RESOLVED FIXED 195013
[JSC] Revert r226885 to make SlotVisitor creation lazy 
https://bugs.webkit.org/show_bug.cgi?id=195013
Summary [JSC] Revert r226885 to make SlotVisitor creation lazy
Yusuke Suzuki
Reported 2019-02-25 13:16:38 PST
[JSC] Revert r226885 to make SlotVisitor creation lazy
Attachments
Patch (6.08 KB, patch)
2019-02-25 13:27 PST, Yusuke Suzuki 		no flags
DetailsFormatted DiffDiff
Patch (6.31 KB, patch)
2019-02-25 13:33 PST, Yusuke Suzuki 		saam: review+
DetailsFormatted DiffDiff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.
Yusuke Suzuki
Comment 1 2019-02-25 13:27:14 PST
Created attachment 362922 [details] Patch
Yusuke Suzuki
Comment 2 2019-02-25 13:33:06 PST
Created attachment 362923 [details] Patch
Saam Barati
Comment 3 2019-02-25 13:35:31 PST
Comment on attachment 362923 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=362923&action=review > Source/JavaScriptCore/ChangeLog:8 > + We once changed SlotVisitor creation apriori to drop the lock. But it turns out that SlotVisitor is memory-consuming. I don't understand what these two sentences have to do with each other. What does "memory-consuming" here mean?
Yusuke Suzuki
Comment 4 2019-02-25 13:40:11 PST
Comment on attachment 362923 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=362923&action=review >> Source/JavaScriptCore/ChangeLog:8 >> + We once changed SlotVisitor creation apriori to drop the lock. But it turns out that SlotVisitor is memory-consuming. > > I don't understand what these two sentences have to do with each other. What does "memory-consuming" here mean? In previous change, we created # of Heap collectors SlotVisitors apriori. But each SlotVisitor takes so much memory (each has two MarkedStack, and each MarkedStack at least has 4KB dirty memory). This patch reverts that apriori creation, and creates SlotVisitors dynamically.
Yusuke Suzuki
Comment 5 2019-02-25 14:18:05 PST
Comment on attachment 362923 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=362923&action=review >>> Source/JavaScriptCore/ChangeLog:8 >>> + We once changed SlotVisitor creation apriori to drop the lock. But it turns out that SlotVisitor is memory-consuming. >> >> I don't understand what these two sentences have to do with each other. What does "memory-consuming" here mean? > > In previous change, we created # of Heap collectors SlotVisitors apriori. But each SlotVisitor takes so much memory (each has two MarkedStack, and each MarkedStack at least has 4KB dirty memory). > This patch reverts that apriori creation, and creates SlotVisitors dynamically. And the previous patch removed the lock because we no longer created SlotVisitors dynamically. But this patch attempts to revert that. Then, the lock comes back. I checked `forEachSlotVisitor()` calls in our JSC code base and the passed lambda does not cause dead-lock.
Saam Barati
Comment 6 2019-02-25 19:22:10 PST
Comment on attachment 362923 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=362923&action=review >>>> Source/JavaScriptCore/ChangeLog:8 >>>> + We once changed SlotVisitor creation apriori to drop the lock. But it turns out that SlotVisitor is memory-consuming. >>> >>> I don't understand what these two sentences have to do with each other. What does "memory-consuming" here mean? >> >> In previous change, we created # of Heap collectors SlotVisitors apriori. But each SlotVisitor takes so much memory (each has two MarkedStack, and each MarkedStack at least has 4KB dirty memory). >> This patch reverts that apriori creation, and creates SlotVisitors dynamically. > > And the previous patch removed the lock because we no longer created SlotVisitors dynamically. But this patch attempts to revert that. Then, the lock comes back. > I checked `forEachSlotVisitor()` calls in our JSC code base and the passed lambda does not cause dead-lock. "But" => "Also, it turns out..."
Yusuke Suzuki
Comment 7 2019-02-25 21:01:56 PST
Comment on attachment 362923 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=362923&action=review Thanks! >>>>> Source/JavaScriptCore/ChangeLog:8 >>>>> + We once changed SlotVisitor creation apriori to drop the lock. But it turns out that SlotVisitor is memory-consuming. >>>> >>>> I don't understand what these two sentences have to do with each other. What does "memory-consuming" here mean? >>> >>> In previous change, we created # of Heap collectors SlotVisitors apriori. But each SlotVisitor takes so much memory (each has two MarkedStack, and each MarkedStack at least has 4KB dirty memory). >>> This patch reverts that apriori creation, and creates SlotVisitors dynamically. >> >> And the previous patch removed the lock because we no longer created SlotVisitors dynamically. But this patch attempts to revert that. Then, the lock comes back. >> I checked `forEachSlotVisitor()` calls in our JSC code base and the passed lambda does not cause dead-lock. > > "But" => "Also, it turns out..." Fixed.
Yusuke Suzuki
Comment 8 2019-02-25 21:37:58 PST
Committed r242070: <https://trac.webkit.org/changeset/242070>
Radar WebKit Bug Importer
Comment 9 2019-02-25 21:41:15 PST
<rdar://problem/48390304>
Yusuke Suzuki
Comment 10 2019-05-28 01:45:08 PDT
Committed r245808: <https://trac.webkit.org/changeset/245808>
Note You need to log in before you can comment on or make changes to this bug.