xref: /llvm-project/llvm/lib/Target/ARM/Thumb2ITBlockPass.cpp (revision a35db2880a488b62a16f269972ad885fd58033f7)
1 //===-- Thumb2ITBlockPass.cpp - Insert Thumb-2 IT blocks ------------------===//
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 #include "ARM.h"
10 #include "ARMMachineFunctionInfo.h"
11 #include "ARMSubtarget.h"
12 #include "Thumb2InstrInfo.h"
13 #include "llvm/ADT/SmallSet.h"
14 #include "llvm/ADT/SmallVector.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/CodeGen/MachineBasicBlock.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/CodeGen/MachineInstr.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineInstrBundle.h"
23 #include "llvm/CodeGen/MachineOperand.h"
24 #include "llvm/IR/DebugLoc.h"
25 #include "llvm/MC/MCInstrDesc.h"
26 #include <cassert>
27 #include <new>
28 
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "thumb2-it"
32 #define PASS_NAME "Thumb IT blocks insertion pass"
33 
34 STATISTIC(NumITs,        "Number of IT blocks inserted");
35 STATISTIC(NumMovedInsts, "Number of predicated instructions moved");
36 
37 using RegisterSet = SmallSet<unsigned, 4>;
38 
39 namespace {
40 
41   class Thumb2ITBlock : public MachineFunctionPass {
42   public:
43     static char ID;
44 
45     bool restrictIT;
46     const Thumb2InstrInfo *TII;
47     const TargetRegisterInfo *TRI;
48     ARMFunctionInfo *AFI;
49 
50     Thumb2ITBlock() : MachineFunctionPass(ID) {}
51 
52     bool runOnMachineFunction(MachineFunction &Fn) override;
53 
54     MachineFunctionProperties getRequiredProperties() const override {
55       return MachineFunctionProperties().set(
56           MachineFunctionProperties::Property::NoVRegs);
57     }
58 
59     StringRef getPassName() const override {
60       return PASS_NAME;
61     }
62 
63   private:
64     bool MoveCopyOutOfITBlock(MachineInstr *MI,
65                               ARMCC::CondCodes CC, ARMCC::CondCodes OCC,
66                               RegisterSet &Defs, RegisterSet &Uses);
67     bool InsertITInstructions(MachineBasicBlock &Block);
68   };
69 
70   char Thumb2ITBlock::ID = 0;
71 
72 } // end anonymous namespace
73 
74 INITIALIZE_PASS(Thumb2ITBlock, DEBUG_TYPE, PASS_NAME, false, false)
75 
76 /// TrackDefUses - Tracking what registers are being defined and used by
77 /// instructions in the IT block. This also tracks "dependencies", i.e. uses
78 /// in the IT block that are defined before the IT instruction.
79 static void TrackDefUses(MachineInstr *MI, RegisterSet &Defs, RegisterSet &Uses,
80                          const TargetRegisterInfo *TRI) {
81   using RegList = SmallVector<unsigned, 4>;
82   RegList LocalDefs;
83   RegList LocalUses;
84 
85   for (auto &MO : MI->operands()) {
86     if (!MO.isReg())
87       continue;
88     Register Reg = MO.getReg();
89     if (!Reg || Reg == ARM::ITSTATE || Reg == ARM::SP)
90       continue;
91     if (MO.isUse())
92       LocalUses.push_back(Reg);
93     else
94       LocalDefs.push_back(Reg);
95   }
96 
97   auto InsertUsesDefs = [&](RegList &Regs, RegisterSet &UsesDefs) {
98     for (unsigned Reg : Regs)
99       for (MCPhysReg Subreg : TRI->subregs_inclusive(Reg))
100         UsesDefs.insert(Subreg);
101   };
102 
103   InsertUsesDefs(LocalDefs, Defs);
104   InsertUsesDefs(LocalUses, Uses);
105 }
106 
107 /// Clear kill flags for any uses in the given set.  This will likely
108 /// conservatively remove more kill flags than are necessary, but removing them
109 /// is safer than incorrect kill flags remaining on instructions.
