xref: /llvm-project/llvm/lib/Target/Hexagon/HexagonRegisterInfo.cpp (revision 6f3f08abdc9faac1fe07018bf72d532443f2ec05)
1 //===-- HexagonRegisterInfo.cpp - Hexagon 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 Hexagon implementation of the TargetRegisterInfo
10 // class.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "HexagonRegisterInfo.h"
15 #include "HexagonMachineFunctionInfo.h"
16 #include "HexagonSubtarget.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/CodeGen/LiveIntervals.h"
21 #include "llvm/CodeGen/LiveRegUnits.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/CodeGen/RegisterScavenging.h"
28 #include "llvm/CodeGen/TargetInstrInfo.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Type.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
36 
37 #define GET_REGINFO_TARGET_DESC
38 #include "HexagonGenRegisterInfo.inc"
39 
40 using namespace llvm;
41 
42 static cl::opt<unsigned> FrameIndexSearchRange(
43     "hexagon-frame-index-search-range", cl::init(32), cl::Hidden,
44     cl::desc("Limit on instruction search range in frame index elimination"));
45 
46 static cl::opt<unsigned> FrameIndexReuseLimit(
47     "hexagon-frame-index-reuse-limit", cl::init(~0), cl::Hidden,
48     cl::desc("Limit on the number of reused registers in frame index "
49     "elimination"));
50 
51 HexagonRegisterInfo::HexagonRegisterInfo(unsigned HwMode)
52     : HexagonGenRegisterInfo(Hexagon::R31, 0/*DwarfFlavor*/, 0/*EHFlavor*/,
53                              0/*PC*/, HwMode) {}
54 
55 
56 bool HexagonRegisterInfo::isEHReturnCalleeSaveReg(Register R) const {
57   return R == Hexagon::R0 || R == Hexagon::R1 || R == Hexagon::R2 ||
58          R == Hexagon::R3 || R == Hexagon::D0 || R == Hexagon::D1;
59 }
60 
61 const MCPhysReg *
62 HexagonRegisterInfo::getCallerSavedRegs(const MachineFunction *MF,
63       const TargetRegisterClass *RC) const {
64   using namespace Hexagon;
65 
66   static const MCPhysReg Int32[] = {
67     R0, R1, R2, R3, R4, R5, R6, R7, R8, R9, R10, R11, R12, R13, R14, R15, 0
68   };
69   static const MCPhysReg Int64[] = {
70     D0, D1, D2, D3, D4, D5, D6, D7, 0
71   };
72   static const MCPhysReg Pred[] = {
73     P0, P1, P2, P3, 0
74   };
75   static const MCPhysReg VecSgl[] = {
76      V0,  V1,  V2,  V3,  V4,  V5,  V6,  V7,  V8,  V9, V10, V11, V12, V13,
77     V14, V15, V16, V17, V18, V19, V20, V21, V22, V23, V24, V25, V26, V27,
78     V28, V29, V30, V31,   0
79   };
80   static const MCPhysReg VecDbl[] = {
81     W0, W1, W2, W3, W4, W5, W6, W7, W8, W9, W10, W11, W12, W13, W14, W15, 0
82   };
83   static const MCPhysReg VecPred[] = {
84     Q0, Q1, Q2, Q3, 0
85   };
86 
87   switch (RC->getID()) {
88     case IntRegsRegClassID:
89       return Int32;
90     case DoubleRegsRegClassID:
91       return Int64;
92     case PredRegsRegClassID:
93       return Pred;
94     case HvxVRRegClassID:
95       return VecSgl;
96     case HvxWRRegClassID:
97       return VecDbl;
98     case HvxQRRegClassID:
99       return VecPred;
100     default:
101       break;
102   }
103 
104   static const MCPhysReg Empty[] = { 0 };
105 #ifndef NDEBUG
106   dbgs() << "Register class: " << getRegClassName(RC) << "\n";
107 #endif
108   llvm_unreachable("Unexpected register class");
109   return Empty;
110 }
111 
112 
113 const MCPhysReg *
114 HexagonRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
115   static const MCPhysReg CalleeSavedRegsV3[] = {
116     Hexagon::R16,   Hexagon::R17,   Hexagon::R18,   Hexagon::R19,
117     Hexagon::R20,   Hexagon::R21,   Hexagon::R22,   Hexagon::R23,
118     Hexagon::R24,   Hexagon::R25,   Hexagon::R26,   Hexagon::R27, 0
119   };
120 
121   // Functions that contain a call to __builtin_eh_return also save the first 4
122   // parameter registers.
