xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64ConditionalCompares.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===-- AArch64ConditionalCompares.cpp --- CCMP formation for AArch64 -----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the AArch64ConditionalCompares pass which reduces
10 // branching and code size by using the conditional compare instructions CCMP,
11 // CCMN, and FCMP.
12 //
13 // The CFG transformations for forming conditional compares are very similar to
14 // if-conversion, and this pass should run immediately before the early
15 // if-conversion pass.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "AArch64.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
23 #include "llvm/CodeGen/MachineDominators.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineLoopInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/MachineTraceMetrics.h"
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/CodeGen/TargetInstrInfo.h"
32 #include "llvm/CodeGen/TargetRegisterInfo.h"
33 #include "llvm/CodeGen/TargetSubtargetInfo.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/raw_ostream.h"
38 
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "aarch64-ccmp"
42 
43 // Absolute maximum number of instructions allowed per speculated block.
44 // This bypasses all other heuristics, so it should be set fairly high.
45 static cl::opt<unsigned> BlockInstrLimit(
46     "aarch64-ccmp-limit", cl::init(30), cl::Hidden,
47     cl::desc("Maximum number of instructions per speculated block."));
48 
49 // Stress testing mode - disable heuristics.
50 static cl::opt<bool> Stress("aarch64-stress-ccmp", cl::Hidden,
51                             cl::desc("Turn all knobs to 11"));
52 
53 STATISTIC(NumConsidered, "Number of ccmps considered");
54 STATISTIC(NumPhiRejs, "Number of ccmps rejected (PHI)");
55 STATISTIC(NumPhysRejs, "Number of ccmps rejected (Physregs)");
56 STATISTIC(NumPhi2Rejs, "Number of ccmps rejected (PHI2)");
57 STATISTIC(NumHeadBranchRejs, "Number of ccmps rejected (Head branch)");
58 STATISTIC(NumCmpBranchRejs, "Number of ccmps rejected (CmpBB branch)");
59 STATISTIC(NumCmpTermRejs, "Number of ccmps rejected (CmpBB is cbz...)");
60 STATISTIC(NumImmRangeRejs, "Number of ccmps rejected (Imm out of range)");
61 STATISTIC(NumLiveDstRejs, "Number of ccmps rejected (Cmp dest live)");
62 STATISTIC(NumMultNZCVUses, "Number of ccmps rejected (NZCV used)");
63 STATISTIC(NumUnknNZCVDefs, "Number of ccmps rejected (NZCV def unknown)");
64 
65 STATISTIC(NumSpeculateRejs, "Number of ccmps rejected (Can't speculate)");
66 
67 STATISTIC(NumConverted, "Number of ccmp instructions created");
68 STATISTIC(NumCompBranches, "Number of cbz/cbnz branches converted");
69 
70 //===----------------------------------------------------------------------===//
71 //                                 SSACCmpConv
72 //===----------------------------------------------------------------------===//
73 //
74 // The SSACCmpConv class performs ccmp-conversion on SSA form machine code
75 // after determining if it is possible. The class contains no heuristics;
76 // external code should be used to determine when ccmp-conversion is a good
77 // idea.
78 //
79 // CCmp-formation works on a CFG representing chained conditions, typically
80 // from C's short-circuit || and && operators:
81 //
82 //   From:         Head            To:         Head
83 //                 / |                         CmpBB
84 //                /  |                         / |
85 //               |  CmpBB                     /  |
86 //               |  / |                    Tail  |
87 //               | /  |                      |   |
88 //              Tail  |                      |   |
89 //                |   |                      |   |
90 //               ... ...                    ... ...
91 //
92 // The Head block is terminated by a br.cond instruction, and the CmpBB block
93 // contains compare + br.cond. Tail must be a successor of both.
94 //
95 // The cmp-conversion turns the compare instruction in CmpBB into a conditional
96 // compare, and merges CmpBB into Head, speculatively executing its
97 // instructions. The AArch64 conditional compare instructions have an immediate
98 // operand that specifies the NZCV flag values when the condition is false and
99 // the compare isn't executed. This makes it possible to chain compares with
100 // different condition codes.
101 //
102 // Example:
103 //
104 //    if (a == 5 || b == 17)
105 //      foo();
106 //
107 //    Head:
108 //       cmp  w0, #5
109 //       b.eq Tail
110 //    CmpBB:
111 //       cmp  w1, #17
112 //       b.eq Tail
113 //    ...
114 //    Tail:
115 //      bl _foo
116 //
117 //  Becomes:
118 //
119 //    Head:
120 //       cmp  w0, #5
121 //       ccmp w1, #17, 4, ne  ; 4 = nZcv
122 //       b.eq Tail
123 //    ...
124 //    Tail:
125 //      bl _foo
126 //
127 // The ccmp condition code is the one that would cause the Head terminator to
128 // branch to CmpBB.
129 //
130 // FIXME: It should also be possible to speculate a block on the critical edge
131 // between Head and Tail, just like if-converting a diamond.
132 //
133 // FIXME: Handle PHIs in Tail by turning them into selects (if-conversion).
134 
135 namespace {
136 class SSACCmpConv {
137   MachineFunction *MF;
138   const TargetInstrInfo *TII;
139   const TargetRegisterInfo *TRI;
140   MachineRegisterInfo *MRI;
141   const MachineBranchProbabilityInfo *MBPI;
142 
143 public:
144   /// The first block containing a conditional branch, dominating everything
145   /// else.
