• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===----- R600Packetizer.cpp - VLIW packetizer ---------------------------===//
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 /// \file
11 /// This pass implements instructions packetization for R600. It unsets isLast
12 /// bit of instructions inside a bundle and substitutes src register with
13 /// PreviousVector when applicable.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "llvm/Support/Debug.h"
18 #include "AMDGPU.h"
19 #include "AMDGPUSubtarget.h"
20 #include "R600InstrInfo.h"
21 #include "llvm/CodeGen/DFAPacketizer.h"
22 #include "llvm/CodeGen/MachineDominators.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineLoopInfo.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/ScheduleDAG.h"
27 #include "llvm/Support/raw_ostream.h"
28 
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "packets"
32 
33 namespace {
34 
35 class R600Packetizer : public MachineFunctionPass {
36 
37 public:
38   static char ID;
R600Packetizer(const TargetMachine & TM)39   R600Packetizer(const TargetMachine &TM) : MachineFunctionPass(ID) {}
40 
getAnalysisUsage(AnalysisUsage & AU) const41   void getAnalysisUsage(AnalysisUsage &AU) const override {
42     AU.setPreservesCFG();
43     AU.addRequired<MachineDominatorTree>();
44     AU.addPreserved<MachineDominatorTree>();
45     AU.addRequired<MachineLoopInfo>();
46     AU.addPreserved<MachineLoopInfo>();
47     MachineFunctionPass::getAnalysisUsage(AU);
48   }
49 
getPassName() const50   const char *getPassName() const override {
51     return "R600 Packetizer";
52   }
53 
54   bool runOnMachineFunction(MachineFunction &Fn) override;
55 };
56 char R600Packetizer::ID = 0;
57 
58 class R600PacketizerList : public VLIWPacketizerList {
59 private:
60   const R600InstrInfo *TII;
61   const R600RegisterInfo &TRI;
62   bool VLIW5;
63   bool ConsideredInstUsesAlreadyWrittenVectorElement;
64 
getSlot(const MachineInstr & MI) const65   unsigned getSlot(const MachineInstr &MI) const {
66     return TRI.getHWRegChan(MI.getOperand(0).getReg());
67   }
68 
69   /// \returns register to PV chan mapping for bundle/single instructions that
70   /// immediately precedes I.
getPreviousVector(MachineBasicBlock::iterator I) const71   DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
72       const {
73     DenseMap<unsigned, unsigned> Result;
74     I--;
75     if (!TII->isALUInstr(I->getOpcode()) && !I->isBundle())
76       return Result;
77     MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
78     if (I->isBundle())
79       BI++;
80     int LastDstChan = -1;
81     do {
82       bool isTrans = false;
83       int BISlot = getSlot(*BI);
84       if (LastDstChan >= BISlot)
85         isTrans = true;
86       LastDstChan = BISlot;
87       if (TII->isPredicated(*BI))
88         continue;
89       int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
90       if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm() == 0)
91         continue;
92       int DstIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::dst);
93       if (DstIdx == -1) {
94         continue;
95       }
96       unsigned Dst = BI->getOperand(DstIdx).getReg();
97       if (isTrans || TII->isTransOnly(*BI)) {
98         Result[Dst] = AMDGPU::PS;
99         continue;
100       }
101       if (BI->getOpcode() == AMDGPU::DOT4_r600 ||
102           BI->getOpcode() == AMDGPU::DOT4_eg) {
103         Result[Dst] = AMDGPU::PV_X;
104         continue;
105       }
106       if (Dst == AMDGPU::OQAP) {
107         continue;
108       }
109       unsigned PVReg = 0;
110       switch (TRI.getHWRegChan(Dst)) {
111       case 0:
112         PVReg = AMDGPU::PV_X;
113         break;
114       case 1:
115         PVReg = AMDGPU::PV_Y;
116         break;
117       case 2:
118         PVReg = AMDGPU::PV_Z;
119         break;
120       case 3:
121         PVReg = AMDGPU::PV_W;
122         break;
123       default:
124         llvm_unreachable("Invalid Chan");
125       }
126       Result[Dst] = PVReg;
127     } while ((++BI)->isBundledWithPred());
128     return Result;
129   }
130 
substitutePV(MachineInstr & MI,const DenseMap<unsigned,unsigned> & PVs) const131   void substitutePV(MachineInstr &MI, const DenseMap<unsigned, unsigned> &PVs)
132       const {
133     unsigned Ops[] = {
134       AMDGPU::OpName::src0,
135       AMDGPU::OpName::src1,
136       AMDGPU::OpName::src2
137     };
138     for (unsigned i = 0; i < 3; i++) {
139       int OperandIdx = TII->getOperandIdx(MI.getOpcode(), Ops[i]);
140       if (OperandIdx < 0)
141         continue;
142       unsigned Src = MI.getOperand(OperandIdx).getReg();
143       const DenseMap<unsigned, unsigned>::const_iterator It = PVs.find(Src);
144       if (It != PVs.end())
145         MI.getOperand(OperandIdx).setReg(It->second);
146     }
147   }
148 public:
149   // Ctor.
