WIP: Implement vectorized m2i on PPC #7941
Conversation
midronij
changed the title
|
|
||
| // set all but least significant bit of each word element to 0 | ||
| generateTrg1ImmInstruction(cg, TR::InstOpCode::vspltisw, node, tmpReg, 1); | ||
| generateTrg1Src2Instruction(cg, TR::InstOpCode::vand, node, tmpReg, srcReg, tmpReg); |
There was a problem hiding this comment.
- isn't java vector mask defined to be in that format already? (i.e. each word element is either 0 or 1)
- code below appeared to assume in-coming element order is not right for LE but right for BE. that is an odd assumption.
such that, i questioned the necessity of code below to flip elements around.
There was a problem hiding this comment.
1.) Vector masks elements are represented with either -1 (i.e.: all 1 bits) for true or 0 for false, while boolean values use 1 for true and 0 for false. So, it's because of this difference that the two instructions highlighted above are needed (just like how with i2m, we need to subtract from 0 to get an all 1/all 0 mask from the given input boolean values)
2.) I think m2i would be affected by endianness the same way that i2m would be, no? For little endian specifically, because the four boolean values are loaded/stored as a single integer value, the byte order is reversed. So if you don't flip it beforehand, when the mask values are stored back to the boolean array, they end up in the wrong order (e.g.: a mask of { false, true, false, true } will end up as a boolean array of { true, false, true, false}.
Signed-off-by: midronij <[email protected]>
3 participants
Implement PPC codegen for m2i (Mask to Integer) on P8+. This operation accepts an IntVector mask (four elements) and converts it into four elements of a boolean array with the corresponding values.