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 
60 private:
61   const R600InstrInfo *TII;
62   const R600RegisterInfo &TRI;
63   bool VLIW5;
64   bool ConsideredInstUsesAlreadyWrittenVectorElement;
65 
getSlot(const MachineInstr * MI) const66   unsigned getSlot(const MachineInstr *MI) const {
67     return TRI.getHWRegChan(MI->getOperand(0).getReg());
68   }
69 
70   /// \returns register to PV chan mapping for bundle/single instructions that
71   /// immediately precedes I.
getPreviousVector(MachineBasicBlock::iterator I) const72   DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
73       const {
74     DenseMap<unsigned, unsigned> Result;
75     I--;
76     if (!TII->isALUInstr(I->getOpcode()) && !I->isBundle())
77       return Result;
78     MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
79     if (I->isBundle())
80       BI++;
81     int LastDstChan = -1;
82     do {
83       bool isTrans = false;
84       int BISlot = getSlot(BI);
85       if (LastDstChan >= BISlot)
86         isTrans = true;
87       LastDstChan = BISlot;
88       if (TII->isPredicated(BI))
89         continue;
90       int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
91       if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm() == 0)
92         continue;
93       int DstIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::dst);
94       if (DstIdx == -1) {
95         continue;
96       }
97       unsigned Dst = BI->getOperand(DstIdx).getReg();
98       if (isTrans || TII->isTransOnly(BI)) {
99         Result[Dst] = AMDGPU::PS;
100         continue;
101       }
102       if (BI->getOpcode() == AMDGPU::DOT4_r600 ||
103           BI->getOpcode() == AMDGPU::DOT4_eg) {
104         Result[Dst] = AMDGPU::PV_X;
105         continue;
106       }
107       if (Dst == AMDGPU::OQAP) {
108         continue;
109       }
110       unsigned PVReg = 0;
111       switch (TRI.getHWRegChan(Dst)) {
112       case 0:
113         PVReg = AMDGPU::PV_X;
114         break;
115       case 1:
116         PVReg = AMDGPU::PV_Y;
117         break;
118       case 2:
119         PVReg = AMDGPU::PV_Z;
120         break;
121       case 3:
122         PVReg = AMDGPU::PV_W;
123         break;
124       default:
125         llvm_unreachable("Invalid Chan");
126       }
127       Result[Dst] = PVReg;
128     } while ((++BI)->isBundledWithPred());
129     return Result;
130   }
131 
substitutePV(MachineInstr * MI,const DenseMap<unsigned,unsigned> & PVs) const132   void substitutePV(MachineInstr *MI, const DenseMap<unsigned, unsigned> &PVs)
133       const {
134     unsigned Ops[] = {
135       AMDGPU::OpName::src0,
136       AMDGPU::OpName::src1,
137       AMDGPU::OpName::src2
138     };
139     for (unsigned i = 0; i < 3; i++) {
140       int OperandIdx = TII->getOperandIdx(MI->getOpcode(), Ops[i]);
141       if (OperandIdx < 0)
142         continue;
143       unsigned Src = MI->getOperand(OperandIdx).getReg();
144       const DenseMap<unsigned, unsigned>::const_iterator It = PVs.find(Src);
145       if (It != PVs.end())
146         MI->getOperand(OperandIdx).setReg(It->second);
147     }
148   }
149 public:
150   // Ctor.
R600PacketizerList(MachineFunction & MF,MachineLoopInfo & MLI)151   R600PacketizerList(MachineFunction &MF, MachineLoopInfo &MLI)
152       : VLIWPacketizerList(MF, MLI, true),
153         TII(static_cast<const R600InstrInfo *>(
154             MF.getSubtarget().getInstrInfo())),
155         TRI(TII->getRegisterInfo()) {
156     VLIW5 = !MF.getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
157   }
158 
159   // initPacketizerState - initialize some internal flags.
initPacketizerState()160   void initPacketizerState() override {
161     ConsideredInstUsesAlreadyWrittenVectorElement = false;
162   }
163 
164   // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
ignorePseudoInstruction(MachineInstr * MI,MachineBasicBlock * MBB)165   bool ignorePseudoInstruction(MachineInstr *MI,
166                                MachineBasicBlock *MBB) override {
167     return false;
168   }
169 
170   // isSoloInstruction - return true if instruction MI can not be packetized
171   // with any other instruction, which means that MI itself is a packet.
isSoloInstruction(MachineInstr * MI)172   bool isSoloInstruction(MachineInstr *MI) override {
173     if (TII->isVector(*MI))
174       return true;
175     if (!TII->isALUInstr(MI->getOpcode()))
176       return true;
177     if (MI->getOpcode() == AMDGPU::GROUP_BARRIER)
178       return true;
179     // XXX: This can be removed once the packetizer properly handles all the
180     // LDS instruction group restrictions.