123   static const MCPhysReg CalleeSavedRegsV3EHReturn[] = {
124     Hexagon::R0,    Hexagon::R1,    Hexagon::R2,    Hexagon::R3,
125     Hexagon::R16,   Hexagon::R17,   Hexagon::R18,   Hexagon::R19,
126     Hexagon::R20,   Hexagon::R21,   Hexagon::R22,   Hexagon::R23,
127     Hexagon::R24,   Hexagon::R25,   Hexagon::R26,   Hexagon::R27, 0
128   };
129 
130   bool HasEHReturn = MF->getInfo<HexagonMachineFunctionInfo>()->hasEHReturn();
131 
132   return HasEHReturn ? CalleeSavedRegsV3EHReturn : CalleeSavedRegsV3;
133 }
134 
135 
136 const uint32_t *HexagonRegisterInfo::getCallPreservedMask(
137       const MachineFunction &MF, CallingConv::ID) const {
138   return HexagonCSR_RegMask;
139 }
140 
141 
142 BitVector HexagonRegisterInfo::getReservedRegs(const MachineFunction &MF)
143       const {
144   BitVector Reserved(getNumRegs());
145   Reserved.set(Hexagon::R29);
146   Reserved.set(Hexagon::R30);
147   Reserved.set(Hexagon::R31);
148   Reserved.set(Hexagon::VTMP);
149 
150   // Guest registers.
151   Reserved.set(Hexagon::GELR);        // G0
152   Reserved.set(Hexagon::GSR);         // G1
153   Reserved.set(Hexagon::GOSP);        // G2
154   Reserved.set(Hexagon::G3);          // G3
155 
156   // Control registers.
157   Reserved.set(Hexagon::SA0);         // C0
158   Reserved.set(Hexagon::LC0);         // C1
159   Reserved.set(Hexagon::SA1);         // C2
160   Reserved.set(Hexagon::LC1);         // C3
161   Reserved.set(Hexagon::P3_0);        // C4
162   Reserved.set(Hexagon::USR);         // C8
163   Reserved.set(Hexagon::PC);          // C9
164   Reserved.set(Hexagon::UGP);         // C10
165   Reserved.set(Hexagon::GP);          // C11
166   Reserved.set(Hexagon::CS0);         // C12
167   Reserved.set(Hexagon::CS1);         // C13
168   Reserved.set(Hexagon::UPCYCLELO);   // C14
169   Reserved.set(Hexagon::UPCYCLEHI);   // C15
170   Reserved.set(Hexagon::FRAMELIMIT);  // C16
171   Reserved.set(Hexagon::FRAMEKEY);    // C17
172   Reserved.set(Hexagon::PKTCOUNTLO);  // C18
173   Reserved.set(Hexagon::PKTCOUNTHI);  // C19
174   Reserved.set(Hexagon::UTIMERLO);    // C30
175   Reserved.set(Hexagon::UTIMERHI);    // C31
176   // Out of the control registers, only C8 is explicitly defined in
177   // HexagonRegisterInfo.td. If others are defined, make sure to add
178   // them here as well.
179   Reserved.set(Hexagon::C8);
180   Reserved.set(Hexagon::USR_OVF);
181 
182   // Leveraging these registers will require more work to recognize
183   // the new semantics posed, Hi/LoVec patterns, etc.
