[llvm] [LV] Change loops' interleave count computation (PR #73766)
Nilanjana Basu via llvm-commits
llvm-commits at lists.llvm.org
Wed Dec 6 18:45:50 PST 2023
================
@@ -5737,10 +5743,19 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
// the InterleaveCount as if vscale is '1', although if some information about
// the vector is known (e.g. min vector size), we can make a better decision.
if (BestKnownTC) {
- MaxInterleaveCount =
- std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount);
- // Make sure MaxInterleaveCount is greater than 0.
- MaxInterleaveCount = std::max(1u, MaxInterleaveCount);
+ unsigned EstimatedVF = VF.getKnownMinValue();
+ if (VF.isScalable()) {
+ if (std::optional<unsigned> VScale = getVScaleForTuning(L, TTI))
+ EstimatedVF *= *VScale;
+ }
+ if (InterleaveSmallLoopScalarReduction || (*BestKnownTC % EstimatedVF == 0))
----------------
nilanjana87 wrote:
> It might be good to do the check (*BestKnownTC % EstimatedVF == 0) only for known trip counts, as the estimated trip count is just an average and may not be a good indication if the actual trip counts are multiples of the EsimatedVF
Point taken. In the new algorithm for loop IC computation, we always use a conservative IC when the exact TC is not known.
https://github.com/llvm/llvm-project/pull/73766
More information about the llvm-commits
mailing list