1 | //===- MachinePipeliner.cpp - Machine Software Pipeliner Pass -------------===// |
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 | // An implementation of the Swing Modulo Scheduling (SMS) software pipeliner. |
10 | // |
11 | // This SMS implementation is a target-independent back-end pass. When enabled, |
12 | // the pass runs just prior to the register allocation pass, while the machine |
13 | // IR is in SSA form. If software pipelining is successful, then the original |
14 | // loop is replaced by the optimized loop. The optimized loop contains one or |
15 | // more prolog blocks, the pipelined kernel, and one or more epilog blocks. If |
16 | // the instructions cannot be scheduled in a given MII, we increase the MII by |
17 | // one and try again. |
18 | // |
19 | // The SMS implementation is an extension of the ScheduleDAGInstrs class. We |
20 | // represent loop carried dependences in the DAG as order edges to the Phi |
21 | // nodes. We also perform several passes over the DAG to eliminate unnecessary |
22 | // edges that inhibit the ability to pipeline. The implementation uses the |
23 | // DFAPacketizer class to compute the minimum initiation interval and the check |
24 | // where an instruction may be inserted in the pipelined schedule. |
25 | // |
26 | // In order for the SMS pass to work, several target specific hooks need to be |
27 | // implemented to get information about the loop structure and to rewrite |
28 | // instructions. |
29 | // |
30 | //===----------------------------------------------------------------------===// |
31 | |
32 | #include "llvm/CodeGen/MachinePipeliner.h" |
33 | #include "llvm/ADT/ArrayRef.h" |
34 | #include "llvm/ADT/BitVector.h" |
35 | #include "llvm/ADT/DenseMap.h" |
36 | #include "llvm/ADT/MapVector.h" |
37 | #include "llvm/ADT/PriorityQueue.h" |
38 | #include "llvm/ADT/STLExtras.h" |
39 | #include "llvm/ADT/SetOperations.h" |
40 | #include "llvm/ADT/SetVector.h" |
41 | #include "llvm/ADT/SmallPtrSet.h" |
42 | #include "llvm/ADT/SmallSet.h" |
43 | #include "llvm/ADT/SmallVector.h" |
44 | #include "llvm/ADT/Statistic.h" |
45 | #include "llvm/ADT/iterator_range.h" |
46 | #include "llvm/Analysis/AliasAnalysis.h" |
47 | #include "llvm/Analysis/CycleAnalysis.h" |
48 | #include "llvm/Analysis/MemoryLocation.h" |
49 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
50 | #include "llvm/Analysis/ValueTracking.h" |
51 | #include "llvm/CodeGen/DFAPacketizer.h" |
52 | #include "llvm/CodeGen/LiveIntervals.h" |
53 | #include "llvm/CodeGen/MachineBasicBlock.h" |
54 | #include "llvm/CodeGen/MachineDominators.h" |
55 | #include "llvm/CodeGen/MachineFunction.h" |
56 | #include "llvm/CodeGen/MachineFunctionPass.h" |
57 | #include "llvm/CodeGen/MachineInstr.h" |
58 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
59 | #include "llvm/CodeGen/MachineLoopInfo.h" |
60 | #include "llvm/CodeGen/MachineMemOperand.h" |
61 | #include "llvm/CodeGen/MachineOperand.h" |
62 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
63 | #include "llvm/CodeGen/ModuloSchedule.h" |
64 | #include "llvm/CodeGen/Register.h" |
65 | #include "llvm/CodeGen/RegisterClassInfo.h" |
66 | #include "llvm/CodeGen/RegisterPressure.h" |
67 | #include "llvm/CodeGen/ScheduleDAG.h" |
68 | #include "llvm/CodeGen/ScheduleDAGMutation.h" |
69 | #include "llvm/CodeGen/TargetInstrInfo.h" |
70 | #include "llvm/CodeGen/TargetOpcodes.h" |
71 | #include "llvm/CodeGen/TargetPassConfig.h" |
72 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
73 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
74 | #include "llvm/Config/llvm-config.h" |
75 | #include "llvm/IR/Attributes.h" |
76 | #include "llvm/IR/Function.h" |
77 | #include "llvm/MC/LaneBitmask.h" |
78 | #include "llvm/MC/MCInstrDesc.h" |
79 | #include "llvm/MC/MCInstrItineraries.h" |
80 | #include "llvm/MC/MCRegisterInfo.h" |
81 | #include "llvm/Pass.h" |
82 | #include "llvm/Support/CommandLine.h" |
83 | #include "llvm/Support/Compiler.h" |
84 | #include "llvm/Support/Debug.h" |
85 | #include "llvm/Support/MathExtras.h" |
86 | #include "llvm/Support/raw_ostream.h" |
87 | #include <algorithm> |
88 | #include <cassert> |
89 | #include <climits> |
90 | #include <cstdint> |
91 | #include <deque> |
92 | #include <functional> |
93 | #include <iomanip> |
94 | #include <iterator> |
95 | #include <map> |
96 | #include <memory> |
97 | #include <sstream> |
98 | #include <tuple> |
99 | #include <utility> |
100 | #include <vector> |
101 | |
102 | using namespace llvm; |
103 | |
104 | #define DEBUG_TYPE "pipeliner" |
105 | |
106 | STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline" ); |
107 | STATISTIC(NumPipelined, "Number of loops software pipelined" ); |
108 | STATISTIC(NumNodeOrderIssues, "Number of node order issues found" ); |
109 | STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch" ); |
110 | STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop" ); |
111 | STATISTIC(, "Pipeliner abort due to missing preheader" ); |
112 | STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large" ); |
113 | STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII" ); |
114 | STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found" ); |
115 | STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage" ); |
116 | STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages" ); |
117 | |
118 | /// A command line option to turn software pipelining on or off. |
119 | static cl::opt<bool> EnableSWP("enable-pipeliner" , cl::Hidden, cl::init(Val: true), |
120 | cl::desc("Enable Software Pipelining" )); |
121 | |
122 | /// A command line option to enable SWP at -Os. |
123 | static cl::opt<bool> EnableSWPOptSize("enable-pipeliner-opt-size" , |
124 | cl::desc("Enable SWP at Os." ), cl::Hidden, |
125 | cl::init(Val: false)); |
126 | |
127 | /// A command line argument to limit minimum initial interval for pipelining. |
128 | static cl::opt<int> SwpMaxMii("pipeliner-max-mii" , |
129 | cl::desc("Size limit for the MII." ), |
130 | cl::Hidden, cl::init(Val: 27)); |
131 | |
132 | /// A command line argument to force pipeliner to use specified initial |
133 | /// interval. |
134 | static cl::opt<int> SwpForceII("pipeliner-force-ii" , |
135 | cl::desc("Force pipeliner to use specified II." ), |
136 | cl::Hidden, cl::init(Val: -1)); |
137 | |
138 | /// A command line argument to limit the number of stages in the pipeline. |
139 | static cl::opt<int> |
140 | SwpMaxStages("pipeliner-max-stages" , |
141 | cl::desc("Maximum stages allowed in the generated scheduled." ), |
142 | cl::Hidden, cl::init(Val: 3)); |
143 | |
144 | /// A command line option to disable the pruning of chain dependences due to |
145 | /// an unrelated Phi. |
146 | static cl::opt<bool> |
147 | SwpPruneDeps("pipeliner-prune-deps" , |
148 | cl::desc("Prune dependences between unrelated Phi nodes." ), |
149 | cl::Hidden, cl::init(Val: true)); |
150 | |
151 | /// A command line option to disable the pruning of loop carried order |
152 | /// dependences. |
153 | static cl::opt<bool> |
154 | SwpPruneLoopCarried("pipeliner-prune-loop-carried" , |
155 | cl::desc("Prune loop carried order dependences." ), |
156 | cl::Hidden, cl::init(Val: true)); |
157 | |
158 | #ifndef NDEBUG |
159 | static cl::opt<int> SwpLoopLimit("pipeliner-max" , cl::Hidden, cl::init(-1)); |
160 | #endif |
161 | |
162 | static cl::opt<bool> SwpIgnoreRecMII("pipeliner-ignore-recmii" , |
163 | cl::ReallyHidden, |
164 | cl::desc("Ignore RecMII" )); |
165 | |
166 | static cl::opt<bool> SwpShowResMask("pipeliner-show-mask" , cl::Hidden, |
167 | cl::init(Val: false)); |
168 | static cl::opt<bool> SwpDebugResource("pipeliner-dbg-res" , cl::Hidden, |
169 | cl::init(Val: false)); |
170 | |
171 | static cl::opt<bool> EmitTestAnnotations( |
172 | "pipeliner-annotate-for-testing" , cl::Hidden, cl::init(Val: false), |
173 | cl::desc("Instead of emitting the pipelined code, annotate instructions " |
174 | "with the generated schedule for feeding into the " |
175 | "-modulo-schedule-test pass" )); |
176 | |
177 | static cl::opt<bool> ExperimentalCodeGen( |
178 | "pipeliner-experimental-cg" , cl::Hidden, cl::init(Val: false), |
179 | cl::desc( |
180 | "Use the experimental peeling code generator for software pipelining" )); |
181 | |
182 | static cl::opt<int> SwpIISearchRange("pipeliner-ii-search-range" , |
183 | cl::desc("Range to search for II" ), |
184 | cl::Hidden, cl::init(Val: 10)); |
185 | |
186 | static cl::opt<bool> |
187 | LimitRegPressure("pipeliner-register-pressure" , cl::Hidden, cl::init(Val: false), |
188 | cl::desc("Limit register pressure of scheduled loop" )); |
189 | |
190 | static cl::opt<int> |
191 | RegPressureMargin("pipeliner-register-pressure-margin" , cl::Hidden, |
192 | cl::init(Val: 5), |
193 | cl::desc("Margin representing the unused percentage of " |
194 | "the register pressure limit" )); |
195 | |
196 | static cl::opt<bool> |
197 | MVECodeGen("pipeliner-mve-cg" , cl::Hidden, cl::init(Val: false), |
198 | cl::desc("Use the MVE code generator for software pipelining" )); |
199 | |
200 | namespace llvm { |
201 | |
202 | // A command line option to enable the CopyToPhi DAG mutation. |
203 | cl::opt<bool> SwpEnableCopyToPhi("pipeliner-enable-copytophi" , cl::ReallyHidden, |
204 | cl::init(Val: true), |
205 | cl::desc("Enable CopyToPhi DAG Mutation" )); |
206 | |
207 | /// A command line argument to force pipeliner to use specified issue |
208 | /// width. |
209 | cl::opt<int> SwpForceIssueWidth( |
210 | "pipeliner-force-issue-width" , |
211 | cl::desc("Force pipeliner to use specified issue width." ), cl::Hidden, |
212 | cl::init(Val: -1)); |
213 | |
214 | /// A command line argument to set the window scheduling option. |
215 | cl::opt<WindowSchedulingFlag> WindowSchedulingOption( |
216 | "window-sched" , cl::Hidden, cl::init(Val: WindowSchedulingFlag::WS_On), |
217 | cl::desc("Set how to use window scheduling algorithm." ), |
218 | cl::values(clEnumValN(WindowSchedulingFlag::WS_Off, "off" , |
219 | "Turn off window algorithm." ), |
220 | clEnumValN(WindowSchedulingFlag::WS_On, "on" , |
221 | "Use window algorithm after SMS algorithm fails." ), |
222 | clEnumValN(WindowSchedulingFlag::WS_Force, "force" , |
223 | "Use window algorithm instead of SMS algorithm." ))); |
224 | |
225 | } // end namespace llvm |
226 | |
227 | unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5; |
228 | char MachinePipeliner::ID = 0; |
229 | #ifndef NDEBUG |
230 | int MachinePipeliner::NumTries = 0; |
231 | #endif |
232 | char &llvm::MachinePipelinerID = MachinePipeliner::ID; |
233 | |
234 | INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE, |
235 | "Modulo Software Pipelining" , false, false) |
236 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
237 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
238 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
239 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
240 | INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE, |
241 | "Modulo Software Pipelining" , false, false) |
242 | |
243 | /// The "main" function for implementing Swing Modulo Scheduling. |
244 | bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) { |
245 | if (skipFunction(F: mf.getFunction())) |
246 | return false; |
247 | |
248 | if (!EnableSWP) |
249 | return false; |
250 | |
251 | if (mf.getFunction().getAttributes().hasFnAttr(Kind: Attribute::OptimizeForSize) && |
252 | !EnableSWPOptSize.getPosition()) |
253 | return false; |
254 | |
255 | if (!mf.getSubtarget().enableMachinePipeliner()) |
256 | return false; |
257 | |
258 | // Cannot pipeline loops without instruction itineraries if we are using |
259 | // DFA for the pipeliner. |
260 | if (mf.getSubtarget().useDFAforSMS() && |
261 | (!mf.getSubtarget().getInstrItineraryData() || |
262 | mf.getSubtarget().getInstrItineraryData()->isEmpty())) |
263 | return false; |
264 | |
265 | MF = &mf; |
266 | MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
267 | MDT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
268 | ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); |
269 | TII = MF->getSubtarget().getInstrInfo(); |
270 | RegClassInfo.runOnMachineFunction(MF: *MF); |
271 | |
272 | for (const auto &L : *MLI) |
273 | scheduleLoop(L&: *L); |
274 | |
275 | return false; |
276 | } |
277 | |
278 | /// Attempt to perform the SMS algorithm on the specified loop. This function is |
279 | /// the main entry point for the algorithm. The function identifies candidate |
280 | /// loops, calculates the minimum initiation interval, and attempts to schedule |
281 | /// the loop. |
282 | bool MachinePipeliner::scheduleLoop(MachineLoop &L) { |
283 | bool Changed = false; |
284 | for (const auto &InnerLoop : L) |
285 | Changed |= scheduleLoop(L&: *InnerLoop); |
286 | |
287 | #ifndef NDEBUG |
288 | // Stop trying after reaching the limit (if any). |
289 | int Limit = SwpLoopLimit; |
290 | if (Limit >= 0) { |
291 | if (NumTries >= SwpLoopLimit) |
292 | return Changed; |
293 | NumTries++; |
294 | } |
295 | #endif |
296 | |
297 | setPragmaPipelineOptions(L); |
298 | if (!canPipelineLoop(L)) { |
299 | LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n" ); |
300 | ORE->emit(RemarkBuilder: [&]() { |
301 | return MachineOptimizationRemarkMissed(DEBUG_TYPE, "canPipelineLoop" , |
302 | L.getStartLoc(), L.getHeader()) |
303 | << "Failed to pipeline loop" ; |
304 | }); |
305 | |
306 | LI.LoopPipelinerInfo.reset(); |
307 | return Changed; |
308 | } |
309 | |
310 | ++NumTrytoPipeline; |
311 | if (useSwingModuloScheduler()) |
312 | Changed = swingModuloScheduler(L); |
313 | |
314 | if (useWindowScheduler(Changed)) |
315 | Changed = runWindowScheduler(L); |
316 | |
317 | LI.LoopPipelinerInfo.reset(); |
318 | return Changed; |
319 | } |
320 | |
321 | void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) { |
322 | // Reset the pragma for the next loop in iteration. |
323 | disabledByPragma = false; |
324 | II_setByPragma = 0; |
325 | |
326 | MachineBasicBlock *LBLK = L.getTopBlock(); |
327 | |
328 | if (LBLK == nullptr) |
329 | return; |
330 | |
331 | const BasicBlock *BBLK = LBLK->getBasicBlock(); |
332 | if (BBLK == nullptr) |
333 | return; |
334 | |
335 | const Instruction *TI = BBLK->getTerminator(); |
336 | if (TI == nullptr) |
337 | return; |
338 | |
339 | MDNode *LoopID = TI->getMetadata(KindID: LLVMContext::MD_loop); |
340 | if (LoopID == nullptr) |
341 | return; |
342 | |
343 | assert(LoopID->getNumOperands() > 0 && "requires atleast one operand" ); |
344 | assert(LoopID->getOperand(0) == LoopID && "invalid loop" ); |
345 | |
346 | for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) { |
347 | MDNode *MD = dyn_cast<MDNode>(Val: MDO); |
348 | |
349 | if (MD == nullptr) |
350 | continue; |
351 | |
352 | MDString *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
353 | |
354 | if (S == nullptr) |
355 | continue; |
356 | |
357 | if (S->getString() == "llvm.loop.pipeline.initiationinterval" ) { |
358 | assert(MD->getNumOperands() == 2 && |
359 | "Pipeline initiation interval hint metadata should have two operands." ); |
360 | II_setByPragma = |
361 | mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 1))->getZExtValue(); |
362 | assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive." ); |
363 | } else if (S->getString() == "llvm.loop.pipeline.disable" ) { |
364 | disabledByPragma = true; |
365 | } |
366 | } |
367 | } |
368 | |
369 | /// Return true if the loop can be software pipelined. The algorithm is |
370 | /// restricted to loops with a single basic block. Make sure that the |
371 | /// branch in the loop can be analyzed. |
372 | bool MachinePipeliner::canPipelineLoop(MachineLoop &L) { |
373 | if (L.getNumBlocks() != 1) { |
374 | ORE->emit(RemarkBuilder: [&]() { |
375 | return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop" , |
376 | L.getStartLoc(), L.getHeader()) |
377 | << "Not a single basic block: " |
378 | << ore::NV("NumBlocks" , L.getNumBlocks()); |
379 | }); |
380 | return false; |
381 | } |
382 | |
383 | if (disabledByPragma) { |
384 | ORE->emit(RemarkBuilder: [&]() { |
385 | return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop" , |
386 | L.getStartLoc(), L.getHeader()) |
387 | << "Disabled by Pragma." ; |
388 | }); |
389 | return false; |
390 | } |
391 | |
392 | // Check if the branch can't be understood because we can't do pipelining |
393 | // if that's the case. |
394 | LI.TBB = nullptr; |
395 | LI.FBB = nullptr; |
396 | LI.BrCond.clear(); |
397 | if (TII->analyzeBranch(MBB&: *L.getHeader(), TBB&: LI.TBB, FBB&: LI.FBB, Cond&: LI.BrCond)) { |
398 | LLVM_DEBUG(dbgs() << "Unable to analyzeBranch, can NOT pipeline Loop\n" ); |
399 | NumFailBranch++; |
400 | ORE->emit(RemarkBuilder: [&]() { |
401 | return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop" , |
402 | L.getStartLoc(), L.getHeader()) |
403 | << "The branch can't be understood" ; |
404 | }); |
405 | return false; |
406 | } |
407 | |
408 | LI.LoopInductionVar = nullptr; |
409 | LI.LoopCompare = nullptr; |
410 | LI.LoopPipelinerInfo = TII->analyzeLoopForPipelining(LoopBB: L.getTopBlock()); |
411 | if (!LI.LoopPipelinerInfo) { |
412 | LLVM_DEBUG(dbgs() << "Unable to analyzeLoop, can NOT pipeline Loop\n" ); |
413 | NumFailLoop++; |
414 | ORE->emit(RemarkBuilder: [&]() { |
415 | return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop" , |
416 | L.getStartLoc(), L.getHeader()) |
417 | << "The loop structure is not supported" ; |
418 | }); |
419 | return false; |
420 | } |
421 | |
422 | if (!L.getLoopPreheader()) { |
423 | LLVM_DEBUG(dbgs() << "Preheader not found, can NOT pipeline Loop\n" ); |
424 | NumFailPreheader++; |
425 | ORE->emit(RemarkBuilder: [&]() { |
426 | return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop" , |
427 | L.getStartLoc(), L.getHeader()) |
428 | << "No loop preheader found" ; |
429 | }); |
430 | return false; |
431 | } |
432 | |
433 | // Remove any subregisters from inputs to phi nodes. |
434 | preprocessPhiNodes(B&: *L.getHeader()); |
435 | return true; |
436 | } |
437 | |
438 | void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) { |
439 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
440 | SlotIndexes &Slots = |
441 | *getAnalysis<LiveIntervalsWrapperPass>().getLIS().getSlotIndexes(); |
442 | |
443 | for (MachineInstr &PI : B.phis()) { |
444 | MachineOperand &DefOp = PI.getOperand(i: 0); |
445 | assert(DefOp.getSubReg() == 0); |
446 | auto *RC = MRI.getRegClass(Reg: DefOp.getReg()); |
447 | |
448 | for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) { |
449 | MachineOperand &RegOp = PI.getOperand(i); |
450 | if (RegOp.getSubReg() == 0) |
451 | continue; |
452 | |
453 | // If the operand uses a subregister, replace it with a new register |
454 | // without subregisters, and generate a copy to the new register. |
455 | Register NewReg = MRI.createVirtualRegister(RegClass: RC); |
456 | MachineBasicBlock &PredB = *PI.getOperand(i: i+1).getMBB(); |
457 | MachineBasicBlock::iterator At = PredB.getFirstTerminator(); |
458 | const DebugLoc &DL = PredB.findDebugLoc(MBBI: At); |
459 | auto Copy = BuildMI(BB&: PredB, I: At, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: NewReg) |
460 | .addReg(RegNo: RegOp.getReg(), flags: getRegState(RegOp), |
