[NFC][InstructionCost] Change LoopVectorizationCostModel::getInstructionCost to return InstructionCost

[david-arm]

Summary

This patch is part of a series of patches that migrate integer
instruction costs to use InstructionCost. In the function
selectVectorizationFactor I have simply asserted that the cost
is valid and extracted the value as is. In future we expect
to encounter invalid costs, but we should filter out those
vectorization factors that lead to such invalid costs.

See this patch for the introduction of the type: https://reviews.llvm.org/D91174
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html

[david-arm]

• Rebased off parent patch.

[ctetreau]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

I think you can get rid of these pairs if you just check validity before assigning it.

5692

ForceVectorization needs to not blow up in the face of invalid InstructionCosts

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

OK. I guess the problem is that we could start off with an invalid cost (ScalarCost) and get another Invalid cost in the loop too. I think we could get rid of the pairs, but it just makes the main loop more complicated that's all. What I could do is add this in the main loop:

if (!C.first.isValid())
  continue
if (!ScalarCost.isValid())
  MinCost = C.first;
// Both costs are now valid.
// ... use MinCost and C somehow ...

The other thing I was trying to do is move from using a float to represent the MinCost in the loop, to using an InstructionCost (integer based) instead. If I get rid of the pairs then it makes more sense to revert to using a float for MinCost I think, since then division (and hence a fractional cost) is involved.

I can try adding control flow to the loop like above?

[ctetreau]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

So I messed around with your patch, and came up with this loop body:

for (unsigned i = 2; i <= MaxVF.getFixedValue(); i *= 2) {
    // Notice that the vector loop needs to be executed less times, so
    // we need to divide the cost of the vector loops by the width of
    // the vector elements.
    VectorizationCostTy C = expectedCost(ElementCount::getFixed(i));

    Optional<float> VectorCost =
        C.first.getValue().map([i](InstructionCost::CostType Cost) {
          return static_cast<float>(Cost) / i;
        });

    // removed debug output noise

    if (auto MinCostVal = MinCost.getValue())
      if (VectorCost && *VectorCost < *MinCostVal) {
        Width = i;
        // requires ctor that takes an Optional, and ctors that convert from number like things
        MinCost = VectorCost;
      }
  }

I haven't ran the tests, but this compiles and I think it should work.

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

Hi @ctetreau, your suggestions for adding a conversion constructor/getter in general seem sensible to me, however I just have one question. In that example above it looks like you're rounding a float (VectorCost) to an integer, which is a change in behaviour I think? Currently the examples I'm thinking of are where we compare a cost of 11 for a width of 2, with a cost of 21 for a width of 4, i.e.

float(11)/2 = 5.5
float(21)/4 = 5.25

Previously we'd choose a width of 4 because 5.25 is less than 5.5. I think with your proposal we'd keep a width of 2 because 5.5 gets rounded down to 5. Unless we make MinCost use floats too, in which case the need for conversion goes away I think?

[ctetreau]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

Yeah, looks like I messed up. If you have MinCost be an Optional<float>, it seems like it should be fine. The main idea is that this whole function should be working with Optional<float> instead of InstructionCost. We can convert all InstructionCost objects to Optional<float> right away, do floating point math with them, and then convert back to unsigned for the return.

[david-arm]

• Removed assert in Instruction::ExtractValue case in getInstructionCost.
• Rebase.

[david-arm]

Hi folks, just a gentle ping to see if anyone has chance to have a quick look?

[sdesmalen]

LGTM, but it's not clear to me if @ctetreau's latest comment still needs to be addressed.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5667–5670

nit: ExpectedScalarCost ?

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5688–5689

Hi @ctetreau, I've changed my patch now to avoid worrying about validity of the cost. Since we now assert the cost is valid we no longer need to use Optionals and so on.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

 #include "llvm/IR/Verifier.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Compiler.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/InstructionCost.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/InjectTLIMappings.h"
 #include "llvm/Transforms/Utils/LoopSimplify.h"
 #include "llvm/Transforms/Utils/LoopUtils.h"
 #include "llvm/Transforms/Utils/LoopVersioning.h"
 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
private:
 
   /// The vectorization cost is a combination of the cost itself and a boolean
   /// indicating whether any of the contributing operations will actually
   /// operate on
   /// vector values after type legalization in the backend. If this latter value
   /// is
   /// false, then all operations will be scalarized (i.e. no vectorization has
   /// actually taken place).
-  using VectorizationCostTy = std::pair<unsigned, bool>;
+  using VectorizationCostTy = std::pair<InstructionCost, bool>;
 
   /// Returns the expected execution cost. The unit of the cost does
   /// not matter because we use the 'cost' units to compare different
   /// vector widths. The cost that is returned is *not* normalized by
   /// the factor width.
   VectorizationCostTy expectedCost(ElementCount VF);
 
