[LV][EVL] Try to put it close to addActiveLaneMask()

[Mel-Chen]

This work try to move EVL lowering into tryToBuildVPlanWithVPRecipes for removing the following TODO.

       // TODO: try to put it close to addActiveLaneMask().

The ultimate goal of this work is to move addExplicitVectorLength to:

    if (useActiveLaneMask(Style)) {
       ...
    } else if (CM.foldTailWithEVL() && !Plan->hasScalarVFOnly()) {
      VPlanTransforms::runPass(VPlanTransforms::addExplicitVectorLength, *Plan,
                              CM.getMaxSafeElements());
    }

As an intermediate step, it will first be moved to the beginning of the following for loop:

    for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) {
      VFRange SubRange = {VF, MaxVFTimes2};
      if (auto Plan = tryToBuildVPlanWithVPRecipes(
              std::unique_ptr<VPlan>(VPlan0->duplicate()), SubRange, &LVer)) {

        /* Move  VPlanTransforms::addExplicitVectorLength to  HERE*/

        bool HasScalarVF = Plan->hasScalarVFOnly();
        // Now optimize the initial VPlan.
        if (!HasScalarVF)
          VPlanTransforms::runPass(VPlanTransforms::truncateToMinimalBitwidths,
                                   *Plan, CM.getMinimalBitwidths());
        //VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);
        VPlanTransforms::runPass(VPlanTransforms::optimize1, *Plan);
        // TODO: try to put it close to addActiveLaneMask().
        if (CM.foldTailWithEVL() && !HasScalarVF)
          VPlanTransforms::runPass(VPlanTransforms::addExplicitVectorLength,
                                   *Plan, CM.getMaxSafeElements());
        VPlanTransforms::runPass(VPlanTransforms::optimize2, *Plan);
        assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");
        VPlans.push_back(std::move(Plan));
      }
      VF = SubRange.End;
    }

and then finally into tryToBuildVPlanWithVPRecipes.

cc. @alexey-bataev @lukel97 @arcbbb @preames @LiqinWeng

[lukel97]

@Mel-Chen have you already started this? I've also got a patch locally for this that I was hoping to submit after #152542

[lukel97]

One thing I've noticed is that after moving the transform to variable stepping, some places need to be changed to use VPEVLBasedIVPhiRecipe instead of Plan.getCanonicalIV(), e.g. legalizeAndOptimizeInductions.

I think we need to come up with a different name for this, e.g. Plan.getCurrentTrip() or Plan.getCurrentIteration(). Since it's not accurate to call it an induction variable anymore given that the step changes.

[Mel-Chen]

@Mel-Chen have you already started this? I've also got a patch locally for this that I was hoping to submit after #152542

Haha, we are already duplicated work.:D
I’m just using bits of free time to move ::optimize after addExplicitVectorLength, and I'm still here. c083a67

One thing I've noticed is that after moving the transform to variable stepping, some places need to be changed to use VPEVLBasedIVPhiRecipe instead of Plan.getCanonicalIV(), e.g. legalizeAndOptimizeInductions.

Yes, that’s indeed one issue. Another one is that tryToBuildVPlanWithVPRecipes calls verifyVPlanIsValid(*Plan) at the end to validate the plan, but currently the EVL doesn’t dominate VPWidenInduction. I do remember I’ve suggested before that maybe turning EVL into a phi form (I forgot where I comment it) could help with this.