Bug 188559 – Use a Variant instead of a union in CSSSelector

• Status: REOPENED
• Component: New Bugs

[Alex Christensen]

Use a Variant instead of a union in CSSSelector

[Alex Christensen]

Created attachment 347086 [details]
Patch

[Alex Christensen]

Created attachment 347092 [details]
Patch

[Antti Koivisto]

Comment on attachment 347092 [details]
Patch

Nice, r=me

At some point we'll need to rewrite these classes completely. We need separate types for Compound/Complex/SimpleSelector instead of the current mess where everything is a CSSSelector.

[Alex Christensen]

I also updated the indentation of CSSSelector.h.
http://trac.webkit.org/r234859

[Radar WebKit Bug Importer]

<rdar://problem/43295037>

[Ross Kirsling]

This broke WinCairo and AppleWin builds. :(

> Source\WebCore\css/CSSSelector.cpp(44): error C2338: CSSSelector should remain small.

(from https://build.webkit.org/builders/WinCairo%2064-bit%20WKL%20Release%20%28Build%29/builds/2478/steps/compile-webkit/logs/stdio)

[Alex Christensen]

What is sizeof(CSSSelector) on Windows (64- or 32-bit)?

[Ross Kirsling]

(In reply to Alex Christensen from comment #7)
> What is sizeof(CSSSelector) on Windows (64- or 32-bit)?

On Win64 and Mac64 alike, I'm seeing 16 for the struct and 12 for the sum (8 + 4).

If I change the static_assert to `== 16`, the compilation proceeds, but we still get this warning quartet over and over:

> Source\WebCore\css\CSSSelector.h(363): warning C4315: 'WebCore::CSSSelector::RareData': 'this' pointer for member 'WebCore::CSSSelector::RareData::m_attribute' may not be aligned 8 as expected by the constructor
> Source\WebCore\css\CSSSelector.h(365): warning C4315: 'WebCore::CSSSelector::RareData': 'this' pointer for member 'WebCore::CSSSelector::RareData::m_argument' may not be aligned 8 as expected by the constructor
> Source\WebCore\css\CSSSelector.h(376): warning C4315: 'WebCore::CSSSelector::NameWithCase': 'this' pointer for member 'WebCore::CSSSelector::NameWithCase::m_originalName' may not be aligned 8 as expected by the constructor
> Source\WebCore\css\CSSSelector.h(376): warning C4315: 'WebCore::CSSSelector::NameWithCase': 'this' pointer for member 'WebCore::CSSSelector::NameWithCase::m_lowercaseLocalName' may not be aligned 8 as expected by the constructor

[Ross Kirsling]

(In reply to Ross Kirsling from comment #8)
> (In reply to Alex Christensen from comment #7)
> > What is sizeof(CSSSelector) on Windows (64- or 32-bit)?
> 
> On Win64 and Mac64 alike, I'm seeing 16 for the struct and 12 for the sum (8
> + 4).
> 
> If I change the static_assert to `== 16`, the compilation proceeds, but we
> still get this warning quartet over and over:
> 
> > Source\WebCore\css\CSSSelector.h(363): warning C4315: 'WebCore::CSSSelector::RareData': 'this' pointer for member 'WebCore::CSSSelector::RareData::m_attribute' may not be aligned 8 as expected by the constructor
> > Source\WebCore\css\CSSSelector.h(365): warning C4315: 'WebCore::CSSSelector::RareData': 'this' pointer for member 'WebCore::CSSSelector::RareData::m_argument' may not be aligned 8 as expected by the constructor
> > Source\WebCore\css\CSSSelector.h(376): warning C4315: 'WebCore::CSSSelector::NameWithCase': 'this' pointer for member 'WebCore::CSSSelector::NameWithCase::m_originalName' may not be aligned 8 as expected by the constructor
> > Source\WebCore\css\CSSSelector.h(376): warning C4315: 'WebCore::CSSSelector::NameWithCase': 'this' pointer for member 'WebCore::CSSSelector::NameWithCase::m_lowercaseLocalName' may not be aligned 8 as expected by the constructor

Oops, spoke too soon. Still encountered the error anyway...

[Alex Christensen]

Feel free to roll it out while you look for a fix.  sizeof(CSSSelector) is definitely 12 for 64-bit Mac and linux.

[Ross Kirsling]

(In reply to Alex Christensen from comment #10)
> Feel free to roll it out while you look for a fix.  sizeof(CSSSelector) is
> definitely 12 for 64-bit Mac and linux.

Alrighty. I hadn't realized right away that sizeof(SameSizeAsCSSSelector) was made irrelevant here, but it looks like sizeof(CSSSelector) is 13 on Windows.

[Fujii Hironori]

Reverted r234859 for reason:

Windows ports can't compile

Committed r234877: <https://trac.webkit.org/changeset/234877>

[Alex Christensen]

Fujii or Ross, could you get this working on Windows 32-bit and 64-bit?  I think it's just a matter of tweaking a few small things, but iterating from EWS is waaay too slow.

[Fujii Hironori]

Created attachment 347150 [details]
Patch

It seems that MSVC doesn't pack char (type of Variant) and unsigned int (the flags).
One possible workaroud is using char for the flags.

[Fujii Hironori]

Comment on attachment 347150 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=347150&action=review

> Source/WTF/ChangeLog:9
> +        Add packing macros to make it so Variant-containing structures don't always have 7 bytes of padding per Variant.

This makes subsequent members after Variant in classes unaligned.
Unaligned access is a undefined behavior.

[Fujii Hironori]

Created attachment 347151 [details]
Patch

* Use unsigned char for the flags
* Reoder flags in order to align 8 bits

[Ross Kirsling]

(In reply to Fujii Hironori from comment #16)
> Created attachment 347151 [details]
> Patch
> 
> * Use unsigned char for the flags
> * Reoder flags in order to align 8 bits

Locally, it seems that:
 - "unsigned" -> "unsigned char" makes the assert pass on Windows
 - reordering fields doesn't help us at all(?)
 - neither of these steps makes the warnings go away
 - Xcode wants the fields to be initialized in the same order that they are declared

[Alex Christensen]

Any luck getting this to work on Windows?  I'd still like to land it.

[Fujii Hironori]

What about an idea adding a new class WTF::CompactVariant<Traits, ...> which uses a template traits to get the position of m_type member variable?

[Alex Christensen]

That seems quite excessive, and it wouldn't benefit all existing users of Variant.

[Fujii Hironori]

Created attachment 349620 [details]
Patch

* Rebased onto ToT
* Removed pragma(pack) of CSSSelector to solve the warning quartet
* Try to solve "field 'm_pseudoType' will be initialized after field 'm_isLastInSelectorList'" on Mac ports

[WebKit Commit Bot]

Comment on attachment 349620 [details]
Patch

Clearing flags on attachment: 349620

Committed r235976: <https://trac.webkit.org/changeset/235976>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.

[Alex Christensen]

Rolled out in http://trac.webkit.org/r236284
We need to do something about alignment...

[Guillaume Emont]

(In reply to Alex Christensen from comment #24)
> Rolled out in http://trac.webkit.org/r236284
> We need to do something about alignment...

Note that a bug was opened about this issue https://bugs.webkit.org/show_bug.cgi?id=189758