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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203450 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 447f5005e555c28238354effb29c31d740a3b5a1..cf2e46dbe3944b1fbd225d014ee9a3f7da759da9 100644 (file)
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/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/test/CodeGen/AArch64/concatvector-bugs.ll b/test/CodeGen/AArch64/concatvector-bugs.ll
new file mode 100644 (file)
index 0000000..5889e22
--- /dev/null
+++ b/test/CodeGen/AArch64/concatvector-bugs.ll
@@ -0,0 +1,68 @@
+; RUN: llc < %s -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mattr=+neon
+; Bug: i8 type in FRP8 register but not registering with register class causes segmentation fault.
+; Fix: Removed i8 type from FPR8 register class.
+
+define void @test_concatvector_v8i8() {
+entry.split:
+  br i1 undef, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry.split
+  unreachable
+
+if.end:                                           ; preds = %entry.split
+  br i1 undef, label %if.then9, label %if.end18
+
+if.then9:                                         ; preds = %if.end
+  unreachable
+
+if.end18:                                         ; preds = %if.end
+  br label %for.body
+
+for.body:                                         ; preds = %for.inc, %if.end18
+  br i1 false, label %if.then30, label %for.inc
+
+if.then30:                                        ; preds = %for.body
+  unreachable
+
+for.inc:                                          ; preds = %for.body
+  br i1 undef, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.inc
+  br label %for.body77
+
+for.body77:                                       ; preds = %for.body77, %for.end
+  br i1 undef, label %for.end106, label %for.body77
+
+for.end106:                                       ; preds = %for.body77
+  br i1 undef, label %for.body130.us.us, label %stmt.for.body130.us.us
+
+stmt.for.body130.us.us:                     ; preds = %stmt.for.body130.us.us, %for.end106
+  %_p_splat.us = shufflevector <1 x i8> zeroinitializer, <1 x i8> undef, <8 x i32> zeroinitializer
+  store <8 x i8> %_p_splat.us, <8 x i8>* undef, align 1
+  br label %stmt.for.body130.us.us
+
+for.body130.us.us:                                ; preds = %for.body130.us.us, %for.end106
+  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
+}

diff --git a/test/CodeGen/AArch64/concatvector-v8i8-bug.ll b/test/CodeGen/AArch64/concatvector-v8i8-bug.ll
deleted file mode 100644 (file)
index f8854c3..0000000
--- a/test/CodeGen/AArch64/concatvector-v8i8-bug.ll
+++ /dev/null
@@ -1,47 +0,0 @@
-; RUN: llc < %s -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mattr=+neon
-; Bug: i8 type in FRP8 register but not registering with register class causes segmentation fault.
-; Fix: Removed i8 type from FPR8 register class.
-
-define void @test_concatvector_v8i8() {
-entry.split:
-  br i1 undef, label %if.then, label %if.end
-
-if.then:                                          ; preds = %entry.split
-  unreachable
-
-if.end:                                           ; preds = %entry.split
-  br i1 undef, label %if.then9, label %if.end18
-
-if.then9:                                         ; preds = %if.end
-  unreachable
-
-if.end18:                                         ; preds = %if.end
-  br label %for.body
-
-for.body:                                         ; preds = %for.inc, %if.end18
-  br i1 false, label %if.then30, label %for.inc
-
-if.then30:                                        ; preds = %for.body
-  unreachable
-
-for.inc:                                          ; preds = %for.body
-  br i1 undef, label %for.end, label %for.body
-
-for.end:                                          ; preds = %for.inc
-  br label %for.body77
-
-for.body77:                                       ; preds = %for.body77, %for.end
-  br i1 undef, label %for.end106, label %for.body77
-
-for.end106:                                       ; preds = %for.body77
-  br i1 undef, label %for.body130.us.us, label %stmt.for.body130.us.us
-
-stmt.for.body130.us.us:                     ; preds = %stmt.for.body130.us.us, %for.end106
-  %_p_splat.us = shufflevector <1 x i8> zeroinitializer, <1 x i8> undef, <8 x i32> zeroinitializer
-  store <8 x i8> %_p_splat.us, <8 x i8>* undef, align 1
-  br label %stmt.for.body130.us.us
-
-for.body130.us.us:                                ; preds = %for.body130.us.us, %for.end106
-  br label %for.body130.us.us
-}
-