146   MachineBasicBlock *Head;
147 
148   /// The block containing cmp+br.cond with a successor shared with Head.
149   MachineBasicBlock *CmpBB;
150 
151   /// The common successor for Head and CmpBB.
152   MachineBasicBlock *Tail;
153 
154   /// The compare instruction in CmpBB that can be converted to a ccmp.
155   MachineInstr *CmpMI;
156 
157 private:
158   /// The branch condition in Head as determined by analyzeBranch.
159   SmallVector<MachineOperand, 4> HeadCond;
160 
161   /// The condition code that makes Head branch to CmpBB.
162   AArch64CC::CondCode HeadCmpBBCC;
163 
164   /// The branch condition in CmpBB.
165   SmallVector<MachineOperand, 4> CmpBBCond;
166 
167   /// The condition code that makes CmpBB branch to Tail.
168   AArch64CC::CondCode CmpBBTailCC;
169 
170   /// Check if the Tail PHIs are trivially convertible.
171   bool trivialTailPHIs();
172 
173   /// Remove CmpBB from the Tail PHIs.
174   void updateTailPHIs();
175 
176   /// Check if an operand defining DstReg is dead.
177   bool isDeadDef(unsigned DstReg);
178 
179   /// Find the compare instruction in MBB that controls the conditional branch.
180   /// Return NULL if a convertible instruction can't be found.
181   MachineInstr *findConvertibleCompare(MachineBasicBlock *MBB);
182 
183   /// Return true if all non-terminator instructions in MBB can be safely
184   /// speculated.
185   bool canSpeculateInstrs(MachineBasicBlock *MBB, const MachineInstr *CmpMI);
186 
187 public:
188   /// runOnMachineFunction - Initialize per-function data structures.
runOnMachineFunction(MachineFunction & MF,const MachineBranchProbabilityInfo * MBPI)189   void runOnMachineFunction(MachineFunction &MF,
190                             const MachineBranchProbabilityInfo *MBPI) {
191     this->MF = &MF;
192     this->MBPI = MBPI;
193     TII = MF.getSubtarget().getInstrInfo();
194     TRI = MF.getSubtarget().getRegisterInfo();
195     MRI = &MF.getRegInfo();
196   }
197 
198   /// If the sub-CFG headed by MBB can be cmp-converted, initialize the
199   /// internal state, and return true.
200   bool canConvert(MachineBasicBlock *MBB);
201 
202   /// Cmo-convert the last block passed to canConvertCmp(), assuming
203   /// it is possible. Add any erased blocks to RemovedBlocks.
204   void convert(SmallVectorImpl<MachineBasicBlock *> &RemovedBlocks);
205 
206   /// Return the expected code size delta if the conversion into a
207   /// conditional compare is performed.
208   int expectedCodeSizeDelta() const;
209 };
210 } // end anonymous namespace
211 
212 // Check that all PHIs in Tail are selecting the same value from Head and CmpBB.
213 // This means that no if-conversion is required when merging CmpBB into Head.
trivialTailPHIs()214 bool SSACCmpConv::trivialTailPHIs() {
215   for (auto &I : *Tail) {
216     if (!I.isPHI())
217       break;
218     unsigned HeadReg = 0, CmpBBReg = 0;
219     // PHI operands come in (VReg, MBB) pairs.
220     for (unsigned oi = 1, oe = I.getNumOperands(); oi != oe; oi += 2) {
221       MachineBasicBlock *MBB = I.getOperand(oi + 1).getMBB();
222       Register Reg = I.getOperand(oi).getReg();
223       if (MBB == Head) {
224         assert((!HeadReg || HeadReg == Reg) && "Inconsistent PHI operands");
225         HeadReg = Reg;
226       }
227       if (MBB == CmpBB) {
228         assert((!CmpBBReg || CmpBBReg == Reg) && "Inconsistent PHI operands");
229         CmpBBReg = Reg;
230       }
231     }
232     if (HeadReg != CmpBBReg)
233       return false;
234   }
235   return true;
236 }
237 
238 // Assuming that trivialTailPHIs() is true, update the Tail PHIs by simply
239 // removing the CmpBB operands. The Head operands will be identical.
updateTailPHIs()240 void SSACCmpConv::updateTailPHIs() {
241   for (auto &I : *Tail) {
242     if (!I.isPHI())
243       break;
244     // I is a PHI. It can have multiple entries for CmpBB.
245     for (unsigned oi = I.getNumOperands(); oi > 2; oi -= 2) {
246       // PHI operands are (Reg, MBB) at (oi-2, oi-1).
247       if (I.getOperand(oi - 1).getMBB() == CmpBB) {
248         I.removeOperand(oi - 1);
249         I.removeOperand(oi - 2);
250       }
251     }
252   }
253 }
254 
255 // This pass runs before the AArch64DeadRegisterDefinitions pass, so compares
256 // are still writing virtual registers without any uses.
isDeadDef(unsigned DstReg)257 bool SSACCmpConv::isDeadDef(unsigned DstReg) {
258   // Writes to the zero register are dead.