110 static void ClearKillFlags(MachineInstr *MI, RegisterSet &Uses) {
111   for (MachineOperand &MO : MI->operands()) {
112     if (!MO.isReg() || MO.isDef() || !MO.isKill())
113       continue;
114     if (!Uses.count(MO.getReg()))
115       continue;
116     MO.setIsKill(false);
117   }
118 }
119 
120 static bool isCopy(MachineInstr *MI) {
121   switch (MI->getOpcode()) {
122   default:
123     return false;
124   case ARM::MOVr:
125   case ARM::MOVr_TC:
126   case ARM::tMOVr:
127   case ARM::t2MOVr:
128     return true;
129   }
130 }
131 
132 bool
133 Thumb2ITBlock::MoveCopyOutOfITBlock(MachineInstr *MI,
134                                     ARMCC::CondCodes CC, ARMCC::CondCodes OCC,
135                                     RegisterSet &Defs, RegisterSet &Uses) {
136   if (!isCopy(MI))
137     return false;
138   // llvm models select's as two-address instructions. That means a copy
139   // is inserted before a t2MOVccr, etc. If the copy is scheduled in
140   // between selects we would end up creating multiple IT blocks.
141   assert(MI->getOperand(0).getSubReg() == 0 &&
142          MI->getOperand(1).getSubReg() == 0 &&
143          "Sub-register indices still around?");
144 
145   Register DstReg = MI->getOperand(0).getReg();
146   Register SrcReg = MI->getOperand(1).getReg();
147 
148   // First check if it's safe to move it.
149   if (Uses.count(DstReg) || Defs.count(SrcReg))
150     return false;
151 
152   // If the CPSR is defined by this copy, then we don't want to move it. E.g.,
153   // if we have:
154   //
155   //   movs  r1, r1
156   //   rsb   r1, 0
157   //   movs  r2, r2
158   //   rsb   r2, 0
159   //
160   // we don't want this to be converted to:
161   //
162   //   movs  r1, r1
163   //   movs  r2, r2
164   //   itt   mi
165   //   rsb   r1, 0
166   //   rsb   r2, 0
167   //
168   const MCInstrDesc &MCID = MI->getDesc();
169   if (MI->hasOptionalDef() &&
170       MI->getOperand(MCID.getNumOperands() - 1).getReg() == ARM::CPSR)
171     return false;
172 
173   // Then peek at the next instruction to see if it's predicated on CC or OCC.
174   // If not, then there is nothing to be gained by moving the copy.
175   MachineBasicBlock::iterator I = MI;
176   ++I;
177   MachineBasicBlock::iterator E = MI->getParent()->end();
178 
179   while (I != E && I->isDebugInstr())
180     ++I;
181 
182   if (I != E) {
183     Register NPredReg;
184     ARMCC::CondCodes NCC = getITInstrPredicate(*I, NPredReg);
185     if (NCC == CC || NCC == OCC)
186       return true;
187   }
188   return false;
189 }
190 
191 bool Thumb2ITBlock::InsertITInstructions(MachineBasicBlock &MBB) {
192   bool Modified = false;
193   RegisterSet Defs, Uses;
194   MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
195 
196   while (MBBI != E) {
197     MachineInstr *MI = &*MBBI;
198     DebugLoc dl = MI->getDebugLoc();
199     Register PredReg;
200     ARMCC::CondCodes CC = getITInstrPredicate(*MI, PredReg);
201     if (CC == ARMCC::AL) {
202       ++MBBI;
203       continue;
204     }
205 
206     Defs.clear();
207     Uses.clear();
208     TrackDefUses(MI, Defs, Uses, TRI);
209 
210     // Insert an IT instruction.
