Reviewers: Benedikt Meurer, Hannes Payer,
Message:
Hi guys, here is the restored overflow check. Hannes, the reason we didn't
see
it was because we cleared the overflow bit on the wrong instruction (the
store,
not the add).
Thanks,
--Michael
Description:
Remove unnecessary overflow check in
HGraphBuilder::BuildCreateAllocationMemento().
With this fix codegen looks like:
mov ecx,[eax+0xf] ;;; <@52,#38> load-named-field
add ecx,0x2 ;;; <@54,#40> add-i
mov [eax+0xf],ecx ;;; <@56,#41> store-named-field
without it there is an overflow check and jump to deopt.
x64 code looks similar, except there is an (annoying) smi-untag then
int32-to-smi around the add operation.
Please review this at https://codereview.chromium.org/104313003/
SVN Base: https://v8.googlecode.com/svn/branches/bleeding_edge
Affected files (+2, -0 lines):
M src/hydrogen.cc
Index: src/hydrogen.cc
diff --git a/src/hydrogen.cc b/src/hydrogen.cc
index
900e07ecdf9fa68a32540fc76be401d0b064a343..ef00703da7551eb0d76c61da21d6759dca4156e0
100644
--- a/src/hydrogen.cc
+++ b/src/hydrogen.cc
@@ -2592,6 +2592,8 @@ void HGraphBuilder::BuildCreateAllocationMemento(
AllocationSite::kMementoCreateCountOffset));
memento_create_count = AddUncasted<HAdd>(
memento_create_count, graph()->GetConstant1());
+ // This smi value is reset to zero after every gc, overflow isn't a
problem.
+ memento_create_count->ClearFlag(HValue::kCanOverflow);
HStoreNamedField* store = Add<HStoreNamedField>(
allocation_site, HObjectAccess::ForAllocationSiteOffset(
AllocationSite::kMementoCreateCountOffset),
memento_create_count);
--
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.
