[InstCombine] Fold binary op of reductions.

Replace binary of of two reductions with one reduction of the binary op
applied to vectors. For example:

```
%v0_red = tail call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> %v0)
%v1_red = tail call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> %v1)
%res = add i32 %v0_red, %v1_red
```
gets transformed to:

```
%1 = add <16 x i32> %v0, %v1
%res = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> %1)
```

Author: Mikhail Gudim

llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

@@ -1516,6 +1516,9 @@ Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
   if (Instruction *X = foldVectorBinop(I))
     return X;
 
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
+
   if (Instruction *Phi = foldBinopWithPhiOperands(I))
     return Phi;
 
@@ -2376,19 +2379,8 @@ Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
     }
   }
 
-  auto m_AddRdx = [](Value *&Vec) {
-    return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
-  };
-  Value *V0, *V1;
-  if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
-      V0->getType() == V1->getType()) {
-    // Difference of sums is sum of differences:
-    // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
-    Value *Sub = Builder.CreateSub(V0, V1);
-    Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
-                                         {Sub->getType()}, {Sub});
-    return replaceInstUsesWith(I, Rdx);
-  }
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
 
   if (Constant *C = dyn_cast<Constant>(Op0)) {
     Value *X;

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

@@ -2388,6 +2388,9 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
   if (Instruction *X = foldVectorBinop(I))
     return X;
 
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
+
   if (Instruction *Phi = foldBinopWithPhiOperands(I))
     return Phi;
 
@@ -3588,6 +3591,9 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
   if (Instruction *X = foldVectorBinop(I))
     return X;
 
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
+
   if (Instruction *Phi = foldBinopWithPhiOperands(I))
     return Phi;
 
@@ -4713,6 +4719,9 @@ Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) {
   if (Instruction *X = foldVectorBinop(I))
     return X;
 
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
+
   if (Instruction *Phi = foldBinopWithPhiOperands(I))
     return Phi;
 

llvm/lib/Transforms/InstCombine/InstCombineInternal.h

@@ -594,6 +594,7 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
 
   /// Canonicalize the position of binops relative to shufflevector.
   Instruction *foldVectorBinop(BinaryOperator &Inst);
+  Instruction *foldBinopOfReductions(BinaryOperator &Inst);
   Instruction *foldVectorSelect(SelectInst &Sel);
   Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf);
 

llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

@@ -205,6 +205,9 @@ Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) {
   if (Instruction *X = foldVectorBinop(I))
     return X;
 
+  if (Instruction *X = foldBinopOfReductions(I))
+    return replaceInstUsesWith(I, X);
+
   if (Instruction *Phi = foldBinopWithPhiOperands(I))
     return Phi;
 

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -2296,6 +2296,58 @@ Instruction *InstCombinerImpl::foldVectorBinop(BinaryOperator &Inst) {
   return nullptr;
 }
 
+static Intrinsic::ID getReductionForBinop(Instruction::BinaryOps Opc) {
+  switch (Opc) {
+  default:
+    break;
+  case Instruction::Add:
+    return Intrinsic::vector_reduce_add;
+  case Instruction::Mul:
+    return Intrinsic::vector_reduce_mul;
+  case Instruction::And:
+    return Intrinsic::vector_reduce_and;
+  case Instruction::Or:
+    return Intrinsic::vector_reduce_or;
+  case Instruction::Xor:
+    return Intrinsic::vector_reduce_xor;
+  }
+  return Intrinsic::num_intrinsics;
+}
+
+Instruction *InstCombinerImpl::foldBinopOfReductions(BinaryOperator &Inst) {
+  IntrinsicInst *II0 = dyn_cast<IntrinsicInst>(Inst.getOperand(0));
+  if (!II0)
+    return nullptr;
+  IntrinsicInst *II1 = dyn_cast<IntrinsicInst>(Inst.getOperand(1));
+  if (!II1)
+    return nullptr;
+
+  Instruction::BinaryOps BinOpOpc = Inst.getOpcode();
+  Intrinsic::ID ReductionIID = getReductionForBinop(BinOpOpc);
+  if (BinOpOpc == Instruction::Sub)
+    ReductionIID = Intrinsic::vector_reduce_add;
+
+  if (ReductionIID == Intrinsic::num_intrinsics)
+    return nullptr;
+  if (II0->getIntrinsicID() != ReductionIID)
+    return nullptr;
+  if (II1->getIntrinsicID() != ReductionIID)
+    return nullptr;
+
+  Value *V0 = II0->getArgOperand(0);
+  Value *V1 = II1->getArgOperand(0);
+  Type *VTy = V0->getType();
+  if (V1->getType() != VTy)
+    return nullptr;
+
+  Value *VectorBO = Builder.CreateBinOp(BinOpOpc, V0, V1);
+  // if (auto *VectorInstBO = dyn_cast<BinaryOperator>(VectorBO))
+  //   VectorInstBO->copyIRFlags(&Inst);
+
+  Instruction *Rdx = Builder.CreateIntrinsic(ReductionIID, {VTy}, {VectorBO});
+  return Rdx;
+}
+
 /// Try to narrow the width of a binop if at least 1 operand is an extend of
 /// of a value. This requires a potentially expensive known bits check to make
 /// sure the narrow op does not overflow.