledPat->addElement(save_op, *fStatus);
         }
         break;
 
 
     case doStar:
         // Normal (greedy) * quantifier.
         // Compiles to
         //       1.   STATE_SAVE   4
         //       2.      body of stuff being iterated over
         //       3.   JMP_SAV      2
@@ skipping 21 matching lines @@
 
             // Check for simple *, where the construct being repeated
             //   compiled to single opcode, and might be optimizable.
             if (topLoc == fRXPat->fCompiledPat->size() - 1) {
                 int32_t repeatedOp = (int32_t)fRXPat->fCompiledPat->elementAti(topLoc);
 
                 if (URX_TYPE(repeatedOp) == URX_SETREF) {
                     // Emit optimized code for a [char set]*
                     int32_t loopOpI = URX_BUILD(URX_LOOP_SR_I, URX_VAL(repeatedOp));
                     fRXPat->fCompiledPat->setElementAt(loopOpI, topLoc);
-                    dataLoc = allocateStackData(1);
+                    dataLoc = fRXPat->fFrameSize;
+                    fRXPat->fFrameSize++;
                     int32_t loopOpC = URX_BUILD(URX_LOOP_C, dataLoc);
-                    appendOp(loopOpC);
+                    fRXPat->fCompiledPat->addElement(loopOpC, *fStatus);
                     break;
                 }
 
                 if (URX_TYPE(repeatedOp) == URX_DOTANY ||
                     URX_TYPE(repeatedOp) == URX_DOTANY_ALL ||
                     URX_TYPE(repeatedOp) == URX_DOTANY_UNIX) {
                     // Emit Optimized code for .* operations.
                     int32_t loopOpI = URX_BUILD(URX_LOOP_DOT_I, 0);
                     if (URX_TYPE(repeatedOp) == URX_DOTANY_ALL) {
                         // URX_LOOP_DOT_I operand is a flag indicating . matches any mode.
                         loopOpI |= 1;
                     }
                     if ((fModeFlags & UREGEX_UNIX_LINES) != 0) {
                         loopOpI |= 2;
                     }
                     fRXPat->fCompiledPat->setElementAt(loopOpI, topLoc);
-                    dataLoc = allocateStackData(1);
+                    dataLoc = fRXPat->fFrameSize;
+                    fRXPat->fFrameSize++;
                     int32_t loopOpC = URX_BUILD(URX_LOOP_C, dataLoc);
-                    appendOp(loopOpC);
+                    fRXPat->fCompiledPat->addElement(loopOpC, *fStatus);
                     break;
                 }
             }
 
             // Emit general case code for this *
             // The optimizations did not apply.
 
             int32_t   saveStateLoc = blockTopLoc(TRUE);
             int32_t   jmpOp        = URX_BUILD(URX_JMP_SAV, saveStateLoc+1);
 
             // Check for minimum match length of zero, which requires
             //    extra loop-breaking code.
             if (minMatchLength(saveStateLoc, fRXPat->fCompiledPat->size()-1) == 0) {
                 insertOp(saveStateLoc);
-                dataLoc = allocateStackData(1);
+                dataLoc =  fRXPat->fFrameSize;
+                fRXPat->fFrameSize++;
 
                 int32_t op = URX_BUILD(URX_STO_INP_LOC, dataLoc);
                 fRXPat->fCompiledPat->setElementAt(op, saveStateLoc+1);
                 jmpOp      = URX_BUILD(URX_JMP_SAV_X, saveStateLoc+2);
             }
 
             // Locate the position in the compiled pattern where the match will continue
             //   after completing the *.   (4 or 5 in the comment above)
             int32_t continueLoc = fRXPat->fCompiledPat->size()+1;
 
             // Put together the save state op store it into the compiled code.
             int32_t saveStateOp = URX_BUILD(URX_STATE_SAVE, continueLoc);
             fRXPat->fCompiledPat->setElementAt(saveStateOp, saveStateLoc);
 
             // Append the URX_JMP_SAV or URX_JMPX operation to the compiled pattern.
-            appendOp(jmpOp);
+            fRXPat->fCompiledPat->addElement(jmpOp, *fStatus);
         }
         break;
 
     case doNGStar:
         // Non-greedy *? quantifier
         // compiles to
         //     1.   JMP    3
         //     2.      body of stuff being iterated over
         //     3.   STATE_SAVE  2
         //     4    ...
         {
             int32_t     jmpLoc  = blockTopLoc(TRUE);                   // loc  1.
             int32_t     saveLoc = fRXPat->fCompiledPat->size();        // loc  3.
             int32_t     jmpOp   = URX_BUILD(URX_JMP, saveLoc);
             int32_t     stateSaveOp = URX_BUILD(URX_STATE_SAVE, jmpLoc+1);
             fRXPat->fCompiledPat->setElementAt(jmpOp, jmpLoc);
-            appendOp(stateSaveOp);
+            fRXPat->fCompiledPat->addElement(stateSaveOp, *fStatus);
         }
         break;
 
 
     case doIntervalInit:
         // The '{' opening an interval quantifier was just scanned.
         // Init the counter varaiables that will accumulate the values as the digits
         //    are scanned.
         fIntervalLow = 0;
         fIntervalUpper = -1;
@@ skipping 48 matching lines @@
             int32_t topLoc = blockTopLoc(FALSE);
 
             // Produce normal looping code.
             compileInterval(URX_CTR_INIT, URX_CTR_LOOP);
 
             // Surround the just-emitted normal looping code with a STO_SP ... LD_SP
             //  just as if the loop was inclosed in atomic parentheses.
 
             // First the STO_SP before the start of the loop
             insertOp(topLoc);
-
-            int32_t  varLoc = allocateData(1);   // Reserve a data location for saving the
-            int32_t  op     = URX_BUILD(URX_STO_SP, varLoc);
+            int32_t  varLoc    = fRXPat->fDataSize;    // Reserve a data location for saving the
+            fRXPat->fDataSize += 1;                    //  state stack ptr.
+            int32_t  op        = URX_BUILD(URX_STO_SP, varLoc);
             fRXPat->fCompiledPat->setElementAt(op, topLoc);
 
             int32_t loopOp = (int32_t)fRXPat->fCompiledPat->popi();
             U_ASSERT(URX_TYPE(loopOp) == URX_CTR_LOOP && URX_VAL(loopOp) == topLoc);
             loopOp++;     // point LoopOp after the just-inserted STO_SP
             fRXPat->fCompiledPat->push(loopOp, *fStatus);
 
             // Then the LD_SP after the end of the loop
             op = URX_BUILD(URX_LD_SP, varLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
 
         break;
 
     case doNGInterval:
         // Finished scanning a non-greedy {lower,upper}? interval.  Generate the code for it.
         compileInterval(URX_CTR_INIT_NG, URX_CTR_LOOP_NG);
         break;
 
     case doIntervalError:
@@ skipping 23 matching lines @@
         {
             fixLiterals(FALSE);
             int32_t   op;
             if (fModeFlags & UREGEX_DOTALL) {
                 op = URX_BUILD(URX_DOTANY_ALL, 0);
             } else if (fModeFlags & UREGEX_UNIX_LINES) {
                 op = URX_BUILD(URX_DOTANY_UNIX, 0);
             } else {
                 op = URX_BUILD(URX_DOTANY, 0);
             }
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
     case doCaret:
         {
             fixLiterals(FALSE);
             int32_t op = 0;
             if (       (fModeFlags & UREGEX_MULTILINE) == 0 && (fModeFlags & UREGEX_UNIX_LINES) == 0) {
                 op = URX_CARET;
             } else if ((fModeFlags & UREGEX_MULTILINE) != 0 && (fModeFlags & UREGEX_UNIX_LINES) == 0) {
                 op = URX_CARET_M;
             } else if ((fModeFlags & UREGEX_MULTILINE) == 0 && (fModeFlags & UREGEX_UNIX_LINES) != 0) {
-                op = URX_CARET;   // Only testing true start of input. 
+                op = URX_CARET;   // Only testing true start of input.
             } else if ((fModeFlags & UREGEX_MULTILINE) != 0 && (fModeFlags & UREGEX_UNIX_LINES) != 0) {
                 op = URX_CARET_M_UNIX;
             }
-            appendOp(URX_BUILD(op, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(op, 0), *fStatus);
         }
         break;
 
     case doDollar:
         {
             fixLiterals(FALSE);
             int32_t op = 0;
             if (       (fModeFlags & UREGEX_MULTILINE) == 0 && (fModeFlags & UREGEX_UNIX_LINES) == 0) {
                 op = URX_DOLLAR;
             } else if ((fModeFlags & UREGEX_MULTILINE) != 0 && (fModeFlags & UREGEX_UNIX_LINES) == 0) {
                 op = URX_DOLLAR_M;
             } else if ((fModeFlags & UREGEX_MULTILINE) == 0 && (fModeFlags & UREGEX_UNIX_LINES) != 0) {
                 op = URX_DOLLAR_D;
             } else if ((fModeFlags & UREGEX_MULTILINE) != 0 && (fModeFlags & UREGEX_UNIX_LINES) != 0) {
                 op = URX_DOLLAR_MD;
             }
-            appendOp(URX_BUILD(op, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(op, 0), *fStatus);
         }
         break;
 