181     if (TII->isLDSInstr(MI->getOpcode()))
182       return true;
183     return false;
184   }
185 
186   // isLegalToPacketizeTogether - Is it legal to packetize SUI and SUJ
187   // together.
isLegalToPacketizeTogether(SUnit * SUI,SUnit * SUJ)188   bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) override {
189     MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
190     if (getSlot(MII) == getSlot(MIJ))
191       ConsideredInstUsesAlreadyWrittenVectorElement = true;
192     // Does MII and MIJ share the same pred_sel ?
193     int OpI = TII->getOperandIdx(MII->getOpcode(), AMDGPU::OpName::pred_sel),
194         OpJ = TII->getOperandIdx(MIJ->getOpcode(), AMDGPU::OpName::pred_sel);
195     unsigned PredI = (OpI > -1)?MII->getOperand(OpI).getReg():0,
196         PredJ = (OpJ > -1)?MIJ->getOperand(OpJ).getReg():0;
197     if (PredI != PredJ)
198       return false;
199     if (SUJ->isSucc(SUI)) {
200       for (unsigned i = 0, e = SUJ->Succs.size(); i < e; ++i) {
201         const SDep &Dep = SUJ->Succs[i];
202         if (Dep.getSUnit() != SUI)
203           continue;
204         if (Dep.getKind() == SDep::Anti)
205           continue;
206         if (Dep.getKind() == SDep::Output)
207           if (MII->getOperand(0).getReg() != MIJ->getOperand(0).getReg())
208             continue;
209         return false;
210       }
211     }
212 
213     bool ARDef = TII->definesAddressRegister(MII) ||
214                  TII->definesAddressRegister(MIJ);
215     bool ARUse = TII->usesAddressRegister(MII) ||
216                  TII->usesAddressRegister(MIJ);
217     if (ARDef && ARUse)
218       return false;
219 
220     return true;
221   }
222 
223   // isLegalToPruneDependencies - Is it legal to prune dependece between SUI
224   // and SUJ.
isLegalToPruneDependencies(SUnit * SUI,SUnit * SUJ)225   bool isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) override {
226     return false;
227   }
228 
setIsLastBit(MachineInstr * MI,unsigned Bit) const229   void setIsLastBit(MachineInstr *MI, unsigned Bit) const {
230     unsigned LastOp = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::last);
231     MI->getOperand(LastOp).setImm(Bit);
232   }
233 
isBundlableWithCurrentPMI(MachineInstr * MI,const DenseMap<unsigned,unsigned> & PV,std::vector<R600InstrInfo::BankSwizzle> & BS,bool & isTransSlot)234   bool isBundlableWithCurrentPMI(MachineInstr *MI,
235                                  const DenseMap<unsigned, unsigned> &PV,
236                                  std::vector<R600InstrInfo::BankSwizzle> &BS,
237                                  bool &isTransSlot) {
238     isTransSlot = TII->isTransOnly(MI);
239     assert (!isTransSlot || VLIW5);
240 
241     // Is the dst reg sequence legal ?
242     if (!isTransSlot && !CurrentPacketMIs.empty()) {
243       if (getSlot(MI) <= getSlot(CurrentPacketMIs.back())) {
244         if (ConsideredInstUsesAlreadyWrittenVectorElement  &&
245             !TII->isVectorOnly(MI) && VLIW5) {
246           isTransSlot = true;
247           DEBUG(dbgs() << "Considering as Trans Inst :"; MI->dump(););
248         }
249         else
250           return false;
251       }
252     }
253 
254     // Are the Constants limitations met ?
255     CurrentPacketMIs.push_back(MI);
256     if (!TII->fitsConstReadLimitations(CurrentPacketMIs)) {
257       DEBUG(
258         dbgs() << "Couldn't pack :\n";
259         MI->dump();
260         dbgs() << "with the following packets :\n";
261         for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
262           CurrentPacketMIs[i]->dump();
263           dbgs() << "\n";
264         }
265         dbgs() << "because of Consts read limitations\n";
266       );
267       CurrentPacketMIs.pop_back();
268       return false;
269     }
270 
271     // Is there a BankSwizzle set that meet Read Port limitations ?
272     if (!TII->fitsReadPortLimitations(CurrentPacketMIs,
273             PV, BS, isTransSlot)) {
274       DEBUG(
275         dbgs() << "Couldn't pack :\n";
276         MI->dump();
277         dbgs() << "with the following packets :\n";
278         for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
279           CurrentPacketMIs[i]->dump();
280           dbgs() << "\n";
281         }
282         dbgs() << "because of Read port limitations\n";
283       );
284       CurrentPacketMIs.pop_back();
285       return false;
286     }
287 
288     // We cannot read LDS source registrs from the Trans slot.
