[PPC] Use xxbrd to speed up bswap64

[Carrot]

Summary

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

[nemanjai]

This is a great idea considering direct moves are so fast on Power9. I guess we just didn't think of this use when we implemented the vector byte reversal. Thanks for doing this. Other than the rather obvious change to generate the faster mfvsrd instruction, this LGTM.

lib/Target/PowerPC/PPCISelLowering.cpp

8571

Extracting LE doubleword 1 is probably better. It'll produce mfvsrd rather than mfvsrld on LE systems. The latter uses the permute pipeline and is potentially a higher-latency instruction. And it shouldn't make a functional difference since you're populating both doublewords.

[Carrot]

Will check in this version.

lib/Target/PowerPC/PPCISelLowering.h

private:
     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
     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;
 
     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
 
     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,

lib/Target/PowerPC/PPCISelLowering.cpp

if (Subtarget.hasFPRND()) {
     setOperationAction(ISD::FROUND, MVT::f64, Legal);
 
     setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
     setOperationAction(ISD::FCEIL,  MVT::f32, Legal);
     setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
     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);
   }
 
   if (Subtarget.hasPOPCNTD() == PPCSubtarget::POPCNTD_Fast) {
     setOperationAction(ISD::CTPOP, MVT::i32  , Legal);
     setOperationAction(ISD::CTPOP, MVT::i64  , Legal);
   } else {
if ((Op.getOpcode() == ISD::SREM && UI->getOpcode() == ISD::SDIV) ||
         (Op.getOpcode() == ISD::UREM && UI->getOpcode() == ISD::UDIV))
       if (UI->getOperand(0) == Op.getOperand(0) &&
           UI->getOperand(1) == Op.getOperand(1))
         return SDValue();
   }
   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);
+  // MFVSRLD or MFVSRD
+  Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Op,
+                   DAG.getTargetConstant(0, dl, MVT::i32));
+  return Op;
+}
+
 SDValue PPCTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
                                                   SelectionDAG &DAG) const {
   SDLoc dl(Op);
   // For v2i64 (VSX), we can pattern patch the v2i32 case (using fp <-> int
   // instructions), but for smaller types, we need to first extend up to v2i32
   // before doing going farther.
   if (Op.getValueType() == MVT::v2i64) {
     EVT ExtVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::RETURNADDR:         return LowerRETURNADDR(Op, DAG);
   case ISD::FRAMEADDR:          return LowerFRAMEADDR(Op, DAG);
 
   case ISD::INTRINSIC_VOID:
     return LowerINTRINSIC_VOID(Op, DAG);
   case ISD::SREM:
   case ISD::UREM:
     return LowerREM(Op, DAG);
+  case ISD::BSWAP:
+    return LowerBSWAP(Op, DAG);
   }
 }
 
 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
                                            SmallVectorImpl<SDValue>&Results,
                                            SelectionDAG &DAG) const {
   SDLoc dl(N);
   switch (N->getOpcode()) {

test/CodeGen/PowerPC/bswap64.ll

+; RUN: llc -verify-machineinstrs < %s -mtriple=ppc64le-- -mcpu=pwr9 | FileCheck %s
+
+declare i64 @llvm.bswap.i64(i64)
+
+; CHECK: mtvsrdd
+; CHECK: xxbrd
+; CHECK: mfvsrld
+define i64 @bswap64(i64 %x) {
+entry:
+  %0 = call i64 @llvm.bswap.i64(i64 %x)
+  ret i64 %0
+}
+