[PowerPC] Skeletal FastISel support for 64-bit PowerPC ELF.

This is the first of many upcoming patches for PowerPC fast
instruction selection support.  This patch implements the minimum
necessary for a functional (but extremely limited) FastISel pass.  It
allows the table-generated portions of the selector to be created and
used, but in most cases selection will fall back to the DAG selector.
None of the block terminator instructions are implemented yet, and
most interesting instructions require some special handling.
Therefore there aren't any new test cases with this patch.  There will
be quite a few tests coming with future patches.

This patch adds the make/CMake support for the new code (including
tablegen -gen-fast-isel) and creates the FastISel object for PPC64 ELF
only.  It instantiates the necessary virtual functions
(TargetSelectInstruction, TargetMaterializeConstant,
TargetMaterializeAlloca, tryToFoldLoadIntoMI, and FastLowerArguments),
but of these, only TargetMaterializeConstant contains any useful
implementation.  This is present since the table-generated code
requires the ability to materialize integer constants for some
instructions.

This patch has been tested by building and running the
projects/test-suite code with -O0.  All tests passed with the
exception of a couple of long-running tests that time out using -O0
code generation.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187399 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/PowerPC/CMakeLists.txt

@@ -7,6 +7,7 @@ tablegen(LLVM PPCGenMCCodeEmitter.inc -gen-emitter -mc-emitter)
 tablegen(LLVM PPCGenRegisterInfo.inc -gen-register-info)
 tablegen(LLVM PPCGenInstrInfo.inc -gen-instr-info)
 tablegen(LLVM PPCGenDAGISel.inc -gen-dag-isel)
+tablegen(LLVM PPCGenFastISel.inc -gen-fast-isel)
 tablegen(LLVM PPCGenCallingConv.inc -gen-callingconv)
 tablegen(LLVM PPCGenSubtargetInfo.inc -gen-subtarget)
 add_public_tablegen_target(PowerPCCommonTableGen)
@@ -20,6 +21,7 @@ add_llvm_target(PowerPCCodeGen
   PPCInstrInfo.cpp
   PPCISelDAGToDAG.cpp
   PPCISelLowering.cpp
+  PPCFastISel.cpp
   PPCFrameLowering.cpp
   PPCJITInfo.cpp
   PPCMCInstLower.cpp

lib/Target/PowerPC/Makefile

@@ -16,7 +16,7 @@ BUILT_SOURCES = PPCGenRegisterInfo.inc PPCGenAsmMatcher.inc \
                 PPCGenAsmWriter.inc  PPCGenCodeEmitter.inc \
                 PPCGenInstrInfo.inc PPCGenDAGISel.inc \
                 PPCGenSubtargetInfo.inc PPCGenCallingConv.inc \
-                PPCGenMCCodeEmitter.inc
+                PPCGenMCCodeEmitter.inc PPCGenFastISel.inc
 
