1 //===----- R600Packetizer.cpp - VLIW packetizer ---------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Support/Debug.h"
19 #include "R600InstrInfo.h"
20 #include "llvm/CodeGen/DFAPacketizer.h"
21 #include "llvm/CodeGen/MachineDominators.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineLoopInfo.h"
24 #include "llvm/CodeGen/Passes.h"
25 #include "llvm/CodeGen/ScheduleDAG.h"
26 #include "llvm/Support/raw_ostream.h"
30 #define DEBUG_TYPE "packets"
34 class R600Packetizer : public MachineFunctionPass {
38 R600Packetizer(const TargetMachine &TM) : MachineFunctionPass(ID) {}
40 void getAnalysisUsage(AnalysisUsage &AU) const override {
42 AU.addRequired<MachineDominatorTree>();
43 AU.addPreserved<MachineDominatorTree>();
44 AU.addRequired<MachineLoopInfo>();
45 AU.addPreserved<MachineLoopInfo>();
46 MachineFunctionPass::getAnalysisUsage(AU);
49 const char *getPassName() const override {
50 return "R600 Packetizer";
53 bool runOnMachineFunction(MachineFunction &Fn) override;
55 char R600Packetizer::ID = 0;
57 class R600PacketizerList : public VLIWPacketizerList {
60 const R600InstrInfo *TII;
61 const R600RegisterInfo &TRI;
63 bool ConsideredInstUsesAlreadyWrittenVectorElement;
65 unsigned getSlot(const MachineInstr *MI) const {
66 return TRI.getHWRegChan(MI->getOperand(0).getReg());
69 /// \returns register to PV chan mapping for bundle/single instructions that
70 /// immediately precedes I.
71 DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
73 DenseMap<unsigned, unsigned> Result;
75 if (!TII->isALUInstr(I->getOpcode()) && !I->isBundle())
77 MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
83 int BISlot = getSlot(BI);
84 if (LastDstChan >= BISlot)
87 if (TII->isPredicated(BI))
89 int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
90 if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm() == 0)
92 int DstIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::dst);
96 unsigned Dst = BI->getOperand(DstIdx).getReg();
97 if (isTrans || TII->isTransOnly(BI)) {
98 Result[Dst] = AMDGPU::PS;
101 if (BI->getOpcode() == AMDGPU::DOT4_r600 ||
102 BI->getOpcode() == AMDGPU::DOT4_eg) {
103 Result[Dst] = AMDGPU::PV_X;
106 if (Dst == AMDGPU::OQAP) {
110 switch (TRI.getHWRegChan(Dst)) {
112 PVReg = AMDGPU::PV_X;
115 PVReg = AMDGPU::PV_Y;
118 PVReg = AMDGPU::PV_Z;
121 PVReg = AMDGPU::PV_W;
124 llvm_unreachable("Invalid Chan");
127 } while ((++BI)->isBundledWithPred());
131 void substitutePV(MachineInstr *MI, const DenseMap<unsigned, unsigned> &PVs)
134 AMDGPU::OpName::src0,
135 AMDGPU::OpName::src1,
138 for (unsigned i = 0; i < 3; i++) {
139 int OperandIdx = TII->getOperandIdx(MI->getOpcode(), Ops[i]);
142 unsigned Src = MI->getOperand(OperandIdx).getReg();
143 const DenseMap<unsigned, unsigned>::const_iterator It = PVs.find(Src);
145 MI->getOperand(OperandIdx).setReg(It->second);
150 R600PacketizerList(MachineFunction &MF, MachineLoopInfo &MLI,
151 MachineDominatorTree &MDT)
152 : VLIWPacketizerList(MF, MLI, MDT, true),
153 TII (static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo())),
154 TRI(TII->getRegisterInfo()) {
155 VLIW5 = !MF.getTarget().getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
158 // initPacketizerState - initialize some internal flags.
159 void initPacketizerState() override {
160 ConsideredInstUsesAlreadyWrittenVectorElement = false;
163 // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
164 bool ignorePseudoInstruction(MachineInstr *MI,
165 MachineBasicBlock *MBB) override {
169 // isSoloInstruction - return true if instruction MI can not be packetized
170 // with any other instruction, which means that MI itself is a packet.
171 bool isSoloInstruction(MachineInstr *MI) override {
172 if (TII->isVector(*MI))
174 if (!TII->isALUInstr(MI->getOpcode()))
176 if (MI->getOpcode() == AMDGPU::GROUP_BARRIER)
178 // XXX: This can be removed once the packetizer properly handles all the
179 // LDS instruction group restrictions.
180 if (TII->isLDSInstr(MI->getOpcode()))
185 // isLegalToPacketizeTogether - Is it legal to packetize SUI and SUJ
187 bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) override {
188 MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
189 if (getSlot(MII) == getSlot(MIJ))
190 ConsideredInstUsesAlreadyWrittenVectorElement = true;
191 // Does MII and MIJ share the same pred_sel ?
