RFR: 8275317: AArch64: Support some type conversion vectorization in SLP

[Tobias Hartmann]

On Thu, 28 Oct 2021 03:39:42 GMT, Fei Gao <duke at openjdk.java.net> wrote:

> Current SLP vectorizer in C2 compiler doesn't support type conversion
> operations. But AArch64 has vector type conversion instructions in
> both NEON and SVE.
> 
> The type conversion involves two kinds of scenarios, conversion between
> the same data sizes and conversion between different data sizes. If we
> want to support casts between different data sizes, we need to amend
> the code part for identifying adjacent memory references and the code
> part for justifying if the combination is profitable. I suppose it
> would be easier to review if we split the whole task to support type
> conversion into two separate patches, one for the same data sizes and
> the other one for different data sizes. The goal of this patch is just
> to support conversions within the same data size, including:
>   int -> float
>   float -> int
>   long -> double
>   double -> long
> 
> A typical test case:
> 
> for (int i = start; i < limit; i++) {
>     b[i] = (float) a[i];
> }
> 
> To implement it, the patch completed the necessary instructions and
> matching rules in the backend and added implementation for SLP in
> the middle end.
> 
> The percentage of performance uplift on aarch64 system:
> Mode: avgt
> Cnt: 15
> Metric: (ns/op)
> 
> benchmark                       percentage change  [(After-Before)/Before] 
> VectorLoop.convertD2L      -48.46%	                 
> VectorLoop.convertF2I        -55.67%
> VectorLoop.convertI2F        -55.27%
> VectorLoop.convertL2D       -48.75%

That looks good to me but x86 supports vector instructions for these operations as well, right? Or am I missing something?

https://github.com/openjdk/jdk/blob/55b36c6f3bb7eb066daaf41f9eba46633afedf08/src/hotspot/cpu/x86/x86.ad#L6701

Do you have perf numbers for x86?

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

PR: https://git.openjdk.java.net/jdk/pull/6145