 DIRS = AsmParser InstPrinter TargetInfo MCTargetDesc
 

lib/Target/PowerPC/PPCFastISel.cpp

@@ -0,0 +1,328 @@
+//===-- PPCFastISel.cpp - PowerPC FastISel implementation -----------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the PowerPC-specific support for the FastISel class. Some
+// of the target-specific code is generated by tablegen in the file
+// PPCGenFastISel.inc, which is #included here.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "ppcfastisel"
+#include "PPC.h"
+#include "PPCISelLowering.h"
+#include "PPCSubtarget.h"
+#include "PPCTargetMachine.h"
+#include "MCTargetDesc/PPCPredicates.h"
+#include "llvm/ADT/Optional.h"
+#include "llvm/CodeGen/CallingConvLower.h"
+#include "llvm/CodeGen/FastISel.h"
+#include "llvm/CodeGen/FunctionLoweringInfo.h"
+#include "llvm/CodeGen/MachineConstantPool.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/CallingConv.h"
+#include "llvm/IR/GlobalAlias.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetMachine.h"
+
+using namespace llvm;
+
+namespace {
+
+typedef struct Address {
+  enum {
+    RegBase,
+    FrameIndexBase
+  } BaseType;
+
+  union {
+    unsigned Reg;
+    int FI;
+  } Base;
+
+  int Offset;
+
+  // Innocuous defaults for our address.
+  Address()
+   : BaseType(RegBase), Offset(0) {
+     Base.Reg = 0;
+   }
+} Address;
+
+class PPCFastISel : public FastISel {
+
+  const TargetMachine &TM;
+  const TargetInstrInfo &TII;
+  const TargetLowering &TLI;
+  const PPCSubtarget &PPCSubTarget;
+  LLVMContext *Context;
+
+  public:
+    explicit PPCFastISel(FunctionLoweringInfo &FuncInfo,
+                         const TargetLibraryInfo *LibInfo)
+    : FastISel(FuncInfo, LibInfo),
+      TM(FuncInfo.MF->getTarget()),
+      TII(*TM.getInstrInfo()),
+      TLI(*TM.getTargetLowering()),
+      PPCSubTarget(
+       *((static_cast<const PPCTargetMachine *>(&TM))->getSubtargetImpl())
+      ),
+      Context(&FuncInfo.Fn->getContext()) { }
+
+  // Backend specific FastISel code.
+  private:
+    virtual bool TargetSelectInstruction(const Instruction *I);
+    virtual unsigned TargetMaterializeConstant(const Constant *C);
+    virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
+    virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+                                     const LoadInst *LI);
+    virtual bool FastLowerArguments();
+
+  // Utility routines.
+  private:
+    unsigned PPCMaterializeFP(const ConstantFP *CFP, MVT VT);
+    unsigned PPCMaterializeInt(const Constant *C, MVT VT);
+    unsigned PPCMaterialize32BitInt(int64_t Imm,
+                                    const TargetRegisterClass *RC);
+    unsigned PPCMaterialize64BitInt(int64_t Imm,
+                                    const TargetRegisterClass *RC);
+
+  private:
+  #include "PPCGenFastISel.inc"
+
+};
+
+} // end anonymous namespace
+
+// Attempt to fast-select an instruction that wasn't handled by
+// the table-generated machinery.  TBD.
+bool PPCFastISel::TargetSelectInstruction(const Instruction *I) {
+  return I && false;
+}
+
+// Materialize a floating-point constant into a register, and return
+// the register number (or zero if we failed to handle it).
+unsigned PPCFastISel::PPCMaterializeFP(const ConstantFP *CFP, MVT VT) {
+  // No plans to handle long double here.
+  if (VT != MVT::f32 && VT != MVT::f64)
+    return 0;
+
+  // All FP constants are loaded from the constant pool.
+  unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
+  assert(Align > 0 && "Unexpectedly missing alignment information!");
+  unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
+  unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
+  CodeModel::Model CModel = TM.getCodeModel();
+
+  MachineMemOperand *MMO =
+    FuncInfo.MF->getMachineMemOperand(
+      MachinePointerInfo::getConstantPool(), MachineMemOperand::MOLoad,
+      (VT == MVT::f32) ? 4 : 8, Align);
+
+  // For small code model, generate a LDtocCPT.
