1 //===- AArch64FalkorHWPFFix.cpp - Avoid HW prefetcher pitfalls on Falkor --===//
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 /// \file For Falkor, we want to avoid HW prefetcher instruction tag collisions
9 /// that may inhibit the HW prefetching.  This is done in two steps.  Before
10 /// ISel, we mark strided loads (i.e. those that will likely benefit from
11 /// prefetching) with metadata.  Then, after opcodes have been finalized, we
12 /// insert MOVs and re-write loads to prevent unintentional tag collisions.
13 // ===---------------------------------------------------------------------===//
14 
15 #include "AArch64.h"
16 #include "AArch64InstrInfo.h"
17 #include "AArch64Subtarget.h"
18 #include "AArch64TargetMachine.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/None.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/ScalarEvolution.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/CodeGen/LiveRegUnits.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineFunctionPass.h"
32 #include "llvm/CodeGen/MachineInstr.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineOperand.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/TargetPassConfig.h"
38 #include "llvm/CodeGen/TargetRegisterInfo.h"
39 #include "llvm/IR/DebugLoc.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/InitializePasses.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/DebugCounter.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include <cassert>
52 #include <iterator>
53 #include <utility>
54 
55 using namespace llvm;
56 
57 #define DEBUG_TYPE "aarch64-falkor-hwpf-fix"
58 
59 STATISTIC(NumStridedLoadsMarked, "Number of strided loads marked");
60 STATISTIC(NumCollisionsAvoided,
61           "Number of HW prefetch tag collisions avoided");
62 STATISTIC(NumCollisionsNotAvoided,
63           "Number of HW prefetch tag collisions not avoided due to lack of registers");
64 DEBUG_COUNTER(FixCounter, "falkor-hwpf",
65               "Controls which tag collisions are avoided");
66 
67 namespace {
68 
69 class FalkorMarkStridedAccesses {
70 public:
FalkorMarkStridedAccesses(LoopInfo & LI,ScalarEvolution & SE)71   FalkorMarkStridedAccesses(LoopInfo &LI, ScalarEvolution &SE)
72       : LI(LI), SE(SE) {}
73 
74   bool run();
75 
76 private:
77   bool runOnLoop(Loop &L);
78 
79   LoopInfo &LI;
80   ScalarEvolution &SE;
81 };
82 
83 class FalkorMarkStridedAccessesLegacy : public FunctionPass {
84 public:
85   static char ID; // Pass ID, replacement for typeid
86 
FalkorMarkStridedAccessesLegacy()87   FalkorMarkStridedAccessesLegacy() : FunctionPass(ID) {
88     initializeFalkorMarkStridedAccessesLegacyPass(
89         *PassRegistry::getPassRegistry());
90   }
91 
getAnalysisUsage(AnalysisUsage & AU) const92   void getAnalysisUsage(AnalysisUsage &AU) const override {
93     AU.addRequired<TargetPassConfig>();
94     AU.addPreserved<DominatorTreeWrapperPass>();
95     AU.addRequired<LoopInfoWrapperPass>();
96     AU.addPreserved<LoopInfoWrapperPass>();
97     AU.addRequired<ScalarEvolutionWrapperPass>();
98     AU.addPreserved<ScalarEvolutionWrapperPass>();
99   }
100 
101   bool runOnFunction(Function &F) override;
102 };
103 
104 } // end anonymous namespace
105 
106 char FalkorMarkStridedAccessesLegacy::ID = 0;
107 
108 INITIALIZE_PASS_BEGIN(FalkorMarkStridedAccessesLegacy, DEBUG_TYPE,
109                       "Falkor HW Prefetch Fix", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)110 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
