SLPVectorizer: Improve the cost model for loop invariant broadcast values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179930 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -83,7 +83,7 @@ struct SLPVectorizer : public FunctionPass {
 
       // Use the bollom up slp vectorizer to construct chains that start with
       // he store instructions.
-      BoUpSLP R(BB, SE, DL, TTI, AA);
+      BoUpSLP R(BB, SE, DL, TTI, AA, LI->getLoopFor(BB));
 
       // Vectorize trees that end at reductions.
       BBChanged |= vectorizeReductions(BB, R);

lib/Transforms/Vectorize/VecUtils.cpp

@@ -18,6 +18,7 @@
 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Analysis/Verifier.h"
+#include "llvm/Analysis/LoopInfo.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Function.h"
@@ -44,8 +45,8 @@ static const unsigned RecursionMaxDepth = 6;
 namespace llvm {
 
 BoUpSLP::BoUpSLP(BasicBlock *Bb, ScalarEvolution *S, DataLayout *Dl,
-                             TargetTransformInfo *Tti, AliasAnalysis *Aa) :
-                             BB(Bb), SE(S), DL(Dl), TTI(Tti), AA(Aa) {
+                 TargetTransformInfo *Tti, AliasAnalysis *Aa, Loop *Lp) :
+  BB(Bb), SE(S), DL(Dl), TTI(Tti), AA(Aa), L(Lp)  {
   numberInstructions();
 }
 
@@ -121,7 +122,7 @@ bool BoUpSLP::vectorizeStoreChain(ValueList &Chain, int CostThreshold) {
     if (Cost < CostThreshold) {
       DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
       vectorizeTree(Operands, VF);
-      i += VF;
+      i += VF - 1;
       Changed = true;
     }
   }
@@ -381,13 +382,15 @@ int BoUpSLP::getTreeCost_rec(ValueList &VL, unsigned Depth) {
   // Check if all of the operands are constants.
   bool AllConst = true;
   bool AllSameScalar = true;
+  bool MustScalarizeFlag = false;
   for (unsigned i = 0, e = VL.size(); i < e; ++i) {
     AllConst &= isa<Constant>(VL[i]);
     AllSameScalar &= (VL[0] == VL[i]);
     // Must have a single use.
     Instruction *I = dyn_cast<Instruction>(VL[i]);
-    // This instruction is outside the basic block or if it is a known hazard.
-    if (MustScalarize.count(VL[i]) || (I && I->getParent() != BB))
+    MustScalarizeFlag |= MustScalarize.count(VL[i]);
+    // This instruction is outside the basic block.
+    if (I && I->getParent() != BB)
       return getScalarizationCost(VecTy);
   }
 
@@ -395,11 +398,23 @@ int BoUpSLP::getTreeCost_rec(ValueList &VL, unsigned Depth) {
   if (AllConst) return 0;
 
   // If all of the operands are identical we can broadcast them.
-  if (AllSameScalar)
-    return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
-
-  // Scalarize unknown structures.
   Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
+  if (AllSameScalar) {
+    // If we are in a loop, and this is not an instruction (e.g. constant or
+    // argument) or the instruction is defined outside the loop then assume
+    // that the cost is zero.
+    if (L && (!VL0 || !L->contains(VL0)))
+      return 0;
+
+    // We need to broadcast the scalar.
+    return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
+  }
+
+  // If this is not a constant, or a scalar from outside the loop then we
+  // need to scalarize it.
+  if (MustScalarizeFlag)
+    return getScalarizationCost(VecTy);
+
   if (!VL0) return getScalarizationCost(VecTy);
   assert(VL0->getParent() == BB && "Wrong BB");
 

lib/Transforms/Vectorize/VecUtils.h

@@ -27,6 +27,7 @@ class BasicBlock; class Instruction; class Type;
 class VectorType; class StoreInst; class Value;
 class ScalarEvolution; class DataLayout;
 class TargetTransformInfo; class AliasAnalysis;
+class Loop;
 
 /// Bottom Up SLP vectorization utility class.
 struct BoUpSLP  {
@@ -37,7 +38,7 @@ struct BoUpSLP  {
 
   // \brief C'tor.
   BoUpSLP(BasicBlock *Bb, ScalarEvolution *Se, DataLayout *Dl,
-         TargetTransformInfo *Tti, AliasAnalysis *Aa);
+         TargetTransformInfo *Tti, AliasAnalysis *Aa, Loop *Lp);
 
