Fix lookbehind with zero-length branch in DFA matching. Fixes oss-fuzz 1959.

ChangeLog

@@ -171,6 +171,9 @@ one: PCRE2_EXTRA_ALLOW_SURROGATE_ESCAPES.
 
 34. Implement newline type PCRE2_NEWLINE_NUL.
 
+35. A lookbehind assertion that had a zero-length branch caused undefined 
+behaviour when processed by pcre2_dfa_match(). This is oss-fuzz issue 1859.
+
 
 Version 10.23 14-February-2017
 ------------------------------

src/pcre2_dfa_match.c

@@ -375,14 +375,10 @@ internal_dfa_match(
 {
 stateblock *active_states, *new_states, *temp_states;
 stateblock *next_active_state, *next_new_state;
-
 const uint8_t *ctypes, *lcc, *fcc;
 PCRE2_SPTR ptr;
 PCRE2_SPTR end_code;
-PCRE2_SPTR first_op;
-
 dfa_recursion_info new_recursive;
-
 int active_count, new_count, match_count;
 
 /* Some fields in the mb block are frequently referenced, so we load them into
@@ -417,21 +413,15 @@ active_states = (stateblock *)(workspace + 2);
 next_new_state = new_states = active_states + wscount;
 new_count = 0;
 
-first_op = this_start_code + 1 + LINK_SIZE +
-  ((*this_start_code == OP_CBRA || *this_start_code == OP_SCBRA ||
-    *this_start_code == OP_CBRAPOS || *this_start_code == OP_SCBRAPOS)
-    ? IMM2_SIZE:0);
-
 /* The first thing in any (sub) pattern is a bracket of some sort. Push all
 the alternative states onto the list, and find out where the end is. This
 makes is possible to use this function recursively, when we want to stop at a
 matching internal ket rather than at the end.
 
-If the first opcode in the first alternative is OP_REVERSE, we are dealing with
-a backward assertion. In that case, we have to find out the maximum amount to
-move back, and set up each alternative appropriately. */
+If we are dealing with a backward assertion we have to find out the maximum
+amount to move back, and set up each alternative appropriately. */
 
-if (*first_op == OP_REVERSE)
+if (*this_start_code == OP_ASSERTBACK || *this_start_code == OP_ASSERTBACK_NOT)
   {
   size_t max_back = 0;
   size_t gone_back;
@@ -476,15 +466,17 @@ if (*first_op == OP_REVERSE)
   if (current_subject < mb->start_used_ptr)
     mb->start_used_ptr = current_subject;
 
-  /* Now we can process the individual branches. */
+  /* Now we can process the individual branches. There will be an OP_REVERSE at
+  the start of each branch, except when the length of the branch is zero. */
 
   end_code = this_start_code;
   do
     {
-    size_t back = (size_t)GET(end_code, 2+LINK_SIZE);
+    uint32_t revlen = (end_code[1+LINK_SIZE] == OP_REVERSE)? 1 + LINK_SIZE : 0;
+    size_t back = (revlen == 0)? 0 : (size_t)GET(end_code, 2+LINK_SIZE);
     if (back <= gone_back)
       {
-      int bstate = (int)(end_code - start_code + 2 + 2*LINK_SIZE);
+      int bstate = (int)(end_code - start_code + 1 + LINK_SIZE + revlen);
       ADD_NEW_DATA(-bstate, 0, (int)(gone_back - back));
       }
     end_code += GET(end_code, 1);

testdata/testinput6

@@ -4929,4 +4929,16 @@
 \= Expect no match
     X\x00Y      
 
+/(?<=abc|)/
+    abcde\=aftertext
+    
+/(?<=|abc)/ 
+    abcde\=aftertext
+
+/(?<=abc|)/endanchored
+    abcde\=aftertext
+    
+/(?<=|abc)/endanchored
+    abcde\=aftertext
+
 # End of testinput6

testdata/testoutput6

@@ -7746,4 +7746,24 @@ No match
     X\x00Y      
 No match
 
+/(?<=abc|)/
+    abcde\=aftertext
+ 0: 
+ 0+ abcde
+    
+/(?<=|abc)/ 
+    abcde\=aftertext
+ 0: 
+ 0+ abcde
+
+/(?<=abc|)/endanchored
+    abcde\=aftertext
+ 0: 
+ 0+ 
+    
+/(?<=|abc)/endanchored
+    abcde\=aftertext
+ 0: 
+ 0+ 
+
 # End of testinput6