184   // Note well: if enabled, they should be restricted to only
185   // where `HST.useHVXOps() && HST.hasV67Ops()` is true.
186   for (auto Reg : Hexagon_MC::GetVectRegRev())
187     Reserved.set(Reg);
188 
189   if (MF.getSubtarget<HexagonSubtarget>().hasReservedR19())
190     Reserved.set(Hexagon::R19);
191 
192   Register AP =
193       MF.getInfo<HexagonMachineFunctionInfo>()->getStackAlignBaseReg();
194   if (AP.isValid())
195     Reserved.set(AP);
196 
197   for (int x = Reserved.find_first(); x >= 0; x = Reserved.find_next(x))
198     markSuperRegs(Reserved, x);
199 
200   return Reserved;
201 }
202 
203 bool HexagonRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
204                                               int SPAdj, unsigned FIOp,
205                                               RegScavenger *RS) const {
206   static unsigned ReuseCount = 0;
207   //
208   // Hexagon_TODO: Do we need to enforce this for Hexagon?
209   assert(SPAdj == 0 && "Unexpected");
210 
211   MachineInstr &MI = *II;
212   MachineBasicBlock &MB = *MI.getParent();
213   MachineFunction &MF = *MB.getParent();
214   auto &HST = MF.getSubtarget<HexagonSubtarget>();
215   auto &HII = *HST.getInstrInfo();
216   auto &HFI = *HST.getFrameLowering();
217 
218   Register BP;
219   int FI = MI.getOperand(FIOp).getIndex();
220   // Select the base pointer (BP) and calculate the actual offset from BP
221   // to the beginning of the object at index FI.
222   int Offset = HFI.getFrameIndexReference(MF, FI, BP).getFixed();
223   // Add the offset from the instruction.
224   int RealOffset = Offset + MI.getOperand(FIOp+1).getImm();
225 
226   unsigned Opc = MI.getOpcode();
227   switch (Opc) {
228     case Hexagon::PS_fia:
229       MI.setDesc(HII.get(Hexagon::A2_addi));
230       MI.getOperand(FIOp).ChangeToImmediate(RealOffset);
231       MI.removeOperand(FIOp+1);
232       return false;
233     case Hexagon::PS_fi:
234       // Set up the instruction for updating below.
235       MI.setDesc(HII.get(Hexagon::A2_addi));
236       break;
237   }
238 
239   if (!HII.isValidOffset(Opc, RealOffset, this)) {
240     // If the offset is not valid, calculate the address in a temporary
241     // register and use it with offset 0.
242     int InstOffset = 0;
243     // The actual base register (BP) is typically shared between many
244     // instructions where frame indices are being replaced. In scalar
245     // instructions the offset range is large, and the need for an extra
246     // add instruction is infrequent. Vector loads/stores, however, have
247     // a much smaller offset range: [-8, 7), or #s4. In those cases it
248     // makes sense to "standardize" the immediate in the "addi" instruction
249     // so that multiple loads/stores could be based on it.
250     bool IsPair = false;
251     switch (MI.getOpcode()) {
252       // All of these instructions have the same format: base+#s4.
253       case Hexagon::PS_vloadrw_ai:
254       case Hexagon::PS_vloadrw_nt_ai:
255       case Hexagon::PS_vstorerw_ai:
256       case Hexagon::PS_vstorerw_nt_ai:
257         IsPair = true;
258         [[fallthrough]];
259       case Hexagon::PS_vloadrv_ai:
260       case Hexagon::PS_vloadrv_nt_ai:
261       case Hexagon::PS_vstorerv_ai:
262       case Hexagon::PS_vstorerv_nt_ai:
263       case Hexagon::V6_vL32b_ai:
264       case Hexagon::V6_vS32b_ai: {
265         unsigned HwLen = HST.getVectorLength();
266         if (RealOffset % HwLen == 0) {
267           int VecOffset = RealOffset / HwLen;
268           // Rewrite the offset as "base + [-8, 7)".