R600PacketizerList(MachineFunction & MF,const R600Subtarget & ST,MachineLoopInfo & MLI)150   R600PacketizerList(MachineFunction &MF, const R600Subtarget &ST,
151                      MachineLoopInfo &MLI)
152       : VLIWPacketizerList(MF, MLI, nullptr),
153         TII(ST.getInstrInfo()),
154         TRI(TII->getRegisterInfo()) {
155     VLIW5 = !ST.hasCaymanISA();
156   }
157 
158   // initPacketizerState - initialize some internal flags.
initPacketizerState()159   void initPacketizerState() override {
160     ConsideredInstUsesAlreadyWrittenVectorElement = false;
161   }
162 
163   // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
ignorePseudoInstruction(const MachineInstr & MI,const MachineBasicBlock * MBB)164   bool ignorePseudoInstruction(const MachineInstr &MI,
165                                const MachineBasicBlock *MBB) override {
166     return false;
167   }
168 
169   // isSoloInstruction - return true if instruction MI can not be packetized
170   // with any other instruction, which means that MI itself is a packet.
isSoloInstruction(const MachineInstr & MI)171   bool isSoloInstruction(const MachineInstr &MI) override {
172     if (TII->isVector(MI))
173       return true;
174     if (!TII->isALUInstr(MI.getOpcode()))
175       return true;
176     if (MI.getOpcode() == AMDGPU::GROUP_BARRIER)
177       return true;
178     // XXX: This can be removed once the packetizer properly handles all the
179     // LDS instruction group restrictions.
180     return TII->isLDSInstr(MI.getOpcode());
181   }
182 
183   // isLegalToPacketizeTogether - Is it legal to packetize SUI and SUJ
184   // together.
isLegalToPacketizeTogether(SUnit * SUI,SUnit * SUJ)185   bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) override {
186     MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
187     if (getSlot(*MII) == getSlot(*MIJ))
188       ConsideredInstUsesAlreadyWrittenVectorElement = true;
189     // Does MII and MIJ share the same pred_sel ?
190     int OpI = TII->getOperandIdx(MII->getOpcode(), AMDGPU::OpName::pred_sel),
191         OpJ = TII->getOperandIdx(MIJ->getOpcode(), AMDGPU::OpName::pred_sel);
192     unsigned PredI = (OpI > -1)?MII->getOperand(OpI).getReg():0,
193         PredJ = (OpJ > -1)?MIJ->getOperand(OpJ).getReg():0;
194     if (PredI != PredJ)
195       return false;
196     if (SUJ->isSucc(SUI)) {
197       for (unsigned i = 0, e = SUJ->Succs.size(); i < e; ++i) {
198         const SDep &Dep = SUJ->Succs[i];
199         if (Dep.getSUnit() != SUI)
200           continue;
201         if (Dep.getKind() == SDep::Anti)
202           continue;
203         if (Dep.getKind() == SDep::Output)
204           if (MII->getOperand(0).getReg() != MIJ->getOperand(0).getReg())
205             continue;
206         return false;
207       }
208     }
209 
210     bool ARDef =
211         TII->definesAddressRegister(*MII) || TII->definesAddressRegister(*MIJ);
212     bool ARUse =
213         TII->usesAddressRegister(*MII) || TII->usesAddressRegister(*MIJ);
214 
215     return !ARDef || !ARUse;
216   }
217 
218   // isLegalToPruneDependencies - Is it legal to prune dependece between SUI
219   // and SUJ.