259   if (DstReg == AArch64::WZR || DstReg == AArch64::XZR)
260     return true;
261   if (!Register::isVirtualRegister(DstReg))
262     return false;
263   // A virtual register def without any uses will be marked dead later, and
264   // eventually replaced by the zero register.
265   return MRI->use_nodbg_empty(DstReg);
266 }
267 
268 // Parse a condition code returned by analyzeBranch, and compute the CondCode
269 // corresponding to TBB.
270 // Return
parseCond(ArrayRef<MachineOperand> Cond,AArch64CC::CondCode & CC)271 static bool parseCond(ArrayRef<MachineOperand> Cond, AArch64CC::CondCode &CC) {
272   // A normal br.cond simply has the condition code.
273   if (Cond[0].getImm() != -1) {
274     assert(Cond.size() == 1 && "Unknown Cond array format");
275     CC = (AArch64CC::CondCode)(int)Cond[0].getImm();
276     return true;
277   }
278   // For tbz and cbz instruction, the opcode is next.
279   switch (Cond[1].getImm()) {
280   default:
281     // This includes tbz / tbnz branches which can't be converted to
282     // ccmp + br.cond.
283     return false;
284   case AArch64::CBZW:
285   case AArch64::CBZX:
286     assert(Cond.size() == 3 && "Unknown Cond array format");
287     CC = AArch64CC::EQ;
288     return true;
289   case AArch64::CBNZW:
290   case AArch64::CBNZX:
291     assert(Cond.size() == 3 && "Unknown Cond array format");
292     CC = AArch64CC::NE;
293     return true;
294   }
295 }
296 
findConvertibleCompare(MachineBasicBlock * MBB)297 MachineInstr *SSACCmpConv::findConvertibleCompare(MachineBasicBlock *MBB) {
298   MachineBasicBlock::iterator I = MBB->getFirstTerminator();
299   if (I == MBB->end())
300     return nullptr;
301   // The terminator must be controlled by the flags.
302   if (!I->readsRegister(AArch64::NZCV, /*TRI=*/nullptr)) {
303     switch (I->getOpcode()) {
304     case AArch64::CBZW:
305     case AArch64::CBZX:
306     case AArch64::CBNZW:
307     case AArch64::CBNZX:
308       // These can be converted into a ccmp against #0.
309       return &*I;
310     }
311     ++NumCmpTermRejs;
312     LLVM_DEBUG(dbgs() << "Flags not used by terminator: " << *I);
313     return nullptr;
314   }
315 
316   // Now find the instruction controlling the terminator.
317   for (MachineBasicBlock::iterator B = MBB->begin(); I != B;) {
318     I = prev_nodbg(I, MBB->begin());
319     assert(!I->isTerminator() && "Spurious terminator");
320     switch (I->getOpcode()) {
321     // cmp is an alias for subs with a dead destination register.
322     case AArch64::SUBSWri:
323     case AArch64::SUBSXri:
324     // cmn is an alias for adds with a dead destination register.
325     case AArch64::ADDSWri:
326     case AArch64::ADDSXri:
327       // Check that the immediate operand is within range, ccmp wants a uimm5.
328       // Rd = SUBSri Rn, imm, shift
329       if (I->getOperand(3).getImm() || !isUInt<5>(I->getOperand(2).getImm())) {
330         LLVM_DEBUG(dbgs() << "Immediate out of range for ccmp: " << *I);
331         ++NumImmRangeRejs;
332         return nullptr;
333       }
334       [[fallthrough]];
335     case AArch64::SUBSWrr:
336     case AArch64::SUBSXrr:
337     case AArch64::ADDSWrr:
338     case AArch64::ADDSXrr:
339       if (isDeadDef(I->getOperand(0).getReg()))
340         return &*I;
341       LLVM_DEBUG(dbgs() << "Can't convert compare with live destination: "
342                         << *I);
343       ++NumLiveDstRejs;
344       return nullptr;
345     case AArch64::FCMPSrr:
346     case AArch64::FCMPDrr:
347     case AArch64::FCMPESrr:
348     case AArch64::FCMPEDrr:
349       return &*I;
350     }
351 
352     // Check for flag reads and clobbers.
353     PhysRegInfo PRI = AnalyzePhysRegInBundle(*I, AArch64::NZCV, TRI);
354 
355     if (PRI.Read) {
356       // The ccmp doesn't produce exactly the same flags as the original
357       // compare, so reject the transform if there are uses of the flags
358       // besides the terminators.
359       LLVM_DEBUG(dbgs() << "Can't create ccmp with multiple uses: " << *I);
360       ++NumMultNZCVUses;
361       return nullptr;
362     }
363 
364     if (PRI.Defined || PRI.Clobbered) {
365       LLVM_DEBUG(dbgs() << "Not convertible compare: " << *I);
366       ++NumUnknNZCVDefs;
367       return nullptr;
368     }
369   }
370   LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB)
371                     << '\n');
372   return nullptr;
373 }
374 
375 /// Determine if all the instructions in MBB can safely
376 /// be speculated. The terminators are not considered.
377 ///
378 /// Only CmpMI is allowed to clobber the flags.
379 ///
canSpeculateInstrs(MachineBasicBlock * MBB,const MachineInstr * CmpMI)380 bool SSACCmpConv::canSpeculateInstrs(MachineBasicBlock *MBB,
381                                      const MachineInstr *CmpMI) {
382   // Reject any live-in physregs. It's probably NZCV/EFLAGS, and very hard to
383   // get right.