461 | SubReg: RegOp.getSubReg()); |
462 | Slots.insertMachineInstrInMaps(MI&: *Copy); |
463 | RegOp.setReg(NewReg); |
464 | RegOp.setSubReg(0); |
465 | } |
466 | } |
467 | } |
468 | |
469 | /// The SMS algorithm consists of the following main steps: |
470 | /// 1. Computation and analysis of the dependence graph. |
471 | /// 2. Ordering of the nodes (instructions). |
472 | /// 3. Attempt to Schedule the loop. |
473 | bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) { |
474 | assert(L.getBlocks().size() == 1 && "SMS works on single blocks only." ); |
475 | |
476 | SwingSchedulerDAG SMS( |
477 | *this, L, getAnalysis<LiveIntervalsWrapperPass>().getLIS(), RegClassInfo, |
478 | II_setByPragma, LI.LoopPipelinerInfo.get()); |
479 | |
480 | MachineBasicBlock *MBB = L.getHeader(); |
481 | // The kernel should not include any terminator instructions. These |
482 | // will be added back later. |
483 | SMS.startBlock(BB: MBB); |
484 | |
485 | // Compute the number of 'real' instructions in the basic block by |
486 | // ignoring terminators. |
487 | unsigned size = MBB->size(); |
488 | for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(), |
489 | E = MBB->instr_end(); |
490 | I != E; ++I, --size) |
491 | ; |
492 | |
493 | SMS.enterRegion(bb: MBB, begin: MBB->begin(), end: MBB->getFirstTerminator(), regioninstrs: size); |
494 | SMS.schedule(); |
495 | SMS.exitRegion(); |
496 | |
497 | SMS.finishBlock(); |
498 | return SMS.hasNewSchedule(); |
499 | } |
500 | |
501 | void MachinePipeliner::getAnalysisUsage(AnalysisUsage &AU) const { |
502 | AU.addRequired<AAResultsWrapperPass>(); |
503 | AU.addPreserved<AAResultsWrapperPass>(); |
504 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
505 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
506 | AU.addRequired<LiveIntervalsWrapperPass>(); |
507 | AU.addRequired<MachineOptimizationRemarkEmitterPass>(); |
508 | AU.addRequired<TargetPassConfig>(); |
509 | MachineFunctionPass::getAnalysisUsage(AU); |
510 | } |
511 | |
512 | bool MachinePipeliner::runWindowScheduler(MachineLoop &L) { |
513 | MachineSchedContext Context; |
514 | Context.MF = MF; |
515 | Context.MLI = MLI; |
516 | Context.MDT = MDT; |
517 | Context.PassConfig = &getAnalysis<TargetPassConfig>(); |
518 | Context.AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
519 | Context.LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
520 | Context.RegClassInfo->runOnMachineFunction(MF: *MF); |
521 | WindowScheduler WS(&Context, L); |
522 | return WS.run(); |
523 | } |
524 | |
525 | bool MachinePipeliner::useSwingModuloScheduler() { |
526 | // SwingModuloScheduler does not work when WindowScheduler is forced. |
527 | return WindowSchedulingOption != WindowSchedulingFlag::WS_Force; |
528 | } |
529 | |
530 | bool MachinePipeliner::useWindowScheduler(bool Changed) { |
531 | // WindowScheduler does not work for following cases: |
532 | // 1. when it is off. |
533 | // 2. when SwingModuloScheduler is successfully scheduled. |
534 | // 3. when pragma II is enabled. |
535 | if (II_setByPragma) { |
536 | LLVM_DEBUG(dbgs() << "Window scheduling is disabled when " |
537 | "llvm.loop.pipeline.initiationinterval is set.\n" ); |
538 | return false; |
539 | } |
540 | |
541 | return WindowSchedulingOption == WindowSchedulingFlag::WS_Force || |
542 | (WindowSchedulingOption == WindowSchedulingFlag::WS_On && !Changed); |
543 | } |
544 | |
545 | void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) { |
546 | if (SwpForceII > 0) |
547 | MII = SwpForceII; |
548 | else if (II_setByPragma > 0) |
549 | MII = II_setByPragma; |
550 | else |
551 | MII = std::max(a: ResMII, b: RecMII); |
552 | } |
553 | |
554 | void SwingSchedulerDAG::setMAX_II() { |
555 | if (SwpForceII > 0) |
556 | MAX_II = SwpForceII; |
557 | else if (II_setByPragma > 0) |
558 | MAX_II = II_setByPragma; |
559 | else |
560 | MAX_II = MII + SwpIISearchRange; |
561 | } |
562 | |
563 | /// We override the schedule function in ScheduleDAGInstrs to implement the |
564 | /// scheduling part of the Swing Modulo Scheduling algorithm. |
565 | void SwingSchedulerDAG::schedule() { |
566 | AliasAnalysis *AA = &Pass.getAnalysis<AAResultsWrapperPass>().getAAResults(); |
567 | buildSchedGraph(AA); |
568 | addLoopCarriedDependences(AA); |
569 | updatePhiDependences(); |
570 | Topo.InitDAGTopologicalSorting(); |
571 | changeDependences(); |
572 | postProcessDAG(); |
573 | LLVM_DEBUG(dump()); |
574 | |
575 | NodeSetType NodeSets; |
576 | findCircuits(NodeSets); |
577 | NodeSetType Circuits = NodeSets; |
578 | |
579 | // Calculate the MII. |
580 | unsigned ResMII = calculateResMII(); |
581 | unsigned RecMII = calculateRecMII(RecNodeSets&: NodeSets); |
582 | |
583 | fuseRecs(NodeSets); |
584 | |
585 | // This flag is used for testing and can cause correctness problems. |
586 | if (SwpIgnoreRecMII) |
587 | RecMII = 0; |
588 | |
589 | setMII(ResMII, RecMII); |
590 | setMAX_II(); |
591 | |
592 | LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II |
593 | << " (rec=" << RecMII << ", res=" << ResMII << ")\n" ); |
594 | |
595 | // Can't schedule a loop without a valid MII. |
596 | if (MII == 0) { |
597 | LLVM_DEBUG(dbgs() << "Invalid Minimal Initiation Interval: 0\n" ); |
598 | NumFailZeroMII++; |
599 | Pass.ORE->emit(RemarkBuilder: [&]() { |
600 | return MachineOptimizationRemarkAnalysis( |
601 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
602 | << "Invalid Minimal Initiation Interval: 0" ; |
603 | }); |
604 | return; |
605 | } |
606 | |
607 | // Don't pipeline large loops. |
608 | if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) { |
609 | LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii |
610 | << ", we don't pipeline large loops\n" ); |
611 | NumFailLargeMaxMII++; |
612 | Pass.ORE->emit(RemarkBuilder: [&]() { |
613 | return MachineOptimizationRemarkAnalysis( |
614 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
615 | << "Minimal Initiation Interval too large: " |
616 | << ore::NV("MII" , (int)MII) << " > " |
617 | << ore::NV("SwpMaxMii" , SwpMaxMii) << "." |
618 | << "Refer to -pipeliner-max-mii." ; |
619 | }); |
620 | return; |
621 | } |
622 | |
623 | computeNodeFunctions(NodeSets); |
624 | |
625 | registerPressureFilter(NodeSets); |
626 | |
627 | colocateNodeSets(NodeSets); |
628 | |
629 | checkNodeSets(NodeSets); |
630 | |
631 | LLVM_DEBUG({ |
632 | for (auto &I : NodeSets) { |
633 | dbgs() << " Rec NodeSet " ; |
634 | I.dump(); |
635 | } |
636 | }); |
637 | |
638 | llvm::stable_sort(Range&: NodeSets, C: std::greater<NodeSet>()); |
639 | |
640 | groupRemainingNodes(NodeSets); |
641 | |
642 | removeDuplicateNodes(NodeSets); |
643 | |
644 | LLVM_DEBUG({ |
645 | for (auto &I : NodeSets) { |
646 | dbgs() << " NodeSet " ; |
647 | I.dump(); |
648 | } |
649 | }); |
650 | |
651 | computeNodeOrder(NodeSets); |
652 | |
653 | // check for node order issues |
654 | checkValidNodeOrder(Circuits); |
655 | |
656 | SMSchedule Schedule(Pass.MF, this); |
657 | Scheduled = schedulePipeline(Schedule); |
658 | |
659 | if (!Scheduled){ |
660 | LLVM_DEBUG(dbgs() << "No schedule found, return\n" ); |
661 | NumFailNoSchedule++; |
662 | Pass.ORE->emit(RemarkBuilder: [&]() { |
663 | return MachineOptimizationRemarkAnalysis( |
664 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
665 | << "Unable to find schedule" ; |
666 | }); |
667 | return; |
668 | } |
669 | |
670 | unsigned numStages = Schedule.getMaxStageCount(); |
671 | // No need to generate pipeline if there are no overlapped iterations. |
672 | if (numStages == 0) { |
673 | LLVM_DEBUG(dbgs() << "No overlapped iterations, skip.\n" ); |
674 | NumFailZeroStage++; |
675 | Pass.ORE->emit(RemarkBuilder: [&]() { |
676 | return MachineOptimizationRemarkAnalysis( |
677 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
678 | << "No need to pipeline - no overlapped iterations in schedule." ; |
679 | }); |
680 | return; |
681 | } |
682 | // Check that the maximum stage count is less than user-defined limit. |
683 | if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) { |
684 | LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages |
685 | << " : too many stages, abort\n" ); |
686 | NumFailLargeMaxStage++; |
687 | Pass.ORE->emit(RemarkBuilder: [&]() { |
688 | return MachineOptimizationRemarkAnalysis( |
689 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
690 | << "Too many stages in schedule: " |
691 | << ore::NV("numStages" , (int)numStages) << " > " |
692 | << ore::NV("SwpMaxStages" , SwpMaxStages) |
693 | << ". Refer to -pipeliner-max-stages." ; |
694 | }); |
695 | return; |
696 | } |
697 | |
698 | Pass.ORE->emit(RemarkBuilder: [&]() { |
699 | return MachineOptimizationRemark(DEBUG_TYPE, "schedule" , Loop.getStartLoc(), |
700 | Loop.getHeader()) |
701 | << "Pipelined succesfully!" ; |
702 | }); |
703 | |
704 | // Generate the schedule as a ModuloSchedule. |
705 | DenseMap<MachineInstr *, int> Cycles, Stages; |
706 | std::vector<MachineInstr *> OrderedInsts; |
707 | for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); |
708 | ++Cycle) { |
709 | for (SUnit *SU : Schedule.getInstructions(cycle: Cycle)) { |
710 | OrderedInsts.push_back(x: SU->getInstr()); |
711 | Cycles[SU->getInstr()] = Cycle; |
712 | Stages[SU->getInstr()] = Schedule.stageScheduled(SU); |
713 | } |
714 | } |
715 | DenseMap<MachineInstr *, std::pair<unsigned, int64_t>> NewInstrChanges; |
716 | for (auto &KV : NewMIs) { |
717 | Cycles[KV.first] = Cycles[KV.second]; |
718 | Stages[KV.first] = Stages[KV.second]; |
719 | NewInstrChanges[KV.first] = InstrChanges[getSUnit(MI: KV.first)]; |
720 | } |
721 | |
722 | ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles), |
723 | std::move(Stages)); |
724 | if (EmitTestAnnotations) { |
725 | assert(NewInstrChanges.empty() && |
726 | "Cannot serialize a schedule with InstrChanges!" ); |
727 | ModuloScheduleTestAnnotater MSTI(MF, MS); |
728 | MSTI.annotate(); |
729 | return; |
730 | } |
731 | // The experimental code generator can't work if there are InstChanges. |
732 | if (ExperimentalCodeGen && NewInstrChanges.empty()) { |
733 | PeelingModuloScheduleExpander MSE(MF, MS, &LIS); |
734 | MSE.expand(); |
735 | } else if (MVECodeGen && NewInstrChanges.empty() && |
736 | LoopPipelinerInfo->isMVEExpanderSupported() && |
737 | ModuloScheduleExpanderMVE::canApply(L&: Loop)) { |
738 | ModuloScheduleExpanderMVE MSE(MF, MS, LIS); |
739 | MSE.expand(); |
740 | } else { |
741 | ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges)); |
742 | MSE.expand(); |
743 | MSE.cleanup(); |
744 | } |
745 | ++NumPipelined; |
746 | } |
747 | |
748 | /// Clean up after the software pipeliner runs. |
749 | void SwingSchedulerDAG::finishBlock() { |
750 | for (auto &KV : NewMIs) |
751 | MF.deleteMachineInstr(MI: KV.second); |
752 | NewMIs.clear(); |
753 | |
754 | // Call the superclass. |
755 | ScheduleDAGInstrs::finishBlock(); |
756 | } |
757 | |
758 | /// Return the register values for the operands of a Phi instruction. |
759 | /// This function assume the instruction is a Phi. |
760 | static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, |
761 | unsigned &InitVal, unsigned &LoopVal) { |
762 | assert(Phi.isPHI() && "Expecting a Phi." ); |
763 | |
764 | InitVal = 0; |
765 | LoopVal = 0; |
766 | for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) |
767 | if (Phi.getOperand(i: i + 1).getMBB() != Loop) |
768 | InitVal = Phi.getOperand(i).getReg(); |
769 | else |
770 | LoopVal = Phi.getOperand(i).getReg(); |
771 | |
772 | assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure." ); |
773 | } |
774 | |
775 | /// Return the Phi register value that comes the loop block. |
776 | static unsigned getLoopPhiReg(const MachineInstr &Phi, |
777 | const MachineBasicBlock *LoopBB) { |
778 | for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) |
779 | if (Phi.getOperand(i: i + 1).getMBB() == LoopBB) |
780 | return Phi.getOperand(i).getReg(); |
781 | return 0; |
782 | } |
783 | |
784 | /// Return true if SUb can be reached from SUa following the chain edges. |
785 | static bool isSuccOrder(SUnit *SUa, SUnit *SUb) { |
786 | SmallPtrSet<SUnit *, 8> Visited; |
787 | SmallVector<SUnit *, 8> Worklist; |
788 | Worklist.push_back(Elt: SUa); |
789 | while (!Worklist.empty()) { |
790 | const SUnit *SU = Worklist.pop_back_val(); |
791 | for (const auto &SI : SU->Succs) { |
792 | SUnit *SuccSU = SI.getSUnit(); |
793 | if (SI.getKind() == SDep::Order) { |
794 | if (Visited.count(Ptr: SuccSU)) |
795 | continue; |
796 | if (SuccSU == SUb) |
797 | return true; |
798 | Worklist.push_back(Elt: SuccSU); |
799 | Visited.insert(Ptr: SuccSU); |
800 | } |
801 | } |
802 | } |
803 | return false; |
804 | } |
805 | |
806 | /// Return true if the instruction causes a chain between memory |
807 | /// references before and after it. |
808 | static bool isDependenceBarrier(MachineInstr &MI) { |
809 | return MI.isCall() || MI.mayRaiseFPException() || |
810 | MI.hasUnmodeledSideEffects() || |
811 | (MI.hasOrderedMemoryRef() && |
812 | (!MI.mayLoad() || !MI.isDereferenceableInvariantLoad())); |
813 | } |
814 | |
815 | /// Return the underlying objects for the memory references of an instruction. |
816 | /// This function calls the code in ValueTracking, but first checks that the |
817 | /// instruction has a memory operand. |
818 | static void getUnderlyingObjects(const MachineInstr *MI, |
819 | SmallVectorImpl<const Value *> &Objs) { |
820 | if (!MI->hasOneMemOperand()) |
821 | return; |
822 | MachineMemOperand *MM = *MI->memoperands_begin(); |
823 | if (!MM->getValue()) |
824 | return; |
825 | getUnderlyingObjects(V: MM->getValue(), Objects&: Objs); |
826 | for (const Value *V : Objs) { |
827 | if (!isIdentifiedObject(V)) { |
828 | Objs.clear(); |
829 | return; |
830 | } |
831 | } |
832 | } |
833 | |
834 | /// Add a chain edge between a load and store if the store can be an |
835 | /// alias of the load on a subsequent iteration, i.e., a loop carried |
836 | /// dependence. This code is very similar to the code in ScheduleDAGInstrs |
837 | /// but that code doesn't create loop carried dependences. |
838 | void SwingSchedulerDAG::addLoopCarriedDependences(AliasAnalysis *AA) { |
839 | MapVector<const Value *, SmallVector<SUnit *, 4>> PendingLoads; |
840 | Value *UnknownValue = |
841 | UndefValue::get(T: Type::getVoidTy(C&: MF.getFunction().getContext())); |
842 | for (auto &SU : SUnits) { |
843 | MachineInstr &MI = *SU.getInstr(); |
844 | if (isDependenceBarrier(MI)) |
845 | PendingLoads.clear(); |
846 | else if (MI.mayLoad()) { |
847 | SmallVector<const Value *, 4> Objs; |
848 | ::getUnderlyingObjects(MI: &MI, Objs); |
849 | if (Objs.empty()) |
850 | Objs.push_back(Elt: UnknownValue); |
851 | for (const auto *V : Objs) { |
852 | SmallVector<SUnit *, 4> &SUs = PendingLoads[V]; |
853 | SUs.push_back(Elt: &SU); |
854 | } |
855 | } else if (MI.mayStore()) { |
856 | SmallVector<const Value *, 4> Objs; |
857 | ::getUnderlyingObjects(MI: &MI, Objs); |
858 | if (Objs.empty()) |
859 | Objs.push_back(Elt: UnknownValue); |
860 | for (const auto *V : Objs) { |
861 | MapVector<const Value *, SmallVector<SUnit *, 4>>::iterator I = |
862 | PendingLoads.find(Key: V); |
863 | if (I == PendingLoads.end()) |
864 | continue; |
865 | for (auto *Load : I->second) { |
866 | if (isSuccOrder(SUa: Load, SUb: &SU)) |
867 | continue; |
868 | MachineInstr &LdMI = *Load->getInstr(); |
869 | // First, perform the cheaper check that compares the base register. |
870 | // If they are the same and the load offset is less than the store |
871 | // offset, then mark the dependence as loop carried potentially. |
872 | const MachineOperand *BaseOp1, *BaseOp2; |
873 | int64_t Offset1, Offset2; |
874 | bool Offset1IsScalable, Offset2IsScalable; |
875 | if (TII->getMemOperandWithOffset(MI: LdMI, BaseOp&: BaseOp1, Offset&: Offset1, |
876 | OffsetIsScalable&: Offset1IsScalable, TRI) && |
877 | TII->getMemOperandWithOffset(MI, BaseOp&: BaseOp2, Offset&: Offset2, |
878 | OffsetIsScalable&: Offset2IsScalable, TRI)) { |
879 | if (BaseOp1->isIdenticalTo(Other: *BaseOp2) && |
880 | Offset1IsScalable == Offset2IsScalable && |
881 | (int)Offset1 < (int)Offset2) { |
882 | assert(TII->areMemAccessesTriviallyDisjoint(LdMI, MI) && |
883 | "What happened to the chain edge?" ); |
884 | SDep Dep(Load, SDep::Barrier); |
885 | Dep.setLatency(1); |
886 | SU.addPred(D: Dep); |
887 | continue; |
888 | } |
889 | } |
890 | // Second, the more expensive check that uses alias analysis on the |
891 | // base registers. If they alias, and the load offset is less than |
892 | // the store offset, the mark the dependence as loop carried. |
893 | if (!AA) { |
894 | SDep Dep(Load, SDep::Barrier); |
895 | Dep.setLatency(1); |
896 | SU.addPred(D: Dep); |
897 | continue; |
898 | } |
899 | MachineMemOperand *MMO1 = *LdMI.memoperands_begin(); |
900 | MachineMemOperand *MMO2 = *MI.memoperands_begin(); |
901 | if (!MMO1->getValue() || !MMO2->getValue()) { |
902 | SDep Dep(Load, SDep::Barrier); |
903 | Dep.setLatency(1); |
904 | SU.addPred(D: Dep); |
905 | continue; |
906 | } |
907 | if (MMO1->getValue() == MMO2->getValue() && |
908 | MMO1->getOffset() <= MMO2->getOffset()) { |
909 | SDep Dep(Load, SDep::Barrier); |
910 | Dep.setLatency(1); |
911 | SU.addPred(D: Dep); |
912 | continue; |
913 | } |
914 | if (!AA->isNoAlias( |
915 | LocA: MemoryLocation::getAfter(Ptr: MMO1->getValue(), AATags: MMO1->getAAInfo()), |
916 | LocB: MemoryLocation::getAfter(Ptr: MMO2->getValue(), |
917 | AATags: MMO2->getAAInfo()))) { |
918 | SDep Dep(Load, SDep::Barrier); |
919 | Dep.setLatency(1); |
920 | SU.addPred(D: Dep); |
921 | } |
922 | } |
923 | } |
924 | } |
925 | } |
926 | } |
927 | |
928 | /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer |
929 | /// processes dependences for PHIs. This function adds true dependences |
930 | /// from a PHI to a use, and a loop carried dependence from the use to the |
931 | /// PHI. The loop carried dependence is represented as an anti dependence |
932 | /// edge. This function also removes chain dependences between unrelated |
933 | /// PHIs. |
934 | void SwingSchedulerDAG::updatePhiDependences() { |
935 | SmallVector<SDep, 4> RemoveDeps; |
936 | const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>(); |
937 | |
938 | // Iterate over each DAG node. |
939 | for (SUnit &I : SUnits) { |
940 | RemoveDeps.clear(); |
941 | // Set to true if the instruction has an operand defined by a Phi. |
942 | unsigned HasPhiUse = 0; |
943 | unsigned HasPhiDef = 0; |
944 | MachineInstr *MI = I.getInstr(); |
945 | // Iterate over each operand, and we process the definitions. |
946 | for (const MachineOperand &MO : MI->operands()) { |
947 | if (!MO.isReg()) |
948 | continue; |
949 | Register Reg = MO.getReg(); |
950 | if (MO.isDef()) { |
951 | // If the register is used by a Phi, then create an anti dependence. |
952 | for (MachineRegisterInfo::use_instr_iterator |
953 | UI = MRI.use_instr_begin(RegNo: Reg), |
954 | UE = MRI.use_instr_end(); |
955 | UI != UE; ++UI) { |
956 | MachineInstr *UseMI = &*UI; |
957 | SUnit *SU = getSUnit(MI: UseMI); |
958 | if (SU != nullptr && UseMI->isPHI()) { |
959 | if (!MI->isPHI()) { |
960 | SDep Dep(SU, SDep::Anti, Reg); |
961 | Dep.setLatency(1); |
962 | I.addPred(D: Dep); |
963 | } else { |
964 | HasPhiDef = Reg; |