     case doBackslashA:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_CARET, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_CARET, 0), *fStatus);
         break;
 
     case doBackslashB:
         {
             #if  UCONFIG_NO_BREAK_ITERATION==1
             if (fModeFlags & UREGEX_UWORD) {
                 error(U_UNSUPPORTED_ERROR);
             }
             #endif
             fixLiterals(FALSE);
             int32_t op = (fModeFlags & UREGEX_UWORD)? URX_BACKSLASH_BU : URX_BACKSLASH_B;
-            appendOp(URX_BUILD(op, 1));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(op, 1), *fStatus);
         }
         break;
 
     case doBackslashb:
         {
             #if  UCONFIG_NO_BREAK_ITERATION==1
             if (fModeFlags & UREGEX_UWORD) {
                 error(U_UNSUPPORTED_ERROR);
             }
             #endif
             fixLiterals(FALSE);
             int32_t op = (fModeFlags & UREGEX_UWORD)? URX_BACKSLASH_BU : URX_BACKSLASH_B;
-            appendOp(URX_BUILD(op, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(op, 0), *fStatus);
         }
         break;
 
     case doBackslashD:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_BACKSLASH_D, 1));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKSLASH_D, 1), *fStatus);
         break;
 
     case doBackslashd:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_BACKSLASH_D, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKSLASH_D, 0), *fStatus);
         break;
 
     case doBackslashG:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_BACKSLASH_G, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKSLASH_G, 0), *fStatus);
         break;
 
     case doBackslashS:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_STAT_SETREF_N, URX_ISSPACE_SET));
+        fRXPat->fCompiledPat->addElement(
+            URX_BUILD(URX_STAT_SETREF_N, URX_ISSPACE_SET), *fStatus);
         break;
 
     case doBackslashs:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_STATIC_SETREF, URX_ISSPACE_SET));
+        fRXPat->fCompiledPat->addElement(
+            URX_BUILD(URX_STATIC_SETREF, URX_ISSPACE_SET), *fStatus);
         break;
 
     case doBackslashW:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_STAT_SETREF_N, URX_ISWORD_SET));
+        fRXPat->fCompiledPat->addElement(
+            URX_BUILD(URX_STAT_SETREF_N, URX_ISWORD_SET), *fStatus);
         break;
 
     case doBackslashw:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_STATIC_SETREF, URX_ISWORD_SET));
+        fRXPat->fCompiledPat->addElement(
+            URX_BUILD(URX_STATIC_SETREF, URX_ISWORD_SET), *fStatus);
         break;
 
     case doBackslashX:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_BACKSLASH_X, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKSLASH_X, 0), *fStatus);
         break;
 
 
     case doBackslashZ:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_DOLLAR, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_DOLLAR, 0), *fStatus);
         break;
 
     case doBackslashz:
         fixLiterals(FALSE);
-        appendOp(URX_BUILD(URX_BACKSLASH_Z, 0));
+        fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKSLASH_Z, 0), *fStatus);
         break;
 
     case doEscapeError:
         error(U_REGEX_BAD_ESCAPE_SEQUENCE);
         break;
 
     case doExit:
         fixLiterals(FALSE);
         returnVal = FALSE;
         break;
 
     case doProperty:
         {
             fixLiterals(FALSE);
             UnicodeSet *theSet = scanProp();
             compileSet(theSet);
         }
         break;
 
     case doNamedChar:
         {
             UChar32 c = scanNamedChar();
             literalChar(c);
         }
         break;
-        
+
 
     case doBackRef:
         // BackReference.  Somewhat unusual in that the front-end can not completely parse
         //                 the regular expression, because the number of digits to be consumed
         //                 depends on the number of capture groups that have been defined.  So
         //                 we have to do it here instead.
         {
             int32_t  numCaptureGroups = fRXPat->fGroupMap->size();
             int32_t  groupNum = 0;
             UChar32  c        = fC.fChar;
@@ skipping 19 matching lines @@
             //  of compilation, it will be changed to the variable's location.
             U_ASSERT(groupNum > 0);  // Shouldn't happen.  '\0' begins an octal escape sequence,
                                      //    and shouldn't enter this code path at all.
             fixLiterals(FALSE);
             int32_t  op;
             if (fModeFlags & UREGEX_CASE_INSENSITIVE) {
                 op = URX_BUILD(URX_BACKREF_I, groupNum);
             } else {
                 op = URX_BUILD(URX_BACKREF, groupNum);
             }
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
 
     case doPossessivePlus:
         // Possessive ++ quantifier.
         // Compiles to
         //       1.   STO_SP
         //       2.      body of stuff being iterated over
         //       3.   STATE_SAVE 5
         //       4.   JMP        2
         //       5.   LD_SP
         //       6.   ...
         //
         //  Note:  TODO:  This is pretty inefficient.  A mass of saved state is built up
         //                then unconditionally discarded.  Perhaps introduce a new opcode.  Ticket 6056
         //
         {
             // Emit the STO_SP
             int32_t   topLoc = blockTopLoc(TRUE);
-            int32_t   stoLoc = allocateData(1);  // Reserve the data location for storing save stack ptr.
+            int32_t   stoLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize++;       // Reserve the data location for storing save stack ptr.
             int32_t   op     = URX_BUILD(URX_STO_SP, stoLoc);
             fRXPat->fCompiledPat->setElementAt(op, topLoc);
 
             // Emit the STATE_SAVE
             op = URX_BUILD(URX_STATE_SAVE, fRXPat->fCompiledPat->size()+2);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Emit the JMP
             op = URX_BUILD(URX_JMP, topLoc+1);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Emit the LD_SP
             op = URX_BUILD(URX_LD_SP, stoLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
     case doPossessiveStar:
         // Possessive *+ quantifier.
         // Compiles to
         //       1.   STO_SP       loc
         //       2.   STATE_SAVE   5
         //       3.      body of stuff being iterated over
         //       4.   JMP          2
         //       5.   LD_SP        loc
         //       6    ...
         // TODO:  do something to cut back the state stack each time through the loop.
         {
             // Reserve two slots at the top of the block.
             int32_t   topLoc = blockTopLoc(TRUE);
             insertOp(topLoc);
 
             // emit   STO_SP     loc
-            int32_t   stoLoc = allocateData(1);    // Reserve the data location for storing save stack ptr.
+            int32_t   stoLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize++;       // Reserve the data location for storing save stack ptr.
             int32_t   op     = URX_BUILD(URX_STO_SP, stoLoc);
             fRXPat->fCompiledPat->setElementAt(op, topLoc);
 
             // Emit the SAVE_STATE   5
             int32_t L7 = fRXPat->fCompiledPat->size()+1;
             op = URX_BUILD(URX_STATE_SAVE, L7);
             fRXPat->fCompiledPat->setElementAt(op, topLoc+1);
 
             // Append the JMP operation.
             op = URX_BUILD(URX_JMP, topLoc+1);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Emit the LD_SP       loc
             op = URX_BUILD(URX_LD_SP, stoLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
     case doPossessiveOpt:
         // Possessive  ?+ quantifier.
         //  Compiles to
         //     1. STO_SP      loc
         //     2. SAVE_STATE  5
         //     3.    body of optional block
         //     4. LD_SP       loc
         //     5. ...
         //
         {
             // Reserve two slots at the top of the block.
             int32_t   topLoc = blockTopLoc(TRUE);
             insertOp(topLoc);
 
             // Emit the STO_SP
-            int32_t   stoLoc = allocateData(1);   // Reserve the data location for storing save stack ptr.
+            int32_t   stoLoc = fRXPat->fDataSize;
+            fRXPat->fDataSize++;       // Reserve the data location for storing save stack ptr.
             int32_t   op     = URX_BUILD(URX_STO_SP, stoLoc);
             fRXPat->fCompiledPat->setElementAt(op, topLoc);
 
             // Emit the SAVE_STATE
             int32_t   continueLoc = fRXPat->fCompiledPat->size()+1;
             op = URX_BUILD(URX_STATE_SAVE, continueLoc);
             fRXPat->fCompiledPat->setElementAt(op, topLoc+1);
 
