1//===-------- X86PadShortFunction.cpp - pad short functions -----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the pass which will pad short functions to prevent
11// a stall if a function returns before the return address is ready. This
12// is needed for some Intel Atom processors.
13//
14//===----------------------------------------------------------------------===//
15
16#include <algorithm>
17
18#include "X86.h"
19#include "X86InstrInfo.h"
20#include "X86Subtarget.h"
21#include "llvm/ADT/Statistic.h"
22#include "llvm/CodeGen/MachineFunctionPass.h"
23#include "llvm/CodeGen/MachineInstrBuilder.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/CodeGen/Passes.h"
26#include "llvm/IR/Function.h"
27#include "llvm/Support/Debug.h"
28#include "llvm/Support/raw_ostream.h"
29#include "llvm/Target/TargetInstrInfo.h"
30
31using namespace llvm;
32
33#define DEBUG_TYPE "x86-pad-short-functions"
34
35STATISTIC(NumBBsPadded, "Number of basic blocks padded");
36
37namespace {
38  struct VisitedBBInfo {
39    // HasReturn - Whether the BB contains a return instruction
40    bool HasReturn;
41
42    // Cycles - Number of cycles until return if HasReturn is true, otherwise
43    // number of cycles until end of the BB
44    unsigned int Cycles;
45
46    VisitedBBInfo() : HasReturn(false), Cycles(0) {}
47    VisitedBBInfo(bool HasReturn, unsigned int Cycles)
48      : HasReturn(HasReturn), Cycles(Cycles) {}
49  };
50
51  struct PadShortFunc : public MachineFunctionPass {
52    static char ID;
53    PadShortFunc() : MachineFunctionPass(ID)
54                   , Threshold(4), STI(nullptr), TII(nullptr) {}
55
56    bool runOnMachineFunction(MachineFunction &MF) override;
57
58    MachineFunctionProperties getRequiredProperties() const override {
59      return MachineFunctionProperties().set(
60          MachineFunctionProperties::Property::AllVRegsAllocated);
61    }
62
63    const char *getPassName() const override {
64      return "X86 Atom pad short functions";
65    }
66
67  private:
68    void findReturns(MachineBasicBlock *MBB,
69                     unsigned int Cycles = 0);
70
71    bool cyclesUntilReturn(MachineBasicBlock *MBB,
72                           unsigned int &Cycles);
73
74    void addPadding(MachineBasicBlock *MBB,
75                    MachineBasicBlock::iterator &MBBI,
76                    unsigned int NOOPsToAdd);
77
78    const unsigned int Threshold;
79
80    // ReturnBBs - Maps basic blocks that return to the minimum number of
81    // cycles until the return, starting from the entry block.
82    DenseMap<MachineBasicBlock*, unsigned int> ReturnBBs;
83
84    // VisitedBBs - Cache of previously visited BBs.
85    DenseMap<MachineBasicBlock*, VisitedBBInfo> VisitedBBs;
86
87    const X86Subtarget *STI;
88    const TargetInstrInfo *TII;
89  };
90
91  char PadShortFunc::ID = 0;
92}
93
94FunctionPass *llvm::createX86PadShortFunctions() {
95  return new PadShortFunc();
96}
97
98/// runOnMachineFunction - Loop over all of the basic blocks, inserting
99/// NOOP instructions before early exits.
100bool PadShortFunc::runOnMachineFunction(MachineFunction &MF) {
101  if (skipFunction(*MF.getFunction()))
102    return false;
103
104  if (MF.getFunction()->optForSize()) {
105    return false;
106  }
107
108  STI = &MF.getSubtarget<X86Subtarget>();
109  if (!STI->padShortFunctions())
110    return false;
111
112  TII = STI->getInstrInfo();
113
114  // Search through basic blocks and mark the ones that have early returns
115  ReturnBBs.clear();
116  VisitedBBs.clear();
117  findReturns(&MF.front());
118
119  bool MadeChange = false;
120
121  MachineBasicBlock *MBB;
122  unsigned int Cycles = 0;
123
124  // Pad the identified basic blocks with NOOPs
125  for (DenseMap<MachineBasicBlock*, unsigned int>::iterator I = ReturnBBs.begin();
126       I != ReturnBBs.end(); ++I) {
127    MBB = I->first;
128    Cycles = I->second;
129
130    if (Cycles < Threshold) {
131      // BB ends in a return. Skip over any DBG_VALUE instructions
132      // trailing the terminator.