   /// Returns the execution time cost of an instruction for a given vector
   /// width. Vector width of one means scalar.
   VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF);
 
   /// The cost-computation logic from getInstructionCost which provides
   /// the vector type as an output parameter.
-  unsigned getInstructionCost(Instruction *I, ElementCount VF, Type *&VectorTy);
+  InstructionCost getInstructionCost(Instruction *I, ElementCount VF,
+                                     Type *&VectorTy);
 
   /// Calculate vectorization cost of memory instruction \p I.
   unsigned getMemoryInstructionCost(Instruction *I, ElementCount VF);
 
   /// The cost computation for scalarized memory instruction.
   unsigned getMemInstScalarizationCost(Instruction *I, ElementCount VF);
 
   /// The cost computation for interleaving group of memory instructions.
private:
   /// Map of scalar integer values to the smallest bitwidth they can be legally
   /// represented as. The vector equivalents of these values should be truncated
   /// to this type.
   MapVector<Instruction *, uint64_t> MinBWs;
 
   /// A type representing the costs for instructions if they were to be
   /// scalarized rather than vectorized. The entries are Instruction-Cost
   /// pairs.
-  using ScalarCostsTy = DenseMap<Instruction *, unsigned>;
+  using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>;
 
   /// A set containing all BasicBlocks that are known to present after
   /// vectorization as a predicated block.
   SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization;
 
   /// Records whether it is allowed to have the original scalar loop execute at
   /// least once. This may be needed as a fallback loop in case runtime
   /// aliasing/dependence checks fail, or to handle the tail/remainder

 
 VectorizationFactor
 LoopVectorizationCostModel::selectVectorizationFactor(ElementCount MaxVF) {
   // FIXME: This can be fixed for scalable vectors later, because at this stage
   // the LoopVectorizer will only consider vectorizing a loop with scalable
   // vectors when the loop has a hint to enable vectorization for a given VF.
   assert(!MaxVF.isScalable() && "scalable vectors not yet supported");
 
-  float Cost = expectedCost(ElementCount::getFixed(1)).first;
-  const float ScalarCost = Cost;
+  InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first;
+  LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n");
+  assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop");
+
   unsigned Width = 1;
-  LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
+  const float ScalarCost = *ExpectedCost.getValue();
+  float Cost = ScalarCost;
 
   bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
   if (ForceVectorization && MaxVF.isVector()) {
     // Ignore scalar width, because the user explicitly wants vectorization.
     // Initialize cost to max so that VF = 2 is, at least, chosen during cost
     // evaluation.
     Cost = std::numeric_limits<float>::max();
   }
 
   for (unsigned i = 2; i <= MaxVF.getFixedValue(); i *= 2) {
     // Notice that the vector loop needs to be executed less times, so
     // we need to divide the cost of the vector loops by the width of
     // the vector elements.
     VectorizationCostTy C = expectedCost(ElementCount::getFixed(i));
-    float VectorCost = C.first / (float)i;
+    assert(C.first.isValid() && "Unexpected invalid cost for vector loop");
+    float VectorCost = *C.first.getValue() / (float)i;
     LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << i
                       << " costs: " << (int)VectorCost << ".\n");
     if (!C.second && !ForceVectorization) {
       LLVM_DEBUG(
           dbgs() << "LV: Not considering vector loop of width " << i
                  << " because it will not generate any vector instructions.\n");
       continue;
     }
 
     // If profitable add it to ProfitableVF list.
     if (VectorCost < ScalarCost) {
unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
   else
     // Make sure IC is greater than 0.
     IC = std::max(1u, IC);
 
   assert(IC > 0 && "Interleave count must be greater than 0.");
 
   // If we did not calculate the cost for VF (because the user selected the VF)
   // then we calculate the cost of VF here.
-  if (LoopCost == 0)
-    LoopCost = expectedCost(VF).first;
+  if (LoopCost == 0) {
+    assert(expectedCost(VF).first.isValid() && "Expected a valid cost");
+    LoopCost = *expectedCost(VF).first.getValue();
+  }
 
   assert(LoopCost && "Non-zero loop cost expected");
 
   // Interleave if we vectorized this loop and there is a reduction that could
   // benefit from interleaving.
   if (VF.isVector() && HasReductions) {
     LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
     return IC;
for (Instruction &I : *BB)
           ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
         // Remember that BB will remain after vectorization.
         PredicatedBBsAfterVectorization.insert(BB);
       }
   }
 }
 
 int LoopVectorizationCostModel::computePredInstDiscount(
-    Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
-    ElementCount VF) {
+    Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) {
   assert(!isUniformAfterVectorization(PredInst, VF) &&
          "Instruction marked uniform-after-vectorization will be predicated");
 