             // Emit the LD_SP
             op = URX_BUILD(URX_LD_SP, stoLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
 
     case doBeginMatchMode:
         fNewModeFlags = fModeFlags;
         fSetModeFlag  = TRUE;
         break;
 
     case doMatchMode:   //  (?i)    and similar
@@ skipping 36 matching lines @@
         // We've got a (?i: or similar.  Begin a parenthesized block, save old
         //   mode flags so they can be restored at the close of the block.
         //
         //   Compile to a
         //      - NOP, which later may be replaced by a save-state if the
         //         parenthesized group gets a * quantifier, followed by
         //      - NOP, which may later be replaced by a save-state if there
         //             is an '|' alternation within the parens.
         {
             fixLiterals(FALSE);
-            appendOp(URX_BUILD(URX_NOP, 0));
-            appendOp(URX_BUILD(URX_NOP, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_NOP, 0), *fStatus);
 
             // On the Parentheses stack, start a new frame and add the postions
             //   of the two NOPs (a normal non-capturing () frame, except for the
             //   saving of the orignal mode flags.)
             fParenStack.push(fModeFlags, *fStatus);
             fParenStack.push(flags, *fStatus);                            // Frame Marker
             fParenStack.push(fRXPat->fCompiledPat->size()-2, *fStatus);   // The first NOP
             fParenStack.push(fRXPat->fCompiledPat->size()-1, *fStatus);   // The second NOP
 
             // Set the current mode flags to the new values.
@@ skipping 140 matching lines @@
         // Finished a complete set expression, including all nested sets.
         //   The close bracket has already triggered clearing out pending set operators,
         //    the operator stack should be empty and the operand stack should have just
         //    one entry, the result set.
         U_ASSERT(fSetOpStack.empty());
         UnicodeSet *theSet = (UnicodeSet *)fSetStack.pop();
         U_ASSERT(fSetStack.empty());
         compileSet(theSet);
         break;
         }
-        
+
     case doSetIntersection2:
         // Have scanned something like [abc&&
         setPushOp(setIntersection2);
         break;
 
     case doSetLiteral:
         // Union the just-scanned literal character into the set being built.
         //    This operation is the highest precedence set operation, so we can always do
         //    it immediately, without waiting to see what follows.  It is necessary to perform
         //    any pending '-' or '&' operation first, because these have the same precedence
-        //    as union-ing in a literal' 
+        //    as union-ing in a literal'
         {
             setEval(setUnion);
             UnicodeSet *s = (UnicodeSet *)fSetStack.peek();
             s->add(fC.fChar);
             fLastSetLiteral = fC.fChar;
             break;
         }
 
     case doSetLiteralEscaped:
         // A back-slash escaped literal character was encountered.
@@ skipping 74 matching lines @@
     case doSetPosixProp:
         {
             UnicodeSet *s = scanPosixProp();
             if (s != NULL) {
                 UnicodeSet *tos = (UnicodeSet *)fSetStack.peek();
                 tos->addAll(*s);
                 delete s;
             }  // else error.  scanProp() reported the error status already.
         }
         break;
-        
+
     case doSetProp:
         //  Scanned a \p \P within [brackets].
         {
             UnicodeSet *s = scanProp();
             if (s != NULL) {
                 UnicodeSet *tos = (UnicodeSet *)fSetStack.peek();
                 tos->addAll(*s);
                 delete s;
             }  // else error.  scanProp() reported the error status already.
         }
         break;
 
 
     case doSetRange:
         // We have scanned literal-literal.  Add the range to the set.
         // The left character is already in the set, and is saved in fLastSetLiteral.
         // The right side is the current character.
         // Lower Limit > Upper limit being an error matches both Java
         //        and ICU UnicodeSet behavior.
         {
         if (fLastSetLiteral > fC.fChar) {
-            error(U_REGEX_INVALID_RANGE);  
+            error(U_REGEX_INVALID_RANGE);
         }
         UnicodeSet *s = (UnicodeSet *)fSetStack.peek();
         s->add(fLastSetLiteral, fC.fChar);
         break;
         }
 
     default:
         U_ASSERT(FALSE);
         error(U_REGEX_INTERNAL_ERROR);
         break;
@@ skipping 38 matching lines @@
 
     // If no literal characters have been scanned but not yet had code generated
     //   for them, nothing needs to be done.
     if (fLiteralChars.length() == 0) {
         return;
     }
 
     int32_t indexOfLastCodePoint = fLiteralChars.moveIndex32(fLiteralChars.length(), -1);
     UChar32 lastCodePoint = fLiteralChars.char32At(indexOfLastCodePoint);
 
-    // Split:  We need to  ensure that the last item in the compiled pattern 
+    // Split:  We need to  ensure that the last item in the compiled pattern
     //     refers only to the last literal scanned in the pattern, so that
     //     quantifiers (*, +, etc.) affect only it, and not a longer string.
     //     Split before case folding for case insensitive matches.
 
     if (split) {
         fLiteralChars.truncate(indexOfLastCodePoint);
         fixLiterals(FALSE);   // Recursive call, emit code to match the first part of the string.
                               //  Note that the truncated literal string may be empty, in which case
                               //  nothing will be emitted.
 
         literalChar(lastCodePoint);  // Re-add the last code point as if it were a new literal.
         fixLiterals(FALSE);          // Second recursive call, code for the final code point.
         return;
     }
 
     // If we are doing case-insensitive matching, case fold the string.  This may expand
     //   the string, e.g. the German sharp-s turns into "ss"
     if (fModeFlags & UREGEX_CASE_INSENSITIVE) {
         fLiteralChars.foldCase();
         indexOfLastCodePoint = fLiteralChars.moveIndex32(fLiteralChars.length(), -1);
         lastCodePoint = fLiteralChars.char32At(indexOfLastCodePoint);
     }
 
     if (indexOfLastCodePoint == 0) {
         // Single character, emit a URX_ONECHAR op to match it.
-        if ((fModeFlags & UREGEX_CASE_INSENSITIVE) && 
+        if ((fModeFlags & UREGEX_CASE_INSENSITIVE) &&
                  u_hasBinaryProperty(lastCodePoint, UCHAR_CASE_SENSITIVE)) {
             op = URX_BUILD(URX_ONECHAR_I, lastCodePoint);
         } else {
             op = URX_BUILD(URX_ONECHAR, lastCodePoint);
         }
-        appendOp(op);
+        fRXPat->fCompiledPat->addElement(op, *fStatus);
     } else {
         // Two or more chars, emit a URX_STRING to match them.
         if (fModeFlags & UREGEX_CASE_INSENSITIVE) {
             op = URX_BUILD(URX_STRING_I, fRXPat->fLiteralText.length());
         } else {
             // TODO here:  add optimization to split case sensitive strings of length two
             //             into two single char ops, for efficiency.
             op = URX_BUILD(URX_STRING, fRXPat->fLiteralText.length());
         }
-        appendOp(op);
+        fRXPat->fCompiledPat->addElement(op, *fStatus);
         op = URX_BUILD(URX_STRING_LEN, fLiteralChars.length());
-        appendOp(op);
+        fRXPat->fCompiledPat->addElement(op, *fStatus);
 
         // Add this string into the accumulated strings of the compiled pattern.
-        // The total size of the accumulated strings must be restricted to 24 bits because
-        // string indexes appear as compiled pattern operand values.
-        // This is the only place that the pattern.fLiteralText string is modified.
-
         fRXPat->fLiteralText.append(fLiteralChars);
-        if (U_SUCCESS(*fStatus) && fRXPat->fLiteralText.length() > 0x00ffffff) {
-            *fStatus = U_REGEX_PATTERN_TOO_BIG;
-        }
     }
 
     fLiteralChars.remove();
 }
 
 
-//------------------------------------------------------------------------------
-//
-//   appendOp()             Append a new instruction onto the compiled pattern
-//                          Includes error checking, limiting the size of the
-//                          pattern to lengths that can be represented in the
-//                          24 bit operand field of an instruction.
-//
-//------------------------------------------------------------------------------
-void RegexCompile::appendOp(int32_t op) {
-    fRXPat->fCompiledPat->addElement(op, *fStatus);
-    if ((fRXPat->fCompiledPat->size() > 0x00fffff0) && U_SUCCESS(*fStatus)) {
-        *fStatus = U_REGEX_PATTERN_TOO_BIG;
-    }
-}
+
+
 
 
 //------------------------------------------------------------------------------
 //
 //   insertOp()             Insert a slot for a new opcode into the already
 //                          compiled pattern code.
 //
 //                          Fill the slot with a NOP.  Our caller will replace it
 //                          with what they really wanted.
 //
@@ skipping 41 matching lines @@
 
     if (fMatchCloseParen > where) {
         fMatchCloseParen++;
     }
     if (fMatchOpenParen > where) {
         fMatchOpenParen++;
     }
 }
 
