Bug 184182 – Make InlineTextBox::createTextRun() take a const lvalue reference String

• Status: RESOLVED
• Resolution: FIXED
• Component: Layout and Rendering

[Said Abou-Hallawa]

InlineTextBox::createTextRun() takes a non-const lvalue reference. It is tempting to change the signature of this method to take a const lvalue reference. But this was done intentionally. TextRun is effectively a StringView: it does not own the passed string. We rely on the compiler to prevent calls like createTextRun("abc").

A possible Solution is to make InlineTextBox own three TextRuns:

class InlineTextBox {

private:
    mutable std::uniqe_ptr<TextRun> m_textRun;
    mutable std::uniqe_ptr<TextRun> m_combinedTextRun;
    mutable std::uniqe_ptr<TextRun> m_hyphenatedTextRun;
};

Then we can replace InlineTextBox::createTextRun(String&) by InlineTextBox::textRun(bool ignoreCombinedText = false, bool ignoreHyphen = false) const.

[Said Abou-Hallawa]

Created attachment 337246 [details]
Patch

[Daniel Bates]

Comment on attachment 337246 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337246&action=review

> Source/WebCore/platform/graphics/FontCascade.cpp:377
> +float FontCascade::widthForSimpleText(const StringView& text) const

This is not necessary. StringView is a small object and should be passed by value.

> Source/WebCore/platform/graphics/TextRun.h:103
> +        m_string = emptyString();

I do not understand the need for TextRun to hold both a non-owning and owning String. It should either hold a non-owned string, StringView, or hold a String, not both.

[Said Abou-Hallawa]

Created attachment 337271 [details]
Patch

[Said Abou-Hallawa]

Comment on attachment 337246 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337246&action=review

>> Source/WebCore/platform/graphics/FontCascade.cpp:377
>> +float FontCascade::widthForSimpleText(const StringView& text) const
> 
> This is not necessary. StringView is a small object and should be passed by value.

Ok. Done.

>> Source/WebCore/platform/graphics/TextRun.h:103
>> +        m_string = emptyString();
> 
> I do not understand the need for TextRun to hold both a non-owning and owning String. It should either hold a non-owned string, StringView, or hold a String, not both.

Yes this is correct. TextRun will either (1) own String or (2) hold a non-owned string StringView.

-- In case (1), TextRun owns a String, m_stringView will encapsulate the owned String. This will make the code cleaner and easier to read since all the methods will operate on m_stringView.
-- In case (2), TextRun holds a non-owned string StringView, m_stringView will be a copy of this StringView. TextRun::m_string should be empty and should never be used in functions like TextRun::string().

This allows code like this:

auto text = this->text();
TextRun textRun = createTextRun(text);

To be replaced by this:

TextRun textRun = createTextRun(text());

We moved the ownership of the underlying of the String from the caller side to the TextRun itself. This will only happen when the caller passes a String instead of a StringView.

[Daniel Bates]

(In reply to Said Abou-Hallawa from comment #4)
> Comment on attachment 337246 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=337246&action=review
> 
> >> Source/WebCore/platform/graphics/FontCascade.cpp:377
> >> +float FontCascade::widthForSimpleText(const StringView& text) const
> > 
> > This is not necessary. StringView is a small object and should be passed by value.
> 
> Ok. Done.
> 
> >> Source/WebCore/platform/graphics/TextRun.h:103
> >> +        m_string = emptyString();
> > 
> > I do not understand the need for TextRun to hold both a non-owning and owning String. It should either hold a non-owned string, StringView, or hold a String, not both.
> 
> Yes this is correct. TextRun will either (1) own String or (2) hold a
> non-owned string StringView.
> 
> -- In case (1), TextRun owns a String, m_stringView will encapsulate the
> owned String. This will make the code cleaner and easier to read since all
> the methods will operate on m_stringView.
> -- In case (2), TextRun holds a non-owned string StringView, m_stringView
> will be a copy of this StringView. TextRun::m_string should be empty and
> should never be used in functions like TextRun::string().
> 

It seems sufficient for TextRun to hold a String. In case 1, holding both means we hold two pointers to the same StringImpl. This seems unnecessary and I would be surprised if there is any efficiency gained in case 2 by avoiding a ref/deref.

