1//===----- R600Packetizer.cpp - VLIW packetizer ---------------------------===//
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/// \file
10/// This pass implements instructions packetization for R600. It unsets isLast
11/// bit of instructions inside a bundle and substitutes src register with
12/// PreviousVector when applicable.
13//
14//===----------------------------------------------------------------------===//
15
16#include "MCTargetDesc/R600MCTargetDesc.h"
17#include "R600.h"
18#include "R600Subtarget.h"
19#include "llvm/CodeGen/DFAPacketizer.h"
20#include "llvm/CodeGen/MachineDominators.h"
21#include "llvm/CodeGen/MachineLoopInfo.h"
22#include "llvm/CodeGen/ScheduleDAG.h"
23
24using namespace llvm;
25
26#define DEBUG_TYPE "packets"
27
28namespace {
29
30class R600Packetizer : public MachineFunctionPass {
31
32public:
33 static char ID;
34 R600Packetizer() : MachineFunctionPass(ID) {}
35
36 void getAnalysisUsage(AnalysisUsage &AU) const override {
37 AU.setPreservesCFG();
38 AU.addRequired<MachineDominatorTree>();
39 AU.addPreserved<MachineDominatorTree>();
40 AU.addRequired<MachineLoopInfo>();
41 AU.addPreserved<MachineLoopInfo>();
42 MachineFunctionPass::getAnalysisUsage(AU);
43 }
44
45 StringRef getPassName() const override { return "R600 Packetizer"; }
46
47 bool runOnMachineFunction(MachineFunction &Fn) override;
48};
49
50class R600PacketizerList : public VLIWPacketizerList {
51private:
52 const R600InstrInfo *TII;
53 const R600RegisterInfo &TRI;
54 bool VLIW5;
55 bool ConsideredInstUsesAlreadyWrittenVectorElement;
56
57 unsigned getSlot(const MachineInstr &MI) const {
58 return TRI.getHWRegChan(reg: MI.getOperand(i: 0).getReg());
59 }
60
61 /// \returns register to PV chan mapping for bundle/single instructions that
62 /// immediately precedes I.
63 DenseMap<unsigned, unsigned> getPreviousVector(MachineBasicBlock::iterator I)
64 const {
65 DenseMap<unsigned, unsigned> Result;
66 I--;
67 if (!TII->isALUInstr(Opcode: I->getOpcode()) && !I->isBundle())
68 return Result;
69 MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
70 if (I->isBundle())
71 BI++;
72 int LastDstChan = -1;
73 do {
74 bool isTrans = false;
75 int BISlot = getSlot(MI: *BI);
76 if (LastDstChan >= BISlot)
77 isTrans = true;
78 LastDstChan = BISlot;
79 if (TII->isPredicated(MI: *BI))
80 continue;
81 int OperandIdx = TII->getOperandIdx(BI->getOpcode(), R600::OpName::write);
82 if (OperandIdx > -1 && BI->getOperand(i: OperandIdx).getImm() == 0)
83 continue;
84 int DstIdx = TII->getOperandIdx(BI->getOpcode(), R600::OpName::dst);
85 if (DstIdx == -1) {
86 continue;
87 }
88 Register Dst = BI->getOperand(i: DstIdx).getReg();
89 if (isTrans || TII->isTransOnly(MI: *BI)) {
90 Result[Dst] = R600::PS;
91 continue;
92 }
93 if (BI->getOpcode() == R600::DOT4_r600 ||
94 BI->getOpcode() == R600::DOT4_eg) {
95 Result[Dst] = R600::PV_X;
96 continue;
97 }
98 if (Dst == R600::OQAP) {
99 continue;
100 }
101 unsigned PVReg = 0;
102 switch (TRI.getHWRegChan(reg: Dst)) {
103 case 0:
104 PVReg = R600::PV_X;
105 break;
106 case 1:
107 PVReg = R600::PV_Y;
108 break;
109 case 2:
110 PVReg = R600::PV_Z;
111 break;
112 case 3:
113 PVReg = R600::PV_W;
114 break;
115 default:
116 llvm_unreachable("Invalid Chan");
117 }
118 Result[Dst] = PVReg;
119 } while ((++BI)->isBundledWithPred());
120 return Result;
121 }
122
123 void substitutePV(MachineInstr &MI, const DenseMap<unsigned, unsigned> &PVs)
124 const {
125 unsigned Ops[] = {
126 R600::OpName::src0,
127 R600::OpName::src1,
128 R600::OpName::src2
129 };
130 for (unsigned Op : Ops) {
131 int OperandIdx = TII->getOperandIdx(MI.getOpcode(), Op);
132 if (OperandIdx < 0)
133 continue;
134 Register Src = MI.getOperand(OperandIdx).getReg();
135 const DenseMap<unsigned, unsigned>::const_iterator It = PVs.find(Src);
136 if (It != PVs.end())
137 MI.getOperand(OperandIdx).setReg(It->second);
138 }
139 }
140public:
141 // Ctor.
