[DAGComb] Do not turn insert_elt into shuffle for single elt vectors.

[fhahn]

Summary

Currently combineInsertEltToShuffle turns insert_vector_elt into a
vector_shuffle, even if the inserted element is a vector with a single
element. In this case, it should be unlikely that the additional shuffle
would be more efficient than a insert_vector_elt.

Additionally, this fixes a infinite cycle in DAGCombine, where
combineInsertEltToShuffle turns a insert_vector_elt into a shuffle,
which gets turned back into a insert_vector_elt/extract_vector_elt by
a custom AArch64 lowering (in visitVECTOR_SHUFFLE).

Such insert_vector_elt and extract_vector_elt combinations can be
lowered efficiently using mov on AArch64.

There are 2 test changes in arm64-neon-copy.ll: we now use one or two
mov instructions instead of a single zip1. The reason that we need a
second mov in ins1f2 is that we have to move the result to the result
register and is not really related to the DAGCombine fold I think.
But in any case, on most uarchs, mov should be cheaper than zip1. On a
Cortex-A75 for example, zip1 is twice as expensive as mov
(https://developer.arm.com/docs/101398/latest/arm-cortex-a75-software-optimization-guide-v20)

[RKSimon]

LGTM

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (InsertVal.getOpcode() != ISD::BITCAST || !InsertVal.hasOneUse() ||
       !InsertVal.getOperand(0).getValueType().isVector())
     return SDValue();
 
   SDValue SubVec = InsertVal.getOperand(0);
   SDValue DestVec = N->getOperand(0);
   EVT SubVecVT = SubVec.getValueType();
   EVT VT = DestVec.getValueType();
   unsigned NumSrcElts = SubVecVT.getVectorNumElements();
+  // If the source only has a single vector element, the cost of creating adding
+  // it to a vector is likely to exceed the cost of a insert_vector_elt.
+  if (NumSrcElts == 1)
+    return SDValue();
   unsigned ExtendRatio = VT.getSizeInBits() / SubVecVT.getSizeInBits();
   unsigned NumMaskVals = ExtendRatio * NumSrcElts;
 
   // Step 1: Create a shuffle mask that implements this insert operation. The
   // vector that we are inserting into will be operand 0 of the shuffle, so
   // those elements are just 'i'. The inserted subvector is in the first
   // positions of operand 1 of the shuffle. Example:
   // insert v4i32 V, (v2i16 X), 2 --> shuffle v8i16 V', X', {0,1,2,3,8,9,6,7}

llvm/test/CodeGen/AArch64/arm64-neon-copy.ll

; CHECK-NEXT:    ret
   %tmp4 = insertelement <4 x float> %tmp2, float %tmp3, i32 1
   ret <4 x float> %tmp4
 }
 
 define <2 x double> @ins1f2(<1 x double> %tmp1, <2 x double> %tmp2) {
 ; CHECK-LABEL: ins1f2:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    // kill: def $d0 killed $d0 def $q0
-; CHECK-NEXT:    zip1 v0.2d, v1.2d, v0.2d
+; CHECK-NEXT:    mov v1.d[1], v0.d[0]
+; CHECK-NEXT:    mov v0.16b, v1.16b
 ; CHECK-NEXT:    ret
   %tmp3 = extractelement <1 x double> %tmp1, i32 0
   %tmp4 = insertelement <2 x double> %tmp2, double %tmp3, i32 1
   ret <2 x double> %tmp4
 }
 
 define <2 x double> @ins1f2_args_flipped(<2 x double> %tmp2, <1 x double> %tmp1) {
 ; CHECK-LABEL: ins1f2_args_flipped:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    // kill:  def $d1 killed $d1 def $q1
-; CHECK-NEXT:    zip1 v0.2d, v0.2d, v1.2d
+; CHECK-NEXT:    mov v0.d[1], v1.d[0]
 ; CHECK-NEXT:    ret
   %tmp3 = extractelement <1 x double> %tmp1, i32 0
   %tmp4 = insertelement <2 x double> %tmp2, double %tmp3, i32 1
   ret <2 x double> %tmp4
 }
 
 define <8 x i8> @ins16b8(<16 x i8> %tmp1, <8 x i8> %tmp2) {
 ; CHECK-LABEL: ins16b8:

llvm/test/CodeGen/AArch64/vector-insert-shuffle-cycle.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc %s -o - | FileCheck %s
+
+target triple = "arm64-apple-ios13.4.0"
+
+; Make we do not get stuck in a cycle in DAGCombiner.
+
+define void @test(i1 %c, <1 x double>* %ptr) {
+; CHECK-LABEL: test:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    movi d0, #0000000000000000
+; CHECK-NEXT:    tbz w0, #0, LBB0_2
+; CHECK-NEXT:  ; %bb.1: ; %bb1
+; CHECK-NEXT:    ldr d0, [x1]
+; CHECK-NEXT:  LBB0_2: ; %bb2
+; CHECK-NEXT:    ldr q1, [x8]
+; CHECK-NEXT:    mov.d v1[0], v0[0]
+; CHECK-NEXT:    str q1, [x8]
+; CHECK-NEXT:    ret
+entry:
+  br i1 %c, label %bb1, label %bb2
+
+bb1:
+  %lv1 = load <1 x double>, <1 x double>* %ptr, align 16
+  br label %bb2
+
+bb2:
+  %p = phi <1 x double> [ %lv1, %bb1 ], [ zeroinitializer, %entry ]
+  %vecext19 = extractelement <1 x double> %p, i32 0
+  %arrayidx21 = getelementptr inbounds [4 x <4 x double>], [4 x <4 x double>]* undef, i64 0, i64 3
+  %lv2 = load <4 x double>, <4 x double>* %arrayidx21, align 16
+  %vecins22 = insertelement <4 x double> %lv2, double %vecext19, i32 2
+  store <4 x double> %vecins22, <4 x double>* %arrayidx21, align 16
+  ret void
+}