LLVM  10.0.0svn
PPCBranchSelector.cpp
Go to the documentation of this file.
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 
18 #include "PPC.h"
19 #include "PPCInstrBuilder.h"
20 #include "PPCInstrInfo.h"
21 #include "PPCSubtarget.h"
22 #include "llvm/ADT/Statistic.h"
28 #include <algorithm>
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "ppc-branch-select"
32 
33 STATISTIC(NumExpanded, "Number of branches expanded to long format");
34 
35 namespace {
36  struct PPCBSel : public MachineFunctionPass {
37  static char ID;
38  PPCBSel() : MachineFunctionPass(ID) {
40  }
41 
42  // The sizes of the basic blocks in the function (the first
43  // element of the pair); the second element of the pair is the amount of the
44  // size that is due to potential padding.
45  std::vector<std::pair<unsigned, unsigned>> BlockSizes;
46 
47  // The first block number which has imprecise instruction address.
48  int FirstImpreciseBlock = -1;
49 
50  unsigned GetAlignmentAdjustment(MachineBasicBlock &MBB, unsigned Offset);
51  unsigned ComputeBlockSizes(MachineFunction &Fn);
52  void modifyAdjustment(MachineFunction &Fn);
53  int computeBranchSize(MachineFunction &Fn,
54  const MachineBasicBlock *Src,
55  const MachineBasicBlock *Dest,
56  unsigned BrOffset);
57 
58  bool runOnMachineFunction(MachineFunction &Fn) override;
59 
60  MachineFunctionProperties getRequiredProperties() const override {
63  }
64 
65  StringRef getPassName() const override { return "PowerPC Branch Selector"; }
66  };
67  char PPCBSel::ID = 0;
68 }
69 
70 INITIALIZE_PASS(PPCBSel, "ppc-branch-select", "PowerPC Branch Selector",
71  false, false)
72 
73 /// createPPCBranchSelectionPass - returns an instance of the Branch Selection
74 /// Pass
75 ///
77  return new PPCBSel();
78 }
79 
80 /// In order to make MBB aligned, we need to add an adjustment value to the
81 /// original Offset.
82 unsigned PPCBSel::GetAlignmentAdjustment(MachineBasicBlock &MBB,
83  unsigned Offset) {
84  unsigned Align = MBB.getAlignment();
85  if (!Align)
86  return 0;
87 
88  unsigned AlignAmt = 1 << Align;
89  unsigned ParentAlign = MBB.getParent()->getAlignment();
90 
91  if (Align <= ParentAlign)
92  return OffsetToAlignment(Offset, AlignAmt);
93 
94  // The alignment of this MBB is larger than the function's alignment, so we
95  // can't tell whether or not it will insert nops. Assume that it will.
96  if (FirstImpreciseBlock < 0)
97  FirstImpreciseBlock = MBB.getNumber();
98  return AlignAmt + OffsetToAlignment(Offset, AlignAmt);
99 }
100 
101 /// We need to be careful about the offset of the first block in the function
102 /// because it might not have the function's alignment. This happens because,
103 /// under the ELFv2 ABI, for functions which require a TOC pointer, we add a
104 /// two-instruction sequence to the start of the function.
105 /// Note: This needs to be synchronized with the check in
106 /// PPCLinuxAsmPrinter::EmitFunctionBodyStart.
107 static inline unsigned GetInitialOffset(MachineFunction &Fn) {
108  unsigned InitialOffset = 0;
109  if (Fn.getSubtarget<PPCSubtarget>().isELFv2ABI() &&
110  !Fn.getRegInfo().use_empty(PPC::X2))
111  InitialOffset = 8;
112  return InitialOffset;
113 }
114 
115 /// Measure each MBB and compute a size for the entire function.
116 unsigned PPCBSel::ComputeBlockSizes(MachineFunction &Fn) {
117  const PPCInstrInfo *TII =
118  static_cast<const PPCInstrInfo *>(Fn.getSubtarget().getInstrInfo());
119  unsigned FuncSize = GetInitialOffset(Fn);
120 
121  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
122  ++MFI) {
123  MachineBasicBlock *MBB = &*MFI;
124 
125  // The end of the previous block may have extra nops if this block has an
126  // alignment requirement.
127  if (MBB->getNumber() > 0) {
128  unsigned AlignExtra = GetAlignmentAdjustment(*MBB, FuncSize);
129 
130  auto &BS = BlockSizes[MBB->getNumber()-1];
131  BS.first += AlignExtra;
132  BS.second = AlignExtra;
133 
134  FuncSize += AlignExtra;
135  }
136 
137  unsigned BlockSize = 0;
138  for (MachineInstr &MI : *MBB) {
139  BlockSize += TII->getInstSizeInBytes(MI);
140  if (MI.isInlineAsm() && (FirstImpreciseBlock < 0))
141  FirstImpreciseBlock = MBB->getNumber();
142  }
143 
144  BlockSizes[MBB->getNumber()].first = BlockSize;
145  FuncSize += BlockSize;
146  }
147 
148  return FuncSize;
149 }
150 
151 /// Modify the basic block align adjustment.