142 R600PacketizerList(MachineFunction &MF, const R600Subtarget &ST,
143 MachineLoopInfo &MLI)
144 : VLIWPacketizerList(MF, MLI, nullptr),
145 TII(ST.getInstrInfo()),
146 TRI(TII->getRegisterInfo()) {
147 VLIW5 = !ST.hasCaymanISA();
148 }
149
150 // initPacketizerState - initialize some internal flags.
151 void initPacketizerState() override {
152 ConsideredInstUsesAlreadyWrittenVectorElement = false;
153 }
154
155 // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
156 bool ignorePseudoInstruction(const MachineInstr &MI,
157 const MachineBasicBlock *MBB) override {
158 return false;
159 }
160
161 // isSoloInstruction - return true if instruction MI can not be packetized
162 // with any other instruction, which means that MI itself is a packet.
163 bool isSoloInstruction(const MachineInstr &MI) override {
164 if (TII->isVector(MI))
165 return true;
166 if (!TII->isALUInstr(Opcode: MI.getOpcode()))
167 return true;
168 if (MI.getOpcode() == R600::GROUP_BARRIER)
169 return true;
170 // XXX: This can be removed once the packetizer properly handles all the
171 // LDS instruction group restrictions.
172 return TII->isLDSInstr(Opcode: MI.getOpcode());
173 }
174
175 // isLegalToPacketizeTogether - Is it legal to packetize SUI and SUJ
176 // together.
177 bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) override {
178 MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
179 if (getSlot(MI: *MII) == getSlot(MI: *MIJ))
180 ConsideredInstUsesAlreadyWrittenVectorElement = true;
181 // Does MII and MIJ share the same pred_sel ?
182 int OpI = TII->getOperandIdx(MII->getOpcode(), R600::OpName::pred_sel),
183 OpJ = TII->getOperandIdx(MIJ->getOpcode(), R600::OpName::pred_sel);
184 Register PredI = (OpI > -1)?MII->getOperand(i: OpI).getReg() : Register(),
185 PredJ = (OpJ > -1)?MIJ->getOperand(i: OpJ).getReg() : Register();
186 if (PredI != PredJ)
187 return false;
188 if (SUJ->isSucc(N: SUI)) {
189 for (unsigned i = 0, e = SUJ->Succs.size(); i < e; ++i) {
190 const SDep &Dep = SUJ->Succs[i];
191 if (Dep.getSUnit() != SUI)
192 continue;
193 if (Dep.getKind() == SDep::Anti)
194 continue;
195 if (Dep.getKind() == SDep::Output)
196 if (MII->getOperand(i: 0).getReg() != MIJ->getOperand(i: 0).getReg())
197 continue;
198 return false;
199 }
200 }
201
202 bool ARDef =
203 TII->definesAddressRegister(MI&: *MII) || TII->definesAddressRegister(MI&: *MIJ);
204 bool ARUse =
205 TII->usesAddressRegister(MI&: *MII) || TII->usesAddressRegister(MI&: *MIJ);
206
207 return !ARDef || !ARUse;
208 }
209
210 // isLegalToPruneDependencies - Is it legal to prune dependency between SUI
211 // and SUJ.