384   if (!MBB->livein_empty()) {
385     LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has live-ins.\n");
386     return false;
387   }
388 
389   unsigned InstrCount = 0;
390 
391   // Check all instructions, except the terminators. It is assumed that
392   // terminators never have side effects or define any used register values.
393   for (auto &I : make_range(MBB->begin(), MBB->getFirstTerminator())) {
394     if (I.isDebugInstr())
395       continue;
396 
397     if (++InstrCount > BlockInstrLimit && !Stress) {
398       LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has more than "
399                         << BlockInstrLimit << " instructions.\n");
400       return false;
401     }
402 
403     // There shouldn't normally be any phis in a single-predecessor block.
404     if (I.isPHI()) {
405       LLVM_DEBUG(dbgs() << "Can't hoist: " << I);
406       return false;
407     }
408 
409     // Don't speculate loads. Note that it may be possible and desirable to
410     // speculate GOT or constant pool loads that are guaranteed not to trap,
411     // but we don't support that for now.
412     if (I.mayLoad()) {
413       LLVM_DEBUG(dbgs() << "Won't speculate load: " << I);
414       return false;
415     }
416 
417     // We never speculate stores, so an AA pointer isn't necessary.
418     bool DontMoveAcrossStore = true;
419     if (!I.isSafeToMove(nullptr, DontMoveAcrossStore)) {
420       LLVM_DEBUG(dbgs() << "Can't speculate: " << I);
421       return false;
422     }
423 
424     // Only CmpMI is allowed to clobber the flags.
425     if (&I != CmpMI && I.modifiesRegister(AArch64::NZCV, TRI)) {
426       LLVM_DEBUG(dbgs() << "Clobbers flags: " << I);
427       return false;
428     }
429   }
430   return true;
431 }
432 
433 /// Analyze the sub-cfg rooted in MBB, and return true if it is a potential
434 /// candidate for cmp-conversion. Fill out the internal state.
435 ///
canConvert(MachineBasicBlock * MBB)436 bool SSACCmpConv::canConvert(MachineBasicBlock *MBB) {
437   Head = MBB;
438   Tail = CmpBB = nullptr;
439 
440   if (Head->succ_size() != 2)
441     return false;
442   MachineBasicBlock *Succ0 = Head->succ_begin()[0];
443   MachineBasicBlock *Succ1 = Head->succ_begin()[1];
444 
445   // CmpBB can only have a single predecessor. Tail is allowed many.
446   if (Succ0->pred_size() != 1)
447     std::swap(Succ0, Succ1);
448 
449   // Succ0 is our candidate for CmpBB.
450   if (Succ0->pred_size() != 1 || Succ0->succ_size() != 2)
451     return false;
452 
453   CmpBB = Succ0;
454   Tail = Succ1;
455 
456   if (!CmpBB->isSuccessor(Tail))
457     return false;
458 
459   // The CFG topology checks out.
460   LLVM_DEBUG(dbgs() << "\nTriangle: " << printMBBReference(*Head) << " -> "
461                     << printMBBReference(*CmpBB) << " -> "
462                     << printMBBReference(*Tail) << '\n');
463   ++NumConsidered;
464 
465   // Tail is allowed to have many predecessors, but we can't handle PHIs yet.
466   //
467   // FIXME: Real PHIs could be if-converted as long as the CmpBB values are
468   // defined before The CmpBB cmp clobbers the flags. Alternatively, it should
469   // always be safe to sink the ccmp down to immediately before the CmpBB
470   // terminators.
471   if (!trivialTailPHIs()) {
472     LLVM_DEBUG(dbgs() << "Can't handle phis in Tail.\n");
473     ++NumPhiRejs;
474     return false;
475   }
476 
477   if (!Tail->livein_empty()) {
478     LLVM_DEBUG(dbgs() << "Can't handle live-in physregs in Tail.\n");
479     ++NumPhysRejs;
480     return false;
481   }
482 
483   // CmpBB should never have PHIs since Head is its only predecessor.
484   // FIXME: Clean them up if it happens.
485   if (!CmpBB->empty() && CmpBB->front().isPHI()) {
486     LLVM_DEBUG(dbgs() << "Can't handle phis in CmpBB.\n");
487     ++NumPhi2Rejs;
488     return false;
489   }
490 
491   if (!CmpBB->livein_empty()) {
492     LLVM_DEBUG(dbgs() << "Can't handle live-in physregs in CmpBB.\n");
493     ++NumPhysRejs;
494     return false;
495   }
496 
497   // The branch we're looking to eliminate must be analyzable.
498   HeadCond.clear();
499   MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
500   if (TII->analyzeBranch(*Head, TBB, FBB, HeadCond)) {
501     LLVM_DEBUG(dbgs() << "Head branch not analyzable.\n");
502     ++NumHeadBranchRejs;
503     return false;
504   }
505 
506   // This is weird, probably some sort of degenerate CFG, or an edge to a
507   // landing pad.
508   if (!TBB || HeadCond.empty()) {
509     LLVM_DEBUG(
510         dbgs() << "analyzeBranch didn't find conditional branch in Head.\n");
511     ++NumHeadBranchRejs;
512     return false;
513   }
514 
515   if (!parseCond(HeadCond, HeadCmpBBCC)) {
516     LLVM_DEBUG(dbgs() << "Unsupported branch type on Head\n");
517     ++NumHeadBranchRejs;
518     return false;
519   }
520 
521   // Make sure the branch direction is right.