111 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
112 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
113 INITIALIZE_PASS_END(FalkorMarkStridedAccessesLegacy, DEBUG_TYPE,
114                     "Falkor HW Prefetch Fix", false, false)
115 
116 FunctionPass *llvm::createFalkorMarkStridedAccessesPass() {
117   return new FalkorMarkStridedAccessesLegacy();
118 }
119 
runOnFunction(Function & F)120 bool FalkorMarkStridedAccessesLegacy::runOnFunction(Function &F) {
121   TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
122   const AArch64Subtarget *ST =
123       TPC.getTM<AArch64TargetMachine>().getSubtargetImpl(F);
124   if (ST->getProcFamily() != AArch64Subtarget::Falkor)
125     return false;
126 
127   if (skipFunction(F))
128     return false;
129 
130   LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
131   ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
132 
133   FalkorMarkStridedAccesses LDP(LI, SE);
134   return LDP.run();
135 }
136 
run()137 bool FalkorMarkStridedAccesses::run() {
138   bool MadeChange = false;
139 
140   for (Loop *L : LI)
141     for (auto LIt = df_begin(L), LE = df_end(L); LIt != LE; ++LIt)
142       MadeChange |= runOnLoop(**LIt);
143 
144   return MadeChange;
145 }
146 
runOnLoop(Loop & L)147 bool FalkorMarkStridedAccesses::runOnLoop(Loop &L) {
148   // Only mark strided loads in the inner-most loop
149   if (!L.isInnermost())
150     return false;
151 
152   bool MadeChange = false;
153 
154   for (BasicBlock *BB : L.blocks()) {
155     for (Instruction &I : *BB) {
156       LoadInst *LoadI = dyn_cast<LoadInst>(&I);
157       if (!LoadI)
158         continue;
159 
160       Value *PtrValue = LoadI->getPointerOperand();
161       if (L.isLoopInvariant(PtrValue))
162         continue;
163 
164       const SCEV *LSCEV = SE.getSCEV(PtrValue);
165       const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
166       if (!LSCEVAddRec || !LSCEVAddRec->isAffine())
167         continue;
168 
169       LoadI->setMetadata(FALKOR_STRIDED_ACCESS_MD,
170                          MDNode::get(LoadI->getContext(), {}));
171       ++NumStridedLoadsMarked;
172       LLVM_DEBUG(dbgs() << "Load: " << I << " marked as strided\n");
173       MadeChange = true;
174     }
175   }
176 
177   return MadeChange;
178 }
179 
180 namespace {
181 
182 class FalkorHWPFFix : public MachineFunctionPass {
183 public:
184   static char ID;
185 
FalkorHWPFFix()186   FalkorHWPFFix() : MachineFunctionPass(ID) {
187     initializeFalkorHWPFFixPass(*PassRegistry::getPassRegistry());
188   }
189 
190   bool runOnMachineFunction(MachineFunction &Fn) override;
191 
getAnalysisUsage(AnalysisUsage & AU) const192   void getAnalysisUsage(AnalysisUsage &AU) const override {
193     AU.setPreservesCFG();
194     AU.addRequired<MachineLoopInfo>();
195     MachineFunctionPass::getAnalysisUsage(AU);
196   }
197 
getRequiredProperties() const198   MachineFunctionProperties getRequiredProperties() const override {
199     return MachineFunctionProperties().set(
200         MachineFunctionProperties::Property::NoVRegs);
201   }
202 
203 private:
204   void runOnLoop(MachineLoop &L, MachineFunction &Fn);
205 
206   const AArch64InstrInfo *TII;
207   const TargetRegisterInfo *TRI;
208   DenseMap<unsigned, SmallVector<MachineInstr *, 4>> TagMap;
209   bool Modified;
210 };
211 
212 /// Bits from load opcodes used to compute HW prefetcher instruction tags.
