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

sjoerdmeijer · sjoerdmeijer · commit 193457ab3be8 · 2024-06-24T15:07:26.000+05:30
For the Neoverse V2, we would like to prefer fixed width over scalable
vectorisation if the cost-model assigns an equal cost to both 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 vectorizater 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
affects 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.
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfo.h b/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);
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h b/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;
diff --git a/llvm/lib/Analysis/TargetTransformInfo.cpp b/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);
diff --git a/llvm/lib/Target/AArch64/AArch64Features.td b/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>;
diff --git a/llvm/lib/Target/AArch64/AArch64Processors.td b/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",
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h
@@ -371,6 +371,10 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
     return TailFoldingStyle::DataWithoutLaneMask;
   }
 
+  bool preferFixedOverScalableIfEqualCost() const {
+    return ST->useFixedOverScalableIfEqualCost();
+  }
+
   bool preferPredicateOverEpilogue(TailFoldingInfo *TFI);
 
   bool supportsScalableVectors() const { return ST->hasSVE(); }
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4780,7 +4780,12 @@ 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;
+  if (!TTI.preferFixedOverScalableIfEqualCost())
+    PreferScalable = !A.Width.isScalable() && B.Width.isScalable();
+  else
+    PreferScalable = A.Width.isScalable() && !B.Width.isScalable();
+
   auto CmpFn = [PreferScalable](const InstructionCost &LHS,
                                 const InstructionCost &RHS) {
     return PreferScalable ? LHS <= RHS : LHS < RHS;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/prefer-fixed-if-equal-to-scalable.ll b/llvm/test/Transforms/LoopVectorize/AArch64/prefer-fixed-if-equal-to-scalable.ll
@@ -0,0 +1,216 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S < %s -passes=loop-vectorize | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
+target triple = "aarch64-unknown-linux-gnu"
+
+@a = dso_local local_unnamed_addr global [32000 x float] zeroinitializer, align 64
+@b = dso_local local_unnamed_addr global [32000 x float] zeroinitializer, align 64
+
+define void @NeoverseV2() local_unnamed_addr #0 {
+; CHECK-LABEL: define void @NeoverseV2(
+; CHECK-SAME: ) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[TMP11:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP18:%.*]] = mul i64 [[TMP11]], 8
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 16000, [[TMP18]]
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[TMP26:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP27:%.*]] = mul i64 [[TMP26]], 8
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 16000, [[TMP27]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 16000, [[N_MOD_VF]]
+; CHECK-NEXT:    [[TMP28:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP28]], 8
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP29:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP30:%.*]] = mul i64 [[TMP29]], 4
+; CHECK-NEXT:    [[TMP9:%.*]] = add i64 [[TMP30]], 0
+; CHECK-NEXT:    [[TMP10:%.*]] = mul i64 [[TMP9]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], [[TMP10]]
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    [[TMP35:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP36:%.*]] = mul i64 [[TMP35]], 4
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i64 [[TMP36]]
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <vscale x 4 x float>, ptr [[TMP4]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <vscale x 4 x float>, ptr [[TMP17]], align 4
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds float, ptr [[TMP6]], i32 0
+; CHECK-NEXT:    [[TMP37:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP22:%.*]] = mul i64 [[TMP37]], 4
+; CHECK-NEXT:    [[TMP23:%.*]] = getelementptr inbounds float, ptr [[TMP6]], i64 [[TMP22]]
+; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <vscale x 4 x float>, ptr [[TMP8]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD3:%.*]] = load <vscale x 4 x float>, ptr [[TMP23]], align 4