522   if (TBB != CmpBB) {
523     assert(TBB == Tail && "Unexpected TBB");
524     HeadCmpBBCC = AArch64CC::getInvertedCondCode(HeadCmpBBCC);
525   }
526 
527   CmpBBCond.clear();
528   TBB = FBB = nullptr;
529   if (TII->analyzeBranch(*CmpBB, TBB, FBB, CmpBBCond)) {
530     LLVM_DEBUG(dbgs() << "CmpBB branch not analyzable.\n");
531     ++NumCmpBranchRejs;
532     return false;
533   }
534 
535   if (!TBB || CmpBBCond.empty()) {
536     LLVM_DEBUG(
537         dbgs() << "analyzeBranch didn't find conditional branch in CmpBB.\n");
538     ++NumCmpBranchRejs;
539     return false;
540   }
541 
542   if (!parseCond(CmpBBCond, CmpBBTailCC)) {
543     LLVM_DEBUG(dbgs() << "Unsupported branch type on CmpBB\n");
544     ++NumCmpBranchRejs;
545     return false;
546   }
547 
548   if (TBB != Tail)
549     CmpBBTailCC = AArch64CC::getInvertedCondCode(CmpBBTailCC);
550 
551   LLVM_DEBUG(dbgs() << "Head->CmpBB on "
552                     << AArch64CC::getCondCodeName(HeadCmpBBCC)
553                     << ", CmpBB->Tail on "
554                     << AArch64CC::getCondCodeName(CmpBBTailCC) << '\n');
555 
556   CmpMI = findConvertibleCompare(CmpBB);
557   if (!CmpMI)
558     return false;
559 
560   if (!canSpeculateInstrs(CmpBB, CmpMI)) {
561     ++NumSpeculateRejs;
562     return false;
563   }
564   return true;
565 }
566 
convert(SmallVectorImpl<MachineBasicBlock * > & RemovedBlocks)567 void SSACCmpConv::convert(SmallVectorImpl<MachineBasicBlock *> &RemovedBlocks) {
568   LLVM_DEBUG(dbgs() << "Merging " << printMBBReference(*CmpBB) << " into "
569                     << printMBBReference(*Head) << ":\n"
570                     << *CmpBB);
571 
572   // All CmpBB instructions are moved into Head, and CmpBB is deleted.
573   // Update the CFG first.
574   updateTailPHIs();
575 
576   // Save successor probabilties before removing CmpBB and Tail from their
577   // parents.
578   BranchProbability Head2CmpBB = MBPI->getEdgeProbability(Head, CmpBB);
579   BranchProbability CmpBB2Tail = MBPI->getEdgeProbability(CmpBB, Tail);
580 
581   Head->removeSuccessor(CmpBB);
582   CmpBB->removeSuccessor(Tail);
583 
584   // If Head and CmpBB had successor probabilties, udpate the probabilities to
585   // reflect the ccmp-conversion.
586   if (Head->hasSuccessorProbabilities() && CmpBB->hasSuccessorProbabilities()) {
587 
588     // Head is allowed two successors. We've removed CmpBB, so the remaining
589     // successor is Tail. We need to increase the successor probability for
590     // Tail to account for the CmpBB path we removed.
591     //
592     // Pr(Tail|Head) += Pr(CmpBB|Head) * Pr(Tail|CmpBB).
593     assert(*Head->succ_begin() == Tail && "Head successor is not Tail");
594     BranchProbability Head2Tail = MBPI->getEdgeProbability(Head, Tail);
595     Head->setSuccProbability(Head->succ_begin(),
596                              Head2Tail + Head2CmpBB * CmpBB2Tail);
597 
598     // We will transfer successors of CmpBB to Head in a moment without
599     // normalizing the successor probabilities. Set the successor probabilites
600     // before doing so.
601     //
602     // Pr(I|Head) = Pr(CmpBB|Head) * Pr(I|CmpBB).
603     for (auto I = CmpBB->succ_begin(), E = CmpBB->succ_end(); I != E; ++I) {
604       BranchProbability CmpBB2I = MBPI->getEdgeProbability(CmpBB, *I);
605       CmpBB->setSuccProbability(I, Head2CmpBB * CmpBB2I);
606     }
607   }
608 
609   Head->transferSuccessorsAndUpdatePHIs(CmpBB);
610   DebugLoc TermDL = Head->getFirstTerminator()->getDebugLoc();
611   TII->removeBranch(*Head);
612 
613   // If the Head terminator was one of the cbz / tbz branches with built-in
614   // compare, we need to insert an explicit compare instruction in its place.
615   if (HeadCond[0].getImm() == -1) {
616     ++NumCompBranches;
617     unsigned Opc = 0;
618     switch (HeadCond[1].getImm()) {
619     case AArch64::CBZW:
620     case AArch64::CBNZW:
621       Opc = AArch64::SUBSWri;
622       break;
623     case AArch64::CBZX:
624     case AArch64::CBNZX:
625       Opc = AArch64::SUBSXri;
626       break;
627     default:
628       llvm_unreachable("Cannot convert Head branch");
629     }
630     const MCInstrDesc &MCID = TII->get(Opc);
631     // Create a dummy virtual register for the SUBS def.
