On 2015-11-12 01:10, John Rose wrote:
On Nov 11, 2015, at 11:53 AM, Claes Redestad <claes.redes...@oracle.com> wrote:
Hi,
following-up on JDK-8141678[1]. So, it appears we want to avoid the fragility
of keeping local copies of Byte/Short/Integer in Wrapper, and instead get the
boxed zeroes lazily, when we actually need them.
It turns out simply fixing Wrapper.zero() would then regress things back a bit,
since java.lang.invoke.LambdaForm itself eagerly generates a number of
LambdaForms and NamedFunctions that touch a couple of Wrapper.zeros. By making
the initialization of these lazy as well we not only avoid regression compared
to JDK-8141678, but further remove another 9 LambdaForms from jigsaw startup
(down from 74 to 65; down to 37 together with JDK-8142334[2]). An unneeded
function (void zero_V) was removed in the process.
Making them lazy is fine, but this change is buggy, due to a large amount of
cut-n-paste.
For example, this insertion looks wrong:
+ private static void createZeroForm(BasicType type) {
+ synchronized (LF_zero) {
+ final int ord = type.ordinal();
+ LambdaForm zeForm = LF_identity[ord]; <<< s.b. LF_zero???
Guilty as charged. I found two places and updated in-place. Maybe should
simply drop the attempt to avoid duplication of effort, since that's
what adds most of the code duplication.
For better maintainability, I think the zero and identity forms should be
created together, not in separate twin code blocks.
I think this is a safer, saner way to inject laziness here:
- private static void createIdentityForms() {
+ // Create LF_zero, LF_identity, etc., for the given type.
+ private static void createIdentityForms(BasicType type) {
It means that groups of LFs get lazily created; if that is tolerable for the
present purpose, it's easier to reason about.
I'll give it a try tomorrow and re-measure.
/Claes
