RFR: 8351833: Unexpected increase in live nodes when splitting Phis through MergeMems in PhiNode::Ideal

[Daniel Lundén]

On Mon, 7 Apr 2025 12:03:29 GMT, Roberto Castañeda Lozano <rcastanedalo at openjdk.org> wrote:

> it would be good, for completeness, to study how many (if any) of those bailouts in Renaissance, SPECjvm, and SPECjbb are due to excessive IGVN node counts, at least for the "baseline" and "target" configurations.

I investigated and it turns out there is one IGVN node count bailout in "target" and zero in "baseline". The bailed out compilation maxes out at 81000 nodes (just above the bailout limit) and looks related to the original failure that resulted in this issue:
- Original compilation failure: `com.sun.org.apache.xerces.internal.impl.dtd.XMLDTDValidator::reset`
- Compilation bailout (in SPECjvm 2008): `com.sun.org.apache.xerces.internal.impl.XMLNSDocumentScannerImpl::reset`

Given that it is such a rare occurrence and not significant in practice, I suggest we investigate if we can avoid the bailout in a separate RFE. I and @robcasloz had a discussion offline, and it looks like the general problem is that we sometimes have far too many alias categories in the memory graph. We should perhaps limit the number of alias categories in some way, or simply retry compilation with aliasing disabled if things get out of hand.

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

PR Comment: https://git.openjdk.org/jdk/pull/24325#issuecomment-2792498724