RFR: 8288425: Footprint regression due MH creation when initializing StringConcatFactory

[Paul Sandoz]

On Wed, 15 Jun 2022 21:30:24 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

>> Avoid doing MH creation when initializing `StringConcatFactory` by lazily initializing to a `@Stable` field instead.
>
> src/java.base/share/classes/java/lang/invoke/StringConcatFactory.java line 729:
> 
>> 727:         MethodHandle base = PREPEND_BASE;
>> 728:         if (base == null) {
>> 729:             base = PREPEND_BASE = MethodHandles.dropArguments(
> 
> I'm curious about this idiom - couldn't this result in the `PREPEND_BASE` field being assigned twice, with objects that are morally identical, but could have different identities? (thus violating `@Stable` contract)

It's not a violation, it's relying on behavior not discussed by the `@Stable` doc, which refers to such behavior as "undefined".

It's a carefully correlated dance between the code, interpreter, and C2 relying on:
1. safe publication of equivalent immutable structures with no dependence on their identity when operating on them;
2. the field being set in the interpreter; and 
3. by the time C2 inlines `prependBase` the field is no longer be updated, thus C2 can replace the field access with its value.

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

PR: https://git.openjdk.org/jdk/pull/9154