review (XS) for 6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
Vladimir Kozlov
vladimir.kozlov at oracle.com
Tue Sep 7 11:22:11 PDT 2010
Tom,
I am fine with your index explicit check, I only want an additional check for array ptr and your suggestion is good.
Your current changes look good.
Thanks,
Vladimir
Tom Rodriguez wrote:
> On Sep 7, 2010, at 10:49 AM, Vladimir Kozlov wrote:
>
>> Looks good but could we also check (in assert?) that address type is array pointer?
>
> I considered adding that as an early check but I wanted an explicit check that we found the index to confirm the shape of the address expression. I could add an early weed out test:
>
> --- a/src/share/vm/opto/loopTransform.cpp Fri Sep 03 13:31:03 2010 -0700
> +++ b/src/share/vm/opto/loopTransform.cpp Tue Sep 07 11:04:07 2010 -0700
> @@ -2417,6 +2417,8 @@ bool PhaseIdealLoop::match_fill_loop(Ide
> Node* value = n->in(MemNode::ValueIn);
> if (!lpt->is_invariant(value)) {
> msg = "variant store value";
> + } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
> + msg = "not array address";
> }
> store = n;
> store_value = value;
>
> tom
>
>> Thanks,
>> Vladimir
>>
>> Tom Rodriguez wrote:
>>> http://cr.openjdk.java.net/~never/6982533
>>> 6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
>>> Reviewed-by:
>>> The logic for matching a byte fill is missing a check for the use of
>>> the index. It normally happens as part of the check for a shift
>>> expression but since a byte array doesn't have a shift the check is
>>> missed. The fix is to make the index check explicit. The reason it
>>> didn't always crash was because of differences in heap size caused by
>>> ergonomics. With a 16m heap it crashes on any machine. Tested with
>>> failing test case.
>>> src/share/vm/opto/loopTransform.cpp
>
More information about the hotspot-compiler-dev
mailing list