1 //===-- AVRRegisterInfo.cpp - AVR Register Information --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the AVR implementation of the TargetRegisterInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "AVRRegisterInfo.h"
14 
15 #include "llvm/ADT/BitVector.h"
16 #include "llvm/CodeGen/MachineFrameInfo.h"
17 #include "llvm/CodeGen/MachineFunction.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/CodeGen/TargetFrameLowering.h"
21 #include "llvm/IR/Function.h"
22 
23 #include "AVR.h"
24 #include "AVRInstrInfo.h"
25 #include "AVRMachineFunctionInfo.h"
26 #include "AVRTargetMachine.h"
27 #include "MCTargetDesc/AVRMCTargetDesc.h"
28 
29 #define GET_REGINFO_TARGET_DESC
30 #include "AVRGenRegisterInfo.inc"
31 
32 namespace llvm {
33 
AVRRegisterInfo()34 AVRRegisterInfo::AVRRegisterInfo() : AVRGenRegisterInfo(0) {}
35 
36 const uint16_t *
getCalleeSavedRegs(const MachineFunction * MF) const37 AVRRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
38   const AVRMachineFunctionInfo *AFI = MF->getInfo<AVRMachineFunctionInfo>();
39   const AVRSubtarget &STI = MF->getSubtarget<AVRSubtarget>();
40   if (STI.hasTinyEncoding())
41     return AFI->isInterruptOrSignalHandler() ? CSR_InterruptsTiny_SaveList
42                                              : CSR_NormalTiny_SaveList;
43   else
44     return AFI->isInterruptOrSignalHandler() ? CSR_Interrupts_SaveList
45                                              : CSR_Normal_SaveList;
46 }
47 
48 const uint32_t *
getCallPreservedMask(const MachineFunction & MF,CallingConv::ID CC) const49 AVRRegisterInfo::getCallPreservedMask(const MachineFunction &MF,
50                                       CallingConv::ID CC) const {
51   const AVRSubtarget &STI = MF.getSubtarget<AVRSubtarget>();
52   return STI.hasTinyEncoding() ? CSR_NormalTiny_RegMask : CSR_Normal_RegMask;
53 }
54 
getReservedRegs(const MachineFunction & MF) const55 BitVector AVRRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
56   BitVector Reserved(getNumRegs());
57 
58   // Reserve the intermediate result registers r1 and r2
59   // The result of instructions like 'mul' is always stored here.
60   // R0/R1/R1R0 are always reserved on both avr and avrtiny.
61   Reserved.set(AVR::R0);
62   Reserved.set(AVR::R1);
63   Reserved.set(AVR::R1R0);
64 
65   // Reserve the stack pointer.
66   Reserved.set(AVR::SPL);
67   Reserved.set(AVR::SPH);
68   Reserved.set(AVR::SP);
69 
70   // Reserve R2~R17 only on avrtiny.
71   if (MF.getSubtarget<AVRSubtarget>().hasTinyEncoding()) {
72     // Reserve 8-bit registers R2~R15, Rtmp(R16) and Zero(R17).
73     for (unsigned Reg = AVR::R2; Reg <= AVR::R17; Reg++)
74       Reserved.set(Reg);
75     // Reserve 16-bit registers R3R2~R18R17.
76     for (unsigned Reg = AVR::R3R2; Reg <= AVR::R18R17; Reg++)
77       Reserved.set(Reg);
78   }
79 
80   // We tenatively reserve the frame pointer register r29:r28 because the
81   // function may require one, but we cannot tell until register allocation
82   // is complete, which can be too late.
83   //
84   // Instead we just unconditionally reserve the Y register.
85   //
86   // TODO: Write a pass to enumerate functions which reserved the Y register
87   //       but didn't end up needing a frame pointer. In these, we can
88   //       convert one or two of the spills inside to use the Y register.
89   Reserved.set(AVR::R28);
90   Reserved.set(AVR::R29);
91   Reserved.set(AVR::R29R28);
92 
93   return Reserved;
94 }
95 
96 const TargetRegisterClass *
getLargestLegalSuperClass(const TargetRegisterClass * RC,const MachineFunction & MF) const97 AVRRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
98                                            const MachineFunction &MF) const {
99   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
100   if (TRI->isTypeLegalForClass(*RC, MVT::i16)) {
101     return &AVR::DREGSRegClass;
102   }
103 
104   if (TRI->isTypeLegalForClass(*RC, MVT::i8)) {
105     return &AVR::GPR8RegClass;
106   }
107 
108   llvm_unreachable("Invalid register size");
109 }
110 
111 /// Fold a frame offset shared between two add instructions into a single one.
foldFrameOffset(MachineBasicBlock::iterator & II,int & Offset,Register DstReg)112 static void foldFrameOffset(MachineBasicBlock::iterator &II, int &Offset,
113                             Register DstReg) {
114   MachineInstr &MI = *II;
115   int Opcode = MI.getOpcode();
116 
117   // Don't bother trying if the next instruction is not an add or a sub.
