RFR: 8347006: LoadRangeNode floats above array guard in arraycopy intrinsic [v4]

[Vladimir Kozlov]

On Thu, 9 Jan 2025 10:22:12 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

>> C2's arraycopy intrinsic adds guards that check that the source and destination objects are arrays:
>> https://github.com/openjdk/jdk/blob/afe543414f58a04832d4f07dea88881d64954a0b/src/hotspot/share/opto/library_call.cpp#L5917-L5919
>> 
>> If these guards pass, the array length is loaded:
>> https://github.com/openjdk/jdk/blob/afe543414f58a04832d4f07dea88881d64954a0b/src/hotspot/share/opto/library_call.cpp#L5930-L5933
>> 
>> But since the `LoadRangeNode` is not pinned, it might float above the array guard:
>> https://github.com/openjdk/jdk/blob/afe543414f58a04832d4f07dea88881d64954a0b/src/hotspot/share/opto/graphKit.cpp#L1214
>> 
>> If the object is not an array, we will read garbage. That's usually fine because the result will not be used (the array guard will trigger) but with `-XX:+UseCompactObjectHeaders` it can happen that the memory right after the header is not mapped and we crash.
>> 
>> The fix is to add a `CheckCastPPNode` to propagate the information that the operand is an array and prevent the load from floating. 
>> 
>> Thanks to @shipilev for identifying the root cause!
>> 
>> I was able to reliably reproduce the issue with `compiler/arraycopy/TestArrayCopyNoInit.java` and `-XX:-UseTLAB -XX:+UnlockExperimentalVMOptions -XX:+UseCompactObjectHeaders` on Linux AArch64 and verified that the fix solves the problem.
>> 
>> Best regards,
>> Tobias
>
> Tobias Hartmann has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Copyright date

src/hotspot/share/opto/library_call.cpp line 4307:

> 4305:     // Keep track of the fact that 'obj' is an array to prevent
> 4306:     // array specific accesses from floating above the guard.
> 4307:     *obj = _gvn.transform(new CastPPNode(is_array_ctrl, *obj, TypeAryPtr::BOTTOM));

Should we do this for above code when layout is known for compiler (`layout_con` is checked)?

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/22967#discussion_r1909125151