1//===-- PPCHazardRecognizers.cpp - PowerPC Hazard Recognizer Impls --------===//
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 implements hazard recognizers for scheduling on PowerPC processors.
10//
11//===----------------------------------------------------------------------===//
12
13#include "PPCHazardRecognizers.h"
14#include "PPCInstrInfo.h"
15#include "PPCSubtarget.h"
16#include "llvm/CodeGen/ScheduleDAG.h"
17#include "llvm/Support/Debug.h"
18#include "llvm/Support/ErrorHandling.h"
19#include "llvm/Support/raw_ostream.h"
20using namespace llvm;
21
22#define DEBUG_TYPE "pre-RA-sched"
23
24bool PPCDispatchGroupSBHazardRecognizer::isLoadAfterStore(SUnit *SU) {
25 // FIXME: Move this.
26 if (isBCTRAfterSet(SU))
27 return true;
28
29 const MCInstrDesc *MCID = DAG->getInstrDesc(SU);
30 if (!MCID)
31 return false;
32
33 if (!MCID->mayLoad())
34 return false;
35
36 // SU is a load; for any predecessors in this dispatch group, that are stores,
37 // and with which we have an ordering dependency, return true.
38 for (unsigned i = 0, ie = (unsigned) SU->Preds.size(); i != ie; ++i) {
39 const MCInstrDesc *PredMCID = DAG->getInstrDesc(SU: SU->Preds[i].getSUnit());
40 if (!PredMCID || !PredMCID->mayStore())
41 continue;
42
43 if (!SU->Preds[i].isNormalMemory() && !SU->Preds[i].isBarrier())
44 continue;
45
46 for (unsigned j = 0, je = CurGroup.size(); j != je; ++j)
47 if (SU->Preds[i].getSUnit() == CurGroup[j])
48 return true;
49 }
50
51 return false;
52}
53
54bool PPCDispatchGroupSBHazardRecognizer::isBCTRAfterSet(SUnit *SU) {
55 const MCInstrDesc *MCID = DAG->getInstrDesc(SU);
56 if (!MCID)
57 return false;
58
59 if (!MCID->isBranch())
60 return false;
61
62 // SU is a branch; for any predecessors in this dispatch group, with which we
63 // have a data dependence and set the counter register, return true.
64 for (unsigned i = 0, ie = (unsigned) SU->Preds.size(); i != ie; ++i) {
65 const MCInstrDesc *PredMCID = DAG->getInstrDesc(SU: SU->Preds[i].getSUnit());
66 if (!PredMCID || PredMCID->getSchedClass() != PPC::Sched::IIC_SprMTSPR)
67 continue;
68
69 if (SU->Preds[i].isCtrl())
70 continue;
71
72 for (unsigned j = 0, je = CurGroup.size(); j != je; ++j)
73 if (SU->Preds[i].getSUnit() == CurGroup[j])
74 return true;
75 }
76
77 return false;
78}
79
80// FIXME: Remove this when we don't need this:
81namespace llvm { namespace PPC { extern int getNonRecordFormOpcode(uint16_t); } }
82
83// FIXME: A lot of code in PPCDispatchGroupSBHazardRecognizer is P7 specific.
84
85bool PPCDispatchGroupSBHazardRecognizer::mustComeFirst(const MCInstrDesc *MCID,
86 unsigned &NSlots) {
87 // FIXME: Indirectly, this information is contained in the itinerary, and
88 // we should derive it from there instead of separately specifying it
89 // here.
90 unsigned IIC = MCID->getSchedClass();
91 switch (IIC) {
92 default:
93 NSlots = 1;
94 break;
95 case PPC::Sched::IIC_IntDivW:
96 case PPC::Sched::IIC_IntDivD:
97 case PPC::Sched::IIC_LdStLoadUpd:
98 case PPC::Sched::IIC_LdStLDU:
99 case PPC::Sched::IIC_LdStLFDU:
100 case PPC::Sched::IIC_LdStLFDUX:
101 case PPC::Sched::IIC_LdStLHA:
102 case PPC::Sched::IIC_LdStLHAU:
103 case PPC::Sched::IIC_LdStLWA:
104 case PPC::Sched::IIC_LdStSTU:
105 case PPC::Sched::IIC_LdStSTFDU:
106 NSlots = 2;
107 break;
108 case PPC::Sched::IIC_LdStLoadUpdX:
109 case PPC::Sched::IIC_LdStLDUX:
110 case PPC::Sched::IIC_LdStLHAUX:
111 case PPC::Sched::IIC_LdStLWARX:
112 case PPC::Sched::IIC_LdStLDARX:
113 case PPC::Sched::IIC_LdStSTUX:
114 case PPC::Sched::IIC_LdStSTDCX:
115 case PPC::Sched::IIC_LdStSTWCX:
116 case PPC::Sched::IIC_BrMCRX: // mtcr
117 // FIXME: Add sync/isync (here and in the itinerary).
