1//===-- HexagonFrameLowering.cpp - Define frame lowering ------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//
9//===----------------------------------------------------------------------===//
10
11#include "HexagonFrameLowering.h"
12#include "Hexagon.h"
13#include "HexagonInstrInfo.h"
14#include "HexagonRegisterInfo.h"
15#include "HexagonSubtarget.h"
16#include "HexagonTargetMachine.h"
17#include "HexagonMachineFunctionInfo.h"
18#include "llvm/Function.h"
19#include "llvm/Type.h"
20#include "llvm/ADT/BitVector.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/CodeGen/AsmPrinter.h"
23#include "llvm/CodeGen/MachineInstrBuilder.h"
24#include "llvm/CodeGen/MachineFunction.h"
25#include "llvm/CodeGen/MachineFunctionPass.h"
26#include "llvm/CodeGen/MachineFrameInfo.h"
27#include "llvm/CodeGen/MachineModuleInfo.h"
28#include "llvm/CodeGen/MachineRegisterInfo.h"
29#include "llvm/CodeGen/RegisterScavenging.h"
30#include "llvm/MC/MachineLocation.h"
31#include "llvm/MC/MCAsmInfo.h"
32#include "llvm/Target/TargetInstrInfo.h"
33#include "llvm/Target/TargetMachine.h"
34#include "llvm/Target/TargetOptions.h"
35#include "llvm/Support/CommandLine.h"
36
37using namespace llvm;
38
39static cl::opt<bool> DisableDeallocRet(
40                       "disable-hexagon-dealloc-ret",
41                       cl::Hidden,
42                       cl::desc("Disable Dealloc Return for Hexagon target"));
43
44/// determineFrameLayout - Determine the size of the frame and maximum call
45/// frame size.
46void HexagonFrameLowering::determineFrameLayout(MachineFunction &MF) const {
47  MachineFrameInfo *MFI = MF.getFrameInfo();
48
49  // Get the number of bytes to allocate from the FrameInfo.
50  unsigned FrameSize = MFI->getStackSize();
51
52  // Get the alignments provided by the target.
53  unsigned TargetAlign = MF.getTarget().getFrameLowering()->getStackAlignment();
54  // Get the maximum call frame size of all the calls.
55  unsigned maxCallFrameSize = MFI->getMaxCallFrameSize();
56
57  // If we have dynamic alloca then maxCallFrameSize needs to be aligned so
58  // that allocations will be aligned.
59  if (MFI->hasVarSizedObjects())
60    maxCallFrameSize = RoundUpToAlignment(maxCallFrameSize, TargetAlign);
61
62  // Update maximum call frame size.
63  MFI->setMaxCallFrameSize(maxCallFrameSize);
64
65  // Include call frame size in total.
66  FrameSize += maxCallFrameSize;
67
68  // Make sure the frame is aligned.
69  FrameSize = RoundUpToAlignment(FrameSize, TargetAlign);
70
71  // Update frame info.
72  MFI->setStackSize(FrameSize);
73}
74
75
76void HexagonFrameLowering::emitPrologue(MachineFunction &MF) const {
77  MachineBasicBlock &MBB = MF.front();
78  MachineFrameInfo *MFI = MF.getFrameInfo();
79  MachineModuleInfo &MMI = MF.getMMI();
80  MachineBasicBlock::iterator MBBI = MBB.begin();
81  const HexagonRegisterInfo *QRI =
82    static_cast<const HexagonRegisterInfo *>(MF.getTarget().getRegisterInfo());
83  DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
84  determineFrameLayout(MF);
85
86  // Check if frame moves are needed for EH.
87  bool needsFrameMoves = MMI.hasDebugInfo() ||
88    !MF.getFunction()->needsUnwindTableEntry();
89
90  // Get the number of bytes to allocate from the FrameInfo.
91  int NumBytes = (int) MFI->getStackSize();
92
93  // LLVM expects allocframe not to be the first instruction in the
94  // basic block.
95  MachineBasicBlock::iterator InsertPt = MBB.begin();
96
97  //
98  // ALLOCA adjust regs.  Iterate over ADJDYNALLOC nodes and change the offset.
