[LV][AArch64] Prefer Fixed over Scalable if cost-model is equal (Neoverse V2)

For the Neoverse V2, we would like to prefer fixed width over scalable
vectorisation if the cost-model assigns an equal cost for certain loops. This
improves 7 kernels from TSVC-2 by about 2x, and does not affect SPEC21017 INT
and FP. This also adds a new TTI new hook that can steer the loop vectoriser
to preferring fixed width vectorization, which can be set per CPU. For now,
this is only enabled for the Neoverse V2.

This tends to benefit small kernels, like the ones in TSVC, for a
number of reasons: processing the predicates does not come entirely
for free, NEON tends to generate slightly less code which can have a
big impact on these small kernels, and then there are second order
effects that SVE codegen is slightly less optimal in some areas.

This codegen strategy to generate more NEON is inline with GCC's codegen
strategy, which is actually even more aggressive in generating NEON when
no predication is required. We could be smarter and more aggressive too
about generating more NEON (and improve performance), but this seems to
be a first good and straight forward step.

Author: sjoerdmeijer

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -1674,6 +1674,11 @@ class TargetTransformInfo {
         false; ///< If op is an fp min/max, whether NaNs may be present.
   };
 
+  /// \returns True if the targets prefers fixed width vectorization if the
+  /// loop vectorizer's cost-model assigns an equal cost to the fixed and
+  /// scalable version of the vectorized loop.
+  bool preferFixedOverScalableIfEqualCost() const;
+
   /// \returns True if the target prefers reductions in loop.
   bool preferInLoopReduction(unsigned Opcode, Type *Ty,
                              ReductionFlags Flags) const;
@@ -2143,6 +2148,7 @@ class TargetTransformInfo::Concept {
   virtual unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
                                         unsigned ChainSizeInBytes,
                                         VectorType *VecTy) const = 0;
+  virtual bool preferFixedOverScalableIfEqualCost() const = 0;
   virtual bool preferInLoopReduction(unsigned Opcode, Type *Ty,
                                      ReductionFlags) const = 0;
   virtual bool preferPredicatedReductionSelect(unsigned Opcode, Type *Ty,
@@ -2873,6 +2879,9 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
                                 VectorType *VecTy) const override {
     return Impl.getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
   }
+  bool preferFixedOverScalableIfEqualCost() const override {
+    return Impl.preferFixedOverScalableIfEqualCost();
+  }
   bool preferInLoopReduction(unsigned Opcode, Type *Ty,
                              ReductionFlags Flags) const override {
     return Impl.preferInLoopReduction(Opcode, Ty, Flags);

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

@@ -913,6 +913,8 @@ class TargetTransformInfoImplBase {
     return VF;
   }
 
+  bool preferFixedOverScalableIfEqualCost() const { return false; }
+
   bool preferInLoopReduction(unsigned Opcode, Type *Ty,
                              TTI::ReductionFlags Flags) const {
     return false;

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -1282,6 +1282,10 @@ unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF,
   return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
 }
 
+bool TargetTransformInfo::preferFixedOverScalableIfEqualCost() const {
+  return TTIImpl->preferFixedOverScalableIfEqualCost();
+}
+
 bool TargetTransformInfo::preferInLoopReduction(unsigned Opcode, Type *Ty,
                                                 ReductionFlags Flags) const {
   return TTIImpl->preferInLoopReduction(Opcode, Ty, Flags);

llvm/lib/Target/AArch64/AArch64Features.td

@@ -244,6 +244,10 @@ def FeatureExperimentalZeroingPseudos
 def FeatureUseScalarIncVL : SubtargetFeature<"use-scalar-inc-vl",
   "UseScalarIncVL", "true", "Prefer inc/dec over add+cnt">;
 
+def FeatureUseFixedOverScalableIfEqualCost: SubtargetFeature<"use-fixed-over-scalable-equal-cost",
+  "UseFixedOverScalableIfEqualCost", "true",
+  "Prefer fixed width loop vectorization over scalable if the cost-model assigns equal costs">;
+
 def FeatureBF16 : Extension<"bf16", "BF16",
     "Enable BFloat16 Extension (FEAT_BF16)", [],
     "FEAT_BF16", "+bf16", 280>;

llvm/lib/Target/AArch64/AArch64Processors.td

@@ -489,6 +489,7 @@ def TuneNeoverseV2 : SubtargetFeature<"neoversev2", "ARMProcFamily", "NeoverseV2
                                       FeatureALULSLFast,
                                       FeaturePostRAScheduler,
                                       FeatureEnableSelectOptimize,
+                                      FeatureUseFixedOverScalableIfEqualCost,
                                       FeaturePredictableSelectIsExpensive]>;
 
 def TuneNeoverseV3 : SubtargetFeature<"neoversev3", "ARMProcFamily", "NeoverseV3",

