[JSC] Add ASCII comparison fast path for IntlCollator

• Component: New Bugs
• Status: RESOLVED FIXED

[Yusuke Suzuki]

[JSC] Add ASCII comparison fast path for IntlCollator

[Yusuke Suzuki]

Created attachment 407175 [details]
Patch

[Yusuke Suzuki]

Created attachment 407176 [details]
Patch

[Yusuke Suzuki]

Created attachment 407177 [details]
Patch

[Radar WebKit Bug Importer]

<rdar://problem/67725253>

[Saam Barati]

Comment on attachment 407177 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407177&action=review

> Source/JavaScriptCore/runtime/IntlCollator.cpp:294
> +        if (x.isAllASCII() && y.isAllASCII()) {

see comment right below, but I'd write this as:

if (canDoASCIIUCADUCETComparison() && x.isAllASCII() && y.isAllASCII())

> Source/JavaScriptCore/runtime/IntlCollator.cpp:295
> +            if (canDoASCIIUCADUCETComparison()) {

shouldn't be the first thing to check for the speed of not having to check the entire strings if they're ASCII when in a UCollator that can't go down the fast path?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:396
> +static bool canDoASCIIUCADUCETComparisonWithUCollator(UCollator& collator)

should some of this code live in WTF?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:409
> +    UErrorCode status = U_ZERO_ERROR;

should this go inside the loop?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:411
> +    int32_t index = 0;

remove

> Source/JavaScriptCore/runtime/IntlCollator.cpp:417
> +        index++;

remove

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:126
> +    auto compareCharacter = [](CharacterType1 lhs, CharacterType2 rhs) {

should we inline this below just to ensure that we don't actually have the "result != 0" branch in case the compiler doesn't optimize it away?

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:137
> +    while (true) {

why not just: "while (position < commonLength)"?

[Darin Adler]

Comment on attachment 407177 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407177&action=review

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:295
>> +            if (canDoASCIIUCADUCETComparison()) {
> 
> shouldn't be the first thing to check for the speed of not having to check the entire strings if they're ASCII when in a UCollator that can't go down the fast path?

No, this has the ucol_strcollUTF8 optimization too, not just compareASCIIWithUACDUCET.

> Source/JavaScriptCore/runtime/IntlCollator.cpp:297
> +                Vector<LChar> xBuffer;
> +                Vector<LChar> yBuffer;

If strings are often short, inline capacity could avoid two malloc/free operations. Also, could we limit the buffers to the shorter of the two lengths, or the shorter of the two lengths + 1? Maybe we don’t need to copy leading characters that are equal in both strings?

I guess a better proposal is to call compareASCIIWithUACDUCET on the characters with the widths they currently have rather than allocating memory to narrow them to 8-bit before calling.

    if (x.is8Bit() && y.is8Bit())
        return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters8(), x.length(), y.characters8(), y.length()));
    if (x.is8Bit())
        return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters8(), x.length(), y.characters16(), y.length()));
    if (y.is8Bit())
        return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters16(), x.length(), y.characters8(), y.length()));
    return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters16(), x.length(), y.characters16(), y.length()));

> Source/JavaScriptCore/runtime/IntlCollator.cpp:298
> +                auto getCharacters8 = [&] (const StringView& view, Vector<LChar>& buffer) -> const LChar* {

Can this use a return value instead of an out argument? Or will the vector operation be too expensive?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:301
> +                    buffer.resize(view.length());

This should be grow rather than resize.

> Source/JavaScriptCore/runtime/IntlCollator.cpp:449
> +    m_canDoASCIIUCADUCETComparison = result ? TriState::True : TriState::False;

m_canDoASCIIUCADUCETComparison = triState(result);

> Source/JavaScriptCore/runtime/IntlCollator.cpp:456
> +void IntlCollator::checkICULocaleInvariants(const HashSet<String>& locales)
> +{
> +    UNUSED_PARAM(locales);
> +#if ASSERT_ENABLED

Small style idea. Consider putting this in the header:

    #if !ASSERT_ENABLED
    inline void IntlCollator::checkICULocaleInvariants(const HashSet<String>&) { }
    #endif

Then here just write:

    #if ASSERT_ENABLED

    void IntlCollator::checkICULocaleInvariants(const HashSet<String>& locales)
    {
    }

    #endif

Without the UNUSED_PARAM.

