[DAGISel][ARM] Fix vector truncate combine for big-endian

This DAG combine was incorrect for big-endian targets, because it
assumes that when a bitcast changes the lane width, the
least-significant bits of the wider lanes are in the lower-numbered
lanes of the smaller type, which is only true for little-endian.

Author: ostannard

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -15495,12 +15495,15 @@ SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
       unsigned BuildVecNumElts =  BuildVect.getNumOperands();
       unsigned TruncVecNumElts = VT.getVectorNumElements();
       unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
+      unsigned FirstElt =
+          DAG.getDataLayout().isBigEndian() ? (TruncEltOffset - 1) : 0;
 
       assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
              "Invalid number of elements");
 
       SmallVector<SDValue, 8> Opnds;
-      for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
+      for (unsigned i = FirstElt, e = BuildVecNumElts; i < e;
+           i += TruncEltOffset)
         Opnds.push_back(BuildVect.getOperand(i));
 
       return DAG.getBuildVector(VT, DL, Opnds);

llvm/test/CodeGen/ARM/big-endian-vector-trunc.ll

@@ -4,8 +4,23 @@
 define i32 @test(i64 %arg1) "target-features"="+neon" {
 ; CHECK-LABEL: test:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    mov r0, #0
+; CHECK-NEXT:    subs r1, r1, #1
+; CHECK-NEXT:    mov r2, #0
+; CHECK-NEXT:    sbcs r0, r0, #0
+; CHECK-NEXT:    vldr s0, .LCPI0_0
+; CHECK-NEXT:    movwhs r2, #1
+; CHECK-NEXT:    cmp r2, #0
+; CHECK-NEXT:    mvnne r2, #0
+; CHECK-NEXT:    vmov s1, r2
+; CHECK-NEXT:    vmovn.i32 d16, q0
+; CHECK-NEXT:    vmovn.i16 d16, q8
+; CHECK-NEXT:    vmov.u8 r0, d16[0]
+; CHECK-NEXT:    and r0, r0, #1
 ; CHECK-NEXT:    bx lr
+; CHECK-NEXT:    .p2align 2
+; CHECK-NEXT:  @ %bb.1:
+; CHECK-NEXT:  .LCPI0_0:
+; CHECK-NEXT:    .long 0xffffffff @ float NaN
 entry:
   %insert_zero = insertelement <8 x i64> poison, i64 %arg1, i64 0
   %splat_zero = shufflevector <8 x i64> %insert_zero, <8 x i64> poison, <8 x i32> zeroinitializer