1
//===-- ARMConstantIslandPass.cpp - ARM constant islands --------*- C++ -*-===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This file contains a pass that splits the constant pool up into 'islands'
11
// which are scattered through-out the function. This is required due to the
12
// limited pc-relative displacements that ARM has.
14
//===----------------------------------------------------------------------===//
16
#define DEBUG_TYPE "arm-cp-islands"
18
#include "ARMAddressingModes.h"
19
#include "ARMMachineFunctionInfo.h"
20
#include "ARMInstrInfo.h"
21
#include "llvm/CodeGen/MachineConstantPool.h"
22
#include "llvm/CodeGen/MachineFunctionPass.h"
23
#include "llvm/CodeGen/MachineInstrBuilder.h"
24
#include "llvm/CodeGen/MachineJumpTableInfo.h"
25
#include "llvm/Target/TargetData.h"
26
#include "llvm/Target/TargetMachine.h"
27
#include "llvm/Support/Debug.h"
28
#include "llvm/Support/ErrorHandling.h"
29
#include "llvm/Support/raw_ostream.h"
30
#include "llvm/ADT/SmallSet.h"
31
#include "llvm/ADT/SmallVector.h"
32
#include "llvm/ADT/STLExtras.h"
33
#include "llvm/ADT/Statistic.h"
34
#include "llvm/Support/CommandLine.h"
38
STATISTIC(NumCPEs, "Number of constpool entries");
39
STATISTIC(NumSplit, "Number of uncond branches inserted");
40
STATISTIC(NumCBrFixed, "Number of cond branches fixed");
41
STATISTIC(NumUBrFixed, "Number of uncond branches fixed");
42
STATISTIC(NumTBs, "Number of table branches generated");
43
STATISTIC(NumT2CPShrunk, "Number of Thumb2 constantpool instructions shrunk");
44
STATISTIC(NumT2BrShrunk, "Number of Thumb2 immediate branches shrunk");
45
STATISTIC(NumCBZ, "Number of CBZ / CBNZ formed");
46
STATISTIC(NumJTMoved, "Number of jump table destination blocks moved");
47
STATISTIC(NumJTInserted, "Number of jump table intermediate blocks inserted");
51
AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true),
52
cl::desc("Adjust basic block layout to better use TB[BH]"));
55
/// ARMConstantIslands - Due to limited PC-relative displacements, ARM
56
/// requires constant pool entries to be scattered among the instructions
57
/// inside a function. To do this, it completely ignores the normal LLVM
58
/// constant pool; instead, it places constants wherever it feels like with
59
/// special instructions.
61
/// The terminology used in this pass includes:
62
/// Islands - Clumps of constants placed in the function.
63
/// Water - Potential places where an island could be formed.
64
/// CPE - A constant pool entry that has been placed somewhere, which
65
/// tracks a list of users.
66
class ARMConstantIslands : public MachineFunctionPass {
67
/// BBSizes - The size of each MachineBasicBlock in bytes of code, indexed
68
/// by MBB Number. The two-byte pads required for Thumb alignment are
69
/// counted as part of the following block (i.e., the offset and size for
70
/// a padded block will both be ==2 mod 4).
71
std::vector<unsigned> BBSizes;
73
/// BBOffsets - the offset of each MBB in bytes, starting from 0.
74
/// The two-byte pads required for Thumb alignment are counted as part of
75
/// the following block.
76
std::vector<unsigned> BBOffsets;
78
/// WaterList - A sorted list of basic blocks where islands could be placed
79
/// (i.e. blocks that don't fall through to the following block, due
80
/// to a return, unreachable, or unconditional branch).
81
std::vector<MachineBasicBlock*> WaterList;
83
/// NewWaterList - The subset of WaterList that was created since the
84
/// previous iteration by inserting unconditional branches.
85
SmallSet<MachineBasicBlock*, 4> NewWaterList;
87
typedef std::vector<MachineBasicBlock*>::iterator water_iterator;
89
/// CPUser - One user of a constant pool, keeping the machine instruction
90
/// pointer, the constant pool being referenced, and the max displacement
91
/// allowed from the instruction to the CP. The HighWaterMark records the
92
/// highest basic block where a new CPEntry can be placed. To ensure this
93
/// pass terminates, the CP entries are initially placed at the end of the
94
/// function and then move monotonically to lower addresses. The
95
/// exception to this rule is when the current CP entry for a particular
96
/// CPUser is out of range, but there is another CP entry for the same
97
/// constant value in range. We want to use the existing in-range CP
98
/// entry, but if it later moves out of range, the search for new water
99
/// should resume where it left off. The HighWaterMark is used to record
104
MachineBasicBlock *HighWaterMark;
108
CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp,
109
bool neg, bool soimm)
110
: MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) {
111
HighWaterMark = CPEMI->getParent();
115
/// CPUsers - Keep track of all of the machine instructions that use various
116
/// constant pools and their max displacement.
117
std::vector<CPUser> CPUsers;
119
/// CPEntry - One per constant pool entry, keeping the machine instruction
120
/// pointer, the constpool index, and the number of CPUser's which
121
/// reference this entry.
126
CPEntry(MachineInstr *cpemi, unsigned cpi, unsigned rc = 0)
127
: CPEMI(cpemi), CPI(cpi), RefCount(rc) {}
130
/// CPEntries - Keep track of all of the constant pool entry machine
131
/// instructions. For each original constpool index (i.e. those that
132
/// existed upon entry to this pass), it keeps a vector of entries.
133
/// Original elements are cloned as we go along; the clones are
134
/// put in the vector of the original element, but have distinct CPIs.
135
std::vector<std::vector<CPEntry> > CPEntries;
137
/// ImmBranch - One per immediate branch, keeping the machine instruction
138
/// pointer, conditional or unconditional, the max displacement,
139
/// and (if isCond is true) the corresponding unconditional branch
143
unsigned MaxDisp : 31;
146
ImmBranch(MachineInstr *mi, unsigned maxdisp, bool cond, int ubr)
147
: MI(mi), MaxDisp(maxdisp), isCond(cond), UncondBr(ubr) {}
150
/// ImmBranches - Keep track of all the immediate branch instructions.
