xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64ConditionOptimizer.cpp (revision 6be3386466ab79a84b48429ae66244f21526d3df)
1 //=- AArch64ConditionOptimizer.cpp - Remove useless comparisons 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 pass tries to make consecutive compares of values use same operands to
10 // allow CSE pass to remove duplicated instructions.  For this it analyzes
11 // branches and adjusts comparisons with immediate values by converting:
12 //  * GE -> GT
13 //  * GT -> GE
14 //  * LT -> LE
15 //  * LE -> LT
16 // and adjusting immediate values appropriately.  It basically corrects two
17 // immediate values towards each other to make them equal.
18 //
19 // Consider the following example in C:
20 //
21 //   if ((a < 5 && ...) || (a > 5 && ...)) {
22 //        ~~~~~             ~~~~~
23 //          ^                 ^
24 //          x                 y
25 //
26 // Here both "x" and "y" expressions compare "a" with "5".  When "x" evaluates
27 // to "false", "y" can just check flags set by the first comparison.  As a
28 // result of the canonicalization employed by
29 // SelectionDAGBuilder::visitSwitchCase, DAGCombine, and other target-specific
30 // code, assembly ends up in the form that is not CSE friendly:
31 //
32 //     ...
33 //     cmp      w8, #4
34 //     b.gt     .LBB0_3
35 //     ...
36 //   .LBB0_3:
37 //     cmp      w8, #6
38 //     b.lt     .LBB0_6
39 //     ...
40 //
41 // Same assembly after the pass:
42 //
43 //     ...
44 //     cmp      w8, #5
45 //     b.ge     .LBB0_3
46 //     ...
47 //   .LBB0_3:
48 //     cmp      w8, #5     // <-- CSE pass removes this instruction
49 //     b.le     .LBB0_6
50 //     ...
51 //
52 // Currently only SUBS and ADDS followed by b.?? are supported.
53 //
54 // TODO: maybe handle TBNZ/TBZ the same way as CMP when used instead for "a < 0"
55 // TODO: handle other conditional instructions (e.g. CSET)
56 // TODO: allow second branching to be anything if it doesn't require adjusting
57 //
58 //===----------------------------------------------------------------------===//
59 
60 #include "AArch64.h"
61 #include "MCTargetDesc/AArch64AddressingModes.h"
62 #include "Utils/AArch64BaseInfo.h"
63 #include "llvm/ADT/ArrayRef.h"
64 #include "llvm/ADT/DepthFirstIterator.h"
65 #include "llvm/ADT/SmallVector.h"
66 #include "llvm/ADT/Statistic.h"
67 #include "llvm/CodeGen/MachineBasicBlock.h"
68 #include "llvm/CodeGen/MachineDominators.h"
69 #include "llvm/CodeGen/MachineFunction.h"
70 #include "llvm/CodeGen/MachineFunctionPass.h"
71 #include "llvm/CodeGen/MachineInstr.h"
72 #include "llvm/CodeGen/MachineInstrBuilder.h"
73 #include "llvm/CodeGen/MachineOperand.h"
74 #include "llvm/CodeGen/MachineRegisterInfo.h"
75 #include "llvm/CodeGen/TargetInstrInfo.h"
76 #include "llvm/CodeGen/TargetSubtargetInfo.h"
77 #include "llvm/InitializePasses.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/raw_ostream.h"
82 #include <cassert>
83 #include <cstdlib>
84 #include <tuple>
85 
86 using namespace llvm;
87 
88 #define DEBUG_TYPE "aarch64-condopt"
89 
90 STATISTIC(NumConditionsAdjusted, "Number of conditions adjusted");
91 
92 namespace {
93 
94 class AArch64ConditionOptimizer : public MachineFunctionPass {
95   const TargetInstrInfo *TII;
96   MachineDominatorTree *DomTree;
97   const MachineRegisterInfo *MRI;
98 
99 public:
100   // Stores immediate, compare instruction opcode and branch condition (in this
101   // order) of adjusted comparison.