> Source/JavaScriptCore/runtime/IntlCollator.cpp:470
> +                            dataLogLn("BAD ", locale, " ", makeString(hex(x)), "(", String(xstring, 1), ") <=> ", makeString(hex(y)), "(", String(ystring, 1), ") ICU:(", static_cast<int32_t>(resultICU), "),WTF:(", static_cast<int32_t>(resultWTF), ")");

These should be StringView, no need to create and  destroy a string each time.

[Ross Kirsling]

Comment on attachment 407177 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407177&action=review

> Source/JavaScriptCore/runtime/IntlCollator.cpp:293
> +    UCollationResult result = ([&]() -> UCollationResult {

How come this is an immediately-invoked lambda? Is there a benefit over using a block?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:456
> +void IntlCollator::checkICULocaleInvariants(const HashSet<String>& locales)
> +{
> +    UNUSED_PARAM(locales);
> +#if ASSERT_ENABLED

I can see how this was useful to figure out ICU's current behavior and how it would provide a way to immediately see if that changes, but it seems rather heavy to be treated as "just some assertions"?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:458
> +        ([&] {

Here too I'm not quite sure about the immediately-invoked lambda.

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:145
> +        ++characters1;
> +        ++characters2;
> +        ++position;

Is this better than just indexing on position?

[Yusuke Suzuki]

Comment on attachment 407177 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407177&action=review

Thanks!

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:293
>> +    UCollationResult result = ([&]() -> UCollationResult {
> 
> How come this is an immediately-invoked lambda? Is there a benefit over using a block?

This is for early returning. This makes code readable.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:294
>> +        if (x.isAllASCII() && y.isAllASCII()) {
> 
> see comment right below, but I'd write this as:
> 
> if (canDoASCIIUCADUCETComparison() && x.isAllASCII() && y.isAllASCII())

There is ucol_strcollUTF8 case. So we anyway need to perform `x.isAllASCII() && y.isAllASCII()` first.

>>> Source/JavaScriptCore/runtime/IntlCollator.cpp:295
>>> +            if (canDoASCIIUCADUCETComparison()) {
>> 
>> shouldn't be the first thing to check for the speed of not having to check the entire strings if they're ASCII when in a UCollator that can't go down the fast path?
> 
> No, this has the ucol_strcollUTF8 optimization too, not just compareASCIIWithUACDUCET.

Yeah, this is necessary since there is ucol_strcollUTF8 path.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:297
>> +                Vector<LChar> yBuffer;
> 
> If strings are often short, inline capacity could avoid two malloc/free operations. Also, could we limit the buffers to the shorter of the two lengths, or the shorter of the two lengths + 1? Maybe we don’t need to copy leading characters that are equal in both strings?
> 
> I guess a better proposal is to call compareASCIIWithUACDUCET on the characters with the widths they currently have rather than allocating memory to narrow them to 8-bit before calling.
> 
>     if (x.is8Bit() && y.is8Bit())
>         return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters8(), x.length(), y.characters8(), y.length()));
>     if (x.is8Bit())
>         return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters8(), x.length(), y.characters16(), y.length()));
>     if (y.is8Bit())
>         return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters16(), x.length(), y.characters8(), y.length()));
>     return static_cast<UCollationResult>(compareASCIIWithUACDUCET(m_collator.get(), x.characters16(), x.length(), y.characters16(), y.length()));

Sounds good. Changing.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:298
>> +                auto getCharacters8 = [&] (const StringView& view, Vector<LChar>& buffer) -> const LChar* {
> 
> Can this use a return value instead of an out argument? Or will the vector operation be too expensive?

