Rewrite regexes where common prefix can be pulled out from alternatio…

…n branches

Thanks to Michael Voříšek for suggesting this optimization.

A new "rewrite" pass has been added to the regex compilation process.
For now, the rewrite pass only optimized one type of regex: those where
every branch of an alternation construct has a common prefix.

In such cases, we rewrite the regex like so (for example):

    (abc|abd|abe) ⇒ (ab(?:c|d|e))

An extra non-capturing group is not introduced if the alternation
is within a non-capturing group (which is not quantified using ?, *, or
a similar suffix). In that case we simply do something like:

    (?:abc|abd|abe) ⇒ ab(?:c|d|e)

In some edge cases, it is possible that rewriting a group with common
alternation prefix might open up the opportunity to pull out more common
prefixes. For example:

    (a(b|c)d|(ab|ac)e)

In that case, if the group '(ab|ac)' was rewritten to pull out the
common prefix, it would then become possible to pull out a common
prefix from the top-level group. However, we do not take advantage of
that opportunity.

Further, we do not perform the rewrite in cases where the prefixes are
semantically equivalent, but parse to a different parsed_pattern
sequence.

Groups which the regex engine might need to backtrack into are never
pulled out, since this could change the order in which the regex
engine considers possible ways of matching the pattern against the
subject string, and could thus change the returned match. For
example, this pattern will not be rewritten:

    ((?:a|b)c|(?:a|b)d)

Also, callouts are never extracted even if they form a common prefix
to an alternation. Some backtracking control verbs, like (*SKIP) and
(*COMMIT), are never extracted either.

A different type of rewrite is performed if an alternation construct
matches only single, literal characters:

    (a|b|c) ⇒ ([a-c])

A new compile option, PCRE2_NO_PATTERN_REWRITE, has been added to
skip the pattern rewrite phase when compiling a pattern.

RunTest

@@ -89,7 +89,9 @@ title23="Test 23: \C disabled test"
 title24="Test 24: Non-UTF pattern conversion tests"
 title25="Test 25: UTF pattern conversion tests"
 title26="Test 26: Auto-generated unicode property tests"
-maxtest=26
+title27="Test 27: Pattern rewriter tests without UTF"
+title28="Test 28: Pattern rewriter tests with UTF"
+maxtest=28
 titleheap="Test 'heap': Environment-specific heap tests"
 
 if [ $# -eq 1 -a "$1" = "list" ]; then
@@ -120,6 +122,8 @@ if [ $# -eq 1 -a "$1" = "list" ]; then
   echo $title24
   echo $title25
   echo $title26
+  echo $title27
+  echo $title28
   echo ""
   echo $titleheap
   echo ""
@@ -255,6 +259,8 @@ do23=no
 do24=no
 do25=no
 do26=no
+do27=no
+do28=no
 doheap=no
 
 while [ $# -gt 0 ] ; do
@@ -286,6 +292,8 @@ while [ $# -gt 0 ] ; do
    24) do24=yes;;
    25) do25=yes;;
    26) do26=yes;;
+   27) do27=yes;;
+   28) do28=yes;;
  heap) doheap=yes;;
    -8) arg8=yes;;
   -16) arg16=yes;;
@@ -437,7 +445,8 @@ if [ $do0  = no -a $do1  = no -a $do2  = no -a $do3  = no -a \
      $do12 = no -a $do13 = no -a $do14 = no -a $do15 = no -a \
      $do16 = no -a $do17 = no -a $do18 = no -a $do19 = no -a \
      $do20 = no -a $do21 = no -a $do22 = no -a $do23 = no -a \
-     $do24 = no -a $do25 = no -a $do26 = no -a $doheap = no \
+     $do24 = no -a $do25 = no -a $do26 = no -a $do27 = no -a \
+     $do28 = no -a $doheap = no \
    ]; then
   do0=yes
   do1=yes
@@ -466,6 +475,8 @@ if [ $do0  = no -a $do1  = no -a $do2  = no -a $do3  = no -a \
   do24=yes
   do25=yes
   do26=yes
+  do27=yes
+  do28=yes
 fi
 
 # Handle any explicit skips at this stage, so that an argument list may consist
@@ -898,6 +909,26 @@ for bmode in "$test8" "$test16" "$test32"; do
     fi
   fi
 
+  # Pattern rewriter tests (without UTF)
+
+  if [ $do27 = yes ] ; then
+    echo $title27
+    $sim $valgrind ./pcre2test -q $setstack $bmode $testdata/testinput27 testtry
+    checkresult $? 27 ""
+  fi
+
+  # Pattern rewriter tests (with UTF)
+
+  if [ $do28 = yes ] ; then
+    echo $title28
+    if [ $utf -eq 0 ] ; then
+      echo "  Skipped because UTF-$bits support is not available"
+    else
+      $sim $valgrind ./pcre2test -q $setstack $bmode $testdata/testinput28 testtry
+      checkresult $? 28 ""
+    fi
+  fi
+
   # Manually selected heap tests - output may vary in different environments,
   # which is why that are not automatically run.
 