152
std::vector<ImmBranch> ImmBranches;
154
/// PushPopMIs - Keep track of all the Thumb push / pop instructions.
156
SmallVector<MachineInstr*, 4> PushPopMIs;
158
/// T2JumpTables - Keep track of all the Thumb2 jumptable instructions.
159
SmallVector<MachineInstr*, 4> T2JumpTables;
161
/// HasFarJump - True if any far jump instruction has been emitted during
162
/// the branch fix up pass.
165
/// HasInlineAsm - True if the function contains inline assembly.
168
const TargetInstrInfo *TII;
169
const ARMSubtarget *STI;
170
ARMFunctionInfo *AFI;
176
ARMConstantIslands() : MachineFunctionPass(&ID) {}
178
virtual bool runOnMachineFunction(MachineFunction &MF);
180
virtual const char *getPassName() const {
181
return "ARM constant island placement and branch shortening pass";
185
void DoInitialPlacement(MachineFunction &MF,
186
std::vector<MachineInstr*> &CPEMIs);
187
CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI);
188
void JumpTableFunctionScan(MachineFunction &MF);
189
void InitialFunctionScan(MachineFunction &MF,
190
const std::vector<MachineInstr*> &CPEMIs);
191
MachineBasicBlock *SplitBlockBeforeInstr(MachineInstr *MI);
192
void UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB);
193
void AdjustBBOffsetsAfter(MachineBasicBlock *BB, int delta);
194
bool DecrementOldEntry(unsigned CPI, MachineInstr* CPEMI);
195
int LookForExistingCPEntry(CPUser& U, unsigned UserOffset);
196
bool LookForWater(CPUser&U, unsigned UserOffset, water_iterator &WaterIter);
197
void CreateNewWater(unsigned CPUserIndex, unsigned UserOffset,
198
MachineBasicBlock *&NewMBB);
199
bool HandleConstantPoolUser(MachineFunction &MF, unsigned CPUserIndex);
200
void RemoveDeadCPEMI(MachineInstr *CPEMI);
201
bool RemoveUnusedCPEntries();
202
bool CPEIsInRange(MachineInstr *MI, unsigned UserOffset,
203
MachineInstr *CPEMI, unsigned Disp, bool NegOk,
204
bool DoDump = false);
205
bool WaterIsInRange(unsigned UserOffset, MachineBasicBlock *Water,
207
bool OffsetIsInRange(unsigned UserOffset, unsigned TrialOffset,
208
unsigned Disp, bool NegativeOK, bool IsSoImm = false);
209
bool BBIsInRange(MachineInstr *MI, MachineBasicBlock *BB, unsigned Disp);
210
bool FixUpImmediateBr(MachineFunction &MF, ImmBranch &Br);
211
bool FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br);
212
bool FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br);
213
bool UndoLRSpillRestore();
214
bool OptimizeThumb2Instructions(MachineFunction &MF);
215
bool OptimizeThumb2Branches(MachineFunction &MF);
216
bool ReorderThumb2JumpTables(MachineFunction &MF);
217
bool OptimizeThumb2JumpTables(MachineFunction &MF);
218
MachineBasicBlock *AdjustJTTargetBlockForward(MachineBasicBlock *BB,
219
MachineBasicBlock *JTBB);
221
unsigned GetOffsetOf(MachineInstr *MI) const;
223
void verify(MachineFunction &MF);
225
char ARMConstantIslands::ID = 0;
228
/// verify - check BBOffsets, BBSizes, alignment of islands
229
void ARMConstantIslands::verify(MachineFunction &MF) {
230
assert(BBOffsets.size() == BBSizes.size());
231
for (unsigned i = 1, e = BBOffsets.size(); i != e; ++i)
232
assert(BBOffsets[i-1]+BBSizes[i-1] == BBOffsets[i]);
236
for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
238
MachineBasicBlock *MBB = MBBI;
240
MBB->begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) {
241
unsigned MBBId = MBB->getNumber();
242
assert(HasInlineAsm ||
243
(BBOffsets[MBBId]%4 == 0 && BBSizes[MBBId]%4 == 0) ||
244
(BBOffsets[MBBId]%4 != 0 && BBSizes[MBBId]%4 != 0));
247
for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) {
248
CPUser &U = CPUsers[i];
249
unsigned UserOffset = GetOffsetOf(U.MI) + (isThumb ? 4 : 8);
250
unsigned CPEOffset = GetOffsetOf(U.CPEMI);
251
unsigned Disp = UserOffset < CPEOffset ? CPEOffset - UserOffset :
252
UserOffset - CPEOffset;
253
assert(Disp <= U.MaxDisp || "Constant pool entry out of range!");
258
/// print block size and offset information - debugging
259
void ARMConstantIslands::dumpBBs() {
260
for (unsigned J = 0, E = BBOffsets.size(); J !=E; ++J) {
261
DEBUG(errs() << "block " << J << " offset " << BBOffsets[J]
262
<< " size " << BBSizes[J] << "\n");
266
/// createARMConstantIslandPass - returns an instance of the constpool
268
FunctionPass *llvm::createARMConstantIslandPass() {
269
return new ARMConstantIslands();
272
bool ARMConstantIslands::runOnMachineFunction(MachineFunction &MF) {
273
MachineConstantPool &MCP = *MF.getConstantPool();
275
TII = MF.getTarget().getInstrInfo();
276
AFI = MF.getInfo<ARMFunctionInfo>();
277
STI = &MF.getTarget().getSubtarget<ARMSubtarget>();
279
isThumb = AFI->isThumbFunction();
280
isThumb1 = AFI->isThumb1OnlyFunction();
281
isThumb2 = AFI->isThumb2Function();
284
HasInlineAsm = false;
286
// Renumber all of the machine basic blocks in the function, guaranteeing that
287
// the numbers agree with the position of the block in the function.