+  if (CModel == CodeModel::Small || CModel == CodeModel::JITDefault)
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::LDtocCPT),
+            DestReg)
+      .addConstantPoolIndex(Idx).addReg(PPC::X2).addMemOperand(MMO);
+  else {
+    // Otherwise we generate LF[SD](Idx[lo], ADDIStocHA(X2, Idx)).
+    unsigned Opc = (VT == MVT::f32) ? PPC::LFS : PPC::LFD;
+    unsigned TmpReg = createResultReg(&PPC::G8RC_and_G8RC_NOX0RegClass);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::ADDIStocHA),
+            TmpReg).addReg(PPC::X2).addConstantPoolIndex(Idx);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
+      .addConstantPoolIndex(Idx, 0, PPCII::MO_TOC_LO)
+      .addReg(TmpReg)
+      .addMemOperand(MMO);
+  }
+
+  return DestReg;
+}
+
+// Materialize a 32-bit integer constant into a register, and return
+// the register number (or zero if we failed to handle it).
+unsigned PPCFastISel::PPCMaterialize32BitInt(int64_t Imm,
+                                             const TargetRegisterClass *RC) {
+  unsigned Lo = Imm & 0xFFFF;
+  unsigned Hi = (Imm >> 16) & 0xFFFF;
+
+  unsigned ResultReg = createResultReg(RC);
+  bool IsGPRC = RC->hasSuperClassEq(&PPC::GPRCRegClass);
+
+  if (isInt<16>(Imm))
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(IsGPRC ? PPC::LI : PPC::LI8), ResultReg)
+      .addImm(Imm);
+  else if (Lo) {
+    // Both Lo and Hi have nonzero bits.
+    unsigned TmpReg = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(IsGPRC ? PPC::LIS : PPC::LIS8), TmpReg)
+      .addImm(Hi);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(IsGPRC ? PPC::ORI : PPC::ORI8), ResultReg)
+      .addReg(TmpReg).addImm(Lo);
+  } else
+    // Just Hi bits.
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(IsGPRC ? PPC::LIS : PPC::LIS8), ResultReg)
+      .addImm(Hi);
+
+  return ResultReg;
+}
+
+// Materialize a 64-bit integer constant into a register, and return
+// the register number (or zero if we failed to handle it).
+unsigned PPCFastISel::PPCMaterialize64BitInt(int64_t Imm,
+                                             const TargetRegisterClass *RC) {
+  unsigned Remainder = 0;
+  unsigned Shift = 0;
+
+  // If the value doesn't fit in 32 bits, see if we can shift it
+  // so that it fits in 32 bits.
+  if (!isInt<32>(Imm)) {
+    Shift = countTrailingZeros<uint64_t>(Imm);
+    int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
+
+    if (isInt<32>(ImmSh))
+      Imm = ImmSh;
+    else {
+      Remainder = Imm;
+      Shift = 32;
+      Imm >>= 32;
+    }
+  }
+
+  // Handle the high-order 32 bits (if shifted) or the whole 32 bits
+  // (if not shifted).
+  unsigned TmpReg1 = PPCMaterialize32BitInt(Imm, RC);
+  if (!Shift)
+    return TmpReg1;
+
+  // If upper 32 bits were not zero, we've built them and need to shift
+  // them into place.
+  unsigned TmpReg2;
+  if (Imm) {
+    TmpReg2 = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::RLDICR),
+            TmpReg2).addReg(TmpReg1).addImm(Shift).addImm(63 - Shift);
+  } else
+    TmpReg2 = TmpReg1;
+
+  unsigned TmpReg3, Hi, Lo;
+  if ((Hi = (Remainder >> 16) & 0xFFFF)) {
+    TmpReg3 = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::ORIS8),
+            TmpReg3).addReg(TmpReg2).addImm(Hi);
+  } else
+    TmpReg3 = TmpReg2;
+
+  if ((Lo = Remainder & 0xFFFF)) {
+    unsigned ResultReg = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(PPC::ORI8),
+            ResultReg).addReg(TmpReg3).addImm(Lo);
+    return ResultReg;
+  }
+
+  return TmpReg3;
+}
+
+
+// Materialize an integer constant into a register, and return
+// the register number (or zero if we failed to handle it).
+unsigned PPCFastISel::PPCMaterializeInt(const Constant *C, MVT VT) {
+
+  if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 &&
+      VT != MVT::i8 && VT != MVT::i1) 
+    return 0;
+
+  const TargetRegisterClass *RC = ((VT == MVT::i64) ? &PPC::G8RCRegClass :
