RFR(S): 8250609: C2 crash in IfNode::fold_compares

[Vladimir Kozlov]

It happens because 'lo' is new node created just now and have no uses yet.
For such new nodes we usually add dummy use to avoid removal from graph:

http://hg.openjdk.java.net/jdk/jdk/file/c379dc750a02/src/hotspot/share/opto/convertnode.cpp#l403

Thanks,
Vladimir K

On 7/27/20 5:27 AM, Yangfei (Felix) wrote:
> Hi,
> 
> Bug: https://bugs.openjdk.java.net/browse/JDK-8250609
> Webrev: http://cr.openjdk.java.net/~fyang/8250609/webrev.00/
> 
> In IfNode::fold_compares_helper, C2 tries to fold 2 CmpI into a single CmpU.
> At the crash site in IfNode::fold_compares_helper:
>   995   if (lo && hi) {
>   996     // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo))
>   997     Node* adjusted_val = igvn->transform(new SubINode(n,  lo));
>   998     if (adjusted_lim == NULL) {
>   999       adjusted_lim = igvn->transform(new SubINode(hi, lo));
> 1000     }
> 
> At line 997, we have:
> (gdb) p lo->dump()
>   641    AddI    === _  513  92  [[]]
> $1 = void
> 
> After the transformation at line 997, we have
> (gdb) p lo->dump()
>   641    AddI    === _ _ _  [[]]   [34200641]
> $3 = void
> 
> Then node 641 was used at line 999, which triggers the crash.
> Patch fixes the issue by delaying transformation in IfNode::fold_compares temporarily.
> Tier1-3 tested on aarch64-linux-gnu & x86_64-linux-gnu.
> Newly added test fail without the patch and pass otherwise.
> Suggestions?
> 
> Thanks,
> Felix
>