965 | // Add a chain edge to a dependent Phi that isn't an existing |
966 | // predecessor. |
967 | if (SU->NodeNum < I.NodeNum && !I.isPred(N: SU)) |
968 | I.addPred(D: SDep(SU, SDep::Barrier)); |
969 | } |
970 | } |
971 | } |
972 | } else if (MO.isUse()) { |
973 | // If the register is defined by a Phi, then create a true dependence. |
974 | MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg); |
975 | if (DefMI == nullptr) |
976 | continue; |
977 | SUnit *SU = getSUnit(MI: DefMI); |
978 | if (SU != nullptr && DefMI->isPHI()) { |
979 | if (!MI->isPHI()) { |
980 | SDep Dep(SU, SDep::Data, Reg); |
981 | Dep.setLatency(0); |
982 | ST.adjustSchedDependency(Def: SU, DefOpIdx: 0, Use: &I, UseOpIdx: MO.getOperandNo(), Dep, |
983 | SchedModel: &SchedModel); |
984 | I.addPred(D: Dep); |
985 | } else { |
986 | HasPhiUse = Reg; |
987 | // Add a chain edge to a dependent Phi that isn't an existing |
988 | // predecessor. |
989 | if (SU->NodeNum < I.NodeNum && !I.isPred(N: SU)) |
990 | I.addPred(D: SDep(SU, SDep::Barrier)); |
991 | } |
992 | } |
993 | } |
994 | } |
995 | // Remove order dependences from an unrelated Phi. |
996 | if (!SwpPruneDeps) |
997 | continue; |
998 | for (auto &PI : I.Preds) { |
999 | MachineInstr *PMI = PI.getSUnit()->getInstr(); |
1000 | if (PMI->isPHI() && PI.getKind() == SDep::Order) { |
1001 | if (I.getInstr()->isPHI()) { |
1002 | if (PMI->getOperand(i: 0).getReg() == HasPhiUse) |
1003 | continue; |
1004 | if (getLoopPhiReg(Phi: *PMI, LoopBB: PMI->getParent()) == HasPhiDef) |
1005 | continue; |
1006 | } |
1007 | RemoveDeps.push_back(Elt: PI); |
1008 | } |
1009 | } |
1010 | for (const SDep &D : RemoveDeps) |
1011 | I.removePred(D); |
1012 | } |
1013 | } |
1014 | |
1015 | /// Iterate over each DAG node and see if we can change any dependences |
1016 | /// in order to reduce the recurrence MII. |
1017 | void SwingSchedulerDAG::changeDependences() { |
1018 | // See if an instruction can use a value from the previous iteration. |
1019 | // If so, we update the base and offset of the instruction and change |
1020 | // the dependences. |
1021 | for (SUnit &I : SUnits) { |
1022 | unsigned BasePos = 0, OffsetPos = 0, NewBase = 0; |
1023 | int64_t NewOffset = 0; |
1024 | if (!canUseLastOffsetValue(MI: I.getInstr(), BasePos, OffsetPos, NewBase, |
1025 | NewOffset)) |
1026 | continue; |
1027 | |
1028 | // Get the MI and SUnit for the instruction that defines the original base. |
1029 | Register OrigBase = I.getInstr()->getOperand(i: BasePos).getReg(); |
1030 | MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg: OrigBase); |
1031 | if (!DefMI) |
1032 | continue; |
1033 | SUnit *DefSU = getSUnit(MI: DefMI); |
1034 | if (!DefSU) |
1035 | continue; |
1036 | // Get the MI and SUnit for the instruction that defins the new base. |
1037 | MachineInstr *LastMI = MRI.getUniqueVRegDef(Reg: NewBase); |
1038 | if (!LastMI) |
1039 | continue; |
1040 | SUnit *LastSU = getSUnit(MI: LastMI); |
1041 | if (!LastSU) |
1042 | continue; |
1043 | |
1044 | if (Topo.IsReachable(SU: &I, TargetSU: LastSU)) |
1045 | continue; |
1046 | |
1047 | // Remove the dependence. The value now depends on a prior iteration. |
1048 | SmallVector<SDep, 4> Deps; |
1049 | for (const SDep &P : I.Preds) |
1050 | if (P.getSUnit() == DefSU) |
1051 | Deps.push_back(Elt: P); |
1052 | for (const SDep &D : Deps) { |
1053 | Topo.RemovePred(M: &I, N: D.getSUnit()); |
1054 | I.removePred(D); |
1055 | } |
1056 | // Remove the chain dependence between the instructions. |
1057 | Deps.clear(); |
1058 | for (auto &P : LastSU->Preds) |
1059 | if (P.getSUnit() == &I && P.getKind() == SDep::Order) |
1060 | Deps.push_back(Elt: P); |
1061 | for (const SDep &D : Deps) { |
1062 | Topo.RemovePred(M: LastSU, N: D.getSUnit()); |
1063 | LastSU->removePred(D); |
1064 | } |
1065 | |
1066 | // Add a dependence between the new instruction and the instruction |
1067 | // that defines the new base. |
1068 | SDep Dep(&I, SDep::Anti, NewBase); |
1069 | Topo.AddPred(Y: LastSU, X: &I); |
1070 | LastSU->addPred(D: Dep); |
1071 | |
1072 | // Remember the base and offset information so that we can update the |
1073 | // instruction during code generation. |
1074 | InstrChanges[&I] = std::make_pair(x&: NewBase, y&: NewOffset); |
1075 | } |
1076 | } |
1077 | |
1078 | /// Create an instruction stream that represents a single iteration and stage of |
1079 | /// each instruction. This function differs from SMSchedule::finalizeSchedule in |
1080 | /// that this doesn't have any side-effect to SwingSchedulerDAG. That is, this |
1081 | /// function is an approximation of SMSchedule::finalizeSchedule with all |
1082 | /// non-const operations removed. |
1083 | static void computeScheduledInsts(const SwingSchedulerDAG *SSD, |
1084 | SMSchedule &Schedule, |
1085 | std::vector<MachineInstr *> &OrderedInsts, |
1086 | DenseMap<MachineInstr *, unsigned> &Stages) { |
1087 | DenseMap<int, std::deque<SUnit *>> Instrs; |
1088 | |
1089 | // Move all instructions to the first stage from the later stages. |
1090 | for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); |
1091 | ++Cycle) { |
1092 | for (int Stage = 0, LastStage = Schedule.getMaxStageCount(); |
1093 | Stage <= LastStage; ++Stage) { |
1094 | for (SUnit *SU : llvm::reverse(C&: Schedule.getInstructions( |
1095 | cycle: Cycle + Stage * Schedule.getInitiationInterval()))) { |
1096 | Instrs[Cycle].push_front(x: SU); |
1097 | } |
1098 | } |
1099 | } |
1100 | |
1101 | for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); |
1102 | ++Cycle) { |
1103 | std::deque<SUnit *> &CycleInstrs = Instrs[Cycle]; |
1104 | CycleInstrs = Schedule.reorderInstructions(SSD, Instrs: CycleInstrs); |
1105 | for (SUnit *SU : CycleInstrs) { |
1106 | MachineInstr *MI = SU->getInstr(); |
1107 | OrderedInsts.push_back(x: MI); |
1108 | Stages[MI] = Schedule.stageScheduled(SU); |
1109 | } |
1110 | } |
1111 | } |
1112 | |
1113 | namespace { |
1114 | |
1115 | // FuncUnitSorter - Comparison operator used to sort instructions by |
1116 | // the number of functional unit choices. |
1117 | struct FuncUnitSorter { |
1118 | const InstrItineraryData *InstrItins; |
1119 | const MCSubtargetInfo *STI; |
1120 | DenseMap<InstrStage::FuncUnits, unsigned> Resources; |
1121 | |
1122 | FuncUnitSorter(const TargetSubtargetInfo &TSI) |
1123 | : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {} |
1124 | |
1125 | // Compute the number of functional unit alternatives needed |
1126 | // at each stage, and take the minimum value. We prioritize the |
1127 | // instructions by the least number of choices first. |
1128 | unsigned minFuncUnits(const MachineInstr *Inst, |
1129 | InstrStage::FuncUnits &F) const { |
1130 | unsigned SchedClass = Inst->getDesc().getSchedClass(); |
1131 | unsigned min = UINT_MAX; |
1132 | if (InstrItins && !InstrItins->isEmpty()) { |
1133 | for (const InstrStage &IS : |
1134 | make_range(x: InstrItins->beginStage(ItinClassIndx: SchedClass), |
1135 | y: InstrItins->endStage(ItinClassIndx: SchedClass))) { |
1136 | InstrStage::FuncUnits funcUnits = IS.getUnits(); |
1137 | unsigned numAlternatives = llvm::popcount(Value: funcUnits); |
1138 | if (numAlternatives < min) { |
1139 | min = numAlternatives; |
1140 | F = funcUnits; |
1141 | } |
1142 | } |
1143 | return min; |
1144 | } |
1145 | if (STI && STI->getSchedModel().hasInstrSchedModel()) { |
1146 | const MCSchedClassDesc *SCDesc = |
1147 | STI->getSchedModel().getSchedClassDesc(SchedClassIdx: SchedClass); |
1148 | if (!SCDesc->isValid()) |
1149 | // No valid Schedule Class Desc for schedClass, should be |
1150 | // Pseudo/PostRAPseudo |
1151 | return min; |
1152 | |
1153 | for (const MCWriteProcResEntry &PRE : |
1154 | make_range(x: STI->getWriteProcResBegin(SC: SCDesc), |
1155 | y: STI->getWriteProcResEnd(SC: SCDesc))) { |
1156 | if (!PRE.ReleaseAtCycle) |
1157 | continue; |
1158 | const MCProcResourceDesc *ProcResource = |
1159 | STI->getSchedModel().getProcResource(ProcResourceIdx: PRE.ProcResourceIdx); |
1160 | unsigned NumUnits = ProcResource->NumUnits; |
1161 | if (NumUnits < min) { |
1162 | min = NumUnits; |
1163 | F = PRE.ProcResourceIdx; |
1164 | } |
1165 | } |
1166 | return min; |
1167 | } |
1168 | llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!" ); |
1169 | } |
1170 | |
1171 | // Compute the critical resources needed by the instruction. This |
1172 | // function records the functional units needed by instructions that |
1173 | // must use only one functional unit. We use this as a tie breaker |
1174 | // for computing the resource MII. The instrutions that require |
1175 | // the same, highly used, functional unit have high priority. |
1176 | void calcCriticalResources(MachineInstr &MI) { |
1177 | unsigned SchedClass = MI.getDesc().getSchedClass(); |
1178 | if (InstrItins && !InstrItins->isEmpty()) { |
1179 | for (const InstrStage &IS : |
1180 | make_range(x: InstrItins->beginStage(ItinClassIndx: SchedClass), |
1181 | y: InstrItins->endStage(ItinClassIndx: SchedClass))) { |
1182 | InstrStage::FuncUnits FuncUnits = IS.getUnits(); |
1183 | if (llvm::popcount(Value: FuncUnits) == 1) |
1184 | Resources[FuncUnits]++; |
1185 | } |
1186 | return; |
1187 | } |
1188 | if (STI && STI->getSchedModel().hasInstrSchedModel()) { |
1189 | const MCSchedClassDesc *SCDesc = |
1190 | STI->getSchedModel().getSchedClassDesc(SchedClassIdx: SchedClass); |
1191 | if (!SCDesc->isValid()) |
1192 | // No valid Schedule Class Desc for schedClass, should be |
1193 | // Pseudo/PostRAPseudo |
1194 | return; |
1195 | |
1196 | for (const MCWriteProcResEntry &PRE : |
1197 | make_range(x: STI->getWriteProcResBegin(SC: SCDesc), |
1198 | y: STI->getWriteProcResEnd(SC: SCDesc))) { |
1199 | if (!PRE.ReleaseAtCycle) |
1200 | continue; |
1201 | Resources[PRE.ProcResourceIdx]++; |
1202 | } |
1203 | return; |
1204 | } |
1205 | llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!" ); |
1206 | } |
1207 | |
1208 | /// Return true if IS1 has less priority than IS2. |
1209 | bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const { |
1210 | InstrStage::FuncUnits F1 = 0, F2 = 0; |
1211 | unsigned MFUs1 = minFuncUnits(Inst: IS1, F&: F1); |
1212 | unsigned MFUs2 = minFuncUnits(Inst: IS2, F&: F2); |
1213 | if (MFUs1 == MFUs2) |
1214 | return Resources.lookup(Val: F1) < Resources.lookup(Val: F2); |
1215 | return MFUs1 > MFUs2; |
1216 | } |
1217 | }; |
1218 | |
1219 | /// Calculate the maximum register pressure of the scheduled instructions stream |
1220 | class HighRegisterPressureDetector { |
1221 | MachineBasicBlock *OrigMBB; |
1222 | const MachineFunction &MF; |
1223 | const MachineRegisterInfo &MRI; |
1224 | const TargetRegisterInfo *TRI; |
1225 | |
1226 | const unsigned PSetNum; |
1227 | |
1228 | // Indexed by PSet ID |
1229 | // InitSetPressure takes into account the register pressure of live-in |
1230 | // registers. It's not depend on how the loop is scheduled, so it's enough to |
1231 | // calculate them once at the beginning. |
1232 | std::vector<unsigned> InitSetPressure; |
1233 | |
1234 | // Indexed by PSet ID |
1235 | // Upper limit for each register pressure set |
1236 | std::vector<unsigned> PressureSetLimit; |
1237 | |
1238 | DenseMap<MachineInstr *, RegisterOperands> ROMap; |
1239 | |
1240 | using Instr2LastUsesTy = DenseMap<MachineInstr *, SmallDenseSet<Register, 4>>; |
1241 | |
1242 | public: |
1243 | using OrderedInstsTy = std::vector<MachineInstr *>; |
1244 | using Instr2StageTy = DenseMap<MachineInstr *, unsigned>; |
1245 | |
1246 | private: |
1247 | static void dumpRegisterPressures(const std::vector<unsigned> &Pressures) { |
1248 | if (Pressures.size() == 0) { |
1249 | dbgs() << "[]" ; |
1250 | } else { |
1251 | char Prefix = '['; |
1252 | for (unsigned P : Pressures) { |
1253 | dbgs() << Prefix << P; |
1254 | Prefix = ' '; |
1255 | } |
1256 | dbgs() << ']'; |
1257 | } |
1258 | } |
1259 | |
1260 | void dumpPSet(Register Reg) const { |
1261 | dbgs() << "Reg=" << printReg(Reg, TRI, SubIdx: 0, MRI: &MRI) << " PSet=" ; |
1262 | for (auto PSetIter = MRI.getPressureSets(RegUnit: Reg); PSetIter.isValid(); |
1263 | ++PSetIter) { |
1264 | dbgs() << *PSetIter << ' '; |
1265 | } |
1266 | dbgs() << '\n'; |
1267 | } |
1268 | |
1269 | void increaseRegisterPressure(std::vector<unsigned> &Pressure, |
1270 | Register Reg) const { |
1271 | auto PSetIter = MRI.getPressureSets(RegUnit: Reg); |
1272 | unsigned Weight = PSetIter.getWeight(); |
1273 | for (; PSetIter.isValid(); ++PSetIter) |
1274 | Pressure[*PSetIter] += Weight; |
1275 | } |
1276 | |
1277 | void decreaseRegisterPressure(std::vector<unsigned> &Pressure, |
1278 | Register Reg) const { |
1279 | auto PSetIter = MRI.getPressureSets(RegUnit: Reg); |
1280 | unsigned Weight = PSetIter.getWeight(); |
1281 | for (; PSetIter.isValid(); ++PSetIter) { |
1282 | auto &P = Pressure[*PSetIter]; |
1283 | assert(P >= Weight && |
1284 | "register pressure must be greater than or equal weight" ); |
1285 | P -= Weight; |
1286 | } |
1287 | } |
1288 | |
1289 | // Return true if Reg is fixed one, for example, stack pointer |
1290 | bool isFixedRegister(Register Reg) const { |
1291 | return Reg.isPhysical() && TRI->isFixedRegister(MF, PhysReg: Reg.asMCReg()); |
1292 | } |
1293 | |
1294 | bool isDefinedInThisLoop(Register Reg) const { |
1295 | return Reg.isVirtual() && MRI.getVRegDef(Reg)->getParent() == OrigMBB; |
1296 | } |
1297 | |
1298 | // Search for live-in variables. They are factored into the register pressure |
1299 | // from the begining. Live-in variables used by every iteration should be |
1300 | // considered as alive throughout the loop. For example, the variable `c` in |
1301 | // following code. \code |
1302 | // int c = ...; |
1303 | // for (int i = 0; i < n; i++) |
1304 | // a[i] += b[i] + c; |
1305 | // \endcode |
1306 | void computeLiveIn() { |
1307 | DenseSet<Register> Used; |
1308 | for (auto &MI : *OrigMBB) { |
1309 | if (MI.isDebugInstr()) |
1310 | continue; |
1311 | for (auto &Use : ROMap[&MI].Uses) { |
1312 | auto Reg = Use.RegUnit; |
1313 | // Ignore the variable that appears only on one side of phi instruction |
1314 | // because it's used only at the first iteration. |
1315 | if (MI.isPHI() && Reg != getLoopPhiReg(Phi: MI, LoopBB: OrigMBB)) |
1316 | continue; |
1317 | if (isFixedRegister(Reg)) |
1318 | continue; |
1319 | if (isDefinedInThisLoop(Reg)) |
1320 | continue; |
1321 | Used.insert(V: Reg); |
1322 | } |
1323 | } |
1324 | |
1325 | for (auto LiveIn : Used) |
1326 | increaseRegisterPressure(Pressure&: InitSetPressure, Reg: LiveIn); |
1327 | } |
1328 | |
1329 | // Calculate the upper limit of each pressure set |
1330 | void computePressureSetLimit(const RegisterClassInfo &RCI) { |
1331 | for (unsigned PSet = 0; PSet < PSetNum; PSet++) |
1332 | PressureSetLimit[PSet] = TRI->getRegPressureSetLimit(MF, Idx: PSet); |
1333 | |
1334 | // We assume fixed registers, such as stack pointer, are already in use. |
1335 | // Therefore subtracting the weight of the fixed registers from the limit of |
1336 | // each pressure set in advance. |
1337 | SmallDenseSet<Register, 8> FixedRegs; |
1338 | for (const TargetRegisterClass *TRC : TRI->regclasses()) { |
1339 | for (const MCPhysReg Reg : *TRC) |
1340 | if (isFixedRegister(Reg)) |
1341 | FixedRegs.insert(V: Reg); |
1342 | } |
1343 | |
1344 | LLVM_DEBUG({ |
1345 | for (auto Reg : FixedRegs) { |
1346 | dbgs() << printReg(Reg, TRI, 0, &MRI) << ": [" ; |
1347 | const int *Sets = TRI->getRegUnitPressureSets(Reg); |
1348 | for (; *Sets != -1; Sets++) { |
1349 | dbgs() << TRI->getRegPressureSetName(*Sets) << ", " ; |
1350 | } |
1351 | dbgs() << "]\n" ; |
1352 | } |
1353 | }); |
1354 | |
1355 | for (auto Reg : FixedRegs) { |
1356 | LLVM_DEBUG(dbgs() << "fixed register: " << printReg(Reg, TRI, 0, &MRI) |
1357 | << "\n" ); |
1358 | auto PSetIter = MRI.getPressureSets(RegUnit: Reg); |
1359 | unsigned Weight = PSetIter.getWeight(); |
1360 | for (; PSetIter.isValid(); ++PSetIter) { |
1361 | unsigned &Limit = PressureSetLimit[*PSetIter]; |
1362 | assert(Limit >= Weight && |
1363 | "register pressure limit must be greater than or equal weight" ); |
1364 | Limit -= Weight; |
1365 | LLVM_DEBUG(dbgs() << "PSet=" << *PSetIter << " Limit=" << Limit |
1366 | << " (decreased by " << Weight << ")\n" ); |
1367 | } |
1368 | } |
1369 | } |
1370 | |
1371 | // There are two patterns of last-use. |
1372 | // - by an instruction of the current iteration |
1373 | // - by a phi instruction of the next iteration (loop carried value) |
1374 | // |
1375 | // Furthermore, following two groups of instructions are executed |
1376 | // simultaneously |
1377 | // - next iteration's phi instructions in i-th stage |
1378 | // - current iteration's instructions in i+1-th stage |
1379 | // |
1380 | // This function calculates the last-use of each register while taking into |
1381 | // account the above two patterns. |
1382 | Instr2LastUsesTy computeLastUses(const OrderedInstsTy &OrderedInsts, |
1383 | Instr2StageTy &Stages) const { |
1384 | // We treat virtual registers that are defined and used in this loop. |
1385 | // Following virtual register will be ignored |
1386 | // - live-in one |
1387 | // - defined but not used in the loop (potentially live-out) |
1388 | DenseSet<Register> TargetRegs; |
1389 | const auto UpdateTargetRegs = [this, &TargetRegs](Register Reg) { |
1390 | if (isDefinedInThisLoop(Reg)) |
1391 | TargetRegs.insert(V: Reg); |
1392 | }; |
1393 | for (MachineInstr *MI : OrderedInsts) { |
1394 | if (MI->isPHI()) { |
1395 | Register Reg = getLoopPhiReg(Phi: *MI, LoopBB: OrigMBB); |
1396 | UpdateTargetRegs(Reg); |
1397 | } else { |
1398 | for (auto &Use : ROMap.find(Val: MI)->getSecond().Uses) |
1399 | UpdateTargetRegs(Use.RegUnit); |
1400 | } |
1401 | } |
1402 | |
1403 | const auto InstrScore = [&Stages](MachineInstr *MI) { |
1404 | return Stages[MI] + MI->isPHI(); |
1405 | }; |
1406 | |
1407 | DenseMap<Register, MachineInstr *> LastUseMI; |
1408 | for (MachineInstr *MI : llvm::reverse(C: OrderedInsts)) { |
1409 | for (auto &Use : ROMap.find(Val: MI)->getSecond().Uses) { |
1410 | auto Reg = Use.RegUnit; |
1411 | if (!TargetRegs.contains(V: Reg)) |
1412 | continue; |
1413 | auto Ite = LastUseMI.find(Val: Reg); |
1414 | if (Ite == LastUseMI.end()) { |
1415 | LastUseMI[Reg] = MI; |
1416 | } else { |
1417 | MachineInstr *Orig = Ite->second; |
1418 | MachineInstr *New = MI; |
1419 | if (InstrScore(Orig) < InstrScore(New)) |
1420 | LastUseMI[Reg] = New; |
1421 | } |
1422 | } |
1423 | } |
1424 | |
1425 | Instr2LastUsesTy LastUses; |
1426 | for (auto &Entry : LastUseMI) |
1427 | LastUses[Entry.second].insert(V: Entry.first); |
1428 | return LastUses; |
1429 | } |
1430 | |
1431 | // Compute the maximum register pressure of the kernel. We'll simulate #Stage |
1432 | // iterations and check the register pressure at the point where all stages |
1433 | // overlapping. |
1434 | // |
1435 | // An example of unrolled loop where #Stage is 4.. |
1436 | // Iter i+0 i+1 i+2 i+3 |
1437 | // ------------------------ |
1438 | // Stage 0 |
1439 | // Stage 1 0 |
1440 | // Stage 2 1 0 |
1441 | // Stage 3 2 1 0 <- All stages overlap |
1442 | // |
1443 | std::vector<unsigned> |
1444 | computeMaxSetPressure(const OrderedInstsTy &OrderedInsts, |
1445 | Instr2StageTy &Stages, |
1446 | const unsigned StageCount) const { |