152 void PPCBSel::modifyAdjustment(MachineFunction &Fn) {
153  unsigned Offset = GetInitialOffset(Fn);
154  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
155  ++MFI) {
156  MachineBasicBlock *MBB = &*MFI;
157 
158  if (MBB->getNumber() > 0) {
159  auto &BS = BlockSizes[MBB->getNumber()-1];
160  BS.first -= BS.second;
161  Offset -= BS.second;
162 
163  unsigned AlignExtra = GetAlignmentAdjustment(*MBB, Offset);
164 
165  BS.first += AlignExtra;
166  BS.second = AlignExtra;
167 
168  Offset += AlignExtra;
169  }
170 
171  Offset += BlockSizes[MBB->getNumber()].first;
172  }
173 }
174 
175 /// Determine the offset from the branch in Src block to the Dest block.
176 /// BrOffset is the offset of the branch instruction inside Src block.
177 int PPCBSel::computeBranchSize(MachineFunction &Fn,
178  const MachineBasicBlock *Src,
179  const MachineBasicBlock *Dest,
180  unsigned BrOffset) {
181  int BranchSize;
182  unsigned MaxAlign = 2;
183  bool NeedExtraAdjustment = false;
184  if (Dest->getNumber() <= Src->getNumber()) {
185  // If this is a backwards branch, the delta is the offset from the
186  // start of this block to this branch, plus the sizes of all blocks
187  // from this block to the dest.
188  BranchSize = BrOffset;
189  MaxAlign = std::max(MaxAlign, Src->getAlignment());
190 
191  int DestBlock = Dest->getNumber();
192  BranchSize += BlockSizes[DestBlock].first;
193  for (unsigned i = DestBlock+1, e = Src->getNumber(); i < e; ++i) {
194  BranchSize += BlockSizes[i].first;
195  MaxAlign = std::max(MaxAlign,
196  Fn.getBlockNumbered(i)->getAlignment());
197  }
198 
199  NeedExtraAdjustment = (FirstImpreciseBlock >= 0) &&
200  (DestBlock >= FirstImpreciseBlock);
201  } else {
202  // Otherwise, add the size of the blocks between this block and the
203  // dest to the number of bytes left in this block.
204  unsigned StartBlock = Src->getNumber();
205  BranchSize = BlockSizes[StartBlock].first - BrOffset;
206 
207  MaxAlign = std::max(MaxAlign, Dest->getAlignment());
208  for (unsigned i = StartBlock+1, e = Dest->getNumber(); i != e; ++i) {
209  BranchSize += BlockSizes[i].first;
210  MaxAlign = std::max(MaxAlign,
211  Fn.getBlockNumbered(i)->getAlignment());
212  }
213 
214  NeedExtraAdjustment = (FirstImpreciseBlock >= 0) &&
215  (Src->getNumber() >= FirstImpreciseBlock);
216  }
217 
218  // We tend to over estimate code size due to large alignment and
219  // inline assembly. Usually it causes larger computed branch offset.
220  // But sometimes it may also causes smaller computed branch offset
221  // than actual branch offset. If the offset is close to the limit of
222  // encoding, it may cause problem at run time.
223  // Following is a simplified example.
224  //
225  // actual estimated
226  // address address
227  // ...
228  // bne Far 100 10c
229  // .p2align 4
230  // Near: 110 110
231  // ...
232  // Far: 8108 8108
233  //
234  // Actual offset: 0x8108 - 0x100 = 0x8008
235  // Computed offset: 0x8108 - 0x10c = 0x7ffc
236  //
237  // This example also shows when we can get the largest gap between
238  // estimated offset and actual offset. If there is an aligned block
239  // ABB between branch and target, assume its alignment is <align>
240  // bits. Now consider the accumulated function size FSIZE till the end
241  // of previous block PBB. If the estimated FSIZE is multiple of
242  // 2^<align>, we don't need any padding for the estimated address of
243  // ABB. If actual FSIZE at the end of PBB is 4 bytes more than
244  // multiple of 2^<align>, then we need (2^<align> - 4) bytes of
245  // padding. It also means the actual branch offset is (2^<align> - 4)
246  // larger than computed offset. Other actual FSIZE needs less padding
247  // bytes, so causes smaller gap between actual and computed offset.
