1//===-- PPCBranchSelector.cpp - Emit long conditional branches ------------===//
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 contains a pass that scans a machine function to determine which
10// conditional branches need more than 16 bits of displacement to reach their
11// target basic block.  It does this in two passes; a calculation of basic block
12// positions pass, and a branch pseudo op to machine branch opcode pass.  This
13// pass should be run last, just before the assembly printer.
14//
15//===----------------------------------------------------------------------===//
16
17#include "MCTargetDesc/PPCPredicates.h"
18#include "PPC.h"
19#include "PPCInstrBuilder.h"
20#include "PPCInstrInfo.h"
21#include "PPCSubtarget.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/CodeGen/MachineFunctionPass.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/CodeGen/TargetSubtargetInfo.h"
26#include "llvm/Support/MathExtras.h"
27#include "llvm/Target/TargetMachine.h"
28#include <algorithm>
29using namespace llvm;
30
31#define DEBUG_TYPE "ppc-branch-select"
32
33STATISTIC(NumExpanded, "Number of branches expanded to long format");
34STATISTIC(NumPrefixed, "Number of prefixed instructions");
35STATISTIC(NumPrefixedAligned,
36          "Number of prefixed instructions that have been aligned");
37
38namespace {
39  struct PPCBSel : public MachineFunctionPass {
40    static char ID;
41    PPCBSel() : MachineFunctionPass(ID) {
42      initializePPCBSelPass(*PassRegistry::getPassRegistry());
43    }
44
45    // The sizes of the basic blocks in the function (the first
46    // element of the pair); the second element of the pair is the amount of the
47    // size that is due to potential padding.
48    std::vector<std::pair<unsigned, unsigned>> BlockSizes;
49
50    // The first block number which has imprecise instruction address.
51    int FirstImpreciseBlock = -1;
52
53    unsigned GetAlignmentAdjustment(MachineBasicBlock &MBB, unsigned Offset);
54    unsigned ComputeBlockSizes(MachineFunction &Fn);
55    void modifyAdjustment(MachineFunction &Fn);
56    int computeBranchSize(MachineFunction &Fn,
57                          const MachineBasicBlock *Src,
58                          const MachineBasicBlock *Dest,
59                          unsigned BrOffset);
60
61    bool runOnMachineFunction(MachineFunction &Fn) override;
62
63    MachineFunctionProperties getRequiredProperties() const override {
64      return MachineFunctionProperties().set(
65          MachineFunctionProperties::Property::NoVRegs);
66    }
67
68    StringRef getPassName() const override { return "PowerPC Branch Selector"; }
69  };
70  char PPCBSel::ID = 0;
71}
72
73INITIALIZE_PASS(PPCBSel, "ppc-branch-select", "PowerPC Branch Selector",
74                false, false)
75
76/// createPPCBranchSelectionPass - returns an instance of the Branch Selection
77/// Pass
78///
79FunctionPass *llvm::createPPCBranchSelectionPass() {
80  return new PPCBSel();
81}
82
83/// In order to make MBB aligned, we need to add an adjustment value to the
84/// original Offset.
85unsigned PPCBSel::GetAlignmentAdjustment(MachineBasicBlock &MBB,
86                                         unsigned Offset) {
87  const Align Alignment = MBB.getAlignment();
88  if (Alignment == Align(1))
89    return 0;
90
91  const Align ParentAlign = MBB.getParent()->getAlignment();
92
93  if (Alignment <= ParentAlign)
94    return offsetToAlignment(Offset, Alignment);
95
96  // The alignment of this MBB is larger than the function's alignment, so we
97  // can't tell whether or not it will insert nops. Assume that it will.
98  if (FirstImpreciseBlock < 0)
99    FirstImpreciseBlock = MBB.getNumber();
100  return Alignment.value() + offsetToAlignment(Offset, Alignment);
101}
102
103/// We need to be careful about the offset of the first block in the function
104/// because it might not have the function's alignment. This happens because,
105/// under the ELFv2 ABI, for functions which require a TOC pointer, we add a
106/// two-instruction sequence to the start of the function.
107/// Note: This needs to be synchronized with the check in
108/// PPCLinuxAsmPrinter::EmitFunctionBodyStart.
109static inline unsigned GetInitialOffset(MachineFunction &Fn) {
110  unsigned InitialOffset = 0;
111  if (Fn.getSubtarget<PPCSubtarget>().isELFv2ABI() &&
112      !Fn.getRegInfo().use_empty(PPC::X2))
113    InitialOffset = 8;
114  return InitialOffset;
115}
116
117/// Measure each MBB and compute a size for the entire function.