[alan]

(In reply to Daniel Bates from comment #5)
> (In reply to Said Abou-Hallawa from comment #4)
> > Comment on attachment 337246 [details]
> > Patch
> > 
> > View in context:
> > https://bugs.webkit.org/attachment.cgi?id=337246&action=review
> > 
> > >> Source/WebCore/platform/graphics/FontCascade.cpp:377
> > >> +float FontCascade::widthForSimpleText(const StringView& text) const
> > > 
> > > This is not necessary. StringView is a small object and should be passed by value.
> > 
> > Ok. Done.
> > 
> > >> Source/WebCore/platform/graphics/TextRun.h:103
> > >> +        m_string = emptyString();
> > > 
> > > I do not understand the need for TextRun to hold both a non-owning and owning String. It should either hold a non-owned string, StringView, or hold a String, not both.
> > 
> > Yes this is correct. TextRun will either (1) own String or (2) hold a
> > non-owned string StringView.
> > 
> > -- In case (1), TextRun owns a String, m_stringView will encapsulate the
> > owned String. This will make the code cleaner and easier to read since all
> > the methods will operate on m_stringView.
> > -- In case (2), TextRun holds a non-owned string StringView, m_stringView
> > will be a copy of this StringView. TextRun::m_string should be empty and
> > should never be used in functions like TextRun::string().
> > 
> 
> It seems sufficient for TextRun to hold a String. In case 1, holding both
> means we hold two pointers to the same StringImpl. This seems unnecessary
> and I would be surprised if there is any efficiency gained in case 2 by
> avoiding a ref/deref.
I very much agree with this.

[Said Abou-Hallawa]

Created attachment 337291 [details]
Patch

[Said Abou-Hallawa]

(In reply to Daniel Bates from comment #5)
> (In reply to Said Abou-Hallawa from comment #4)
> > Comment on attachment 337246 [details]
> > Patch
> > 
> > View in context:
> > https://bugs.webkit.org/attachment.cgi?id=337246&action=review
> > 
> > >> Source/WebCore/platform/graphics/FontCascade.cpp:377
> > >> +float FontCascade::widthForSimpleText(const StringView& text) const
> > > 
> > > This is not necessary. StringView is a small object and should be passed by value.
> > 
> > Ok. Done.
> > 
> > >> Source/WebCore/platform/graphics/TextRun.h:103
> > >> +        m_string = emptyString();
> > > 
> > > I do not understand the need for TextRun to hold both a non-owning and owning String. It should either hold a non-owned string, StringView, or hold a String, not both.
> > 
> > Yes this is correct. TextRun will either (1) own String or (2) hold a
> > non-owned string StringView.
> > 
> > -- In case (1), TextRun owns a String, m_stringView will encapsulate the
> > owned String. This will make the code cleaner and easier to read since all
> > the methods will operate on m_stringView.
> > -- In case (2), TextRun holds a non-owned string StringView, m_stringView
> > will be a copy of this StringView. TextRun::m_string should be empty and
> > should never be used in functions like TextRun::string().
> > 
> 
> It seems sufficient for TextRun to hold a String. In case 1, holding both
> means we hold two pointers to the same StringImpl. This seems unnecessary
> and I would be surprised if there is any efficiency gained in case 2 by
> avoiding a ref/deref.

My understanding was TextRun was changed such that it holds a StringView to avoid copying the underlying buffer of the String given that the caller guarantees the String outlives the TextRun. For instance GraphicsContext::drawBidiText() slices the passed TextRun into smaller sub-TextRuns without having to allocate and copy the text of each sub-TextRun.

This is why I kept the StringView member after adding a String member to the TextRun. But I've just learnt that StringView::toStringWithoutCopying() can solve this problem by creating a string which shares the same underlying buffer of the StringView and does not make a new copy of the buffer. So the returned String is effectively acts like a StringView in this case.

So I changed TextRun to hold only a String member which will be either a copy of another String or created by StringView::toStringWithoutCopying().

[alan]

Comment on attachment 337291 [details]
Patch

I'd prefer doing one thing per patch (not just because it's easier to review but also becasue the unrelated changes don't get lost if this gets rolled out).

[Said Abou-Hallawa]

(In reply to zalan from comment #9)
> Comment on attachment 337291 [details]
> Patch
> 
> I'd prefer doing one thing per patch (not just because it's easier to review
> but also becasue the unrelated changes don't get lost if this gets rolled
> out).

What do you mean by "one thing per patch"? What are these things?