+; CHECK-NEXT:    [[TMP24:%.*]] = fadd fast <vscale x 4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP25:%.*]] = fadd fast <vscale x 4 x float> [[WIDE_LOAD3]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP12:%.*]] = add nuw nsw i64 [[TMP0]], 16000
+; CHECK-NEXT:    [[TMP13:%.*]] = add nuw nsw i64 [[TMP1]], 16000
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP12]]
+; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP13]]
+; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds float, ptr [[TMP14]], i32 0
+; CHECK-NEXT:    [[TMP31:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP32:%.*]] = mul i64 [[TMP31]], 4
+; CHECK-NEXT:    [[TMP33:%.*]] = getelementptr inbounds float, ptr [[TMP14]], i64 [[TMP32]]
+; CHECK-NEXT:    store <vscale x 4 x float> [[TMP24]], ptr [[TMP16]], align 4
+; CHECK-NEXT:    store <vscale x 4 x float> [[TMP25]], ptr [[TMP33]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
+; CHECK-NEXT:    [[TMP34:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP34]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 16000, [[N_VEC]]
+; CHECK-NEXT:    br i1 [[CMP_N]], label %[[FOR_COND_CLEANUP:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[FOR_BODY:.*]]
+; CHECK:       [[FOR_COND_CLEANUP]]:
+; CHECK-NEXT:    ret void
+; CHECK:       [[FOR_BODY]]:
+; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_BODY]] ]
+; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    [[TMP19:%.*]] = load float, ptr [[ARRAYIDX]], align 4
+; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    [[TMP20:%.*]] = load float, ptr [[ARRAYIDX2]], align 4
+; CHECK-NEXT:    [[ADD:%.*]] = fadd fast float [[TMP20]], [[TMP19]]
+; CHECK-NEXT:    [[TMP21:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 16000
+; CHECK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP21]]
+; CHECK-NEXT:    store float [[ADD]], ptr [[ARRAYIDX5]], align 4
+; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 16000
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label %[[FOR_COND_CLEANUP]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+;
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv
+  %0 = load float, ptr %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 %indvars.iv
+  %1 = load float, ptr %arrayidx2, align 4
+  %add = fadd fast float %1, %0
+  %2 = add nuw nsw i64 %indvars.iv, 16000
+  %arrayidx5 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %2
+  store float %add, ptr %arrayidx5, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, 16000
+  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
+}
+
+define void @NeoverseV1() #1 {
+; CHECK-LABEL: define void @NeoverseV1(
+; CHECK-SAME: ) #[[ATTR1:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 8
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 16000, [[TMP1]]
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i64 [[TMP2]], 8
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 16000, [[TMP3]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 16000, [[N_MOD_VF]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 8
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP6:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP8:%.*]] = mul i64 [[TMP7]], 4
+; CHECK-NEXT:    [[TMP9:%.*]] = add i64 [[TMP8]], 0
+; CHECK-NEXT:    [[TMP10:%.*]] = mul i64 [[TMP9]], 1
+; CHECK-NEXT:    [[TMP11:%.*]] = add i64 [[INDEX]], [[TMP10]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP11]]
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds float, ptr [[TMP12]], i32 0
+; CHECK-NEXT:    [[TMP15:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP16:%.*]] = mul i64 [[TMP15]], 4
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds float, ptr [[TMP12]], i64 [[TMP16]]
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <vscale x 4 x float>, ptr [[TMP14]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <vscale x 4 x float>, ptr [[TMP17]], align 4
+; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[TMP11]]
+; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr inbounds float, ptr [[TMP18]], i32 0
+; CHECK-NEXT:    [[TMP21:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP22:%.*]] = mul i64 [[TMP21]], 4
+; CHECK-NEXT:    [[TMP23:%.*]] = getelementptr inbounds float, ptr [[TMP18]], i64 [[TMP22]]
+; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <vscale x 4 x float>, ptr [[TMP20]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD3:%.*]] = load <vscale x 4 x float>, ptr [[TMP23]], align 4
