[VPlan] Don't reset canonical IV start value.

Instead of re-setting the start value of the canonical IV when
vectorizing the epilogue we can emit an Add VPInstruction to provide
canonical IV value, adjusted by the resume value from the main loop.

This is in preparation to make the canonical IV a VPValue defined by
loop regions. It ensures that the canonical IV always starts at 0.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -9549,13 +9549,12 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
   VPBasicBlock *Header = VectorLoop->getEntryBasicBlock();
   Header->setName("vec.epilog.vector.body");
 
-  // Ensure that the start values for all header phi recipes are updated before
-  // vectorizing the epilogue loop.
   VPCanonicalIVPHIRecipe *IV = Plan.getCanonicalIV();
-  // When vectorizing the epilogue loop, the canonical induction start
-  // value needs to be changed from zero to the value after the main
-  // vector loop. Find the resume value created during execution of the main
-  // VPlan. It must be the first phi in the loop preheader.
+  // When vectorizing the epilogue loop, the canonical induction needs to be
+  // adjusted by the value after the main vector loop. Find the resume value
+  // created during execution of the main VPlan. It must be the first phi in the
+  // loop preheader. Use the value to increment the canonical IV, and update all
+  // users in the loop region to use the adjusted value.
   // FIXME: Improve modeling for canonical IV start values in the epilogue
   // loop.
   using namespace llvm::PatternMatch;
@@ -9590,10 +9589,16 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
                 }) &&
          "the canonical IV should only be used by its increment or "
          "ScalarIVSteps when resetting the start value");
-  IV->setOperand(0, VPV);
+  VPBuilder Builder(Header, Header->getFirstNonPhi());
+  VPInstruction *Add = Builder.createNaryOp(Instruction::Add, {IV, VPV});
+  IV->replaceAllUsesWith(Add);
+  Add->setOperand(0, IV);
 
   DenseMap<Value *, Value *> ToFrozen;
   SmallVector<Instruction *> InstsToMove;
+  // Ensure that the start values for all header phi recipes are updated before
+  // vectorizing the epilogue loop. Skip the canonical IV, which has been
+  // handled above.
   for (VPRecipeBase &R : drop_begin(Header->phis())) {
     Value *ResumeV = nullptr;
     // TODO: Move setting of resume values to prepareToExecute.

llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h

@@ -431,6 +431,12 @@ m_c_Binary(const Op0_t &Op0, const Op1_t &Op1) {
   return AllRecipe_commutative_match<Opcode, Op0_t, Op1_t>(Op0, Op1);
 }
 
+template <typename Op0_t, typename Op1_t>
+inline AllRecipe_match<Instruction::Add, Op0_t, Op1_t> m_Add(const Op0_t &Op0,
+                                                             const Op1_t &Op1) {
+  return m_Binary<Instruction::Add, Op0_t, Op1_t>(Op0, Op1);
+}
+
 template <typename Op0_t, typename Op1_t>
 inline AllRecipe_commutative_match<Instruction::Add, Op0_t, Op1_t>
 m_c_Add(const Op0_t &Op0, const Op1_t &Op1) {

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1234,6 +1234,17 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
   if (!Plan->isUnrolled())
     return;
 
+  if (match(Def, m_Add(m_VPValue(X), m_VPValue(Y))) && Y->isLiveIn() &&
+      isa<VPPhi>(X)) {
+    auto *Phi = cast<VPPhi>(X);
+    if (Phi->getOperand(1) != Def && match(Phi->getOperand(0), m_ZeroInt()) &&
+        Phi->getNumUsers() == 1 && (*Phi->user_begin() == &R)) {
+      Phi->setOperand(0, Y);
+      Def->replaceAllUsesWith(Phi);
+      return;
+    }
+  }
+
   // VPVectorPointer for part 0 can be replaced by their start pointer.
   if (auto *VecPtr = dyn_cast<VPVectorPointerRecipe>(&R)) {
     if (VecPtr->isFirstPart()) {