RFR: 8268362: [REDO] C2 crash when compile negative Arrays.copyOf length after loop

[Dean Long]

On Sun, 13 Jun 2021 12:06:36 GMT, Hui Shi <hshi at openjdk.org> wrote:

> This redos PR https://github.com/openjdk/jdk/pull/4238
> 
> Some closed tests fail due to original PR, the main reasons are:
> - 8268325/8268345
>   - CastIINode created in AllocateArrayNode::make_ideal_length apply on a negative array length.  GVN transform converts these  CastIINode to top. This breaks assumptions, eg., array index node type is TypeInt.
> - 8268345
>   - Some CastIINode is created in LoadRange::Ideal, which doesn't apply GVN transform. CastIINode can not be found in GVN hashtable and cause assertion failure.
> 
> Based on previous PR https://github.com/openjdk/jdk/pull/4238, this PR unifies CastIINode processing post AllocateArrayNode::make_ideal_length. Setup type and add into igvn worklist. Avoid apply GVN transform to break different assumptions in parser.
> 
> Based on closed test failrue message, two new tests are added
> 1. TestNegArrayLengthAsIndex1 : fail in debug build even without previous PR https://github.com/openjdk/jdk/pull/4238, it loads a negative array length and apply GVN transform (In GraphKit::load_array_length) which covert length to top node. It also assert in Parse::array_addressing.
> 
> 2. TestNegArrayLengthAsIndex2 : similar with failrue in https://bugs.openjdk.java.net/browse/JDK-8268301

@huishi-hs I submitted a test job for your changes.  Also, I noticed your branch seems to be based off of "jdk" (jdk18).  There is now a separate jdk17 repo.   Unless you plan to push your fix to 18 and than backport to 17, it might be easier to base your PR on jdk17.

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

PR: https://git.openjdk.java.net/jdk/pull/4480