+; CHECK-NEXT:    [[TMP24:%.*]] = fadd fast <vscale x 4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP25:%.*]] = fadd fast <vscale x 4 x float> [[WIDE_LOAD3]], [[WIDE_LOAD1]]
+; CHECK-NEXT:    [[TMP26:%.*]] = add nuw nsw i64 [[TMP6]], 16000
+; CHECK-NEXT:    [[TMP27:%.*]] = add nuw nsw i64 [[TMP11]], 16000
+; CHECK-NEXT:    [[TMP28:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP26]]
+; CHECK-NEXT:    [[TMP29:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP27]]
+; CHECK-NEXT:    [[TMP30:%.*]] = getelementptr inbounds float, ptr [[TMP28]], i32 0
+; CHECK-NEXT:    [[TMP31:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP32:%.*]] = mul i64 [[TMP31]], 4
+; CHECK-NEXT:    [[TMP33:%.*]] = getelementptr inbounds float, ptr [[TMP28]], i64 [[TMP32]]
+; CHECK-NEXT:    store <vscale x 4 x float> [[TMP24]], ptr [[TMP30]], align 4
+; CHECK-NEXT:    store <vscale x 4 x float> [[TMP25]], ptr [[TMP33]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
+; CHECK-NEXT:    [[TMP34:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP34]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 16000, [[N_VEC]]
+; CHECK-NEXT:    br i1 [[CMP_N]], label %[[FOR_COND_CLEANUP:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[FOR_BODY:.*]]
+; CHECK:       [[FOR_COND_CLEANUP]]:
+; CHECK-NEXT:    ret void
+; CHECK:       [[FOR_BODY]]:
+; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_BODY]] ]
+; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    [[TMP35:%.*]] = load float, ptr [[ARRAYIDX]], align 4
+; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    [[TMP36:%.*]] = load float, ptr [[ARRAYIDX2]], align 4
+; CHECK-NEXT:    [[ADD:%.*]] = fadd fast float [[TMP36]], [[TMP35]]
+; CHECK-NEXT:    [[TMP37:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 16000
+; CHECK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 [[TMP37]]
+; CHECK-NEXT:    store float [[ADD]], ptr [[ARRAYIDX5]], align 4
+; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 16000
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label %[[FOR_COND_CLEANUP]], label %[[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
+;
+entry:
+  br label %for.body
+
+for.cond.cleanup:
+  ret void
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv
+  %0 = load float, ptr %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds [32000 x float], ptr @b, i64 0, i64 %indvars.iv
+  %1 = load float, ptr %arrayidx2, align 4
+  %add = fadd fast float %1, %0
+  %2 = add nuw nsw i64 %indvars.iv, 16000
+  %arrayidx5 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %2
+  store float %add, ptr %arrayidx5, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, 16000
+  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
+}
+
+attributes #0 = { mustprogress nofree norecurse nosync nounwind memory(readwrite, argmem: none, inaccessiblemem: none) uwtable vscale_range(1,16) "approx-func-fp-math"="true" "frame-pointer"="non-leaf" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="neoverse-v2" "target-features"="+sve,+sve2,+v9a" "unsafe-fp-math"="true" }
+
+attributes #1 = { mustprogress nofree norecurse nosync nounwind memory(readwrite, argmem: none, inaccessiblemem: none) uwtable vscale_range(1,16) "approx-func-fp-math"="true" "frame-pointer"="non-leaf" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+sve,+v9a" "unsafe-fp-math"="true" }
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
+; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META2]], [[META1]]}
+;.

Original file line number	Diff line number	Diff line change
`@@ -913,6 +913,8 @@ class TargetTransformInfoImplBase {`
`913`	`913`	`return VF;`
`914`	`914`	`}`
`915`	`915`
	`916`	`+ bool preferFixedOverScalableIfEqualCost() const { return false; }`
	`917`	`+`
`916`	`918`	`bool preferInLoopReduction(unsigned Opcode, Type *Ty,`
`917`	`919`	`TTI::ReductionFlags Flags) const {`
`918`	`920`	`return false;`
Original file line number	Diff line number	Diff line change
`@@ -371,6 +371,10 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {`
`371`	`371`	`return TailFoldingStyle::DataWithoutLaneMask;`
`372`	`372`	`}`
`373`	`373`
	`374`	`+ bool preferFixedOverScalableIfEqualCost() const {`
	`375`	`+ return ST->useFixedOverScalableIfEqualCost();`
	`376`	`+ }`
	`377`	`+`
`374`	`378`	`bool preferPredicateOverEpilogue(TailFoldingInfo *TFI);`
`375`	`379`
`376`	`380`	`bool supportsScalableVectors() const { return ST->hasSVE(); }`