1447 | using RegSetTy = SmallDenseSet<Register, 16>; |
1448 | |
1449 | // Indexed by #Iter. To treat "local" variables of each stage separately, we |
1450 | // manage the liveness of the registers independently by iterations. |
1451 | SmallVector<RegSetTy> LiveRegSets(StageCount); |
1452 | |
1453 | auto CurSetPressure = InitSetPressure; |
1454 | auto MaxSetPressure = InitSetPressure; |
1455 | auto LastUses = computeLastUses(OrderedInsts, Stages); |
1456 | |
1457 | LLVM_DEBUG({ |
1458 | dbgs() << "Ordered instructions:\n" ; |
1459 | for (MachineInstr *MI : OrderedInsts) { |
1460 | dbgs() << "Stage " << Stages[MI] << ": " ; |
1461 | MI->dump(); |
1462 | } |
1463 | }); |
1464 | |
1465 | const auto InsertReg = [this, &CurSetPressure](RegSetTy &RegSet, |
1466 | Register Reg) { |
1467 | if (!Reg.isValid() || isFixedRegister(Reg)) |
1468 | return; |
1469 | |
1470 | bool Inserted = RegSet.insert(V: Reg).second; |
1471 | if (!Inserted) |
1472 | return; |
1473 | |
1474 | LLVM_DEBUG(dbgs() << "insert " << printReg(Reg, TRI, 0, &MRI) << "\n" ); |
1475 | increaseRegisterPressure(Pressure&: CurSetPressure, Reg); |
1476 | LLVM_DEBUG(dumpPSet(Reg)); |
1477 | }; |
1478 | |
1479 | const auto EraseReg = [this, &CurSetPressure](RegSetTy &RegSet, |
1480 | Register Reg) { |
1481 | if (!Reg.isValid() || isFixedRegister(Reg)) |
1482 | return; |
1483 | |
1484 | // live-in register |
1485 | if (!RegSet.contains(V: Reg)) |
1486 | return; |
1487 | |
1488 | LLVM_DEBUG(dbgs() << "erase " << printReg(Reg, TRI, 0, &MRI) << "\n" ); |
1489 | RegSet.erase(V: Reg); |
1490 | decreaseRegisterPressure(Pressure&: CurSetPressure, Reg); |
1491 | LLVM_DEBUG(dumpPSet(Reg)); |
1492 | }; |
1493 | |
1494 | for (unsigned I = 0; I < StageCount; I++) { |
1495 | for (MachineInstr *MI : OrderedInsts) { |
1496 | const auto Stage = Stages[MI]; |
1497 | if (I < Stage) |
1498 | continue; |
1499 | |
1500 | const unsigned Iter = I - Stage; |
1501 | |
1502 | for (auto &Def : ROMap.find(Val: MI)->getSecond().Defs) |
1503 | InsertReg(LiveRegSets[Iter], Def.RegUnit); |
1504 | |
1505 | for (auto LastUse : LastUses[MI]) { |
1506 | if (MI->isPHI()) { |
1507 | if (Iter != 0) |
1508 | EraseReg(LiveRegSets[Iter - 1], LastUse); |
1509 | } else { |
1510 | EraseReg(LiveRegSets[Iter], LastUse); |
1511 | } |
1512 | } |
1513 | |
1514 | for (unsigned PSet = 0; PSet < PSetNum; PSet++) |
1515 | MaxSetPressure[PSet] = |
1516 | std::max(a: MaxSetPressure[PSet], b: CurSetPressure[PSet]); |
1517 | |
1518 | LLVM_DEBUG({ |
1519 | dbgs() << "CurSetPressure=" ; |
1520 | dumpRegisterPressures(CurSetPressure); |
1521 | dbgs() << " iter=" << Iter << " stage=" << Stage << ":" ; |
1522 | MI->dump(); |
1523 | }); |
1524 | } |
1525 | } |
1526 | |
1527 | return MaxSetPressure; |
1528 | } |
1529 | |
1530 | public: |
1531 | HighRegisterPressureDetector(MachineBasicBlock *OrigMBB, |
1532 | const MachineFunction &MF) |
1533 | : OrigMBB(OrigMBB), MF(MF), MRI(MF.getRegInfo()), |
1534 | TRI(MF.getSubtarget().getRegisterInfo()), |
1535 | PSetNum(TRI->getNumRegPressureSets()), InitSetPressure(PSetNum, 0), |
1536 | PressureSetLimit(PSetNum, 0) {} |
1537 | |
1538 | // Used to calculate register pressure, which is independent of loop |
1539 | // scheduling. |
1540 | void init(const RegisterClassInfo &RCI) { |
1541 | for (MachineInstr &MI : *OrigMBB) { |
1542 | if (MI.isDebugInstr()) |
1543 | continue; |
1544 | ROMap[&MI].collect(MI, TRI: *TRI, MRI, TrackLaneMasks: false, IgnoreDead: true); |
1545 | } |
1546 | |
1547 | computeLiveIn(); |
1548 | computePressureSetLimit(RCI); |
1549 | } |
1550 | |
1551 | // Calculate the maximum register pressures of the loop and check if they |
1552 | // exceed the limit |
1553 | bool detect(const SwingSchedulerDAG *SSD, SMSchedule &Schedule, |
1554 | const unsigned MaxStage) const { |
1555 | assert(0 <= RegPressureMargin && RegPressureMargin <= 100 && |
1556 | "the percentage of the margin must be between 0 to 100" ); |
1557 | |
1558 | OrderedInstsTy OrderedInsts; |
1559 | Instr2StageTy Stages; |
1560 | computeScheduledInsts(SSD, Schedule, OrderedInsts, Stages); |
1561 | const auto MaxSetPressure = |
1562 | computeMaxSetPressure(OrderedInsts, Stages, StageCount: MaxStage + 1); |
1563 | |
1564 | LLVM_DEBUG({ |
1565 | dbgs() << "Dump MaxSetPressure:\n" ; |
1566 | for (unsigned I = 0; I < MaxSetPressure.size(); I++) { |
1567 | dbgs() << format("MaxSetPressure[%d]=%d\n" , I, MaxSetPressure[I]); |
1568 | } |
1569 | dbgs() << '\n'; |
1570 | }); |
1571 | |
1572 | for (unsigned PSet = 0; PSet < PSetNum; PSet++) { |
1573 | unsigned Limit = PressureSetLimit[PSet]; |
1574 | unsigned Margin = Limit * RegPressureMargin / 100; |
1575 | LLVM_DEBUG(dbgs() << "PSet=" << PSet << " Limit=" << Limit |
1576 | << " Margin=" << Margin << "\n" ); |
1577 | if (Limit < MaxSetPressure[PSet] + Margin) { |
1578 | LLVM_DEBUG( |
1579 | dbgs() |
1580 | << "Rejected the schedule because of too high register pressure\n" ); |
1581 | return true; |
1582 | } |
1583 | } |
1584 | return false; |
1585 | } |
1586 | }; |
1587 | |
1588 | } // end anonymous namespace |
1589 | |
1590 | /// Calculate the resource constrained minimum initiation interval for the |
1591 | /// specified loop. We use the DFA to model the resources needed for |
1592 | /// each instruction, and we ignore dependences. A different DFA is created |
1593 | /// for each cycle that is required. When adding a new instruction, we attempt |
1594 | /// to add it to each existing DFA, until a legal space is found. If the |
1595 | /// instruction cannot be reserved in an existing DFA, we create a new one. |
1596 | unsigned SwingSchedulerDAG::calculateResMII() { |
1597 | LLVM_DEBUG(dbgs() << "calculateResMII:\n" ); |
1598 | ResourceManager RM(&MF.getSubtarget(), this); |
1599 | return RM.calculateResMII(); |
1600 | } |
1601 | |
1602 | /// Calculate the recurrence-constrainted minimum initiation interval. |
1603 | /// Iterate over each circuit. Compute the delay(c) and distance(c) |
1604 | /// for each circuit. The II needs to satisfy the inequality |
1605 | /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest |
1606 | /// II that satisfies the inequality, and the RecMII is the maximum |
1607 | /// of those values. |
1608 | unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) { |
1609 | unsigned RecMII = 0; |
1610 | |
1611 | for (NodeSet &Nodes : NodeSets) { |
1612 | if (Nodes.empty()) |
1613 | continue; |
1614 | |
1615 | unsigned Delay = Nodes.getLatency(); |
1616 | unsigned Distance = 1; |
1617 | |
1618 | // ii = ceil(delay / distance) |
1619 | unsigned CurMII = (Delay + Distance - 1) / Distance; |
1620 | Nodes.setRecMII(CurMII); |
1621 | if (CurMII > RecMII) |
1622 | RecMII = CurMII; |
1623 | } |
1624 | |
1625 | return RecMII; |
1626 | } |
1627 | |
1628 | /// Swap all the anti dependences in the DAG. That means it is no longer a DAG, |
1629 | /// but we do this to find the circuits, and then change them back. |
1630 | static void swapAntiDependences(std::vector<SUnit> &SUnits) { |
1631 | SmallVector<std::pair<SUnit *, SDep>, 8> DepsAdded; |
1632 | for (SUnit &SU : SUnits) { |
1633 | for (SDep &Pred : SU.Preds) |
1634 | if (Pred.getKind() == SDep::Anti) |
1635 | DepsAdded.push_back(Elt: std::make_pair(x: &SU, y&: Pred)); |
1636 | } |
1637 | for (std::pair<SUnit *, SDep> &P : DepsAdded) { |
1638 | // Remove this anti dependency and add one in the reverse direction. |
1639 | SUnit *SU = P.first; |
1640 | SDep &D = P.second; |
1641 | SUnit *TargetSU = D.getSUnit(); |
1642 | unsigned Reg = D.getReg(); |
1643 | unsigned Lat = D.getLatency(); |
1644 | SU->removePred(D); |
1645 | SDep Dep(SU, SDep::Anti, Reg); |
1646 | Dep.setLatency(Lat); |
1647 | TargetSU->addPred(D: Dep); |
1648 | } |
1649 | } |
1650 | |
1651 | /// Create the adjacency structure of the nodes in the graph. |
1652 | void SwingSchedulerDAG::Circuits::createAdjacencyStructure( |
1653 | SwingSchedulerDAG *DAG) { |
1654 | BitVector Added(SUnits.size()); |
1655 | DenseMap<int, int> OutputDeps; |
1656 | for (int i = 0, e = SUnits.size(); i != e; ++i) { |
1657 | Added.reset(); |
1658 | // Add any successor to the adjacency matrix and exclude duplicates. |
1659 | for (auto &SI : SUnits[i].Succs) { |
1660 | // Only create a back-edge on the first and last nodes of a dependence |
1661 | // chain. This records any chains and adds them later. |
1662 | if (SI.getKind() == SDep::Output) { |
1663 | int N = SI.getSUnit()->NodeNum; |
1664 | int BackEdge = i; |
1665 | auto Dep = OutputDeps.find(Val: BackEdge); |
1666 | if (Dep != OutputDeps.end()) { |
1667 | BackEdge = Dep->second; |
1668 | OutputDeps.erase(I: Dep); |
1669 | } |
1670 | OutputDeps[N] = BackEdge; |
1671 | } |
1672 | // Do not process a boundary node, an artificial node. |
1673 | // A back-edge is processed only if it goes to a Phi. |
1674 | if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() || |
1675 | (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI())) |
1676 | continue; |
1677 | int N = SI.getSUnit()->NodeNum; |
1678 | if (!Added.test(Idx: N)) { |
1679 | AdjK[i].push_back(Elt: N); |
1680 | Added.set(N); |
1681 | } |
1682 | } |
1683 | // A chain edge between a store and a load is treated as a back-edge in the |
1684 | // adjacency matrix. |
1685 | for (auto &PI : SUnits[i].Preds) { |
1686 | if (!SUnits[i].getInstr()->mayStore() || |
1687 | !DAG->isLoopCarriedDep(Source: &SUnits[i], Dep: PI, isSucc: false)) |
1688 | continue; |
1689 | if (PI.getKind() == SDep::Order && PI.getSUnit()->getInstr()->mayLoad()) { |
1690 | int N = PI.getSUnit()->NodeNum; |
1691 | if (!Added.test(Idx: N)) { |
1692 | AdjK[i].push_back(Elt: N); |
1693 | Added.set(N); |
1694 | } |
1695 | } |
1696 | } |
1697 | } |
1698 | // Add back-edges in the adjacency matrix for the output dependences. |
1699 | for (auto &OD : OutputDeps) |
1700 | if (!Added.test(Idx: OD.second)) { |
1701 | AdjK[OD.first].push_back(Elt: OD.second); |
1702 | Added.set(OD.second); |
1703 | } |
1704 | } |
1705 | |
1706 | /// Identify an elementary circuit in the dependence graph starting at the |
1707 | /// specified node. |
1708 | bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets, |
1709 | bool HasBackedge) { |
1710 | SUnit *SV = &SUnits[V]; |
1711 | bool F = false; |
1712 | Stack.insert(X: SV); |
1713 | Blocked.set(V); |
1714 | |
1715 | for (auto W : AdjK[V]) { |
1716 | if (NumPaths > MaxPaths) |
1717 | break; |
1718 | if (W < S) |
1719 | continue; |
1720 | if (W == S) { |
1721 | if (!HasBackedge) |
1722 | NodeSets.push_back(Elt: NodeSet(Stack.begin(), Stack.end())); |
1723 | F = true; |
1724 | ++NumPaths; |
1725 | break; |
1726 | } else if (!Blocked.test(Idx: W)) { |
1727 | if (circuit(V: W, S, NodeSets, |
1728 | HasBackedge: Node2Idx->at(n: W) < Node2Idx->at(n: V) ? true : HasBackedge)) |
1729 | F = true; |
1730 | } |
1731 | } |
1732 | |
1733 | if (F) |
1734 | unblock(U: V); |
1735 | else { |
1736 | for (auto W : AdjK[V]) { |
1737 | if (W < S) |
1738 | continue; |
1739 | B[W].insert(Ptr: SV); |
1740 | } |
1741 | } |
1742 | Stack.pop_back(); |
1743 | return F; |
1744 | } |
1745 | |
1746 | /// Unblock a node in the circuit finding algorithm. |
1747 | void SwingSchedulerDAG::Circuits::unblock(int U) { |
1748 | Blocked.reset(Idx: U); |
1749 | SmallPtrSet<SUnit *, 4> &BU = B[U]; |
1750 | while (!BU.empty()) { |
1751 | SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin(); |
1752 | assert(SI != BU.end() && "Invalid B set." ); |
1753 | SUnit *W = *SI; |
1754 | BU.erase(Ptr: W); |
1755 | if (Blocked.test(Idx: W->NodeNum)) |
1756 | unblock(U: W->NodeNum); |
1757 | } |
1758 | } |
1759 | |
1760 | /// Identify all the elementary circuits in the dependence graph using |
1761 | /// Johnson's circuit algorithm. |
1762 | void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) { |
1763 | // Swap all the anti dependences in the DAG. That means it is no longer a DAG, |
1764 | // but we do this to find the circuits, and then change them back. |
1765 | swapAntiDependences(SUnits); |
1766 | |
1767 | Circuits Cir(SUnits, Topo); |
1768 | // Create the adjacency structure. |
1769 | Cir.createAdjacencyStructure(DAG: this); |
1770 | for (int i = 0, e = SUnits.size(); i != e; ++i) { |
1771 | Cir.reset(); |
1772 | Cir.circuit(V: i, S: i, NodeSets); |
1773 | } |
1774 | |
1775 | // Change the dependences back so that we've created a DAG again. |
1776 | swapAntiDependences(SUnits); |
1777 | } |
1778 | |
1779 | // Create artificial dependencies between the source of COPY/REG_SEQUENCE that |
1780 | // is loop-carried to the USE in next iteration. This will help pipeliner avoid |
1781 | // additional copies that are needed across iterations. An artificial dependence |
1782 | // edge is added from USE to SOURCE of COPY/REG_SEQUENCE. |
1783 | |
1784 | // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried) |
1785 | // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE |
1786 | // PHI-------True-Dep------> USEOfPhi |
1787 | |
1788 | // The mutation creates |
1789 | // USEOfPHI -------Artificial-Dep---> SRCOfCopy |
1790 | |
1791 | // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy |
1792 | // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled |
1793 | // late to avoid additional copies across iterations. The possible scheduling |
1794 | // order would be |
1795 | // USEOfPHI --- SRCOfCopy--- COPY/REG_SEQUENCE. |
1796 | |
1797 | void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) { |
1798 | for (SUnit &SU : DAG->SUnits) { |
1799 | // Find the COPY/REG_SEQUENCE instruction. |
1800 | if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence()) |
1801 | continue; |
1802 | |
1803 | // Record the loop carried PHIs. |
1804 | SmallVector<SUnit *, 4> PHISUs; |
1805 | // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions. |
1806 | SmallVector<SUnit *, 4> SrcSUs; |
1807 | |
1808 | for (auto &Dep : SU.Preds) { |
1809 | SUnit *TmpSU = Dep.getSUnit(); |
1810 | MachineInstr *TmpMI = TmpSU->getInstr(); |
1811 | SDep::Kind DepKind = Dep.getKind(); |
1812 | // Save the loop carried PHI. |
1813 | if (DepKind == SDep::Anti && TmpMI->isPHI()) |
1814 | PHISUs.push_back(Elt: TmpSU); |
1815 | // Save the source of COPY/REG_SEQUENCE. |
1816 | // If the source has no pre-decessors, we will end up creating cycles. |
1817 | else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0) |
1818 | SrcSUs.push_back(Elt: TmpSU); |
1819 | } |
1820 | |
1821 | if (PHISUs.size() == 0 || SrcSUs.size() == 0) |
1822 | continue; |
1823 | |
1824 | // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this |
1825 | // SUnit to the container. |
1826 | SmallVector<SUnit *, 8> UseSUs; |
1827 | // Do not use iterator based loop here as we are updating the container. |
1828 | for (size_t Index = 0; Index < PHISUs.size(); ++Index) { |
1829 | for (auto &Dep : PHISUs[Index]->Succs) { |
1830 | if (Dep.getKind() != SDep::Data) |
1831 | continue; |
1832 | |
1833 | SUnit *TmpSU = Dep.getSUnit(); |
1834 | MachineInstr *TmpMI = TmpSU->getInstr(); |
1835 | if (TmpMI->isPHI() || TmpMI->isRegSequence()) { |
1836 | PHISUs.push_back(Elt: TmpSU); |
1837 | continue; |
1838 | } |
1839 | UseSUs.push_back(Elt: TmpSU); |
1840 | } |
1841 | } |
1842 | |
1843 | if (UseSUs.size() == 0) |
1844 | continue; |
1845 | |
1846 | SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(Val: DAG); |
1847 | // Add the artificial dependencies if it does not form a cycle. |
1848 | for (auto *I : UseSUs) { |
1849 | for (auto *Src : SrcSUs) { |
1850 | if (!SDAG->Topo.IsReachable(SU: I, TargetSU: Src) && Src != I) { |
1851 | Src->addPred(D: SDep(I, SDep::Artificial)); |
1852 | SDAG->Topo.AddPred(Y: Src, X: I); |
1853 | } |
1854 | } |
1855 | } |
1856 | } |
1857 | } |
1858 | |
1859 | /// Return true for DAG nodes that we ignore when computing the cost functions. |
1860 | /// We ignore the back-edge recurrence in order to avoid unbounded recursion |
1861 | /// in the calculation of the ASAP, ALAP, etc functions. |
1862 | static bool ignoreDependence(const SDep &D, bool isPred) { |
1863 | if (D.isArtificial() || D.getSUnit()->isBoundaryNode()) |
1864 | return true; |
1865 | return D.getKind() == SDep::Anti && isPred; |
1866 | } |
1867 | |
1868 | /// Compute several functions need to order the nodes for scheduling. |
1869 | /// ASAP - Earliest time to schedule a node. |
1870 | /// ALAP - Latest time to schedule a node. |
1871 | /// MOV - Mobility function, difference between ALAP and ASAP. |
1872 | /// D - Depth of each node. |
1873 | /// H - Height of each node. |
1874 | void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) { |
1875 | ScheduleInfo.resize(new_size: SUnits.size()); |
1876 | |
1877 | LLVM_DEBUG({ |
1878 | for (int I : Topo) { |
1879 | const SUnit &SU = SUnits[I]; |
1880 | dumpNode(SU); |
1881 | } |
1882 | }); |
1883 | |
1884 | int maxASAP = 0; |
1885 | // Compute ASAP and ZeroLatencyDepth. |
1886 | for (int I : Topo) { |
1887 | int asap = 0; |
1888 | int zeroLatencyDepth = 0; |
1889 | SUnit *SU = &SUnits[I]; |
1890 | for (const SDep &P : SU->Preds) { |
1891 | SUnit *pred = P.getSUnit(); |
1892 | if (P.getLatency() == 0) |
1893 | zeroLatencyDepth = |
1894 | std::max(a: zeroLatencyDepth, b: getZeroLatencyDepth(Node: pred) + 1); |
1895 | if (ignoreDependence(D: P, isPred: true)) |
1896 | continue; |
1897 | asap = std::max(a: asap, b: (int)(getASAP(Node: pred) + P.getLatency() - |
1898 | getDistance(U: pred, V: SU, Dep: P) * MII)); |
1899 | } |
1900 | maxASAP = std::max(a: maxASAP, b: asap); |
1901 | ScheduleInfo[I].ASAP = asap; |
1902 | ScheduleInfo[I].ZeroLatencyDepth = zeroLatencyDepth; |
1903 | } |
1904 | |
1905 | // Compute ALAP, ZeroLatencyHeight, and MOV. |
1906 | for (int I : llvm::reverse(C&: Topo)) { |
1907 | int alap = maxASAP; |
1908 | int zeroLatencyHeight = 0; |
1909 | SUnit *SU = &SUnits[I]; |
1910 | for (const SDep &S : SU->Succs) { |
1911 | SUnit *succ = S.getSUnit(); |
1912 | if (succ->isBoundaryNode()) |
1913 | continue; |
1914 | if (S.getLatency() == 0) |
1915 | zeroLatencyHeight = |
1916 | std::max(a: zeroLatencyHeight, b: getZeroLatencyHeight(Node: succ) + 1); |
1917 | if (ignoreDependence(D: S, isPred: true)) |
1918 | continue; |
1919 | alap = std::min(a: alap, b: (int)(getALAP(Node: succ) - S.getLatency() + |
1920 | getDistance(U: SU, V: succ, Dep: S) * MII)); |
1921 | } |
1922 | |
1923 | ScheduleInfo[I].ALAP = alap; |
1924 | ScheduleInfo[I].ZeroLatencyHeight = zeroLatencyHeight; |
1925 | } |
1926 | |
1927 | // After computing the node functions, compute the summary for each node set. |
1928 | for (NodeSet &I : NodeSets) |
1929 | I.computeNodeSetInfo(SSD: this); |
1930 | |
1931 | LLVM_DEBUG({ |
1932 | for (unsigned i = 0; i < SUnits.size(); i++) { |
1933 | dbgs() << "\tNode " << i << ":\n" ; |
1934 | dbgs() << "\t ASAP = " << getASAP(&SUnits[i]) << "\n" ; |
1935 | dbgs() << "\t ALAP = " << getALAP(&SUnits[i]) << "\n" ; |
1936 | dbgs() << "\t MOV = " << getMOV(&SUnits[i]) << "\n" ; |
1937 | dbgs() << "\t D = " << getDepth(&SUnits[i]) << "\n" ; |
1938 | dbgs() << "\t H = " << getHeight(&SUnits[i]) << "\n" ; |
1939 | dbgs() << "\t ZLD = " << getZeroLatencyDepth(&SUnits[i]) << "\n" ; |
1940 | dbgs() << "\t ZLH = " << getZeroLatencyHeight(&SUnits[i]) << "\n" ; |
1941 | } |
1942 | }); |
1943 | } |
1944 | |
1945 | /// Compute the Pred_L(O) set, as defined in the paper. The set is defined |