192 int OpI = TII->getOperandIdx(MII->getOpcode(), AMDGPU::OpName::pred_sel),
193 OpJ = TII->getOperandIdx(MIJ->getOpcode(), AMDGPU::OpName::pred_sel);
194 unsigned PredI = (OpI > -1)?MII->getOperand(OpI).getReg():0,
195 PredJ = (OpJ > -1)?MIJ->getOperand(OpJ).getReg():0;
198 if (SUJ->isSucc(SUI)) {
199 for (unsigned i = 0, e = SUJ->Succs.size(); i < e; ++i) {
200 const SDep &Dep = SUJ->Succs[i];
201 if (Dep.getSUnit() != SUI)
203 if (Dep.getKind() == SDep::Anti)
205 if (Dep.getKind() == SDep::Output)
206 if (MII->getOperand(0).getReg() != MIJ->getOperand(0).getReg())
212 bool ARDef = TII->definesAddressRegister(MII) ||
213 TII->definesAddressRegister(MIJ);
214 bool ARUse = TII->usesAddressRegister(MII) ||
215 TII->usesAddressRegister(MIJ);
222 // isLegalToPruneDependencies - Is it legal to prune dependece between SUI
224 bool isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) override {
228 void setIsLastBit(MachineInstr *MI, unsigned Bit) const {
229 unsigned LastOp = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::last);
230 MI->getOperand(LastOp).setImm(Bit);
233 bool isBundlableWithCurrentPMI(MachineInstr *MI,
234 const DenseMap<unsigned, unsigned> &PV,
235 std::vector<R600InstrInfo::BankSwizzle> &BS,
237 isTransSlot = TII->isTransOnly(MI);
238 assert (!isTransSlot || VLIW5);
240 // Is the dst reg sequence legal ?
241 if (!isTransSlot && !CurrentPacketMIs.empty()) {
242 if (getSlot(MI) <= getSlot(CurrentPacketMIs.back())) {
243 if (ConsideredInstUsesAlreadyWrittenVectorElement &&
244 !TII->isVectorOnly(MI) && VLIW5) {
246 DEBUG(dbgs() << "Considering as Trans Inst :"; MI->dump(););
253 // Are the Constants limitations met ?
254 CurrentPacketMIs.push_back(MI);
255 if (!TII->fitsConstReadLimitations(CurrentPacketMIs)) {
257 dbgs() << "Couldn't pack :\n";
259 dbgs() << "with the following packets :\n";
260 for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
261 CurrentPacketMIs[i]->dump();
264 dbgs() << "because of Consts read limitations\n";
266 CurrentPacketMIs.pop_back();
270 // Is there a BankSwizzle set that meet Read Port limitations ?
271 if (!TII->fitsReadPortLimitations(CurrentPacketMIs,
272 PV, BS, isTransSlot)) {
274 dbgs() << "Couldn't pack :\n";
276 dbgs() << "with the following packets :\n";
277 for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
278 CurrentPacketMIs[i]->dump();
281 dbgs() << "because of Read port limitations\n";
283 CurrentPacketMIs.pop_back();
287 // We cannot read LDS source registrs from the Trans slot.
288 if (isTransSlot && TII->readsLDSSrcReg(MI))
291 CurrentPacketMIs.pop_back();
295 MachineBasicBlock::iterator addToPacket(MachineInstr *MI) override {
296 MachineBasicBlock::iterator FirstInBundle =
297 CurrentPacketMIs.empty() ? MI : CurrentPacketMIs.front();
298 const DenseMap<unsigned, unsigned> &PV =
299 getPreviousVector(FirstInBundle);
300 std::vector<R600InstrInfo::BankSwizzle> BS;
303 if (isBundlableWithCurrentPMI(MI, PV, BS, isTransSlot)) {
304 for (unsigned i = 0, e = CurrentPacketMIs.size(); i < e; i++) {
305 MachineInstr *MI = CurrentPacketMIs[i];
306 unsigned Op = TII->getOperandIdx(MI->getOpcode(),
307 AMDGPU::OpName::bank_swizzle);
308 MI->getOperand(Op).setImm(BS[i]);
310 unsigned Op = TII->getOperandIdx(MI->getOpcode(),
311 AMDGPU::OpName::bank_swizzle);
312 MI->getOperand(Op).setImm(BS.back());
313 if (!CurrentPacketMIs.empty())
314 setIsLastBit(CurrentPacketMIs.back(), 0);
315 substitutePV(MI, PV);
316 MachineBasicBlock::iterator It = VLIWPacketizerList::addToPacket(MI);
318 endPacket(std::next(It)->getParent(), std::next(It));
322 endPacket(MI->getParent(), MI);
323 if (TII->isTransOnly(MI))
325 return VLIWPacketizerList::addToPacket(MI);
329 bool R600Packetizer::runOnMachineFunction(MachineFunction &Fn) {
330 const TargetInstrInfo *TII = Fn.getTarget().getInstrInfo();
331 MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
332 MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
334 // Instantiate the packetizer.
335 R600PacketizerList Packetizer(Fn, MLI, MDT);
337 // DFA state table should not be empty.
338 assert(Packetizer.getResourceTracker() && "Empty DFA table!");
341 // Loop over all basic blocks and remove KILL pseudo-instructions
342 // These instructions confuse the dependence analysis. Consider:
344 // R0 = KILL R0, D0 (Insn 1)
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
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();
355 if (MI->isKill() || MI->getOpcode() == AMDGPU::IMPLICIT_DEF ||
356 (MI->getOpcode() == AMDGPU::CF_ALU && !MI->getOperand(8).getImm())) {
357 MachineBasicBlock::iterator DeleteMI = MI;
359 MBB->erase(DeleteMI);
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))
383 // Skip empty scheduling regions.
384 if (I == RegionEnd) {
385 RegionEnd = std::prev(RegionEnd);
389 // Skip regions with one instruction.
390 if (I == std::prev(RegionEnd)) {
391 RegionEnd = std::prev(RegionEnd);
395 Packetizer.PacketizeMIs(MBB, I, RegionEnd);
404 } // end anonymous namespace
406 llvm::FunctionPass *llvm::createR600Packetizer(TargetMachine &tm) {
407 return new R600Packetizer(tm);