632     Register DestReg =
633         MRI->createVirtualRegister(TII->getRegClass(MCID, 0, TRI, *MF));
634     // Insert a SUBS Rn, #0 instruction instead of the cbz / cbnz.
635     BuildMI(*Head, Head->end(), TermDL, MCID)
636         .addReg(DestReg, RegState::Define | RegState::Dead)
637         .add(HeadCond[2])
638         .addImm(0)
639         .addImm(0);
640     // SUBS uses the GPR*sp register classes.
641     MRI->constrainRegClass(HeadCond[2].getReg(),
642                            TII->getRegClass(MCID, 1, TRI, *MF));
643   }
644 
645   Head->splice(Head->end(), CmpBB, CmpBB->begin(), CmpBB->end());
646 
647   // Now replace CmpMI with a ccmp instruction that also considers the incoming
648   // flags.
649   unsigned Opc = 0;
650   unsigned FirstOp = 1;   // First CmpMI operand to copy.
651   bool isZBranch = false; // CmpMI is a cbz/cbnz instruction.
652   switch (CmpMI->getOpcode()) {
653   default:
654     llvm_unreachable("Unknown compare opcode");
655   case AArch64::SUBSWri:    Opc = AArch64::CCMPWi; break;
656   case AArch64::SUBSWrr:    Opc = AArch64::CCMPWr; break;
657   case AArch64::SUBSXri:    Opc = AArch64::CCMPXi; break;
658   case AArch64::SUBSXrr:    Opc = AArch64::CCMPXr; break;
659   case AArch64::ADDSWri:    Opc = AArch64::CCMNWi; break;
660   case AArch64::ADDSWrr:    Opc = AArch64::CCMNWr; break;
661   case AArch64::ADDSXri:    Opc = AArch64::CCMNXi; break;
662   case AArch64::ADDSXrr:    Opc = AArch64::CCMNXr; break;
663   case AArch64::FCMPSrr:    Opc = AArch64::FCCMPSrr; FirstOp = 0; break;
664   case AArch64::FCMPDrr:    Opc = AArch64::FCCMPDrr; FirstOp = 0; break;
665   case AArch64::FCMPESrr:   Opc = AArch64::FCCMPESrr; FirstOp = 0; break;
666   case AArch64::FCMPEDrr:   Opc = AArch64::FCCMPEDrr; FirstOp = 0; break;
667   case AArch64::CBZW:
668   case AArch64::CBNZW:
669     Opc = AArch64::CCMPWi;
670     FirstOp = 0;
671     isZBranch = true;
672     break;
673   case AArch64::CBZX:
674   case AArch64::CBNZX:
675     Opc = AArch64::CCMPXi;
676     FirstOp = 0;
677     isZBranch = true;
678     break;
679   }
680 
681   // The ccmp instruction should set the flags according to the comparison when
682   // Head would have branched to CmpBB.
683   // The NZCV immediate operand should provide flags for the case where Head
684   // would have branched to Tail. These flags should cause the new Head
685   // terminator to branch to tail.
686   unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(CmpBBTailCC);
687   const MCInstrDesc &MCID = TII->get(Opc);
688   MRI->constrainRegClass(CmpMI->getOperand(FirstOp).getReg(),
689                          TII->getRegClass(MCID, 0, TRI, *MF));
690   if (CmpMI->getOperand(FirstOp + 1).isReg())
691     MRI->constrainRegClass(CmpMI->getOperand(FirstOp + 1).getReg(),
692                            TII->getRegClass(MCID, 1, TRI, *MF));
693   MachineInstrBuilder MIB = BuildMI(*Head, CmpMI, CmpMI->getDebugLoc(), MCID)
694                                 .add(CmpMI->getOperand(FirstOp)); // Register Rn
695   if (isZBranch)
696     MIB.addImm(0); // cbz/cbnz Rn -> ccmp Rn, #0
697   else
698     MIB.add(CmpMI->getOperand(FirstOp + 1)); // Register Rm / Immediate
699   MIB.addImm(NZCV).addImm(HeadCmpBBCC);
700 
701   // If CmpMI was a terminator, we need a new conditional branch to replace it.
702   // This now becomes a Head terminator.
703   if (isZBranch) {
704     bool isNZ = CmpMI->getOpcode() == AArch64::CBNZW ||
705                 CmpMI->getOpcode() == AArch64::CBNZX;
706     BuildMI(*Head, CmpMI, CmpMI->getDebugLoc(), TII->get(AArch64::Bcc))
707         .addImm(isNZ ? AArch64CC::NE : AArch64CC::EQ)
708         .add(CmpMI->getOperand(1)); // Branch target.
709   }
710   CmpMI->eraseFromParent();
711   Head->updateTerminator(CmpBB->getNextNode());
712 
713   RemovedBlocks.push_back(CmpBB);
714   CmpBB->eraseFromParent();
715   LLVM_DEBUG(dbgs() << "Result:\n" << *Head);
716   ++NumConverted;
717 }
718 
expectedCodeSizeDelta() const719 int SSACCmpConv::expectedCodeSizeDelta() const {
720   int delta = 0;
721   // If the Head terminator was one of the cbz / tbz branches with built-in
722   // compare, we need to insert an explicit compare instruction in its place
723   // plus a branch instruction.