118   if ((Opcode != AVR::SUBIWRdK) && (Opcode != AVR::ADIWRdK)) {
119     return;
120   }
121 
122   // Check that DstReg matches with next instruction, otherwise the instruction
123   // is not related to stack address manipulation.
124   if (DstReg != MI.getOperand(0).getReg()) {
125     return;
126   }
127 
128   // Add the offset in the next instruction to our offset.
129   switch (Opcode) {
130   case AVR::SUBIWRdK:
131     Offset += -MI.getOperand(2).getImm();
132     break;
133   case AVR::ADIWRdK:
134     Offset += MI.getOperand(2).getImm();
135     break;
136   }
137 
138   // Finally remove the instruction.
139   II++;
140   MI.eraseFromParent();
141 }
142 
eliminateFrameIndex(MachineBasicBlock::iterator II,int SPAdj,unsigned FIOperandNum,RegScavenger * RS) const143 bool AVRRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
144                                           int SPAdj, unsigned FIOperandNum,
145                                           RegScavenger *RS) const {
146   assert(SPAdj == 0 && "Unexpected SPAdj value");
147 
148   MachineInstr &MI = *II;
149   DebugLoc dl = MI.getDebugLoc();
150   MachineBasicBlock &MBB = *MI.getParent();
151   const MachineFunction &MF = *MBB.getParent();
152   const AVRTargetMachine &TM = (const AVRTargetMachine &)MF.getTarget();
153   const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
154   const MachineFrameInfo &MFI = MF.getFrameInfo();
155   const TargetFrameLowering *TFI = TM.getSubtargetImpl()->getFrameLowering();
156   const AVRSubtarget &STI = MF.getSubtarget<AVRSubtarget>();
157   int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
158   int Offset = MFI.getObjectOffset(FrameIndex);
159 
160   // Add one to the offset because SP points to an empty slot.
161   Offset += MFI.getStackSize() - TFI->getOffsetOfLocalArea() + 1;
162   // Fold incoming offset.
163   Offset += MI.getOperand(FIOperandNum + 1).getImm();
164 
165   // This is actually "load effective address" of the stack slot
166   // instruction. We have only two-address instructions, thus we need to
167   // expand it into move + add.
168   if (MI.getOpcode() == AVR::FRMIDX) {
169     Register DstReg = MI.getOperand(0).getReg();
170     assert(DstReg != AVR::R29R28 && "Dest reg cannot be the frame pointer");
171 
172     // Copy the frame pointer.
173     if (STI.hasMOVW()) {
174       BuildMI(MBB, MI, dl, TII.get(AVR::MOVWRdRr), DstReg)
175           .addReg(AVR::R29R28);
176     } else {
177       Register DstLoReg, DstHiReg;
178       splitReg(DstReg, DstLoReg, DstHiReg);
179       BuildMI(MBB, MI, dl, TII.get(AVR::MOVRdRr), DstLoReg)
180           .addReg(AVR::R28);
181       BuildMI(MBB, MI, dl, TII.get(AVR::MOVRdRr), DstHiReg)
182           .addReg(AVR::R29);
183     }
184 
185     assert(Offset > 0 && "Invalid offset");
186 
187     // We need to materialize the offset via an add instruction.
188     unsigned Opcode;
189 
190     II++; // Skip over the FRMIDX instruction.
191 
192     // Generally, to load a frame address two add instructions are emitted that
193     // could get folded into a single one:
194     //  movw    r31:r30, r29:r28
195     //  adiw    r31:r30, 29
196     //  adiw    r31:r30, 16
197     // to:
198     //  movw    r31:r30, r29:r28
199     //  adiw    r31:r30, 45
200     if (II != MBB.end())
201       foldFrameOffset(II, Offset, DstReg);
202 
203     // Select the best opcode based on DstReg and the offset size.
204     switch (DstReg) {
205     case AVR::R25R24:
206     case AVR::R27R26:
207     case AVR::R31R30: {
208       if (isUInt<6>(Offset) && STI.hasADDSUBIW()) {
209         Opcode = AVR::ADIWRdK;
210         break;
211       }
212       [[fallthrough]];
213     }
214     default: {
215       // This opcode will get expanded into a pair of subi/sbci.
216       Opcode = AVR::SUBIWRdK;
217       Offset = -Offset;
218       break;
219     }
220     }
221 
222     MachineInstr *New = BuildMI(MBB, II, dl, TII.get(Opcode), DstReg)
223                             .addReg(DstReg, RegState::Kill)
224                             .addImm(Offset);
225     New->getOperand(3).setIsDead();
226 
227     MI.eraseFromParent(); // remove FRMIDX
228 
229     return false;
230   }
231 
232   // On most AVRs, we can use an offset up to 62 for load/store with
233   // displacement (63 for byte values, 62 for word values). However, the
234   // "reduced tiny" cores don't support load/store with displacement. So for
235   // them, we force an offset of 0 meaning that any positive offset will require
236   // adjusting the frame pointer.
