[ARM] Fix DAG combine loop in reduction distribution

Given a constant operand, the MVE and DAGCombine combines could fight,
each redistributing in the opposite order. Add a guard to the MVE
vecreduce distribution to prevent that.

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 79cf91f..9e7f403 100644 (file)
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -13089,7 +13089,8 @@ static SDValue TryDistrubutionADDVecReduce(SDNode *N, SelectionDAG &DAG) {
     //   add(add(X, vecreduce(Y)), vecreduce(Z))
     // to make better use of vaddva style instructions.
     if (VT == MVT::i32 && N1.getOpcode() == ISD::ADD && !IsVecReduce(N0) &&
-        IsVecReduce(N1.getOperand(0)) && IsVecReduce(N1.getOperand(1))) {
+        IsVecReduce(N1.getOperand(0)) && IsVecReduce(N1.getOperand(1)) &&
+        !isa<ConstantSDNode>(N0)) {
       SDValue Add0 = DAG.getNode(ISD::ADD, dl, VT, N0, N1.getOperand(0));
       return DAG.getNode(ISD::ADD, dl, VT, Add0, N1.getOperand(1));
     }

diff --git a/llvm/test/CodeGen/Thumb2/mve-vecreduce-slp.ll b/llvm/test/CodeGen/Thumb2/mve-vecreduce-slp.ll
index 44b8005..465457b 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/mve-vecreduce-slp.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-vecreduce-slp.ll
@@ -3772,6 +3772,53 @@ entry:
   ret i8 %52
 }
 
+
+define arm_aapcs_vfpcc i32 @add_two_const(<4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: add_two_const:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u32 r0, q1
+; CHECK-NEXT:    vaddva.u32 r0, q0
+; CHECK-NEXT:    adds r0, #10
+; CHECK-NEXT:    bx lr
+entry:
+  %a = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %x)
+  %b = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %y)
+  %c = add i32 %a, %b
+  %d = add i32 %c, 10
+  ret i32 %d
+}
+
+define arm_aapcs_vfpcc i32 @add_two_const2(<4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: add_two_const2:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u32 r0, q1
+; CHECK-NEXT:    vaddva.u32 r0, q0
+; CHECK-NEXT:    adds r0, #10
+; CHECK-NEXT:    bx lr
+entry:
+  %a = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %x)
+  %b = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %y)
+  %c = add i32 %a, 10
+  %d = add i32 %c, %b
+  ret i32 %d
+}
+
+define arm_aapcs_vfpcc i32 @add_two_const3(<4 x i32> %x, <4 x i32> %y) {
+; CHECK-LABEL: add_two_const3:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u32 r0, q0
+; CHECK-NEXT:    vaddva.u32 r0, q1
+; CHECK-NEXT:    adds r0, #20
+; CHECK-NEXT:    bx lr
+entry:
+  %a = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %x)
+  %b = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %y)
+  %c = add i32 %a, 10
+  %d = add i32 %b, 10
+  %e = add i32 %c, %d
+  ret i32 %e
+}
+
 declare i32 @llvm.vector.reduce.add.v4i32(<4 x i32>)
 declare i32 @llvm.vector.reduce.add.v8i32(<8 x i32>)
 declare i32 @llvm.vector.reduce.add.v16i32(<16 x i32>)