290
// Try to reorder and otherwise adjust the block layout to make good use
291
// of the TB[BH] instructions.
292
bool MadeChange = false;
293
if (isThumb2 && AdjustJumpTableBlocks) {
294
JumpTableFunctionScan(MF);
295
MadeChange |= ReorderThumb2JumpTables(MF);
296
// Data is out of date, so clear it. It'll be re-computed later.
297
T2JumpTables.clear();
298
// Blocks may have shifted around. Keep the numbering up to date.
302
// Thumb1 functions containing constant pools get 4-byte alignment.
303
// This is so we can keep exact track of where the alignment padding goes.
305
// ARM and Thumb2 functions need to be 4-byte aligned.
307
MF.EnsureAlignment(2); // 2 = log2(4)
309
// Perform the initial placement of the constant pool entries. To start with,
310
// we put them all at the end of the function.
311
std::vector<MachineInstr*> CPEMIs;
312
if (!MCP.isEmpty()) {
313
DoInitialPlacement(MF, CPEMIs);
315
MF.EnsureAlignment(2); // 2 = log2(4)
318
/// The next UID to take is the first unused one.
319
AFI->initConstPoolEntryUId(CPEMIs.size());
321
// Do the initial scan of the function, building up information about the
322
// sizes of each block, the location of all the water, and finding all of the
323
// constant pool users.
324
InitialFunctionScan(MF, CPEMIs);
327
/// Remove dead constant pool entries.
328
RemoveUnusedCPEntries();
330
// Iteratively place constant pool entries and fix up branches until there
332
unsigned NoCPIters = 0, NoBRIters = 0;
334
bool CPChange = false;
335
for (unsigned i = 0, e = CPUsers.size(); i != e; ++i)
336
CPChange |= HandleConstantPoolUser(MF, i);
337
if (CPChange && ++NoCPIters > 30)
338
llvm_unreachable("Constant Island pass failed to converge!");
341
// Clear NewWaterList now. If we split a block for branches, it should
342
// appear as "new water" for the next iteration of constant pool placement.
343
NewWaterList.clear();
345
bool BRChange = false;
346
for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i)
347
BRChange |= FixUpImmediateBr(MF, ImmBranches[i]);
348
if (BRChange && ++NoBRIters > 30)
349
llvm_unreachable("Branch Fix Up pass failed to converge!");
352
if (!CPChange && !BRChange)
357
// Shrink 32-bit Thumb2 branch, load, and store instructions.
359
MadeChange |= OptimizeThumb2Instructions(MF);
361
// After a while, this might be made debug-only, but it is not expensive.
364
// If LR has been forced spilled and no far jumps (i.e. BL) has been issued.
365
// Undo the spill / restore of LR if possible.
366
if (isThumb && !HasFarJump && AFI->isLRSpilledForFarJump())
367
MadeChange |= UndoLRSpillRestore();
376
T2JumpTables.clear();
381
/// DoInitialPlacement - Perform the initial placement of the constant pool
382
/// entries. To start with, we put them all at the end of the function.
383
void ARMConstantIslands::DoInitialPlacement(MachineFunction &MF,
384
std::vector<MachineInstr*> &CPEMIs) {
385
// Create the basic block to hold the CPE's.
386
MachineBasicBlock *BB = MF.CreateMachineBasicBlock();
389
// Add all of the constants from the constant pool to the end block, use an
390
// identity mapping of CPI's to CPE's.
391
const std::vector<MachineConstantPoolEntry> &CPs =
392
MF.getConstantPool()->getConstants();
394
const TargetData &TD = *MF.getTarget().getTargetData();
395
for (unsigned i = 0, e = CPs.size(); i != e; ++i) {
396
unsigned Size = TD.getTypeAllocSize(CPs[i].getType());
397
// Verify that all constant pool entries are a multiple of 4 bytes. If not,
398
// we would have to pad them out or something so that instructions stay
400
assert((Size & 3) == 0 && "CP Entry not multiple of 4 bytes!");
401
MachineInstr *CPEMI =
402
BuildMI(BB, DebugLoc::getUnknownLoc(), TII->get(ARM::CONSTPOOL_ENTRY))
403
.addImm(i).addConstantPoolIndex(i).addImm(Size);
404
CPEMIs.push_back(CPEMI);
406
// Add a new CPEntry, but no corresponding CPUser yet.
407
std::vector<CPEntry> CPEs;
408
CPEs.push_back(CPEntry(CPEMI, i));
409
CPEntries.push_back(CPEs);
411
DEBUG(errs() << "Moved CPI#" << i << " to end of function as #" << i
416
/// BBHasFallthrough - Return true if the specified basic block can fallthrough
417
/// into the block immediately after it.
418
static bool BBHasFallthrough(MachineBasicBlock *MBB) {
419
// Get the next machine basic block in the function.
420
MachineFunction::iterator MBBI = MBB;
421
if (llvm::next(MBBI) == MBB->getParent()->end()) // Can't fall off end of function.
424
MachineBasicBlock *NextBB = llvm::next(MBBI);
425
for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
426
E = MBB->succ_end(); I != E; ++I)
433
/// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI,
434
/// look up the corresponding CPEntry.
435
ARMConstantIslands::CPEntry
436
*ARMConstantIslands::findConstPoolEntry(unsigned CPI,
437
const MachineInstr *CPEMI) {
438
std::vector<CPEntry> &CPEs = CPEntries[CPI];
439
// Number of entries per constpool index should be small, just do a
441
for (unsigned i = 0, e = CPEs.size(); i != e; ++i) {
442
if (CPEs[i].CPEMI == CPEMI)
448
/// JumpTableFunctionScan - Do a scan of the function, building up
449
/// information about the sizes of each block and the locations of all
451
void ARMConstantIslands::JumpTableFunctionScan(MachineFunction &MF) {
452
for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
454
MachineBasicBlock &MBB = *MBBI;
456
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
458
if (I->getDesc().isBranch() && I->getOpcode() == ARM::t2BR_JT)
459
T2JumpTables.push_back(I);
463
/// InitialFunctionScan - Do the initial scan of the function, building up
464
/// information about the sizes of each block, the location of all the water,
465
/// and finding all of the constant pool users.