118unsigned PPCBSel::ComputeBlockSizes(MachineFunction &Fn) {
119  const PPCInstrInfo *TII =
120      static_cast<const PPCInstrInfo *>(Fn.getSubtarget().getInstrInfo());
121  unsigned FuncSize = GetInitialOffset(Fn);
122
123  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
124       ++MFI) {
125    MachineBasicBlock *MBB = &*MFI;
126
127    // The end of the previous block may have extra nops if this block has an
128    // alignment requirement.
129    if (MBB->getNumber() > 0) {
130      unsigned AlignExtra = GetAlignmentAdjustment(*MBB, FuncSize);
131
132      auto &BS = BlockSizes[MBB->getNumber()-1];
133      BS.first += AlignExtra;
134      BS.second = AlignExtra;
135
136      FuncSize += AlignExtra;
137    }
138
139    unsigned BlockSize = 0;
140    unsigned UnalignedBytesRemaining = 0;
141    for (MachineInstr &MI : *MBB) {
142      unsigned MINumBytes = TII->getInstSizeInBytes(MI);
143      if (MI.isInlineAsm() && (FirstImpreciseBlock < 0))
144        FirstImpreciseBlock = MBB->getNumber();
145      if (TII->isPrefixed(MI.getOpcode())) {
146        NumPrefixed++;
147
148        // All 8 byte instructions may require alignment. Each 8 byte
149        // instruction may be aligned by another 4 bytes.
150        // This means that an 8 byte instruction may require 12 bytes
151        // (8 for the instruction itself and 4 for the alignment nop).
152        // This will happen if an 8 byte instruction can be aligned to 64 bytes
153        // by only adding a 4 byte nop.
154        // We don't know the alignment at this point in the code so we have to
155        // adopt a more pessimistic approach. If an instruction may need
156        // alignment we assume that it does need alignment and add 4 bytes to
157        // it. As a result we may end up with more long branches than before
158        // but we are in the safe position where if we need a long branch we
159        // have one.
160        // The if statement checks to make sure that two 8 byte instructions
161        // are at least 64 bytes away from each other. It is not possible for
162        // two instructions that both need alignment to be within 64 bytes of
163        // each other.
164        if (!UnalignedBytesRemaining) {
165          BlockSize += 4;
166          UnalignedBytesRemaining = 60;
167          NumPrefixedAligned++;
168        }
169      }
170      UnalignedBytesRemaining -= std::min(UnalignedBytesRemaining, MINumBytes);
171      BlockSize += MINumBytes;
172    }
173
174    BlockSizes[MBB->getNumber()].first = BlockSize;
175    FuncSize += BlockSize;
176  }
177
178  return FuncSize;
179}
180
181/// Modify the basic block align adjustment.
182void PPCBSel::modifyAdjustment(MachineFunction &Fn) {
183  unsigned Offset = GetInitialOffset(Fn);
184  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
185       ++MFI) {
186    MachineBasicBlock *MBB = &*MFI;
187
188    if (MBB->getNumber() > 0) {
189      auto &BS = BlockSizes[MBB->getNumber()-1];
190      BS.first -= BS.second;
191      Offset -= BS.second;
192
193      unsigned AlignExtra = GetAlignmentAdjustment(*MBB, Offset);
194
195      BS.first += AlignExtra;
196      BS.second = AlignExtra;
197
198      Offset += AlignExtra;
199    }
200
201    Offset += BlockSizes[MBB->getNumber()].first;
202  }
203}
204
205/// Determine the offset from the branch in Src block to the Dest block.
206/// BrOffset is the offset of the branch instruction inside Src block.
207int PPCBSel::computeBranchSize(MachineFunction &Fn,
208                               const MachineBasicBlock *Src,
209                               const MachineBasicBlock *Dest,
210                               unsigned BrOffset) {
211  int BranchSize;
212  Align MaxAlign = Align(4);
213  bool NeedExtraAdjustment = false;
214  if (Dest->getNumber() <= Src->getNumber()) {
215    // If this is a backwards branch, the delta is the offset from the
216    // start of this block to this branch, plus the sizes of all blocks
217    // from this block to the dest.
