On Wed, 18 May 2022 14:51:02 GMT, Jorn Vernee <jver...@openjdk.org> wrote:
>> I wasn't suggesting to add more bindings. I was more suggesting to filter
>> out the load/store from the set of bindings (since these are virtualized
>> anyways) that are passed to `doBindings`. Then, before executing a set of
>> bindings, (if we are in downcall mode) we load the corresponding input local
>> var. After executing bindings (if we are in upcall mode) we store result in
>> corresponding var.
>>
>> E.g. make the logic that load locals and store locals explicit in the
>> `specialize` method, rather than have parts of it execute "in disguise" as
>> "binding interpretation".
>
> It's not quite that simple since a binding recipe for a single parameter can
> have multiple VMStores for instance if a struct is decomposed into multiple
> values.
>
> It can be done by pulling the binding loops up to the `specialize` method,
> and have if statements for VMStore and VMLoad, like this:
>
> for (Binding binding : callingSequence.argumentBindings(i)) {
> if (binding instanceof Binding.VMStore vms &&
> callingSequence.forDowncall()) {
> emitSetOutput(vms.type());
> } else if (binding instanceof Binding.VMLoad &&
> callingSequence.forUpcall()) {
> emitGetInput();
> } else {
> doBinding(binding);
> }
> }
>
> And for returns:
>
> for (Binding binding : callingSequence.returnBindings()) {
> if (binding instanceof Binding.VMLoad vml &&
> callingSequence.forDowncall()) {
> if (!callingSequence.needsReturnBuffer()) {
> emitRestoreReturnValue(vml.type());
> } else {
> emitReturnBufferLoad(vml);
> }
> } else if (binding instanceof Binding.VMStore vms &&
> callingSequence.forUpcall()) {
> if (!callingSequence.needsReturnBuffer()) {
> emitSaveReturnValue(vms.type());
> } else {
> emitReturnBufferStore(vms);
> }
> } else {
> doBinding(binding);
> }
> }
>
> But, maybe that's better?
I think someone might look at this and think "why aren't these bindings handled
by `doBinding`"?
-------------
PR: https://git.openjdk.java.net/jdk/pull/8685