211     MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(ARM::t2IT))
212       .addImm(CC);
213 
214     // Add implicit use of ITSTATE to IT block instructions.
215     MI->addOperand(MachineOperand::CreateReg(ARM::ITSTATE, false/*ifDef*/,
216                                              true/*isImp*/, false/*isKill*/));
217 
218     MachineInstr *LastITMI = MI;
219     MachineBasicBlock::iterator InsertPos = MIB.getInstr();
220     ++MBBI;
221 
222     // Form IT block.
223     ARMCC::CondCodes OCC = ARMCC::getOppositeCondition(CC);
224     unsigned Mask = 0, Pos = 3;
225 
226     // IT blocks are limited to one conditional op if -arm-restrict-it
227     // is set: skip the loop
228     if (!restrictIT) {
229       LLVM_DEBUG(dbgs() << "Allowing complex IT block\n");
230       // Branches, including tricky ones like LDM_RET, need to end an IT
231       // block so check the instruction we just put in the block.
232       for (; MBBI != E && Pos &&
233              (!MI->isBranch() && !MI->isReturn()) ; ++MBBI) {
234         if (MBBI->isDebugInstr())
235           continue;
236 
237         MachineInstr *NMI = &*MBBI;
238         MI = NMI;
239 
240         Register NPredReg;
241         ARMCC::CondCodes NCC = getITInstrPredicate(*NMI, NPredReg);
242         if (NCC == CC || NCC == OCC) {
243           Mask |= ((NCC ^ CC) & 1) << Pos;
244           // Add implicit use of ITSTATE.
245           NMI->addOperand(MachineOperand::CreateReg(ARM::ITSTATE, false/*ifDef*/,
246                                                  true/*isImp*/, false/*isKill*/));
247           LastITMI = NMI;
248         } else {
249           if (NCC == ARMCC::AL &&
250               MoveCopyOutOfITBlock(NMI, CC, OCC, Defs, Uses)) {
251             --MBBI;
252             MBB.remove(NMI);
253             MBB.insert(InsertPos, NMI);
254             ClearKillFlags(MI, Uses);
255             ++NumMovedInsts;
256             continue;
257           }
258           break;
259         }
260         TrackDefUses(NMI, Defs, Uses, TRI);
261         --Pos;
262       }
263     }
264 
265     // Finalize IT mask.
266     Mask |= (1 << Pos);
267     MIB.addImm(Mask);
268 
269     // Last instruction in IT block kills ITSTATE.
270     LastITMI->findRegisterUseOperand(ARM::ITSTATE, /*TRI=*/nullptr)
271         ->setIsKill();
272 
273     // Finalize the bundle.
274     finalizeBundle(MBB, InsertPos.getInstrIterator(),
275                    ++LastITMI->getIterator());
276 
277     Modified = true;
278     ++NumITs;
279   }
280 
281   return Modified;
282 }
283 
284 bool Thumb2ITBlock::runOnMachineFunction(MachineFunction &Fn) {
285   const ARMSubtarget &STI = Fn.getSubtarget<ARMSubtarget>();
286   if (!STI.isThumb2())
287     return false;
288   AFI = Fn.getInfo<ARMFunctionInfo>();
289   TII = static_cast<const Thumb2InstrInfo *>(STI.getInstrInfo());
290   TRI = STI.getRegisterInfo();
291   restrictIT = STI.restrictIT();
292 
293   if (!AFI->isThumbFunction())
294     return false;
295 
296   bool Modified = false;
297   for (auto &MBB : Fn )
298     Modified |= InsertITInstructions(MBB);
299 
300   if (Modified)
301     AFI->setHasITBlocks(true);
302 
303   return Modified;
304 }
305 
306 /// createThumb2ITBlockPass - Returns an instance of the Thumb2 IT blocks
307 /// insertion pass.
308 FunctionPass *llvm::createThumb2ITBlockPass() { return new Thumb2ITBlock(); }
309