213 struct LoadInfo {
214   LoadInfo() = default;
215 
216   Register DestReg;
217   Register BaseReg;
218   int BaseRegIdx = -1;
219   const MachineOperand *OffsetOpnd = nullptr;
220   bool IsPrePost = false;
221 };
222 
223 } // end anonymous namespace
224 
225 char FalkorHWPFFix::ID = 0;
226 
227 INITIALIZE_PASS_BEGIN(FalkorHWPFFix, "aarch64-falkor-hwpf-fix-late",
228                       "Falkor HW Prefetch Fix Late Phase", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)229 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
230 INITIALIZE_PASS_END(FalkorHWPFFix, "aarch64-falkor-hwpf-fix-late",
231                     "Falkor HW Prefetch Fix Late Phase", false, false)
232 
233 static unsigned makeTag(unsigned Dest, unsigned Base, unsigned Offset) {
234   return (Dest & 0xf) | ((Base & 0xf) << 4) | ((Offset & 0x3f) << 8);
235 }
236 
getLoadInfo(const MachineInstr & MI)237 static Optional<LoadInfo> getLoadInfo(const MachineInstr &MI) {
238   int DestRegIdx;
239   int BaseRegIdx;
240   int OffsetIdx;
241   bool IsPrePost;
242 
243   switch (MI.getOpcode()) {
244   default:
245     return None;
246 
247   case AArch64::LD1i64:
248   case AArch64::LD2i64:
249     DestRegIdx = 0;
250     BaseRegIdx = 3;
251     OffsetIdx = -1;
252     IsPrePost = false;
253     break;
254 
255   case AArch64::LD1i8:
256   case AArch64::LD1i16:
257   case AArch64::LD1i32:
258   case AArch64::LD2i8:
259   case AArch64::LD2i16:
260   case AArch64::LD2i32:
261   case AArch64::LD3i8:
262   case AArch64::LD3i16:
263   case AArch64::LD3i32:
264   case AArch64::LD3i64:
265   case AArch64::LD4i8:
266   case AArch64::LD4i16:
267   case AArch64::LD4i32:
268   case AArch64::LD4i64:
269     DestRegIdx = -1;
270     BaseRegIdx = 3;
271     OffsetIdx = -1;
272     IsPrePost = false;
273     break;
274 
275   case AArch64::LD1Onev1d:
276   case AArch64::LD1Onev2s:
277   case AArch64::LD1Onev4h:
278   case AArch64::LD1Onev8b:
279   case AArch64::LD1Onev2d:
280   case AArch64::LD1Onev4s:
281   case AArch64::LD1Onev8h:
282   case AArch64::LD1Onev16b:
283   case AArch64::LD1Rv1d:
284   case AArch64::LD1Rv2s:
285   case AArch64::LD1Rv4h:
286   case AArch64::LD1Rv8b:
287   case AArch64::LD1Rv2d:
288   case AArch64::LD1Rv4s:
289   case AArch64::LD1Rv8h:
290   case AArch64::LD1Rv16b:
291     DestRegIdx = 0;
292     BaseRegIdx = 1;
293     OffsetIdx = -1;
294     IsPrePost = false;
295     break;
296 
297   case AArch64::LD1Twov1d:
298   case AArch64::LD1Twov2s:
299   case AArch64::LD1Twov4h:
300   case AArch64::LD1Twov8b:
301   case AArch64::LD1Twov2d:
302   case AArch64::LD1Twov4s:
303   case AArch64::LD1Twov8h:
304   case AArch64::LD1Twov16b:
305   case AArch64::LD1Threev1d:
306   case AArch64::LD1Threev2s:
307   case AArch64::LD1Threev4h:
308   case AArch64::LD1Threev8b:
309   case AArch64::LD1Threev2d:
310   case AArch64::LD1Threev4s:
311   case AArch64::LD1Threev8h:
312   case AArch64::LD1Threev16b:
313   case AArch64::LD1Fourv1d:
314   case AArch64::LD1Fourv2s:
315   case AArch64::LD1Fourv4h:
316   case AArch64::LD1Fourv8b:
317   case AArch64::LD1Fourv2d:
318   case AArch64::LD1Fourv4s:
319   case AArch64::LD1Fourv8h:
320   case AArch64::LD1Fourv16b:
321   case AArch64::LD2Twov2s:
322   case AArch64::LD2Twov4s:
323   case AArch64::LD2Twov8b:
324   case AArch64::LD2Twov2d:
325   case AArch64::LD2Twov4h:
326   case AArch64::LD2Twov8h:
327   case AArch64::LD2Twov16b:
328   case AArch64::LD2Rv1d:
329   case AArch64::LD2Rv2s:
330   case AArch64::LD2Rv4s:
331   case AArch64::LD2Rv8b:
332   case AArch64::LD2Rv2d:
333   case AArch64::LD2Rv4h:
334   case AArch64::LD2Rv8h:
335   case AArch64::LD2Rv16b:
336   case AArch64::LD3Threev2s:
337   case AArch64::LD3Threev4h:
338   case AArch64::LD3Threev8b:
339   case AArch64::LD3Threev2d:
340   case AArch64::LD3Threev4s:
341   case AArch64::LD3Threev8h:
342   case AArch64::LD3Threev16b:
343   case AArch64::LD3Rv1d:
344   case AArch64::LD3Rv2s:
345   case AArch64::LD3Rv4h:
346   case AArch64::LD3Rv8b:
347   case AArch64::LD3Rv2d:
348   case AArch64::LD3Rv4s:
349   case AArch64::LD3Rv8h:
350   case AArch64::LD3Rv16b:
351   case AArch64::LD4Fourv2s:
352   case AArch64::LD4Fourv4h:
353   case AArch64::LD4Fourv8b:
354   case AArch64::LD4Fourv2d:
355   case AArch64::LD4Fourv4s:
356   case AArch64::LD4Fourv8h:
357   case AArch64::LD4Fourv16b:
358   case AArch64::LD4Rv1d:
359   case AArch64::LD4Rv2s:
360   case AArch64::LD4Rv4h:
361   case AArch64::LD4Rv8b:
362   case AArch64::LD4Rv2d:
363   case AArch64::LD4Rv4s:
364   case AArch64::LD4Rv8h:
365   case AArch64::LD4Rv16b:
366     DestRegIdx = -1;
367     BaseRegIdx = 1;
368     OffsetIdx = -1;
369     IsPrePost = false;
370     break;
371 
372   case AArch64::LD1i64_POST:
373   case AArch64::LD2i64_POST:
374     DestRegIdx = 1;
375     BaseRegIdx = 4;
376     OffsetIdx = 5;
377     IsPrePost = true;
378     break;
379 
380   case AArch64::LD1i8_POST:
381   case AArch64::LD1i16_POST:
382   case AArch64::LD1i32_POST:
383   case AArch64::LD2i8_POST:
384   case AArch64::LD2i16_POST:
385   case AArch64::LD2i32_POST:
386   case AArch64::LD3i8_POST:
387   case AArch64::LD3i16_POST:
388   case AArch64::LD3i32_POST:
389   case AArch64::LD3i64_POST:
390   case AArch64::LD4i8_POST:
391   case AArch64::LD4i16_POST:
392   case AArch64::LD4i32_POST:
393   case AArch64::LD4i64_POST:
394     DestRegIdx = -1;
395     BaseRegIdx = 4;
396     OffsetIdx = 5;
397     IsPrePost = true;
398     break;
399 
400   case AArch64::LD1Onev1d_POST:
401   case AArch64::LD1Onev2s_POST:
402   case AArch64::LD1Onev4h_POST:
403   case AArch64::LD1Onev8b_POST:
404   case AArch64::LD1Onev2d_POST:
405   case AArch64::LD1Onev4s_POST:
406   case AArch64::LD1Onev8h_POST:
407   case AArch64::LD1Onev16b_POST:
408   case AArch64::LD1Rv1d_POST:
409   case AArch64::LD1Rv2s_POST:
410   case AArch64::LD1Rv4h_POST:
411   case AArch64::LD1Rv8b_POST:
412   case AArch64::LD1Rv2d_POST:
413   case AArch64::LD1Rv4s_POST:
414   case AArch64::LD1Rv8h_POST:
415   case AArch64::LD1Rv16b_POST:
416     DestRegIdx = 1;
417     BaseRegIdx = 2;
418     OffsetIdx = 3;
419     IsPrePost = true;
420     break;
421 
422   case AArch64::LD1Twov1d_POST:
423   case AArch64::LD1Twov2s_POST:
424   case AArch64::LD1Twov4h_POST:
425   case AArch64::LD1Twov8b_POST:
426   case AArch64::LD1Twov2d_POST:
427   case AArch64::LD1Twov4s_POST:
428   case AArch64::LD1Twov8h_POST:
