RFR: 8333258: C2: high memory usage in PhaseCFG::insert_anti_dependences()

[Emanuel Peter]

On Wed, 19 Jun 2024 12:54:26 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> In a debug build, `PhaseCFG::insert_anti_dependences()` is called
> twice for a single node: once for actual processing, once for
> verification.
> 
> In TestAntiDependenciesHighMemUsage, the test has a `Region` that
> merges 337 incoming path. It also has one `Phi` per memory slice that
> are stored to: 1000 `Phi` nodes. Each `Phi` node has 337 inputs that
> are identical except for one. The common input is the memory state on
> method entry. The test has 60 `Load` that needs to be processed for
> anti dependences. All `Load` share the same memory input: the memory
> state on method entry. For each `Load`, all `Phi` nodes are pushed 336
> times on the work lists for anti dependence processing because all of
> them appear multiple times as uses of each `Load`s memory state: `Phi`s
> are pushed 336 000 on 2 work lists. Memory is not reclaimed on exit
> from `PhaseCFG::insert_anti_dependences()` so memory usage grows as
> `Load` nodes are processed:
> 
> 336000 * 2 work lists * 60 loads * 8 bytes pointer = 322 MB. 
> 
> The fix I propose for this is to not push `Phi` nodes more than once
> when they have the same inputs multiple times.
> 
> In TestAntiDependenciesHighMemUsage2, the test has 4000 loads. For
> each of them, when processed for anti dependences, all 4000 loads are
> pushed on the work lists because they share the same memory
> input. Then when they are popped from the work list, they are
> discarded because only stores are of interest:
> 
> 4000 loads processed * 4000 loads pushed * 2 work lists * 8 bytes pointer = 256 MB. 
> 
> The fix I propose for this is to test before pushing on the work list
> whether a node is a store or not.
> 
> Finally, I propose adding a `ResourceMark` so memory doesn't
> accumulate over calls to `PhaseCFG::insert_anti_dependences()`.

First a question:
Why does your `already_enqueued` exit early if `worklist_mem.at(j-1) != mem`? Are we sure that there will not be a `mem, store` pair at a lower index `i`?

An idea: this really looks like we are storing a `Pair<mem, store>`. Could we not use a set of some kind, maybe a `ResourceHashtable`? That would allow a possibly faster lookup. Ah, or maybe you can combine the `idx` of `mem` and `store` to a unique combined idx, and use that for a `VectorSet`?

I just think that a hand-crafted `already_enqueued` is not really super nice, especially if it has this early-exit mechanism that is not really predictable from the name.

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

PR Comment: https://git.openjdk.org/jdk/pull/19791#issuecomment-2180581945