[PPC] Use xxbrd to speed up bswap64

Power doesn't have bswap instructions, so llvm generates following code sequence for bswap64.

  rotldi   5, 3, 16
  rotldi   4, 3, 8
  rotldi   9, 3, 24
  rotldi   10, 3, 32
  rotldi   11, 3, 48
  rotldi   12, 3, 56
  rldimi 4, 5, 8, 48
  rldimi 4, 9, 16, 40
  rldimi 4, 10, 24, 32
  rldimi 4, 11, 40, 16
  rldimi 4, 12, 48, 8
  rldimi 4, 3, 56, 0

But Power9 has vector bswap instructions, they can also be used to speed up scalar bswap intrinsic. With this patch, bswap64 can be translated to:

  mtvsrdd 34, 3, 3
  xxbrd 34, 34
  mfvsrld 3, 34

Differential Revision: https://reviews.llvm.org/D39510



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@317499 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 37 insertions, 2 deletions

diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index f3e7b4af45d..15ff7740bbd 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -291,14 +291,16 @@ PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
     setOperationAction(ISD::FROUND, MVT::f32, Legal);
   }
 
-  // PowerPC does not have BSWAP
+  // PowerPC does not have BSWAP, but we can use vector BSWAP instruction xxbrd
+  // to speed up scalar BSWAP64.
   // CTPOP or CTTZ were introduced in P8/P9 respectivelly
   setOperationAction(ISD::BSWAP, MVT::i32  , Expand);
-  setOperationAction(ISD::BSWAP, MVT::i64  , Expand);
   if (Subtarget.isISA3_0()) {
+    setOperationAction(ISD::BSWAP, MVT::i64  , Custom);
     setOperationAction(ISD::CTTZ , MVT::i32  , Legal);
     setOperationAction(ISD::CTTZ , MVT::i64  , Legal);
   } else {
+    setOperationAction(ISD::BSWAP, MVT::i64  , Expand);
     setOperationAction(ISD::CTTZ , MVT::i32  , Expand);
     setOperationAction(ISD::CTTZ , MVT::i64  , Expand);
   }
@@ -8675,6 +8677,23 @@ SDValue PPCTargetLowering::LowerREM(SDValue Op, SelectionDAG &DAG) const {
   return Op;
 }
 
+// Lower scalar BSWAP64 to xxbrd.
+SDValue PPCTargetLowering::LowerBSWAP(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc dl(Op);
+  // MTVSRDD
+  Op = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, Op.getOperand(0),
+                   Op.getOperand(0));
+  // XXBRD
+  Op = DAG.getNode(PPCISD::XXREVERSE, dl, MVT::v2i64, Op);
+  // MFVSRD
+  int VectorIndex = 0;
+  if (Subtarget.isLittleEndian())
+    VectorIndex = 1;
+  Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Op,
+                   DAG.getTargetConstant(VectorIndex, dl, MVT::i32));
+  return Op;
+}
+
 SDValue PPCTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
                                                   SelectionDAG &DAG) const {
   SDLoc dl(Op);
@@ -9146,6 +9165,8 @@ SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SREM:
   case ISD::UREM:
     return LowerREM(Op, DAG);
+  case ISD::BSWAP:
+    return LowerBSWAP(Op, DAG);
   }
 }
 
diff --git a/lib/Target/PowerPC/PPCISelLowering.h b/lib/Target/PowerPC/PPCISelLowering.h
index 1a5efeba4cf..d06c10adb7e 100644
--- a/lib/Target/PowerPC/PPCISelLowering.h
+++ b/lib/Target/PowerPC/PPCISelLowering.h
@@ -953,6 +953,7 @@ namespace llvm {
     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerREM(SDValue Op, SelectionDAG &DAG) const;
+    SDValue LowerBSWAP(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
diff --git a/test/CodeGen/PowerPC/bswap64.ll b/test/CodeGen/PowerPC/bswap64.ll
new file mode 100644
index 00000000000..0a78aa2dc54
--- /dev/null
+++ b/test/CodeGen/PowerPC/bswap64.ll
@@ -0,0 +1,13 @@
+; RUN: llc -verify-machineinstrs < %s -mtriple=ppc64le-- -mcpu=pwr9 | FileCheck %s
+
+declare i64 @llvm.bswap.i64(i64)
+
+; CHECK: mtvsrdd
+; CHECK: xxbrd
+; CHECK: mfvsrd
+define i64 @bswap64(i64 %x) {
+entry:
+  %0 = call i64 @llvm.bswap.i64(i64 %x)
+  ret i64 %0
+}
+