429   case AArch64::LD1Twov16b_POST:
430   case AArch64::LD1Threev1d_POST:
431   case AArch64::LD1Threev2s_POST:
432   case AArch64::LD1Threev4h_POST:
433   case AArch64::LD1Threev8b_POST:
434   case AArch64::LD1Threev2d_POST:
435   case AArch64::LD1Threev4s_POST:
436   case AArch64::LD1Threev8h_POST:
437   case AArch64::LD1Threev16b_POST:
438   case AArch64::LD1Fourv1d_POST:
439   case AArch64::LD1Fourv2s_POST:
440   case AArch64::LD1Fourv4h_POST:
441   case AArch64::LD1Fourv8b_POST:
442   case AArch64::LD1Fourv2d_POST:
443   case AArch64::LD1Fourv4s_POST:
444   case AArch64::LD1Fourv8h_POST:
445   case AArch64::LD1Fourv16b_POST:
446   case AArch64::LD2Twov2s_POST:
447   case AArch64::LD2Twov4s_POST:
448   case AArch64::LD2Twov8b_POST:
449   case AArch64::LD2Twov2d_POST:
450   case AArch64::LD2Twov4h_POST:
451   case AArch64::LD2Twov8h_POST:
452   case AArch64::LD2Twov16b_POST:
453   case AArch64::LD2Rv1d_POST:
454   case AArch64::LD2Rv2s_POST:
455   case AArch64::LD2Rv4s_POST:
456   case AArch64::LD2Rv8b_POST:
457   case AArch64::LD2Rv2d_POST:
458   case AArch64::LD2Rv4h_POST:
459   case AArch64::LD2Rv8h_POST:
460   case AArch64::LD2Rv16b_POST:
461   case AArch64::LD3Threev2s_POST:
462   case AArch64::LD3Threev4h_POST:
463   case AArch64::LD3Threev8b_POST:
464   case AArch64::LD3Threev2d_POST:
465   case AArch64::LD3Threev4s_POST:
466   case AArch64::LD3Threev8h_POST:
467   case AArch64::LD3Threev16b_POST:
468   case AArch64::LD3Rv1d_POST:
469   case AArch64::LD3Rv2s_POST:
470   case AArch64::LD3Rv4h_POST:
471   case AArch64::LD3Rv8b_POST:
472   case AArch64::LD3Rv2d_POST:
473   case AArch64::LD3Rv4s_POST:
474   case AArch64::LD3Rv8h_POST:
475   case AArch64::LD3Rv16b_POST:
476   case AArch64::LD4Fourv2s_POST:
477   case AArch64::LD4Fourv4h_POST:
478   case AArch64::LD4Fourv8b_POST:
479   case AArch64::LD4Fourv2d_POST:
480   case AArch64::LD4Fourv4s_POST:
481   case AArch64::LD4Fourv8h_POST:
482   case AArch64::LD4Fourv16b_POST:
483   case AArch64::LD4Rv1d_POST:
484   case AArch64::LD4Rv2s_POST:
485   case AArch64::LD4Rv4h_POST:
486   case AArch64::LD4Rv8b_POST:
487   case AArch64::LD4Rv2d_POST:
488   case AArch64::LD4Rv4s_POST:
489   case AArch64::LD4Rv8h_POST:
490   case AArch64::LD4Rv16b_POST:
491     DestRegIdx = -1;
492     BaseRegIdx = 2;
493     OffsetIdx = 3;
494     IsPrePost = true;
495     break;
496 
497   case AArch64::LDRBBroW:
498   case AArch64::LDRBBroX:
499   case AArch64::LDRBBui:
500   case AArch64::LDRBroW:
501   case AArch64::LDRBroX:
502   case AArch64::LDRBui:
503   case AArch64::LDRDl:
504   case AArch64::LDRDroW:
505   case AArch64::LDRDroX:
506   case AArch64::LDRDui:
507   case AArch64::LDRHHroW:
508   case AArch64::LDRHHroX:
509   case AArch64::LDRHHui:
510   case AArch64::LDRHroW:
511   case AArch64::LDRHroX:
512   case AArch64::LDRHui:
513   case AArch64::LDRQl:
514   case AArch64::LDRQroW:
515   case AArch64::LDRQroX:
516   case AArch64::LDRQui:
517   case AArch64::LDRSBWroW:
518   case AArch64::LDRSBWroX:
519   case AArch64::LDRSBWui:
520   case AArch64::LDRSBXroW:
521   case AArch64::LDRSBXroX:
522   case AArch64::LDRSBXui:
523   case AArch64::LDRSHWroW:
524   case AArch64::LDRSHWroX:
525   case AArch64::LDRSHWui:
526   case AArch64::LDRSHXroW:
527   case AArch64::LDRSHXroX:
528   case AArch64::LDRSHXui:
529   case AArch64::LDRSWl:
530   case AArch64::LDRSWroW:
531   case AArch64::LDRSWroX:
532   case AArch64::LDRSWui:
533   case AArch64::LDRSl:
534   case AArch64::LDRSroW:
535   case AArch64::LDRSroX:
536   case AArch64::LDRSui:
537   case AArch64::LDRWl:
538   case AArch64::LDRWroW:
539   case AArch64::LDRWroX:
540   case AArch64::LDRWui:
541   case AArch64::LDRXl:
542   case AArch64::LDRXroW:
543   case AArch64::LDRXroX:
544   case AArch64::LDRXui:
545   case AArch64::LDURBBi:
546   case AArch64::LDURBi:
547   case AArch64::LDURDi:
548   case AArch64::LDURHHi:
549   case AArch64::LDURHi:
550   case AArch64::LDURQi:
551   case AArch64::LDURSBWi:
552   case AArch64::LDURSBXi:
553   case AArch64::LDURSHWi:
554   case AArch64::LDURSHXi:
555   case AArch64::LDURSWi:
556   case AArch64::LDURSi:
557   case AArch64::LDURWi:
558   case AArch64::LDURXi:
559     DestRegIdx = 0;
560     BaseRegIdx = 1;
561     OffsetIdx = 2;
562     IsPrePost = false;
563     break;
564 
565   case AArch64::LDRBBpost:
566   case AArch64::LDRBBpre:
567   case AArch64::LDRBpost:
568   case AArch64::LDRBpre:
569   case AArch64::LDRDpost:
570   case AArch64::LDRDpre:
571   case AArch64::LDRHHpost:
572   case AArch64::LDRHHpre:
573   case AArch64::LDRHpost:
574   case AArch64::LDRHpre:
575   case AArch64::LDRQpost:
576   case AArch64::LDRQpre:
577   case AArch64::LDRSBWpost:
578   case AArch64::LDRSBWpre:
579   case AArch64::LDRSBXpost:
580   case AArch64::LDRSBXpre:
581   case AArch64::LDRSHWpost:
582   case AArch64::LDRSHWpre:
583   case AArch64::LDRSHXpost:
584   case AArch64::LDRSHXpre:
585   case AArch64::LDRSWpost:
586   case AArch64::LDRSWpre:
587   case AArch64::LDRSpost:
588   case AArch64::LDRSpre:
589   case AArch64::LDRWpost:
590   case AArch64::LDRWpre:
591   case AArch64::LDRXpost:
592   case AArch64::LDRXpre:
593     DestRegIdx = 1;
594     BaseRegIdx = 2;
595     OffsetIdx = 3;
596     IsPrePost = true;
597     break;
598 
599   case AArch64::LDNPDi:
600   case AArch64::LDNPQi:
601   case AArch64::LDNPSi:
602   case AArch64::LDPQi:
603   case AArch64::LDPDi:
604   case AArch64::LDPSi:
605     DestRegIdx = -1;
606     BaseRegIdx = 2;
607     OffsetIdx = 3;
608     IsPrePost = false;
609     break;
610 
611   case AArch64::LDPSWi:
612   case AArch64::LDPWi:
613   case AArch64::LDPXi:
614     DestRegIdx = 0;
615     BaseRegIdx = 2;
616     OffsetIdx = 3;
617     IsPrePost = false;
618     break;
619 
620   case AArch64::LDPQpost:
621   case AArch64::LDPQpre:
622   case AArch64::LDPDpost:
623   case AArch64::LDPDpre:
624   case AArch64::LDPSpost:
625   case AArch64::LDPSpre:
626     DestRegIdx = -1;
627     BaseRegIdx = 3;
628     OffsetIdx = 4;
629     IsPrePost = true;
630     break;
631 
632   case AArch64::LDPSWpost:
633   case AArch64::LDPSWpre:
634   case AArch64::LDPWpost:
635   case AArch64::LDPWpre:
636   case AArch64::LDPXpost:
637   case AArch64::LDPXpre:
638     DestRegIdx = 1;
639     BaseRegIdx = 3;
640     OffsetIdx = 4;
641     IsPrePost = true;
642     break;
643   }
644 
645   // Loads from the stack pointer don't get prefetched.