 
+
 //------------------------------------------------------------------------------
 //
-//   allocateData()        Allocate storage in the matcher's static data area.
-//                         Return the index for the newly allocated data.
-//                         The storage won't actually exist until we are running a match
-//                         operation, but the storage indexes are inserted into various
-//                         opcodes while compiling the pattern.
-//
-//------------------------------------------------------------------------------
-int32_t RegexCompile::allocateData(int32_t size) {
-    if (U_FAILURE(*fStatus)) {
-        return 0;
-    }
-    if (size <= 0 || size > 0x100 || fRXPat->fDataSize < 0) {
-        *fStatus = U_REGEX_INTERNAL_ERROR;
-        return 0;
-    }
-    int32_t dataIndex = fRXPat->fDataSize;
-    fRXPat->fDataSize += size;
-    if (fRXPat->fDataSize >= 0x00fffff0) {
-        *fStatus = U_REGEX_PATTERN_TOO_BIG;
-    }
-    return dataIndex;
-}
-
-
-//------------------------------------------------------------------------------
-//
-//   allocateStackData()   Allocate space in the back-tracking stack frame.
-//                         Return the index for the newly allocated data.
-//                         The frame indexes are inserted into various
-//                         opcodes while compiling the pattern, meaning that frame
-//                         size must be restricted to the size that will fit
-//                         as an operand (24 bits).
-//
-//------------------------------------------------------------------------------
-int32_t RegexCompile::allocateStackData(int32_t size) {
-    if (U_FAILURE(*fStatus)) {
-        return 0;
-    }
-    if (size <= 0 || size > 0x100 || fRXPat->fFrameSize < 0) {
-        *fStatus = U_REGEX_INTERNAL_ERROR;
-        return 0;
-    }
-    int32_t dataIndex = fRXPat->fFrameSize;
-    fRXPat->fFrameSize += size;
-    if (fRXPat->fFrameSize >= 0x00fffff0) {
-        *fStatus = U_REGEX_PATTERN_TOO_BIG;
-    }
-    return dataIndex;
-}
-
-
-//------------------------------------------------------------------------------
-//
 //   blockTopLoc()          Find or create a location in the compiled pattern
 //                          at the start of the operation or block that has
 //                          just been compiled.  Needed when a quantifier (* or
 //                          whatever) appears, and we need to add an operation
 //                          at the start of the thing being quantified.
 //
 //                          (Parenthesized Blocks) have a slot with a NOP that
 //                          is reserved for this purpose.  .* or similar don't
 //                          and a slot needs to be added.
 //
@@ skipping 95 matching lines @@
         // Capturing Parentheses.
         //   Insert a End Capture op into the pattern.
         //   The frame offset of the variables for this cg is obtained from the
         //       start capture op and put it into the end-capture op.
         {
             int32_t   captureOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen+1);
             U_ASSERT(URX_TYPE(captureOp) == URX_START_CAPTURE);
 
             int32_t   frameVarLocation = URX_VAL(captureOp);
             int32_t   endCaptureOp = URX_BUILD(URX_END_CAPTURE, frameVarLocation);
-            appendOp(endCaptureOp);
+            fRXPat->fCompiledPat->addElement(endCaptureOp, *fStatus);
         }
         break;
     case atomic:
         // Atomic Parenthesis.
         //   Insert a LD_SP operation to restore the state stack to the position
         //   it was when the atomic parens were entered.
         {
             int32_t   stoOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen+1);
             U_ASSERT(URX_TYPE(stoOp) == URX_STO_SP);
             int32_t   stoLoc = URX_VAL(stoOp);
             int32_t   ldOp   = URX_BUILD(URX_LD_SP, stoLoc);
-            appendOp(ldOp);
+            fRXPat->fCompiledPat->addElement(ldOp, *fStatus);
         }
         break;
 
     case lookAhead:
         {
             int32_t  startOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen-5);
             U_ASSERT(URX_TYPE(startOp) == URX_LA_START);
             int32_t dataLoc  = URX_VAL(startOp);
             int32_t op       = URX_BUILD(URX_LA_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
         }
         break;
 
     case negLookAhead:
         {
             // See comment at doOpenLookAheadNeg
             int32_t  startOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen-1);
             U_ASSERT(URX_TYPE(startOp) == URX_LA_START);
             int32_t dataLoc  = URX_VAL(startOp);
             int32_t op       = URX_BUILD(URX_LA_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
             op               = URX_BUILD(URX_BACKTRACK, 0);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
             op               = URX_BUILD(URX_LA_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Patch the URX_SAVE near the top of the block.
             // The destination of the SAVE is the final LA_END that was just added.
             int32_t saveOp   = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen);
             U_ASSERT(URX_TYPE(saveOp) == URX_STATE_SAVE);
             int32_t dest     = fRXPat->fCompiledPat->size()-1;
             saveOp           = URX_BUILD(URX_STATE_SAVE, dest);
             fRXPat->fCompiledPat->setElementAt(saveOp, fMatchOpenParen);
         }
         break;
 
     case lookBehind:
         {
             // See comment at doOpenLookBehind.
 
             // Append the URX_LB_END and URX_LA_END to the compiled pattern.
             int32_t  startOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen-4);
             U_ASSERT(URX_TYPE(startOp) == URX_LB_START);
             int32_t dataLoc  = URX_VAL(startOp);
             int32_t op       = URX_BUILD(URX_LB_END, dataLoc);
-            appendOp(op);
-            op = URX_BUILD(URX_LA_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
+                    op       = URX_BUILD(URX_LA_END, dataLoc);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Determine the min and max bounds for the length of the
             //  string that the pattern can match.
             //  An unbounded upper limit is an error.
             int32_t patEnd   = fRXPat->fCompiledPat->size() - 1;
             int32_t minML    = minMatchLength(fMatchOpenParen, patEnd);
             int32_t maxML    = maxMatchLength(fMatchOpenParen, patEnd);
-            if (URX_TYPE(maxML) != 0) {
-                error(U_REGEX_LOOK_BEHIND_LIMIT);
-                break;
-            }
             if (maxML == INT32_MAX) {
                 error(U_REGEX_LOOK_BEHIND_LIMIT);
                 break;
             }
             U_ASSERT(minML <= maxML);
 
             // Insert the min and max match len bounds into the URX_LB_CONT op that
             //  appears at the top of the look-behind block, at location fMatchOpenParen+1
             fRXPat->fCompiledPat->setElementAt(minML,  fMatchOpenParen-2);
             fRXPat->fCompiledPat->setElementAt(maxML,  fMatchOpenParen-1);
 
         }
         break;
 
 
 
     case lookBehindN:
         {
             // See comment at doOpenLookBehindNeg.
 
             // Append the URX_LBN_END to the compiled pattern.
             int32_t  startOp = (int32_t)fRXPat->fCompiledPat->elementAti(fMatchOpenParen-5);
             U_ASSERT(URX_TYPE(startOp) == URX_LB_START);
             int32_t dataLoc  = URX_VAL(startOp);
             int32_t op       = URX_BUILD(URX_LBN_END, dataLoc);
-            appendOp(op);
+            fRXPat->fCompiledPat->addElement(op, *fStatus);
 
             // Determine the min and max bounds for the length of the
             //  string that the pattern can match.
             //  An unbounded upper limit is an error.
             int32_t patEnd   = fRXPat->fCompiledPat->size() - 1;
             int32_t minML    = minMatchLength(fMatchOpenParen, patEnd);
             int32_t maxML    = maxMatchLength(fMatchOpenParen, patEnd);
-            if (URX_TYPE(maxML) != 0) {
-                error(U_REGEX_LOOK_BEHIND_LIMIT);
-                break;
-            }
             if (maxML == INT32_MAX) {
                 error(U_REGEX_LOOK_BEHIND_LIMIT);
                 break;
             }
             U_ASSERT(minML <= maxML);
 
             // Insert the min and max match len bounds into the URX_LB_CONT op that
             //  appears at the top of the look-behind block, at location fMatchOpenParen+1
             fRXPat->fCompiledPat->setElementAt(minML,  fMatchOpenParen-3);
             fRXPat->fCompiledPat->setElementAt(maxML,  fMatchOpenParen-2);
@@ skipping 34 matching lines @@
     //  There shoudn't be any, but just in case.
     //     (Case Closure can add them; if we had a simple case closure avaialble that
     //      ignored strings, that would be better.)
     theSet->removeAllStrings();
     int32_t  setSize = theSet->size();
 
     switch (setSize) {
     case 0:
         {
             // Set of no elements.   Always fails to match.
-            appendOp(URX_BUILD(URX_BACKTRACK, 0));
+            fRXPat->fCompiledPat->addElement(URX_BUILD(URX_BACKTRACK, 0), *fStatus);
             delete theSet;
         }
         break;
 
     case 1:
         {
             // The set contains only a single code point.  Put it into
             //   the compiled pattern as a single char operation rather
             //   than a set, and discard the set itself.
             literalChar(theSet->charAt(0));
             delete theSet;
         }
         break;
 
     default:
         {
             //  The set contains two or more chars.  (the normal case)
             //  Put it into the compiled pattern as a set.
             int32_t setNumber = fRXPat->fSets->size();
             fRXPat->fSets->addElement(theSet, *fStatus);
             int32_t setOp = URX_BUILD(URX_SETREF, setNumber);
-            appendOp(setOp);
+            fRXPat->fCompiledPat->addElement(setOp, *fStatus);
         }
     }
 }
 
