On Mon, 19 Sep 2022 08:51:24 GMT, Xiaohong Gong <xg...@openjdk.org> wrote:
> "`VectorSupport.indexVector()`" is used to compute a vector that contains the
> index values based on a given vector and a scale value (`i.e. index = vec +
> iota * scale`). This function is widely used in other APIs like
> "`VectorMask.indexInRange`" which is useful to the tail loop vectorization.
> And it can be easily implemented with the vector instructions.
>
> This patch adds the vector intrinsic implementation of it. The steps are:
>
> 1) Load the const "iota" vector.
>
> We extend the "`vector_iota_indices`" stubs from byte to other integral
> types. For floating point vectors, it needs an additional vector cast to get
> the right iota values.
>
> 2) Compute indexes with "`vec + iota * scale`"
>
> Here is the performance result to the new added micro benchmark on ARM NEON:
>
> Benchmark Gain
> IndexVectorBenchmark.byteIndexVector 1.477
> IndexVectorBenchmark.doubleIndexVector 5.031
> IndexVectorBenchmark.floatIndexVector 5.342
> IndexVectorBenchmark.intIndexVector 5.529
> IndexVectorBenchmark.longIndexVector 3.177
> IndexVectorBenchmark.shortIndexVector 5.841
>
>
> Please help to review and share the feedback! Thanks in advance!
src/hotspot/share/opto/vectorIntrinsics.cpp line 2949:
> 2947: } else if (elem_bt == T_DOUBLE) {
> 2948: iota = gvn().transform(new VectorCastL2XNode(iota, vt));
> 2949: }
Since we are loading constants from stub initialized memory locations, defining
new stubs for floating point iota indices may eliminate need for costly
conversion instructions. Specially on X86 conversion b/w Long and Double is
only supported by AVX512DQ targets and intrinsification may fail for legacy
targets.
src/hotspot/share/opto/vectorIntrinsics.cpp line 2978:
> 2976: case T_DOUBLE: {
> 2977: scale = gvn().transform(new ConvI2LNode(scale));
> 2978: scale = gvn().transform(new ConvL2DNode(scale));
Prior target support check for these IR nodes may prevent surprises in the
backend.
src/hotspot/share/opto/vectorIntrinsics.cpp line 2978:
> 2976: case T_DOUBLE: {
> 2977: scale = gvn().transform(new ConvI2LNode(scale));
> 2978: scale = gvn().transform(new ConvL2DNode(scale));
Any specific reason for not directly using ConvI2D for double case.
-------------
PR: https://git.openjdk.org/jdk/pull/10332