isLegalToPruneDependencies(SUnit * SUI,SUnit * SUJ)220   bool isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) override {
221     return false;
222   }
223 
setIsLastBit(MachineInstr * MI,unsigned Bit) const224   void setIsLastBit(MachineInstr *MI, unsigned Bit) const {
225     unsigned LastOp = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::last);
226     MI->getOperand(LastOp).setImm(Bit);
227   }
228 
isBundlableWithCurrentPMI(MachineInstr & MI,const DenseMap<unsigned,unsigned> & PV,std::vector<R600InstrInfo::BankSwizzle> & BS,bool & isTransSlot)229   bool isBundlableWithCurrentPMI(MachineInstr &MI,
230                                  const DenseMap<unsigned, unsigned> &PV,
231                                  std::vector<R600InstrInfo::BankSwizzle> &BS,
232                                  bool &isTransSlot) {
233     isTransSlot = TII->isTransOnly(MI);
234     assert (!isTransSlot || VLIW5);
235 
236     // Is the dst reg sequence legal ?
237     if (!isTransSlot && !CurrentPacketMIs.empty()) {
238       if (getSlot(MI) <= getSlot(*CurrentPacketMIs.back())) {
239         if (ConsideredInstUsesAlreadyWrittenVectorElement &&
240             !TII->isVectorOnly(MI) && VLIW5) {
241           isTransSlot = true;
242           DEBUG({
243             dbgs() << "Considering as Trans Inst :";
244             MI.dump();
245           });
246         }
247         else
248           return false;
249       }
250     }
251 
252     // Are the Constants limitations met ?
253     CurrentPacketMIs.push_back(&MI);
254     if (!TII->fitsConstReadLimitations(CurrentPacketMIs)) {
255       DEBUG({
256         dbgs() << "Couldn't pack :\n";
257         MI.dump();
258         dbgs() << "with the following packets :\n";
259         for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
260           CurrentPacketMIs[i]->dump();
261           dbgs() << "\n";
262         }
263         dbgs() << "because of Consts read limitations\n";
264       });
265       CurrentPacketMIs.pop_back();
266       return false;
267     }
268 
269     // Is there a BankSwizzle set that meet Read Port limitations ?
270     if (!TII->fitsReadPortLimitations(CurrentPacketMIs,
271             PV, BS, isTransSlot)) {
272       DEBUG({
273         dbgs() << "Couldn't pack :\n";
274         MI.dump();
275         dbgs() << "with the following packets :\n";
276         for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
277           CurrentPacketMIs[i]->dump();
278           dbgs() << "\n";
279         }
280         dbgs() << "because of Read port limitations\n";
281       });
282       CurrentPacketMIs.pop_back();
283       return false;
284     }
285 
286     // We cannot read LDS source registrs from the Trans slot.
