Bug 143833

Summary: DFG bytecode parser should record the places where bytecode operands die and this should be preserved by write(SideState)/NodeMustGenerate
Product: WebKit Reporter: Filip Pizlo <fpizlo>
Component: JavaScriptCoreAssignee: Filip Pizlo <fpizlo>
Status: NEW    
Severity: Normal    
Priority: P2    
Version: 528+ (Nightly build)   
Hardware: All   
OS: All   
Bug Depends on:    
Bug Blocks: 143734    
Attachments:
Description Flags
work in progress
none
passing some tests none

Filip Pizlo
Reported 2015-04-16 11:27:44 PDT
...
Attachments
work in progress (47.03 KB, patch)
2015-04-16 15:56 PDT, Filip Pizlo 		no flags
DetailsFormatted DiffDiff
passing some tests (61.86 KB, patch)
2015-04-19 14:19 PDT, Filip Pizlo 		no flags
DetailsFormatted DiffDiff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.
Filip Pizlo
Comment 1 2015-04-16 15:56:35 PDT
Created attachment 250962 [details] work in progress Starting to pass tests.
Filip Pizlo
Comment 2 2015-04-19 14:19:15 PDT
Created attachment 251128 [details] passing some tests Still need to test more things and run benchmarks. The perf impact of this is the insertion of additional nodes and the fact that block->last() is no longer a terminal. Getting the terminal requires a small backwards search.
Filip Pizlo
Comment 3 2015-04-19 14:47:33 PDT
I'm no longer sure about this. It appears to require some dirty logic for inserting the KillHints in the "right" places and then maintaining them through transformations. It would be super awesome if we could just use FullBytecodeLiveness.
Filip Pizlo
Comment 4 2015-04-19 15:28:31 PDT
The point of inserting KillHints was that this was supposed to be faster than consulting FullBytecodeLiveness late in the game. But, as far as I can tell, even inserting KillHints incurs a 1% loss on SunSpider. I'm going to factor out the sloppy-terminal part of this patch. Hopefully that is not responsible for any slow-down. If it's not, then I'll move on to writing a Phantom inserter based on FullBytecodeLiveness. Then I'll try to make that as fast as possible. It seems like it shouldn't have to be slow - after all, liveness over bytecode is just fun with bitvectors.
Note You need to log in before you can comment on or make changes to this bug.