SLPVectorizer: vectorize compares and selects.

llvm-svn: 184282

2 files changed, 107 insertions, 3 deletions

diff --git a/llvm/lib/Transforms/Vectorize/VecUtils.cpp b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
index 658162aa9f6..2f6b7df21b6 100644
--- a/llvm/lib/Transforms/Vectorize/VecUtils.cpp
+++ b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
@@ -384,6 +384,9 @@ void BoUpSLP::getTreeUses_rec(ArrayRef<Value *> VL, unsigned Depth) {
     case Instruction::Trunc:
     case Instruction::FPTrunc:
     case Instruction::BitCast:
+    case Instruction::Select:
+    case Instruction::ICmp:
+    case Instruction::FCmp:
     case Instruction::Add:
     case Instruction::FAdd:
     case Instruction::Sub:
@@ -541,6 +544,18 @@ int BoUpSLP::getTreeCost_rec(ArrayRef<Value *> VL, unsigned Depth) {
     Cost += (VecCost - ScalarCost);
     return Cost;
   }
+  case Instruction::FCmp:
+  case Instruction::ICmp: {
+    // Check that all of the compares have the same predicate.
+    CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
+    for (unsigned i = 1, e = VL.size(); i < e; ++i) {
+      CmpInst *Cmp = cast<CmpInst>(VL[i]);
+      if (Cmp->getPredicate() != P0)
+        return getScalarizationCost(VecTy);
+    }
+    // Fall through.
+  }
+  case Instruction::Select:
   case Instruction::Add:
   case Instruction::FAdd:
   case Instruction::Sub:
@@ -572,10 +587,19 @@ int BoUpSLP::getTreeCost_rec(ArrayRef<Value *> VL, unsigned Depth) {
     }
 
     // Calculate the cost of this instruction.
-    int ScalarCost = VecTy->getNumElements() *
+    int ScalarCost = 0;
+    int VecCost = 0;
+    if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
+        Opcode == Instruction::Select) {
+      VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
+      ScalarCost = VecTy->getNumElements() *
+        TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
+      VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
+    } else {
+      ScalarCost = VecTy->getNumElements() *
       TTI->getArithmeticInstrCost(Opcode, ScalarTy);
-
-    int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
+      VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
+    }
     Cost += (VecCost - ScalarCost);
     return Cost;
   }
@@ -772,6 +796,54 @@ Value *BoUpSLP::vectorizeTree_rec(ArrayRef<Value *> VL, int VF) {
 
     return V;
   }
+  case Instruction::FCmp:
+  case Instruction::ICmp: {
+    // Check that all of the compares have the same predicate.
+    CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
+    for (unsigned i = 1, e = VF; i < e; ++i) {
+      CmpInst *Cmp = cast<CmpInst>(VL[i]);
+      if (Cmp->getPredicate() != P0)
+        return Scalarize(VL, VecTy);
+    }
+
+    ValueList LHSV, RHSV;
+    for (int i = 0; i < VF; ++i) {
+      LHSV.push_back(cast<Instruction>(VL[i])->getOperand(0));
+      RHSV.push_back(cast<Instruction>(VL[i])->getOperand(1));
+    }
+
+    Value *L = vectorizeTree_rec(LHSV, VF);
+    Value *R = vectorizeTree_rec(RHSV, VF);
+    Value *V;
+    if (VL0->getOpcode() == Instruction::FCmp)
+      V = Builder.CreateFCmp(P0, L, R);
+    else
+      V = Builder.CreateICmp(P0, L, R);
+
+    for (int i = 0; i < VF; ++i)
+      VectorizedValues[VL[i]] = V;
+
+    return V;
+
+  }
+  case Instruction::Select: {
+    ValueList TrueVec, FalseVec, CondVec;
+    for (int i = 0; i < VF; ++i) {
+      CondVec.push_back(cast<Instruction>(VL[i])->getOperand(0));
+      TrueVec.push_back(cast<Instruction>(VL[i])->getOperand(1));
+      FalseVec.push_back(cast<Instruction>(VL[i])->getOperand(2));
+    }
+
+    Value *True = vectorizeTree_rec(TrueVec, VF);
+    Value *False = vectorizeTree_rec(FalseVec, VF);
+    Value *Cond = vectorizeTree_rec(CondVec, VF);
+    Value *V = Builder.CreateSelect(Cond, True, False);
+
+    for (int i = 0; i < VF; ++i)
+      VectorizedValues[VL[i]] = V;
+
+    return V;
+  }
   case Instruction::Add:
   case Instruction::FAdd:
   case Instruction::Sub:
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/cmp_sel.ll b/llvm/test/Transforms/SLPVectorizer/X86/cmp_sel.ll
new file mode 100644
index 00000000000..3bf4f19517a
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/cmp_sel.ll
@@ -0,0 +1,32 @@
+; 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"
+
+; int foo(double * restrict A, double * restrict B, double G) {
+;   A[0] = (B[10] ? G : 1);
+;   A[1] = (B[11] ? G : 1);
+; }
+
+;CHECK: @foo
+;CHECK: load <2 x double>
+;CHECK: fcmp une <2 x double>
+;CHECK: select <2 x i1>
+;CHECK: store <2 x double>
+;CHECK: ret i32 undef
+define i32 @foo(double* noalias nocapture %A, double* noalias nocapture %B, double %G) {
+entry:
+  %arrayidx = getelementptr inbounds double* %B, i64 10
+  %0 = load double* %arrayidx, align 8
+  %tobool = fcmp une double %0, 0.000000e+00
+  %cond = select i1 %tobool, double %G, double 1.000000e+00
+  store double %cond, double* %A, align 8
+  %arrayidx2 = getelementptr inbounds double* %B, i64 11
+  %1 = load double* %arrayidx2, align 8
+  %tobool3 = fcmp une double %1, 0.000000e+00
+  %cond7 = select i1 %tobool3, double %G, double 1.000000e+00
+  %arrayidx8 = getelementptr inbounds double* %A, i64 1
+  store double %cond7, double* %arrayidx8, align 8
+  ret i32 undef
+}
+