 
 //------------------------------------------------------------------------------
 //
 //   compileInterval    Generate the code for a {min, max} style interval quantifier.
 //                      Except for the specific opcodes used, the code is the same
 //                      for all three types (greedy, non-greedy, possessive) of
@@ skipping 18 matching lines @@
     int32_t   topOfBlock = blockTopLoc(TRUE);
     insertOp(topOfBlock);
     insertOp(topOfBlock);
     insertOp(topOfBlock);
 
     // The operands for the CTR_INIT opcode include the index in the matcher data
     //   of the counter.  Allocate it now. There are two data items
     //        counterLoc   -->  Loop counter
     //               +1    -->  Input index (for breaking non-progressing loops)
     //                          (Only present if unbounded upper limit on loop)
-    int32_t   dataSize = fIntervalUpper < 0 ? 2 : 1;
-    int32_t   counterLoc = allocateStackData(dataSize);
+    int32_t   counterLoc = fRXPat->fFrameSize;
+    fRXPat->fFrameSize++;
+    if (fIntervalUpper < 0) {
+        fRXPat->fFrameSize++;
+    }
 
     int32_t   op = URX_BUILD(InitOp, counterLoc);
     fRXPat->fCompiledPat->setElementAt(op, topOfBlock);
 
     // The second operand of CTR_INIT is the location following the end of the loop.
     //   Must put in as a URX_RELOC_OPRND so that the value will be adjusted if the
     //   compilation of something later on causes the code to grow and the target
     //   position to move.
     int32_t loopEnd = fRXPat->fCompiledPat->size();
     op = URX_BUILD(URX_RELOC_OPRND, loopEnd);
     fRXPat->fCompiledPat->setElementAt(op, topOfBlock+1);
 
     // Followed by the min and max counts.
     fRXPat->fCompiledPat->setElementAt(fIntervalLow, topOfBlock+2);
     fRXPat->fCompiledPat->setElementAt(fIntervalUpper, topOfBlock+3);
 
     // Apend the CTR_LOOP op.  The operand is the location of the CTR_INIT op.
     //   Goes at end of the block being looped over, so just append to the code so far.
     op = URX_BUILD(LoopOp, topOfBlock);
-    appendOp(op);
+    fRXPat->fCompiledPat->addElement(op, *fStatus);
 
     if ((fIntervalLow & 0xff000000) != 0 ||
         (fIntervalUpper > 0 && (fIntervalUpper & 0xff000000) != 0)) {
             error(U_REGEX_NUMBER_TOO_BIG);
         }
 
     if (fIntervalLow > fIntervalUpper && fIntervalUpper != -1) {
         error(U_REGEX_MAX_LT_MIN);
     }
 }
 
 
 
 UBool RegexCompile::compileInlineInterval() {
     if (fIntervalUpper > 10 || fIntervalUpper < fIntervalLow) {
         // Too big to inline.  Fail, which will cause looping code to be generated.
         //   (Upper < Lower picks up unbounded upper and errors, both.)
         return FALSE;
     }
 
     int32_t   topOfBlock = blockTopLoc(FALSE);
     if (fIntervalUpper == 0) {
         // Pathological case.  Attempt no matches, as if the block doesn't exist.
-        // Discard the generated code for the block.
-        // If the block included parens, discard the info pertaining to them as well.
         fRXPat->fCompiledPat->setSize(topOfBlock);
-        if (fMatchOpenParen >= topOfBlock) {
-            fMatchOpenParen = -1;
-        }
-        if (fMatchCloseParen >= topOfBlock) {
-            fMatchCloseParen = -1;
-        }
         return TRUE;
     }
 
     if (topOfBlock != fRXPat->fCompiledPat->size()-1 && fIntervalUpper != 1) {
         // The thing being repeated is not a single op, but some
         //   more complex block.  Do it as a loop, not inlines.
         //   Note that things "repeated" a max of once are handled as inline, because
         //     the one copy of the code already generated is just fine.
         return FALSE;
     }
@@ skipping 14 matching lines @@
     }
 
 
 
     //  Loop, emitting the op for the thing being repeated each time.
     //    Loop starts at 1 because one instance of the op already exists in the pattern,
     //    it was put there when it was originally encountered.
     int32_t i;
     for (i=1; i<fIntervalUpper; i++ ) {
         if (i == fIntervalLow) {
-            appendOp(saveOp);
+            fRXPat->fCompiledPat->addElement(saveOp, *fStatus);
         }
         if (i > fIntervalLow) {
-            appendOp(saveOp);
+            fRXPat->fCompiledPat->addElement(saveOp, *fStatus);
         }
-        appendOp(op);
+        fRXPat->fCompiledPat->addElement(op, *fStatus);
     }
     return TRUE;
 }
 
 
 