1946 | /// as the predecessors of the elements of NodeOrder that are not also in |
1947 | /// NodeOrder. |
1948 | static bool pred_L(SetVector<SUnit *> &NodeOrder, |
1949 | SmallSetVector<SUnit *, 8> &Preds, |
1950 | const NodeSet *S = nullptr) { |
1951 | Preds.clear(); |
1952 | for (const SUnit *SU : NodeOrder) { |
1953 | for (const SDep &Pred : SU->Preds) { |
1954 | if (S && S->count(SU: Pred.getSUnit()) == 0) |
1955 | continue; |
1956 | if (ignoreDependence(D: Pred, isPred: true)) |
1957 | continue; |
1958 | if (NodeOrder.count(key: Pred.getSUnit()) == 0) |
1959 | Preds.insert(X: Pred.getSUnit()); |
1960 | } |
1961 | // Back-edges are predecessors with an anti-dependence. |
1962 | for (const SDep &Succ : SU->Succs) { |
1963 | if (Succ.getKind() != SDep::Anti) |
1964 | continue; |
1965 | if (S && S->count(SU: Succ.getSUnit()) == 0) |
1966 | continue; |
1967 | if (NodeOrder.count(key: Succ.getSUnit()) == 0) |
1968 | Preds.insert(X: Succ.getSUnit()); |
1969 | } |
1970 | } |
1971 | return !Preds.empty(); |
1972 | } |
1973 | |
1974 | /// Compute the Succ_L(O) set, as defined in the paper. The set is defined |
1975 | /// as the successors of the elements of NodeOrder that are not also in |
1976 | /// NodeOrder. |
1977 | static bool succ_L(SetVector<SUnit *> &NodeOrder, |
1978 | SmallSetVector<SUnit *, 8> &Succs, |
1979 | const NodeSet *S = nullptr) { |
1980 | Succs.clear(); |
1981 | for (const SUnit *SU : NodeOrder) { |
1982 | for (const SDep &Succ : SU->Succs) { |
1983 | if (S && S->count(SU: Succ.getSUnit()) == 0) |
1984 | continue; |
1985 | if (ignoreDependence(D: Succ, isPred: false)) |
1986 | continue; |
1987 | if (NodeOrder.count(key: Succ.getSUnit()) == 0) |
1988 | Succs.insert(X: Succ.getSUnit()); |
1989 | } |
1990 | for (const SDep &Pred : SU->Preds) { |
1991 | if (Pred.getKind() != SDep::Anti) |
1992 | continue; |
1993 | if (S && S->count(SU: Pred.getSUnit()) == 0) |
1994 | continue; |
1995 | if (NodeOrder.count(key: Pred.getSUnit()) == 0) |
1996 | Succs.insert(X: Pred.getSUnit()); |
1997 | } |
1998 | } |
1999 | return !Succs.empty(); |
2000 | } |
2001 | |
2002 | /// Return true if there is a path from the specified node to any of the nodes |
2003 | /// in DestNodes. Keep track and return the nodes in any path. |
2004 | static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path, |
2005 | SetVector<SUnit *> &DestNodes, |
2006 | SetVector<SUnit *> &Exclude, |
2007 | SmallPtrSet<SUnit *, 8> &Visited) { |
2008 | if (Cur->isBoundaryNode()) |
2009 | return false; |
2010 | if (Exclude.contains(key: Cur)) |
2011 | return false; |
2012 | if (DestNodes.contains(key: Cur)) |
2013 | return true; |
2014 | if (!Visited.insert(Ptr: Cur).second) |
2015 | return Path.contains(key: Cur); |
2016 | bool FoundPath = false; |
2017 | for (auto &SI : Cur->Succs) |
2018 | if (!ignoreDependence(D: SI, isPred: false)) |
2019 | FoundPath |= |
2020 | computePath(Cur: SI.getSUnit(), Path, DestNodes, Exclude, Visited); |
2021 | for (auto &PI : Cur->Preds) |
2022 | if (PI.getKind() == SDep::Anti) |
2023 | FoundPath |= |
2024 | computePath(Cur: PI.getSUnit(), Path, DestNodes, Exclude, Visited); |
2025 | if (FoundPath) |
2026 | Path.insert(X: Cur); |
2027 | return FoundPath; |
2028 | } |
2029 | |
2030 | /// Compute the live-out registers for the instructions in a node-set. |
2031 | /// The live-out registers are those that are defined in the node-set, |
2032 | /// but not used. Except for use operands of Phis. |
2033 | static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker, |
2034 | NodeSet &NS) { |
2035 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
2036 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
2037 | SmallVector<RegisterMaskPair, 8> LiveOutRegs; |
2038 | SmallSet<unsigned, 4> Uses; |
2039 | for (SUnit *SU : NS) { |
2040 | const MachineInstr *MI = SU->getInstr(); |
2041 | if (MI->isPHI()) |
2042 | continue; |
2043 | for (const MachineOperand &MO : MI->all_uses()) { |
2044 | Register Reg = MO.getReg(); |
2045 | if (Reg.isVirtual()) |
2046 | Uses.insert(V: Reg); |
2047 | else if (MRI.isAllocatable(PhysReg: Reg)) |
2048 | for (MCRegUnit Unit : TRI->regunits(Reg: Reg.asMCReg())) |
2049 | Uses.insert(V: Unit); |
2050 | } |
2051 | } |
2052 | for (SUnit *SU : NS) |
2053 | for (const MachineOperand &MO : SU->getInstr()->all_defs()) |
2054 | if (!MO.isDead()) { |
2055 | Register Reg = MO.getReg(); |
2056 | if (Reg.isVirtual()) { |
2057 | if (!Uses.count(V: Reg)) |
2058 | LiveOutRegs.push_back(Elt: RegisterMaskPair(Reg, |
2059 | LaneBitmask::getNone())); |
2060 | } else if (MRI.isAllocatable(PhysReg: Reg)) { |
2061 | for (MCRegUnit Unit : TRI->regunits(Reg: Reg.asMCReg())) |
2062 | if (!Uses.count(V: Unit)) |
2063 | LiveOutRegs.push_back( |
2064 | Elt: RegisterMaskPair(Unit, LaneBitmask::getNone())); |
2065 | } |
2066 | } |
2067 | RPTracker.addLiveRegs(Regs: LiveOutRegs); |
2068 | } |
2069 | |
2070 | /// A heuristic to filter nodes in recurrent node-sets if the register |
2071 | /// pressure of a set is too high. |
2072 | void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) { |
2073 | for (auto &NS : NodeSets) { |
2074 | // Skip small node-sets since they won't cause register pressure problems. |
2075 | if (NS.size() <= 2) |
2076 | continue; |
2077 | IntervalPressure RecRegPressure; |
2078 | RegPressureTracker RecRPTracker(RecRegPressure); |
2079 | RecRPTracker.init(mf: &MF, rci: &RegClassInfo, lis: &LIS, mbb: BB, pos: BB->end(), TrackLaneMasks: false, TrackUntiedDefs: true); |
2080 | computeLiveOuts(MF, RPTracker&: RecRPTracker, NS); |
2081 | RecRPTracker.closeBottom(); |
2082 | |
2083 | std::vector<SUnit *> SUnits(NS.begin(), NS.end()); |
2084 | llvm::sort(C&: SUnits, Comp: [](const SUnit *A, const SUnit *B) { |
2085 | return A->NodeNum > B->NodeNum; |
2086 | }); |
2087 | |
2088 | for (auto &SU : SUnits) { |
2089 | // Since we're computing the register pressure for a subset of the |
2090 | // instructions in a block, we need to set the tracker for each |
2091 | // instruction in the node-set. The tracker is set to the instruction |
2092 | // just after the one we're interested in. |
2093 | MachineBasicBlock::const_iterator CurInstI = SU->getInstr(); |
2094 | RecRPTracker.setPos(std::next(x: CurInstI)); |
2095 | |
2096 | RegPressureDelta RPDelta; |
2097 | ArrayRef<PressureChange> CriticalPSets; |
2098 | RecRPTracker.getMaxUpwardPressureDelta(MI: SU->getInstr(), PDiff: nullptr, Delta&: RPDelta, |
2099 | CriticalPSets, |
2100 | MaxPressureLimit: RecRegPressure.MaxSetPressure); |
2101 | if (RPDelta.Excess.isValid()) { |
2102 | LLVM_DEBUG( |
2103 | dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") " |
2104 | << TRI->getRegPressureSetName(RPDelta.Excess.getPSet()) |
2105 | << ":" << RPDelta.Excess.getUnitInc() << "\n" ); |
2106 | NS.setExceedPressure(SU); |
2107 | break; |
2108 | } |
2109 | RecRPTracker.recede(); |
2110 | } |
2111 | } |
2112 | } |
2113 | |
2114 | /// A heuristic to colocate node sets that have the same set of |
2115 | /// successors. |
2116 | void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) { |
2117 | unsigned Colocate = 0; |
2118 | for (int i = 0, e = NodeSets.size(); i < e; ++i) { |
2119 | NodeSet &N1 = NodeSets[i]; |
2120 | SmallSetVector<SUnit *, 8> S1; |
2121 | if (N1.empty() || !succ_L(NodeOrder&: N1, Succs&: S1)) |
2122 | continue; |
2123 | for (int j = i + 1; j < e; ++j) { |
2124 | NodeSet &N2 = NodeSets[j]; |
2125 | if (N1.compareRecMII(RHS&: N2) != 0) |
2126 | continue; |
2127 | SmallSetVector<SUnit *, 8> S2; |
2128 | if (N2.empty() || !succ_L(NodeOrder&: N2, Succs&: S2)) |
2129 | continue; |
2130 | if (llvm::set_is_subset(S1, S2) && S1.size() == S2.size()) { |
2131 | N1.setColocate(++Colocate); |
2132 | N2.setColocate(Colocate); |
2133 | break; |
2134 | } |
2135 | } |
2136 | } |
2137 | } |
2138 | |
2139 | /// Check if the existing node-sets are profitable. If not, then ignore the |
2140 | /// recurrent node-sets, and attempt to schedule all nodes together. This is |
2141 | /// a heuristic. If the MII is large and all the recurrent node-sets are small, |
2142 | /// then it's best to try to schedule all instructions together instead of |
2143 | /// starting with the recurrent node-sets. |
2144 | void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) { |
2145 | // Look for loops with a large MII. |
2146 | if (MII < 17) |
2147 | return; |
2148 | // Check if the node-set contains only a simple add recurrence. |
2149 | for (auto &NS : NodeSets) { |
2150 | if (NS.getRecMII() > 2) |
2151 | return; |
2152 | if (NS.getMaxDepth() > MII) |
2153 | return; |
2154 | } |
2155 | NodeSets.clear(); |
2156 | LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n" ); |
2157 | } |
2158 | |
2159 | /// Add the nodes that do not belong to a recurrence set into groups |
2160 | /// based upon connected components. |
2161 | void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) { |
2162 | SetVector<SUnit *> NodesAdded; |
2163 | SmallPtrSet<SUnit *, 8> Visited; |
2164 | // Add the nodes that are on a path between the previous node sets and |
2165 | // the current node set. |
2166 | for (NodeSet &I : NodeSets) { |
2167 | SmallSetVector<SUnit *, 8> N; |
2168 | // Add the nodes from the current node set to the previous node set. |
2169 | if (succ_L(NodeOrder&: I, Succs&: N)) { |
2170 | SetVector<SUnit *> Path; |
2171 | for (SUnit *NI : N) { |
2172 | Visited.clear(); |
2173 | computePath(Cur: NI, Path, DestNodes&: NodesAdded, Exclude&: I, Visited); |
2174 | } |
2175 | if (!Path.empty()) |
2176 | I.insert(S: Path.begin(), E: Path.end()); |
2177 | } |
2178 | // Add the nodes from the previous node set to the current node set. |
2179 | N.clear(); |
2180 | if (succ_L(NodeOrder&: NodesAdded, Succs&: N)) { |
2181 | SetVector<SUnit *> Path; |
2182 | for (SUnit *NI : N) { |
2183 | Visited.clear(); |
2184 | computePath(Cur: NI, Path, DestNodes&: I, Exclude&: NodesAdded, Visited); |
2185 | } |
2186 | if (!Path.empty()) |
2187 | I.insert(S: Path.begin(), E: Path.end()); |
2188 | } |
2189 | NodesAdded.insert(Start: I.begin(), End: I.end()); |
2190 | } |
2191 | |
2192 | // Create a new node set with the connected nodes of any successor of a node |
2193 | // in a recurrent set. |
2194 | NodeSet NewSet; |
2195 | SmallSetVector<SUnit *, 8> N; |
2196 | if (succ_L(NodeOrder&: NodesAdded, Succs&: N)) |
2197 | for (SUnit *I : N) |
2198 | addConnectedNodes(SU: I, NewSet, NodesAdded); |
2199 | if (!NewSet.empty()) |
2200 | NodeSets.push_back(Elt: NewSet); |
2201 | |
2202 | // Create a new node set with the connected nodes of any predecessor of a node |
2203 | // in a recurrent set. |
2204 | NewSet.clear(); |
2205 | if (pred_L(NodeOrder&: NodesAdded, Preds&: N)) |
2206 | for (SUnit *I : N) |
2207 | addConnectedNodes(SU: I, NewSet, NodesAdded); |
2208 | if (!NewSet.empty()) |
2209 | NodeSets.push_back(Elt: NewSet); |
2210 | |
2211 | // Create new nodes sets with the connected nodes any remaining node that |
2212 | // has no predecessor. |
2213 | for (SUnit &SU : SUnits) { |
2214 | if (NodesAdded.count(key: &SU) == 0) { |
2215 | NewSet.clear(); |
2216 | addConnectedNodes(SU: &SU, NewSet, NodesAdded); |
2217 | if (!NewSet.empty()) |
2218 | NodeSets.push_back(Elt: NewSet); |
2219 | } |
2220 | } |
2221 | } |
2222 | |
2223 | /// Add the node to the set, and add all of its connected nodes to the set. |
2224 | void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet, |
2225 | SetVector<SUnit *> &NodesAdded) { |
2226 | NewSet.insert(SU); |
2227 | NodesAdded.insert(X: SU); |
2228 | for (auto &SI : SU->Succs) { |
2229 | SUnit *Successor = SI.getSUnit(); |
2230 | if (!SI.isArtificial() && !Successor->isBoundaryNode() && |
2231 | NodesAdded.count(key: Successor) == 0) |
2232 | addConnectedNodes(SU: Successor, NewSet, NodesAdded); |
2233 | } |
2234 | for (auto &PI : SU->Preds) { |
2235 | SUnit *Predecessor = PI.getSUnit(); |
2236 | if (!PI.isArtificial() && NodesAdded.count(key: Predecessor) == 0) |
2237 | addConnectedNodes(SU: Predecessor, NewSet, NodesAdded); |
2238 | } |
2239 | } |
2240 | |
2241 | /// Return true if Set1 contains elements in Set2. The elements in common |
2242 | /// are returned in a different container. |
2243 | static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2, |
2244 | SmallSetVector<SUnit *, 8> &Result) { |
2245 | Result.clear(); |
2246 | for (SUnit *SU : Set1) { |
2247 | if (Set2.count(SU) != 0) |
2248 | Result.insert(X: SU); |
2249 | } |
2250 | return !Result.empty(); |
2251 | } |
2252 | |
2253 | /// Merge the recurrence node sets that have the same initial node. |
2254 | void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) { |
2255 | for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; |
2256 | ++I) { |
2257 | NodeSet &NI = *I; |
2258 | for (NodeSetType::iterator J = I + 1; J != E;) { |
2259 | NodeSet &NJ = *J; |
2260 | if (NI.getNode(i: 0)->NodeNum == NJ.getNode(i: 0)->NodeNum) { |
2261 | if (NJ.compareRecMII(RHS&: NI) > 0) |
2262 | NI.setRecMII(NJ.getRecMII()); |
2263 | for (SUnit *SU : *J) |
2264 | I->insert(SU); |
2265 | NodeSets.erase(CI: J); |
2266 | E = NodeSets.end(); |
2267 | } else { |
2268 | ++J; |
2269 | } |
2270 | } |
2271 | } |
2272 | } |
2273 | |
2274 | /// Remove nodes that have been scheduled in previous NodeSets. |
2275 | void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) { |
2276 | for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; |
2277 | ++I) |
2278 | for (NodeSetType::iterator J = I + 1; J != E;) { |
2279 | J->remove_if(P: [&](SUnit *SUJ) { return I->count(SU: SUJ); }); |
2280 | |
2281 | if (J->empty()) { |
2282 | NodeSets.erase(CI: J); |
2283 | E = NodeSets.end(); |
2284 | } else { |
2285 | ++J; |
2286 | } |
2287 | } |
2288 | } |
2289 | |
2290 | /// Compute an ordered list of the dependence graph nodes, which |
2291 | /// indicates the order that the nodes will be scheduled. This is a |
2292 | /// two-level algorithm. First, a partial order is created, which |
2293 | /// consists of a list of sets ordered from highest to lowest priority. |
2294 | void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) { |
2295 | SmallSetVector<SUnit *, 8> R; |
2296 | NodeOrder.clear(); |
2297 | |
2298 | for (auto &Nodes : NodeSets) { |
2299 | LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n" ); |
2300 | OrderKind Order; |
2301 | SmallSetVector<SUnit *, 8> N; |
2302 | if (pred_L(NodeOrder, Preds&: N) && llvm::set_is_subset(S1: N, S2: Nodes)) { |
2303 | R.insert(Start: N.begin(), End: N.end()); |
2304 | Order = BottomUp; |
2305 | LLVM_DEBUG(dbgs() << " Bottom up (preds) " ); |
2306 | } else if (succ_L(NodeOrder, Succs&: N) && llvm::set_is_subset(S1: N, S2: Nodes)) { |
2307 | R.insert(Start: N.begin(), End: N.end()); |
2308 | Order = TopDown; |
2309 | LLVM_DEBUG(dbgs() << " Top down (succs) " ); |
2310 | } else if (isIntersect(Set1&: N, Set2: Nodes, Result&: R)) { |
2311 | // If some of the successors are in the existing node-set, then use the |
2312 | // top-down ordering. |
2313 | Order = TopDown; |
2314 | LLVM_DEBUG(dbgs() << " Top down (intersect) " ); |
2315 | } else if (NodeSets.size() == 1) { |
2316 | for (const auto &N : Nodes) |
2317 | if (N->Succs.size() == 0) |
2318 | R.insert(X: N); |
2319 | Order = BottomUp; |
2320 | LLVM_DEBUG(dbgs() << " Bottom up (all) " ); |
2321 | } else { |
2322 | // Find the node with the highest ASAP. |
2323 | SUnit *maxASAP = nullptr; |
2324 | for (SUnit *SU : Nodes) { |
2325 | if (maxASAP == nullptr || getASAP(Node: SU) > getASAP(Node: maxASAP) || |
2326 | (getASAP(Node: SU) == getASAP(Node: maxASAP) && SU->NodeNum > maxASAP->NodeNum)) |
2327 | maxASAP = SU; |
2328 | } |
2329 | R.insert(X: maxASAP); |
2330 | Order = BottomUp; |
2331 | LLVM_DEBUG(dbgs() << " Bottom up (default) " ); |
2332 | } |
2333 | |
2334 | while (!R.empty()) { |
2335 | if (Order == TopDown) { |
2336 | // Choose the node with the maximum height. If more than one, choose |
2337 | // the node wiTH the maximum ZeroLatencyHeight. If still more than one, |
2338 | // choose the node with the lowest MOV. |
2339 | while (!R.empty()) { |
2340 | SUnit *maxHeight = nullptr; |
2341 | for (SUnit *I : R) { |
2342 | if (maxHeight == nullptr || getHeight(Node: I) > getHeight(Node: maxHeight)) |
2343 | maxHeight = I; |
2344 | else if (getHeight(Node: I) == getHeight(Node: maxHeight) && |
2345 | getZeroLatencyHeight(Node: I) > getZeroLatencyHeight(Node: maxHeight)) |
2346 | maxHeight = I; |
2347 | else if (getHeight(Node: I) == getHeight(Node: maxHeight) && |
2348 | getZeroLatencyHeight(Node: I) == |
2349 | getZeroLatencyHeight(Node: maxHeight) && |
2350 | getMOV(Node: I) < getMOV(Node: maxHeight)) |
2351 | maxHeight = I; |
2352 | } |
2353 | NodeOrder.insert(X: maxHeight); |
2354 | LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " " ); |
2355 | R.remove(X: maxHeight); |
2356 | for (const auto &I : maxHeight->Succs) { |
2357 | if (Nodes.count(SU: I.getSUnit()) == 0) |
2358 | continue; |
2359 | if (NodeOrder.contains(key: I.getSUnit())) |
2360 | continue; |
2361 | if (ignoreDependence(D: I, isPred: false)) |
2362 | continue; |
2363 | R.insert(X: I.getSUnit()); |
2364 | } |
2365 | // Back-edges are predecessors with an anti-dependence. |
2366 | for (const auto &I : maxHeight->Preds) { |
2367 | if (I.getKind() != SDep::Anti) |
2368 | continue; |
2369 | if (Nodes.count(SU: I.getSUnit()) == 0) |
2370 | continue; |
2371 | if (NodeOrder.contains(key: I.getSUnit())) |
2372 | continue; |
2373 | R.insert(X: I.getSUnit()); |
2374 | } |
2375 | } |
2376 | Order = BottomUp; |
2377 | LLVM_DEBUG(dbgs() << "\n Switching order to bottom up " ); |
2378 | SmallSetVector<SUnit *, 8> N; |
2379 | if (pred_L(NodeOrder, Preds&: N, S: &Nodes)) |
2380 | R.insert(Start: N.begin(), End: N.end()); |
2381 | } else { |
2382 | // Choose the node with the maximum depth. If more than one, choose |
2383 | // the node with the maximum ZeroLatencyDepth. If still more than one, |
2384 | // choose the node with the lowest MOV. |
2385 | while (!R.empty()) { |
2386 | SUnit *maxDepth = nullptr; |
2387 | for (SUnit *I : R) { |
2388 | if (maxDepth == nullptr || getDepth(Node: I) > getDepth(Node: maxDepth)) |
2389 | maxDepth = I; |
2390 | else if (getDepth(Node: I) == getDepth(Node: maxDepth) && |
2391 | getZeroLatencyDepth(Node: I) > getZeroLatencyDepth(Node: maxDepth)) |
2392 | maxDepth = I; |
2393 | else if (getDepth(Node: I) == getDepth(Node: maxDepth) && |
2394 | getZeroLatencyDepth(Node: I) == getZeroLatencyDepth(Node: maxDepth) && |
2395 | getMOV(Node: I) < getMOV(Node: maxDepth)) |
2396 | maxDepth = I; |
2397 | } |
2398 | NodeOrder.insert(X: maxDepth); |
2399 | LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " " ); |
2400 | R.remove(X: maxDepth); |
2401 | if (Nodes.isExceedSU(SU: maxDepth)) { |
2402 | Order = TopDown; |
2403 | R.clear(); |
2404 | R.insert(X: Nodes.getNode(i: 0)); |
2405 | break; |
2406 | } |
2407 | for (const auto &I : maxDepth->Preds) { |
2408 | if (Nodes.count(SU: I.getSUnit()) == 0) |
2409 | continue; |
2410 | if (NodeOrder.contains(key: I.getSUnit())) |
2411 | continue; |
2412 | R.insert(X: I.getSUnit()); |
2413 | } |
2414 | // Back-edges are predecessors with an anti-dependence. |
2415 | for (const auto &I : maxDepth->Succs) { |
2416 | if (I.getKind() != SDep::Anti) |
2417 | continue; |
2418 | if (Nodes.count(SU: I.getSUnit()) == 0) |