This is because we do not want to allocate Vector if view is 8Bit. I'll specialize function instead of doing this here.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:301
>> +                    buffer.resize(view.length());
> 
> This should be grow rather than resize.

I'll change getCharacters16 side.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:396
>> +static bool canDoASCIIUCADUCETComparisonWithUCollator(UCollator& collator)
> 
> should some of this code live in WTF?

For now, I think putting this function in IntlCollator is better because the following options are derived from the default options used in Intl.Collator.
And currently, I don't think any other WTF client wants to use it. If we have such a client, then I think we should move it to WTF.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:409
>> +    UErrorCode status = U_ZERO_ERROR;
> 
> should this go inside the loop?

Moved.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:411
>> +    int32_t index = 0;
> 
> remove

Removed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:417
>> +        index++;
> 
> remove

Removed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:449
>> +    m_canDoASCIIUCADUCETComparison = result ? TriState::True : TriState::False;
> 
> m_canDoASCIIUCADUCETComparison = triState(result);

Fixed.

>>> Source/JavaScriptCore/runtime/IntlCollator.cpp:456
>>> +#if ASSERT_ENABLED
>> 
>> Small style idea. Consider putting this in the header:
>> 
>>     #if !ASSERT_ENABLED
>>     inline void IntlCollator::checkICULocaleInvariants(const HashSet<String>&) { }
>>     #endif
>> 
>> Then here just write:
>> 
>>     #if ASSERT_ENABLED
>> 
>>     void IntlCollator::checkICULocaleInvariants(const HashSet<String>& locales)
>>     {
>>     }
>> 
>>     #endif
>> 
>> Without the UNUSED_PARAM.
> 
> I can see how this was useful to figure out ICU's current behavior and how it would provide a way to immediately see if that changes, but it seems rather heavy to be treated as "just some assertions"?

I think this function is very important assertion. If the invariant is broken, we should fix the code, and without this exhaustive check, we cannot notice ICU is changed.
We should exhaustively check ICU behavior here to catch future changes in ICU.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:458
>> +        ([&] {
> 
> Here too I'm not quite sure about the immediately-invoked lambda.

In this case, I can use `continue` instead. Changed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:470
>> +                            dataLogLn("BAD ", locale, " ", makeString(hex(x)), "(", String(xstring, 1), ") <=> ", makeString(hex(y)), "(", String(ystring, 1), ") ICU:(", static_cast<int32_t>(resultICU), "),WTF:(", static_cast<int32_t>(resultWTF), ")");
> 
> These should be StringView, no need to create and  destroy a string each time.

Fixed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:126
>> +    auto compareCharacter = [](CharacterType1 lhs, CharacterType2 rhs) {
> 
> should we inline this below just to ensure that we don't actually have the "result != 0" branch in case the compiler doesn't optimize it away?

Sounds like inlining this manually is better. Changed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:137
>> +    while (true) {
> 
> why not just: "while (position < commonLength)"?

This sounds good, fixed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:145
>> +        ++position;
> 
> Is this better than just indexing on position?

We can just use position.

[Yusuke Suzuki]

Created attachment 407219 [details]
Patch

[Saam Barati]

Comment on attachment 407177 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407177&action=review

>>>> Source/JavaScriptCore/runtime/IntlCollator.cpp:295
>>>> +            if (canDoASCIIUCADUCETComparison()) {
>>> 
>>> shouldn't be the first thing to check for the speed of not having to check the entire strings if they're ASCII when in a UCollator that can't go down the fast path?
>> 
>> No, this has the ucol_strcollUTF8 optimization too, not just compareASCIIWithUACDUCET.
> 
> Yeah, this is necessary since there is ucol_strcollUTF8 path.

yeah I missed that path. Ignore comment

[Yusuke Suzuki]

Created attachment 407236 [details]
Patch

[Yusuke Suzuki]

Created attachment 407239 [details]
Patch

[Yusuke Suzuki]

Created attachment 407240 [details]
Patch

[Darin Adler]