 //------------------------------------------------------------------------------
 //
+//   caseInsensitiveStart  given a single code point from a pattern string, determine the 
+//                         set of characters that could potentially begin a case-insensitive 
+//                         match of a string beginning with that character, using full Unicode
+//                         case insensitive matching.
+//
+//          This is used in optimizing find().
+//
+//          closeOver(USET_CASE_INSENSITIVE) does most of what is needed, but
+//          misses cases like this:
+//             A string from the pattern begins with 'ss' (although all we know
+//                 in this context is that it begins with 's')
+//             The pattern could match a string beginning with a German sharp-s
+//
+//           To the ordinary case closure for a character c, we add all other
+//           characters cx where the case closure of cx incudes a string form that begins
+//           with the original character c.
+//
+//           This function could be made smarter. The full pattern string is available
+//           and it would be possible to verify that the extra characters being added
+//           to the starting set fully match, rather than having just a first-char of the
+//           folded form match.
+//
+//------------------------------------------------------------------------------
+void  RegexCompile::findCaseInsensitiveStarters(UChar32 c, UnicodeSet *starterChars) {
+
+// Machine Generated below.
+// It may need updating with new versions of Unicode.
+// Intltest test RegexTest::TestCaseInsensitiveStarters will fail if an update is needed.
+// The update tool is here: svn+ssh://source.icu-project.org/repos/icu/tools/trunk/unicode/c/genregexcasing
+
+// Machine Generated Data. Do not hand edit.
+    static const UChar32 RECaseFixCodePoints[] = {
+        0x61, 0x66, 0x68, 0x69, 0x6a, 0x73, 0x74, 0x77, 0x79, 0x2bc, 
+        0x3ac, 0x3ae, 0x3b1, 0x3b7, 0x3b9, 0x3c1, 0x3c5, 0x3c9, 0x3ce, 0x565, 
+        0x574, 0x57e, 0x1f00, 0x1f01, 0x1f02, 0x1f03, 0x1f04, 0x1f05, 0x1f06, 0x1f07, 
+        0x1f20, 0x1f21, 0x1f22, 0x1f23, 0x1f24, 0x1f25, 0x1f26, 0x1f27, 0x1f60, 0x1f61, 
+        0x1f62, 0x1f63, 0x1f64, 0x1f65, 0x1f66, 0x1f67, 0x1f70, 0x1f74, 0x1f7c, 0x110000};
+
+    static const int16_t RECaseFixStringOffsets[] = {
+        0x0, 0x1, 0x6, 0x7, 0x8, 0x9, 0xd, 0xe, 0xf, 0x10, 
+        0x11, 0x12, 0x13, 0x17, 0x1b, 0x20, 0x21, 0x2a, 0x2e, 0x2f, 
+        0x30, 0x34, 0x35, 0x37, 0x39, 0x3b, 0x3d, 0x3f, 0x41, 0x43, 
+        0x45, 0x47, 0x49, 0x4b, 0x4d, 0x4f, 0x51, 0x53, 0x55, 0x57, 
+        0x59, 0x5b, 0x5d, 0x5f, 0x61, 0x63, 0x65, 0x66, 0x67, 0};
+
+    static const int16_t RECaseFixCounts[] = {
+        0x1, 0x5, 0x1, 0x1, 0x1, 0x4, 0x1, 0x1, 0x1, 0x1, 
+        0x1, 0x1, 0x4, 0x4, 0x5, 0x1, 0x9, 0x4, 0x1, 0x1, 
+        0x4, 0x1, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 
+        0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 
+        0x2, 0x2, 0x2, 0x2, 0x2, 0x2, 0x1, 0x1, 0x1, 0};
+
+    static const UChar RECaseFixData[] = {
+        0x1e9a, 0xfb00, 0xfb01, 0xfb02, 0xfb03, 0xfb04, 0x1e96, 0x130, 0x1f0, 0xdf, 
+        0x1e9e, 0xfb05, 0xfb06, 0x1e97, 0x1e98, 0x1e99, 0x149, 0x1fb4, 0x1fc4, 0x1fb3, 
+        0x1fb6, 0x1fb7, 0x1fbc, 0x1fc3, 0x1fc6, 0x1fc7, 0x1fcc, 0x390, 0x1fd2, 0x1fd3, 
+        0x1fd6, 0x1fd7, 0x1fe4, 0x3b0, 0x1f50, 0x1f52, 0x1f54, 0x1f56, 0x1fe2, 0x1fe3, 
+        0x1fe6, 0x1fe7, 0x1ff3, 0x1ff6, 0x1ff7, 0x1ffc, 0x1ff4, 0x587, 0xfb13, 0xfb14, 
+        0xfb15, 0xfb17, 0xfb16, 0x1f80, 0x1f88, 0x1f81, 0x1f89, 0x1f82, 0x1f8a, 0x1f83, 
+        0x1f8b, 0x1f84, 0x1f8c, 0x1f85, 0x1f8d, 0x1f86, 0x1f8e, 0x1f87, 0x1f8f, 0x1f90, 
+        0x1f98, 0x1f91, 0x1f99, 0x1f92, 0x1f9a, 0x1f93, 0x1f9b, 0x1f94, 0x1f9c, 0x1f95, 
+        0x1f9d, 0x1f96, 0x1f9e, 0x1f97, 0x1f9f, 0x1fa0, 0x1fa8, 0x1fa1, 0x1fa9, 0x1fa2, 
+        0x1faa, 0x1fa3, 0x1fab, 0x1fa4, 0x1fac, 0x1fa5, 0x1fad, 0x1fa6, 0x1fae, 0x1fa7, 
+        0x1faf, 0x1fb2, 0x1fc2, 0x1ff2, 0};
+
+// End of machine generated data.
+
+    if (u_hasBinaryProperty(c, UCHAR_CASE_SENSITIVE)) {
+        UChar32 caseFoldedC  = u_foldCase(c, U_FOLD_CASE_DEFAULT);
+        starterChars->set(caseFoldedC, caseFoldedC);
+
+        int32_t i;
+        for (i=0; RECaseFixCodePoints[i]<c ; i++) {
+            // Simple linear search through the sorted list of interesting code points.
+        }
+
+        if (RECaseFixCodePoints[i] == c) {
+            int32_t dataIndex = RECaseFixStringOffsets[i];
+            int32_t numCharsToAdd = RECaseFixCounts[i];
+            UChar32 cpToAdd = 0;
+            for (int32_t j=0; j<numCharsToAdd; j++) {
+                U16_NEXT_UNSAFE(RECaseFixData, dataIndex, cpToAdd);
+                starterChars->add(cpToAdd);
+            }
+        }
+
+        starterChars->closeOver(USET_CASE_INSENSITIVE);
+        starterChars->removeAllStrings();
+    } else {
+        // Not a cased character. Just return it alone.
+        starterChars->set(c, c);
+    }
+}
+
+
+
+
+//------------------------------------------------------------------------------
+//
 //   matchStartType    Determine how a match can start.
 //                     Used to optimize find() operations.
 //
 //                     Operation is very similar to minMatchLength().  Walk the compiled
 //                     pattern, keeping an on-going minimum-match-length.  For any
 //                     op where the min match coming in is zero, add that ops possible
 //                     starting matches to the possible starts for the overall pattern.
 //
 //------------------------------------------------------------------------------
 void   RegexCompile::matchStartType() {
@@ skipping 51 matching lines @@
         case URX_BACKSLASH_G:
         case URX_BACKSLASH_Z:
         case URX_DOLLAR:
         case URX_DOLLAR_M:
         case URX_DOLLAR_D:
         case URX_DOLLAR_MD:
         case URX_RELOC_OPRND:
         case URX_STO_INP_LOC:
         case URX_BACKREF:         // BackRef.  Must assume that it might be a zero length match
         case URX_BACKREF_I:
-                
+
         case URX_STO_SP:          // Setup for atomic or possessive blocks.  Doesn't change what can match.
         case URX_LD_SP:
             break;
 
         case URX_CARET:
             if (atStart) {
                 fRXPat->fStartType = START_START;
             }
             break;
 
@@ skipping 96 matching lines @@
             currentLen++;
             atStart = FALSE;
             break;
 
 
         case URX_ONECHAR_I:
             // Case Insensitive Single Character.
             if (currentLen == 0) {
                 UChar32  c = URX_VAL(op);
                 if (u_hasBinaryProperty(c, UCHAR_CASE_SENSITIVE)) {
-
-                    // Disable optimizations on first char of match.
-                    // TODO: Compute the set of chars that case fold to this char, or to
-                    //       a string that begins with this char.
-                    //       For simple case folding, this code worked:
-                    //   UnicodeSet s(c, c);
-                    //   s.closeOver(USET_CASE_INSENSITIVE);
-                    //   fRXPat->fInitialChars->addAll(s);
-
-                    fRXPat->fInitialChars->clear();
-                    fRXPat->fInitialChars->complement();
+                    UnicodeSet starters(c, c);
+                    starters.closeOver(USET_CASE_INSENSITIVE);
+                    // findCaseInsensitiveStarters(c, &starters);
+                    //   For ONECHAR_I, no need to worry about text chars that expand on folding into strings.
+                    //   The expanded folding can't match the pattern.
+                    fRXPat->fInitialChars->addAll(starters);
                 } else {
                     // Char has no case variants.  Just add it as-is to the
                     //   set of possible starting chars.
                     fRXPat->fInitialChars->add(c);
                 }
                 numInitialStrings += 2;
             }
             currentLen++;
             atStart = FALSE;
             break;
@@ skipping 102 matching lines @@
                 loc++;
                 int32_t stringLenOp = (int32_t)fRXPat->fCompiledPat->elementAti(loc);
                 int32_t stringLen   = URX_VAL(stringLenOp);
                 U_ASSERT(URX_TYPE(stringLenOp) == URX_STRING_LEN);
                 U_ASSERT(stringLenOp >= 2);
                 if (currentLen == 0) {
                     // Add the starting character of this string to the set of possible starting
                     //   characters for this pattern.
                     int32_t stringStartIdx = URX_VAL(op);
                     UChar32  c = fRXPat->fLiteralText.char32At(stringStartIdx);
-                    UnicodeSet s(c, c);
-
-                    // TODO:  compute correct set of starting chars for full case folding.
-                    //        For the moment, say any char can start.
-                    // s.closeOver(USET_CASE_INSENSITIVE);
-                    s.clear();
-                    s.complement();
-
+                    UnicodeSet s;
+                    findCaseInsensitiveStarters(c, &s);
                     fRXPat->fInitialChars->addAll(s);
                     numInitialStrings += 2;  // Matching on an initial string not possible.
                 }
                 currentLen += stringLen;
                 atStart = FALSE;
             }
             break;
 
         case URX_CTR_INIT:
         case URX_CTR_INIT_NG:
@@ skipping 35 matching lines @@
             //   don't change the minimum match
             atStart = FALSE;
             break;
 
 
         case URX_LA_START:
         case URX_LB_START:
             {
                 // Look-around.  Scan forward until the matching look-ahead end,
                 //   without processing the look-around block.  This is overly pessimistic.
-                
+
                 // Keep track of the nesting depth of look-around blocks.  Boilerplate code for
                 //   lookahead contains two LA_END instructions, so count goes up by two
                 //   for each LA_START.
                 int32_t  depth = (opType == URX_LA_START? 2: 1);
                 for (;;) {
                     loc++;
                     op = (int32_t)fRXPat->fCompiledPat->elementAti(loc);
                     if (URX_TYPE(op) == URX_LA_START) {
                         depth+=2;
                     }
@@ skipping 539 matching lines @@
 
         case URX_STRING_I:
             // TODO:  This code assumes that any user string that matches will be no longer
             //        than our compiled string, with case insensitive matching.
             //        Our compiled string has been case-folded already.
             //
             //        Any matching user string will have no more code points than our
             //        compiled (folded) string.  Folding may add code points, but
             //        not remove them.
             //
-            //        There is a potential problem if a supplemental code point 
+            //        There is a potential problem if a supplemental code point
             //        case-folds to a BMP code point.  In this case our compiled string
             //        could be shorter (in code units) than a matching user string.
             //
             //        At this time (Unicode 6.1) there are no such characters, and this case
             //        is not being handled.  A test, intltest regex/Bug9283, will fail if
             //        any problematic characters are added to Unicode.
             //
             //        If this happens, we can make a set of the BMP chars that the
             //        troublesome supplementals fold to, scan our string, and bump the
             //        currentLen one extra for each that is found.
@@ skipping 10 matching lines @@
             // For Loops, recursively call this function on the pattern for the loop body,
             //   then multiply the result by the maximum loop count.
             {
                 int32_t  loopEndLoc = URX_VAL(fRXPat->fCompiledPat->elementAti(loc+1));
                 if (loopEndLoc == loc+4) {
                     // Loop has an empty body. No affect on max match length.
                     // Continue processing with code after the loop end.
                     loc = loopEndLoc;
                     break;
                 }
-                
+
                 int32_t maxLoopCount = fRXPat->fCompiledPat->elementAti(loc+3);
                 if (maxLoopCount == -1) {
                     // Unbounded Loop. No upper bound on match length.
                     currentLen = INT32_MAX;
                     break;
                 }
 