102   using CmpInfo = std::tuple<int, unsigned, AArch64CC::CondCode>;
103 
104   static char ID;
105 
106   AArch64ConditionOptimizer() : MachineFunctionPass(ID) {
107     initializeAArch64ConditionOptimizerPass(*PassRegistry::getPassRegistry());
108   }
109 
110   void getAnalysisUsage(AnalysisUsage &AU) const override;
111   MachineInstr *findSuitableCompare(MachineBasicBlock *MBB);
112   CmpInfo adjustCmp(MachineInstr *CmpMI, AArch64CC::CondCode Cmp);
113   void modifyCmp(MachineInstr *CmpMI, const CmpInfo &Info);
114   bool adjustTo(MachineInstr *CmpMI, AArch64CC::CondCode Cmp, MachineInstr *To,
115                 int ToImm);
116   bool runOnMachineFunction(MachineFunction &MF) override;
117 
118   StringRef getPassName() const override {
119     return "AArch64 Condition Optimizer";
120   }
121 };
122 
123 } // end anonymous namespace
124 
125 char AArch64ConditionOptimizer::ID = 0;
126 
127 INITIALIZE_PASS_BEGIN(AArch64ConditionOptimizer, "aarch64-condopt",
128                       "AArch64 CondOpt Pass", false, false)
129 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
130 INITIALIZE_PASS_END(AArch64ConditionOptimizer, "aarch64-condopt",
131                     "AArch64 CondOpt Pass", false, false)
132 
133 FunctionPass *llvm::createAArch64ConditionOptimizerPass() {
134   return new AArch64ConditionOptimizer();
135 }
136 
137 void AArch64ConditionOptimizer::getAnalysisUsage(AnalysisUsage &AU) const {
138   AU.addRequired<MachineDominatorTree>();
139   AU.addPreserved<MachineDominatorTree>();
140   MachineFunctionPass::getAnalysisUsage(AU);
141 }
142 
143 // Finds compare instruction that corresponds to supported types of branching.
144 // Returns the instruction or nullptr on failures or detecting unsupported
145 // instructions.
146 MachineInstr *AArch64ConditionOptimizer::findSuitableCompare(
147     MachineBasicBlock *MBB) {
148   MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
149   if (Term == MBB->end())
150     return nullptr;
151 
152   if (Term->getOpcode() != AArch64::Bcc)
153     return nullptr;
154 
155   // Since we may modify cmp of this MBB, make sure NZCV does not live out.
156   for (auto SuccBB : MBB->successors())
157     if (SuccBB->isLiveIn(AArch64::NZCV))
158       return nullptr;
159 
160   // Now find the instruction controlling the terminator.
161   for (MachineBasicBlock::iterator B = MBB->begin(), It = Term; It != B;) {
162     It = prev_nodbg(It, B);
163     MachineInstr &I = *It;
164     assert(!I.isTerminator() && "Spurious terminator");
165     // Check if there is any use of NZCV between CMP and Bcc.
166     if (I.readsRegister(AArch64::NZCV))
167       return nullptr;
168     switch (I.getOpcode()) {
169     // cmp is an alias for subs with a dead destination register.
170     case AArch64::SUBSWri:
171     case AArch64::SUBSXri:
172     // cmn is an alias for adds with a dead destination register.
173     case AArch64::ADDSWri:
174     case AArch64::ADDSXri: {
175       unsigned ShiftAmt = AArch64_AM::getShiftValue(I.getOperand(3).getImm());
176       if (!I.getOperand(2).isImm()) {
177         LLVM_DEBUG(dbgs() << "Immediate of cmp is symbolic, " << I << '\n');
178         return nullptr;
179       } else if (I.getOperand(2).getImm() << ShiftAmt >= 0xfff) {
180         LLVM_DEBUG(dbgs() << "Immediate of cmp may be out of range, " << I
181                           << '\n');
182         return nullptr;
183       } else if (!MRI->use_nodbg_empty(I.getOperand(0).getReg())) {
184         LLVM_DEBUG(dbgs() << "Destination of cmp is not dead, " << I << '\n');
185         return nullptr;
186       }
187       return &I;
188     }
189     // Prevent false positive case like:
190     // cmp      w19, #0
191     // cinc     w0, w19, gt
192     // ...
