t0(extant, precise A) = new A
t1 = instanceof t0(extant, precise A), A
it would be nice to simplify this to:
t0(precise A) = new A
t1 = 1
however, we can't do this as t0 could be holding null - although from the code it clearly can't.
This issue is opened following r14138 and r14133. An alternate approach is to use expression folding - currently disabled both locally and globally.
Description
Following inlining we often get code of the form:
t0(extant, precise A) = new A
t1 = instanceof t0(extant, precise A), A
it would be nice to simplify this to:
t0(precise A) = new A
t1 = 1
however, we can't do this as t0 could be holding null - although from the code it clearly can't.
This issue is opened following r14138 and r14133. An alternate approach is to use expression folding - currently disabled both locally and globally.
BC2IR tries to catch this when it generates IR for the instanceof bytecode. Is the problem that the non-null guard on the scratch field isn't being propagated correctly across inlined method boundaries?
Of course we could get this kind of IR after BC2IR is complete and we do subsequent optimization passes, but if it is really as simple as the only definiton that flows into an instanceof being from a new statement, then BC2IR should be able to handle it without the non-null bit.
David Grove added a comment - 21/Apr/08 10:29 AM BC2IR tries to catch this when it generates IR for the instanceof bytecode. Is the problem that the non-null guard on the scratch field isn't being propagated correctly across inlined method boundaries?
Of course we could get this kind of IR after BC2IR is complete and we do subsequent optimization passes, but if it is really as simple as the only definiton that flows into an instanceof being from a new statement, then BC2IR should be able to handle it without the non-null bit.
The code where I'm hitting this maybe being conservative because of RVM-444. Thanks for the suggestion, I'll see if I can work around the problem in BC2IR.
Ian Rogers added a comment - 21/Apr/08 10:49 AM The code where I'm hitting this maybe being conservative because of RVM-444. Thanks for the suggestion, I'll see if I can work around the problem in BC2IR.
Improvement
Open
Major
Add non-nullness information to register operands
Of course we could get this kind of IR after BC2IR is complete and we do subsequent optimization passes, but if it is really as simple as the only definiton that flows into an instanceof being from a new statement, then BC2IR should be able to handle it without the non-null bit.