Fixed atomic group backtracking bug.

ChangeLog

@@ -112,6 +112,12 @@ such as the 04 in 10.04 incorrectly and hence giving wrong results.
 scripts and characters, this involved re-jigging the grapheme break property
 algorithm because Unicode has changed the way emojis are handled.
 
+25. Fixed an obscure bug that struck when there were two atomic groups not 
+separated by something with a backtracking point. There could be an incorrect 
+backtrack into the first of the atomic groups. A complicated example is
+/(?>a(*:1))(?>b)(*SKIP:1)x|.*/ matched against "abc", where the *SKIP 
+shouldn't find a MARK (because is in an atomic group), but it did.
+      
 
 Version 10.31 12-February-2018
 ------------------------------

src/pcre2_match.c

@@ -5509,7 +5509,7 @@ fprintf(stderr, "++ op=%d\n", *Fecode);
       frame so that it points to the final branch. */
 
       case OP_ONCE:
-      Fback_frame = ((char *)F - (char *)P) + frame_size;
+      Fback_frame = ((char *)F - (char *)P);
       for (;;)
         {
         uint32_t y = GET(P->ecode,1);

testdata/testinput1

@@ -6203,4 +6203,10 @@ ef) x/x,mark
 /a(?:(*:X))(*SKIP:X)(*F)|(.)/
     abc
 
+/(?>a(*:1))(?>b(*:1))(*SKIP:1)x|.*/no_start_optimize
+    abc
+
+/(?>a(*:1))(?>b)(*SKIP:1)x|.*/no_start_optimize
+    abc
+
 # End of testinput1 

testdata/testoutput1

@@ -9846,4 +9846,12 @@ No match
  0: b
  1: b
 
+/(?>a(*:1))(?>b(*:1))(*SKIP:1)x|.*/no_start_optimize
+    abc
+ 0: abc
+
+/(?>a(*:1))(?>b)(*SKIP:1)x|.*/no_start_optimize
+    abc
+ 0: abc
+
 # End of testinput1