Changeset 60267 in webkit

Timestamp:

May 26, 2010 8:44:28 PM (14 years ago)

Author:

barraclough@apple.com

Message:

Bug 39795 - Add support for YARR JIT generation of greedy quantified parens at the end of the main disjunction.

Reviewed by Oliver Hunt.

If the last item in a main disjunction is a quantified set of parentheses,
this is easier to code generate for than the general case for quantified
parentheses. This is because we never need to backtrack into the parentheses

• the first match will be the final and accepted match.

This patch also somewhat reverts a recent change to when fallback to PCRE
occurs. At the minute the compiler is tracking on patterns which will
require JIT fallback. This is handy from a performance perspective (it saves
the failed attempt at JIT compilation), but it means introducing knowledge
of the JITs capabilities into the other layers of the regex compilers. For
the specific feature of back-references, add a flag tracking their presence
on the pattern, and make these expressions fallback without attempting to
JIT. For parentheses, return to detecting which cases are have or have not
been handled during JIT compilation.

18% progression on tagcloud, ~1.5% overall on sunspidey.

• yarr/RegexCompiler.cpp:

(JSC::Yarr::RegexPatternConstructor::atomBackReference):
(JSC::Yarr::RegexPatternConstructor::quantifyAtom):

• yarr/RegexJIT.cpp:

(JSC::Yarr::RegexGenerator::TermGenerationState::isLastTerm):
(JSC::Yarr::RegexGenerator::TermGenerationState::isMainDisjunction):
(JSC::Yarr::RegexGenerator::generateParenthesesGreedyNoBacktrack):
(JSC::Yarr::RegexGenerator::generateTerm):
(JSC::Yarr::RegexGenerator::RegexGenerator):
(JSC::Yarr::RegexGenerator::shouldFallBack):
(JSC::Yarr::jitCompileRegex):

• yarr/RegexPattern.h:

(JSC::Yarr::RegexPattern::RegexPattern):
(JSC::Yarr::RegexPattern::reset):

Location:

trunk/JavaScriptCore

Files:

• ChangeLog (modified) (1 diff)
• yarr/RegexCompiler.cpp (modified) (2 diffs)
• yarr/RegexJIT.cpp (modified) (8 diffs)
• yarr/RegexPattern.h (modified) (3 diffs)

trunk/JavaScriptCore/ChangeLog

@@ +1 @@
+2010-05-26  Gavin Barraclough  <barraclough@apple.com>
+
+        Reviewed by Oliver Hunt.
+
+        Bug 39795 - Add support for YARR JIT generation of greedy quantified parens at the end of the main disjunction.
+
+        If the last item in a main disjunction is a quantified set of parentheses,
+        this is easier to code generate for than the general case for quantified
+        parentheses. This is because we never need to backtrack into the parentheses
+        - the first match will be the final and accepted match.
+
+        This patch also somewhat reverts a recent change to when fallback to PCRE
+        occurs. At the minute the compiler is tracking on patterns which will
+        require JIT fallback. This is handy from a performance perspective (it saves
+        the failed attempt at JIT compilation), but it means introducing knowledge
+        of the JITs capabilities into the other layers of the regex compilers. For
+        the specific feature of back-references, add a flag tracking their presence
+        on the pattern, and make these expressions fallback without attempting to
+        JIT. For parentheses, return to detecting which cases are have or have not
+        been handled during JIT compilation.
+
+        18% progression on tagcloud, ~1.5% overall on sunspidey.
+
+        * yarr/RegexCompiler.cpp:
+        (JSC::Yarr::RegexPatternConstructor::atomBackReference):
+        (JSC::Yarr::RegexPatternConstructor::quantifyAtom):
+        * yarr/RegexJIT.cpp:
+        (JSC::Yarr::RegexGenerator::TermGenerationState::isLastTerm):
+        (JSC::Yarr::RegexGenerator::TermGenerationState::isMainDisjunction):
+        (JSC::Yarr::RegexGenerator::generateParenthesesGreedyNoBacktrack):
+        (JSC::Yarr::RegexGenerator::generateTerm):
+        (JSC::Yarr::RegexGenerator::RegexGenerator):
+        (JSC::Yarr::RegexGenerator::shouldFallBack):
+        (JSC::Yarr::jitCompileRegex):
+        * yarr/RegexPattern.h:
+        (JSC::Yarr::RegexPattern::RegexPattern):
+        (JSC::Yarr::RegexPattern::reset):
+
 2010-05-26  Geoffrey Garen  <ggaren@apple.com>
 