218    BranchSize = BrOffset;
219    MaxAlign = std::max(MaxAlign, Src->getAlignment());
220
221    int DestBlock = Dest->getNumber();
222    BranchSize += BlockSizes[DestBlock].first;
223    for (unsigned i = DestBlock+1, e = Src->getNumber(); i < e; ++i) {
224      BranchSize += BlockSizes[i].first;
225      MaxAlign = std::max(MaxAlign, Fn.getBlockNumbered(i)->getAlignment());
226    }
227
228    NeedExtraAdjustment = (FirstImpreciseBlock >= 0) &&
229                          (DestBlock >= FirstImpreciseBlock);
230  } else {
231    // Otherwise, add the size of the blocks between this block and the
232    // dest to the number of bytes left in this block.
233    unsigned StartBlock = Src->getNumber();
234    BranchSize = BlockSizes[StartBlock].first - BrOffset;
235
236    MaxAlign = std::max(MaxAlign, Dest->getAlignment());
237    for (unsigned i = StartBlock+1, e = Dest->getNumber(); i != e; ++i) {
238      BranchSize += BlockSizes[i].first;
239      MaxAlign = std::max(MaxAlign, Fn.getBlockNumbered(i)->getAlignment());
240    }
241
242    NeedExtraAdjustment = (FirstImpreciseBlock >= 0) &&
243                          (Src->getNumber() >= FirstImpreciseBlock);
244  }
245
246  // We tend to over estimate code size due to large alignment and
247  // inline assembly. Usually it causes larger computed branch offset.
248  // But sometimes it may also causes smaller computed branch offset
249  // than actual branch offset. If the offset is close to the limit of
250  // encoding, it may cause problem at run time.
251  // Following is a simplified example.
252  //
253  //              actual        estimated
254  //              address        address
255  //    ...
256  //   bne Far      100            10c
257  //   .p2align 4
258  //   Near:        110            110
259  //    ...
260  //   Far:        8108           8108
261  //
262  //   Actual offset:    0x8108 - 0x100 = 0x8008
263  //   Computed offset:  0x8108 - 0x10c = 0x7ffc
264  //
265  // This example also shows when we can get the largest gap between
266  // estimated offset and actual offset. If there is an aligned block
267  // ABB between branch and target, assume its alignment is <align>
268  // bits. Now consider the accumulated function size FSIZE till the end
269  // of previous block PBB. If the estimated FSIZE is multiple of
270  // 2^<align>, we don't need any padding for the estimated address of
271  // ABB. If actual FSIZE at the end of PBB is 4 bytes more than
272  // multiple of 2^<align>, then we need (2^<align> - 4) bytes of
273  // padding. It also means the actual branch offset is (2^<align> - 4)
274  // larger than computed offset. Other actual FSIZE needs less padding
275  // bytes, so causes smaller gap between actual and computed offset.
276  //
277  // On the other hand, if the inline asm or large alignment occurs
278  // between the branch block and destination block, the estimated address
279  // can be <delta> larger than actual address. If padding bytes are
280  // needed for a later aligned block, the actual number of padding bytes
281  // is at most <delta> more than estimated padding bytes. So the actual
282  // aligned block address is less than or equal to the estimated aligned
283  // block address. So the actual branch offset is less than or equal to
284  // computed branch offset.
285  //
286  // The computed offset is at most ((1 << alignment) - 4) bytes smaller
287  // than actual offset. So we add this number to the offset for safety.
288  if (NeedExtraAdjustment)
289    BranchSize += MaxAlign.value() - 4;
290
291  return BranchSize;
292}
293
294bool PPCBSel::runOnMachineFunction(MachineFunction &Fn) {
295  const PPCInstrInfo *TII =
296      static_cast<const PPCInstrInfo *>(Fn.getSubtarget().getInstrInfo());
297  // Give the blocks of the function a dense, in-order, numbering.
298  Fn.RenumberBlocks();
299  BlockSizes.resize(Fn.getNumBlockIDs());
300  FirstImpreciseBlock = -1;
301
302  // Measure each MBB and compute a size for the entire function.
303  unsigned FuncSize = ComputeBlockSizes(Fn);
304
305  // If the entire function is smaller than the displacement of a branch field,
306  // we know we don't need to shrink any branches in this function.  This is a
307  // common case.
308  if (FuncSize < (1 << 15)) {
309    BlockSizes.clear();
310    return false;
311  }
312
313  // For each conditional branch, if the offset to its destination is larger
314  // than the offset field allows, transform it into a long branch sequence
315  // like this:
316  //   short branch:
317  //     bCC MBB
318  //   long branch:
319  //     b!CC $PC+8
320  //     b MBB
321  //
322  bool MadeChange = true;
323  bool EverMadeChange = false;
324  while (MadeChange) {
325    // Iteratively expand branches until we reach a fixed point.