This patch is basically rolling back http://trac.webkit.org/changeset/174228 but with a better fix. TextRun was holding a character pointer and length. Using StringView was a nice approach but it introduced the underlying String lifetime issue and the caller has to do tricks like:

    auto text = this->text();
    TextRun textRun = createTextRun(text);

instead of just
    TextRun textRun = createTextRun(text());

The new patch uses String as a replacement instead of StringView giving that no underlying charter buffer will be ever copied which was the purpose of using the StringView in the first time.

I do not think this patch can be split more than that. It replaces a single member of TextRun and fixes the code to compile, link and work as expected without perf regression. The rest of the patch are clean-ups which are very easy to follow.

[alan]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

> Source/WebCore/ChangeLog:3
> +        Fix the lifetime of the TextRun

This title is misleading. You don't change the lifetime of the TextRun.

> Source/WebCore/platform/graphics/TextRun.h:69
> +        RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(startOffset < m_text.length());

This line has nothing to do with this patch. Please remove it.

> Source/WebCore/platform/graphics/TextRun.h:83
> -    UChar operator[](unsigned i) const { ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); return m_text[i]; }
> -    const LChar* data8(unsigned i) const { ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(is8Bit()); return &m_text.characters8()[i]; }
> -    const UChar* data16(unsigned i) const { ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(!is8Bit()); return &m_text.characters16()[i]; }
> +    UChar operator[](unsigned i) const { RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); return m_text[i]; }
> +    const LChar* data8(unsigned i) const { RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(is8Bit()); return &m_text.characters8()[i]; }
> +    const UChar* data16(unsigned i) const { RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(!is8Bit()); return &m_text.characters16()[i]; }

These lines have nothing to do with this patch. Please remove them.

[alan]

> without perf regression. The rest of the patch are clean-ups which are very
> easy to follow.
I've wasted too much time looking at unrelated "easy to follow, clean-up" changes at regression points and I'd rather use that time for something else.

[Daniel Bates]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

> Source/WebCore/rendering/InlineTextBox.cpp:1217
> +TextRun InlineTextBox::createTextRun(bool ignoreCombinedText, bool ignoreHypheng) const

ignoreHypheng => ignoreHyphen

> Source/WebCore/rendering/InlineTextBox.cpp:1220
> +    TextRun textRun { text(ignoreCombinedText, ignoreHypheng), textPos(), expansion(), expansionBehavior(), direction(), dirOverride() || style.rtlOrdering() == VisualOrder, !renderer().canUseSimpleFontCodePath() };

Ditto.

> Source/WebCore/rendering/SimpleLineLayout.cpp:184
> +            TextRun run(String(textRenderer.text()));

Is it necessary to explicitly call the String constructor?

> Source/WebCore/rendering/SimpleLineLayoutTextFragmentIterator.cpp:52
> +    , hyphenStringWidth(shouldHyphenate ? (useSimplifiedTextMeasuring ? font.widthForSimpleText(style.hyphenString()) : font.width(TextRun(String(style.hyphenString())))) : 0)

Is it necessary to explicitly call the String constructor?

[Daniel Bates]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

>> Source/WebCore/platform/graphics/TextRun.h:69
>> +        RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(startOffset < m_text.length());
> 
> This line has nothing to do with this patch. Please remove it.

I agree with zalan, please make these changes in a separate bug.

>> Source/WebCore/platform/graphics/TextRun.h:83
>> +    const UChar* data16(unsigned i) const { RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(!is8Bit()); return &m_text.characters16()[i]; }
> 
> These lines have nothing to do with this patch. Please remove them.

Ditto.

[Daniel Bates]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

For your consideration, I suggest using uniform initializer syntax throughout this patch.

> Source/WebCore/platform/graphics/TextRun.h:91
> +    void setText(const LChar* c, unsigned len) { setText(StringView(c, len)); }

c  => text
len  => length

(Since we are touching this line anyway).

> Source/WebCore/platform/graphics/TextRun.h:92
> +    void setText(const UChar* c, unsigned len) { setText(StringView(c, len)); }

Ditto.

> Source/WebCore/platform/graphics/TextRun.h:93
> +    void setText(StringView stringView) { m_text = stringView.toStringWithoutCopying(); }

stringView => text

(Typically we name the argument of the setter to match the name of the setter and underlying instance variable)

[Daniel Bates]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

> Source/WebCore/ChangeLog:19
> +           will make a copy of this string and assign it to m_text.