287     if (isTransSlot && TII->readsLDSSrcReg(MI))
288       return false;
289 
290     CurrentPacketMIs.pop_back();
291     return true;
292   }
293 
addToPacket(MachineInstr & MI)294   MachineBasicBlock::iterator addToPacket(MachineInstr &MI) override {
295     MachineBasicBlock::iterator FirstInBundle =
296         CurrentPacketMIs.empty() ? &MI : CurrentPacketMIs.front();
297     const DenseMap<unsigned, unsigned> &PV =
298         getPreviousVector(FirstInBundle);
299     std::vector<R600InstrInfo::BankSwizzle> BS;
300     bool isTransSlot;
301 
302     if (isBundlableWithCurrentPMI(MI, PV, BS, isTransSlot)) {
303       for (unsigned i = 0, e = CurrentPacketMIs.size(); i < e; i++) {
304         MachineInstr *MI = CurrentPacketMIs[i];
305         unsigned Op = TII->getOperandIdx(MI->getOpcode(),
306             AMDGPU::OpName::bank_swizzle);
307         MI->getOperand(Op).setImm(BS[i]);
308       }
309       unsigned Op =
310           TII->getOperandIdx(MI.getOpcode(), AMDGPU::OpName::bank_swizzle);
311       MI.getOperand(Op).setImm(BS.back());
312       if (!CurrentPacketMIs.empty())
313         setIsLastBit(CurrentPacketMIs.back(), 0);
314       substitutePV(MI, PV);
315       MachineBasicBlock::iterator It = VLIWPacketizerList::addToPacket(MI);
316       if (isTransSlot) {
317         endPacket(std::next(It)->getParent(), std::next(It));
318       }
319       return It;
320     }
321     endPacket(MI.getParent(), MI);
322     if (TII->isTransOnly(MI))
323       return MI;
324     return VLIWPacketizerList::addToPacket(MI);
325   }
326 };
327 
runOnMachineFunction(MachineFunction & Fn)328 bool R600Packetizer::runOnMachineFunction(MachineFunction &Fn) {
329   const R600Subtarget &ST = Fn.getSubtarget<R600Subtarget>();
330   const R600InstrInfo *TII = ST.getInstrInfo();
331 
332   MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
333 
334   // Instantiate the packetizer.
335   R600PacketizerList Packetizer(Fn, ST, MLI);
336 
337   // DFA state table should not be empty.
338   assert(Packetizer.getResourceTracker() && "Empty DFA table!");
339 
340   if (Packetizer.getResourceTracker()->getInstrItins()->isEmpty())
341     return false;
342 
343   //
344   // Loop over all basic blocks and remove KILL pseudo-instructions
345   // These instructions confuse the dependence analysis. Consider:
346   // D0 = ...   (Insn 0)
347   // R0 = KILL R0, D0 (Insn 1)
348   // R0 = ... (Insn 2)
349   // Here, Insn 1 will result in the dependence graph not emitting an output
350   // dependence between Insn 0 and Insn 2. This can lead to incorrect
351   // packetization
352   //
353   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
354        MBB != MBBe; ++MBB) {
355     MachineBasicBlock::iterator End = MBB->end();
356     MachineBasicBlock::iterator MI = MBB->begin();
357     while (MI != End) {
358       if (MI->isKill() || MI->getOpcode() == AMDGPU::IMPLICIT_DEF ||
359           (MI->getOpcode() == AMDGPU::CF_ALU && !MI->getOperand(8).getImm())) {
360         MachineBasicBlock::iterator DeleteMI = MI;
361         ++MI;
362         MBB->erase(DeleteMI);
363         End = MBB->end();
364         continue;
365       }
366       ++MI;
367     }
368   }
369 
370   // Loop over all of the basic blocks.
371   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
372        MBB != MBBe; ++MBB) {
373     // Find scheduling regions and schedule / packetize each region.
374     unsigned RemainingCount = MBB->size();
375     for(MachineBasicBlock::iterator RegionEnd = MBB->end();
376         RegionEnd != MBB->begin();) {
377       // The next region starts above the previous region. Look backward in the
378       // instruction stream until we find the nearest boundary.
379       MachineBasicBlock::iterator I = RegionEnd;
380       for(;I != MBB->begin(); --I, --RemainingCount) {
381         if (TII->isSchedulingBoundary(*std::prev(I), &*MBB, Fn))
382           break;
383       }
384       I = MBB->begin();
385 
386       // Skip empty scheduling regions.
387       if (I == RegionEnd) {
388         RegionEnd = std::prev(RegionEnd);
389         --RemainingCount;
390         continue;
391       }
392       // Skip regions with one instruction.
393       if (I == std::prev(RegionEnd)) {
394         RegionEnd = std::prev(RegionEnd);
395         continue;
396       }
397 
398       Packetizer.PacketizeMIs(&*MBB, &*I, RegionEnd);
399       RegionEnd = I;
400     }
401   }
402 
403   return true;
404 
405 }
406 
407 } // end anonymous namespace
408 
createR600Packetizer(TargetMachine & tm)409 llvm::FunctionPass *llvm::createR600Packetizer(TargetMachine &tm) {
410   return new R600Packetizer(tm);
411 }
412