|Summary:||Enable LLVM TailDup|
|Product:||WebKit||Reporter:||Filip Pizlo <fpizlo>|
|Severity:||Normal||CC:||atrick, barraclough, ggaren, mark.lam, mhahnenberg, mmirman, msaboff, nrotem, oliver, sam|
|Version:||528+ (Nightly build)|
|Bug Depends on:|
Description Filip Pizlo 2014-01-18 20:17:58 PST
Comment 2 Filip Pizlo 2014-01-18 20:21:15 PST
Created attachment 221579 [details] llvm patch This just reverts LLVM r138841, and then updates the code a bit to various internal API changes.
Comment 3 Geoffrey Garen 2014-01-18 22:33:14 PST
Comment on attachment 221578 [details] jsc patch r=me
Comment 4 Geoffrey Garen 2014-01-18 22:35:07 PST
Comment on attachment 221579 [details] llvm patch Seems like it would be best to get this into LLVM trunk first.
Comment 5 Filip Pizlo 2014-01-18 22:39:16 PST
(In reply to comment #4) > (From update of attachment 221579 [details]) > Seems like it would be best to get this into LLVM trunk first. Yeah - that's definitely a blocker. At the time that they removed TailDup there seemed to be multiple good justifications for doing so: - It de-SSAifies the IR in a weird non-canonical way. - It has demonstrably bad heuristics for important things like switch. - It was unused in LLVM. So it may be a challenge to get it pushed back in. I suspect that this optimization is only good for JS because JS is so weird.