Comment on attachment 407239 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407239&action=review

> Source/JavaScriptCore/runtime/IntlCollator.cpp:297
> +        if (x.isAllSpecialCharacters<static_cast<bool*(UChar)>(canUseFastPath)>() && y.isAllSpecialCharacters<static_cast<bool*(UChar)>(canUseFastPath)>()) {

Do the type conversion with another local variable? Could get rid of the need for static_cast.

    bool (*canUseFastPathPredicate)(UChar) = canUseFastPath;

> Source/JavaScriptCore/runtime/IntlCollator.cpp:309
> +            if (x.is8Bit() && y.is8Bit())
> +                return ucol_strcollUTF8(m_collator.get(), bitwise_cast<const char*>(x.characters8()), x.length(), bitwise_cast<const char*>(y.characters8()), y.length(), &status);

This could be done even in cases involving control characters. Is it worth doing that?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:324
> +        auto getCharacters16 = [&] (const StringView& view, Vector<UChar>& buffer) -> const UChar* {
>              if (!view.is8Bit())
>                  return view.characters16();
> -            buffer.resize(view.length());
> +            buffer.grow(view.length());
>              StringImpl::copyCharacters(buffer.data(), view.characters8(), view.length());
>              return buffer.data();
>          };
>  
>          Vector<UChar> xBuffer;
>          Vector<UChar> yBuffer;
> -        const UChar* xCharacters = getCharacters(x, xBuffer);
> -        const UChar* yCharacters = getCharacters(y, yBuffer);
> -        result = ucol_strcoll(m_collator.get(), xCharacters, x.length(), yCharacters, y.length());
> -    }
> +        const UChar* xCharacters = getCharacters16(x, xBuffer);
> +        const UChar* yCharacters = getCharacters16(y, yBuffer);
> +        return ucol_strcoll(m_collator.get(), xCharacters, x.length(), yCharacters, y.length());

The StringView::upconvertedCharacters function is designed to that uses like this are easier to write:

    return ucol_strcoll(m_collator.get(), x.upconvertedCharacters(), x.length(), y.upconvertedCharacters(), y.length());

Will that work?

> Source/JavaScriptCore/runtime/IntlCollator.cpp:403
> +        // We do not check UCOL_NORMALIZATION_MODE status since FCD normalization do nothing for ASCII strings.

"do nothing" -> "does nothing"

> Source/JavaScriptCore/runtime/IntlCollator.cpp:408
> +        UColAttributeValue result = ucol_getAttribute(&collator, pair.first, &status);

auto

> Source/JavaScriptCore/runtime/IntlCollator.cpp:414
> +    // Check existence of tailoring rules. If it does not exist, collation algorithm is UAC DUCET.

"it does not exist" -> "they do not exist"

> Source/JavaScriptCore/runtime/IntlCollator.h:71
> +        if (m_canDoASCIIUCADUCETComparison != TriState::Indeterminate)
> +            return m_canDoASCIIUCADUCETComparison == TriState::True;
> +        return canDoASCIIUCADUCETComparisonSlow();

I prefer to a style like this:

    if (m_canDoASCIIUCADUCETComparison == TriState::Indeterminate)
        updateCanDoASCIIUCADUCETComparison();
    return m_canDoASCIIUCADUCETComparison == TriState::True;

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:121
> +inline int compareASCIIWithUACDUCET(const CharacterType1* characters1, unsigned length1, const CharacterType2* characters2, unsigned length2)

Caller wants UCollationResult. Why not return it here instead of putting the typecast at the call location.