118 NSlots = 4;
119 break;
120 }
121
122 // FIXME: record-form instructions need a different itinerary class.
123 if (NSlots == 1 && PPC::getNonRecordFormOpcode(MCID->getOpcode()) != -1)
124 NSlots = 2;
125
126 switch (IIC) {
127 default:
128 // All multi-slot instructions must come first.
129 return NSlots > 1;
130 case PPC::Sched::IIC_BrCR: // cr logicals
131 case PPC::Sched::IIC_SprMFCR:
132 case PPC::Sched::IIC_SprMFCRF:
133 case PPC::Sched::IIC_SprMTSPR:
134 return true;
135 }
136}
137
138ScheduleHazardRecognizer::HazardType
139PPCDispatchGroupSBHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
140 if (Stalls == 0 && isLoadAfterStore(SU))
141 return NoopHazard;
142
143 return ScoreboardHazardRecognizer::getHazardType(SU, Stalls);
144}
145
146bool PPCDispatchGroupSBHazardRecognizer::ShouldPreferAnother(SUnit *SU) {
147 const MCInstrDesc *MCID = DAG->getInstrDesc(SU);
148 unsigned NSlots;
149 if (MCID && mustComeFirst(MCID, NSlots) && CurSlots)
150 return true;
151
152 return ScoreboardHazardRecognizer::ShouldPreferAnother(SU);
153}
154
155unsigned PPCDispatchGroupSBHazardRecognizer::PreEmitNoops(SUnit *SU) {
156 // We only need to fill out a maximum of 5 slots here: The 6th slot could
157 // only be a second branch, and otherwise the next instruction will start a
158 // new group.
159 if (isLoadAfterStore(SU) && CurSlots < 6) {
160 unsigned Directive =
161 DAG->MF.getSubtarget<PPCSubtarget>().getCPUDirective();
162 // If we're using a special group-terminating nop, then we need only one.
163 // FIXME: the same for P9 as previous gen until POWER9 scheduling is ready
164 if (Directive == PPC::DIR_PWR6 || Directive == PPC::DIR_PWR7 ||
165 Directive == PPC::DIR_PWR8 || Directive == PPC::DIR_PWR9)
166 return 1;
167
168 return 5 - CurSlots;
169 }
170
171 return ScoreboardHazardRecognizer::PreEmitNoops(SU);
172}
173
174void PPCDispatchGroupSBHazardRecognizer::EmitInstruction(SUnit *SU) {
175 const MCInstrDesc *MCID = DAG->getInstrDesc(SU);
176 if (MCID) {
177 if (CurSlots == 5 || (MCID->isBranch() && CurBranches == 1)) {
178 CurGroup.clear();
179 CurSlots = CurBranches = 0;
180 } else {
181 LLVM_DEBUG(dbgs() << "**** Adding to dispatch group: ");
182 LLVM_DEBUG(DAG->dumpNode(*SU));
183
184 unsigned NSlots;
185 bool MustBeFirst = mustComeFirst(MCID, NSlots);
186
187 // If this instruction must come first, but does not, then it starts a
188 // new group.
189 if (MustBeFirst && CurSlots) {
190 CurSlots = CurBranches = 0;
191 CurGroup.clear();
192 }
193
194 CurSlots += NSlots;
195 CurGroup.push_back(Elt: SU);
196
197 if (MCID->isBranch())
198 ++CurBranches;
199 }
200 }
201
202 return ScoreboardHazardRecognizer::EmitInstruction(SU);
203}
204
205void PPCDispatchGroupSBHazardRecognizer::AdvanceCycle() {
206 return ScoreboardHazardRecognizer::AdvanceCycle();
207}
208
209void PPCDispatchGroupSBHazardRecognizer::RecedeCycle() {
210 llvm_unreachable("Bottom-up scheduling not supported");
211}
212
213void PPCDispatchGroupSBHazardRecognizer::Reset() {
214 CurGroup.clear();
215 CurSlots = CurBranches = 0;
216 return ScoreboardHazardRecognizer::Reset();
217}
218
219void PPCDispatchGroupSBHazardRecognizer::EmitNoop() {
220 unsigned Directive =
221 DAG->MF.getSubtarget<PPCSubtarget>().getCPUDirective();
222 // If the group has now filled all of its slots, or if we're using a special
223 // group-terminating nop, the group is complete.
