RFR: 8324724: Add Stub routines for FP16 conversions on aarch64

[Nick Gasson]

On Mon, 5 Feb 2024 09:12:23 GMT, Bhavana Kilambi <bkilambi at openjdk.org> wrote:

> This commit - https://github.com/openjdk/jdk/commit/8cfd74f76afc9e5d50c52104fef9974784718dd4 adds stub routines for FP16 conversions for float/float16 constants on x86 to get an accurate compile time value of the nodes. This task adds similar stub routines for aarch64 as well.
> 
> With this patch, if the inputs to the conversion functions are constants, the stub routines are executed to determine the compile time values of the ConvHF2F/ConvF2HF nodes (in their respective Value() functions) and the ConvHF2F/ConvF2HF nodes are replaced with ConI/ConF nodes. This might help in further compiler optimizations like constant folding.
> 
> The following testcase was used to test the disassembly -
> 
> 
> public class FloatConv {
> 
>     private static final short sconst;
>     private static final float fconst;
> 
>     static {
>         sconst = Short.MAX_VALUE;
>         fconst = Float.MIN_VALUE;
>     }
> 
>     @Benchmark
>     public float hf2f() {
>         return Float.float16ToFloat(sconst);
>     }
> 
>     @Benchmark
>     public short f2hf() {
>         return Float.floatToFloat16(fconst);
>     }
> }
> 
> 
> **Disassembly without patch :**
> 
> **FloatConv.f2hf() :-**
> 
> ...
> ldr   s17, 0x0000ffff918cec80
> fcvt  h16, s17
> smov  x0, v16.h[0]
> ...
> ret
> 
> 
> **FloatConv.hf2f() :-**
> 
> ...
> orr   w11, wzr, #0x7fff
> mov   v16.h[0], w11
> fcvt  s0, h16
> ...
> ret
> 
> 
> **Disassembly with patch :**
> 
> **FloatConv.hf2f() :-**
> 
> ...
> ldr   s0, 0x0000ffffb58ce880
> ...
> ret
> 
> 
> **FloatConv.f2hf() :-**
> 
> ...
> mov   w0, wzr
> ...
> ret
> 
> 
> With this patch, the conversion computation is done well in advance and the ConvHF2F/ConvF2HF nodes are replaced with the ConI/ConF nodes and thus the constant values are just loaded into registers and returned.
> 
> The tests in - "hotspot/jtreg/compiler/intrinsics/float16" pass on both aarch64 and x86.

Marked as reviewed by ngasson (Reviewer).

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

PR Review: https://git.openjdk.org/jdk/pull/17706#pullrequestreview-1862309920