[DAGISel][ARM] Fix vector truncate combine for big-endian (#118101)

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

@@ -15498,12 +15498,14 @@ SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
       unsigned BuildVecNumElts =  BuildVect.getNumOperands();
       unsigned TruncVecNumElts = VT.getVectorNumElements();
       unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
+      unsigned FirstElt = isLE ? 0 : (TruncEltOffset - 1);
 
       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/vector-trunc.ll

@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=armv7-unknown-none-eabihf -mattr=+neon < %s | FileCheck %s --check-prefix=LE
+; RUN: llc -mtriple=armebv7-unknown-none-eabihf -mattr=+neon < %s | FileCheck %s --check-prefix=BE
+
+define i32 @test(i64 %arg1) {
+; LE-LABEL: test:
+; LE:       @ %bb.0: @ %entry
+; LE-NEXT:    subs r0, r0, #1
+; LE-NEXT:    mov r2, #0
+; LE-NEXT:    sbcs r0, r1, #0
+; LE-NEXT:    vldr s0, .LCPI0_0
+; LE-NEXT:    movwhs r2, #1
+; LE-NEXT:    cmp r2, #0
+; LE-NEXT:    mvnne r2, #0
+; LE-NEXT:    vmov s1, r2
+; LE-NEXT:    vmovn.i32 d16, q0
+; LE-NEXT:    vmovn.i16 d16, q8
+; LE-NEXT:    vmov.u8 r0, d16[0]
+; LE-NEXT:    and r0, r0, #1
+; LE-NEXT:    bx lr
+; LE-NEXT:    .p2align 2
+; LE-NEXT:  @ %bb.1:
+; LE-NEXT:  .LCPI0_0:
+; LE-NEXT:    .long 0xffffffff @ float NaN
+;
+; BE-LABEL: test:
+; BE:       @ %bb.0: @ %entry
+; BE-NEXT:    subs r1, r1, #1
+; BE-NEXT:    mov r2, #0
+; BE-NEXT:    sbcs r0, r0, #0
+; BE-NEXT:    vldr s0, .LCPI0_0
+; BE-NEXT:    movwhs r2, #1
+; BE-NEXT:    cmp r2, #0
+; BE-NEXT:    mvnne r2, #0
+; BE-NEXT:    vmov s1, r2
+; BE-NEXT:    vmovn.i32 d16, q0
+; BE-NEXT:    vmovn.i16 d16, q8
+; BE-NEXT:    vmov.u8 r0, d16[0]
+; BE-NEXT:    and r0, r0, #1
+; BE-NEXT:    bx lr
+; BE-NEXT:    .p2align 2
+; BE-NEXT:  @ %bb.1:
+; BE-NEXT:  .LCPI0_0:
+; BE-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
+  %cmp_vec = icmp ule <8 x i64> <i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6, i64 7>, %splat_zero
+  %first_cmp = extractelement <8 x i1> %cmp_vec, i32 0
+  %ext = zext i1 %first_cmp to i32
+  ret i32 %ext
+}