224 // FIXME: the same for P9 as previous gen until POWER9 scheduling is ready
225 if (Directive == PPC::DIR_PWR6 || Directive == PPC::DIR_PWR7 ||
226 Directive == PPC::DIR_PWR8 || Directive == PPC::DIR_PWR9 ||
227 CurSlots == 6) {
228 CurGroup.clear();
229 CurSlots = CurBranches = 0;
230 } else {
231 CurGroup.push_back(Elt: nullptr);
232 ++CurSlots;
233 }
234}
235
236//===----------------------------------------------------------------------===//
237// PowerPC 970 Hazard Recognizer
238//
239// This models the dispatch group formation of the PPC970 processor. Dispatch
240// groups are bundles of up to five instructions that can contain various mixes
241// of instructions. The PPC970 can dispatch a peak of 4 non-branch and one
242// branch instruction per-cycle.
243//
244// There are a number of restrictions to dispatch group formation: some
245// instructions can only be issued in the first slot of a dispatch group, & some
246// instructions fill an entire dispatch group. Additionally, only branches can
247// issue in the 5th (last) slot.
248//
249// Finally, there are a number of "structural" hazards on the PPC970. These
250// conditions cause large performance penalties due to misprediction, recovery,
251// and replay logic that has to happen. These cases include setting a CTR and
252// branching through it in the same dispatch group, and storing to an address,
253// then loading from the same address within a dispatch group. To avoid these
254// conditions, we insert no-op instructions when appropriate.
255//
256// FIXME: This is missing some significant cases:
257// 1. Modeling of microcoded instructions.
258// 2. Handling of serialized operations.
259// 3. Handling of the esoteric cases in "Resource-based Instruction Grouping".
260//
261
262PPCHazardRecognizer970::PPCHazardRecognizer970(const ScheduleDAG &DAG)
263 : DAG(DAG) {
264 EndDispatchGroup();
265}
266
267void PPCHazardRecognizer970::EndDispatchGroup() {
268 LLVM_DEBUG(errs() << "=== Start of dispatch group\n");
269 NumIssued = 0;
270
271 // Structural hazard info.
272 HasCTRSet = false;
273 NumStores = 0;
274}
275
276
277PPCII::PPC970_Unit
278PPCHazardRecognizer970::GetInstrType(unsigned Opcode,
279 bool &isFirst, bool &isSingle,
280 bool &isCracked,
281 bool &isLoad, bool &isStore) {
282 const MCInstrDesc &MCID = DAG.TII->get(Opcode);
283
284 isLoad = MCID.mayLoad();
285 isStore = MCID.mayStore();
286
287 uint64_t TSFlags = MCID.TSFlags;
288
289 isFirst = TSFlags & PPCII::PPC970_First;
290 isSingle = TSFlags & PPCII::PPC970_Single;
291 isCracked = TSFlags & PPCII::PPC970_Cracked;
292 return (PPCII::PPC970_Unit)(TSFlags & PPCII::PPC970_Mask);
293}
294
295/// isLoadOfStoredAddress - If we have a load from the previously stored pointer
296/// as indicated by StorePtr1/StorePtr2/StoreSize, return true.
297bool PPCHazardRecognizer970::
298isLoadOfStoredAddress(uint64_t LoadSize, int64_t LoadOffset,
299 const Value *LoadValue) const {
300 for (unsigned i = 0, e = NumStores; i != e; ++i) {
301 // Handle exact and commuted addresses.
302 if (LoadValue == StoreValue[i] && LoadOffset == StoreOffset[i])
303 return true;
304
305 // Okay, we don't have an exact match, if this is an indexed offset, see if
306 // we have overlap (which happens during fp->int conversion for example).
307 if (StoreValue[i] == LoadValue) {
308 // Okay the base pointers match, so we have [c1+r] vs [c2+r]. Check
309 // to see if the load and store actually overlap.
310 if (StoreOffset[i] < LoadOffset) {
311 if (int64_t(StoreOffset[i]+StoreSize[i]) > LoadOffset) return true;
312 } else {
313 if (int64_t(LoadOffset+LoadSize) > StoreOffset[i]) return true;
314 }
315 }
316 }
317 return false;
318}
319
320/// getHazardType - We return hazard for any non-branch instruction that would
321/// terminate the dispatch group. We turn NoopHazard for any
322/// instructions that wouldn't terminate the dispatch group that would cause a
323/// pipeline flush.