+                                   &PPC::GPRCRegClass);
+
+  // If the constant is in range, use a load-immediate.
+  const ConstantInt *CI = cast<ConstantInt>(C);
+  if (isInt<16>(CI->getSExtValue())) {
+    unsigned Opc = (VT == MVT::i64) ? PPC::LI8 : PPC::LI;
+    unsigned ImmReg = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ImmReg)
+      .addImm(CI->getSExtValue());
+    return ImmReg;
+  }
+
+  // Construct the constant piecewise.
+  int64_t Imm = CI->getZExtValue();
+
+  if (VT == MVT::i64)
+    return PPCMaterialize64BitInt(Imm, RC);
+  else if (VT == MVT::i32)
+    return PPCMaterialize32BitInt(Imm, RC);
+
+  return 0;
+}
+
+// Materialize a constant into a register, and return the register
+// number (or zero if we failed to handle it).
+unsigned PPCFastISel::TargetMaterializeConstant(const Constant *C) {
+  EVT CEVT = TLI.getValueType(C->getType(), true);
+
+  // Only handle simple types.
+  if (!CEVT.isSimple()) return 0;
+  MVT VT = CEVT.getSimpleVT();
+
+  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
+    return PPCMaterializeFP(CFP, VT);
+  else if (isa<ConstantInt>(C))
+    return PPCMaterializeInt(C, VT);
+  // TBD: Global values.
+
+  return 0;
+}
+
+// Materialize the address created by an alloca into a register, and
+// return the register number (or zero if we failed to handle it).  TBD.
+unsigned PPCFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
+  return AI && 0;
+}
+
+// Fold loads into extends when possible.  TBD.
+bool PPCFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+                                      const LoadInst *LI) {
+  return MI && OpNo && LI && false;
+}
+
+// Attempt to lower call arguments in a faster way than done by
+// the selection DAG code.
+bool PPCFastISel::FastLowerArguments() {
+  // Defer to normal argument lowering for now.  It's reasonably
+  // efficient.  Consider doing something like ARM to handle the
+  // case where all args fit in registers, no varargs, no float
+  // or vector args.
+  return false;
+}
+
+namespace llvm {
+  // Create the fast instruction selector for PowerPC64 ELF.
+  FastISel *PPC::createFastISel(FunctionLoweringInfo &FuncInfo,
+                                const TargetLibraryInfo *LibInfo) {
+    const TargetMachine &TM = FuncInfo.MF->getTarget();
+
+    // Only available on 64-bit ELF for now.
+    const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
+    if (Subtarget->isPPC64() && Subtarget->isSVR4ABI())
+      return new PPCFastISel(FuncInfo, LibInfo);
+
+    return 0;
+  }
+}

lib/Target/PowerPC/PPCISelLowering.cpp

@@ -7885,3 +7885,9 @@ Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
   return Sched::ILP;
 }
 
+// Create a fast isel object.
+FastISel *
+PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo,
+                                  const TargetLibraryInfo *LibInfo) const {
+  return PPC::createFastISel(FuncInfo, LibInfo);
+}

lib/Target/PowerPC/PPCISelLowering.h

@@ -465,6 +465,11 @@ namespace llvm {
     /// expanded to fmul + fadd.
     virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const;
 
+    /// createFastISel - This method returns a target-specific FastISel object,
+    /// or null if the target does not support "fast" instruction selection.
+    virtual FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
+                                     const TargetLibraryInfo *LibInfo) const;
+
   private:
     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
@@ -624,6 +629,11 @@ namespace llvm {
     SDValue DAGCombineFastRecipFSQRT(SDValue Op, DAGCombinerInfo &DCI) const;
   };
 
+  namespace PPC {
+    FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
+                             const TargetLibraryInfo *LibInfo);
+  }
+
   bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
                                   CCValAssign::LocInfo &LocInfo,
                                   ISD::ArgFlagsTy &ArgFlags,