99  //
100  HexagonMachineFunctionInfo *FuncInfo =
101    MF.getInfo<HexagonMachineFunctionInfo>();
102  const std::vector<MachineInstr*>& AdjustRegs =
103    FuncInfo->getAllocaAdjustInsts();
104  for (std::vector<MachineInstr*>::const_iterator i = AdjustRegs.begin(),
105         e = AdjustRegs.end();
106       i != e; ++i) {
107    MachineInstr* MI = *i;
108    assert((MI->getOpcode() == Hexagon::ADJDYNALLOC) &&
109           "Expected adjust alloca node");
110
111    MachineOperand& MO = MI->getOperand(2);
112    assert(MO.isImm() && "Expected immediate");
113    MO.setImm(MFI->getMaxCallFrameSize());
114  }
115
116 std::vector<MachineMove> &Moves = MMI.getFrameMoves();
117
118 if (needsFrameMoves) {
119   // Advance CFA. DW_CFA_def_cfa
120   unsigned FPReg = QRI->getFrameRegister();
121   unsigned RAReg = QRI->getRARegister();
122
123   MachineLocation Dst(MachineLocation::VirtualFP);
124   MachineLocation Src(FPReg, -8);
125   Moves.push_back(MachineMove(0, Dst, Src));
126
127   // R31 = (R31 - #4)
128   MachineLocation LRDst(RAReg, -4);
129   MachineLocation LRSrc(RAReg);
130   Moves.push_back(MachineMove(0, LRDst, LRSrc));
131
132   // R30 = (R30 - #8)
133   MachineLocation SPDst(FPReg, -8);
134   MachineLocation SPSrc(FPReg);
135   Moves.push_back(MachineMove(0, SPDst, SPSrc));
136 }
137
138  //
139  // Only insert ALLOCFRAME if we need to.
140  //
141  if (hasFP(MF)) {
142    // Check for overflow.
143    // Hexagon_TODO: Ugh! hardcoding. Is there an API that can be used?
144    const int ALLOCFRAME_MAX = 16384;
145    const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
146
147    if (NumBytes >= ALLOCFRAME_MAX) {
148      // Emit allocframe(#0).
149      BuildMI(MBB, InsertPt, dl, TII.get(Hexagon::ALLOCFRAME)).addImm(0);
150
151      // Subtract offset from frame pointer.
152      BuildMI(MBB, InsertPt, dl, TII.get(Hexagon::CONST32_Int_Real),
153                                      HEXAGON_RESERVED_REG_1).addImm(NumBytes);
154      BuildMI(MBB, InsertPt, dl, TII.get(Hexagon::SUB_rr),
155                                      QRI->getStackRegister()).
156                                      addReg(QRI->getStackRegister()).
157                                      addReg(HEXAGON_RESERVED_REG_1);
158    } else {
159      BuildMI(MBB, InsertPt, dl, TII.get(Hexagon::ALLOCFRAME)).addImm(NumBytes);
160    }
161  }
162}
163// Returns true if MBB has a machine instructions that indicates a tail call
164// in the block.
165bool HexagonFrameLowering::hasTailCall(MachineBasicBlock &MBB) const {
166  MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
167  unsigned RetOpcode = MBBI->getOpcode();
168
169  return RetOpcode == Hexagon::TCRETURNtg || RetOpcode == Hexagon::TCRETURNtext;}
170
171void HexagonFrameLowering::emitEpilogue(MachineFunction &MF,
172                                     MachineBasicBlock &MBB) const {
173  MachineBasicBlock::iterator MBBI = prior(MBB.end());
174  DebugLoc dl = MBBI->getDebugLoc();
175  //
176  // Only insert deallocframe if we need to.
177  //
178  if (hasFP(MF)) {
179    MachineBasicBlock::iterator MBBI = prior(MBB.end());
180    MachineBasicBlock::iterator MBBI_end = MBB.end();
181    //
182    // For Hexagon, we don't need the frame size.
183    //
184    MachineFrameInfo *MFI = MF.getFrameInfo();
185    int NumBytes = (int) MFI->getStackSize();
186
187    const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
188
189    // Replace 'jumpr r31' instruction with dealloc_return for V4 and higher
190    // versions.
191    if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPR
192                        && !DisableDeallocRet) {
193      // Remove jumpr node.
194      MBB.erase(MBBI);
195      // Add dealloc_return.
196      BuildMI(MBB, MBBI_end, dl, TII.get(Hexagon::DEALLOC_RET_V4))
197        .addImm(NumBytes);
198    } else { // Add deallocframe for V2 and V3.
199      BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)).addImm(NumBytes);
200    }
201  }
202}
203
204bool HexagonFrameLowering::hasFP(const MachineFunction &MF) const {
205  const MachineFrameInfo *MFI = MF.getFrameInfo();
206  const HexagonMachineFunctionInfo *FuncInfo =
207    MF.getInfo<HexagonMachineFunctionInfo>();
208  return (MFI->hasCalls() || (MFI->getStackSize() > 0) ||
209          FuncInfo->hasClobberLR() );
210}
211
212bool
213HexagonFrameLowering::spillCalleeSavedRegisters(
214                                        MachineBasicBlock &MBB,
215                                        MachineBasicBlock::iterator MI,
216                                        const std::vector<CalleeSavedInfo> &CSI,
217                                        const TargetRegisterInfo *TRI) const {
218  MachineFunction *MF = MBB.getParent();
219  const TargetInstrInfo &TII = *MF->getTarget().getInstrInfo();
220
221  if (CSI.empty()) {
222    return false;
223  }
224
225  // We can only schedule double loads if we spill contiguous callee-saved regs
226  // For instance, we cannot scheduled double-word loads if we spill r24,
227  // r26, and r27.
