From df08543f4890b5ef42afcfab0d04d4b7ced4d71a Mon Sep 17 00:00:00 2001
From: Ahmed Bougacha <ahmed.bougacha@gmail.com>
Date: Tue, 17 Mar 2015 03:23:09 +0000
Subject: [PATCH] [AArch64] Use intermediate step for concat_vectors of illegal
 truncs.

Optimize concat_vectors of truncated vectors, where the intermediate
type is illegal, to avoid said illegality,  e.g.,
  (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
                         (v2i16 (truncate (v2i64)))))
->
  (v4i16 (truncate (v4i32 (concat_vectors (v2i32 (truncate (v2i64))),
                                          (v2i32 (truncate (v2i64)))))))

This isn't really target-specific, and, as such, would best go in the
DAGCombiner.  However, ISD::TRUNCATE legality isn't keyed on both input
and result type, so we might generate worse code when we don't know
better.  On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
rdar://20022387


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@232459 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AArch64/AArch64ISelLowering.cpp    | 31 +++++++++++++++++++
 .../AArch64/concat_vector-truncate-combine.ll | 31 +++++++++++++++++++
 2 files changed, 62 insertions(+)
 create mode 100644 test/CodeGen/AArch64/concat_vector-truncate-combine.ll

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 43f443fc400..f1073beb32f 100644
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -7180,6 +7180,37 @@ static SDValue performConcatVectorsCombine(SDNode *N,
   EVT VT = N->getValueType(0);
   SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
 
+  // Optimize concat_vectors of truncated vectors, where the intermediate
+  // type is illegal, to avoid said illegality,  e.g.,
+  //   (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
+  //                          (v2i16 (truncate (v2i64)))))
+  // ->
+  //   (v4i16 (truncate (v4i32 (concat_vectors (v2i32 (truncate (v2i64))),
+  //                                           (v2i32 (truncate (v2i64)))))))
+  // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
+  // on both input and result type, so we might generate worse code.
+  // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
+  if (N->getNumOperands() == 2 &&
+      N0->getOpcode() == ISD::TRUNCATE &&
+      N1->getOpcode() == ISD::TRUNCATE) {
+    SDValue N00 = N0->getOperand(0);
+    SDValue N10 = N1->getOperand(0);
+    EVT N00VT = N00.getValueType();
+
+    if (N00VT == N10.getValueType() &&
+        (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
+        N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
+      MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v2i32 : MVT::v4i16);
+      MVT ConcatMidVT = MVT::getVectorVT(MidVT.getVectorElementType(),
+                                         MidVT.getVectorNumElements() * 2);
+      return DAG.getNode(
+          ISD::TRUNCATE, dl, VT,
+          DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatMidVT,
+                      DAG.getNode(ISD::TRUNCATE, dl, MidVT, N00),
+                      DAG.getNode(ISD::TRUNCATE, dl, MidVT, N10)));
+    }
+  }
+
   // Wait 'til after everything is legalized to try this. That way we have
   // legal vector types and such.
   if (DCI.isBeforeLegalizeOps())
diff --git a/test/CodeGen/AArch64/concat_vector-truncate-combine.ll b/test/CodeGen/AArch64/concat_vector-truncate-combine.ll
new file mode 100644
index 00000000000..468aa42f75d
--- /dev/null
+++ b/test/CodeGen/AArch64/concat_vector-truncate-combine.ll
@@ -0,0 +1,31 @@
+; RUN: llc < %s -mtriple arm64-apple-darwin -asm-verbose=false | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+
+define <4 x i16> @test_concat_truncate_v2i64_to_v4i16(<2 x i64> %a, <2 x i64> %b) #0 {
+entry:
+; CHECK-LABEL: test_concat_truncate_v2i64_to_v4i16:
+; CHECK-NEXT: xtn.2s  v0, v0
+; CHECK-NEXT: xtn2.4s v0, v1
+; CHECK-NEXT: xtn.4h  v0, v0
+; CHECK-NEXT: ret
+  %at = trunc <2 x i64> %a to <2 x i16>
+  %bt = trunc <2 x i64> %b to <2 x i16>
+  %shuffle = shufflevector <2 x i16> %at, <2 x i16> %bt, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  ret <4 x i16> %shuffle
+}
+
+define <8 x i8> @test_concat_truncate_v4i32_to_v8i8(<4 x i32> %a, <4 x i32> %b) #0 {
+entry:
+; CHECK-LABEL: test_concat_truncate_v4i32_to_v8i8:
+; CHECK-NEXT: xtn.4h  v0, v0
+; CHECK-NEXT: xtn2.8h v0, v1
+; CHECK-NEXT: xtn.8b  v0, v0
+; CHECK-NEXT: ret
+  %at = trunc <4 x i32> %a to <4 x i8>
+  %bt = trunc <4 x i32> %b to <4 x i8>
+  %shuffle = shufflevector <4 x i8> %at, <4 x i8> %bt, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  ret <8 x i8> %shuffle
+}
+
+attributes #0 = { nounwind }
-- 
2.34.1