724   if (HeadCond[0].getImm() == -1) {
725     switch (HeadCond[1].getImm()) {
726     case AArch64::CBZW:
727     case AArch64::CBNZW:
728     case AArch64::CBZX:
729     case AArch64::CBNZX:
730       // Therefore delta += 1
731       delta = 1;
732       break;
733     default:
734       llvm_unreachable("Cannot convert Head branch");
735     }
736   }
737   // If the Cmp terminator was one of the cbz / tbz branches with
738   // built-in compare, it will be turned into a compare instruction
739   // into Head, but we do not save any instruction.
740   // Otherwise, we save the branch instruction.
741   switch (CmpMI->getOpcode()) {
742   default:
743     --delta;
744     break;
745   case AArch64::CBZW:
746   case AArch64::CBNZW:
747   case AArch64::CBZX:
748   case AArch64::CBNZX:
749     break;
750   }
751   return delta;
752 }
753 
754 //===----------------------------------------------------------------------===//
755 //                       AArch64ConditionalCompares Pass
756 //===----------------------------------------------------------------------===//
757 
758 namespace {
759 class AArch64ConditionalCompares : public MachineFunctionPass {
760   const MachineBranchProbabilityInfo *MBPI;
761   const TargetInstrInfo *TII;
762   const TargetRegisterInfo *TRI;
763   MCSchedModel SchedModel;
764   // Does the proceeded function has Oz attribute.
765   bool MinSize;
766   MachineRegisterInfo *MRI;
767   MachineDominatorTree *DomTree;
768   MachineLoopInfo *Loops;
769   MachineTraceMetrics *Traces;
770   MachineTraceMetrics::Ensemble *MinInstr;
771   SSACCmpConv CmpConv;
772 
773 public:
774   static char ID;
AArch64ConditionalCompares()775   AArch64ConditionalCompares() : MachineFunctionPass(ID) {
776     initializeAArch64ConditionalComparesPass(*PassRegistry::getPassRegistry());
777   }
778   void getAnalysisUsage(AnalysisUsage &AU) const override;
779   bool runOnMachineFunction(MachineFunction &MF) override;
getPassName() const780   StringRef getPassName() const override {
781     return "AArch64 Conditional Compares";
782   }
783 
784 private:
785   bool tryConvert(MachineBasicBlock *);
786   void updateDomTree(ArrayRef<MachineBasicBlock *> Removed);
787   void updateLoops(ArrayRef<MachineBasicBlock *> Removed);
788   void invalidateTraces();
789   bool shouldConvert();
790 };
791 } // end anonymous namespace
792 
793 char AArch64ConditionalCompares::ID = 0;
794 
795 INITIALIZE_PASS_BEGIN(AArch64ConditionalCompares, "aarch64-ccmp",
796                       "AArch64 CCMP Pass", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)797 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
798 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
799 INITIALIZE_PASS_DEPENDENCY(MachineTraceMetrics)
800 INITIALIZE_PASS_END(AArch64ConditionalCompares, "aarch64-ccmp",
801                     "AArch64 CCMP Pass", false, false)
802 
803 FunctionPass *llvm::createAArch64ConditionalCompares() {
804   return new AArch64ConditionalCompares();
805 }
806 
getAnalysisUsage(AnalysisUsage & AU) const807 void AArch64ConditionalCompares::getAnalysisUsage(AnalysisUsage &AU) const {
808   AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
809   AU.addRequired<MachineDominatorTreeWrapperPass>();
810   AU.addPreserved<MachineDominatorTreeWrapperPass>();
811   AU.addRequired<MachineLoopInfoWrapperPass>();
812   AU.addPreserved<MachineLoopInfoWrapperPass>();
813   AU.addRequired<MachineTraceMetrics>();
814   AU.addPreserved<MachineTraceMetrics>();
815   MachineFunctionPass::getAnalysisUsage(AU);
816 }
817 
818 /// Update the dominator tree after if-conversion erased some blocks.
updateDomTree(ArrayRef<MachineBasicBlock * > Removed)819 void AArch64ConditionalCompares::updateDomTree(
820     ArrayRef<MachineBasicBlock *> Removed) {
821   // convert() removes CmpBB which was previously dominated by Head.
822   // CmpBB children should be transferred to Head.
823   MachineDomTreeNode *HeadNode = DomTree->getNode(CmpConv.Head);
824   for (MachineBasicBlock *RemovedMBB : Removed) {
825     MachineDomTreeNode *Node = DomTree->getNode(RemovedMBB);
826     assert(Node != HeadNode && "Cannot erase the head node");
827     assert(Node->getIDom() == HeadNode && "CmpBB should be dominated by Head");
828     while (Node->getNumChildren())
829       DomTree->changeImmediateDominator(Node->back(), HeadNode);
830     DomTree->eraseNode(RemovedMBB);
831   }
832 }
833 
834 /// Update LoopInfo after if-conversion.