269           VecOffset += 8;
270           // Pairs are expanded into two instructions: make sure that both
271           // can use the same base (i.e. VecOffset+1 is not a different
272           // multiple of 16 than VecOffset).
273           if (!IsPair || (VecOffset + 1) % 16 != 0) {
274             RealOffset = (VecOffset & -16) * HwLen;
275             InstOffset = (VecOffset % 16 - 8) * HwLen;
276           }
277         }
278       }
279     }
280 
281     // Search backwards in the block for "Reg = A2_addi BP, RealOffset".
282     // This will give us a chance to avoid creating a new register.
283     Register ReuseBP;
284 
285     if (ReuseCount < FrameIndexReuseLimit) {
286       unsigned SearchCount = 0, SearchRange = FrameIndexSearchRange;
287       SmallSet<Register,2> SeenVRegs;
288       bool PassedCall = false;
289       LiveRegUnits Defs(*this), Uses(*this);
290 
291       for (auto I = std::next(II.getReverse()), E = MB.rend(); I != E; ++I) {
292         if (SearchCount == SearchRange)
293           break;
294         ++SearchCount;
295         const MachineInstr &BI = *I;
296         LiveRegUnits::accumulateUsedDefed(BI, Defs, Uses, this);
297         PassedCall |= BI.isCall();
298         for (const MachineOperand &Op : BI.operands()) {
299           if (SeenVRegs.size() > 1)
300             break;
301           if (Op.isReg() && Op.getReg().isVirtual())
302             SeenVRegs.insert(Op.getReg());
303         }
304         if (BI.getOpcode() != Hexagon::A2_addi)
305           continue;
306         if (BI.getOperand(1).getReg() != BP)
307           continue;
308         const auto &Op2 = BI.getOperand(2);
309         if (!Op2.isImm() || Op2.getImm() != RealOffset)
310           continue;
311 
312         Register R = BI.getOperand(0).getReg();
313         if (R.isPhysical()) {
314           if (Defs.available(R))
315             ReuseBP = R;
316         } else if (R.isVirtual()) {
317           // Extending a range of a virtual register can be dangerous,
318           // since the scavenger will need to find a physical register
319           // for it. Avoid extending the range past a function call,
320           // and avoid overlapping it with another virtual register.
321           if (!PassedCall && SeenVRegs.size() <= 1)
322             ReuseBP = R;
323         }
324         break;
325       }
326       if (ReuseBP)
327         ++ReuseCount;
328     }
329 
330     auto &MRI = MF.getRegInfo();
331     if (!ReuseBP) {
332       ReuseBP = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass);
333       const DebugLoc &DL = MI.getDebugLoc();
334       BuildMI(MB, II, DL, HII.get(Hexagon::A2_addi), ReuseBP)
335         .addReg(BP)
336         .addImm(RealOffset);
337     }
338     BP = ReuseBP;
339     RealOffset = InstOffset;
340   }
341 
342   MI.getOperand(FIOp).ChangeToRegister(BP, false, false, false);
343   MI.getOperand(FIOp+1).ChangeToImmediate(RealOffset);
344   return false;
345 }
346 
347 
348 bool HexagonRegisterInfo::shouldCoalesce(MachineInstr *MI,
349       const TargetRegisterClass *SrcRC, unsigned SubReg,
350       const TargetRegisterClass *DstRC, unsigned DstSubReg,
351       const TargetRegisterClass *NewRC, LiveIntervals &LIS) const {
352   // Coalescing will extend the live interval of the destination register.
353   // If the destination register is a vector pair, avoid introducing function
354   // calls into the interval, since it could result in a spilling of a pair
355   // instead of a single vector.
