RFR: JDK-8325264: two compiler/intrinsics/float16 tests fail after JDK-8324724 [v2]

[Vladimir Kozlov]

On Thu, 8 Feb 2024 17:36:15 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> Bhavana Kilambi has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   Make assert the first statement in the function
>
> Good.
> 
> C2 can call `StubRoutines::f2hf()` and `hf2f()` when the value is constant to get constant value.
> Usually C2 executed C++ code from `SharedRuntime` for such cases.
> But to get the same result in Interpreter, C1 and C2 I added call to stub  [JDK-8302976](https://bugs.openjdk.org/browse/JDK-8302976).
> An other solution would be just remove this optimization so that only compiled code will call these stubs. As I remember this is the only place it is called from runtime code.

> @vnkozlov Thank you for the review and feedback comments. If we make changes so that the compiled code calls these stubs then the Conv* nodes wouldn't get reduced to Con* nodes early on (when their Value() functions are called) right? This could help with more compiler optimizations further down the lane when these nodes are used..

Your fix is perfectly fine.

I was just grumbling that if we did not have this optimization we would not need this fix.
Yes, the optimization helps to simplify C2's IR graph.

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

PR Comment: https://git.openjdk.org/jdk/pull/17733#issuecomment-1936269771