trunk/JavaScriptCore/yarr/RegexCompiler.cpp

@@ -373 +373 @@
     {
         ASSERT(subpatternId);
-        m_pattern.m_shouldFallBack = true;
+        m_pattern.m_containsBackreferences = true;
         m_pattern.m_maxBackReference = std::max(m_pattern.m_maxBackReference, subpatternId);
 
@@ -448 +448 @@
             return;
         }
-
-        if (max > 1 && term.type == PatternTerm::TypeParenthesesSubpattern)
-            m_pattern.m_shouldFallBack = true;
 
         if (min == 0)

trunk/JavaScriptCore/yarr/RegexJIT.cpp

@@ -346 +346 @@
             return alternative()->m_terms[t];
         }
+        bool isLastTerm()
+        {
+            ASSERT(alternativeValid());
+            return (t + 1) == alternative()->m_terms.size();
+        }
+        bool isMainDisjunction()
+        {
+            return !disjunction->m_parent;
+        }
 
         PatternTerm& lookaheadTerm()
@@ -902 +911 @@
         PatternDisjunction* disjunction = term.parentheses.disjunction;
         ASSERT(term.quantityCount == 1);
+
+        if (!term.parentheses.isCopy) {
+            m_shouldFallBack = true;
+            return;
+        }
 
         unsigned preCheckedCount = ((term.quantityCount == 1) && (term.quantityType == QuantifierFixedCount)) ? disjunction->m_minimumSize : 0;
@@ -988 +1002 @@
             success.link(this);
         }
+    }
+
+    void generateParenthesesGreedyNoBacktrack(TermGenerationState& state)
+    {
+        PatternTerm& parenthesesTerm = state.term();
+        PatternDisjunction* disjunction = parenthesesTerm.parentheses.disjunction;
+        ASSERT(parenthesesTerm.type == PatternTerm::TypeParenthesesSubpattern);
+        ASSERT(parenthesesTerm.quantityCount != 1); // Handled by generateParenthesesSingle.
+
+        // Capturing not yet implemented!
+        if (parenthesesTerm.invertOrCapture) {
+            m_shouldFallBack = true;
+            return;
+        }
+
+        // Quantification limit not yet implemented!
+        if (parenthesesTerm.quantityCount != 0xffffffff) {
+            m_shouldFallBack = true;
+            return;
+        }
+
+        TermGenerationState parenthesesState(disjunction, state.checkedTotal);
+
+        Label matchAgain(this);
+        for (parenthesesState.resetAlternative(); parenthesesState.alternativeValid(); parenthesesState.nextAlternative()) {
+
+            PatternAlternative* alternative = parenthesesState.alternative();
+            optimizeAlternative(alternative);
+
+            int countToCheck = alternative->m_minimumSize;
+            if (countToCheck) {
+                parenthesesState.addBacktrackJump(jumpIfNoAvailableInput(countToCheck));
+                parenthesesState.checkedTotal += countToCheck;
+            }
+
+            for (parenthesesState.resetTerm(); parenthesesState.termValid(); parenthesesState.nextTerm())
+                generateTerm(parenthesesState);
+
+            // If we get here, we matched! Limit not yet supported, so just try to match more!
+            jump(matchAgain);
+            
+            parenthesesState.linkAlternativeBacktracks(this);
+            // We get here if the alternative fails to match - fall through to the next iteration, or out of the loop.
+
+            if (countToCheck) {
+                sub32(Imm32(countToCheck), index);
+                parenthesesState.checkedTotal -= countToCheck;
+            }
+        }
+
+        // If the last alternative falls through to here, we have a failed match...
+        // Which means that we match whatever we have matched up to this point (even if nothing).
     }
 