This is not accurate. The word “string” is usually interpreted as a character buffer. This is not the same as a String object. Copying of a String refs the underlying StringImpl. That is, we do not make a copy of the underlying character data.

> Source/WebCore/ChangeLog:23
> +           will not use  StringView::toString(). Instead We will use

“Instead We”  => “Instead we”

The last sentence does not read well.

[Said Abou-Hallawa]

Comment on attachment 337291 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=337291&action=review

>> Source/WebCore/ChangeLog:3
>> +        Fix the lifetime of the TextRun
> 
> This title is misleading. You don't change the lifetime of the TextRun.

Title was changed to "Make InlineTextBox::createTextRun() take a const lvalue reference String"

>> Source/WebCore/ChangeLog:19
>> +           will make a copy of this string and assign it to m_text.
> 
> This is not accurate. The word “string” is usually interpreted as a character buffer. This is not the same as a String object. Copying of a String refs the underlying StringImpl. That is, we do not make a copy of the underlying character data.

This was changed to "This constructor will addRef the underlying StringImpl when assigning it to m_text".

>> Source/WebCore/ChangeLog:23
>> +           will not use  StringView::toString(). Instead We will use
> 
> “Instead We”  => “Instead we”
> 
> The last sentence does not read well.

Fixed.

>>> Source/WebCore/platform/graphics/TextRun.h:69
>>> +        RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(startOffset < m_text.length());
>> 
>> This line has nothing to do with this patch. Please remove it.
> 
> I agree with zalan, please make these changes in a separate bug.

Removed.

>>> Source/WebCore/platform/graphics/TextRun.h:83
>>> +    const UChar* data16(unsigned i) const { RELEASE_ASSERT_WITH_SECURITY_IMPLICATION(i < m_text.length()); ASSERT(!is8Bit()); return &m_text.characters16()[i]; }
>> 
>> These lines have nothing to do with this patch. Please remove them.
> 
> Ditto.

Removed.

>> Source/WebCore/platform/graphics/TextRun.h:91
>> +    void setText(const LChar* c, unsigned len) { setText(StringView(c, len)); }
> 
> c  => text
> len  => length
> 
> (Since we are touching this line anyway).

Fixed.

>> Source/WebCore/platform/graphics/TextRun.h:92
>> +    void setText(const UChar* c, unsigned len) { setText(StringView(c, len)); }
> 
> Ditto.

Fixed.

>> Source/WebCore/platform/graphics/TextRun.h:93
>> +    void setText(StringView stringView) { m_text = stringView.toStringWithoutCopying(); }
> 
> stringView => text
> 
> (Typically we name the argument of the setter to match the name of the setter and underlying instance variable)

Fixed.

>> Source/WebCore/rendering/InlineTextBox.cpp:1217
>> +TextRun InlineTextBox::createTextRun(bool ignoreCombinedText, bool ignoreHypheng) const
> 
> ignoreHypheng => ignoreHyphen

Fixed.

>> Source/WebCore/rendering/InlineTextBox.cpp:1220
>> +    TextRun textRun { text(ignoreCombinedText, ignoreHypheng), textPos(), expansion(), expansionBehavior(), direction(), dirOverride() || style.rtlOrdering() == VisualOrder, !renderer().canUseSimpleFontCodePath() };
> 
> Ditto.

Fixed.

>> Source/WebCore/rendering/SimpleLineLayout.cpp:184
>> +            TextRun run(String(textRenderer.text()));
> 
> Is it necessary to explicitly call the String constructor?

Yes. Because RenderText::text() returns StringImpl. Since both String and StringView can be implicitly constructed given a StringImpl, the compiler wants us to be more explicit about which constructor of TextRun to be call.

>> Source/WebCore/rendering/SimpleLineLayoutTextFragmentIterator.cpp:52
>> +    , hyphenStringWidth(shouldHyphenate ? (useSimplifiedTextMeasuring ? font.widthForSimpleText(style.hyphenString()) : font.width(TextRun(String(style.hyphenString())))) : 0)
> 
> Is it necessary to explicitly call the String constructor?

Yes. Because RenderStyle::hyphenString() returns an AtomicString.

[Said Abou-Hallawa]

Created attachment 337508 [details]
Patch

[WebKit Commit Bot]

Comment on attachment 337508 [details]
Patch

Clearing flags on attachment: 337508

Committed r230452: <https://trac.webkit.org/changeset/230452>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.

[Radar WebKit Bug Importer]

<rdar://problem/39296994>