Bug 194953 – DFGBytecodeParser should not declare that a node won't clobberExit if DFGFixupPhase can later declare it does clobberExit

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Robin Morisset]

Or we fairly naturally hit validation errors out of DFGFixupPhase..
There is already such a rule for GetByVal, but we must also be conservative for other nodes that:
- (a) may or may not clobberExit depending on their arrayMode
- (b) and get their arrayMode from profiling information in DFGBytecodeParser
- (c) and can have their arrayMode refined by DFGFixupPhase.

[Robin Morisset]

<rdar://problem/47595253>

[Robin Morisset]

Created attachment 362738 [details]
Patch

[Saam Barati]

Comment on attachment 362738 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=362738&action=review

r=me

> Source/JavaScriptCore/ChangeLog:14
> +        Otherwise we are likely to hit validation failure after fixup.

instead of "likely" I think you should describe precisely when/why this happens.

> Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:7216
> +        m_exitOK = false; // PutByVal and PutByValDirect must be treated as if they clobber exit state, since FixupPhase may make them generic.

This one isn't actually needed since PutByVal always clobbers exit unless it's in ForceOSR exit mode. And if it's in that mode, I don't think we can ever refine it to be wider? I'm ok with keeping this here conservatively, but this isn't the right comment to have.

[Robin Morisset]

(In reply to Saam Barati from comment #3)
> Comment on attachment 362738 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=362738&action=review
> 
> r=me
> 
> > Source/JavaScriptCore/ChangeLog:14
> > +        Otherwise we are likely to hit validation failure after fixup.
> 
> instead of "likely" I think you should describe precisely when/why this
> happens.

OK, I am replacing this line by the following:
"Otherwise we will hit a validation failure after fixup if the next node was marked ExitValid and exits to the same semantic origin."

> > Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:7216
> > +        m_exitOK = false; // PutByVal and PutByValDirect must be treated as if they clobber exit state, since FixupPhase may make them generic.
> 
> This one isn't actually needed since PutByVal always clobbers exit unless
> it's in ForceOSR exit mode. And if it's in that mode, I don't think we can
> ever refine it to be wider? I'm ok with keeping this here conservatively,
> but this isn't the right comment to have.

I am not entirely sure whether or not we can refine it to be wider. Looking at ArrayMode::refine, it will be widened to Generic whenever "base && index && !isInt32Speculation(index)". I cannot tell if this is likely to ever happen or not, but it is not blatantly impossible.

[Robin Morisset]

Created attachment 362758 [details]
Patch

Improved a sentence in the Changelog, per Saam's suggestion.

[WebKit Commit Bot]

Comment on attachment 362758 [details]
Patch

Clearing flags on attachment: 362758

Committed r241968: <https://trac.webkit.org/changeset/241968>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.