193     // fcmp     d8, #0.0
194     // b.gt     .LBB0_5
195     case AArch64::FCMPDri:
196     case AArch64::FCMPSri:
197     case AArch64::FCMPESri:
198     case AArch64::FCMPEDri:
199 
200     case AArch64::SUBSWrr:
201     case AArch64::SUBSXrr:
202     case AArch64::ADDSWrr:
203     case AArch64::ADDSXrr:
204     case AArch64::FCMPSrr:
205     case AArch64::FCMPDrr:
206     case AArch64::FCMPESrr:
207     case AArch64::FCMPEDrr:
208       // Skip comparison instructions without immediate operands.
209       return nullptr;
210     }
211   }
212   LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB)
213                     << '\n');
214   return nullptr;
215 }
216 
217 // Changes opcode adds <-> subs considering register operand width.
218 static int getComplementOpc(int Opc) {
219   switch (Opc) {
220   case AArch64::ADDSWri: return AArch64::SUBSWri;
221   case AArch64::ADDSXri: return AArch64::SUBSXri;
222   case AArch64::SUBSWri: return AArch64::ADDSWri;
223   case AArch64::SUBSXri: return AArch64::ADDSXri;
224   default:
225     llvm_unreachable("Unexpected opcode");
226   }
227 }
228 
229 // Changes form of comparison inclusive <-> exclusive.
230 static AArch64CC::CondCode getAdjustedCmp(AArch64CC::CondCode Cmp) {
231   switch (Cmp) {
232   case AArch64CC::GT: return AArch64CC::GE;
233   case AArch64CC::GE: return AArch64CC::GT;
234   case AArch64CC::LT: return AArch64CC::LE;
235   case AArch64CC::LE: return AArch64CC::LT;
236   default:
237     llvm_unreachable("Unexpected condition code");
238   }
239 }
240 
241 // Transforms GT -> GE, GE -> GT, LT -> LE, LE -> LT by updating comparison
242 // operator and condition code.
243 AArch64ConditionOptimizer::CmpInfo AArch64ConditionOptimizer::adjustCmp(
244     MachineInstr *CmpMI, AArch64CC::CondCode Cmp) {
245   unsigned Opc = CmpMI->getOpcode();
246 
247   // CMN (compare with negative immediate) is an alias to ADDS (as
248   // "operand - negative" == "operand + positive")
249   bool Negative = (Opc == AArch64::ADDSWri || Opc == AArch64::ADDSXri);
250 
251   int Correction = (Cmp == AArch64CC::GT) ? 1 : -1;
252   // Negate Correction value for comparison with negative immediate (CMN).
253   if (Negative) {
254     Correction = -Correction;
255   }
256 
257   const int OldImm = (int)CmpMI->getOperand(2).getImm();
258   const int NewImm = std::abs(OldImm + Correction);
259 
260   // Handle +0 -> -1 and -0 -> +1 (CMN with 0 immediate) transitions by
261   // adjusting compare instruction opcode.
262   if (OldImm == 0 && ((Negative && Correction == 1) ||
263                       (!Negative && Correction == -1))) {
264     Opc = getComplementOpc(Opc);
265   }
266 
267   return CmpInfo(NewImm, Opc, getAdjustedCmp(Cmp));
268 }
269 
270 // Applies changes to comparison instruction suggested by adjustCmp().
271 void AArch64ConditionOptimizer::modifyCmp(MachineInstr *CmpMI,
272     const CmpInfo &Info) {
273   int Imm;
274   unsigned Opc;
275   AArch64CC::CondCode Cmp;
276   std::tie(Imm, Opc, Cmp) = Info;
277 
278   MachineBasicBlock *const MBB = CmpMI->getParent();
279 
280   // Change immediate in comparison instruction (ADDS or SUBS).