466
void ARMConstantIslands::InitialFunctionScan(MachineFunction &MF,
467
const std::vector<MachineInstr*> &CPEMIs) {
468
// First thing, see if the function has any inline assembly in it. If so,
469
// we have to be conservative about alignment assumptions, as we don't
470
// know for sure the size of any instructions in the inline assembly.
471
for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
473
MachineBasicBlock &MBB = *MBBI;
474
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
476
if (I->getOpcode() == ARM::INLINEASM)
480
// Now go back through the instructions and build up our data structures
482
for (MachineFunction::iterator MBBI = MF.begin(), E = MF.end();
484
MachineBasicBlock &MBB = *MBBI;
486
// If this block doesn't fall through into the next MBB, then this is
487
// 'water' that a constant pool island could be placed.
488
if (!BBHasFallthrough(&MBB))
489
WaterList.push_back(&MBB);
491
unsigned MBBSize = 0;
492
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
494
// Add instruction size to MBBSize.
495
MBBSize += TII->GetInstSizeInBytes(I);
497
int Opc = I->getOpcode();
498
if (I->getDesc().isBranch()) {
505
continue; // Ignore other JT branches
507
// A Thumb1 table jump may involve padding; for the offsets to
508
// be right, functions containing these must be 4-byte aligned.
509
MF.EnsureAlignment(2U);
510
if ((Offset+MBBSize)%4 != 0 || HasInlineAsm)
511
// FIXME: Add a pseudo ALIGN instruction instead.
512
MBBSize += 2; // padding
513
continue; // Does not get an entry in ImmBranches
515
T2JumpTables.push_back(I);
516
continue; // Does not get an entry in ImmBranches
547
// Record this immediate branch.
548
unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale;
549
ImmBranches.push_back(ImmBranch(I, MaxOffs, isCond, UOpc));
552
if (Opc == ARM::tPUSH || Opc == ARM::tPOP_RET)
553
PushPopMIs.push_back(I);
555
if (Opc == ARM::CONSTPOOL_ENTRY)
558
// Scan the instructions for constant pool operands.
559
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op)
560
if (I->getOperand(op).isCPI()) {
561
// We found one. The addressing mode tells us the max displacement
562
// from the PC that this instruction permits.
564
// Basic size info comes from the TSFlags field.
568
bool IsSoImm = false;
572
llvm_unreachable("Unknown addressing mode for CP reference!");
575
// Taking the address of a CP entry.
577
// This takes a SoImm, which is 8 bit immediate rotated. We'll
578
// pretend the maximum offset is 255 * 4. Since each instruction
579
// 4 byte wide, this is always correct. We'll check for other
580
// displacements that fits in a SoImm as well.
586
case ARM::t2LEApcrel:
598
Bits = 12; // +-offset_12
605
Scale = 4; // +(offset_8*4)
611
Scale = 4; // +-(offset_8*4)
616
// Remember that this is a user of a CP entry.
617
unsigned CPI = I->getOperand(op).getIndex();
618
MachineInstr *CPEMI = CPEMIs[CPI];
619
unsigned MaxOffs = ((1 << Bits)-1) * Scale;
620
CPUsers.push_back(CPUser(I, CPEMI, MaxOffs, NegOk, IsSoImm));
622
// Increment corresponding CPEntry reference count.
623
CPEntry *CPE = findConstPoolEntry(CPI, CPEMI);
624
assert(CPE && "Cannot find a corresponding CPEntry!");
627
// Instructions can only use one CP entry, don't bother scanning the
628
// rest of the operands.
633
// In thumb mode, if this block is a constpool island, we may need padding
634
// so it's aligned on 4 byte boundary.
637
MBB.begin()->getOpcode() == ARM::CONSTPOOL_ENTRY &&
638
((Offset%4) != 0 || HasInlineAsm))
641
BBSizes.push_back(MBBSize);
642
BBOffsets.push_back(Offset);
647
/// GetOffsetOf - Return the current offset of the specified machine instruction
648
/// from the start of the function. This offset changes as stuff is moved
649
/// around inside the function.
650
unsigned ARMConstantIslands::GetOffsetOf(MachineInstr *MI) const {
651
MachineBasicBlock *MBB = MI->getParent();
653
// The offset is composed of two things: the sum of the sizes of all MBB's
654
// before this instruction's block, and the offset from the start of the block
656
unsigned Offset = BBOffsets[MBB->getNumber()];
658
// If we're looking for a CONSTPOOL_ENTRY in Thumb, see if this block has
659
// alignment padding, and compensate if so.
661
MI->getOpcode() == ARM::CONSTPOOL_ENTRY &&
662
(Offset%4 != 0 || HasInlineAsm))
665
// Sum instructions before MI in MBB.
666
for (MachineBasicBlock::iterator I = MBB->begin(); ; ++I) {
667
assert(I != MBB->end() && "Didn't find MI in its own basic block?");
668
if (&*I == MI) return Offset;
669
Offset += TII->GetInstSizeInBytes(I);
673
/// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB
675
static bool CompareMBBNumbers(const MachineBasicBlock *LHS,
676
const MachineBasicBlock *RHS) {
677
return LHS->getNumber() < RHS->getNumber();
680
/// UpdateForInsertedWaterBlock - When a block is newly inserted into the
681
/// machine function, it upsets all of the block numbers. Renumber the blocks
682
/// and update the arrays that parallel this numbering.
683
void ARMConstantIslands::UpdateForInsertedWaterBlock(MachineBasicBlock *NewBB) {
684
// Renumber the MBB's to keep them consequtive.