356   MachineFunction &MF = *MI->getParent()->getParent();
357   const HexagonSubtarget &HST = MF.getSubtarget<HexagonSubtarget>();
358   if (!HST.useHVXOps() || NewRC->getID() != Hexagon::HvxWRRegClass.getID())
359     return true;
360   bool SmallSrc = SrcRC->getID() == Hexagon::HvxVRRegClass.getID();
361   bool SmallDst = DstRC->getID() == Hexagon::HvxVRRegClass.getID();
362   if (!SmallSrc && !SmallDst)
363     return true;
364 
365   Register DstReg = MI->getOperand(0).getReg();
366   Register SrcReg = MI->getOperand(1).getReg();
367   const SlotIndexes &Indexes = *LIS.getSlotIndexes();
368   auto HasCall = [&Indexes] (const LiveInterval::Segment &S) {
369     for (SlotIndex I = S.start.getBaseIndex(), E = S.end.getBaseIndex();
370          I != E; I = I.getNextIndex()) {
371       if (const MachineInstr *MI = Indexes.getInstructionFromIndex(I))
372         if (MI->isCall())
373           return true;
374     }
375     return false;
376   };
377 
378   if (SmallSrc == SmallDst) {
379     // Both must be true, because the case for both being false was
380     // checked earlier. Both registers will be coalesced into a register
381     // of a wider class (HvxWR), and we don't want its live range to
382     // span over calls.
383     return !any_of(LIS.getInterval(DstReg), HasCall) &&
384            !any_of(LIS.getInterval(SrcReg), HasCall);
385   }
386 
387   // If one register is large (HvxWR) and the other is small (HvxVR), then
388   // coalescing is ok if the large is already live across a function call,
389   // or if the small one is not.
390   Register SmallReg = SmallSrc ? SrcReg : DstReg;
391   Register LargeReg = SmallSrc ? DstReg : SrcReg;
392   return  any_of(LIS.getInterval(LargeReg), HasCall) ||
393          !any_of(LIS.getInterval(SmallReg), HasCall);
394 }
395 
396 
397 Register HexagonRegisterInfo::getFrameRegister(const MachineFunction
398                                                &MF) const {
399   const HexagonFrameLowering *TFI = getFrameLowering(MF);
400   if (TFI->hasFP(MF))
401     return getFrameRegister();
402   return getStackRegister();
403 }
404 
405 
406 Register HexagonRegisterInfo::getFrameRegister() const {
407   return Hexagon::R30;
408 }
409 
410 
411 Register HexagonRegisterInfo::getStackRegister() const {
412   return Hexagon::R29;
413 }
414 
415 
416 unsigned HexagonRegisterInfo::getHexagonSubRegIndex(
417       const TargetRegisterClass &RC, unsigned GenIdx) const {
418   assert(GenIdx == Hexagon::ps_sub_lo || GenIdx == Hexagon::ps_sub_hi);
419 
420   static const unsigned ISub[] = { Hexagon::isub_lo, Hexagon::isub_hi };
421   static const unsigned VSub[] = { Hexagon::vsub_lo, Hexagon::vsub_hi };
422   static const unsigned WSub[] = { Hexagon::wsub_lo, Hexagon::wsub_hi };
423 
424   switch (RC.getID()) {
425     case Hexagon::CtrRegs64RegClassID:
426     case Hexagon::DoubleRegsRegClassID:
427       return ISub[GenIdx];
428     case Hexagon::HvxWRRegClassID:
429       return VSub[GenIdx];
430     case Hexagon::HvxVQRRegClassID:
431       return WSub[GenIdx];
432   }
433 
434   if (!RC.superclasses().empty())
435     return getHexagonSubRegIndex(*getRegClass(*RC.superclasses().begin()),
436                                  GenIdx);
437 
438   llvm_unreachable("Invalid register class");
439 }
440 
441 bool HexagonRegisterInfo::useFPForScavengingIndex(const MachineFunction &MF)
442       const {
443   return MF.getSubtarget<HexagonSubtarget>().getFrameLowering()->hasFP(MF);
444 }
445 
446 const TargetRegisterClass *
447 HexagonRegisterInfo::getPointerRegClass(const MachineFunction &MF,
448                                         unsigned Kind) const {
449   return &Hexagon::IntRegsRegClass;
450 }
451