289     if (isTransSlot && TII->readsLDSSrcReg(MI))
290       return false;
291 
292     CurrentPacketMIs.pop_back();
293     return true;
294   }
295 
addToPacket(MachineInstr * MI)296   MachineBasicBlock::iterator addToPacket(MachineInstr *MI) override {
297     MachineBasicBlock::iterator FirstInBundle =
298         CurrentPacketMIs.empty() ? MI : CurrentPacketMIs.front();
299     const DenseMap<unsigned, unsigned> &PV =
300         getPreviousVector(FirstInBundle);
301     std::vector<R600InstrInfo::BankSwizzle> BS;
302     bool isTransSlot;
303 
304     if (isBundlableWithCurrentPMI(MI, PV, BS, isTransSlot)) {
305       for (unsigned i = 0, e = CurrentPacketMIs.size(); i < e; i++) {
306         MachineInstr *MI = CurrentPacketMIs[i];
307         unsigned Op = TII->getOperandIdx(MI->getOpcode(),
308             AMDGPU::OpName::bank_swizzle);
309         MI->getOperand(Op).setImm(BS[i]);
310       }
311       unsigned Op = TII->getOperandIdx(MI->getOpcode(),
312           AMDGPU::OpName::bank_swizzle);
313       MI->getOperand(Op).setImm(BS.back());
314       if (!CurrentPacketMIs.empty())
315         setIsLastBit(CurrentPacketMIs.back(), 0);
316       substitutePV(MI, PV);
317       MachineBasicBlock::iterator It = VLIWPacketizerList::addToPacket(MI);
318       if (isTransSlot) {
319         endPacket(std::next(It)->getParent(), std::next(It));
320       }
321       return It;
322     }
323     endPacket(MI->getParent(), MI);
324     if (TII->isTransOnly(MI))
325       return MI;
326     return VLIWPacketizerList::addToPacket(MI);
327   }
328 };
329 
runOnMachineFunction(MachineFunction & Fn)330 bool R600Packetizer::runOnMachineFunction(MachineFunction &Fn) {
331   const TargetInstrInfo *TII = Fn.getSubtarget().getInstrInfo();
332   MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
333 
334   // Instantiate the packetizer.
335   R600PacketizerList Packetizer(Fn, MLI);
336 
337   // DFA state table should not be empty.
338   assert(Packetizer.getResourceTracker() && "Empty DFA table!");
339 
340   //
341   // Loop over all basic blocks and remove KILL pseudo-instructions
342   // These instructions confuse the dependence analysis. Consider:
343   // D0 = ...   (Insn 0)
344   // R0 = KILL R0, D0 (Insn 1)
345   // R0 = ... (Insn 2)
346   // Here, Insn 1 will result in the dependence graph not emitting an output
347   // dependence between Insn 0 and Insn 2. This can lead to incorrect
348   // packetization
349   //
350   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
351        MBB != MBBe; ++MBB) {
352     MachineBasicBlock::iterator End = MBB->end();
353     MachineBasicBlock::iterator MI = MBB->begin();
354     while (MI != End) {
355       if (MI->isKill() || MI->getOpcode() == AMDGPU::IMPLICIT_DEF ||
356           (MI->getOpcode() == AMDGPU::CF_ALU && !MI->getOperand(8).getImm())) {
357         MachineBasicBlock::iterator DeleteMI = MI;
358         ++MI;
359         MBB->erase(DeleteMI);
360         End = MBB->end();
361         continue;
362       }
363       ++MI;
364     }
365   }
366 
367   // Loop over all of the basic blocks.
368   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
369        MBB != MBBe; ++MBB) {
370     // Find scheduling regions and schedule / packetize each region.
371     unsigned RemainingCount = MBB->size();
372     for(MachineBasicBlock::iterator RegionEnd = MBB->end();
373         RegionEnd != MBB->begin();) {
374       // The next region starts above the previous region. Look backward in the
375       // instruction stream until we find the nearest boundary.
376       MachineBasicBlock::iterator I = RegionEnd;
377       for(;I != MBB->begin(); --I, --RemainingCount) {
378         if (TII->isSchedulingBoundary(std::prev(I), MBB, Fn))
379           break;
380       }
381       I = MBB->begin();
382 
383       // Skip empty scheduling regions.
384       if (I == RegionEnd) {
385         RegionEnd = std::prev(RegionEnd);
386         --RemainingCount;
387         continue;
388       }
389       // Skip regions with one instruction.
390       if (I == std::prev(RegionEnd)) {
391         RegionEnd = std::prev(RegionEnd);
392         continue;
393       }
394 
395       Packetizer.PacketizeMIs(MBB, I, RegionEnd);
396       RegionEnd = I;
397     }
398   }
399 
400   return true;
401 
402 }
403 
404 } // end anonymous namespace
405 
createR600Packetizer(TargetMachine & tm)406 llvm::FunctionPass *llvm::createR600Packetizer(TargetMachine &tm) {
407   return new R600Packetizer(tm);
408 }
409