685
NewBB->getParent()->RenumberBlocks(NewBB);
687
// Insert a size into BBSizes to align it properly with the (newly
688
// renumbered) block numbers.
689
BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0);
691
// Likewise for BBOffsets.
692
BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0);
694
// Next, update WaterList. Specifically, we need to add NewMBB as having
695
// available water after it.
697
std::lower_bound(WaterList.begin(), WaterList.end(), NewBB,
699
WaterList.insert(IP, NewBB);
703
/// Split the basic block containing MI into two blocks, which are joined by
704
/// an unconditional branch. Update data structures and renumber blocks to
705
/// account for this change and returns the newly created block.
706
MachineBasicBlock *ARMConstantIslands::SplitBlockBeforeInstr(MachineInstr *MI) {
707
MachineBasicBlock *OrigBB = MI->getParent();
708
MachineFunction &MF = *OrigBB->getParent();
710
// Create a new MBB for the code after the OrigBB.
711
MachineBasicBlock *NewBB =
712
MF.CreateMachineBasicBlock(OrigBB->getBasicBlock());
713
MachineFunction::iterator MBBI = OrigBB; ++MBBI;
714
MF.insert(MBBI, NewBB);
716
// Splice the instructions starting with MI over to NewBB.
717
NewBB->splice(NewBB->end(), OrigBB, MI, OrigBB->end());
719
// Add an unconditional branch from OrigBB to NewBB.
720
// Note the new unconditional branch is not being recorded.
721
// There doesn't seem to be meaningful DebugInfo available; this doesn't
722
// correspond to anything in the source.
723
unsigned Opc = isThumb ? (isThumb2 ? ARM::t2B : ARM::tB) : ARM::B;
724
BuildMI(OrigBB, DebugLoc::getUnknownLoc(), TII->get(Opc)).addMBB(NewBB);
727
// Update the CFG. All succs of OrigBB are now succs of NewBB.
728
while (!OrigBB->succ_empty()) {
729
MachineBasicBlock *Succ = *OrigBB->succ_begin();
730
OrigBB->removeSuccessor(Succ);
731
NewBB->addSuccessor(Succ);
733
// This pass should be run after register allocation, so there should be no
734
// PHI nodes to update.
735
assert((Succ->empty() || !Succ->begin()->isPHI())
736
&& "PHI nodes should be eliminated by now!");
739
// OrigBB branches to NewBB.
740
OrigBB->addSuccessor(NewBB);
742
// Update internal data structures to account for the newly inserted MBB.
743
// This is almost the same as UpdateForInsertedWaterBlock, except that
744
// the Water goes after OrigBB, not NewBB.
745
MF.RenumberBlocks(NewBB);
747
// Insert a size into BBSizes to align it properly with the (newly
748
// renumbered) block numbers.
749
BBSizes.insert(BBSizes.begin()+NewBB->getNumber(), 0);
751
// Likewise for BBOffsets.
752
BBOffsets.insert(BBOffsets.begin()+NewBB->getNumber(), 0);
754
// Next, update WaterList. Specifically, we need to add OrigMBB as having
755
// available water after it (but not if it's already there, which happens
756
// when splitting before a conditional branch that is followed by an
757
// unconditional branch - in that case we want to insert NewBB).
759
std::lower_bound(WaterList.begin(), WaterList.end(), OrigBB,
761
MachineBasicBlock* WaterBB = *IP;
762
if (WaterBB == OrigBB)
763
WaterList.insert(llvm::next(IP), NewBB);
765
WaterList.insert(IP, OrigBB);
766
NewWaterList.insert(OrigBB);
768
// Figure out how large the first NewMBB is. (It cannot
769
// contain a constpool_entry or tablejump.)
770
unsigned NewBBSize = 0;
771
for (MachineBasicBlock::iterator I = NewBB->begin(), E = NewBB->end();
773
NewBBSize += TII->GetInstSizeInBytes(I);
775
unsigned OrigBBI = OrigBB->getNumber();
776
unsigned NewBBI = NewBB->getNumber();
777
// Set the size of NewBB in BBSizes.
778
BBSizes[NewBBI] = NewBBSize;
780
// We removed instructions from UserMBB, subtract that off from its size.
781
// Add 2 or 4 to the block to count the unconditional branch we added to it.
782
int delta = isThumb1 ? 2 : 4;
783
BBSizes[OrigBBI] -= NewBBSize - delta;
785
// ...and adjust BBOffsets for NewBB accordingly.
786
BBOffsets[NewBBI] = BBOffsets[OrigBBI] + BBSizes[OrigBBI];
788
// All BBOffsets following these blocks must be modified.
789
AdjustBBOffsetsAfter(NewBB, delta);
794
/// OffsetIsInRange - Checks whether UserOffset (the location of a constant pool
795
/// reference) is within MaxDisp of TrialOffset (a proposed location of a
796
/// constant pool entry).
797
bool ARMConstantIslands::OffsetIsInRange(unsigned UserOffset,
798
unsigned TrialOffset, unsigned MaxDisp,
799
bool NegativeOK, bool IsSoImm) {
800
// On Thumb offsets==2 mod 4 are rounded down by the hardware for
801
// purposes of the displacement computation; compensate for that here.
802
// Effectively, the valid range of displacements is 2 bytes smaller for such
804
unsigned TotalAdj = 0;
805
if (isThumb && UserOffset%4 !=0) {
809
// CPEs will be rounded up to a multiple of 4.
810
if (isThumb && TrialOffset%4 != 0) {
815
// In Thumb2 mode, later branch adjustments can shift instructions up and
816
// cause alignment change. In the worst case scenario this can cause the
817
// user's effective address to be subtracted by 2 and the CPE's address to
819
if (isThumb2 && TotalAdj != 4)
820
MaxDisp -= (4 - TotalAdj);
822
if (UserOffset <= TrialOffset) {
823
// User before the Trial.