2419 | continue; |
2420 | if (NodeOrder.contains(key: I.getSUnit())) |
2421 | continue; |
2422 | R.insert(X: I.getSUnit()); |
2423 | } |
2424 | } |
2425 | Order = TopDown; |
2426 | LLVM_DEBUG(dbgs() << "\n Switching order to top down " ); |
2427 | SmallSetVector<SUnit *, 8> N; |
2428 | if (succ_L(NodeOrder, Succs&: N, S: &Nodes)) |
2429 | R.insert(Start: N.begin(), End: N.end()); |
2430 | } |
2431 | } |
2432 | LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n" ); |
2433 | } |
2434 | |
2435 | LLVM_DEBUG({ |
2436 | dbgs() << "Node order: " ; |
2437 | for (SUnit *I : NodeOrder) |
2438 | dbgs() << " " << I->NodeNum << " " ; |
2439 | dbgs() << "\n" ; |
2440 | }); |
2441 | } |
2442 | |
2443 | /// Process the nodes in the computed order and create the pipelined schedule |
2444 | /// of the instructions, if possible. Return true if a schedule is found. |
2445 | bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) { |
2446 | |
2447 | if (NodeOrder.empty()){ |
2448 | LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" ); |
2449 | return false; |
2450 | } |
2451 | |
2452 | bool scheduleFound = false; |
2453 | std::unique_ptr<HighRegisterPressureDetector> HRPDetector; |
2454 | if (LimitRegPressure) { |
2455 | HRPDetector = |
2456 | std::make_unique<HighRegisterPressureDetector>(args: Loop.getHeader(), args&: MF); |
2457 | HRPDetector->init(RCI: RegClassInfo); |
2458 | } |
2459 | // Keep increasing II until a valid schedule is found. |
2460 | for (unsigned II = MII; II <= MAX_II && !scheduleFound; ++II) { |
2461 | Schedule.reset(); |
2462 | Schedule.setInitiationInterval(II); |
2463 | LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n" ); |
2464 | |
2465 | SetVector<SUnit *>::iterator NI = NodeOrder.begin(); |
2466 | SetVector<SUnit *>::iterator NE = NodeOrder.end(); |
2467 | do { |
2468 | SUnit *SU = *NI; |
2469 | |
2470 | // Compute the schedule time for the instruction, which is based |
2471 | // upon the scheduled time for any predecessors/successors. |
2472 | int EarlyStart = INT_MIN; |
2473 | int LateStart = INT_MAX; |
2474 | Schedule.computeStart(SU, MaxEarlyStart: &EarlyStart, MinLateStart: &LateStart, II, DAG: this); |
2475 | LLVM_DEBUG({ |
2476 | dbgs() << "\n" ; |
2477 | dbgs() << "Inst (" << SU->NodeNum << ") " ; |
2478 | SU->getInstr()->dump(); |
2479 | dbgs() << "\n" ; |
2480 | }); |
2481 | LLVM_DEBUG( |
2482 | dbgs() << format("\tes: %8x ls: %8x\n" , EarlyStart, LateStart)); |
2483 | |
2484 | if (EarlyStart > LateStart) |
2485 | scheduleFound = false; |
2486 | else if (EarlyStart != INT_MIN && LateStart == INT_MAX) |
2487 | scheduleFound = |
2488 | Schedule.insert(SU, StartCycle: EarlyStart, EndCycle: EarlyStart + (int)II - 1, II); |
2489 | else if (EarlyStart == INT_MIN && LateStart != INT_MAX) |
2490 | scheduleFound = |
2491 | Schedule.insert(SU, StartCycle: LateStart, EndCycle: LateStart - (int)II + 1, II); |
2492 | else if (EarlyStart != INT_MIN && LateStart != INT_MAX) { |
2493 | LateStart = std::min(a: LateStart, b: EarlyStart + (int)II - 1); |
2494 | // When scheduling a Phi it is better to start at the late cycle and |
2495 | // go backwards. The default order may insert the Phi too far away |
2496 | // from its first dependence. |
2497 | // Also, do backward search when all scheduled predecessors are |
2498 | // loop-carried output/order dependencies. Empirically, there are also |
2499 | // cases where scheduling becomes possible with backward search. |
2500 | if (SU->getInstr()->isPHI() || |
2501 | Schedule.onlyHasLoopCarriedOutputOrOrderPreds(SU, DAG: this)) |
2502 | scheduleFound = Schedule.insert(SU, StartCycle: LateStart, EndCycle: EarlyStart, II); |
2503 | else |
2504 | scheduleFound = Schedule.insert(SU, StartCycle: EarlyStart, EndCycle: LateStart, II); |
2505 | } else { |
2506 | int FirstCycle = Schedule.getFirstCycle(); |
2507 | scheduleFound = Schedule.insert(SU, StartCycle: FirstCycle + getASAP(Node: SU), |
2508 | EndCycle: FirstCycle + getASAP(Node: SU) + II - 1, II); |
2509 | } |
2510 | |
2511 | // Even if we find a schedule, make sure the schedule doesn't exceed the |
2512 | // allowable number of stages. We keep trying if this happens. |
2513 | if (scheduleFound) |
2514 | if (SwpMaxStages > -1 && |
2515 | Schedule.getMaxStageCount() > (unsigned)SwpMaxStages) |
2516 | scheduleFound = false; |
2517 | |
2518 | LLVM_DEBUG({ |
2519 | if (!scheduleFound) |
2520 | dbgs() << "\tCan't schedule\n" ; |
2521 | }); |
2522 | } while (++NI != NE && scheduleFound); |
2523 | |
2524 | // If a schedule is found, ensure non-pipelined instructions are in stage 0 |
2525 | if (scheduleFound) |
2526 | scheduleFound = |
2527 | Schedule.normalizeNonPipelinedInstructions(SSD: this, PLI: LoopPipelinerInfo); |
2528 | |
2529 | // If a schedule is found, check if it is a valid schedule too. |
2530 | if (scheduleFound) |
2531 | scheduleFound = Schedule.isValidSchedule(SSD: this); |
2532 | |
2533 | // If a schedule was found and the option is enabled, check if the schedule |
2534 | // might generate additional register spills/fills. |
2535 | if (scheduleFound && LimitRegPressure) |
2536 | scheduleFound = |
2537 | !HRPDetector->detect(SSD: this, Schedule, MaxStage: Schedule.getMaxStageCount()); |
2538 | } |
2539 | |
2540 | LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound |
2541 | << " (II=" << Schedule.getInitiationInterval() |
2542 | << ")\n" ); |
2543 | |
2544 | if (scheduleFound) { |
2545 | scheduleFound = LoopPipelinerInfo->shouldUseSchedule(SSD&: *this, SMS&: Schedule); |
2546 | if (!scheduleFound) |
2547 | LLVM_DEBUG(dbgs() << "Target rejected schedule\n" ); |
2548 | } |
2549 | |
2550 | if (scheduleFound) { |
2551 | Schedule.finalizeSchedule(SSD: this); |
2552 | Pass.ORE->emit(RemarkBuilder: [&]() { |
2553 | return MachineOptimizationRemarkAnalysis( |
2554 | DEBUG_TYPE, "schedule" , Loop.getStartLoc(), Loop.getHeader()) |
2555 | << "Schedule found with Initiation Interval: " |
2556 | << ore::NV("II" , Schedule.getInitiationInterval()) |
2557 | << ", MaxStageCount: " |
2558 | << ore::NV("MaxStageCount" , Schedule.getMaxStageCount()); |
2559 | }); |
2560 | } else |
2561 | Schedule.reset(); |
2562 | |
2563 | return scheduleFound && Schedule.getMaxStageCount() > 0; |
2564 | } |
2565 | |
2566 | /// Return true if we can compute the amount the instruction changes |
2567 | /// during each iteration. Set Delta to the amount of the change. |
2568 | bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) { |
2569 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
2570 | const MachineOperand *BaseOp; |
2571 | int64_t Offset; |
2572 | bool OffsetIsScalable; |
2573 | if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) |
2574 | return false; |
2575 | |
2576 | // FIXME: This algorithm assumes instructions have fixed-size offsets. |
2577 | if (OffsetIsScalable) |
2578 | return false; |
2579 | |
2580 | if (!BaseOp->isReg()) |
2581 | return false; |
2582 | |
2583 | Register BaseReg = BaseOp->getReg(); |
2584 | |
2585 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
2586 | // Check if there is a Phi. If so, get the definition in the loop. |
2587 | MachineInstr *BaseDef = MRI.getVRegDef(Reg: BaseReg); |
2588 | if (BaseDef && BaseDef->isPHI()) { |
2589 | BaseReg = getLoopPhiReg(Phi: *BaseDef, LoopBB: MI.getParent()); |
2590 | BaseDef = MRI.getVRegDef(Reg: BaseReg); |
2591 | } |
2592 | if (!BaseDef) |
2593 | return false; |
2594 | |
2595 | int D = 0; |
2596 | if (!TII->getIncrementValue(MI: *BaseDef, Value&: D) && D >= 0) |
2597 | return false; |
2598 | |
2599 | Delta = D; |
2600 | return true; |
2601 | } |
2602 | |
2603 | /// Check if we can change the instruction to use an offset value from the |
2604 | /// previous iteration. If so, return true and set the base and offset values |
2605 | /// so that we can rewrite the load, if necessary. |
2606 | /// v1 = Phi(v0, v3) |
2607 | /// v2 = load v1, 0 |
2608 | /// v3 = post_store v1, 4, x |
2609 | /// This function enables the load to be rewritten as v2 = load v3, 4. |
2610 | bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI, |
2611 | unsigned &BasePos, |
2612 | unsigned &OffsetPos, |
2613 | unsigned &NewBase, |
2614 | int64_t &Offset) { |
2615 | // Get the load instruction. |
2616 | if (TII->isPostIncrement(MI: *MI)) |
2617 | return false; |
2618 | unsigned BasePosLd, OffsetPosLd; |
2619 | if (!TII->getBaseAndOffsetPosition(MI: *MI, BasePos&: BasePosLd, OffsetPos&: OffsetPosLd)) |
2620 | return false; |
2621 | Register BaseReg = MI->getOperand(i: BasePosLd).getReg(); |
2622 | |
2623 | // Look for the Phi instruction. |
2624 | MachineRegisterInfo &MRI = MI->getMF()->getRegInfo(); |
2625 | MachineInstr *Phi = MRI.getVRegDef(Reg: BaseReg); |
2626 | if (!Phi || !Phi->isPHI()) |
2627 | return false; |
2628 | // Get the register defined in the loop block. |
2629 | unsigned PrevReg = getLoopPhiReg(Phi: *Phi, LoopBB: MI->getParent()); |
2630 | if (!PrevReg) |
2631 | return false; |
2632 | |
2633 | // Check for the post-increment load/store instruction. |
2634 | MachineInstr *PrevDef = MRI.getVRegDef(Reg: PrevReg); |
2635 | if (!PrevDef || PrevDef == MI) |
2636 | return false; |
2637 | |
2638 | if (!TII->isPostIncrement(MI: *PrevDef)) |
2639 | return false; |
2640 | |
2641 | unsigned BasePos1 = 0, OffsetPos1 = 0; |
2642 | if (!TII->getBaseAndOffsetPosition(MI: *PrevDef, BasePos&: BasePos1, OffsetPos&: OffsetPos1)) |
2643 | return false; |
2644 | |
2645 | // Make sure that the instructions do not access the same memory location in |
2646 | // the next iteration. |
2647 | int64_t LoadOffset = MI->getOperand(i: OffsetPosLd).getImm(); |
2648 | int64_t StoreOffset = PrevDef->getOperand(i: OffsetPos1).getImm(); |
2649 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: MI); |
2650 | NewMI->getOperand(i: OffsetPosLd).setImm(LoadOffset + StoreOffset); |
2651 | bool Disjoint = TII->areMemAccessesTriviallyDisjoint(MIa: *NewMI, MIb: *PrevDef); |
2652 | MF.deleteMachineInstr(MI: NewMI); |
2653 | if (!Disjoint) |
2654 | return false; |
2655 | |
2656 | // Set the return value once we determine that we return true. |
2657 | BasePos = BasePosLd; |
2658 | OffsetPos = OffsetPosLd; |
2659 | NewBase = PrevReg; |
2660 | Offset = StoreOffset; |
2661 | return true; |
2662 | } |
2663 | |
2664 | /// Apply changes to the instruction if needed. The changes are need |
2665 | /// to improve the scheduling and depend up on the final schedule. |
2666 | void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI, |
2667 | SMSchedule &Schedule) { |
2668 | SUnit *SU = getSUnit(MI); |
2669 | DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = |
2670 | InstrChanges.find(Val: SU); |
2671 | if (It != InstrChanges.end()) { |
2672 | std::pair<unsigned, int64_t> RegAndOffset = It->second; |
2673 | unsigned BasePos, OffsetPos; |
2674 | if (!TII->getBaseAndOffsetPosition(MI: *MI, BasePos, OffsetPos)) |
2675 | return; |
2676 | Register BaseReg = MI->getOperand(i: BasePos).getReg(); |
2677 | MachineInstr *LoopDef = findDefInLoop(Reg: BaseReg); |
2678 | int DefStageNum = Schedule.stageScheduled(SU: getSUnit(MI: LoopDef)); |
2679 | int DefCycleNum = Schedule.cycleScheduled(SU: getSUnit(MI: LoopDef)); |
2680 | int BaseStageNum = Schedule.stageScheduled(SU); |
2681 | int BaseCycleNum = Schedule.cycleScheduled(SU); |
2682 | if (BaseStageNum < DefStageNum) { |
2683 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: MI); |
2684 | int OffsetDiff = DefStageNum - BaseStageNum; |
2685 | if (DefCycleNum < BaseCycleNum) { |
2686 | NewMI->getOperand(i: BasePos).setReg(RegAndOffset.first); |
2687 | if (OffsetDiff > 0) |
2688 | --OffsetDiff; |
2689 | } |
2690 | int64_t NewOffset = |
2691 | MI->getOperand(i: OffsetPos).getImm() + RegAndOffset.second * OffsetDiff; |
2692 | NewMI->getOperand(i: OffsetPos).setImm(NewOffset); |
2693 | SU->setInstr(NewMI); |
2694 | MISUnitMap[NewMI] = SU; |
2695 | NewMIs[MI] = NewMI; |
2696 | } |
2697 | } |
2698 | } |
2699 | |
2700 | /// Return the instruction in the loop that defines the register. |
2701 | /// If the definition is a Phi, then follow the Phi operand to |
2702 | /// the instruction in the loop. |
2703 | MachineInstr *SwingSchedulerDAG::findDefInLoop(Register Reg) { |
2704 | SmallPtrSet<MachineInstr *, 8> Visited; |
2705 | MachineInstr *Def = MRI.getVRegDef(Reg); |
2706 | while (Def->isPHI()) { |
2707 | if (!Visited.insert(Ptr: Def).second) |
2708 | break; |
2709 | for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) |
2710 | if (Def->getOperand(i: i + 1).getMBB() == BB) { |
2711 | Def = MRI.getVRegDef(Reg: Def->getOperand(i).getReg()); |
2712 | break; |
2713 | } |
2714 | } |
2715 | return Def; |
2716 | } |
2717 | |
2718 | /// Return true for an order or output dependence that is loop carried |
2719 | /// potentially. A dependence is loop carried if the destination defines a value |
2720 | /// that may be used or defined by the source in a subsequent iteration. |
2721 | bool SwingSchedulerDAG::isLoopCarriedDep(SUnit *Source, const SDep &Dep, |
2722 | bool isSucc) { |
2723 | if ((Dep.getKind() != SDep::Order && Dep.getKind() != SDep::Output) || |
2724 | Dep.isArtificial() || Dep.getSUnit()->isBoundaryNode()) |
2725 | return false; |
2726 | |
2727 | if (!SwpPruneLoopCarried) |
2728 | return true; |
2729 | |
2730 | if (Dep.getKind() == SDep::Output) |
2731 | return true; |
2732 | |
2733 | MachineInstr *SI = Source->getInstr(); |
2734 | MachineInstr *DI = Dep.getSUnit()->getInstr(); |
2735 | if (!isSucc) |
2736 | std::swap(a&: SI, b&: DI); |
2737 | assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI." ); |
2738 | |
2739 | // Assume ordered loads and stores may have a loop carried dependence. |
2740 | if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() || |
2741 | SI->mayRaiseFPException() || DI->mayRaiseFPException() || |
2742 | SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef()) |
2743 | return true; |
2744 | |
2745 | if (!DI->mayLoadOrStore() || !SI->mayLoadOrStore()) |
2746 | return false; |
2747 | |
2748 | // The conservative assumption is that a dependence between memory operations |
2749 | // may be loop carried. The following code checks when it can be proved that |
2750 | // there is no loop carried dependence. |
2751 | unsigned DeltaS, DeltaD; |
2752 | if (!computeDelta(MI&: *SI, Delta&: DeltaS) || !computeDelta(MI&: *DI, Delta&: DeltaD)) |
2753 | return true; |
2754 | |
2755 | const MachineOperand *BaseOpS, *BaseOpD; |
2756 | int64_t OffsetS, OffsetD; |
2757 | bool OffsetSIsScalable, OffsetDIsScalable; |
2758 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
2759 | if (!TII->getMemOperandWithOffset(MI: *SI, BaseOp&: BaseOpS, Offset&: OffsetS, OffsetIsScalable&: OffsetSIsScalable, |
2760 | TRI) || |
2761 | !TII->getMemOperandWithOffset(MI: *DI, BaseOp&: BaseOpD, Offset&: OffsetD, OffsetIsScalable&: OffsetDIsScalable, |
2762 | TRI)) |
2763 | return true; |
2764 | |
2765 | assert(!OffsetSIsScalable && !OffsetDIsScalable && |
2766 | "Expected offsets to be byte offsets" ); |
2767 | |
2768 | MachineInstr *DefS = MRI.getVRegDef(Reg: BaseOpS->getReg()); |
2769 | MachineInstr *DefD = MRI.getVRegDef(Reg: BaseOpD->getReg()); |
2770 | if (!DefS || !DefD || !DefS->isPHI() || !DefD->isPHI()) |
2771 | return true; |
2772 | |
2773 | unsigned InitValS = 0; |
2774 | unsigned LoopValS = 0; |
2775 | unsigned InitValD = 0; |
2776 | unsigned LoopValD = 0; |
2777 | getPhiRegs(Phi&: *DefS, Loop: BB, InitVal&: InitValS, LoopVal&: LoopValS); |
2778 | getPhiRegs(Phi&: *DefD, Loop: BB, InitVal&: InitValD, LoopVal&: LoopValD); |
2779 | MachineInstr *InitDefS = MRI.getVRegDef(Reg: InitValS); |
2780 | MachineInstr *InitDefD = MRI.getVRegDef(Reg: InitValD); |
2781 | |
2782 | if (!InitDefS->isIdenticalTo(Other: *InitDefD)) |
2783 | return true; |
2784 | |
2785 | // Check that the base register is incremented by a constant value for each |
2786 | // iteration. |
2787 | MachineInstr *LoopDefS = MRI.getVRegDef(Reg: LoopValS); |
2788 | int D = 0; |
2789 | if (!LoopDefS || !TII->getIncrementValue(MI: *LoopDefS, Value&: D)) |
2790 | return true; |
2791 | |
2792 | LocationSize AccessSizeS = (*SI->memoperands_begin())->getSize(); |
2793 | LocationSize AccessSizeD = (*DI->memoperands_begin())->getSize(); |
2794 | |
2795 | // This is the main test, which checks the offset values and the loop |
2796 | // increment value to determine if the accesses may be loop carried. |
2797 | if (!AccessSizeS.hasValue() || !AccessSizeD.hasValue()) |
2798 | return true; |
2799 | |
2800 | if (DeltaS != DeltaD || DeltaS < AccessSizeS.getValue() || |
2801 | DeltaD < AccessSizeD.getValue()) |
2802 | return true; |
2803 | |
2804 | return (OffsetS + (int64_t)AccessSizeS.getValue() < |
2805 | OffsetD + (int64_t)AccessSizeD.getValue()); |
2806 | } |
2807 | |
2808 | void SwingSchedulerDAG::postProcessDAG() { |
2809 | for (auto &M : Mutations) |
2810 | M->apply(DAG: this); |
2811 | } |
2812 | |
2813 | /// Try to schedule the node at the specified StartCycle and continue |
2814 | /// until the node is schedule or the EndCycle is reached. This function |
2815 | /// returns true if the node is scheduled. This routine may search either |
2816 | /// forward or backward for a place to insert the instruction based upon |
2817 | /// the relative values of StartCycle and EndCycle. |
2818 | bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) { |
2819 | bool forward = true; |
2820 | LLVM_DEBUG({ |
2821 | dbgs() << "Trying to insert node between " << StartCycle << " and " |
2822 | << EndCycle << " II: " << II << "\n" ; |
2823 | }); |
2824 | if (StartCycle > EndCycle) |
2825 | forward = false; |
2826 | |
2827 | // The terminating condition depends on the direction. |
2828 | int termCycle = forward ? EndCycle + 1 : EndCycle - 1; |
2829 | for (int curCycle = StartCycle; curCycle != termCycle; |
2830 | forward ? ++curCycle : --curCycle) { |
2831 | |
2832 | if (ST.getInstrInfo()->isZeroCost(Opcode: SU->getInstr()->getOpcode()) || |
2833 | ProcItinResources.canReserveResources(SU&: *SU, Cycle: curCycle)) { |
2834 | LLVM_DEBUG({ |
2835 | dbgs() << "\tinsert at cycle " << curCycle << " " ; |
2836 | SU->getInstr()->dump(); |
2837 | }); |
2838 | |
2839 | if (!ST.getInstrInfo()->isZeroCost(Opcode: SU->getInstr()->getOpcode())) |
2840 | ProcItinResources.reserveResources(SU&: *SU, Cycle: curCycle); |
2841 | ScheduledInstrs[curCycle].push_back(x: SU); |
2842 | InstrToCycle.insert(x: std::make_pair(x&: SU, y&: curCycle)); |
2843 | if (curCycle > LastCycle) |
2844 | LastCycle = curCycle; |
2845 | if (curCycle < FirstCycle) |
2846 | FirstCycle = curCycle; |
2847 | return true; |
2848 | } |
2849 | LLVM_DEBUG({ |
2850 | dbgs() << "\tfailed to insert at cycle " << curCycle << " " ; |
2851 | SU->getInstr()->dump(); |
2852 | }); |
2853 | } |
2854 | return false; |
2855 | } |
2856 | |
2857 | // Return the cycle of the earliest scheduled instruction in the chain. |
2858 | int SMSchedule::earliestCycleInChain(const SDep &Dep) { |
2859 | SmallPtrSet<SUnit *, 8> Visited; |
2860 | SmallVector<SDep, 8> Worklist; |
2861 | Worklist.push_back(Elt: Dep); |
2862 | int EarlyCycle = INT_MAX; |
2863 | while (!Worklist.empty()) { |
2864 | const SDep &Cur = Worklist.pop_back_val(); |
2865 | SUnit *PrevSU = Cur.getSUnit(); |
2866 | if (Visited.count(Ptr: PrevSU)) |
2867 | continue; |
