[lworld] RFR: 8359370: [lworld] allow instance fields of identity classes to be readable in the prologue phase [v19]

Maurizio Cimadamore mcimadamore at openjdk.org
Mon Sep 1 12:20:01 UTC 2025


On Mon, 1 Sep 2025 11:42:53 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

> Do we still need the changes in `Attr.checkAssignable` -- e.g. flexible constructor bodies used to check some extra properties in there. Now that we have an LHS variable in the prologue scanner, I wonder if we can unify the checks?

I did an experiment, and adding this:


            if (isInLHS && !insideLambdaOrClassDef) {
                // Check instance field assignments that appear in constructor prologues
                if (isEarlyReference(localEnv, tree.hasTag(SELECT) ? ((JCFieldAccess)tree).selected : null, sym)) {
                    // Field may not be inherited from a superclass
                    if (sym.owner != localEnv.enclClass.sym) {
                        log.error(tree, Errors.CantRefBeforeCtorCalled(sym));
                        return;
                    }

                    // Field may not have an initializer
                    if ((sym.flags() & HASINIT) != 0) {
                        log.error(tree, Errors.CantAssignInitializedBeforeCtorCalled(sym));
                        return;
                    }
                }
                return;
            };
 ```
 
 At the start of `analyzeSymbol` seems to work fine. Maybe we want to break `analyzeSymbol` in different parts, one for reads and one for writes.

But one nice consequence of this is that now `isEarlyReference` is only called inside this visitor -- this means we can probably avoid having to pass `isPrologue = true` and, we can also avoid this dance:


tree.hasTag(SELECT) ? ((JCFieldAccess)tree).selected : null


Which is required at every callsite.

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

PR Comment: https://git.openjdk.org/valhalla/pull/1523#issuecomment-3242145986


More information about the valhalla-dev mailing list