doc/pcre2_compile.3

@@ -65,6 +65,7 @@ The primary option bits are:
                             theses (named ones available)
   PCRE2_NO_AUTO_POSSESS    Disable auto-possessification
   PCRE2_NO_DOTSTAR_ANCHOR  Disable automatic anchoring for .*
+  PCRE2_NO_PATTERN_REWRITE Disable pattern rewriting optimizations
   PCRE2_NO_START_OPTIMIZE  Disable match-time start optimizations
   PCRE2_NO_UTF_CHECK       Do not check the pattern for UTF validity
                              (only relevant if PCRE2_UTF is set)

doc/pcre2api.3

@@ -1768,6 +1768,18 @@ automatically anchored if PCRE2_DOTALL is set for all the .* items and
 PCRE2_MULTILINE is not set for any ^ items. Otherwise, the fact that any match
 must start either at the start of the subject or following a newline is
 remembered. Like other optimizations, this can cause callouts to be skipped.
+.sp
+  PCRE2_NO_PATTERN_REWRITE
+.sp
+This option disables all optimizations which occur during the pattern rewriting
+phase (after parsing but before compilation). Pattern rewriting may remove
+redundant items, coalesce items, adjust group structure, or replace some
+constructs with an equivalent construct. Pattern rewriting will never affect
+which strings are and are not matched, or what substrings are captured by
+capture groups. However, since it may change the structure of a pattern,
+if you are tracing the matching process, you might prefer PCRE2 to use the
+original pattern without rewriting. This option is also useful for testing.
+Pattern rewriting is also disabled if PCRE2_AUTO_CALLOUT is set.
 .sp
   PCRE2_NO_START_OPTIMIZE
 .sp

doc/pcre2callout.3

@@ -83,7 +83,9 @@ Callouts can be useful for tracking the progress of pattern matching. The
 program has a pattern qualifier (/auto_callout) that sets automatic callouts.
 When any callouts are present, the output from \fBpcre2test\fP indicates how
 the pattern is being matched. This is useful information when you are trying to
-optimize the performance of a particular pattern.
+optimize the performance of a particular pattern. However, note that some
+optimizations which adjust the structure of the pattern are disabled when
+automatic callouts are enabled.
 .
 .
 .SH "MISSING CALLOUTS"

doc/pcre2syntax.3

@@ -414,6 +414,7 @@ appear. For the first three, d is a decimal number.
   (*NO_AUTO_POSSESS) no auto-possessification (PCRE2_NO_AUTO_POSSESS)
   (*NO_DOTSTAR_ANCHOR) no .* anchoring (PCRE2_NO_DOTSTAR_ANCHOR)
   (*NO_JIT)       disable JIT optimization
+  (*NO_PATTERN_REWRITE) disable pattern rewriting optimizations (PCRE2_NO_PATTERN_REWRITE)
   (*NO_START_OPT) no start-match optimization (PCRE2_NO_START_OPTIMIZE)
   (*UTF)          set appropriate UTF mode for the library in use
   (*UCP)          set PCRE2_UCP (use Unicode properties for \ed etc)

doc/pcre2test.1

@@ -623,6 +623,7 @@ for a description of the effects of these options.
   /n  no_auto_capture           set PCRE2_NO_AUTO_CAPTURE
       no_auto_possess           set PCRE2_NO_AUTO_POSSESS
       no_dotstar_anchor         set PCRE2_NO_DOTSTAR_ANCHOR
+      no_pattern_rewrite        set PCRE2_NO_PATTERN_REWRITE
       no_start_optimize         set PCRE2_NO_START_OPTIMIZE
       no_utf_check              set PCRE2_NO_UTF_CHECK
       ucp                       set PCRE2_UCP

doc/pcre2test.txt

@@ -604,6 +604,7 @@ PATTERN MODIFIERS
          /n  no_auto_capture           set PCRE2_NO_AUTO_CAPTURE
              no_auto_possess           set PCRE2_NO_AUTO_POSSESS
              no_dotstar_anchor         set PCRE2_NO_DOTSTAR_ANCHOR
+             no_pattern_rewrite        set PCRE2_NO_PATTERN_REWRITE
              no_start_optimize         set PCRE2_NO_START_OPTIMIZE
              no_utf_check              set PCRE2_NO_UTF_CHECK
              ucp                       set PCRE2_UCP

src/pcre2.h.generic

@@ -143,6 +143,7 @@ D   is inspected during pcre2_dfa_match() execution
 #define PCRE2_EXTENDED_MORE       0x01000000u  /* C       */
 #define PCRE2_LITERAL             0x02000000u  /* C       */
 #define PCRE2_MATCH_INVALID_UTF   0x04000000u  /*   J M D */
+#define PCRE2_NO_PATTERN_REWRITE  0x08000000u  /* C       */
 
 /* An additional compile options word is available in the compile context. */
 

src/pcre2.h.in

@@ -143,6 +143,7 @@ D   is inspected during pcre2_dfa_match() execution
 #define PCRE2_EXTENDED_MORE       0x01000000u  /* C       */
 #define PCRE2_LITERAL             0x02000000u  /* C       */
 #define PCRE2_MATCH_INVALID_UTF   0x04000000u  /*   J M D */
+#define PCRE2_NO_PATTERN_REWRITE  0x08000000u  /* C       */
 
 /* An additional compile options word is available in the compile context. */