248  //
249  // On the other hand, if the inline asm or large alignment occurs
250  // between the branch block and destination block, the estimated address
251  // can be <delta> larger than actual address. If padding bytes are
252  // needed for a later aligned block, the actual number of padding bytes
253  // is at most <delta> more than estimated padding bytes. So the actual
254  // aligned block address is less than or equal to the estimated aligned
255  // block address. So the actual branch offset is less than or equal to
256  // computed branch offset.
257  //
258  // The computed offset is at most ((1 << alignment) - 4) bytes smaller
259  // than actual offset. So we add this number to the offset for safety.
260  if (NeedExtraAdjustment)
261  BranchSize += (1 << MaxAlign) - 4;
262 
263  return BranchSize;
264 }
265 
266 bool PPCBSel::runOnMachineFunction(MachineFunction &Fn) {
267  const PPCInstrInfo *TII =
268  static_cast<const PPCInstrInfo *>(Fn.getSubtarget().getInstrInfo());
269  // Give the blocks of the function a dense, in-order, numbering.
270  Fn.RenumberBlocks();
271  BlockSizes.resize(Fn.getNumBlockIDs());
272  FirstImpreciseBlock = -1;
273 
274  // Measure each MBB and compute a size for the entire function.
275  unsigned FuncSize = ComputeBlockSizes(Fn);
276 
277  // If the entire function is smaller than the displacement of a branch field,
278  // we know we don't need to shrink any branches in this function. This is a
279  // common case.
280  if (FuncSize < (1 << 15)) {
281  BlockSizes.clear();
282  return false;
283  }
284 
285  // For each conditional branch, if the offset to its destination is larger
286  // than the offset field allows, transform it into a long branch sequence
287  // like this:
288  // short branch:
289  // bCC MBB
290  // long branch:
291  // b!CC $PC+8
292  // b MBB
293  //
294  bool MadeChange = true;
295  bool EverMadeChange = false;
296  while (MadeChange) {
297  // Iteratively expand branches until we reach a fixed point.
298  MadeChange = false;
299 
300  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
301  ++MFI) {
302  MachineBasicBlock &MBB = *MFI;
303  unsigned MBBStartOffset = 0;
304  for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
305  I != E; ++I) {
306  MachineBasicBlock *Dest = nullptr;
307  if (I->getOpcode() == PPC::BCC && !I->getOperand(2).isImm())
308  Dest = I->getOperand(2).getMBB();
309  else if ((I->getOpcode() == PPC::BC || I->getOpcode() == PPC::BCn) &&
310  !I->getOperand(1).isImm())
311  Dest = I->getOperand(1).getMBB();
312  else if ((I->getOpcode() == PPC::BDNZ8 || I->getOpcode() == PPC::BDNZ ||
313  I->getOpcode() == PPC::BDZ8 || I->getOpcode() == PPC::BDZ) &&
314  !I->getOperand(0).isImm())
315  Dest = I->getOperand(0).getMBB();
316 
317  if (!Dest) {
318  MBBStartOffset += TII->getInstSizeInBytes(*I);
319  continue;
320  }
321 
322  // Determine the offset from the current branch to the destination
323  // block.
324  int BranchSize = computeBranchSize(Fn, &MBB, Dest, MBBStartOffset);
325 
326  // If this branch is in range, ignore it.
327  if (isInt<16>(BranchSize)) {
328  MBBStartOffset += 4;
329  continue;
330  }
331 
332  // Otherwise, we have to expand it to a long branch.
333  MachineInstr &OldBranch = *I;
334  DebugLoc dl = OldBranch.getDebugLoc();
335 
336  if (I->getOpcode() == PPC::BCC) {
337  // The BCC operands are:
338  // 0. PPC branch predicate
339  // 1. CR register
340  // 2. Target MBB
341  PPC::Predicate Pred = (PPC::Predicate)I->getOperand(0).getImm();
342  unsigned CRReg = I->getOperand(1).getReg();
343 
344  // Jump over the uncond branch inst (i.e. $PC+8) on opposite condition.