@@ -1101 +1167 @@
 
         case PatternTerm::TypeBackReference:
-            ASSERT_NOT_REACHED();
+            m_shouldFallBack = true;
             break;
 
@@ -1108 +1174 @@
 
         case PatternTerm::TypeParenthesesSubpattern:
-            ASSERT((term.quantityCount == 1) && !term.parentheses.isCopy); // must fallback to pcre before this point
-            generateParenthesesSingle(state);
+            if (term.quantityCount == 1) {
+                generateParenthesesSingle(state);
+                break;
+            } else if (state.isLastTerm() && state.isMainDisjunction()) { // Is this is the last term of the main disjunction?
+                // If this has a greedy quantifier, then it will never need to backtrack!
+                if (term.quantityType == QuantifierGreedy) {
+                    generateParenthesesGreedyNoBacktrack(state);
+                    break;
+                }
+            }
+            m_shouldFallBack = true;
             break;
 
@@ -1362 +1437 @@
     RegexGenerator(RegexPattern& pattern)
         : m_pattern(pattern)
+        , m_shouldFallBack(false)
     {
     }
@@ -1391 +1467 @@
     }
 
+    bool shouldFallBack()
+    {
+        return m_shouldFallBack;
+    }
+
 private:
     RegexPattern& m_pattern;
+    bool m_shouldFallBack;
     Vector<AlternativeBacktrackRecord> m_backtrackRecords;
 };
@@ -1399 +1481 @@
 {
     RegexPattern pattern(ignoreCase, multiline);
-
     if ((error = compileRegex(patternString, pattern)))
         return;
-
     numSubpatterns = pattern.m_numSubpatterns;
 
-    if (pattern.m_shouldFallBack) {
-        JSRegExpIgnoreCaseOption ignoreCaseOption = ignoreCase ? JSRegExpIgnoreCase : JSRegExpDoNotIgnoreCase;
-        JSRegExpMultilineOption multilineOption = multiline ? JSRegExpMultiline : JSRegExpSingleLine;
-        jitObject.setFallback(jsRegExpCompile(reinterpret_cast<const UChar*>(patternString.data()), patternString.size(), ignoreCaseOption, multilineOption, &numSubpatterns, &error));
-    } else {
+    if (!pattern.m_containsBackreferences) {
         RegexGenerator generator(pattern);
         generator.compile(globalData, jitObject);
-    }
+        if (!generator.shouldFallBack())
+            return;
+    }
+
+    JSRegExpIgnoreCaseOption ignoreCaseOption = ignoreCase ? JSRegExpIgnoreCase : JSRegExpDoNotIgnoreCase;
+    JSRegExpMultilineOption multilineOption = multiline ? JSRegExpMultiline : JSRegExpSingleLine;
+    jitObject.setFallback(jsRegExpCompile(reinterpret_cast<const UChar*>(patternString.data()), patternString.size(), ignoreCaseOption, multilineOption, &numSubpatterns, &error));
 }
 

trunk/JavaScriptCore/yarr/RegexPattern.h

@@ -272 +272 @@
         , m_numSubpatterns(0)
         , m_maxBackReference(0)
-        , m_shouldFallBack(false)
+        , m_containsBackreferences(false)
         , newlineCached(0)
         , digitsCached(0)
@@ -294 +294 @@
         m_maxBackReference = 0;
 
-        m_shouldFallBack = false;
+        m_containsBackreferences = false;
 
         newlineCached = 0;
@@ -362 +362 @@
     unsigned m_numSubpatterns;
     unsigned m_maxBackReference;
-    bool m_shouldFallBack;
+    bool m_containsBackreferences;
     PatternDisjunction* m_body;
     Vector<PatternDisjunction*, 4> m_disjunctions;