AArch64: Fix loads to lower NEON vector lanes using GPR registers

[MatzeB]

Summary

The ISelLowering code turned insertion turned the element for the
lowest lane of a BUILD_VECTOR into an INSERT_SUBREG, this prohibited
the patterns for SCALAR_TO_VECTOR(Load) to match later. Restrict this
to cases without a load argument.

Reported in rdar://22223823

[jmolloy]

Hi Matthias,

This looks fine to me.

Cheers,

James

llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp

if (!isConstant && !usesOnlyOneValue) {
     SDValue Vec = DAG.getUNDEF(VT);
     SDValue Op0 = Op.getOperand(0);
     unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
     unsigned i = 0;
     // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
     // a) Avoid a RMW dependency on the full vector register, and
     // b) Allow the register coalescer to fold away the copy if the
     //    value is already in an S or D register.
-    if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
+    // Do not do this for UNDEF/LOAD nodes because we have better patterns
+    // for those avoiding the SCALAR_TO_VECTOR/BUILD_VECTOR.
+    if (Op0.getOpcode() != ISD::UNDEF && Op0.getOpcode() != ISD::LOAD &&
+        (ElemSize == 32 || ElemSize == 64)) {
       unsigned SubIdx = ElemSize == 32 ? AArch64::ssub : AArch64::dsub;
       MachineSDNode *N =
           DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
                              DAG.getTargetConstant(SubIdx, dl, MVT::i32));
       Vec = SDValue(N, 0);
       ++i;
     }
     for (; i < NumElts; ++i) {

llvm/trunk/test/CodeGen/AArch64/arm64-indexed-vector-ldst.ll

; CHECK: ld1.h  { v0 }[1], [x0], x{{[0-9]+}}
   store i16* %tmp3, i16** %ptr
   %dl =  load <2 x i32>,  <2 x i32>* %d
   %dr = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> %dl)
   store <2 x i32> %dr, <2 x i32>* %d
   ret <4 x i16> %tmp2
 }
 
 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32>)
+
+; CHECK-LABEL: test_ld1lane_build:
+; CHECK-DAG: ld1.s { [[REG0:v[0-9]+]] }[0], [x0]
+; CHECK-DAG: ld1.s { [[REG0:v[0-9]+]] }[1], [x1]
+; CHECK-DAG: ld1.s { [[REG1:v[0-9]+]] }[0], [x2]
+; CHECK-DAG: ld1.s { [[REG1:v[0-9]+]] }[1], [x3]
+; CHECK: sub.2s v[[REGNUM2:[0-9]+]], [[REG0]], [[REG1]]
+; CHECK-NEXT: str d[[REGNUM2]], [x4]
+; CHECK-NEXT: ret
+define void @test_ld1lane_build(i32* %ptr0, i32* %ptr1, i32* %ptr2, i32* %ptr3, <2 x i32>* %out) {
+  %load0 = load i32, i32* %ptr0, align 4
+  %load1 = load i32, i32* %ptr1, align 4
+  %vec0_0 = insertelement <2 x i32> undef, i32 %load0, i32 0
+  %vec0_1 = insertelement <2 x i32> %vec0_0, i32 %load1, i32 1
+
+  %load2 = load i32, i32* %ptr2, align 4
+  %load3 = load i32, i32* %ptr3, align 4
+  %vec1_0 = insertelement <2 x i32> undef, i32 %load2, i32 0
+  %vec1_1 = insertelement <2 x i32> %vec1_0, i32 %load3, i32 1
+
+  %sub = sub nsw <2 x i32> %vec0_1, %vec1_1
+  store <2 x i32> %sub, <2 x i32>* %out, align 16
+  ret void
+}