LoopVectorizer: Pass OperandValueKind information to the cost model

Pass down the fact that an operand is going to be a vector of constants.

This should bring the performance of MultiSource/Benchmarks/PAQ8p/paq8p on x86
back. It had degraded to scalar performance due to my pervious shift cost change
that made all shifts expensive on x86.

radar://13576547

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178809 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 41 insertions, 2 deletions

diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp
index 930d9c412f..acf2b819b8 100644
--- a/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -3331,8 +3331,19 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
   case Instruction::AShr:
   case Instruction::And:
   case Instruction::Or:
-  case Instruction::Xor:
-    return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy);
+  case Instruction::Xor: {
+    // Certain instructions can be cheaper to vectorize if they have a constant
+    // second vector operand. One example of this are shifts on x86.
+    TargetTransformInfo::OperandValueKind Op1VK =
+      TargetTransformInfo::OK_AnyValue;
+    TargetTransformInfo::OperandValueKind Op2VK =
+      TargetTransformInfo::OK_AnyValue;
+
+    if (isa<ConstantInt>(I->getOperand(1)))
+      Op2VK = TargetTransformInfo::OK_UniformConstantValue;
+
+    return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
+  }
   case Instruction::Select: {
     SelectInst *SI = cast<SelectInst>(I);
     const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
diff --git a/test/Transforms/LoopVectorize/X86/constant-vector-operand.ll b/test/Transforms/LoopVectorize/X86/constant-vector-operand.ll
new file mode 100644
index 0000000000..6c924409af
--- /dev/null
+++ b/test/Transforms/LoopVectorize/X86/constant-vector-operand.ll
@@ -0,0 +1,28 @@
+; RUN: opt -mtriple=x86_64-apple-darwin -mcpu=core2 -loop-vectorize -dce -instcombine -S < %s | FileCheck %s
+
+@B = common global [1024 x i32] zeroinitializer, align 16
+@A = common global [1024 x i32] zeroinitializer, align 16
+
+; We use to not vectorize this loop because the shift was deemed to expensive.
+; Now that we differentiate shift cost base on the operand value kind, we will
+; vectorize this loop.
+; CHECK: ashr <4 x i32>
+define void @f() {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds [1024 x i32]* @B, i64 0, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4
+  %shl = ashr i32 %0, 3
+  %arrayidx2 = getelementptr inbounds [1024 x i32]* @A, i64 0, i64 %indvars.iv
+  store i32 %shl, i32* %arrayidx2, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:
+  ret void
+}