824
if (TrialOffset - UserOffset <= MaxDisp)
826
// FIXME: Make use full range of soimm values.
827
} else if (NegativeOK) {
828
if (UserOffset - TrialOffset <= MaxDisp)
830
// FIXME: Make use full range of soimm values.
835
/// WaterIsInRange - Returns true if a CPE placed after the specified
836
/// Water (a basic block) will be in range for the specific MI.
838
bool ARMConstantIslands::WaterIsInRange(unsigned UserOffset,
839
MachineBasicBlock* Water, CPUser &U) {
840
unsigned MaxDisp = U.MaxDisp;
841
unsigned CPEOffset = BBOffsets[Water->getNumber()] +
842
BBSizes[Water->getNumber()];
844
// If the CPE is to be inserted before the instruction, that will raise
845
// the offset of the instruction.
846
if (CPEOffset < UserOffset)
847
UserOffset += U.CPEMI->getOperand(2).getImm();
849
return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, U.NegOk, U.IsSoImm);
852
/// CPEIsInRange - Returns true if the distance between specific MI and
853
/// specific ConstPool entry instruction can fit in MI's displacement field.
854
bool ARMConstantIslands::CPEIsInRange(MachineInstr *MI, unsigned UserOffset,
855
MachineInstr *CPEMI, unsigned MaxDisp,
856
bool NegOk, bool DoDump) {
857
unsigned CPEOffset = GetOffsetOf(CPEMI);
858
assert((CPEOffset%4 == 0 || HasInlineAsm) && "Misaligned CPE");
861
DEBUG(errs() << "User of CPE#" << CPEMI->getOperand(0).getImm()
862
<< " max delta=" << MaxDisp
863
<< " insn address=" << UserOffset
864
<< " CPE address=" << CPEOffset
865
<< " offset=" << int(CPEOffset-UserOffset) << "\t" << *MI);
868
return OffsetIsInRange(UserOffset, CPEOffset, MaxDisp, NegOk);
872
/// BBIsJumpedOver - Return true of the specified basic block's only predecessor
873
/// unconditionally branches to its only successor.
874
static bool BBIsJumpedOver(MachineBasicBlock *MBB) {
875
if (MBB->pred_size() != 1 || MBB->succ_size() != 1)
878
MachineBasicBlock *Succ = *MBB->succ_begin();
879
MachineBasicBlock *Pred = *MBB->pred_begin();
880
MachineInstr *PredMI = &Pred->back();
881
if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB
882
|| PredMI->getOpcode() == ARM::t2B)
883
return PredMI->getOperand(0).getMBB() == Succ;
888
void ARMConstantIslands::AdjustBBOffsetsAfter(MachineBasicBlock *BB,
890
MachineFunction::iterator MBBI = BB; MBBI = llvm::next(MBBI);
891
for(unsigned i = BB->getNumber()+1, e = BB->getParent()->getNumBlockIDs();
893
BBOffsets[i] += delta;
894
// If some existing blocks have padding, adjust the padding as needed, a
895
// bit tricky. delta can be negative so don't use % on that.
898
MachineBasicBlock *MBB = MBBI;
899
if (!MBB->empty() && !HasInlineAsm) {
900
// Constant pool entries require padding.
901
if (MBB->begin()->getOpcode() == ARM::CONSTPOOL_ENTRY) {
902
unsigned OldOffset = BBOffsets[i] - delta;
903
if ((OldOffset%4) == 0 && (BBOffsets[i]%4) != 0) {
907
} else if ((OldOffset%4) != 0 && (BBOffsets[i]%4) == 0) {
908
// remove existing padding
913
// Thumb1 jump tables require padding. They should be at the end;
914
// following unconditional branches are removed by AnalyzeBranch.
915
MachineInstr *ThumbJTMI = prior(MBB->end());
916
if (ThumbJTMI->getOpcode() == ARM::tBR_JTr) {
917
unsigned NewMIOffset = GetOffsetOf(ThumbJTMI);
918
unsigned OldMIOffset = NewMIOffset - delta;
919
if ((OldMIOffset%4) == 0 && (NewMIOffset%4) != 0) {
920
// remove existing padding
923
} else if ((OldMIOffset%4) != 0 && (NewMIOffset%4) == 0) {
932
MBBI = llvm::next(MBBI);
936
/// DecrementOldEntry - find the constant pool entry with index CPI
937
/// and instruction CPEMI, and decrement its refcount. If the refcount
938
/// becomes 0 remove the entry and instruction. Returns true if we removed
939
/// the entry, false if we didn't.
941
bool ARMConstantIslands::DecrementOldEntry(unsigned CPI, MachineInstr *CPEMI) {
942
// Find the old entry. Eliminate it if it is no longer used.
943
CPEntry *CPE = findConstPoolEntry(CPI, CPEMI);
944
assert(CPE && "Unexpected!");
945
if (--CPE->RefCount == 0) {
946
RemoveDeadCPEMI(CPEMI);
954
/// LookForCPEntryInRange - see if the currently referenced CPE is in range;
955
/// if not, see if an in-range clone of the CPE is in range, and if so,
956
/// change the data structures so the user references the clone. Returns:
957
/// 0 = no existing entry found
958
/// 1 = entry found, and there were no code insertions or deletions
959
/// 2 = entry found, and there were code insertions or deletions
960
int ARMConstantIslands::LookForExistingCPEntry(CPUser& U, unsigned UserOffset)
962
MachineInstr *UserMI = U.MI;
963
MachineInstr *CPEMI = U.CPEMI;
965
// Check to see if the CPE is already in-range.
966
if (CPEIsInRange(UserMI, UserOffset, CPEMI, U.MaxDisp, U.NegOk, true)) {
967
DEBUG(errs() << "In range\n");
971
// No. Look for previously created clones of the CPE that are in range.
