[code-reflection] Integrated: Cache kernel call sites based on lambdaOp location

[Gary Frost]

On Wed, 24 Dec 2025 13:04:01 GMT, Gary Frost <gfrost at openjdk.org> wrote:

> Recently discovered a bug with MINIMIZE_COPIES turned on (Game of life) 
> 
> First time we saw the compute dispatch we located the compute entrypoint which  had no pre/post access.
> 
> We then created compute and kernel callgraphs and injected pre/post mutate calls in the whole compute callgraph (including the lambda) we were processing. 
> 
> On the first call the kernel executed just fine.
> 
> On subsequent compute calls we tried to crack the compute lambda again...
> This time we tripped over the now 'mutated' lambda (with pre/post access mutations) 
> 
> Turned out that we were doing more work than needed on each dispatch. 
> 
> So now  we cache the callsite and only look through the model to crack the lambda and process the callgraphs  once.
> 
> So this turned out to be not only more correct (and stable) it also brought a not insignificant perf gain.

This pull request has now been integrated.

Changeset: 2ac0eff7
Author:    Gary Frost <gfrost at openjdk.org>
URL:       https://git.openjdk.org/babylon/commit/2ac0eff7345088426dd22ac577cedb9b3ae7753c
Stats:     66 lines in 6 files changed: 13 ins; 30 del; 23 mod

Cache kernel call sites based on lambdaOp location

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

PR: https://git.openjdk.org/babylon/pull/786