From 6b96a388eda37bc4e2ca71e029f483de53090868 Mon Sep 17 00:00:00 2001
From: Ahmed Bougacha <ahmed.bougacha@gmail.com>
Date: Fri, 17 Apr 2015 23:43:33 +0000
Subject: [PATCH] [AArch64] Don't force MVT::Untyped when selecting
 LD1LANEpost.

The result is either an Untyped reg sequence, on ldN with N > 1, or
just the type of the input vector, on ld1.  Don't force Untyped.
Instead, just use the type of the reg sequence.

This mirrors the behavior of createTuple, which feeds the LD1*_POST.

The narrow code path wasn't actually covered by tests, because V64
insert_vector_elt are widened to V128 before the LD1LANEpost combine
has the chance to run, usually.

The only case where it does run on V64 vectors is if the vector ops
legalizer ran.  So, tickle the code with a ctpop.

Fixes PR23265.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@235243 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64ISelDAGToDAG.cpp    |  2 +-
 .../AArch64/arm64-indexed-vector-ldst.ll      | 22 +++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp b/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
index f75700d6d33..610e7cf63c9 100644
--- a/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
+++ b/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
@@ -1234,7 +1234,7 @@ SDNode *AArch64DAGToDAGISel::SelectPostLoadLane(SDNode *N, unsigned NumVecs,
   SDValue RegSeq = createQTuple(Regs);
 
   const EVT ResTys[] = {MVT::i64, // Type of the write back register
-                        MVT::Untyped, MVT::Other};
+                        RegSeq->getValueType(0), MVT::Other};
 
   unsigned LaneNo =
       cast<ConstantSDNode>(N->getOperand(NumVecs + 1))->getZExtValue();
diff --git a/test/CodeGen/AArch64/arm64-indexed-vector-ldst.ll b/test/CodeGen/AArch64/arm64-indexed-vector-ldst.ll
index 3cfbb1433c5..ba31513172d 100644
--- a/test/CodeGen/AArch64/arm64-indexed-vector-ldst.ll
+++ b/test/CodeGen/AArch64/arm64-indexed-vector-ldst.ll
@@ -6193,3 +6193,25 @@ define <4 x float> @test_v4f32_post_reg_ld1lane_dep_vec_on_load(float* %bar, flo
   store float* %tmp3, float** %ptr
   ret <4 x float> %tmp2
 }
+
+; Make sure that we test the narrow V64 code path.
+; The tests above don't, because there, 64-bit insert_vector_elt nodes will be
+; widened to 128-bit before the LD1LANEpost combine has the chance to run,
+; making it avoid narrow vector types.
+; One way to trick that combine into running early is to force the vector ops
+; legalizer to run.  We achieve that using the ctpop.
+; PR23265
+define <4 x i16> @test_v4i16_post_reg_ld1lane_forced_narrow(i16* %bar, i16** %ptr, i64 %inc, <4 x i16> %A, <2 x i32>* %d) {
+; CHECK-LABEL: test_v4i16_post_reg_ld1lane_forced_narrow:
+; CHECK: ld1.h  { v0 }[1], [x0], x{{[0-9]+}}
+  %tmp1 = load i16, i16* %bar
+  %tmp2 = insertelement <4 x i16> %A, i16 %tmp1, i32 1
+  %tmp3 = getelementptr i16, i16* %bar, i64 %inc
+  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>)