Request for reviews (S): 6876276: assert(!is_visited, "visit only once")
Changpeng Fang
Changpeng.Fang at Sun.COM
Thu Aug 27 17:15:04 PDT 2009
http://cr.openjdk.java.net/~cfang/6876276/webrev.00/
Problem:
SuperWord scheduling of loads uses the memory state of the last executed
load in the packet
for the SuperWord load. This is not correct because memory order may be
violated. The test case
that has problem is like this:
store1
load1
store2
load2
after scheduling, it becomes:
store1
store2
load1
load2
It turns out that store2 depends on load1, so the memory edge for the
superword load (combination of load1 and load2)
is not correct.
Solution:
Schedule the superword loads based on dependence (we have already done
this for superword stores, and I don't know
why I didn't do this for the loads at that time).
Minor changes: *Use "97" as the exit code (instead of "-1") for the
Tests in test/compiler/6636138*
Tests:
JPRT, CompileTheWord, and *Tests in test/compiler/6636138*
Thanks,
Changpeng
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20090827/45067984/attachment.html
More information about the hotspot-compiler-dev
mailing list