WIP: Implement vectorized l2m on PPC #7909
Conversation
midronij
changed the title
midronij
force-pushed
the
l2m
branch
2 times, most recently
from
1107bfb to
58fb965
Compare
midronij
changed the title
midronij
changed the title
midronij
changed the title
|
|
||
| // move to VRF | ||
| generateTrg1Src1Instruction(cg, TR::InstOpCode::mtvsrd, node, dstReg, srcReg); | ||
|
|
There was a problem hiding this comment.
why not pre-request that the in-coming double-word is in the right byte order for both LE and BE? and, it can be done very cheaply without heavy lifting below. ld and ldbrx instructions come to mind respectively for BE & LE.
There was a problem hiding this comment.
This is a very good point. *2m and m2* are a bit tricky opcodes. Essentially, they are needed for reading/storing a mask from/to a boolean array. We can think if we can combine them with the actual load/store from/to the array but I am sure a few details will need to be worked out.
There was a problem hiding this comment.
Of course, we can also apply the optimization above if the load is available as the child and has reference count of 1 (as we often do).
There was a problem hiding this comment.
@zl-wang please let us know what you think about treating it as an optimization (see above) ?
There was a problem hiding this comment.
yes, i agree with the conclusion that this is essentially an codegen optimization, peeking into the children trees in order to decide what best-performing instructions to generate.
There was a problem hiding this comment.
@zl-wang @gita-omr I've implemented this optimization for the case where the refcount of the child lload node is 1, and it seems to work without any issues. However, as I understand it, the fix is not quite as simple for the case where refcount is greater than 1, which is less likely to occur but certainly still something we need to take into account.
Earlier we discussed the possibility of essentially un-commoning the lload node, and simply using ldbrx to get the input boolean array without setting the register, but as I understand it, this is a somewhat risky approach to take, and there isn't any past precedent for it anywhere else in the codebase. As well, since there is a way to reverse byte order in a register on P9 and higher (using the xxbrd instruction), the un-commoning approach would only really be relevant to P8 and below.
Since it seems like the conditions in which the un-comming approach would actually be used (refcount >1, P8 and lower) are pretty narrow, is this something we want to pursue? Or alternatively, in the interest of getting these changes merged but still making sure we avoid that cumbersome multi-instruction sequence to manually reverse the element order of the boolean array, maybe it's something we want to add in a different PR?
There was a problem hiding this comment.
I agree thatl2mopcode (and similar) are always used with the corresponding scalar load. It's very unlikely that the load is commoned but l2m is not. So it's a very rare situation that, even if addressed, better to be handled in a separate PR.
midronij
force-pushed
the
l2m
branch
2 times, most recently
from
b6f49ad to
ad5dc2b
Compare
midronij
force-pushed
the
l2m
branch
3 times, most recently
from
84e7eb9 to
9560324
Compare
midronij
force-pushed
the
l2m
branch
2 times, most recently
from
32aff9c to
50233d1
Compare
Signed-off-by: midronij <[email protected]>
| // Case (1) | ||
| if (cg->comp()->target().cpu.isLittleEndian() && child->getReferenceCount() == 1 && child->getRegister() == NULL) { | ||
| srcReg = cg->allocateRegister(); | ||
| TR::LoadStoreHandler::generateLoadNodeSequence(cg, srcReg, child, TR::InstOpCode::ldbrx, 8, true); |
There was a problem hiding this comment.
you don't need to test if the child node is definitely a memory (load or store) operation? could it be an lregload already, for example?
| generateTrg1Src2Instruction(cg, TR::InstOpCode::vsubuhm, node, dstReg, tmpReg, dstReg); | ||
|
|
||
| cg->stopUsingRegister(tmpReg); | ||
| cg->decReferenceCount(child); |
There was a problem hiding this comment.
since srcReg is possibly not set on child node, decReferenceCount on child node doesn't provide the functionality of managing its liveness. so, you might need to do it here.
midronij
changed the title
4 participants
Implement PPC codegen for
l2m(Long to Mask) on P8+. This operation accepts eight byte elements of a given boolean array (read from memory using a doubleword load) and converts it into a ShortVector mask with the corresponding boolean values.