P4
A PRT run of solaris amd64 was failing in specJBB. I eventually discovered that
it is because of the way callee save registers are treated on 64bit platforms.
(sparc is basically immune because it uses L registers). When we deoptimize because
c2 thinks of registers as 32bit items we create 2 entries in the callee save area
of the vframe array. Each entry is 64bits wide (ought to be 32bits). So this
means that we store two copies of the callee save register. The gc will only
update the lower numbered slot. When the c2i adapter gets setup it copies
BOTH slots out of the vframe array and since the high numbered slot is
written last (the un-gc'd slot) it "wins". This causes random bad oop failures.
it is because of the way callee save registers are treated on 64bit platforms.
(sparc is basically immune because it uses L registers). When we deoptimize because
c2 thinks of registers as 32bit items we create 2 entries in the callee save area
of the vframe array. Each entry is 64bits wide (ought to be 32bits). So this
means that we store two copies of the callee save register. The gc will only
update the lower numbered slot. When the c2i adapter gets setup it copies
BOTH slots out of the vframe array and since the high numbered slot is
written last (the un-gc'd slot) it "wins". This causes random bad oop failures.
- duplicates
-
JDK-6358341 JCK-runtime-14a RSACertStreamTEST_006 fails for 142_11b1 with -d64 -Xcomp [RSA algorithm not found]
-
- Closed
-