212 bool isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) override {
213 return false;
214 }
215
216 void setIsLastBit(MachineInstr *MI, unsigned Bit) const {
217 unsigned LastOp = TII->getOperandIdx(MI->getOpcode(), R600::OpName::last);
218 MI->getOperand(i: LastOp).setImm(Bit);
219 }
220
221 bool isBundlableWithCurrentPMI(MachineInstr &MI,
222 const DenseMap<unsigned, unsigned> &PV,
223 std::vector<R600InstrInfo::BankSwizzle> &BS,
224 bool &isTransSlot) {
225 isTransSlot = TII->isTransOnly(MI);
226 assert (!isTransSlot || VLIW5);
227
228 // Is the dst reg sequence legal ?
229 if (!isTransSlot && !CurrentPacketMIs.empty()) {
230 if (getSlot(MI) <= getSlot(MI: *CurrentPacketMIs.back())) {
231 if (ConsideredInstUsesAlreadyWrittenVectorElement &&
232 !TII->isVectorOnly(MI) && VLIW5) {
233 isTransSlot = true;
234 LLVM_DEBUG({
235 dbgs() << "Considering as Trans Inst :";
236 MI.dump();
237 });
238 }
239 else
240 return false;
241 }
242 }
243
244 // Are the Constants limitations met ?
245 CurrentPacketMIs.push_back(x: &MI);
246 if (!TII->fitsConstReadLimitations(CurrentPacketMIs)) {
247 LLVM_DEBUG({
248 dbgs() << "Couldn't pack :\n";
249 MI.dump();
250 dbgs() << "with the following packets :\n";
251 for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
252 CurrentPacketMIs[i]->dump();
253 dbgs() << "\n";
254 }
255 dbgs() << "because of Consts read limitations\n";
256 });
257 CurrentPacketMIs.pop_back();
258 return false;
259 }
260
261 // Is there a BankSwizzle set that meet Read Port limitations ?
262 if (!TII->fitsReadPortLimitations(CurrentPacketMIs,
263 PV, BS, isTransSlot)) {
264 LLVM_DEBUG({
265 dbgs() << "Couldn't pack :\n";
266 MI.dump();
267 dbgs() << "with the following packets :\n";
268 for (unsigned i = 0, e = CurrentPacketMIs.size() - 1; i < e; i++) {
269 CurrentPacketMIs[i]->dump();
270 dbgs() << "\n";
271 }
272 dbgs() << "because of Read port limitations\n";
273 });
274 CurrentPacketMIs.pop_back();
275 return false;
276 }
277
278 // We cannot read LDS source registers from the Trans slot.
279 if (isTransSlot && TII->readsLDSSrcReg(MI))
280 return false;
281
282 CurrentPacketMIs.pop_back();
283 return true;
284 }
285
286 MachineBasicBlock::iterator addToPacket(MachineInstr &MI) override {
287 MachineBasicBlock::iterator FirstInBundle =
288 CurrentPacketMIs.empty() ? &MI : CurrentPacketMIs.front();
289 const DenseMap<unsigned, unsigned> &PV =
290 getPreviousVector(I: FirstInBundle);
291 std::vector<R600InstrInfo::BankSwizzle> BS;
292 bool isTransSlot;
293
294 if (isBundlableWithCurrentPMI(MI, PV, BS, isTransSlot)) {
295 for (unsigned i = 0, e = CurrentPacketMIs.size(); i < e; i++) {
296 MachineInstr *MI = CurrentPacketMIs[i];
297 unsigned Op = TII->getOperandIdx(MI->getOpcode(),
298 R600::OpName::bank_swizzle);
299 MI->getOperand(i: Op).setImm(BS[i]);
300 }
301 unsigned Op =
302 TII->getOperandIdx(MI.getOpcode(), R600::OpName::bank_swizzle);
303 MI.getOperand(i: Op).setImm(BS.back());
304 if (!CurrentPacketMIs.empty())
305 setIsLastBit(MI: CurrentPacketMIs.back(), Bit: 0);
306 substitutePV(MI, PVs: PV);
307 MachineBasicBlock::iterator It = VLIWPacketizerList::addToPacket(MI);
308 if (isTransSlot) {
309 endPacket(MBB: std::next(x: It)->getParent(), MI: std::next(x: It));
310 }
311 return It;
312 }
313 endPacket(MBB: MI.getParent(), MI);
314 if (TII->isTransOnly(MI))
315 return MI;
316 return VLIWPacketizerList::addToPacket(MI);
317 }
318};
319
320bool R600Packetizer::runOnMachineFunction(MachineFunction &Fn) {
321 const R600Subtarget &ST = Fn.getSubtarget<R600Subtarget>();
322 const R600InstrInfo *TII = ST.getInstrInfo();
323
324 MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
325
326 // Instantiate the packetizer.