835 void
updateLoops(ArrayRef<MachineBasicBlock * > Removed)836 AArch64ConditionalCompares::updateLoops(ArrayRef<MachineBasicBlock *> Removed) {
837   if (!Loops)
838     return;
839   for (MachineBasicBlock *RemovedMBB : Removed)
840     Loops->removeBlock(RemovedMBB);
841 }
842 
843 /// Invalidate MachineTraceMetrics before if-conversion.
invalidateTraces()844 void AArch64ConditionalCompares::invalidateTraces() {
845   Traces->invalidate(CmpConv.Head);
846   Traces->invalidate(CmpConv.CmpBB);
847 }
848 
849 /// Apply cost model and heuristics to the if-conversion in IfConv.
850 /// Return true if the conversion is a good idea.
851 ///
shouldConvert()852 bool AArch64ConditionalCompares::shouldConvert() {
853   // Stress testing mode disables all cost considerations.
854   if (Stress)
855     return true;
856   if (!MinInstr)
857     MinInstr = Traces->getEnsemble(MachineTraceStrategy::TS_MinInstrCount);
858 
859   // Head dominates CmpBB, so it is always included in its trace.
860   MachineTraceMetrics::Trace Trace = MinInstr->getTrace(CmpConv.CmpBB);
861 
862   // If code size is the main concern
863   if (MinSize) {
864     int CodeSizeDelta = CmpConv.expectedCodeSizeDelta();
865     LLVM_DEBUG(dbgs() << "Code size delta:  " << CodeSizeDelta << '\n');
866     // If we are minimizing the code size, do the conversion whatever
867     // the cost is.
868     if (CodeSizeDelta < 0)
869       return true;
870     if (CodeSizeDelta > 0) {
871       LLVM_DEBUG(dbgs() << "Code size is increasing, give up on this one.\n");
872       return false;
873     }
874     // CodeSizeDelta == 0, continue with the regular heuristics
875   }
876 
877   // Heuristic: The compare conversion delays the execution of the branch
878   // instruction because we must wait for the inputs to the second compare as
879   // well. The branch has no dependent instructions, but delaying it increases
880   // the cost of a misprediction.
881   //
882   // Set a limit on the delay we will accept.
883   unsigned DelayLimit = SchedModel.MispredictPenalty * 3 / 4;
884 
885   // Instruction depths can be computed for all trace instructions above CmpBB.
886   unsigned HeadDepth =
887       Trace.getInstrCycles(*CmpConv.Head->getFirstTerminator()).Depth;
888   unsigned CmpBBDepth =
889       Trace.getInstrCycles(*CmpConv.CmpBB->getFirstTerminator()).Depth;
890   LLVM_DEBUG(dbgs() << "Head depth:  " << HeadDepth
891                     << "\nCmpBB depth: " << CmpBBDepth << '\n');
892   if (CmpBBDepth > HeadDepth + DelayLimit) {
893     LLVM_DEBUG(dbgs() << "Branch delay would be larger than " << DelayLimit
894                       << " cycles.\n");
895     return false;
896   }
897 
898   // Check the resource depth at the bottom of CmpBB - these instructions will
899   // be speculated.
900   unsigned ResDepth = Trace.getResourceDepth(true);
901   LLVM_DEBUG(dbgs() << "Resources:   " << ResDepth << '\n');
902 
903   // Heuristic: The speculatively executed instructions must all be able to
904   // merge into the Head block. The Head critical path should dominate the
905   // resource cost of the speculated instructions.
906   if (ResDepth > HeadDepth) {
907     LLVM_DEBUG(dbgs() << "Too many instructions to speculate.\n");
908     return false;
909   }
910   return true;
911 }
912 
tryConvert(MachineBasicBlock * MBB)913 bool AArch64ConditionalCompares::tryConvert(MachineBasicBlock *MBB) {
914   bool Changed = false;
915   while (CmpConv.canConvert(MBB) && shouldConvert()) {
916     invalidateTraces();
917     SmallVector<MachineBasicBlock *, 4> RemovedBlocks;
918     CmpConv.convert(RemovedBlocks);
919     Changed = true;
920     updateDomTree(RemovedBlocks);
921     updateLoops(RemovedBlocks);
922   }
923   return Changed;
924 }
925 
runOnMachineFunction(MachineFunction & MF)926 bool AArch64ConditionalCompares::runOnMachineFunction(MachineFunction &MF) {
927   LLVM_DEBUG(dbgs() << "********** AArch64 Conditional Compares **********\n"
928                     << "********** Function: " << MF.getName() << '\n');
929   if (skipFunction(MF.getFunction()))
930     return false;
931 
932   TII = MF.getSubtarget().getInstrInfo();
933   TRI = MF.getSubtarget().getRegisterInfo();
934   SchedModel = MF.getSubtarget().getSchedModel();
935   MRI = &MF.getRegInfo();
936   DomTree = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
937   Loops = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
938   MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
939   Traces = &getAnalysis<MachineTraceMetrics>();
940   MinInstr = nullptr;
941   MinSize = MF.getFunction().hasMinSize();
942 
943   bool Changed = false;
944   CmpConv.runOnMachineFunction(MF, MBPI);
945 
946   // Visit blocks in dominator tree pre-order. The pre-order enables multiple
947   // cmp-conversions from the same head block.
948   // Note that updateDomTree() modifies the children of the DomTree node
949   // currently being visited. The df_iterator supports that; it doesn't look at
950   // child_begin() / child_end() until after a node has been visited.
951   for (auto *I : depth_first(DomTree))
952     if (tryConvert(I->getBlock()))
953       Changed = true;
954 
955   return Changed;
956 }
957