228  // Hexagon_TODO: We can try to double-word align odd registers for -O2 and
229  // above.
230  bool ContiguousRegs = true;
231
232  for (unsigned i = 0; i < CSI.size(); ++i) {
233    unsigned Reg = CSI[i].getReg();
234
235    //
236    // Check if we can use a double-word store.
237    //
238    const uint16_t* SuperReg = TRI->getSuperRegisters(Reg);
239
240    // Assume that there is exactly one superreg.
241    assert(SuperReg[0] && !SuperReg[1] && "Expected exactly one superreg");
242    bool CanUseDblStore = false;
243    const TargetRegisterClass* SuperRegClass = 0;
244
245    if (ContiguousRegs && (i < CSI.size()-1)) {
246      const uint16_t* SuperRegNext = TRI->getSuperRegisters(CSI[i+1].getReg());
247      assert(SuperRegNext[0] && !SuperRegNext[1] &&
248             "Expected exactly one superreg");
249      SuperRegClass = TRI->getMinimalPhysRegClass(SuperReg[0]);
250      CanUseDblStore = (SuperRegNext[0] == SuperReg[0]);
251    }
252
253
254    if (CanUseDblStore) {
255      TII.storeRegToStackSlot(MBB, MI, SuperReg[0], true,
256                              CSI[i+1].getFrameIdx(), SuperRegClass, TRI);
257      MBB.addLiveIn(SuperReg[0]);
258      ++i;
259    } else {
260      // Cannot use a double-word store.
261      ContiguousRegs = false;
262      const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
263      TII.storeRegToStackSlot(MBB, MI, Reg, true, CSI[i].getFrameIdx(), RC,
264                              TRI);
265      MBB.addLiveIn(Reg);
266    }
267  }
268  return true;
269}
270
271
272bool HexagonFrameLowering::restoreCalleeSavedRegisters(
273                                        MachineBasicBlock &MBB,
274                                        MachineBasicBlock::iterator MI,
275                                        const std::vector<CalleeSavedInfo> &CSI,
276                                        const TargetRegisterInfo *TRI) const {
277
278  MachineFunction *MF = MBB.getParent();
279  const TargetInstrInfo &TII = *MF->getTarget().getInstrInfo();
280
281  if (CSI.empty()) {
282    return false;
283  }
284
285  // We can only schedule double loads if we spill contiguous callee-saved regs
286  // For instance, we cannot scheduled double-word loads if we spill r24,
287  // r26, and r27.
288  // Hexagon_TODO: We can try to double-word align odd registers for -O2 and
289  // above.
290  bool ContiguousRegs = true;
291
292  for (unsigned i = 0; i < CSI.size(); ++i) {
293    unsigned Reg = CSI[i].getReg();
294
295    //
296    // Check if we can use a double-word load.
297    //
298    const uint16_t* SuperReg = TRI->getSuperRegisters(Reg);
299    const TargetRegisterClass* SuperRegClass = 0;
300
301    // Assume that there is exactly one superreg.
302    assert(SuperReg[0] && !SuperReg[1] && "Expected exactly one superreg");
303    bool CanUseDblLoad = false;
304    if (ContiguousRegs && (i < CSI.size()-1)) {
305      const uint16_t* SuperRegNext = TRI->getSuperRegisters(CSI[i+1].getReg());
306      assert(SuperRegNext[0] && !SuperRegNext[1] &&
307             "Expected exactly one superreg");
308      SuperRegClass = TRI->getMinimalPhysRegClass(SuperReg[0]);
309      CanUseDblLoad = (SuperRegNext[0] == SuperReg[0]);
310    }
311
312
313    if (CanUseDblLoad) {
314      TII.loadRegFromStackSlot(MBB, MI, SuperReg[0], CSI[i+1].getFrameIdx(),
315                               SuperRegClass, TRI);
316      MBB.addLiveIn(SuperReg[0]);
317      ++i;
318    } else {
319      // Cannot use a double-word load.
320      ContiguousRegs = false;
321      const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
322      TII.loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(), RC, TRI);
323      MBB.addLiveIn(Reg);
324    }
325  }
326  return true;
327}
328
329int HexagonFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
330                                              int FI) const {
331  return MF.getFrameInfo()->getObjectOffset(FI);
332}
333