2868 | std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(x: PrevSU); |
2869 | if (it == InstrToCycle.end()) |
2870 | continue; |
2871 | EarlyCycle = std::min(a: EarlyCycle, b: it->second); |
2872 | for (const auto &PI : PrevSU->Preds) |
2873 | if (PI.getKind() == SDep::Order || PI.getKind() == SDep::Output) |
2874 | Worklist.push_back(Elt: PI); |
2875 | Visited.insert(Ptr: PrevSU); |
2876 | } |
2877 | return EarlyCycle; |
2878 | } |
2879 | |
2880 | // Return the cycle of the latest scheduled instruction in the chain. |
2881 | int SMSchedule::latestCycleInChain(const SDep &Dep) { |
2882 | SmallPtrSet<SUnit *, 8> Visited; |
2883 | SmallVector<SDep, 8> Worklist; |
2884 | Worklist.push_back(Elt: Dep); |
2885 | int LateCycle = INT_MIN; |
2886 | while (!Worklist.empty()) { |
2887 | const SDep &Cur = Worklist.pop_back_val(); |
2888 | SUnit *SuccSU = Cur.getSUnit(); |
2889 | if (Visited.count(Ptr: SuccSU) || SuccSU->isBoundaryNode()) |
2890 | continue; |
2891 | std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(x: SuccSU); |
2892 | if (it == InstrToCycle.end()) |
2893 | continue; |
2894 | LateCycle = std::max(a: LateCycle, b: it->second); |
2895 | for (const auto &SI : SuccSU->Succs) |
2896 | if (SI.getKind() == SDep::Order || SI.getKind() == SDep::Output) |
2897 | Worklist.push_back(Elt: SI); |
2898 | Visited.insert(Ptr: SuccSU); |
2899 | } |
2900 | return LateCycle; |
2901 | } |
2902 | |
2903 | /// If an instruction has a use that spans multiple iterations, then |
2904 | /// return true. These instructions are characterized by having a back-ege |
2905 | /// to a Phi, which contains a reference to another Phi. |
2906 | static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) { |
2907 | for (auto &P : SU->Preds) |
2908 | if (DAG->isBackedge(Source: SU, Dep: P) && P.getSUnit()->getInstr()->isPHI()) |
2909 | for (auto &S : P.getSUnit()->Succs) |
2910 | if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI()) |
2911 | return P.getSUnit(); |
2912 | return nullptr; |
2913 | } |
2914 | |
2915 | /// Compute the scheduling start slot for the instruction. The start slot |
2916 | /// depends on any predecessor or successor nodes scheduled already. |
2917 | void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart, |
2918 | int II, SwingSchedulerDAG *DAG) { |
2919 | // Iterate over each instruction that has been scheduled already. The start |
2920 | // slot computation depends on whether the previously scheduled instruction |
2921 | // is a predecessor or successor of the specified instruction. |
2922 | for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) { |
2923 | |
2924 | // Iterate over each instruction in the current cycle. |
2925 | for (SUnit *I : getInstructions(cycle)) { |
2926 | // Because we're processing a DAG for the dependences, we recognize |
2927 | // the back-edge in recurrences by anti dependences. |
2928 | for (unsigned i = 0, e = (unsigned)SU->Preds.size(); i != e; ++i) { |
2929 | const SDep &Dep = SU->Preds[i]; |
2930 | if (Dep.getSUnit() == I) { |
2931 | if (!DAG->isBackedge(Source: SU, Dep)) { |
2932 | int EarlyStart = cycle + Dep.getLatency() - |
2933 | DAG->getDistance(U: Dep.getSUnit(), V: SU, Dep) * II; |
2934 | *MaxEarlyStart = std::max(a: *MaxEarlyStart, b: EarlyStart); |
2935 | if (DAG->isLoopCarriedDep(Source: SU, Dep, isSucc: false)) { |
2936 | int End = earliestCycleInChain(Dep) + (II - 1); |
2937 | *MinLateStart = std::min(a: *MinLateStart, b: End); |
2938 | } |
2939 | } else { |
2940 | int LateStart = cycle - Dep.getLatency() + |
2941 | DAG->getDistance(U: SU, V: Dep.getSUnit(), Dep) * II; |
2942 | *MinLateStart = std::min(a: *MinLateStart, b: LateStart); |
2943 | } |
2944 | } |
2945 | // For instruction that requires multiple iterations, make sure that |
2946 | // the dependent instruction is not scheduled past the definition. |
2947 | SUnit *BE = multipleIterations(SU: I, DAG); |
2948 | if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() && |
2949 | !SU->isPred(N: I)) |
2950 | *MinLateStart = std::min(a: *MinLateStart, b: cycle); |
2951 | } |
2952 | for (unsigned i = 0, e = (unsigned)SU->Succs.size(); i != e; ++i) { |
2953 | if (SU->Succs[i].getSUnit() == I) { |
2954 | const SDep &Dep = SU->Succs[i]; |
2955 | if (!DAG->isBackedge(Source: SU, Dep)) { |
2956 | int LateStart = cycle - Dep.getLatency() + |
2957 | DAG->getDistance(U: SU, V: Dep.getSUnit(), Dep) * II; |
2958 | *MinLateStart = std::min(a: *MinLateStart, b: LateStart); |
2959 | if (DAG->isLoopCarriedDep(Source: SU, Dep)) { |
2960 | int Start = latestCycleInChain(Dep) + 1 - II; |
2961 | *MaxEarlyStart = std::max(a: *MaxEarlyStart, b: Start); |
2962 | } |
2963 | } else { |
2964 | int EarlyStart = cycle + Dep.getLatency() - |
2965 | DAG->getDistance(U: Dep.getSUnit(), V: SU, Dep) * II; |
2966 | *MaxEarlyStart = std::max(a: *MaxEarlyStart, b: EarlyStart); |
2967 | } |
2968 | } |
2969 | } |
2970 | } |
2971 | } |
2972 | } |
2973 | |
2974 | /// Order the instructions within a cycle so that the definitions occur |
2975 | /// before the uses. Returns true if the instruction is added to the start |
2976 | /// of the list, or false if added to the end. |
2977 | void SMSchedule::orderDependence(const SwingSchedulerDAG *SSD, SUnit *SU, |
2978 | std::deque<SUnit *> &Insts) const { |
2979 | MachineInstr *MI = SU->getInstr(); |
2980 | bool OrderBeforeUse = false; |
2981 | bool OrderAfterDef = false; |
2982 | bool OrderBeforeDef = false; |
2983 | unsigned MoveDef = 0; |
2984 | unsigned MoveUse = 0; |
2985 | int StageInst1 = stageScheduled(SU); |
2986 | |
2987 | unsigned Pos = 0; |
2988 | for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E; |
2989 | ++I, ++Pos) { |
2990 | for (MachineOperand &MO : MI->operands()) { |
2991 | if (!MO.isReg() || !MO.getReg().isVirtual()) |
2992 | continue; |
2993 | |
2994 | Register Reg = MO.getReg(); |
2995 | unsigned BasePos, OffsetPos; |
2996 | if (ST.getInstrInfo()->getBaseAndOffsetPosition(MI: *MI, BasePos, OffsetPos)) |
2997 | if (MI->getOperand(i: BasePos).getReg() == Reg) |
2998 | if (unsigned NewReg = SSD->getInstrBaseReg(SU)) |
2999 | Reg = NewReg; |
3000 | bool Reads, Writes; |
3001 | std::tie(args&: Reads, args&: Writes) = |
3002 | (*I)->getInstr()->readsWritesVirtualRegister(Reg); |
3003 | if (MO.isDef() && Reads && stageScheduled(SU: *I) <= StageInst1) { |
3004 | OrderBeforeUse = true; |
3005 | if (MoveUse == 0) |
3006 | MoveUse = Pos; |
3007 | } else if (MO.isDef() && Reads && stageScheduled(SU: *I) > StageInst1) { |
3008 | // Add the instruction after the scheduled instruction. |
3009 | OrderAfterDef = true; |
3010 | MoveDef = Pos; |
3011 | } else if (MO.isUse() && Writes && stageScheduled(SU: *I) == StageInst1) { |
3012 | if (cycleScheduled(SU: *I) == cycleScheduled(SU) && !(*I)->isSucc(N: SU)) { |
3013 | OrderBeforeUse = true; |
3014 | if (MoveUse == 0) |
3015 | MoveUse = Pos; |
3016 | } else { |
3017 | OrderAfterDef = true; |
3018 | MoveDef = Pos; |
3019 | } |
3020 | } else if (MO.isUse() && Writes && stageScheduled(SU: *I) > StageInst1) { |
3021 | OrderBeforeUse = true; |
3022 | if (MoveUse == 0) |
3023 | MoveUse = Pos; |
3024 | if (MoveUse != 0) { |
3025 | OrderAfterDef = true; |
3026 | MoveDef = Pos - 1; |
3027 | } |
3028 | } else if (MO.isUse() && Writes && stageScheduled(SU: *I) < StageInst1) { |
3029 | // Add the instruction before the scheduled instruction. |
3030 | OrderBeforeUse = true; |
3031 | if (MoveUse == 0) |
3032 | MoveUse = Pos; |
3033 | } else if (MO.isUse() && stageScheduled(SU: *I) == StageInst1 && |
3034 | isLoopCarriedDefOfUse(SSD, Def: (*I)->getInstr(), MO)) { |
3035 | if (MoveUse == 0) { |
3036 | OrderBeforeDef = true; |
3037 | MoveUse = Pos; |
3038 | } |
3039 | } |
3040 | } |
3041 | // Check for order dependences between instructions. Make sure the source |
3042 | // is ordered before the destination. |
3043 | for (auto &S : SU->Succs) { |
3044 | if (S.getSUnit() != *I) |
3045 | continue; |
3046 | if (S.getKind() == SDep::Order && stageScheduled(SU: *I) == StageInst1) { |
3047 | OrderBeforeUse = true; |
3048 | if (Pos < MoveUse) |
3049 | MoveUse = Pos; |
3050 | } |
3051 | // We did not handle HW dependences in previous for loop, |
3052 | // and we normally set Latency = 0 for Anti deps, |
3053 | // so may have nodes in same cycle with Anti denpendent on HW regs. |
3054 | else if (S.getKind() == SDep::Anti && stageScheduled(SU: *I) == StageInst1) { |
3055 | OrderBeforeUse = true; |
3056 | if ((MoveUse == 0) || (Pos < MoveUse)) |
3057 | MoveUse = Pos; |
3058 | } |
3059 | } |
3060 | for (auto &P : SU->Preds) { |
3061 | if (P.getSUnit() != *I) |
3062 | continue; |
3063 | if (P.getKind() == SDep::Order && stageScheduled(SU: *I) == StageInst1) { |
3064 | OrderAfterDef = true; |
3065 | MoveDef = Pos; |
3066 | } |
3067 | } |
3068 | } |
3069 | |
3070 | // A circular dependence. |
3071 | if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef) |
3072 | OrderBeforeUse = false; |
3073 | |
3074 | // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due |
3075 | // to a loop-carried dependence. |
3076 | if (OrderBeforeDef) |
3077 | OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef); |
3078 | |
3079 | // The uncommon case when the instruction order needs to be updated because |
3080 | // there is both a use and def. |
3081 | if (OrderBeforeUse && OrderAfterDef) { |
3082 | SUnit *UseSU = Insts.at(n: MoveUse); |
3083 | SUnit *DefSU = Insts.at(n: MoveDef); |
3084 | if (MoveUse > MoveDef) { |
3085 | Insts.erase(position: Insts.begin() + MoveUse); |
3086 | Insts.erase(position: Insts.begin() + MoveDef); |
3087 | } else { |
3088 | Insts.erase(position: Insts.begin() + MoveDef); |
3089 | Insts.erase(position: Insts.begin() + MoveUse); |
3090 | } |
3091 | orderDependence(SSD, SU: UseSU, Insts); |
3092 | orderDependence(SSD, SU, Insts); |
3093 | orderDependence(SSD, SU: DefSU, Insts); |
3094 | return; |
3095 | } |
3096 | // Put the new instruction first if there is a use in the list. Otherwise, |
3097 | // put it at the end of the list. |
3098 | if (OrderBeforeUse) |
3099 | Insts.push_front(x: SU); |
3100 | else |
3101 | Insts.push_back(x: SU); |
3102 | } |
3103 | |
3104 | /// Return true if the scheduled Phi has a loop carried operand. |
3105 | bool SMSchedule::isLoopCarried(const SwingSchedulerDAG *SSD, |
3106 | MachineInstr &Phi) const { |
3107 | if (!Phi.isPHI()) |
3108 | return false; |
3109 | assert(Phi.isPHI() && "Expecting a Phi." ); |
3110 | SUnit *DefSU = SSD->getSUnit(MI: &Phi); |
3111 | unsigned DefCycle = cycleScheduled(SU: DefSU); |
3112 | int DefStage = stageScheduled(SU: DefSU); |
3113 | |
3114 | unsigned InitVal = 0; |
3115 | unsigned LoopVal = 0; |
3116 | getPhiRegs(Phi, Loop: Phi.getParent(), InitVal, LoopVal); |
3117 | SUnit *UseSU = SSD->getSUnit(MI: MRI.getVRegDef(Reg: LoopVal)); |
3118 | if (!UseSU) |
3119 | return true; |
3120 | if (UseSU->getInstr()->isPHI()) |
3121 | return true; |
3122 | unsigned LoopCycle = cycleScheduled(SU: UseSU); |
3123 | int LoopStage = stageScheduled(SU: UseSU); |
3124 | return (LoopCycle > DefCycle) || (LoopStage <= DefStage); |
3125 | } |
3126 | |
3127 | /// Return true if the instruction is a definition that is loop carried |
3128 | /// and defines the use on the next iteration. |
3129 | /// v1 = phi(v2, v3) |
3130 | /// (Def) v3 = op v1 |
3131 | /// (MO) = v1 |
3132 | /// If MO appears before Def, then v1 and v3 may get assigned to the same |
3133 | /// register. |
3134 | bool SMSchedule::isLoopCarriedDefOfUse(const SwingSchedulerDAG *SSD, |
3135 | MachineInstr *Def, |
3136 | MachineOperand &MO) const { |
3137 | if (!MO.isReg()) |
3138 | return false; |
3139 | if (Def->isPHI()) |
3140 | return false; |
3141 | MachineInstr *Phi = MRI.getVRegDef(Reg: MO.getReg()); |
3142 | if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent()) |
3143 | return false; |
3144 | if (!isLoopCarried(SSD, Phi&: *Phi)) |
3145 | return false; |
3146 | unsigned LoopReg = getLoopPhiReg(Phi: *Phi, LoopBB: Phi->getParent()); |
3147 | for (MachineOperand &DMO : Def->all_defs()) { |
3148 | if (DMO.getReg() == LoopReg) |
3149 | return true; |
3150 | } |
3151 | return false; |
3152 | } |
3153 | |
3154 | /// Return true if all scheduled predecessors are loop-carried output/order |
3155 | /// dependencies. |
3156 | bool SMSchedule::onlyHasLoopCarriedOutputOrOrderPreds( |
3157 | SUnit *SU, SwingSchedulerDAG *DAG) const { |
3158 | for (const SDep &Pred : SU->Preds) |
3159 | if (InstrToCycle.count(x: Pred.getSUnit()) && !DAG->isBackedge(Source: SU, Dep: Pred)) |
3160 | return false; |
3161 | for (const SDep &Succ : SU->Succs) |
3162 | if (InstrToCycle.count(x: Succ.getSUnit()) && DAG->isBackedge(Source: SU, Dep: Succ)) |
3163 | return false; |
3164 | return true; |
3165 | } |
3166 | |
3167 | /// Determine transitive dependences of unpipelineable instructions |
3168 | SmallSet<SUnit *, 8> SMSchedule::computeUnpipelineableNodes( |
3169 | SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { |
3170 | SmallSet<SUnit *, 8> DoNotPipeline; |
3171 | SmallVector<SUnit *, 8> Worklist; |
3172 | |
3173 | for (auto &SU : SSD->SUnits) |
3174 | if (SU.isInstr() && PLI->shouldIgnoreForPipelining(MI: SU.getInstr())) |
3175 | Worklist.push_back(Elt: &SU); |
3176 | |
3177 | while (!Worklist.empty()) { |
3178 | auto SU = Worklist.pop_back_val(); |
3179 | if (DoNotPipeline.count(Ptr: SU)) |
3180 | continue; |
3181 | LLVM_DEBUG(dbgs() << "Do not pipeline SU(" << SU->NodeNum << ")\n" ); |
3182 | DoNotPipeline.insert(Ptr: SU); |
3183 | for (auto &Dep : SU->Preds) |
3184 | Worklist.push_back(Elt: Dep.getSUnit()); |
3185 | if (SU->getInstr()->isPHI()) |
3186 | for (auto &Dep : SU->Succs) |
3187 | if (Dep.getKind() == SDep::Anti) |
3188 | Worklist.push_back(Elt: Dep.getSUnit()); |
3189 | } |
3190 | return DoNotPipeline; |
3191 | } |
3192 | |
3193 | // Determine all instructions upon which any unpipelineable instruction depends |
3194 | // and ensure that they are in stage 0. If unable to do so, return false. |
3195 | bool SMSchedule::normalizeNonPipelinedInstructions( |
3196 | SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { |
3197 | SmallSet<SUnit *, 8> DNP = computeUnpipelineableNodes(SSD, PLI); |
3198 | |
3199 | int NewLastCycle = INT_MIN; |
3200 | for (SUnit &SU : SSD->SUnits) { |
3201 | if (!SU.isInstr()) |
3202 | continue; |
3203 | if (!DNP.contains(Ptr: &SU) || stageScheduled(SU: &SU) == 0) { |
3204 | NewLastCycle = std::max(a: NewLastCycle, b: InstrToCycle[&SU]); |
3205 | continue; |
3206 | } |
3207 | |
3208 | // Put the non-pipelined instruction as early as possible in the schedule |
3209 | int NewCycle = getFirstCycle(); |
3210 | for (auto &Dep : SU.Preds) |
3211 | NewCycle = std::max(a: InstrToCycle[Dep.getSUnit()], b: NewCycle); |
3212 | |
3213 | int OldCycle = InstrToCycle[&SU]; |
3214 | if (OldCycle != NewCycle) { |
3215 | InstrToCycle[&SU] = NewCycle; |
3216 | auto &OldS = getInstructions(cycle: OldCycle); |
3217 | llvm::erase(C&: OldS, V: &SU); |
3218 | getInstructions(cycle: NewCycle).emplace_back(args: &SU); |
3219 | LLVM_DEBUG(dbgs() << "SU(" << SU.NodeNum |
3220 | << ") is not pipelined; moving from cycle " << OldCycle |
3221 | << " to " << NewCycle << " Instr:" << *SU.getInstr()); |
3222 | } |
3223 | NewLastCycle = std::max(a: NewLastCycle, b: NewCycle); |
3224 | } |
3225 | LastCycle = NewLastCycle; |
3226 | return true; |
3227 | } |
3228 | |
3229 | // Check if the generated schedule is valid. This function checks if |
3230 | // an instruction that uses a physical register is scheduled in a |
3231 | // different stage than the definition. The pipeliner does not handle |
3232 | // physical register values that may cross a basic block boundary. |
3233 | // Furthermore, if a physical def/use pair is assigned to the same |
3234 | // cycle, orderDependence does not guarantee def/use ordering, so that |
3235 | // case should be considered invalid. (The test checks for both |
3236 | // earlier and same-cycle use to be more robust.) |
3237 | bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) { |
3238 | for (SUnit &SU : SSD->SUnits) { |
3239 | if (!SU.hasPhysRegDefs) |
3240 | continue; |
3241 | int StageDef = stageScheduled(SU: &SU); |
3242 | int CycleDef = InstrToCycle[&SU]; |
3243 | assert(StageDef != -1 && "Instruction should have been scheduled." ); |
3244 | for (auto &SI : SU.Succs) |
3245 | if (SI.isAssignedRegDep() && !SI.getSUnit()->isBoundaryNode()) |
3246 | if (Register::isPhysicalRegister(Reg: SI.getReg())) { |
3247 | if (stageScheduled(SU: SI.getSUnit()) != StageDef) |
3248 | return false; |
3249 | if (InstrToCycle[SI.getSUnit()] <= CycleDef) |
3250 | return false; |
3251 | } |
3252 | } |
3253 | return true; |
3254 | } |
3255 | |
3256 | /// A property of the node order in swing-modulo-scheduling is |
3257 | /// that for nodes outside circuits the following holds: |
3258 | /// none of them is scheduled after both a successor and a |
3259 | /// predecessor. |
3260 | /// The method below checks whether the property is met. |
3261 | /// If not, debug information is printed and statistics information updated. |
3262 | /// Note that we do not use an assert statement. |
3263 | /// The reason is that although an invalid node oder may prevent |
3264 | /// the pipeliner from finding a pipelined schedule for arbitrary II, |
3265 | /// it does not lead to the generation of incorrect code. |
3266 | void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const { |
3267 | |
3268 | // a sorted vector that maps each SUnit to its index in the NodeOrder |
3269 | typedef std::pair<SUnit *, unsigned> UnitIndex; |
3270 | std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(x: nullptr, y: 0)); |
3271 | |
3272 | for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) |
3273 | Indices.push_back(x: std::make_pair(x: NodeOrder[i], y&: i)); |
3274 | |
3275 | auto CompareKey = [](UnitIndex i1, UnitIndex i2) { |
3276 | return std::get<0>(in&: i1) < std::get<0>(in&: i2); |
3277 | }; |
3278 | |
3279 | // sort, so that we can perform a binary search |
3280 | llvm::sort(C&: Indices, Comp: CompareKey); |
3281 | |
3282 | bool Valid = true; |
3283 | (void)Valid; |
3284 | // for each SUnit in the NodeOrder, check whether |
3285 | // it appears after both a successor and a predecessor |
3286 | // of the SUnit. If this is the case, and the SUnit |
3287 | // is not part of circuit, then the NodeOrder is not |
3288 | // valid. |
3289 | for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) { |
3290 | SUnit *SU = NodeOrder[i]; |
3291 | unsigned Index = i; |
3292 | |
3293 | bool PredBefore = false; |
3294 | bool SuccBefore = false; |
3295 | |
3296 | SUnit *Succ; |
3297 | SUnit *Pred; |
3298 | (void)Succ; |
3299 | (void)Pred; |
3300 | |
3301 | for (SDep &PredEdge : SU->Preds) { |
3302 | SUnit *PredSU = PredEdge.getSUnit(); |
3303 | unsigned PredIndex = std::get<1>( |
3304 | in&: *llvm::lower_bound(Range&: Indices, Value: std::make_pair(x&: PredSU, y: 0), C: CompareKey)); |
3305 | if (!PredSU->getInstr()->isPHI() && PredIndex < Index) { |
3306 | PredBefore = true; |
3307 | Pred = PredSU; |
3308 | break; |
3309 | } |
3310 | } |
3311 | |
3312 | for (SDep &SuccEdge : SU->Succs) { |
3313 | SUnit *SuccSU = SuccEdge.getSUnit(); |
3314 | // Do not process a boundary node, it was not included in NodeOrder, |
