Support v2m opcode on Power

[gita-omr]

• TR::v2m opcode converts boolean vector into mask
• least-significant bit of each byte element should be left-extended
• mload/mloadi opcodes don't have to extend the bit since they load correct
  mask representation stored by mstore/mstorei

[dsouzai]

jenkins build plinux,aix

[BradleyWood]

on Power, vectors and masks reside in the same register it can be a passThrough evaluator

That doesn't sound right to me. This opcode is is meant for converting a mask, in the form a boolean array (when 16+ lanes are present) into a mask. The mask (as a vector) cannot be in the form of a boolean array. Each lane should be all 1's or all 0's. Furthermore, the number of boolean[] elements at 8-bit in a 128-bit vector is 16. What if the elements in your mask operations are 16-bit? For instance, vmaddVector256Int16, it would have 16 elements, therefore, the vector you load is Vector128Int8 -- half the size, and incorrect lane type.

[gita-omr]

Made WIP to clarify the definition of v2m opcode.

[gita-omr]

I agree with some points in #7737 (comment) but I think some don't apply to Power.

v2m is indeed meant for loading a mask from a boolean array. It is only needed when the size of the array is bigger than 8 bytes. In these cases, we use a vector load of the right size to load the whole array. Since vector element cannot be smaller than boolean there is always a vector load available on a given platform to load a full boolean array representing a mask.

For example, since vector length on Power is 128-bit, the biggest mask will contain 16 elements (if vector elements are bytes). So we will use vloadiVector128Int8 for a byte mask and we'll use other opcodes for smaller masks (iloadi, lloadi, etc). So I think there is no size issue on Power.

Then, vloadiVector128Int8 becomes a child of v2m to convert the vector containing a boolean array into a mask. I agree that the least significant bit needs to be left-extended by v2m since boolean array might contain value 1 for true, but we need all bits in the lane to be set. We already have such code for mload on Power. But I think it's not necessary there and needs to be moved to v2m.

mload/mloadi does not need the bit to be extended since it assumes that proper mask was stored by mstore/mstorei.

[gita-omr]

Changed code and commit comment according to the comment above.

[gita-omr]

Removed the extension code in vloadEvaluator.

[gita-omr]

Improved commit message.

[gita-omr]

Ready for another review. @BradleyWood @dsouzai

[dsouzai]

Looks OK to me; will wait for @BradleyWood's OK.

[dsouzai]

OMR Jenkins instance is still down; could you run an internal build (or run the OMR tests on a P machine manually) to make sure there's no issues?

[dsouzai]

jenkins build plinux,aix