[9] RFR(S): 8140574: C2 must re-execute checks after deoptimizing from merged uncommon traps
Roland Westrelin
roland.westrelin at oracle.com
Tue Oct 27 16:17:14 UTC 2015
> I see. So the only pattern we are looking for is where 'throw' is replaced with unstable_if uncommon trap.
Actually, thinking about this more, I don’t think testing for unstable_if is sufficient. We could have something like this:
if (i < 0) {
if (some_condition) {
// never taken so unc unstable_if
}
}
if (i >= length) {
// never taken so unc unstable_if
}
Then if the compiler proves some_condition is always true, we have:
if (i < 0) {
// never taken so unc unstable_if
}
if (i >= length) {
// never taken so unc unstable_if
}
and we go ahead with the folding but the deoptimizatin in the first branch would have us resume execution after the i < 0 test.
So I think we have to restrict when we apply that transformation even more, maybe matching the expected bytecode pattern more closely?
Roland.
>
> Okay then.
>
> Thanks,
> Vladimir
>
> On 10/27/15 11:09 PM, Roland Westrelin wrote:
>>> Should we just additionally check re-execute flag?
>>>
>>> dom_unc->jvms()->should_reexecute()
>>>
>>> Unfortunately wee can't change reexecute flag since JVM state is already recorded according to it.
>>>
>>> Roland, what do you think?
>>
>> Is checking for should_reexecute() sufficient? Couldn’t we reexecute some bytecode in one branch that caused the branch to become dead but not the if itself?
>>
>> This change was put in so we optimize the pattern:
>>
>> if (i < 0 || i >= length) { throw… }
>>
>> of explicit array bound checks. Restricting it to unstable_if uncommon traps, doesn’t break that scenario so I would say that fix is good enough and feels safe.
>>
>> Roland.
>>
More information about the hotspot-compiler-dev
mailing list