Title: [152881] trunk/Source/WTF
Revision
152881
Author
[email protected]
Date
2013-07-18 20:24:43 -0700 (Thu, 18 Jul 2013)

Log Message

Little cleaning of StringImpl::lower() and StringImpl::upper() for ARM
https://bugs.webkit.org/show_bug.cgi?id=118831

Reviewed by Gavin Barraclough.

Clean lower() and upper() before trying some optimizations:
-Prefix the code with the empty() case. Otherwise, each descending loop starts with
 a check for length == 0.
-Change ored to a 32 bits registers. ARM only has 32 bits registers and access to full word.
 To keep the value a UChar, the compiler is adding a bunch of useless value & 0xffff in the loops.
-Change the pointer based loops for index based loops. The compiler does a mighty job at optimizing
 those for each architecture. No need to make the code more complex.
-Don't perform the check for int32 unless we need to. The only code limited to int is Unicode::toLower.

* wtf/text/StringImpl.cpp:
(WTF::StringImpl::lower):
(WTF::StringImpl::upper):

Modified Paths

Diff

Modified: trunk/Source/WTF/ChangeLog (152880 => 152881)


--- trunk/Source/WTF/ChangeLog	2013-07-19 03:04:47 UTC (rev 152880)
+++ trunk/Source/WTF/ChangeLog	2013-07-19 03:24:43 UTC (rev 152881)
@@ -1,3 +1,23 @@
+2013-07-18  Benjamin Poulain  <[email protected]>
+
+        Little cleaning of StringImpl::lower() and StringImpl::upper() for ARM
+        https://bugs.webkit.org/show_bug.cgi?id=118831
+
+        Reviewed by Gavin Barraclough.
+
+        Clean lower() and upper() before trying some optimizations:
+        -Prefix the code with the empty() case. Otherwise, each descending loop starts with
+         a check for length == 0.
+        -Change ored to a 32 bits registers. ARM only has 32 bits registers and access to full word.
+         To keep the value a UChar, the compiler is adding a bunch of useless value & 0xffff in the loops.
+        -Change the pointer based loops for index based loops. The compiler does a mighty job at optimizing
+         those for each architecture. No need to make the code more complex.
+        -Don't perform the check for int32 unless we need to. The only code limited to int is Unicode::toLower.
+
+        * wtf/text/StringImpl.cpp:
+        (WTF::StringImpl::lower):
+        (WTF::StringImpl::upper):
+
 2013-07-18  Andreas Kling  <[email protected]>
 
         CodeBlock::m_argumentValueProfiles wastes a lot of memory.

Modified: trunk/Source/WTF/wtf/text/StringImpl.cpp (152880 => 152881)


--- trunk/Source/WTF/wtf/text/StringImpl.cpp	2013-07-19 03:04:47 UTC (rev 152880)
+++ trunk/Source/WTF/wtf/text/StringImpl.cpp	2013-07-19 03:24:43 UTC (rev 152881)
@@ -389,67 +389,63 @@
 {
     // Note: This is a hot function in the Dromaeo benchmark, specifically the
     // no-op code path up through the first 'return' statement.
-    
+
     // First scan the string for uppercase and non-ASCII characters:
     bool noUpper = true;
-    UChar ored = 0;
+    unsigned ored = 0;
     if (is8Bit()) {
-        const LChar* end = m_data8 + m_length;
-        for (const LChar* chp = m_data8; chp != end; ++chp) {
-            if (UNLIKELY(isASCIIUpper(*chp)))
+        for (unsigned i = 0; i < m_length; ++i) {
+            LChar character = m_data8[i];
+            if (UNLIKELYisASCIIUpper(character)))
                 noUpper = false;
-            ored |= *chp;
+            ored |= character;
         }
         // Nothing to do if the string is all ASCII with no uppercase.
         if (noUpper && !(ored & ~0x7F))
             return this;
 
-        if (m_length > static_cast<unsigned>(numeric_limits<int32_t>::max()))
-            CRASH();
-        int32_t length = m_length;
-
         LChar* data8;
-        RefPtr<StringImpl> newImpl = createUninitialized(length, data8);
+        RefPtr<StringImpl> newImpl = createUninitialized(m_length, data8);
 
         if (!(ored & ~0x7F)) {
-            for (int32_t i = 0; i < length; ++i)
+            for (unsigned i = 0; i < m_length; ++i)
                 data8[i] = toASCIILower(m_data8[i]);
 
             return newImpl.release();
         }
 
         // Do a slower implementation for cases that include non-ASCII Latin-1 characters.
-        for (int32_t i = 0; i < length; ++i)
+        for (unsigned i = 0; i < m_length; ++i)
             data8[i] = static_cast<LChar>(Unicode::toLower(m_data8[i]));
 
         return newImpl.release();
     }
 
-    const UChar *end = m_data16 + m_length;
-    for (const UChar* chp = m_data16; chp != end; ++chp) {
-        if (UNLIKELY(isASCIIUpper(*chp)))
+    for (unsigned i = 0; i < m_length; ++i) {
+        UChar character = m_data16[i];
+        if (UNLIKELY(isASCIIUpper(character)))
             noUpper = false;
-        ored |= *chp;
+        ored |= character;
     }
     // Nothing to do if the string is all ASCII with no uppercase.
     if (noUpper && !(ored & ~0x7F))
         return this;
 
-    if (m_length > static_cast<unsigned>(numeric_limits<int32_t>::max()))
-        CRASH();
-    int32_t length = m_length;
-
     if (!(ored & ~0x7F)) {
         UChar* data16;
         RefPtr<StringImpl> newImpl = createUninitialized(m_length, data16);
         
-        for (int32_t i = 0; i < length; ++i) {
+        for (unsigned i = 0; i < m_length; ++i) {
             UChar c = m_data16[i];
             data16[i] = toASCIILower(c);
         }
         return newImpl.release();
     }
-    
+
+    if (m_length > static_cast<unsigned>(numeric_limits<int32_t>::max()))
+        CRASH();
+    int32_t length = m_length;
+
     // Do a slower implementation for cases that include non-ASCII characters.
     UChar* data16;
     RefPtr<StringImpl> newImpl = createUninitialized(m_length, data16);
@@ -481,7 +477,7 @@
         RefPtr<StringImpl> newImpl = createUninitialized(m_length, data8);
         
         // Do a faster loop for the case where all the characters are ASCII.
-        LChar ored = 0;
+        unsigned ored = 0;
         for (int i = 0; i < length; ++i) {
             LChar c = m_data8[i];
             ored |= c;
@@ -542,7 +538,7 @@
     RefPtr<StringImpl> newImpl = createUninitialized(m_length, data16);
     
     // Do a faster loop for the case where all the characters are ASCII.
-    UChar ored = 0;
+    unsigned ored = 0;
     for (int i = 0; i < length; ++i) {
         UChar c = source16[i];
         ored |= c;
_______________________________________________
webkit-changes mailing list
[email protected]
https://lists.webkit.org/mailman/listinfo/webkit-changes

Reply via email to