From 2a661f3f7376b80a9adbf0aaa94c6ecd5cf7b940 Mon Sep 17 00:00:00 2001
From: Tim Northover <tnorthover@apple.com>
Date: Mon, 10 Mar 2014 09:34:07 +0000
Subject: [PATCH] AArch64: fix LowerCONCAT_VECTORS for new CodeGen.

The function was making too many assumptions about its input:

1. The NEON_VDUP optimisation was far too aggressive, assuming (I
think) that the input would always be BUILD_VECTOR.

2. We were treating most unknown concats as legal (by returning Op
rather than SDValue()). I think only concats of pairs of vectors are
actually legal.

http://llvm.org/PR19094

llvm-svn: 203450
---
 llvm/lib/Target/AArch64/AArch64ISelLowering.cpp     | 21 +++++++++++----------
 ...oncatvector-v8i8-bug.ll => concatvector-bugs.ll} | 21 +++++++++++++++++++++
 2 files changed, 32 insertions(+), 10 deletions(-)
 rename llvm/test/CodeGen/AArch64/{concatvector-v8i8-bug.ll => concatvector-bugs.ll} (68%)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 447f500..cf2e46d 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -2281,19 +2281,20 @@ static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
   // We custom lower concat_vectors with 4, 8, or 16 operands that are all the
   // same operand and of type v1* using the DUP instruction.
   unsigned NumOps = Op->getNumOperands();
-  if (NumOps != 4 && NumOps != 8 && NumOps != 16)
+  if (NumOps == 2) {
+    assert(Op.getValueType().getSizeInBits() == 128 && "unexpected concat");
     return Op;
+  }
+
+  if (NumOps != 4 && NumOps != 8 && NumOps != 16)
+    return SDValue();
 
   // Must be a single value for VDUP.
-  bool isConstant = true;
   SDValue Op0 = Op.getOperand(0);
   for (unsigned i = 1; i < NumOps; ++i) {
     SDValue OpN = Op.getOperand(i);
     if (Op0 != OpN)
-      return Op;
-
-    if (!isa<ConstantSDNode>(OpN->getOperand(0)))
-      isConstant = false;
+      return SDValue();
   }
 
   // Verify the value type.
@@ -2302,22 +2303,22 @@ static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
   default: llvm_unreachable("Unexpected number of operands");
   case 4:
     if (EltVT != MVT::v1i16 && EltVT != MVT::v1i32)
-      return Op;
+      return SDValue();
     break;
   case 8:
     if (EltVT != MVT::v1i8 && EltVT != MVT::v1i16)
-      return Op;
+      return SDValue();
     break;
   case 16:
     if (EltVT != MVT::v1i8)
-      return Op;
+      return SDValue();
     break;
   }
 
   SDLoc DL(Op);
   EVT VT = Op.getValueType();
   // VDUP produces better code for constants.
-  if (isConstant)
+  if (Op0->getOpcode() == ISD::BUILD_VECTOR)
     return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Op0->getOperand(0));
   return DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, Op0,
                      DAG.getConstant(0, MVT::i64));
diff --git a/llvm/test/CodeGen/AArch64/concatvector-v8i8-bug.ll b/llvm/test/CodeGen/AArch64/concatvector-bugs.ll
similarity index 68%
rename from llvm/test/CodeGen/AArch64/concatvector-v8i8-bug.ll
rename to llvm/test/CodeGen/AArch64/concatvector-bugs.ll
index f8854c3..5889e22 100644
--- a/llvm/test/CodeGen/AArch64/concatvector-v8i8-bug.ll
+++ b/llvm/test/CodeGen/AArch64/concatvector-bugs.ll
@@ -45,3 +45,24 @@ for.body130.us.us:                                ; preds = %for.body130.us.us,
   br label %for.body130.us.us
 }
 
+declare <1 x i16> @llvm.aarch64.neon.vuqrshrn.v1i16(<1 x i32>, i32)
+
+define <8 x i16> @test_splat(i32 %l) nounwind {
+; CHECK-LABEL: test_splat:
+; CHECK: ret
+  %lhs = insertelement <1 x i32> undef, i32 %l, i32 0
+  %shift = tail call <1 x i16> @llvm.aarch64.neon.vuqrshrn.v1i16(<1 x i32> %lhs, i32 11)
+  %vec = shufflevector <1 x i16> %shift, <1 x i16> undef, <8 x i32> zeroinitializer
+  ret <8 x i16> %vec
+}
+
+
+define <8 x i16> @test_notsplat(<8 x i16> %a, <8 x i16> %b, i32 %l) nounwind {
+; CHECK-LABEL: test_notsplat:
+; CHECK: ret
+entry:
+  %lhs = insertelement <1 x i32> undef, i32 %l, i32 0
+  %shift = tail call <1 x i16> @llvm.aarch64.neon.vuqrshrn.v1i16(<1 x i32> %lhs, i32 11)
+  %vec = shufflevector <1 x i16> %shift, <1 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 0, i32 0, i32 0, i32 0>
+  ret <8 x i16> %vec
+}
-- 
2.7.4