ART: Add Mterp export pc poison testing mode
One of the optimizations of the Mterp assembly interpreter is
that it attempts to limit the number of times it updates the saved
dex_pc_ value stored in the shadow frame. Instead, it maintains
a pointer to the current dex instruction in a dedicated register.
However, whenever execution flow exits the immediate mterp
environment and some other part of the runtime needs to know the
current dex pc (for example - exception throws), mterp must
save - or "export" the dex_pc back into it's home location in
the shadow frame.
Forgetting to do this is a typical bug when developing mterp
instruction handlers. And, the failures show up in odd and
sometimes hard to track down ways. This CL adds a testing mode
that poisons the saved dex_pc prior to each instruction
interpretation. Thus, if any individual handler neglects to
properly export the current dex pc and then goes on to use it
we'll segfault out with an identifiable fault address: 0xdead00ff.
TEST: test-art-target (Nexus 5x, Nexus 6p), test-art-host. Verified
expected failure by artificially neglecting an EXPORT_PC.
Bug: 26572899
Change-Id: I704fbb8ece693b94341aed0ffbb51a6acfb89697
16 files changed