237   int MaxOffset = STI.hasTinyEncoding() ? 0 : 62;
238 
239   // If the offset is too big we have to adjust and restore the frame pointer
240   // to materialize a valid load/store with displacement.
241   //: TODO: consider using only one adiw/sbiw chain for more than one frame
242   //: index
243   if (Offset > MaxOffset) {
244     unsigned AddOpc = AVR::ADIWRdK, SubOpc = AVR::SBIWRdK;
245     int AddOffset = Offset - MaxOffset;
246 
247     // For huge offsets where adiw/sbiw cannot be used use a pair of subi/sbci.
248     if ((Offset - MaxOffset) > 63 || !STI.hasADDSUBIW()) {
249       AddOpc = AVR::SUBIWRdK;
250       SubOpc = AVR::SUBIWRdK;
251       AddOffset = -AddOffset;
252     }
253 
254     // It is possible that the spiller places this frame instruction in between
255     // a compare and branch, invalidating the contents of SREG set by the
256     // compare instruction because of the add/sub pairs. Conservatively save and
257     // restore SREG before and after each add/sub pair.
258     BuildMI(MBB, II, dl, TII.get(AVR::INRdA), STI.getTmpRegister())
259         .addImm(STI.getIORegSREG());
260 
261     MachineInstr *New = BuildMI(MBB, II, dl, TII.get(AddOpc), AVR::R29R28)
262                             .addReg(AVR::R29R28, RegState::Kill)
263                             .addImm(AddOffset);
264     New->getOperand(3).setIsDead();
265 
266     // Restore SREG.
267     BuildMI(MBB, std::next(II), dl, TII.get(AVR::OUTARr))
268         .addImm(STI.getIORegSREG())
269         .addReg(STI.getTmpRegister(), RegState::Kill);
270 
271     // No need to set SREG as dead here otherwise if the next instruction is a
272     // cond branch it will be using a dead register.
273     BuildMI(MBB, std::next(II), dl, TII.get(SubOpc), AVR::R29R28)
274         .addReg(AVR::R29R28, RegState::Kill)
275         .addImm(Offset - MaxOffset);
276 
277     Offset = MaxOffset;
278   }
279 
280   MI.getOperand(FIOperandNum).ChangeToRegister(AVR::R29R28, false);
281   assert(isUInt<6>(Offset) && "Offset is out of range");
282   MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Offset);
283   return false;
284 }
285 
getFrameRegister(const MachineFunction & MF) const286 Register AVRRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
287   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
288   if (TFI->hasFP(MF)) {
289     // The Y pointer register
290     return AVR::R28;
291   }
292 
293   return AVR::SP;
294 }
295 
296 const TargetRegisterClass *
getPointerRegClass(const MachineFunction & MF,unsigned Kind) const297 AVRRegisterInfo::getPointerRegClass(const MachineFunction &MF,
298                                     unsigned Kind) const {
299   // FIXME: Currently we're using avr-gcc as reference, so we restrict
300   // ptrs to Y and Z regs. Though avr-gcc has buggy implementation
301   // of memory constraint, so we can fix it and bit avr-gcc here ;-)
302   return &AVR::PTRDISPREGSRegClass;
303 }
304 
splitReg(Register Reg,Register & LoReg,Register & HiReg) const305 void AVRRegisterInfo::splitReg(Register Reg, Register &LoReg,
306                                Register &HiReg) const {
307   assert(AVR::DREGSRegClass.contains(Reg) && "can only split 16-bit registers");
308 
309   LoReg = getSubReg(Reg, AVR::sub_lo);
310   HiReg = getSubReg(Reg, AVR::sub_hi);
311 }
312 
shouldCoalesce(MachineInstr * MI,const TargetRegisterClass * SrcRC,unsigned SubReg,const TargetRegisterClass * DstRC,unsigned DstSubReg,const TargetRegisterClass * NewRC,LiveIntervals & LIS) const313 bool AVRRegisterInfo::shouldCoalesce(
314     MachineInstr *MI, const TargetRegisterClass *SrcRC, unsigned SubReg,
315     const TargetRegisterClass *DstRC, unsigned DstSubReg,
316     const TargetRegisterClass *NewRC, LiveIntervals &LIS) const {
317   if (this->getRegClass(AVR::PTRDISPREGSRegClassID)->hasSubClassEq(NewRC)) {
318     return false;
319   }
320 
321   return TargetRegisterInfo::shouldCoalesce(MI, SrcRC, SubReg, DstRC, DstSubReg,
322                                             NewRC, LIS);
323 }
324 
325 } // end of namespace llvm
326