972
unsigned CPI = CPEMI->getOperand(1).getIndex();
973
std::vector<CPEntry> &CPEs = CPEntries[CPI];
974
for (unsigned i = 0, e = CPEs.size(); i != e; ++i) {
975
// We already tried this one
976
if (CPEs[i].CPEMI == CPEMI)
978
// Removing CPEs can leave empty entries, skip
979
if (CPEs[i].CPEMI == NULL)
981
if (CPEIsInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.MaxDisp, U.NegOk)) {
982
DEBUG(errs() << "Replacing CPE#" << CPI << " with CPE#"
983
<< CPEs[i].CPI << "\n");
984
// Point the CPUser node to the replacement
985
U.CPEMI = CPEs[i].CPEMI;
986
// Change the CPI in the instruction operand to refer to the clone.
987
for (unsigned j = 0, e = UserMI->getNumOperands(); j != e; ++j)
988
if (UserMI->getOperand(j).isCPI()) {
989
UserMI->getOperand(j).setIndex(CPEs[i].CPI);
992
// Adjust the refcount of the clone...
994
// ...and the original. If we didn't remove the old entry, none of the
995
// addresses changed, so we don't need another pass.
996
return DecrementOldEntry(CPI, CPEMI) ? 2 : 1;
1002
/// getUnconditionalBrDisp - Returns the maximum displacement that can fit in
1003
/// the specific unconditional branch instruction.
1004
static inline unsigned getUnconditionalBrDisp(int Opc) {
1007
return ((1<<10)-1)*2;
1009
return ((1<<23)-1)*2;
1014
return ((1<<23)-1)*4;
1017
/// LookForWater - Look for an existing entry in the WaterList in which
1018
/// we can place the CPE referenced from U so it's within range of U's MI.
1019
/// Returns true if found, false if not. If it returns true, WaterIter
1020
/// is set to the WaterList entry. For Thumb, prefer water that will not
1021
/// introduce padding to water that will. To ensure that this pass
1022
/// terminates, the CPE location for a particular CPUser is only allowed to
1023
/// move to a lower address, so search backward from the end of the list and
1024
/// prefer the first water that is in range.
1025
bool ARMConstantIslands::LookForWater(CPUser &U, unsigned UserOffset,
1026
water_iterator &WaterIter) {
1027
if (WaterList.empty())
1030
bool FoundWaterThatWouldPad = false;
1031
water_iterator IPThatWouldPad;
1032
for (water_iterator IP = prior(WaterList.end()),
1033
B = WaterList.begin();; --IP) {
1034
MachineBasicBlock* WaterBB = *IP;
1035
// Check if water is in range and is either at a lower address than the
1036
// current "high water mark" or a new water block that was created since
1037
// the previous iteration by inserting an unconditional branch. In the
1038
// latter case, we want to allow resetting the high water mark back to
1039
// this new water since we haven't seen it before. Inserting branches
1040
// should be relatively uncommon and when it does happen, we want to be
1041
// sure to take advantage of it for all the CPEs near that block, so that
1042
// we don't insert more branches than necessary.
1043
if (WaterIsInRange(UserOffset, WaterBB, U) &&
1044
(WaterBB->getNumber() < U.HighWaterMark->getNumber() ||
1045
NewWaterList.count(WaterBB))) {
1046
unsigned WBBId = WaterBB->getNumber();
1048
(BBOffsets[WBBId] + BBSizes[WBBId])%4 != 0) {
1049
// This is valid Water, but would introduce padding. Remember
1050
// it in case we don't find any Water that doesn't do this.
1051
if (!FoundWaterThatWouldPad) {
1052
FoundWaterThatWouldPad = true;
1053
IPThatWouldPad = IP;
1063
if (FoundWaterThatWouldPad) {
1064
WaterIter = IPThatWouldPad;
1070
/// CreateNewWater - No existing WaterList entry will work for
1071
/// CPUsers[CPUserIndex], so create a place to put the CPE. The end of the
1072
/// block is used if in range, and the conditional branch munged so control
1073
/// flow is correct. Otherwise the block is split to create a hole with an
1074
/// unconditional branch around it. In either case NewMBB is set to a
1075
/// block following which the new island can be inserted (the WaterList
1076
/// is not adjusted).
1077
void ARMConstantIslands::CreateNewWater(unsigned CPUserIndex,
1078
unsigned UserOffset,
1079
MachineBasicBlock *&NewMBB) {
1080
CPUser &U = CPUsers[CPUserIndex];
1081
MachineInstr *UserMI = U.MI;
1082
MachineInstr *CPEMI = U.CPEMI;
1083
MachineBasicBlock *UserMBB = UserMI->getParent();
1084
unsigned OffsetOfNextBlock = BBOffsets[UserMBB->getNumber()] +
1085
BBSizes[UserMBB->getNumber()];
1086
assert(OffsetOfNextBlock== BBOffsets[UserMBB->getNumber()+1]);
1088
// If the block does not end in an unconditional branch already, and if the
1089
// end of the block is within range, make new water there. (The addition
1090
// below is for the unconditional branch we will be adding: 4 bytes on ARM +
1091
// Thumb2, 2 on Thumb1. Possible Thumb1 alignment padding is allowed for
1092
// inside OffsetIsInRange.
1093
if (BBHasFallthrough(UserMBB) &&
1094
OffsetIsInRange(UserOffset, OffsetOfNextBlock + (isThumb1 ? 2: 4),
1095
U.MaxDisp, U.NegOk, U.IsSoImm)) {
1096
DEBUG(errs() << "Split at end of block\n");
1097
if (&UserMBB->back() == UserMI)
1098
assert(BBHasFallthrough(UserMBB) && "Expected a fallthrough BB!");
1099
NewMBB = llvm::next(MachineFunction::iterator(UserMBB));
1100
// Add an unconditional branch from UserMBB to fallthrough block.
1101
// Record it for branch lengthening; this new branch will not get out of
1102
// range, but if the preceding conditional branch is out of range, the
1103
// targets will be exchanged, and the altered branch may be out of
1104
// range, so the machinery has to know about it.