   /// \brief Take the pointer operand from the Load/Store instruction.
   /// \returns NULL if this is not a valid Load/Store instruction.
@@ -112,7 +113,7 @@ private:
 
   /// \returns a vector from a collection of scalars in \p VL.
   Value *Scalarize(ValueList &VL, VectorType *Ty);
-  
+
 private:
   /// Maps instructions to numbers and back.
   SmallDenseMap<Value*, int> InstrIdx;
@@ -155,6 +156,7 @@ private:
   DataLayout *DL;
   TargetTransformInfo *TTI;
   AliasAnalysis *AA;
+  Loop *L;
 };
 
 } // end of namespace

test/Transforms/SLPVectorizer/X86/loopinvariant.ll

@@ -0,0 +1,73 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+;CHECK: @foo
+;CHECK: load <4 x i32>
+;CHECK: add <4 x i32>
+;CHECK: store <4 x i32>
+;CHECK: load <4 x i32>
+;CHECK: add <4 x i32>
+;CHECK: store <4 x i32>
+;CHECK: ret
+define i32 @foo(i32* nocapture %A, i32 %n) #0 {
+entry:
+  %cmp62 = icmp sgt i32 %n, 0
+  br i1 %cmp62, label %for.body, label %for.end
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4, !tbaa !0
+  %add1 = add nsw i32 %0, %n
+  store i32 %add1, i32* %arrayidx, align 4, !tbaa !0
+  %1 = or i64 %indvars.iv, 1
+  %arrayidx4 = getelementptr inbounds i32* %A, i64 %1
+  %2 = load i32* %arrayidx4, align 4, !tbaa !0
+  %add5 = add nsw i32 %2, %n
+  store i32 %add5, i32* %arrayidx4, align 4, !tbaa !0
+  %3 = or i64 %indvars.iv, 2
+  %arrayidx8 = getelementptr inbounds i32* %A, i64 %3
+  %4 = load i32* %arrayidx8, align 4, !tbaa !0
+  %add9 = add nsw i32 %4, %n
+  store i32 %add9, i32* %arrayidx8, align 4, !tbaa !0
+  %5 = or i64 %indvars.iv, 3
+  %arrayidx12 = getelementptr inbounds i32* %A, i64 %5
+  %6 = load i32* %arrayidx12, align 4, !tbaa !0
+  %add13 = add nsw i32 %6, %n
+  store i32 %add13, i32* %arrayidx12, align 4, !tbaa !0
+  %7 = or i64 %indvars.iv, 4
+  %arrayidx16 = getelementptr inbounds i32* %A, i64 %7
+  %8 = load i32* %arrayidx16, align 4, !tbaa !0
+  %add17 = add nsw i32 %8, %n
+  store i32 %add17, i32* %arrayidx16, align 4, !tbaa !0
+  %9 = or i64 %indvars.iv, 5
+  %arrayidx20 = getelementptr inbounds i32* %A, i64 %9
+  %10 = load i32* %arrayidx20, align 4, !tbaa !0
+  %add21 = add nsw i32 %10, %n
+  store i32 %add21, i32* %arrayidx20, align 4, !tbaa !0
+  %11 = or i64 %indvars.iv, 6
+  %arrayidx24 = getelementptr inbounds i32* %A, i64 %11
+  %12 = load i32* %arrayidx24, align 4, !tbaa !0
+  %add25 = add nsw i32 %12, %n
+  store i32 %add25, i32* %arrayidx24, align 4, !tbaa !0
+  %13 = or i64 %indvars.iv, 7
+  %arrayidx28 = getelementptr inbounds i32* %A, i64 %13
+  %14 = load i32* %arrayidx28, align 4, !tbaa !0
+  %add29 = add nsw i32 %14, %n
+  store i32 %add29, i32* %arrayidx28, align 4, !tbaa !0
+  %indvars.iv.next = add i64 %indvars.iv, 8
+  %15 = trunc i64 %indvars.iv.next to i32
+  %cmp = icmp slt i32 %15, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.end:                                          ; preds = %for.body, %entry
+  ret i32 undef
+}
+
+attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!0 = metadata !{metadata !"int", metadata !1}
+!1 = metadata !{metadata !"omnipotent char", metadata !2}
+!2 = metadata !{metadata !"Simple C/C++ TBAA"}