3315 | // hence not in Indices either, call to std::lower_bound() below will |
3316 | // return Indices.end(). |
3317 | if (SuccSU->isBoundaryNode()) |
3318 | continue; |
3319 | unsigned SuccIndex = std::get<1>( |
3320 | in&: *llvm::lower_bound(Range&: Indices, Value: std::make_pair(x&: SuccSU, y: 0), C: CompareKey)); |
3321 | if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) { |
3322 | SuccBefore = true; |
3323 | Succ = SuccSU; |
3324 | break; |
3325 | } |
3326 | } |
3327 | |
3328 | if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) { |
3329 | // instructions in circuits are allowed to be scheduled |
3330 | // after both a successor and predecessor. |
3331 | bool InCircuit = llvm::any_of( |
3332 | Range: Circuits, P: [SU](const NodeSet &Circuit) { return Circuit.count(SU); }); |
3333 | if (InCircuit) |
3334 | LLVM_DEBUG(dbgs() << "In a circuit, predecessor " ;); |
3335 | else { |
3336 | Valid = false; |
3337 | NumNodeOrderIssues++; |
3338 | LLVM_DEBUG(dbgs() << "Predecessor " ;); |
3339 | } |
3340 | LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum |
3341 | << " are scheduled before node " << SU->NodeNum |
3342 | << "\n" ;); |
3343 | } |
3344 | } |
3345 | |
3346 | LLVM_DEBUG({ |
3347 | if (!Valid) |
3348 | dbgs() << "Invalid node order found!\n" ; |
3349 | }); |
3350 | } |
3351 | |
3352 | /// Attempt to fix the degenerate cases when the instruction serialization |
3353 | /// causes the register lifetimes to overlap. For example, |
3354 | /// p' = store_pi(p, b) |
3355 | /// = load p, offset |
3356 | /// In this case p and p' overlap, which means that two registers are needed. |
3357 | /// Instead, this function changes the load to use p' and updates the offset. |
3358 | void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) { |
3359 | unsigned OverlapReg = 0; |
3360 | unsigned NewBaseReg = 0; |
3361 | for (SUnit *SU : Instrs) { |
3362 | MachineInstr *MI = SU->getInstr(); |
3363 | for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { |
3364 | const MachineOperand &MO = MI->getOperand(i); |
3365 | // Look for an instruction that uses p. The instruction occurs in the |
3366 | // same cycle but occurs later in the serialized order. |
3367 | if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) { |
3368 | // Check that the instruction appears in the InstrChanges structure, |
3369 | // which contains instructions that can have the offset updated. |
3370 | DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = |
3371 | InstrChanges.find(Val: SU); |
3372 | if (It != InstrChanges.end()) { |
3373 | unsigned BasePos, OffsetPos; |
3374 | // Update the base register and adjust the offset. |
3375 | if (TII->getBaseAndOffsetPosition(MI: *MI, BasePos, OffsetPos)) { |
3376 | MachineInstr *NewMI = MF.CloneMachineInstr(Orig: MI); |
3377 | NewMI->getOperand(i: BasePos).setReg(NewBaseReg); |
3378 | int64_t NewOffset = |
3379 | MI->getOperand(i: OffsetPos).getImm() - It->second.second; |
3380 | NewMI->getOperand(i: OffsetPos).setImm(NewOffset); |
3381 | SU->setInstr(NewMI); |
3382 | MISUnitMap[NewMI] = SU; |
3383 | NewMIs[MI] = NewMI; |
3384 | } |
3385 | } |
3386 | OverlapReg = 0; |
3387 | NewBaseReg = 0; |
3388 | break; |
3389 | } |
3390 | // Look for an instruction of the form p' = op(p), which uses and defines |
3391 | // two virtual registers that get allocated to the same physical register. |
3392 | unsigned TiedUseIdx = 0; |
3393 | if (MI->isRegTiedToUseOperand(DefOpIdx: i, UseOpIdx: &TiedUseIdx)) { |
3394 | // OverlapReg is p in the example above. |
3395 | OverlapReg = MI->getOperand(i: TiedUseIdx).getReg(); |
3396 | // NewBaseReg is p' in the example above. |
3397 | NewBaseReg = MI->getOperand(i).getReg(); |
3398 | break; |
3399 | } |
3400 | } |
3401 | } |
3402 | } |
3403 | |
3404 | std::deque<SUnit *> |
3405 | SMSchedule::reorderInstructions(const SwingSchedulerDAG *SSD, |
3406 | const std::deque<SUnit *> &Instrs) const { |
3407 | std::deque<SUnit *> NewOrderPhi; |
3408 | for (SUnit *SU : Instrs) { |
3409 | if (SU->getInstr()->isPHI()) |
3410 | NewOrderPhi.push_back(x: SU); |
3411 | } |
3412 | std::deque<SUnit *> NewOrderI; |
3413 | for (SUnit *SU : Instrs) { |
3414 | if (!SU->getInstr()->isPHI()) |
3415 | orderDependence(SSD, SU, Insts&: NewOrderI); |
3416 | } |
3417 | llvm::append_range(C&: NewOrderPhi, R&: NewOrderI); |
3418 | return NewOrderPhi; |
3419 | } |
3420 | |
3421 | /// After the schedule has been formed, call this function to combine |
3422 | /// the instructions from the different stages/cycles. That is, this |
3423 | /// function creates a schedule that represents a single iteration. |
3424 | void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) { |
3425 | // Move all instructions to the first stage from later stages. |
3426 | for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { |
3427 | for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage; |
3428 | ++stage) { |
3429 | std::deque<SUnit *> &cycleInstrs = |
3430 | ScheduledInstrs[cycle + (stage * InitiationInterval)]; |
3431 | for (SUnit *SU : llvm::reverse(C&: cycleInstrs)) |
3432 | ScheduledInstrs[cycle].push_front(x: SU); |
3433 | } |
3434 | } |
3435 | |
3436 | // Erase all the elements in the later stages. Only one iteration should |
3437 | // remain in the scheduled list, and it contains all the instructions. |
3438 | for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle) |
3439 | ScheduledInstrs.erase(Val: cycle); |
3440 | |
3441 | // Change the registers in instruction as specified in the InstrChanges |
3442 | // map. We need to use the new registers to create the correct order. |
3443 | for (const SUnit &SU : SSD->SUnits) |
3444 | SSD->applyInstrChange(MI: SU.getInstr(), Schedule&: *this); |
3445 | |
3446 | // Reorder the instructions in each cycle to fix and improve the |
3447 | // generated code. |
3448 | for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) { |
3449 | std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle]; |
3450 | cycleInstrs = reorderInstructions(SSD, Instrs: cycleInstrs); |
3451 | SSD->fixupRegisterOverlaps(Instrs&: cycleInstrs); |
3452 | } |
3453 | |
3454 | LLVM_DEBUG(dump();); |
3455 | } |
3456 | |
3457 | void NodeSet::print(raw_ostream &os) const { |
3458 | os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV |
3459 | << " depth " << MaxDepth << " col " << Colocate << "\n" ; |
3460 | for (const auto &I : Nodes) |
3461 | os << " SU(" << I->NodeNum << ") " << *(I->getInstr()); |
3462 | os << "\n" ; |
3463 | } |
3464 | |
3465 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
3466 | /// Print the schedule information to the given output. |
3467 | void SMSchedule::print(raw_ostream &os) const { |
3468 | // Iterate over each cycle. |
3469 | for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { |
3470 | // Iterate over each instruction in the cycle. |
3471 | const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle); |
3472 | for (SUnit *CI : cycleInstrs->second) { |
3473 | os << "cycle " << cycle << " (" << stageScheduled(CI) << ") " ; |
3474 | os << "(" << CI->NodeNum << ") " ; |
3475 | CI->getInstr()->print(os); |
3476 | os << "\n" ; |
3477 | } |
3478 | } |
3479 | } |
3480 | |
3481 | /// Utility function used for debugging to print the schedule. |
3482 | LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); } |
3483 | LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); } |
3484 | |
3485 | void ResourceManager::dumpMRT() const { |
3486 | LLVM_DEBUG({ |
3487 | if (UseDFA) |
3488 | return; |
3489 | std::stringstream SS; |
3490 | SS << "MRT:\n" ; |
3491 | SS << std::setw(4) << "Slot" ; |
3492 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) |
3493 | SS << std::setw(3) << I; |
3494 | SS << std::setw(7) << "#Mops" |
3495 | << "\n" ; |
3496 | for (int Slot = 0; Slot < InitiationInterval; ++Slot) { |
3497 | SS << std::setw(4) << Slot; |
3498 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) |
3499 | SS << std::setw(3) << MRT[Slot][I]; |
3500 | SS << std::setw(7) << NumScheduledMops[Slot] << "\n" ; |
3501 | } |
3502 | dbgs() << SS.str(); |
3503 | }); |
3504 | } |
3505 | #endif |
3506 | |
3507 | void ResourceManager::initProcResourceVectors( |
3508 | const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) { |
3509 | unsigned ProcResourceID = 0; |
3510 | |
3511 | // We currently limit the resource kinds to 64 and below so that we can use |
3512 | // uint64_t for Masks |
3513 | assert(SM.getNumProcResourceKinds() < 64 && |
3514 | "Too many kinds of resources, unsupported" ); |
3515 | // Create a unique bitmask for every processor resource unit. |
3516 | // Skip resource at index 0, since it always references 'InvalidUnit'. |
3517 | Masks.resize(N: SM.getNumProcResourceKinds()); |
3518 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { |
3519 | const MCProcResourceDesc &Desc = *SM.getProcResource(ProcResourceIdx: I); |
3520 | if (Desc.SubUnitsIdxBegin) |
3521 | continue; |
3522 | Masks[I] = 1ULL << ProcResourceID; |
3523 | ProcResourceID++; |
3524 | } |
3525 | // Create a unique bitmask for every processor resource group. |
3526 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { |
3527 | const MCProcResourceDesc &Desc = *SM.getProcResource(ProcResourceIdx: I); |
3528 | if (!Desc.SubUnitsIdxBegin) |
3529 | continue; |
3530 | Masks[I] = 1ULL << ProcResourceID; |
3531 | for (unsigned U = 0; U < Desc.NumUnits; ++U) |
3532 | Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]]; |
3533 | ProcResourceID++; |
3534 | } |
3535 | LLVM_DEBUG({ |
3536 | if (SwpShowResMask) { |
3537 | dbgs() << "ProcResourceDesc:\n" ; |
3538 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { |
3539 | const MCProcResourceDesc *ProcResource = SM.getProcResource(I); |
3540 | dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n" , |
3541 | ProcResource->Name, I, Masks[I], |
3542 | ProcResource->NumUnits); |
3543 | } |
3544 | dbgs() << " -----------------\n" ; |
3545 | } |
3546 | }); |
3547 | } |
3548 | |
3549 | bool ResourceManager::canReserveResources(SUnit &SU, int Cycle) { |
3550 | LLVM_DEBUG({ |
3551 | if (SwpDebugResource) |
3552 | dbgs() << "canReserveResources:\n" ; |
3553 | }); |
3554 | if (UseDFA) |
3555 | return DFAResources[positiveModulo(Dividend: Cycle, Divisor: InitiationInterval)] |
3556 | ->canReserveResources(MID: &SU.getInstr()->getDesc()); |
3557 | |
3558 | const MCSchedClassDesc *SCDesc = DAG->getSchedClass(SU: &SU); |
3559 | if (!SCDesc->isValid()) { |
3560 | LLVM_DEBUG({ |
3561 | dbgs() << "No valid Schedule Class Desc for schedClass!\n" ; |
3562 | dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n" ; |
3563 | }); |
3564 | return true; |
3565 | } |
3566 | |
3567 | reserveResources(SCDesc, Cycle); |
3568 | bool Result = !isOverbooked(); |
3569 | unreserveResources(SCDesc, Cycle); |
3570 | |
3571 | LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return " << Result << "\n\n" ;); |
3572 | return Result; |
3573 | } |
3574 | |
3575 | void ResourceManager::reserveResources(SUnit &SU, int Cycle) { |
3576 | LLVM_DEBUG({ |
3577 | if (SwpDebugResource) |
3578 | dbgs() << "reserveResources:\n" ; |
3579 | }); |
3580 | if (UseDFA) |
3581 | return DFAResources[positiveModulo(Dividend: Cycle, Divisor: InitiationInterval)] |
3582 | ->reserveResources(MID: &SU.getInstr()->getDesc()); |
3583 | |
3584 | const MCSchedClassDesc *SCDesc = DAG->getSchedClass(SU: &SU); |
3585 | if (!SCDesc->isValid()) { |
3586 | LLVM_DEBUG({ |
3587 | dbgs() << "No valid Schedule Class Desc for schedClass!\n" ; |
3588 | dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n" ; |
3589 | }); |
3590 | return; |
3591 | } |
3592 | |
3593 | reserveResources(SCDesc, Cycle); |
3594 | |
3595 | LLVM_DEBUG({ |
3596 | if (SwpDebugResource) { |
3597 | dumpMRT(); |
3598 | dbgs() << "reserveResources: done!\n\n" ; |
3599 | } |
3600 | }); |
3601 | } |
3602 | |
3603 | void ResourceManager::reserveResources(const MCSchedClassDesc *SCDesc, |
3604 | int Cycle) { |
3605 | assert(!UseDFA); |
3606 | for (const MCWriteProcResEntry &PRE : make_range( |
3607 | x: STI->getWriteProcResBegin(SC: SCDesc), y: STI->getWriteProcResEnd(SC: SCDesc))) |
3608 | for (int C = Cycle; C < Cycle + PRE.ReleaseAtCycle; ++C) |
3609 | ++MRT[positiveModulo(Dividend: C, Divisor: InitiationInterval)][PRE.ProcResourceIdx]; |
3610 | |
3611 | for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) |
3612 | ++NumScheduledMops[positiveModulo(Dividend: C, Divisor: InitiationInterval)]; |
3613 | } |
3614 | |
3615 | void ResourceManager::unreserveResources(const MCSchedClassDesc *SCDesc, |
3616 | int Cycle) { |
3617 | assert(!UseDFA); |
3618 | for (const MCWriteProcResEntry &PRE : make_range( |
3619 | x: STI->getWriteProcResBegin(SC: SCDesc), y: STI->getWriteProcResEnd(SC: SCDesc))) |
3620 | for (int C = Cycle; C < Cycle + PRE.ReleaseAtCycle; ++C) |
3621 | --MRT[positiveModulo(Dividend: C, Divisor: InitiationInterval)][PRE.ProcResourceIdx]; |
3622 | |
3623 | for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) |
3624 | --NumScheduledMops[positiveModulo(Dividend: C, Divisor: InitiationInterval)]; |
3625 | } |
3626 | |
3627 | bool ResourceManager::isOverbooked() const { |
3628 | assert(!UseDFA); |
3629 | for (int Slot = 0; Slot < InitiationInterval; ++Slot) { |
3630 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { |
3631 | const MCProcResourceDesc *Desc = SM.getProcResource(ProcResourceIdx: I); |
3632 | if (MRT[Slot][I] > Desc->NumUnits) |
3633 | return true; |
3634 | } |
3635 | if (NumScheduledMops[Slot] > IssueWidth) |
3636 | return true; |
3637 | } |
3638 | return false; |
3639 | } |
3640 | |
3641 | int ResourceManager::calculateResMIIDFA() const { |
3642 | assert(UseDFA); |
3643 | |
3644 | // Sort the instructions by the number of available choices for scheduling, |
3645 | // least to most. Use the number of critical resources as the tie breaker. |
3646 | FuncUnitSorter FUS = FuncUnitSorter(*ST); |
3647 | for (SUnit &SU : DAG->SUnits) |
3648 | FUS.calcCriticalResources(MI&: *SU.getInstr()); |
3649 | PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter> |
3650 | FuncUnitOrder(FUS); |
3651 | |
3652 | for (SUnit &SU : DAG->SUnits) |
3653 | FuncUnitOrder.push(x: SU.getInstr()); |
3654 | |
3655 | SmallVector<std::unique_ptr<DFAPacketizer>, 8> Resources; |
3656 | Resources.push_back( |
3657 | Elt: std::unique_ptr<DFAPacketizer>(TII->CreateTargetScheduleState(*ST))); |
3658 | |
3659 | while (!FuncUnitOrder.empty()) { |
3660 | MachineInstr *MI = FuncUnitOrder.top(); |
3661 | FuncUnitOrder.pop(); |
3662 | if (TII->isZeroCost(Opcode: MI->getOpcode())) |
3663 | continue; |
3664 | |
3665 | // Attempt to reserve the instruction in an existing DFA. At least one |
3666 | // DFA is needed for each cycle. |
3667 | unsigned NumCycles = DAG->getSUnit(MI)->Latency; |
3668 | unsigned ReservedCycles = 0; |
3669 | auto *RI = Resources.begin(); |
3670 | auto *RE = Resources.end(); |
3671 | LLVM_DEBUG({ |
3672 | dbgs() << "Trying to reserve resource for " << NumCycles |
3673 | << " cycles for \n" ; |
3674 | MI->dump(); |
3675 | }); |
3676 | for (unsigned C = 0; C < NumCycles; ++C) |
3677 | while (RI != RE) { |
3678 | if ((*RI)->canReserveResources(MI&: *MI)) { |
3679 | (*RI)->reserveResources(MI&: *MI); |
3680 | ++ReservedCycles; |
3681 | break; |
3682 | } |
3683 | RI++; |
3684 | } |
3685 | LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles |
3686 | << ", NumCycles:" << NumCycles << "\n" ); |
3687 | // Add new DFAs, if needed, to reserve resources. |
3688 | for (unsigned C = ReservedCycles; C < NumCycles; ++C) { |
3689 | LLVM_DEBUG(if (SwpDebugResource) dbgs() |
3690 | << "NewResource created to reserve resources" |
3691 | << "\n" ); |
3692 | auto *NewResource = TII->CreateTargetScheduleState(*ST); |
3693 | assert(NewResource->canReserveResources(*MI) && "Reserve error." ); |
3694 | NewResource->reserveResources(MI&: *MI); |
3695 | Resources.push_back(Elt: std::unique_ptr<DFAPacketizer>(NewResource)); |
3696 | } |
3697 | } |
3698 | |
3699 | int Resmii = Resources.size(); |
3700 | LLVM_DEBUG(dbgs() << "Return Res MII:" << Resmii << "\n" ); |
3701 | return Resmii; |
3702 | } |
3703 | |
3704 | int ResourceManager::calculateResMII() const { |
3705 | if (UseDFA) |
3706 | return calculateResMIIDFA(); |
3707 | |
3708 | // Count each resource consumption and divide it by the number of units. |
3709 | // ResMII is the max value among them. |
3710 | |
3711 | int NumMops = 0; |
3712 | SmallVector<uint64_t> ResourceCount(SM.getNumProcResourceKinds()); |
3713 | for (SUnit &SU : DAG->SUnits) { |
3714 | if (TII->isZeroCost(Opcode: SU.getInstr()->getOpcode())) |
3715 | continue; |
3716 | |
3717 | const MCSchedClassDesc *SCDesc = DAG->getSchedClass(SU: &SU); |
3718 | if (!SCDesc->isValid()) |
3719 | continue; |
3720 | |
3721 | LLVM_DEBUG({ |
3722 | if (SwpDebugResource) { |
3723 | DAG->dumpNode(SU); |
3724 | dbgs() << " #Mops: " << SCDesc->NumMicroOps << "\n" |
3725 | << " WriteProcRes: " ; |
3726 | } |
3727 | }); |
3728 | NumMops += SCDesc->NumMicroOps; |
3729 | for (const MCWriteProcResEntry &PRE : |
3730 | make_range(x: STI->getWriteProcResBegin(SC: SCDesc), |
3731 | y: STI->getWriteProcResEnd(SC: SCDesc))) { |
3732 | LLVM_DEBUG({ |
3733 | if (SwpDebugResource) { |
3734 | const MCProcResourceDesc *Desc = |
3735 | SM.getProcResource(PRE.ProcResourceIdx); |
3736 | dbgs() << Desc->Name << ": " << PRE.ReleaseAtCycle << ", " ; |
3737 | } |
3738 | }); |
3739 | ResourceCount[PRE.ProcResourceIdx] += PRE.ReleaseAtCycle; |
3740 | } |
3741 | LLVM_DEBUG(if (SwpDebugResource) dbgs() << "\n" ); |
3742 | } |
3743 | |
3744 | int Result = (NumMops + IssueWidth - 1) / IssueWidth; |
3745 | LLVM_DEBUG({ |
3746 | if (SwpDebugResource) |
3747 | dbgs() << "#Mops: " << NumMops << ", " |
3748 | << "IssueWidth: " << IssueWidth << ", " |
3749 | << "Cycles: " << Result << "\n" ; |
3750 | }); |
3751 | |
3752 | LLVM_DEBUG({ |
3753 | if (SwpDebugResource) { |
3754 | std::stringstream SS; |
3755 | SS << std::setw(2) << "ID" << std::setw(16) << "Name" << std::setw(10) |
3756 | << "Units" << std::setw(10) << "Consumed" << std::setw(10) << "Cycles" |
3757 | << "\n" ; |
3758 | dbgs() << SS.str(); |
3759 | } |
3760 | }); |
3761 | for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { |
3762 | const MCProcResourceDesc *Desc = SM.getProcResource(ProcResourceIdx: I); |
3763 | int Cycles = (ResourceCount[I] + Desc->NumUnits - 1) / Desc->NumUnits; |
3764 | LLVM_DEBUG({ |
3765 | if (SwpDebugResource) { |
3766 | std::stringstream SS; |
3767 | SS << std::setw(2) << I << std::setw(16) << Desc->Name << std::setw(10) |
3768 | << Desc->NumUnits << std::setw(10) << ResourceCount[I] |
3769 | << std::setw(10) << Cycles << "\n" ; |
3770 | dbgs() << SS.str(); |
3771 | } |
3772 | }); |
3773 | if (Cycles > Result) |
3774 | Result = Cycles; |
3775 | } |
3776 | return Result; |
3777 | } |
3778 | |
3779 | void ResourceManager::init(int II) { |
3780 | InitiationInterval = II; |
3781 | DFAResources.clear(); |
3782 | DFAResources.resize(N: II); |
3783 | for (auto &I : DFAResources) |
3784 | I.reset(p: ST->getInstrInfo()->CreateTargetScheduleState(*ST)); |
3785 | MRT.clear(); |
3786 | MRT.resize(N: II, NV: SmallVector<uint64_t>(SM.getNumProcResourceKinds())); |
3787 | NumScheduledMops.clear(); |
3788 | NumScheduledMops.resize(N: II); |
3789 | } |
3790 | |