Attached src.zip contains some benchmarks for return of value type.
Here are results:
Benchmark Mode Cnt Score Error Units
Ret4.bench_invokeArgInt avgt 5 4.439 ± 0.037 ns/op
Ret4.bench_invokeArgVal_loc0 avgt 5 4.725 ± 0.106 ns/op
Ret4.bench_invokeArgVal_loc1 avgt 5 4.729 ± 0.093 ns/op
Ret4.bench_invokeArgVal_loc2 avgt 5 4.704 ± 0.056 ns/op
Ret4.bench_invokeIF avgt 5 7.349 ± 0.360 ns/op
Ret4.bench_invokeNew avgt 5 4.185 ± 0.083 ns/op
Ret4.bench_invokeSF avgt 5 4.453 ± 0.041 ns/op

Benchmarks cover different behavior:
invokeNew() - return new Value(<const>);
invokeSF() - return <static field>
invokeIF() - return <instance field>
invokeArgInt(int) - return new Value(<arg>);
invokeArgVal(VInteger v) - return <arg>;

All benchmarks behave the same except "invokeIF".
Code generated for "invokeIF" does unnecessary value buffering:

Ret4.bench_invokeArgInt:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op
Ret4.bench_invokeArgVal_loc0:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op
Ret4.bench_invokeArgVal_loc1:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op
Ret4.bench_invokeArgVal_loc2:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op
Ret4.bench_invokeIF:·gc.alloc.rate.norm avgt 5 16.001 ± 0.001 B/op
Ret4.bench_invokeNew:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op
Ret4.bench_invokeSF:·gc.alloc.rate.norm avgt 5 ≈ 10⁻⁶ B/op