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

[Luke Lau via llvm-commits]

================
@@ -5543,6 +5552,170 @@ void VPlanTransforms::makeMemOpWideningDecisions(VPlan &Plan, VFRange &Range,
                            });
 }
 
+void VPlanTransforms::multiversionForUnitStridedMemOps(
+    VPlan &Plan, VPCostContext &CostCtx, VFRange &Range,
+    ArrayRef<VPInstruction *> MemOps) {
+  ScalarEvolution *SE = CostCtx.PSE.getSE();
+  SCEVUnionPredicate StridePredicates({}, *SE);
+
+  for (VPInstruction *VPI : MemOps) {
+    VPValue *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;
+      // Don't speculate unit-strideness if it won't result in any unit-strided
+      // loads, as we'd pay the price of not taking vector loop if the runtime
+      // condition is false for no benefits.
+      if (!CostCtx.Config.isLegalMaskedLoadOrStore(IsLoad, ScalarTy,
+                                                   getLoadStoreAlignment(I),
+                                                   getLoadStoreAddressSpace(I)))
+        continue;
+    }
+
+    if (isa<SCEVConstant>(Stride))
+      continue;
+
+    const auto *TypeSize = cast<SCEVConstant>(SE->getSizeOfExpr(
+        Stride->getType(), SE->getDataLayout().getTypeAllocSize(ScalarTy)));
+
+    const SCEVConstant *StrideConstantMultiplier;
+    const SCEV *StrideNonConstantMultiplier;
+
+    const SCEV *ToMultiVersion = Stride;
+    const SCEV *MVConst = TypeSize;
+    if (match(Stride, m_scev_c_Mul(m_SCEVConstant(StrideConstantMultiplier),
+                                   m_SCEV(StrideNonConstantMultiplier)))) {
+      if (TypeSize != StrideConstantMultiplier) {
+        // TODO: Support `TypeSize = N * StrideConstantMultiplier`,
+        // including negative `N`. For now, only process when they're equal,
+        // which matches the useful part of the legacy behavior that
+        // multiversiones GEP index for stride one.
+        continue;
+      }
+      ToMultiVersion = StrideNonConstantMultiplier;
+      MVConst = SE->getOne(ToMultiVersion->getType());
+    } else if (!TypeSize->isOne()) {
+      // Likewise - try to match legacy behavior.
+      continue;
+    }
+
+    while (auto *C = dyn_cast<SCEVIntegralCastExpr>(ToMultiVersion)) {
+      ToMultiVersion = C->getOperand();
+      MVConst = SE->getTruncateOrSignExtend(MVConst, ToMultiVersion->getType());
+    }
+
+    if (match(ToMultiVersion, m_scev_UndefOrPoison()))
+      continue;
+
+    if (!isa<SCEVUnknown>(ToMultiVersion)) {
+      // Match legacy behavior.
+      // If/when changed, make sure that explicit poison/undef in the defining
+      // expression doesn't cause any issues.
+      continue;
+    }
+
+    Value *StrideVal = cast<SCEVUnknown>(ToMultiVersion)->getValue();
+
+    const SCEVPredicate *NewPred =
+        SE->getComparePredicate(CmpInst::ICMP_EQ, ToMultiVersion, MVConst);
+
+    // Check if new predicate implies that vectorizing a loop of such trip count
+    // is meaningless. The heuristic here could probably be improved.
+    auto *PredicatedMaxBTC = SE->rewriteUsingPredicate(
+        SE->getSymbolicMaxBackedgeTakenCount(CostCtx.L), CostCtx.L,
+        StridePredicates.getUnionWith(NewPred, *SE)
+            .getUnionWith(&CostCtx.PSE.getPredicate(), *SE));
+
+    if (LoopVectorizationPlanner::getDecisionAndClampRange(
+            [&](ElementCount VF) {
+              return SE->isKnownPredicate(
+                  ICmpInst::ICMP_ULT, PredicatedMaxBTC,
+                  SE->getConstant(PredicatedMaxBTC->getType(),
+                                  Plan.hasTailFolded() || VF.isScalable()
+                                      ? 1
+                                      : VF.getFixedValue() - 1));
+            },
+            Range))
+      continue;
+
+    StridePredicates = StridePredicates.getUnionWith(NewPred, *SE);
+
+    auto ReplaceMVUses = [&](Value *V) {
+      VPValue *From = Plan.getLiveIn(V);
+      if (!From)
+        return;
+      VPValue *To = Plan.getConstantInt(
+          From->getScalarType(),
+          cast<SCEVConstant>(MVConst)->getAPInt().getLimitedValue());
+
+      // Original scalar loop can still use `From`, make sure to only rewrite
+      // uses inside the vector loop that we guard with the checks.
+      From->replaceUsesWithIf(To, [&](VPUser &U, unsigned) {
+        auto *R = cast<VPRecipeBase>(&U);
+        return R->getRegion() ||
+               R->getParent() ==
+                   Plan.getVectorLoopRegion()->getSinglePredecessor();
+      });
+    };
+
+    ReplaceMVUses(StrideVal);
+    for (auto *U : StrideVal->users())
+      if (isa<SExtInst, ZExtInst, TruncInst>(U))
+        ReplaceMVUses(U);
----------------
lukel97 wrote:

This is an edge case, but I don't think it's correct to RAUW a sext of i1 with 1, e.g. this test case:

```llvm
define void @i1_stride_with_sext_use(ptr noalias %p.out, ptr %p, i1 %b) {
entry:
  %stride = zext i1 %b to i64
  %e = sext i1 %b to i32
  br label %loop

loop:
  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
  %mul = mul i64 %iv, %stride
  %gep = getelementptr i32, ptr %p, i64 %mul
  %v = load i32, ptr %gep, align 4
  %add = add i32 %v, %e ; GETS RAUW'd TO i32 1, SHOULD TO BE i32 -1
  %gep.out = getelementptr i32, ptr %p.out, i64 %iv
  store i32 %add, ptr %gep.out, align 4
  %iv.next = add i64 %iv, 1
  %ec = icmp eq i64 %iv.next, 128
  br i1 %ec, label %exit, label %loop

exit:
  ret void
}
```

I think we need to adjust the lambda to be something like

```c++
auto ReplaceUsesInVectorLoop = [&](Value *From, VPValue *To) { ... }
```

And then in the users() loop construct the appropriate sext/zext/trunc VPValue based off of MVConst. 

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