133      assert(MBB->size() > 0 &&
134             "Basic block should contain at least a RET but is empty");
135      MachineBasicBlock::iterator ReturnLoc = --MBB->end();
136
137      while (ReturnLoc->isDebugValue())
138        --ReturnLoc;
139      assert(ReturnLoc->isReturn() && !ReturnLoc->isCall() &&
140             "Basic block does not end with RET");
141
142      addPadding(MBB, ReturnLoc, Threshold - Cycles);
143      NumBBsPadded++;
144      MadeChange = true;
145    }
146  }
147
148  return MadeChange;
149}
150
151/// findReturn - Starting at MBB, follow control flow and add all
152/// basic blocks that contain a return to ReturnBBs.
153void PadShortFunc::findReturns(MachineBasicBlock *MBB, unsigned int Cycles) {
154  // If this BB has a return, note how many cycles it takes to get there.
155  bool hasReturn = cyclesUntilReturn(MBB, Cycles);
156  if (Cycles >= Threshold)
157    return;
158
159  if (hasReturn) {
160    ReturnBBs[MBB] = std::max(ReturnBBs[MBB], Cycles);
161    return;
162  }
163
164  // Follow branches in BB and look for returns
165  for (MachineBasicBlock::succ_iterator I = MBB->succ_begin();
166       I != MBB->succ_end(); ++I) {
167    if (*I == MBB)
168      continue;
169    findReturns(*I, Cycles);
170  }
171}
172
173/// cyclesUntilReturn - return true if the MBB has a return instruction,
174/// and return false otherwise.
175/// Cycles will be incremented by the number of cycles taken to reach the
176/// return or the end of the BB, whichever occurs first.
177bool PadShortFunc::cyclesUntilReturn(MachineBasicBlock *MBB,
178                                     unsigned int &Cycles) {
179  // Return cached result if BB was previously visited
180  DenseMap<MachineBasicBlock*, VisitedBBInfo>::iterator it
181    = VisitedBBs.find(MBB);
182  if (it != VisitedBBs.end()) {
183    VisitedBBInfo BBInfo = it->second;
184    Cycles += BBInfo.Cycles;
185    return BBInfo.HasReturn;
186  }
187
188  unsigned int CyclesToEnd = 0;
189
190  for (MachineInstr &MI : *MBB) {
191    // Mark basic blocks with a return instruction. Calls to other
192    // functions do not count because the called function will be padded,
193    // if necessary.
194    if (MI.isReturn() && !MI.isCall()) {
195      VisitedBBs[MBB] = VisitedBBInfo(true, CyclesToEnd);
196      Cycles += CyclesToEnd;
197      return true;
198    }
199
200    CyclesToEnd += TII->getInstrLatency(STI->getInstrItineraryData(), MI);
201  }
202
203  VisitedBBs[MBB] = VisitedBBInfo(false, CyclesToEnd);
204  Cycles += CyclesToEnd;
205  return false;
206}
207
208/// addPadding - Add the given number of NOOP instructions to the function
209/// just prior to the return at MBBI
210void PadShortFunc::addPadding(MachineBasicBlock *MBB,
211                              MachineBasicBlock::iterator &MBBI,
212                              unsigned int NOOPsToAdd) {
213  DebugLoc DL = MBBI->getDebugLoc();
214
215  while (NOOPsToAdd-- > 0) {
216    BuildMI(*MBB, MBBI, DL, TII->get(X86::NOOP));
217    BuildMI(*MBB, MBBI, DL, TII->get(X86::NOOP));
218  }
219}
220