[AArch64] Prefer UZP for concat_vector of illegal truncs.

Follow-up to r232459: prefer a UZP shuffle to the intermediate truncs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@232871 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 93387b00e225838289ad166aa6ae87053c674e2c..25a3b14a64b461f26ed8ec9c888e463b38547ed0 100644 (file)
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -7187,8 +7187,9 @@ static SDValue performConcatVectorsCombine(SDNode *N,
   //   (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
   //                          (v2i16 (truncate (v2i64)))))
   // ->
-  //   (v4i16 (truncate (v4i32 (concat_vectors (v2i32 (truncate (v2i64))),
-  //                                           (v2i32 (truncate (v2i64)))))))
+  //   (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
+  //                                    (v4i32 (bitcast (v2i64))),
+  //                                    <0, 2, 4, 6>)))
   // 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.
@@ -7202,20 +7203,15 @@ static SDValue performConcatVectorsCombine(SDNode *N,
     if (N00VT == N10.getValueType() &&
         (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
         N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
-      MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v2i32 : MVT::v4i16);
-#if defined(__GNUC__)
-#if __GNUC__ == 4 && __GNUC_MINOR__ == 7 && __GNUC_PATCHLEVEL__ == 2
-      // FIXME: g++-4.7.2 might miscompile PerformDAGCombine().
-      asm volatile("":::"memory");
-#endif
-#endif
-      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)));
+      MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
+      SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
+      for (size_t i = 0; i < Mask.size(); ++i)
+        Mask[i] = i * 2;
+      return DAG.getNode(ISD::TRUNCATE, dl, VT,
+                         DAG.getVectorShuffle(
+                             MidVT, dl,
+                             DAG.getNode(ISD::BITCAST, dl, MidVT, N00),
+                             DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask));
     }
   }
 

diff --git a/test/CodeGen/AArch64/concat_vector-truncate-combine.ll b/test/CodeGen/AArch64/concat_vector-truncate-combine.ll
index 468aa42f75d7fb7441f5d3822691deef4497e06c..c510e279fd09c32eb0c69c3d8ee962cab5cfc008 100644 (file)
--- a/test/CodeGen/AArch64/concat_vector-truncate-combine.ll
+++ b/test/CodeGen/AArch64/concat_vector-truncate-combine.ll
@@ -5,9 +5,8 @@ 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: uzp1.4s v0, 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>
@@ -18,9 +17,8 @@ entry:
 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: uzp1.8h v0, 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>
@@ -28,4 +26,16 @@ entry:
   ret <8 x i8> %shuffle
 }
 
+define <8 x i16> @test_concat_truncate_v4i32_to_v8i16(<4 x i32> %a, <4 x i32> %b) #0 {
+entry:
+; CHECK-LABEL: test_concat_truncate_v4i32_to_v8i16:
+; CHECK-NEXT: xtn.4h v0, v0
+; CHECK-NEXT: xtn2.8h v0, v1
+; CHECK-NEXT: ret
+  %at = trunc <4 x i32> %a to <4 x i16>
+  %bt = trunc <4 x i32> %b to <4 x i16>
+  %shuffle = shufflevector <4 x i16> %at, <4 x i16> %bt, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  ret <8 x i16> %shuffle
+}
+
 attributes #0 = { nounwind }