345  BuildMI(MBB, I, dl, TII->get(PPC::BCC))
346  .addImm(PPC::InvertPredicate(Pred)).addReg(CRReg).addImm(2);
347  } else if (I->getOpcode() == PPC::BC) {
348  unsigned CRBit = I->getOperand(0).getReg();
349  BuildMI(MBB, I, dl, TII->get(PPC::BCn)).addReg(CRBit).addImm(2);
350  } else if (I->getOpcode() == PPC::BCn) {
351  unsigned CRBit = I->getOperand(0).getReg();
352  BuildMI(MBB, I, dl, TII->get(PPC::BC)).addReg(CRBit).addImm(2);
353  } else if (I->getOpcode() == PPC::BDNZ) {
354  BuildMI(MBB, I, dl, TII->get(PPC::BDZ)).addImm(2);
355  } else if (I->getOpcode() == PPC::BDNZ8) {
356  BuildMI(MBB, I, dl, TII->get(PPC::BDZ8)).addImm(2);
357  } else if (I->getOpcode() == PPC::BDZ) {
358  BuildMI(MBB, I, dl, TII->get(PPC::BDNZ)).addImm(2);
359  } else if (I->getOpcode() == PPC::BDZ8) {
360  BuildMI(MBB, I, dl, TII->get(PPC::BDNZ8)).addImm(2);
361  } else {
362  llvm_unreachable("Unhandled branch type!");
363  }
364 
365  // Uncond branch to the real destination.
366  I = BuildMI(MBB, I, dl, TII->get(PPC::B)).addMBB(Dest);
367 
368  // Remove the old branch from the function.
369  OldBranch.eraseFromParent();
370 
371  // Remember that this instruction is 8-bytes, increase the size of the
372  // block by 4, remember to iterate.
373  BlockSizes[MBB.getNumber()].first += 4;
374  MBBStartOffset += 8;
375  ++NumExpanded;
376  MadeChange = true;
377  }
378  }
379 
380  if (MadeChange) {
381  // If we're going to iterate again, make sure we've updated our
382  // padding-based contributions to the block sizes.
383  modifyAdjustment(Fn);
384  }
385 
386  EverMadeChange |= MadeChange;
387  }
388 
389  BlockSizes.clear();
390  return true;
391 }
constexpr char Align[]
Key for Kernel::Arg::Metadata::mAlign.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
This class represents lattice values for constants.
Definition: AllocatorList.h:23
void RenumberBlocks(MachineBasicBlock *MBBFrom=nullptr)
RenumberBlocks - This discards all of the MachineBasicBlock numbers and recomputes them...
unsigned getNumBlockIDs() const
getNumBlockIDs - Return the number of MBB ID&#39;s allocated.
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition: MachineInstr.h:385
constexpr bool isInt< 16 >(int64_t x)
Definition: MathExtras.h:305
STATISTIC(NumFunctions, "Total number of functions")
A debug info location.
Definition: DebugLoc.h:33
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
const HexagonInstrInfo * TII
void initializePPCBSelPass(PassRegistry &)
void eraseFromParent()
Unlink &#39;this&#39; from the containing basic block and delete it.
unsigned getAlignment() const
getAlignment - Return the alignment (log2, not bytes) of the function.
CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based loops.
bool isELFv2ABI() const
int getNumber() const
MachineBasicBlocks are uniquely numbered at the function level, unless they&#39;re not in a MachineFuncti...
virtual const TargetInstrInfo * getInstrInfo() const
static unsigned GetInitialOffset(MachineFunction &Fn)
We need to be careful about the offset of the first block in the function because it might not have t...
FunctionPass * createPPCBranchSelectionPass()
MachineInstrBuilder BuildMI(MachineFunction &MF, const DebugLoc &DL, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
MachineBasicBlock * getBlockNumbered(unsigned N) const
getBlockNumbered - MachineBasicBlocks are automatically numbered when they are inserted into the mach...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
unsigned getAlignment() const
Return alignment of the basic block.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
static const int BlockSize
Definition: TarWriter.cpp:33
Iterator for intrusive lists based on ilist_node.
Predicate
Predicate - These are "(BI << 5) | BO" for various predicates.
Definition: PPCPredicates.h:26
bool use_empty(unsigned RegNo) const
use_empty - Return true if there are no instructions using the specified register.
MachineFunctionProperties & set(Property P)
Representation of each machine instruction.
Definition: MachineInstr.h:64
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
Predicate InvertPredicate(Predicate Opcode)
Invert the specified predicate. != -> ==, < -> >=.
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
#define I(x, y, z)
Definition: MD5.cpp:58
unsigned getInstSizeInBytes(const MachineInstr &MI) const override
GetInstSize - Return the number of bytes of code the specified instruction may be.
INITIALIZE_PASS(PPCBSel, "ppc-branch-select", "PowerPC Branch Selector", false, false) FunctionPass *llvm
createPPCBranchSelectionPass - returns an instance of the Branch Selection Pass
uint64_t OffsetToAlignment(uint64_t Value, uint64_t Align)
Returns the offset to the next integer (mod 2**64) that is greater than or equal to Value and is a mu...
Definition: MathExtras.h:731
IRTranslator LLVM IR MI
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
Properties which a MachineFunction may have at a given point in time.