327 R600PacketizerList Packetizer(Fn, ST, MLI);
328
329 // DFA state table should not be empty.
330 assert(Packetizer.getResourceTracker() && "Empty DFA table!");
331 assert(Packetizer.getResourceTracker()->getInstrItins());
332
333 if (Packetizer.getResourceTracker()->getInstrItins()->isEmpty())
334 return false;
335
336 //
337 // Loop over all basic blocks and remove KILL pseudo-instructions
338 // These instructions confuse the dependence analysis. Consider:
339 // D0 = ... (Insn 0)
340 // R0 = KILL R0, D0 (Insn 1)
341 // R0 = ... (Insn 2)
342 // Here, Insn 1 will result in the dependence graph not emitting an output
343 // dependence between Insn 0 and Insn 2. This can lead to incorrect
344 // packetization
345 //
346 for (MachineBasicBlock &MBB : Fn) {
347 for (MachineInstr &MI : llvm::make_early_inc_range(Range&: MBB)) {
348 if (MI.isKill() || MI.getOpcode() == R600::IMPLICIT_DEF ||
349 (MI.getOpcode() == R600::CF_ALU && !MI.getOperand(i: 8).getImm()))
350 MBB.erase(I: MI);
351 }
352 }
353
354 // Loop over all of the basic blocks.
355 for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
356 MBB != MBBe; ++MBB) {
357 // Find scheduling regions and schedule / packetize each region.
358 unsigned RemainingCount = MBB->size();
359 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
360 RegionEnd != MBB->begin();) {
361 // The next region starts above the previous region. Look backward in the
362 // instruction stream until we find the nearest boundary.
363 MachineBasicBlock::iterator I = RegionEnd;
364 for(;I != MBB->begin(); --I, --RemainingCount) {
365 if (TII->isSchedulingBoundary(*std::prev(x: I), &*MBB, Fn))
366 break;
367 }
368 I = MBB->begin();
369
370 // Skip empty scheduling regions.
371 if (I == RegionEnd) {
372 RegionEnd = std::prev(x: RegionEnd);
373 --RemainingCount;
374 continue;
375 }
376 // Skip regions with one instruction.
377 if (I == std::prev(x: RegionEnd)) {
378 RegionEnd = std::prev(x: RegionEnd);
379 continue;
380 }
381
382 Packetizer.PacketizeMIs(MBB: &*MBB, BeginItr: &*I, EndItr: RegionEnd);
383 RegionEnd = I;
384 }
385 }
386
387 return true;
388
389}
390
391} // end anonymous namespace
392
393INITIALIZE_PASS_BEGIN(R600Packetizer, DEBUG_TYPE,
394 "R600 Packetizer", false, false)
395INITIALIZE_PASS_END(R600Packetizer, DEBUG_TYPE,
396 "R600 Packetizer", false, false)
397
398char R600Packetizer::ID = 0;
399
400char &llvm::R600PacketizerID = R600Packetizer::ID;
401
402llvm::FunctionPass *llvm::createR600Packetizer() {
403 return new R600Packetizer();
404}
405

source code of llvm/lib/Target/AMDGPU/R600Packetizer.cpp