ARM cost model: Account for zero cost scalar SROA instructions

By vectorizing a series of srl, or, ... instructions we have obfuscated the
intention so much that the backend does not know how to fold this code away.

radar://15336950

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193573 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 85 insertions, 6 deletions

diff --git a/lib/Target/ARM/ARMTargetTransformInfo.cpp b/lib/Target/ARM/ARMTargetTransformInfo.cpp
index 9dc1cd1..86b6215 100644
--- a/lib/Target/ARM/ARMTargetTransformInfo.cpp
+++ b/lib/Target/ARM/ARMTargetTransformInfo.cpp
@@ -523,8 +523,20 @@ unsigned ARMTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueK
   if (Idx != -1)
     return LT.first * CostTbl[Idx].Cost;
 
-
-  return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
-                                                     Op2Info);
+  unsigned Cost =
+      TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
+
+  // This is somewhat of a hack. The problem that we are facing is that SROA
+  // creates a sequence of shift, and, or instructions to construct values.
+  // These sequences are recognized by the ISel and have zero-cost. Not so for
+  // the vectorized code. Because we have support for v2i64 but not i64 those
+  // sequences look particularily beneficial to vectorize.
+  // To work around this we increase the cost of v2i64 operations to make them
+  // seem less beneficial.
+  if (LT.second == MVT::v2i64 &&
+      Op2Info == TargetTransformInfo::OK_UniformConstantValue)
+    Cost += 4;
+
+  return Cost;
 }
 
diff --git a/lib/Transforms/Vectorize/SLPVectorizer.cpp b/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 012521a..2797a21 100644
--- a/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -1013,9 +1013,24 @@ int BoUpSLP::getEntryCost(TreeEntry *E) {
         TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
         VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
       } else {
-        ScalarCost = VecTy->getNumElements() *
-        TTI->getArithmeticInstrCost(Opcode, ScalarTy);
-        VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
+        // Certain instructions can be cheaper to vectorize if they have a
+        // constant second vector operand.
+        TargetTransformInfo::OperandValueKind Op1VK =
+            TargetTransformInfo::OK_AnyValue;
+        TargetTransformInfo::OperandValueKind Op2VK =
+            TargetTransformInfo::OK_UniformConstantValue;
+
+        // Check whether all second operands are constant.
+        for (unsigned i = 0; i < VL.size(); ++i)
+          if (!isa<ConstantInt>(cast<Instruction>(VL[i])->getOperand(1))) {
+            Op2VK = TargetTransformInfo::OK_AnyValue;
+            break;
+          }
+
+        ScalarCost =
+            VecTy->getNumElements() *
+            TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
+        VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
       }
       return VecCost - ScalarCost;
     }
diff --git a/test/Transforms/SLPVectorizer/ARM/sroa.ll b/test/Transforms/SLPVectorizer/ARM/sroa.ll
new file mode 100644
index 0000000..e0c75b1
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/ARM/sroa.ll
@@ -0,0 +1,52 @@
+; RUN: opt -S -mcpu=swift -mtriple=thumbv7-apple-ios -basicaa -slp-vectorizer < %s | FileCheck %s
+
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
+
+%class.Complex = type { double, double }
+
+; Code like this is the result of SROA. Make sure we don't vectorize this
+; because the in the scalar version of this the shl/or are handled by the
+; backend and disappear, the vectorized code stays.
+
+; CHECK-LABEL: SROAed
+; CHECK-NOT: shl <2 x i64>
+; CHECK-NOT: or <2 x i64>
+
+define void @SROAed(%class.Complex* noalias nocapture sret %agg.result, [4 x i32] %a.coerce, [4 x i32] %b.coerce) {
+entry:
+  %a.coerce.fca.0.extract = extractvalue [4 x i32] %a.coerce, 0
+  %a.sroa.0.0.insert.ext = zext i32 %a.coerce.fca.0.extract to i64
+  %a.coerce.fca.1.extract = extractvalue [4 x i32] %a.coerce, 1
+  %a.sroa.0.4.insert.ext = zext i32 %a.coerce.fca.1.extract to i64
+  %a.sroa.0.4.insert.shift = shl nuw i64 %a.sroa.0.4.insert.ext, 32
+  %a.sroa.0.4.insert.insert = or i64 %a.sroa.0.4.insert.shift, %a.sroa.0.0.insert.ext
+  %0 = bitcast i64 %a.sroa.0.4.insert.insert to double
+  %a.coerce.fca.2.extract = extractvalue [4 x i32] %a.coerce, 2
+  %a.sroa.3.8.insert.ext = zext i32 %a.coerce.fca.2.extract to i64
+  %a.coerce.fca.3.extract = extractvalue [4 x i32] %a.coerce, 3
+  %a.sroa.3.12.insert.ext = zext i32 %a.coerce.fca.3.extract to i64
+  %a.sroa.3.12.insert.shift = shl nuw i64 %a.sroa.3.12.insert.ext, 32
+  %a.sroa.3.12.insert.insert = or i64 %a.sroa.3.12.insert.shift, %a.sroa.3.8.insert.ext
+  %1 = bitcast i64 %a.sroa.3.12.insert.insert to double
+  %b.coerce.fca.0.extract = extractvalue [4 x i32] %b.coerce, 0
+  %b.sroa.0.0.insert.ext = zext i32 %b.coerce.fca.0.extract to i64
+  %b.coerce.fca.1.extract = extractvalue [4 x i32] %b.coerce, 1
+  %b.sroa.0.4.insert.ext = zext i32 %b.coerce.fca.1.extract to i64
+  %b.sroa.0.4.insert.shift = shl nuw i64 %b.sroa.0.4.insert.ext, 32
+  %b.sroa.0.4.insert.insert = or i64 %b.sroa.0.4.insert.shift, %b.sroa.0.0.insert.ext
+  %2 = bitcast i64 %b.sroa.0.4.insert.insert to double
+  %b.coerce.fca.2.extract = extractvalue [4 x i32] %b.coerce, 2
+  %b.sroa.3.8.insert.ext = zext i32 %b.coerce.fca.2.extract to i64
+  %b.coerce.fca.3.extract = extractvalue [4 x i32] %b.coerce, 3
+  %b.sroa.3.12.insert.ext = zext i32 %b.coerce.fca.3.extract to i64
+  %b.sroa.3.12.insert.shift = shl nuw i64 %b.sroa.3.12.insert.ext, 32
+  %b.sroa.3.12.insert.insert = or i64 %b.sroa.3.12.insert.shift, %b.sroa.3.8.insert.ext
+  %3 = bitcast i64 %b.sroa.3.12.insert.insert to double
+  %add = fadd double %0, %2
+  %add3 = fadd double %1, %3
+  %re.i.i = getelementptr inbounds %class.Complex* %agg.result, i32 0, i32 0
+  store double %add, double* %re.i.i, align 4
+  %im.i.i = getelementptr inbounds %class.Complex* %agg.result, i32 0, i32 1
+  store double %add3, double* %im.i.i, align 4
+  ret void
+}