                 U_ASSERT(loopEndLoc >= loc+4);
                 int32_t  blockLen = maxMatchLength(loc+4, loopEndLoc-1);  // Recursive call.
                 if (blockLen == INT32_MAX) {
@@ skipping 97 matching lines @@
     //   will be offset at each location in the original code.
     int32_t   loc;
     int32_t   d = 0;
     for (loc=0; loc<end; loc++) {
         deltas.addElement(d, *fStatus);
         int32_t op = (int32_t)fRXPat->fCompiledPat->elementAti(loc);
         if (URX_TYPE(op) == URX_NOP) {
             d++;
         }
     }
-    
+
     UnicodeString caseStringBuffer;
 
     // Make a second pass over the code, removing the NOPs by moving following
     //  code up, and patching operands that refer to code locations that
     //  are being moved.  The array of offsets from the first step is used
     //  to compute the new operand values.
     int32_t src;
     int32_t dst = 0;
     for (src=0; src<end; src++) {
         int32_t op = (int32_t)fRXPat->fCompiledPat->elementAti(src);
@@ skipping 26 matching lines @@
             {
                 int32_t where = URX_VAL(op);
                 if (where > fRXPat->fGroupMap->size()) {
                     error(U_REGEX_INVALID_BACK_REF);
                     break;
                 }
                 where = fRXPat->fGroupMap->elementAti(where-1);
                 op    = URX_BUILD(opType, where);
                 fRXPat->fCompiledPat->setElementAt(op, dst);
                 dst++;
-                
+
                 fRXPat->fNeedsAltInput = TRUE;
                 break;
             }
         case URX_RESERVED_OP:
         case URX_RESERVED_OP_N:
         case URX_BACKTRACK:
         case URX_END:
         case URX_ONECHAR:
         case URX_STRING:
         case URX_STRING_LEN:
@@ skipping 70 matching lines @@
         if (fLineNum > 0x7FFFFFFF) {
             fParseErr->line   = 0;
             fParseErr->offset = -1;
         } else if (fCharNum > 0x7FFFFFFF) {
             fParseErr->line   = (int32_t)fLineNum;
             fParseErr->offset = -1;
         } else {
             fParseErr->line   = (int32_t)fLineNum;
             fParseErr->offset = (int32_t)fCharNum;
         }
-        
+
         UErrorCode status = U_ZERO_ERROR; // throwaway status for extracting context
 
         // Fill in the context.
         //   Note: extractBetween() pins supplied indicies to the string bounds.
         uprv_memset(fParseErr->preContext,  0, sizeof(fParseErr->preContext));
         uprv_memset(fParseErr->postContext, 0, sizeof(fParseErr->postContext));
         utext_extract(fRXPat->fPattern, fScanIndex-U_PARSE_CONTEXT_LEN+1, fScanIndex, fParseErr->preContext, U_PARSE_CONTEXT_LEN, &status);
         utext_extract(fRXPat->fPattern, fScanIndex, fScanIndex+U_PARSE_CONTEXT_LEN-1, fParseErr->postContext, U_PARSE_CONTEXT_LEN, &status);
     }
 }
@@ skipping 33 matching lines @@
 //
 //------------------------------------------------------------------------------
 UChar32  RegexCompile::nextCharLL() {
     UChar32       ch;
 
     if (fPeekChar != -1) {
         ch = fPeekChar;
         fPeekChar = -1;
         return ch;
     }
-    
+
     // assume we're already in the right place
     ch = UTEXT_NEXT32(fRXPat->fPattern);
     if (ch == U_SENTINEL) {
         return ch;
     }
 
     if (ch == chCR ||
         ch == chNEL ||
         ch == chLS   ||
         (ch == chLF && fLastChar != chCR)) {
@@ skipping 35 matching lines @@
 //
 //------------------------------------------------------------------------------
 void RegexCompile::nextChar(RegexPatternChar &c) {
 
     fScanIndex = UTEXT_GETNATIVEINDEX(fRXPat->fPattern);
     c.fChar    = nextCharLL();
     c.fQuoted  = FALSE;
 
     if (fQuoteMode) {
         c.fQuoted = TRUE;
-        if ((c.fChar==chBackSlash && peekCharLL()==chE && ((fModeFlags & UREGEX_LITERAL) == 0)) || 
+        if ((c.fChar==chBackSlash && peekCharLL()==chE && ((fModeFlags & UREGEX_LITERAL) == 0)) ||
             c.fChar == (UChar32)-1) {
             fQuoteMode = FALSE;  //  Exit quote mode,
             nextCharLL();        // discard the E
             nextChar(c);         // recurse to get the real next char
         }
     }
     else if (fInBackslashQuote) {
         // The current character immediately follows a '\'
         // Don't check for any further escapes, just return it as-is.
         // Don't set c.fQuoted, because that would prevent the state machine from
@@ skipping 40 matching lines @@
         if (c.fChar == chBackSlash) {
             int64_t pos = UTEXT_GETNATIVEINDEX(fRXPat->fPattern);
             if (RegexStaticSets::gStaticSets->fUnescapeCharSet.contains(peekCharLL())) {
                 //
                 // A '\' sequence that is handled by ICU's standard unescapeAt function.
                 //   Includes \uxxxx, \n, \r, many others.
                 //   Return the single equivalent character.
                 //
                 nextCharLL();                 // get & discard the peeked char.
                 c.fQuoted = TRUE;
-                
+
                 if (UTEXT_FULL_TEXT_IN_CHUNK(fRXPat->fPattern, fPatternLength)) {
                     int32_t endIndex = (int32_t)pos;
                     c.fChar = u_unescapeAt(uregex_ucstr_unescape_charAt, &endIndex, (int32_t)fPatternLength, (void *)fRXPat->fPattern->chunkContents);
-                    
+
                     if (endIndex == pos) {
                         error(U_REGEX_BAD_ESCAPE_SEQUENCE);
                     }
                     fCharNum += endIndex - pos;
                     UTEXT_SETNATIVEINDEX(fRXPat->fPattern, endIndex);
                 } else {
                     int32_t offset = 0;
                     struct URegexUTextUnescapeCharContext context = U_REGEX_UTEXT_UNESCAPE_CONTEXT(fRXPat->fPattern);
-                    
+
                     UTEXT_SETNATIVEINDEX(fRXPat->fPattern, pos);
                     c.fChar = u_unescapeAt(uregex_utext_unescape_charAt, &offset, INT32_MAX, &context);
 
                     if (offset == 0) {
                         error(U_REGEX_BAD_ESCAPE_SEQUENCE);
                     } else if (context.lastOffset == offset) {
                         UTEXT_PREVIOUS32(fRXPat->fPattern);
                     } else if (context.lastOffset != offset-1) {
                         utext_moveIndex32(fRXPat->fPattern, offset - context.lastOffset - 1);
                     }
@@ skipping 22 matching lines @@
                     }
                     c.fChar <<= 3;
                     c.fChar += ch&7;
                     if (c.fChar <= 255) {
                         nextCharLL();
                     } else {
                         // The last digit made the number too big.  Forget we saw it.
                         c.fChar >>= 3;
                     }
                 }
-                c.fQuoted = TRUE; 
-            } 
+                c.fQuoted = TRUE;
+            }
             else if (peekCharLL() == chQ) {
                 //  "\Q"  enter quote mode, which will continue until "\E"
                 fQuoteMode = TRUE;
                 nextCharLL();       // discard the 'Q'.
                 nextChar(c);        // recurse to get the real next char.
             }
             else
             {
                 // We are in a '\' escape that will be handled by the state table scanner.
                 // Just return the backslash, but remember that the following char is to
@@ skipping 27 matching lines @@
 UChar32  RegexCompile::scanNamedChar() {
     if (U_FAILURE(*fStatus)) {
         return 0;
     }
 
     nextChar(fC);
     if (fC.fChar != chLBrace) {
         error(U_REGEX_PROPERTY_SYNTAX);
         return 0;
     }
-    
+
     UnicodeString  charName;
     for (;;) {
         nextChar(fC);
         if (fC.fChar == chRBrace) {
             break;
         }
         if (fC.fChar == -1) {
             error(U_REGEX_PROPERTY_SYNTAX);
             return 0;
         }
         charName.append(fC.fChar);
     }
-    
+
     char name[100];
     if (!uprv_isInvariantUString(charName.getBuffer(), charName.length()) ||
          (uint32_t)charName.length()>=sizeof(name)) {
         // All Unicode character names have only invariant characters.
         // The API to get a character, given a name, accepts only char *, forcing us to convert,
         //   which requires this error check
         error(U_REGEX_PROPERTY_SYNTAX);
         return 0;
     }
     charName.extract(0, charName.length(), name, sizeof(name), US_INV);
@@ skipping 18 matching lines @@
 //             Return a UnicodeSet, constructed from the \P pattern,
 //             or NULL if the pattern is invalid.
 //
 //------------------------------------------------------------------------------
 UnicodeSet *RegexCompile::scanProp() {
     UnicodeSet    *uset = NULL;
 
     if (U_FAILURE(*fStatus)) {
         return NULL;
     }
+    (void)chLowerP;   // Suppress compiler unused variable warning.
     U_ASSERT(fC.fChar == chLowerP || fC.fChar == chP);
     UBool negated = (fC.fChar == chP);
 