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h

@@ -371,6 +371,19 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
     return TailFoldingStyle::DataWithoutLaneMask;
   }
 
+  bool preferFixedOverScalableIfEqualCost() const {
+    // TODO: Ideally we only check getVScaleForTuning() == 1, but we do
+    // also check if the CPU has the useFixed feature enabled, which was
+    // introduced to reduce the impact of this for other targets.
+    //
+    // With the getVScaleForTuning() == 1 check, we ask if we're tuning based
+    // on the assumption the SVE registers are no bigger than the NEON ones.
+    // If this is the case, and the loop vectorisation cost-model is a tie,
+    // we prefer NEON as there should be no advantage of using SVE.
+    return ST->useFixedOverScalableIfEqualCost() &&
+           ST->getVScaleForTuning() == 1;
+  }
+
   bool preferPredicateOverEpilogue(TailFoldingInfo *TFI);
 
   bool supportsScalableVectors() const { return ST->hasSVE(); }

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -447,10 +447,13 @@ class LoopVectorizationPlanner {
   VectorizationFactor
   selectVectorizationFactor(const ElementCountSet &CandidateVFs);
 
+  bool preferFixedOverScalableIfEqualCost(const Loop *L, ElementCount VF,
+                                          unsigned IC) const;
+
   /// Returns true if the per-lane cost of VectorizationFactor A is lower than
   /// that of B.
   bool isMoreProfitable(const VectorizationFactor &A,
-                        const VectorizationFactor &B) const;
+                        const VectorizationFactor &B, unsigned IC = 0) const;
 
   /// Determines if we have the infrastructure to vectorize the loop and its
   /// epilogue, assuming the main loop is vectorized by \p VF.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -4760,8 +4760,42 @@ getVScaleForTuning(const Loop *L, const TargetTransformInfo &TTI) {
   return TTI.getVScaleForTuning();
 }
 
-bool LoopVectorizationPlanner::isMoreProfitable(
-    const VectorizationFactor &A, const VectorizationFactor &B) const {
+bool LoopVectorizationPlanner::preferFixedOverScalableIfEqualCost(
+    const Loop *L, ElementCount VF, unsigned IC) const {
+  // Check if the Subtarget has the feature enabled that it might prefer fixed
+  // over scalable vectorisation.
+  if (!TTI.preferFixedOverScalableIfEqualCost())
+    return false;
+
+  // With an interleaving count of 1, we don't expect the potential use of
+  // LDP/STP, which are instructions that SVE lacks, to make a difference for
+  // fixed with vectorisation.
+  if (IC == 1)
+    return false;
+
+  for (BasicBlock *BB : L->blocks()) {
+    for (Instruction &I : *BB) {
+      if (!(isa<LoadInst>(I) || isa<StoreInst>(I)))
+        continue;
+
+      // TODO: This could be more sophisiticated, but the initial idea here is
+      // that if the cost-model is a tie, and gathers/scatters or predication
+      // is required, then SVE is probably more efficient so favour SVE in
+      // these cases.
+      auto Decision = CM.getWideningDecision(&I, VF);
+      if (Decision == LoopVectorizationCostModel::CM_GatherScatter)
+        return false;
+      else if (Decision == LoopVectorizationCostModel::CM_Widen)
+        return !Legal->isMaskRequired(&I);
+    }
+  }
+
+  return false;
+}
+
+bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A,
+                                                const VectorizationFactor &B,
+                                                unsigned IC) const {
   InstructionCost CostA = A.Cost;
   InstructionCost CostB = B.Cost;
 
@@ -4780,7 +4814,10 @@ bool LoopVectorizationPlanner::isMoreProfitable(
   // Assume vscale may be larger than 1 (or the value being tuned for),
   // so that scalable vectorization is slightly favorable over fixed-width
   // vectorization.
-  bool PreferScalable = A.Width.isScalable() && !B.Width.isScalable();
+  bool PreferScalable = false;
+  if (!preferFixedOverScalableIfEqualCost(OrigLoop, A.Width, IC))
+    PreferScalable = A.Width.isScalable() && !B.Width.isScalable();
+
   auto CmpFn = [PreferScalable](const InstructionCost &LHS,
                                 const InstructionCost &RHS) {
     return PreferScalable ? LHS <= RHS : LHS < RHS;
@@ -5100,7 +5137,7 @@ VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(
         continue;
     }
 
-    if (Result.Width.isScalar() || isMoreProfitable(NextVF, Result))
+    if (Result.Width.isScalar() || isMoreProfitable(NextVF, Result, IC))
       Result = NextVF;
   }