324ScheduleHazardRecognizer::HazardType PPCHazardRecognizer970::
325getHazardType(SUnit *SU, int Stalls) {
326 assert(Stalls == 0 && "PPC hazards don't support scoreboard lookahead");
327
328 MachineInstr *MI = SU->getInstr();
329
330 if (MI->isDebugInstr())
331 return NoHazard;
332
333 unsigned Opcode = MI->getOpcode();
334 bool isFirst, isSingle, isCracked, isLoad, isStore;
335 PPCII::PPC970_Unit InstrType =
336 GetInstrType(Opcode, isFirst, isSingle, isCracked,
337 isLoad, isStore);
338 if (InstrType == PPCII::PPC970_Pseudo) return NoHazard;
339
340 // We can only issue a PPC970_First/PPC970_Single instruction (such as
341 // crand/mtspr/etc) if this is the first cycle of the dispatch group.
342 if (NumIssued != 0 && (isFirst || isSingle))
343 return Hazard;
344
345 // If this instruction is cracked into two ops by the decoder, we know that
346 // it is not a branch and that it cannot issue if 3 other instructions are
347 // already in the dispatch group.
348 if (isCracked && NumIssued > 2)
349 return Hazard;
350
351 switch (InstrType) {
352 default: llvm_unreachable("Unknown instruction type!");
353 case PPCII::PPC970_FXU:
354 case PPCII::PPC970_LSU:
355 case PPCII::PPC970_FPU:
356 case PPCII::PPC970_VALU:
357 case PPCII::PPC970_VPERM:
358 // We can only issue a branch as the last instruction in a group.
359 if (NumIssued == 4) return Hazard;
360 break;
361 case PPCII::PPC970_CRU:
362 // We can only issue a CR instruction in the first two slots.
363 if (NumIssued >= 2) return Hazard;
364 break;
365 case PPCII::PPC970_BRU:
366 break;
367 }
368
369 // Do not allow MTCTR and BCTRL to be in the same dispatch group.
370 if (HasCTRSet && Opcode == PPC::BCTRL)
371 return NoopHazard;
372
373 // If this is a load following a store, make sure it's not to the same or
374 // overlapping address.
375 if (isLoad && NumStores && !MI->memoperands_empty()) {
376 MachineMemOperand *MO = *MI->memoperands_begin();
377 if (MO->getSize().hasValue() &&
378 isLoadOfStoredAddress(LoadSize: MO->getSize().getValue(), LoadOffset: MO->getOffset(),
379 LoadValue: MO->getValue()))
380 return NoopHazard;
381 }
382
383 return NoHazard;
384}
385
386void PPCHazardRecognizer970::EmitInstruction(SUnit *SU) {
387 MachineInstr *MI = SU->getInstr();
388
389 if (MI->isDebugInstr())
390 return;
391
392 unsigned Opcode = MI->getOpcode();
393 bool isFirst, isSingle, isCracked, isLoad, isStore;
394 PPCII::PPC970_Unit InstrType =
395 GetInstrType(Opcode, isFirst, isSingle, isCracked,
396 isLoad, isStore);
397 if (InstrType == PPCII::PPC970_Pseudo) return;
398
399 // Update structural hazard information.
400 if (Opcode == PPC::MTCTR || Opcode == PPC::MTCTR8) HasCTRSet = true;
401
402 // Track the address stored to.
403 if (isStore && NumStores < 4 && !MI->memoperands_empty() &&
404 (*MI->memoperands_begin())->getSize().hasValue()) {
405 MachineMemOperand *MO = *MI->memoperands_begin();
406 StoreSize[NumStores] = MO->getSize().getValue();
407 StoreOffset[NumStores] = MO->getOffset();
408 StoreValue[NumStores] = MO->getValue();
409 ++NumStores;
410 }
411
412 if (InstrType == PPCII::PPC970_BRU || isSingle)
413 NumIssued = 4; // Terminate a d-group.
414 ++NumIssued;
415
416 // If this instruction is cracked into two ops by the decoder, remember that
417 // we issued two pieces.
418 if (isCracked)
419 ++NumIssued;
420
421 if (NumIssued == 5)
422 EndDispatchGroup();
423}
424
425void PPCHazardRecognizer970::AdvanceCycle() {
426 assert(NumIssued < 5 && "Illegal dispatch group!");
427 ++NumIssued;
428 if (NumIssued == 5)
429 EndDispatchGroup();
430}
431
432void PPCHazardRecognizer970::Reset() {
433 EndDispatchGroup();
434}
435
436