646   Register BaseReg = MI.getOperand(BaseRegIdx).getReg();
647   if (BaseReg == AArch64::SP || BaseReg == AArch64::WSP)
648     return None;
649 
650   LoadInfo LI;
651   LI.DestReg = DestRegIdx == -1 ? Register() : MI.getOperand(DestRegIdx).getReg();
652   LI.BaseReg = BaseReg;
653   LI.BaseRegIdx = BaseRegIdx;
654   LI.OffsetOpnd = OffsetIdx == -1 ? nullptr : &MI.getOperand(OffsetIdx);
655   LI.IsPrePost = IsPrePost;
656   return LI;
657 }
658 
getTag(const TargetRegisterInfo * TRI,const MachineInstr & MI,const LoadInfo & LI)659 static Optional<unsigned> getTag(const TargetRegisterInfo *TRI,
660                                  const MachineInstr &MI, const LoadInfo &LI) {
661   unsigned Dest = LI.DestReg ? TRI->getEncodingValue(LI.DestReg) : 0;
662   unsigned Base = TRI->getEncodingValue(LI.BaseReg);
663   unsigned Off;
664   if (LI.OffsetOpnd == nullptr)
665     Off = 0;
666   else if (LI.OffsetOpnd->isGlobal() || LI.OffsetOpnd->isSymbol() ||
667            LI.OffsetOpnd->isCPI())
668     return None;
669   else if (LI.OffsetOpnd->isReg())
670     Off = (1 << 5) | TRI->getEncodingValue(LI.OffsetOpnd->getReg());
671   else
672     Off = LI.OffsetOpnd->getImm() >> 2;
673 
674   return makeTag(Dest, Base, Off);
675 }
676 
runOnLoop(MachineLoop & L,MachineFunction & Fn)677 void FalkorHWPFFix::runOnLoop(MachineLoop &L, MachineFunction &Fn) {
678   // Build the initial tag map for the whole loop.
679   TagMap.clear();
680   for (MachineBasicBlock *MBB : L.getBlocks())
681     for (MachineInstr &MI : *MBB) {
682       Optional<LoadInfo> LInfo = getLoadInfo(MI);
683       if (!LInfo)
684         continue;
685       Optional<unsigned> Tag = getTag(TRI, MI, *LInfo);
686       if (!Tag)
687         continue;
688       TagMap[*Tag].push_back(&MI);
689     }
690 
691   bool AnyCollisions = false;
692   for (auto &P : TagMap) {
693     auto Size = P.second.size();
694     if (Size > 1) {
695       for (auto *MI : P.second) {
696         if (TII->isStridedAccess(*MI)) {
697           AnyCollisions = true;
698           break;
699         }
700       }
701     }
702     if (AnyCollisions)
703       break;
704   }
705   // Nothing to fix.
706   if (!AnyCollisions)
707     return;
708 
709   MachineRegisterInfo &MRI = Fn.getRegInfo();
710 
711   // Go through all the basic blocks in the current loop and fix any streaming
712   // loads to avoid collisions with any other loads.
713   LiveRegUnits LR(*TRI);
714   for (MachineBasicBlock *MBB : L.getBlocks()) {
715     LR.clear();
716     LR.addLiveOuts(*MBB);
717     for (auto I = MBB->rbegin(); I != MBB->rend(); LR.stepBackward(*I), ++I) {
718       MachineInstr &MI = *I;
719       if (!TII->isStridedAccess(MI))
720         continue;
721 
722       Optional<LoadInfo> OptLdI = getLoadInfo(MI);
723       if (!OptLdI)
724         continue;
725       LoadInfo LdI = *OptLdI;
726       Optional<unsigned> OptOldTag = getTag(TRI, MI, LdI);
727       if (!OptOldTag)
728         continue;
729       auto &OldCollisions = TagMap[*OptOldTag];
730       if (OldCollisions.size() <= 1)
731         continue;
732 
733       bool Fixed = false;
734       LLVM_DEBUG(dbgs() << "Attempting to fix tag collision: " << MI);
735 
736       if (!DebugCounter::shouldExecute(FixCounter)) {
737         LLVM_DEBUG(dbgs() << "Skipping fix due to debug counter:\n  " << MI);
738         continue;
739       }
740 
741       // Add the non-base registers of MI as live so we don't use them as
742       // scratch registers.