326    MadeChange = false;
327
328    for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
329         ++MFI) {
330      MachineBasicBlock &MBB = *MFI;
331      unsigned MBBStartOffset = 0;
332      for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
333           I != E; ++I) {
334        MachineBasicBlock *Dest = nullptr;
335        if (I->getOpcode() == PPC::BCC && !I->getOperand(2).isImm())
336          Dest = I->getOperand(2).getMBB();
337        else if ((I->getOpcode() == PPC::BC || I->getOpcode() == PPC::BCn) &&
338                 !I->getOperand(1).isImm())
339          Dest = I->getOperand(1).getMBB();
340        else if ((I->getOpcode() == PPC::BDNZ8 || I->getOpcode() == PPC::BDNZ ||
341                  I->getOpcode() == PPC::BDZ8  || I->getOpcode() == PPC::BDZ) &&
342                 !I->getOperand(0).isImm())
343          Dest = I->getOperand(0).getMBB();
344
345        if (!Dest) {
346          MBBStartOffset += TII->getInstSizeInBytes(*I);
347          continue;
348        }
349
350        // Determine the offset from the current branch to the destination
351        // block.
352        int BranchSize = computeBranchSize(Fn, &MBB, Dest, MBBStartOffset);
353
354        // If this branch is in range, ignore it.
355        if (isInt<16>(BranchSize)) {
356          MBBStartOffset += 4;
357          continue;
358        }
359
360        // Otherwise, we have to expand it to a long branch.
361        MachineInstr &OldBranch = *I;
362        DebugLoc dl = OldBranch.getDebugLoc();
363
364        if (I->getOpcode() == PPC::BCC) {
365          // The BCC operands are:
366          // 0. PPC branch predicate
367          // 1. CR register
368          // 2. Target MBB
369          PPC::Predicate Pred = (PPC::Predicate)I->getOperand(0).getImm();
370          Register CRReg = I->getOperand(1).getReg();
371
372          // Jump over the uncond branch inst (i.e. $PC+8) on opposite condition.
373          BuildMI(MBB, I, dl, TII->get(PPC::BCC))
374            .addImm(PPC::InvertPredicate(Pred)).addReg(CRReg).addImm(2);
375        } else if (I->getOpcode() == PPC::BC) {
376          Register CRBit = I->getOperand(0).getReg();
377          BuildMI(MBB, I, dl, TII->get(PPC::BCn)).addReg(CRBit).addImm(2);
378        } else if (I->getOpcode() == PPC::BCn) {
379          Register CRBit = I->getOperand(0).getReg();
380          BuildMI(MBB, I, dl, TII->get(PPC::BC)).addReg(CRBit).addImm(2);
381        } else if (I->getOpcode() == PPC::BDNZ) {
382          BuildMI(MBB, I, dl, TII->get(PPC::BDZ)).addImm(2);
383        } else if (I->getOpcode() == PPC::BDNZ8) {
384          BuildMI(MBB, I, dl, TII->get(PPC::BDZ8)).addImm(2);
385        } else if (I->getOpcode() == PPC::BDZ) {
386          BuildMI(MBB, I, dl, TII->get(PPC::BDNZ)).addImm(2);
387        } else if (I->getOpcode() == PPC::BDZ8) {
388          BuildMI(MBB, I, dl, TII->get(PPC::BDNZ8)).addImm(2);
389        } else {
390           llvm_unreachable("Unhandled branch type!");
391        }
392
393        // Uncond branch to the real destination.
394        I = BuildMI(MBB, I, dl, TII->get(PPC::B)).addMBB(Dest);
395
396        // Remove the old branch from the function.
397        OldBranch.eraseFromParent();
398
399        // Remember that this instruction is 8-bytes, increase the size of the
400        // block by 4, remember to iterate.
401        BlockSizes[MBB.getNumber()].first += 4;
402        MBBStartOffset += 8;
403        ++NumExpanded;
404        MadeChange = true;
405      }
406    }
407
408    if (MadeChange) {
409      // If we're going to iterate again, make sure we've updated our
410      // padding-based contributions to the block sizes.
411      modifyAdjustment(Fn);
412    }
413
414    EverMadeChange |= MadeChange;
415  }
416
417  BlockSizes.clear();
418  return true;
419}
420