281   BuildMI(*MBB, CmpMI, CmpMI->getDebugLoc(), TII->get(Opc))
282       .add(CmpMI->getOperand(0))
283       .add(CmpMI->getOperand(1))
284       .addImm(Imm)
285       .add(CmpMI->getOperand(3));
286   CmpMI->eraseFromParent();
287 
288   // The fact that this comparison was picked ensures that it's related to the
289   // first terminator instruction.
290   MachineInstr &BrMI = *MBB->getFirstTerminator();
291 
292   // Change condition in branch instruction.
293   BuildMI(*MBB, BrMI, BrMI.getDebugLoc(), TII->get(AArch64::Bcc))
294       .addImm(Cmp)
295       .add(BrMI.getOperand(1));
296   BrMI.eraseFromParent();
297 
298   ++NumConditionsAdjusted;
299 }
300 
301 // Parse a condition code returned by analyzeBranch, and compute the CondCode
302 // corresponding to TBB.
303 // Returns true if parsing was successful, otherwise false is returned.
304 static bool parseCond(ArrayRef<MachineOperand> Cond, AArch64CC::CondCode &CC) {
305   // A normal br.cond simply has the condition code.
306   if (Cond[0].getImm() != -1) {
307     assert(Cond.size() == 1 && "Unknown Cond array format");
308     CC = (AArch64CC::CondCode)(int)Cond[0].getImm();
309     return true;
310   }
311   return false;
312 }
313 
314 // Adjusts one cmp instruction to another one if result of adjustment will allow
315 // CSE.  Returns true if compare instruction was changed, otherwise false is
316 // returned.
317 bool AArch64ConditionOptimizer::adjustTo(MachineInstr *CmpMI,
318   AArch64CC::CondCode Cmp, MachineInstr *To, int ToImm)
319 {
320   CmpInfo Info = adjustCmp(CmpMI, Cmp);
321   if (std::get<0>(Info) == ToImm && std::get<1>(Info) == To->getOpcode()) {
322     modifyCmp(CmpMI, Info);
323     return true;
324   }
325   return false;
326 }
327 
328 bool AArch64ConditionOptimizer::runOnMachineFunction(MachineFunction &MF) {
329   LLVM_DEBUG(dbgs() << "********** AArch64 Conditional Compares **********\n"
330                     << "********** Function: " << MF.getName() << '\n');
331   if (skipFunction(MF.getFunction()))
332     return false;
333 
334   TII = MF.getSubtarget().getInstrInfo();
335   DomTree = &getAnalysis<MachineDominatorTree>();
336   MRI = &MF.getRegInfo();
337 
338   bool Changed = false;
339 
340   // Visit blocks in dominator tree pre-order. The pre-order enables multiple
341   // cmp-conversions from the same head block.
342   // Note that updateDomTree() modifies the children of the DomTree node
343   // currently being visited. The df_iterator supports that; it doesn't look at
344   // child_begin() / child_end() until after a node has been visited.
345   for (MachineDomTreeNode *I : depth_first(DomTree)) {
346     MachineBasicBlock *HBB = I->getBlock();
347 
348     SmallVector<MachineOperand, 4> HeadCond;
349     MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
350     if (TII->analyzeBranch(*HBB, TBB, FBB, HeadCond)) {
351       continue;
352     }
353 
354     // Equivalence check is to skip loops.