743       for (unsigned OpI = 0, OpE = MI.getNumOperands(); OpI < OpE; ++OpI) {
744         if (OpI == static_cast<unsigned>(LdI.BaseRegIdx))
745           continue;
746         MachineOperand &MO = MI.getOperand(OpI);
747         if (MO.isReg() && MO.readsReg())
748           LR.addReg(MO.getReg());
749       }
750 
751       for (unsigned ScratchReg : AArch64::GPR64RegClass) {
752         if (!LR.available(ScratchReg) || MRI.isReserved(ScratchReg))
753           continue;
754 
755         LoadInfo NewLdI(LdI);
756         NewLdI.BaseReg = ScratchReg;
757         unsigned NewTag = *getTag(TRI, MI, NewLdI);
758         // Scratch reg tag would collide too, so don't use it.
759         if (TagMap.count(NewTag))
760           continue;
761 
762         LLVM_DEBUG(dbgs() << "Changing base reg to: "
763                           << printReg(ScratchReg, TRI) << '\n');
764 
765         // Rewrite:
766         //   Xd = LOAD Xb, off
767         // to:
768         //   Xc = MOV Xb
769         //   Xd = LOAD Xc, off
770         DebugLoc DL = MI.getDebugLoc();
771         BuildMI(*MBB, &MI, DL, TII->get(AArch64::ORRXrs), ScratchReg)
772             .addReg(AArch64::XZR)
773             .addReg(LdI.BaseReg)
774             .addImm(0);
775         MachineOperand &BaseOpnd = MI.getOperand(LdI.BaseRegIdx);
776         BaseOpnd.setReg(ScratchReg);
777 
778         // If the load does a pre/post increment, then insert a MOV after as
779         // well to update the real base register.
780         if (LdI.IsPrePost) {
781           LLVM_DEBUG(dbgs() << "Doing post MOV of incremented reg: "
782                             << printReg(ScratchReg, TRI) << '\n');
783           MI.getOperand(0).setReg(
784               ScratchReg); // Change tied operand pre/post update dest.
785           BuildMI(*MBB, std::next(MachineBasicBlock::iterator(MI)), DL,
786                   TII->get(AArch64::ORRXrs), LdI.BaseReg)
787               .addReg(AArch64::XZR)
788               .addReg(ScratchReg)
789               .addImm(0);
790         }
791 
792         for (int I = 0, E = OldCollisions.size(); I != E; ++I)
793           if (OldCollisions[I] == &MI) {
794             std::swap(OldCollisions[I], OldCollisions[E - 1]);
795             OldCollisions.pop_back();
796             break;
797           }
798 
799         // Update TagMap to reflect instruction changes to reduce the number
800         // of later MOVs to be inserted.  This needs to be done after
801         // OldCollisions is updated since it may be relocated by this
802         // insertion.
803         TagMap[NewTag].push_back(&MI);
804         ++NumCollisionsAvoided;
805         Fixed = true;
806         Modified = true;
807         break;
808       }
809       if (!Fixed)
810         ++NumCollisionsNotAvoided;
811     }
812   }
813 }
814 
runOnMachineFunction(MachineFunction & Fn)815 bool FalkorHWPFFix::runOnMachineFunction(MachineFunction &Fn) {
816   auto &ST = static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
817   if (ST.getProcFamily() != AArch64Subtarget::Falkor)
818     return false;
819 
820   if (skipFunction(Fn.getFunction()))
821     return false;
822 
823   TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
824   TRI = ST.getRegisterInfo();
825 
826   MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
827 
828   Modified = false;
829 
830   for (MachineLoop *I : LI)
831     for (auto L = df_begin(I), LE = df_end(I); L != LE; ++L)
832       // Only process inner-loops
833       if (L->isInnermost())
834         runOnLoop(**L, Fn);
835 
836   return Modified;
837 }
838 
createFalkorHWPFFixPass()839 FunctionPass *llvm::createFalkorHWPFFixPass() { return new FalkorHWPFFix(); }
840