From 10eb2dd9df81f7d65f8157e4c61e41773c50c767 Mon Sep 17 00:00:00 2001
From: Elena Demikhovsky <elena.demikhovsky@intel.com>
Date: Wed, 3 Jun 2015 11:21:01 +0000
Subject: [PATCH] AVX-512: VSHUFPD instruction selection - code improvements

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@238918 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp | 70 +++++++++++++++---------------
 test/CodeGen/X86/avx512-shuffle.ll |  8 ++++
 2 files changed, 44 insertions(+), 34 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5e1441ff7f8..0b9162511a6 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9468,6 +9468,37 @@ static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
   return true;
 }
 
+static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
+                                            ArrayRef<int> Mask, SDValue V1,
+                                            SDValue V2, SelectionDAG &DAG) {
+
+  // Mask for V8F64: 0/1,  8/9,  2/3,  10/11, 4/5, ..
+  // Mask for V4F64; 0/1,  4/5,  2/3,  6/7..
+  assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
+  int NumElts = VT.getVectorNumElements();
+  bool ShufpdMask = true;
+  bool CommutableMask = true;
+  unsigned Immediate = 0;
+  for (int i = 0; i < NumElts; ++i) {
+    if (Mask[i] < 0)
+      continue;
+    int Val = (i & 6) + NumElts * (i & 1);
+    int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
+    if (Mask[i] < Val ||  Mask[i] > Val + 1)
+      ShufpdMask = false;
+    if (Mask[i] < CommutVal ||  Mask[i] > CommutVal + 1)
+      CommutableMask = false;
+    Immediate |= (Mask[i] % 2) << i;
+  }
+  if (ShufpdMask)
+    return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
+                       DAG.getConstant(Immediate, DL, MVT::i8));
+  if (CommutableMask)
+    return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
+                       DAG.getConstant(Immediate, DL, MVT::i8));
+  return SDValue();
+}
+
 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
 ///
 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
@@ -9532,24 +9563,9 @@ static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
     return Blend;
 
   // Check if the blend happens to exactly fit that of SHUFPD.
-  if ((Mask[0] == -1 || Mask[0] < 2) &&
-      (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
-      (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
-      (Mask[3] == -1 || Mask[3] >= 6)) {
-    unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
-                          ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
-    return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
-                       DAG.getConstant(SHUFPDMask, DL, MVT::i8));
-  }
-  if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
-      (Mask[1] == -1 || Mask[1] < 2) &&
-      (Mask[2] == -1 || Mask[2] >= 6) &&
-      (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
-    unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
-                          ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
-    return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
-                       DAG.getConstant(SHUFPDMask, DL, MVT::i8));
-  }
+  if (SDValue Op =
+      lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
+    return Op;
 
   // Try to simplify this by merging 128-bit lanes to enable a lane-based
   // shuffle. However, if we have AVX2 and either inputs are already in place,
@@ -10156,22 +10172,8 @@ static SDValue lowerV8X64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
   if (SDValue Op = lowerVectorShuffleWithVALIGN(DL, VT, Mask, V1, V2, DAG))
     return Op;
 
-  // VSHUFPD instruction - mask 0/1, 8/9, 2/3, 10/11, 4/5, 12/13, 6/7, 14/15
-  bool ShufpdMask = true;
-  unsigned Immediate = 0;
-  for (int i = 0; i < 8; ++i) {
-    if (Mask[i] < 0)
-      continue;
-    int Val = (i & 6) + 8 * (i & 1);
-    if (Mask[i] < Val ||  Mask[i] > Val+1) {
-      ShufpdMask = false;
-      break;
-    }
-    Immediate |= (Mask[i]%2) << i;
-  }
-  if (ShufpdMask)
-    return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
-                       DAG.getConstant(Immediate, DL, MVT::i8));
+  if (SDValue Op = lowerVectorShuffleWithSHUFPD(DL, VT, Mask, V1, V2, DAG))
+    return Op;
 
   // PERMILPD instruction - mask 0/1, 0/1, 2/3, 2/3, 4/5, 4/5, 6/7, 6/7
   if (isSingleInputShuffleMask(Mask)) {
diff --git a/test/CodeGen/X86/avx512-shuffle.ll b/test/CodeGen/X86/avx512-shuffle.ll
index 5d0de6f2d4e..8411fee1502 100644
--- a/test/CodeGen/X86/avx512-shuffle.ll
+++ b/test/CodeGen/X86/avx512-shuffle.ll
@@ -242,3 +242,11 @@ define <16 x i32> @test31(<16 x i32> %a, <16 x i32> %b) nounwind {
   %c = shufflevector <16 x i32> %a, <16 x i32> %b, <16 x i32> <i32 3, i32 4, i32 5, i32 undef, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18>
   ret <16 x i32> %c
 }
+
+; CHECK-LABEL: test32
+; CHECK: vshufpd $99, %zmm0, %zmm1
+; CHECK: ret
+define <8 x double> @test32(<8 x double> %a, <8 x double> %b) nounwind {
+  %c = shufflevector <8 x double> %a, <8 x double> %b, <8 x i32> <i32 9, i32 1, i32 10, i32 2, i32 undef, i32 5, i32 15, i32 undef>
+  ret <8 x double> %c
+}
-- 
2.34.1