   // Initialize the discount to zero, meaning that the scalar version and the
   // vector version cost the same.
-  int Discount = 0;
+  InstructionCost Discount = 0;
 
   // Holds instructions to analyze. The instructions we visit are mapped in
   // ScalarCosts. Those instructions are the ones that would be scalarized if
   // we find that the scalar version costs less.
   SmallVector<Instruction *, 8> Worklist;
 
   // Returns true if the given instruction can be scalarized.
   auto canBeScalarized = [&](Instruction *I) -> bool {
while (!Worklist.empty()) {
     Instruction *I = Worklist.pop_back_val();
 
     // If we've already analyzed the instruction, there's nothing to do.
     if (ScalarCosts.find(I) != ScalarCosts.end())
       continue;
 
     // Compute the cost of the vector instruction. Note that this cost already
     // includes the scalarization overhead of the predicated instruction.
-    unsigned VectorCost = getInstructionCost(I, VF).first;
+    InstructionCost VectorCost = getInstructionCost(I, VF).first;
 
     // Compute the cost of the scalarized instruction. This cost is the cost of
     // the instruction as if it wasn't if-converted and instead remained in the
     // predicated block. We will scale this cost by block probability after
     // computing the scalarization overhead.
     assert(!VF.isScalable() && "scalable vectors not yet supported.");
-    unsigned ScalarCost =
+    InstructionCost ScalarCost =
         VF.getKnownMinValue() *
         getInstructionCost(I, ElementCount::getFixed(1)).first;
 
     // Compute the scalarization overhead of needed insertelement instructions
     // and phi nodes.
     if (isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
       ScalarCost += TTI.getScalarizationOverhead(
           cast<VectorType>(ToVectorTy(I->getType(), VF)),
while (!Worklist.empty()) {
     ScalarCost /= getReciprocalPredBlockProb();
 
     // Compute the discount. A non-negative discount means the vector version
     // of the instruction costs more, and scalarizing would be beneficial.
     Discount += VectorCost - ScalarCost;
     ScalarCosts[I] = ScalarCost;
   }
 
-  return Discount;
+  return *(Discount.getValue());
 }
 
 LoopVectorizationCostModel::VectorizationCostTy
 LoopVectorizationCostModel::expectedCost(ElementCount VF) {
   VectorizationCostTy Cost;
 
   // For each block.
   for (BasicBlock *BB : TheLoop->blocks()) {
     VectorizationCostTy BlockCost;
 
     // For each instruction in the old loop.
     for (Instruction &I : BB->instructionsWithoutDebug()) {
       // Skip ignored values.
       if (ValuesToIgnore.count(&I) ||
           (VF.isVector() && VecValuesToIgnore.count(&I)))
         continue;
 
       VectorizationCostTy C = getInstructionCost(&I, VF);
 
       // Check if we should override the cost.
       if (ForceTargetInstructionCost.getNumOccurrences() > 0)
-        C.first = ForceTargetInstructionCost;
+        C.first = InstructionCost(ForceTargetInstructionCost);
 
       BlockCost.first += C.first;
       BlockCost.second |= C.second;
       LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first
                         << " for VF " << VF << " For instruction: " << I
                         << '\n');
     }
 
if (VF.isVector() && ForcedScalar != ForcedScalars.end()) {
     if (InstSet.count(I))
       return VectorizationCostTy(
           (getInstructionCost(I, ElementCount::getFixed(1)).first *
            VF.getKnownMinValue()),
           false);
   }
 
   Type *VectorTy;
-  unsigned C = getInstructionCost(I, VF, VectorTy);
+  InstructionCost C = getInstructionCost(I, VF, VectorTy);
 
   bool TypeNotScalarized =
       VF.isVector() && VectorTy->isVectorTy() &&
       TTI.getNumberOfParts(VectorTy) < VF.getKnownMinValue();
   return VectorizationCostTy(C, TypeNotScalarized);
 }
 
 unsigned LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I,
if (isa<LoadInst>(I)) {
       }
     } else
       // Make sure I gets scalarized and a cost estimate without
       // scalarization overhead.
       ForcedScalars[VF].insert(I);
   }
 }
 
-unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
-                                                        ElementCount VF,
-                                                        Type *&VectorTy) {
+InstructionCost
+LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF,
   Type *RetTy = I->getType();
   if (canTruncateToMinimalBitwidth(I, VF))
     RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
   VectorTy = isScalarAfterVectorization(I, VF) ? RetTy : ToVectorTy(RetTy, VF);
   auto SE = PSE.getSE();
   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
 
   // TODO: We need to estimate the cost of intrinsic calls.