[x86][aarch64] ask the backend whether it has a vector blend instruction

The code to match and produce more x86 vector blends was enabled for all
architectures even though the transform may pessimize the code for other
architectures that do not provide a vector blend instruction.

Added an aarch64 testcase to check that a VZIP instruction is generated instead
of byte movs.

Differential Revision: https://reviews.llvm.org/D44118

llvm-svn: 327132

5 files changed, 43 insertions, 26 deletions

diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 7e03f42445b..29b0384be8c 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -2117,6 +2117,9 @@ public:
     return false;
   }
 
+  /// Return true if the target has a vector blend instruction.
+  virtual bool hasVectorBlend() const { return false; }
+
   /// \brief Get the maximum supported factor for interleaved memory accesses.
   /// Default to be the minimum interleave factor: 2.
   virtual unsigned getMaxSupportedInterleaveFactor() const { return 2; }
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 9d0ea385792..aac79075e6e 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -1557,33 +1557,35 @@ SDValue SelectionDAG::getVectorShuffle(EVT VT, const SDLoc &dl, SDValue N1,
   if (N1.isUndef())
     commuteShuffle(N1, N2, MaskVec);
 
-  // If shuffling a splat, try to blend the splat instead. We do this here so
-  // that even when this arises during lowering we don't have to re-handle it.
-  auto BlendSplat = [&](BuildVectorSDNode *BV, int Offset) {
-    BitVector UndefElements;
-    SDValue Splat = BV->getSplatValue(&UndefElements);
-    if (!Splat)
-      return;
+  if (TLI->hasVectorBlend()) {
+    // If shuffling a splat, try to blend the splat instead. We do this here so
+    // that even when this arises during lowering we don't have to re-handle it.
+    auto BlendSplat = [&](BuildVectorSDNode *BV, int Offset) {
+      BitVector UndefElements;
+      SDValue Splat = BV->getSplatValue(&UndefElements);
+      if (!Splat)
+        return;
 
-    for (int i = 0; i < NElts; ++i) {
-      if (MaskVec[i] < Offset || MaskVec[i] >= (Offset + NElts))
-        continue;
+      for (int i = 0; i < NElts; ++i) {
+        if (MaskVec[i] < Offset || MaskVec[i] >= (Offset + NElts))
+          continue;
 
-      // If this input comes from undef, mark it as such.
-      if (UndefElements[MaskVec[i] - Offset]) {
-        MaskVec[i] = -1;
-        continue;
-      }
+        // If this input comes from undef, mark it as such.
+        if (UndefElements[MaskVec[i] - Offset]) {
+          MaskVec[i] = -1;
+          continue;
+        }
 
-      // If we can blend a non-undef lane, use that instead.
-      if (!UndefElements[i])
-        MaskVec[i] = i + Offset;
-    }
-  };
-  if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
-    BlendSplat(N1BV, 0);
-  if (auto *N2BV = dyn_cast<BuildVectorSDNode>(N2))
-    BlendSplat(N2BV, NElts);
+        // If we can blend a non-undef lane, use that instead.
+        if (!UndefElements[i])
+          MaskVec[i] = i + Offset;
+      }
+    };
+    if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
+      BlendSplat(N1BV, 0);
+    if (auto *N2BV = dyn_cast<BuildVectorSDNode>(N2))
+      BlendSplat(N2BV, NElts);
+  }
 
   // Canonicalize all index into lhs, -> shuffle lhs, undef
   // Canonicalize all index into rhs, -> shuffle rhs, undef
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index df82c5a5b82..6181dbd2a6a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -1098,6 +1098,8 @@ namespace llvm {
 
     StringRef getStackProbeSymbolName(MachineFunction &MF) const override;
 
+    bool hasVectorBlend() const override { return true; }
+
     unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
 
     /// \brief Lower interleaved load(s) into target specific
diff --git a/llvm/test/CodeGen/AArch64/aarch64-vuzp.ll b/llvm/test/CodeGen/AArch64/aarch64-vuzp.ll
index cc91873a75b..a7b20f25557 100644
--- a/llvm/test/CodeGen/AArch64/aarch64-vuzp.ll
+++ b/llvm/test/CodeGen/AArch64/aarch64-vuzp.ll
@@ -58,3 +58,13 @@ entry:
   store <4 x i32> %y, <4 x i32>* %z, align 4
   ret void
 }
+
+; Check that this pattern is recognized as a VZIP and
+; that the vector blend transform does not scramble the pattern.
+; CHECK-LABEL: vzipNoBlend:
+; CHECK: zip1
+define <8 x i8> @vzipNoBlend(<8 x i8>* %A, <8 x i16>* %B) nounwind {
+  %t = load <8 x i8>, <8 x i8>* %A
+  %vzip = shufflevector <8 x i8> %t, <8 x i8> <i8 0, i8 0, i8 0, i8 0, i8 undef, i8 undef, i8 undef, i8 undef>, <8 x i32> <i32 0, i32 8, i32 1, i32 9, i32 2, i32 10, i32 3, i32 11>
+  ret <8 x i8> %vzip
+}
diff --git a/llvm/test/CodeGen/AArch64/arm64-collect-loh.ll b/llvm/test/CodeGen/AArch64/arm64-collect-loh.ll
index c7ba989d933..dc34ee243b2 100644
--- a/llvm/test/CodeGen/AArch64/arm64-collect-loh.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-collect-loh.ll
@@ -638,13 +638,13 @@ define void @setL(<1 x i8> %t) {
 ; CHECK: [[LOH_LABEL1:Lloh[0-9]+]]:
 ; CHECK: ldr q[[IDX:[0-9]+]], {{\[}}[[ADRP_REG]], [[CONSTPOOL]]@PAGEOFF]
 ; The tuple comes from the next instruction.
-; CHECK-NEXT: tbl.16b v{{[0-9]+}}, { v{{[0-9]+}}, v{{[0-9]+}} }, v[[IDX]]
+; CHECK: ext.16b v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, #1
 ; CHECK: ret
 ; CHECK: .loh AdrpLdr [[LOH_LABEL0]], [[LOH_LABEL1]]
 define void @uninterestingSub(i8* nocapture %row) #0 {
   %tmp = bitcast i8* %row to <16 x i8>*
   %tmp1 = load <16 x i8>, <16 x i8>* %tmp, align 16
-  %vext43 = shufflevector <16 x i8> <i8 undef, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0>, <16 x i8> %tmp1, <16 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16>
+  %vext43 = shufflevector <16 x i8> <i8 undef, i8 16, i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2>, <16 x i8> %tmp1, <16 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16>
   %add.i.414 = add <16 x i8> zeroinitializer, %vext43
   store <16 x i8> %add.i.414, <16 x i8>* %tmp, align 16
   %add.ptr51 = getelementptr inbounds i8, i8* %row, i64 16