355     if (!TBB || TBB == HBB) {
356       continue;
357     }
358 
359     SmallVector<MachineOperand, 4> TrueCond;
360     MachineBasicBlock *TBB_TBB = nullptr, *TBB_FBB = nullptr;
361     if (TII->analyzeBranch(*TBB, TBB_TBB, TBB_FBB, TrueCond)) {
362       continue;
363     }
364 
365     MachineInstr *HeadCmpMI = findSuitableCompare(HBB);
366     if (!HeadCmpMI) {
367       continue;
368     }
369 
370     MachineInstr *TrueCmpMI = findSuitableCompare(TBB);
371     if (!TrueCmpMI) {
372       continue;
373     }
374 
375     AArch64CC::CondCode HeadCmp;
376     if (HeadCond.empty() || !parseCond(HeadCond, HeadCmp)) {
377       continue;
378     }
379 
380     AArch64CC::CondCode TrueCmp;
381     if (TrueCond.empty() || !parseCond(TrueCond, TrueCmp)) {
382       continue;
383     }
384 
385     const int HeadImm = (int)HeadCmpMI->getOperand(2).getImm();
386     const int TrueImm = (int)TrueCmpMI->getOperand(2).getImm();
387 
388     LLVM_DEBUG(dbgs() << "Head branch:\n");
389     LLVM_DEBUG(dbgs() << "\tcondition: " << AArch64CC::getCondCodeName(HeadCmp)
390                       << '\n');
391     LLVM_DEBUG(dbgs() << "\timmediate: " << HeadImm << '\n');
392 
393     LLVM_DEBUG(dbgs() << "True branch:\n");
394     LLVM_DEBUG(dbgs() << "\tcondition: " << AArch64CC::getCondCodeName(TrueCmp)
395                       << '\n');
396     LLVM_DEBUG(dbgs() << "\timmediate: " << TrueImm << '\n');
397 
398     if (((HeadCmp == AArch64CC::GT && TrueCmp == AArch64CC::LT) ||
399          (HeadCmp == AArch64CC::LT && TrueCmp == AArch64CC::GT)) &&
400         std::abs(TrueImm - HeadImm) == 2) {
401       // This branch transforms machine instructions that correspond to
402       //
403       // 1) (a > {TrueImm} && ...) || (a < {HeadImm} && ...)
404       // 2) (a < {TrueImm} && ...) || (a > {HeadImm} && ...)
405       //
406       // into
407       //
408       // 1) (a >= {NewImm} && ...) || (a <= {NewImm} && ...)
409       // 2) (a <= {NewImm} && ...) || (a >= {NewImm} && ...)
410 
411       CmpInfo HeadCmpInfo = adjustCmp(HeadCmpMI, HeadCmp);
412       CmpInfo TrueCmpInfo = adjustCmp(TrueCmpMI, TrueCmp);
413       if (std::get<0>(HeadCmpInfo) == std::get<0>(TrueCmpInfo) &&
414           std::get<1>(HeadCmpInfo) == std::get<1>(TrueCmpInfo)) {
415         modifyCmp(HeadCmpMI, HeadCmpInfo);
416         modifyCmp(TrueCmpMI, TrueCmpInfo);
417         Changed = true;
418       }
419     } else if (((HeadCmp == AArch64CC::GT && TrueCmp == AArch64CC::GT) ||
420                 (HeadCmp == AArch64CC::LT && TrueCmp == AArch64CC::LT)) &&
421                 std::abs(TrueImm - HeadImm) == 1) {
422       // This branch transforms machine instructions that correspond to
423       //
424       // 1) (a > {TrueImm} && ...) || (a > {HeadImm} && ...)
425       // 2) (a < {TrueImm} && ...) || (a < {HeadImm} && ...)
426       //
427       // into
428       //
429       // 1) (a <= {NewImm} && ...) || (a >  {NewImm} && ...)
430       // 2) (a <  {NewImm} && ...) || (a >= {NewImm} && ...)
431 
432       // GT -> GE transformation increases immediate value, so picking the
433       // smaller one; LT -> LE decreases immediate value so invert the choice.
434       bool adjustHeadCond = (HeadImm < TrueImm);
435       if (HeadCmp == AArch64CC::LT) {
436           adjustHeadCond = !adjustHeadCond;
437       }
438 
439       if (adjustHeadCond) {
440         Changed |= adjustTo(HeadCmpMI, HeadCmp, TrueCmpMI, TrueImm);
441       } else {
442         Changed |= adjustTo(TrueCmpMI, TrueCmp, HeadCmpMI, HeadImm);
443       }
444     }
445     // Other transformation cases almost never occur due to generation of < or >
446     // comparisons instead of <= and >=.
447   }
448 
449   return Changed;
450 }
451