[Yusuke Suzuki]

Created attachment 407241 [details]
Patch

[Yusuke Suzuki]

Created attachment 407243 [details]
Patch

[Yusuke Suzuki]

Comment on attachment 407239 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407239&action=review

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:297
>> +        if (x.isAllSpecialCharacters<static_cast<bool*(UChar)>(canUseFastPath)>() && y.isAllSpecialCharacters<static_cast<bool*(UChar)>(canUseFastPath)>()) {
> 
> Do the type conversion with another local variable? Could get rid of the need for static_cast.
> 
>     bool (*canUseFastPathPredicate)(UChar) = canUseFastPath;

It turned out this is not compiled well in GCC. I'll just define a function.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:309
>> +                return ucol_strcollUTF8(m_collator.get(), bitwise_cast<const char*>(x.characters8()), x.length(), bitwise_cast<const char*>(y.characters8()), y.length(), &status);
> 
> This could be done even in cases involving control characters. Is it worth doing that?

I think this is not worth doing. Control characters are very rare. Calling isAllASCII for control characters could hurt UTF-16 case performance.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:324
>> +        return ucol_strcoll(m_collator.get(), xCharacters, x.length(), yCharacters, y.length());
> 
> The StringView::upconvertedCharacters function is designed to that uses like this are easier to write:
> 
>     return ucol_strcoll(m_collator.get(), x.upconvertedCharacters(), x.length(), y.upconvertedCharacters(), y.length());
> 
> Will that work?

It seems that upconvertedCharacters can just work. Changed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:403
>> +        // We do not check UCOL_NORMALIZATION_MODE status since FCD normalization do nothing for ASCII strings.
> 
> "do nothing" -> "does nothing"

Fixed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:408
>> +        UColAttributeValue result = ucol_getAttribute(&collator, pair.first, &status);
> 
> auto

Changed.

>> Source/JavaScriptCore/runtime/IntlCollator.cpp:414
>> +    // Check existence of tailoring rules. If it does not exist, collation algorithm is UAC DUCET.
> 
> "it does not exist" -> "they do not exist"

Fixed.

>> Source/JavaScriptCore/runtime/IntlCollator.h:71
>> +        return canDoASCIIUCADUCETComparisonSlow();
> 
> I prefer to a style like this:
> 
>     if (m_canDoASCIIUCADUCETComparison == TriState::Indeterminate)
>         updateCanDoASCIIUCADUCETComparison();
>     return m_canDoASCIIUCADUCETComparison == TriState::True;

Changed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:121
>> +inline int compareASCIIWithUACDUCET(const CharacterType1* characters1, unsigned length1, const CharacterType2* characters2, unsigned length2)
> 
> Caller wants UCollationResult. Why not return it here instead of putting the typecast at the call location.

Originally, it was aligned to WTF's generic codePointCompare. But since this function is only used by IntlCollator, using UCollationResult is OK. Changed.

[Yusuke Suzuki]

Created attachment 407246 [details]
Patch

[Yusuke Suzuki]

Created attachment 407261 [details]
Patch

[Darin Adler]

Comment on attachment 407261 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407261&action=review

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:125
> +    static_assert(UCOL_EQUAL == 0);
> +    static_assert(UCOL_GREATER == 1);
> +    static_assert(UCOL_LESS == -1);

Don’t need these any more.

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:129
> +        CharacterType1 lhs = characters1[position];
> +        CharacterType2 rhs = characters2[position];

auto

> Source/JavaScriptCore/runtime/IntlObjectInlines.h:144
> +    return (length1 > length2) ? UCOL_GREATER : UCOL_LESS;

No parentheses needed here (match the weight if statement above).

[Yusuke Suzuki]

Comment on attachment 407261 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=407261&action=review

Thanks!

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:125
>> +    static_assert(UCOL_LESS == -1);
> 
> Don’t need these any more.

Removed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:129
>> +        CharacterType2 rhs = characters2[position];
> 
> auto

Fixed.

>> Source/JavaScriptCore/runtime/IntlObjectInlines.h:144
>> +    return (length1 > length2) ? UCOL_GREATER : UCOL_LESS;
> 
> No parentheses needed here (match the weight if statement above).

Removed

[Yusuke Suzuki]

Committed r266178: <https://trac.webkit.org/changeset/266178>

[Yusuke Suzuki]

*** Bug 230199 has been marked as a duplicate of this bug. ***