     UnicodeString propertyName;
     nextChar(fC);
     if (fC.fChar != chLBrace) {
         error(U_REGEX_PROPERTY_SYNTAX);
         return NULL;
     }
     for (;;) {
@@ skipping 49 matching lines @@
     UBool       savedInBackslashQuote = fInBackslashQuote;
     UBool       savedEOLComments      = fEOLComments;
     int64_t     savedLineNum          = fLineNum;
     int64_t     savedCharNum          = fCharNum;
     UChar32     savedLastChar         = fLastChar;
     UChar32     savedPeekChar         = fPeekChar;
     RegexPatternChar savedfC          = fC;
 
     // Scan for a closing ].   A little tricky because there are some perverse
     //   edge cases possible.  "[:abc\Qdef:] \E]"  is a valid non-property expression,
-    //   ending on the second closing ]. 
+    //   ending on the second closing ].
 
     UnicodeString propName;
     UBool         negated  = FALSE;
 
     // Check for and consume the '^' in a negated POSIX property, e.g.  [:^Letter:]
     nextChar(fC);
     if (fC.fChar == chUp) {
        negated = TRUE;
        nextChar(fC);
     }
-    
+
     // Scan for the closing ":]", collecting the property name along the way.
     UBool  sawPropSetTerminator = FALSE;
     for (;;) {
         propName.append(fC.fChar);
         nextChar(fC);
         if (fC.fQuoted || fC.fChar == -1) {
             // Escaped characters or end of input - either says this isn't a [:Property:]
             break;
         }
         if (fC.fChar == chColon) {
             nextChar(fC);
             if (fC.fChar == chRBracket) {
                 sawPropSetTerminator = TRUE;
             }
             break;
         }
     }
-    
+
     if (sawPropSetTerminator) {
         uset = createSetForProperty(propName, negated);
     }
     else
     {
         // No closing ":]".
         //  Restore the original scan position.
         //  The main scanner will retry the input as a normal set expression,
         //    not a [:Property:] expression.
         fScanIndex        = savedScanIndex;
@@ skipping 12 matching lines @@
 
 static inline void addIdentifierIgnorable(UnicodeSet *set, UErrorCode& ec) {
     set->add(0, 8).add(0x0e, 0x1b).add(0x7f, 0x9f);
     addCategory(set, U_GC_CF_MASK, ec);
 }
 
 //
 //  Create a Unicode Set from a Unicode Property expression.
 //     This is common code underlying both \p{...} ane [:...:] expressions.
 //     Includes trying the Java "properties" that aren't supported as
-//     normal ICU UnicodeSet properties 
+//     normal ICU UnicodeSet properties
 //
 static const UChar posSetPrefix[] = {0x5b, 0x5c, 0x70, 0x7b, 0}; // "[\p{"
 static const UChar negSetPrefix[] = {0x5b, 0x5c, 0x50, 0x7b, 0}; // "[\P{"
 UnicodeSet *RegexCompile::createSetForProperty(const UnicodeString &propName, UBool negated) {
     UnicodeString   setExpr;
     UnicodeSet      *set;
     uint32_t        usetFlags = 0;
-    
+
     if (U_FAILURE(*fStatus)) {
         return NULL;
     }
 
     //
     //  First try the property as we received it
     //
     if (negated) {
         setExpr.append(negSetPrefix, -1);
     } else {
         setExpr.append(posSetPrefix, -1);
     }
     setExpr.append(propName);
     setExpr.append(chRBrace);
     setExpr.append(chRBracket);
     if (fModeFlags & UREGEX_CASE_INSENSITIVE) {
         usetFlags |= USET_CASE_INSENSITIVE;
     }
     set = new UnicodeSet(setExpr, usetFlags, NULL, *fStatus);
     if (U_SUCCESS(*fStatus)) {
        return set;
     }
     delete set;
     set = NULL;
-    
+
     //
     //  The property as it was didn't work.
 
-    //  Do [:word:]. It is not recognized as a property by UnicodeSet.  "word" not standard POSIX 
+    //  Do [:word:]. It is not recognized as a property by UnicodeSet.  "word" not standard POSIX
     //     or standard Java, but many other regular expression packages do recognize it.
-    
+
     if (propName.caseCompare(UNICODE_STRING_SIMPLE("word"), 0) == 0) {
         *fStatus = U_ZERO_ERROR;
         set = new UnicodeSet(*(fRXPat->fStaticSets[URX_ISWORD_SET]));
         if (set == NULL) {
             *fStatus = U_MEMORY_ALLOCATION_ERROR;
             return set;
         }
         if (negated) {
             set->complement();
         }
         return set;
     }
 
 
     //    Do Java fixes -
     //       InGreek -> InGreek or Coptic, that being the official Unicode name for that block.
     //       InCombiningMarksforSymbols -> InCombiningDiacriticalMarksforSymbols.
     //
     //       Note on Spaces:  either "InCombiningMarksForSymbols" or "InCombining Marks for Symbols"
-    //                        is accepted by Java.  The property part of the name is compared 
+    //                        is accepted by Java.  The property part of the name is compared
     //                        case-insenstively.  The spaces must be exactly as shown, either
     //                        all there, or all omitted, with exactly one at each position
     //                        if they are present.  From checking against JDK 1.6
     //
     //       This code should be removed when ICU properties support the Java  compatibility names
     //          (ICU 4.0?)
     //
     UnicodeString mPropName = propName;
     if (mPropName.caseCompare(UNICODE_STRING_SIMPLE("InGreek"), 0) == 0) {
         mPropName = UNICODE_STRING_SIMPLE("InGreek and Coptic");
     }
     if (mPropName.caseCompare(UNICODE_STRING_SIMPLE("InCombining Marks for Symbols"), 0) == 0 ||
         mPropName.caseCompare(UNICODE_STRING_SIMPLE("InCombiningMarksforSymbols"), 0) == 0) {
         mPropName = UNICODE_STRING_SIMPLE("InCombining Diacritical Marks for Symbols");
     }
     else if (mPropName.compare(UNICODE_STRING_SIMPLE("all")) == 0) {
         mPropName = UNICODE_STRING_SIMPLE("javaValidCodePoint");
     }
-    
+
     //    See if the property looks like a Java "InBlockName", which
     //    we will recast as "Block=BlockName"
     //
     static const UChar IN[] = {0x49, 0x6E, 0};  // "In"
     static const UChar BLOCK[] = {0x42, 0x6C, 0x6f, 0x63, 0x6b, 0x3d, 00};  // "Block="
     if (mPropName.startsWith(IN, 2) && propName.length()>=3) {
         setExpr.truncate(4);   // Leaves "[\p{", or "[\P{"
         setExpr.append(BLOCK, -1);
         setExpr.append(UnicodeString(mPropName, 2));  // Property with the leading "In" removed.
         setExpr.append(chRBrace);
@@ skipping 103 matching lines @@
             }
             if (negated) {
                 set->complement();
             }
             return set;
         }
         delete set;
         set = NULL;
     }
     error(*fStatus);
-    return NULL; 
+    return NULL;
 }
 
 
 
 //
 //  SetEval   Part of the evaluation of [set expressions].
 //            Perform any pending (stacked) operations with precedence
 //            equal or greater to that of the next operator encountered
 //            in the expression.
 //
@@ skipping 47 matching lines @@
 
 void RegexCompile::setPushOp(int32_t op) {
     setEval(op);
     fSetOpStack.push(op, *fStatus);
     fSetStack.push(new UnicodeSet(), *fStatus);
 }
 
 U_NAMESPACE_END
 #endif  // !UCONFIG_NO_REGULAR_EXPRESSIONS