[lworld+fp16] RFR: 8341414: Add support for FP16 conversion routines [v2]

[Jatin Bhateja]

On Thu, 14 Nov 2024 09:41:14 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Bhavana Kilambi has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   Remove intrinsification of conversion methods in Float16
>
> src/hotspot/share/opto/cfgnode.cpp line 1907:
> 
>> 1905:     } else if (convert_op == Op_ConvHF2F || convert_op == Op_ConvF2HF || convert_op == Op_ConvHF2D || convert_op == Op_ConvHF2L) {
>> 1906:       return T_SHORT;
>> 1907:     }
> 
> This transformation is moving conversion across the Phi node thereby saving expensive conversion instructions.
> 
> `Phi (ConvHF2F src1) (ConvHF2F src2) =>   ConvHF2F  (Phi src1  src2)
> `
> 
> For the Op_ConvF2HF case, type parameter will be of float type and thus we will never enter control flow between L1905 - L1907. 
> 
> I think it will be harmless if we remove convert_op == Op_ConvF2HF check...

We can also update existing test for new conversion

test/hotspot/jtreg/compiler/c2/irTests/TestPhiDuplicatedConversion.java

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

PR Review Comment: https://git.openjdk.org/valhalla/pull/1283#discussion_r1841879971