[llvm] [VPlan] Implement VPlan-based unit-strideness speculation (PR #182595)

[Andrei Elovikov via llvm-commits]

================
@@ -5521,6 +5531,184 @@ void VPlanTransforms::makeMemOpWideningDecisions(VPlan &Plan, VFRange &Range,
                            });
 }
 
+void VPlanTransforms::multiversionForUnitStridedMemOps(
+    VPlan &Plan, VPCostContext &CostCtx, VPRecipeBuilder &RecipeBuilder,
+    VFRange &Range, SmallVectorImpl<VPInstruction *> &MemOps) {
+  ScalarEvolution *SE = CostCtx.PSE.getSE();
+  PredicatedScalarEvolution StrideMVPSE(*SE, const_cast<Loop &>(*CostCtx.L));
+  SCEVUnionPredicate StridePredicates({}, *SE);
+
+  for (VPInstruction *VPI : MemOps) {
+    auto *PtrOp = VPI->getOpcode() == Instruction::Load ? VPI->getOperand(0)
+                                                        : VPI->getOperand(1);
+
+    const SCEV *PtrSCEV =
+        vputils::getSCEVExprForVPValue(PtrOp, CostCtx.PSE, CostCtx.L);
+    const SCEV *Start, *Stride;
+
+    if (!match(PtrSCEV, m_scev_AffineAddRec(m_SCEV(Start), m_SCEV(Stride),
+                                            m_SpecificLoop(CostCtx.L))))
+      continue;
+
+    Type *ScalarTy = VPI->getOpcode() == Instruction::Load
+                         ? VPI->getScalarType()
+                         : VPI->getOperand(0)->getScalarType();
+
+    if (VPI->getMask()) {
+      Instruction *I = VPI->getUnderlyingInstr();
+      bool IsLoad = VPI->getOpcode() == Instruction::Load;
+      if (!LoopVectorizationPlanner::getDecisionAndClampRange(
+              [&](ElementCount VF) -> bool {
+                return CostCtx.Config.isLegalMaskedLoadOrStore(
+                    IsLoad, ScalarTy, getLoadStoreAlignment(I),
+                    getLoadStoreAddressSpace(I));
+              },
+              Range))
+        continue;
+    }
+
----------------
eas wrote:

> Can we add a comment to explain that. 

Sure, will do later today.

> Also since this will clamp the VF, should we move this down further after the SCEV checks?

Can you explain what's the benefit in doing so?

https://github.com/llvm/llvm-project/pull/182595