1105
int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B;
1106
BuildMI(UserMBB, DebugLoc::getUnknownLoc(),
1107
TII->get(UncondBr)).addMBB(NewMBB);
1108
unsigned MaxDisp = getUnconditionalBrDisp(UncondBr);
1109
ImmBranches.push_back(ImmBranch(&UserMBB->back(),
1110
MaxDisp, false, UncondBr));
1111
int delta = isThumb1 ? 2 : 4;
1112
BBSizes[UserMBB->getNumber()] += delta;
1113
AdjustBBOffsetsAfter(UserMBB, delta);
1115
// What a big block. Find a place within the block to split it.
1116
// This is a little tricky on Thumb1 since instructions are 2 bytes
1117
// and constant pool entries are 4 bytes: if instruction I references
1118
// island CPE, and instruction I+1 references CPE', it will
1119
// not work well to put CPE as far forward as possible, since then
1120
// CPE' cannot immediately follow it (that location is 2 bytes
1121
// farther away from I+1 than CPE was from I) and we'd need to create
1122
// a new island. So, we make a first guess, then walk through the
1123
// instructions between the one currently being looked at and the
1124
// possible insertion point, and make sure any other instructions
1125
// that reference CPEs will be able to use the same island area;
1126
// if not, we back up the insertion point.
1128
// The 4 in the following is for the unconditional branch we'll be
1129
// inserting (allows for long branch on Thumb1). Alignment of the
1130
// island is handled inside OffsetIsInRange.
1131
unsigned BaseInsertOffset = UserOffset + U.MaxDisp -4;
1132
// This could point off the end of the block if we've already got
1133
// constant pool entries following this block; only the last one is
1134
// in the water list. Back past any possible branches (allow for a
1135
// conditional and a maximally long unconditional).
1136
if (BaseInsertOffset >= BBOffsets[UserMBB->getNumber()+1])
1137
BaseInsertOffset = BBOffsets[UserMBB->getNumber()+1] -
1139
unsigned EndInsertOffset = BaseInsertOffset +
1140
CPEMI->getOperand(2).getImm();
1141
MachineBasicBlock::iterator MI = UserMI;
1143
unsigned CPUIndex = CPUserIndex+1;
1144
for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI);
1145
Offset < BaseInsertOffset;
1146
Offset += TII->GetInstSizeInBytes(MI),
1147
MI = llvm::next(MI)) {
1148
if (CPUIndex < CPUsers.size() && CPUsers[CPUIndex].MI == MI) {
1149
CPUser &U = CPUsers[CPUIndex];
1150
if (!OffsetIsInRange(Offset, EndInsertOffset,
1151
U.MaxDisp, U.NegOk, U.IsSoImm)) {
1152
BaseInsertOffset -= (isThumb1 ? 2 : 4);
1153
EndInsertOffset -= (isThumb1 ? 2 : 4);
1155
// This is overly conservative, as we don't account for CPEMIs
1156
// being reused within the block, but it doesn't matter much.
1157
EndInsertOffset += CPUsers[CPUIndex].CPEMI->getOperand(2).getImm();
1161
DEBUG(errs() << "Split in middle of big block\n");
1162
NewMBB = SplitBlockBeforeInstr(prior(MI));
1166
/// HandleConstantPoolUser - Analyze the specified user, checking to see if it
1167
/// is out-of-range. If so, pick up the constant pool value and move it some
1168
/// place in-range. Return true if we changed any addresses (thus must run
1169
/// another pass of branch lengthening), false otherwise.
1170
bool ARMConstantIslands::HandleConstantPoolUser(MachineFunction &MF,
1171
unsigned CPUserIndex) {
1172
CPUser &U = CPUsers[CPUserIndex];
1173
MachineInstr *UserMI = U.MI;
1174
MachineInstr *CPEMI = U.CPEMI;
1175
unsigned CPI = CPEMI->getOperand(1).getIndex();
1176
unsigned Size = CPEMI->getOperand(2).getImm();
1177
// Compute this only once, it's expensive. The 4 or 8 is the value the
1178
// hardware keeps in the PC.
1179
unsigned UserOffset = GetOffsetOf(UserMI) + (isThumb ? 4 : 8);
1181
// See if the current entry is within range, or there is a clone of it
1183
int result = LookForExistingCPEntry(U, UserOffset);
1184
if (result==1) return false;
1185
else if (result==2) return true;
1187
// No existing clone of this CPE is within range.
1188
// We will be generating a new clone. Get a UID for it.
1189
unsigned ID = AFI->createConstPoolEntryUId();
1191
// Look for water where we can place this CPE.
1192
MachineBasicBlock *NewIsland = MF.CreateMachineBasicBlock();
1193
MachineBasicBlock *NewMBB;
1195
if (LookForWater(U, UserOffset, IP)) {
1196
DEBUG(errs() << "found water in range\n");
1197
MachineBasicBlock *WaterBB = *IP;
1199
// If the original WaterList entry was "new water" on this iteration,
1200
// propagate that to the new island. This is just keeping NewWaterList
1201
// updated to match the WaterList, which will be updated below.
1202
if (NewWaterList.count(WaterBB)) {
1203
NewWaterList.erase(WaterBB);
1204
NewWaterList.insert(NewIsland);
1206
// The new CPE goes before the following block (NewMBB).
1207
NewMBB = llvm::next(MachineFunction::iterator(WaterBB));
1211
DEBUG(errs() << "No water found\n");
1212
CreateNewWater(CPUserIndex, UserOffset, NewMBB);
1214
// SplitBlockBeforeInstr adds to WaterList, which is important when it is
1215
// called while handling branches so that the water will be seen on the
1216
// next iteration for constant pools, but in this context, we don't want
1217
// it. Check for this so it will be removed from the WaterList.
1218
// Also remove any entry from NewWaterList.
1219
MachineBasicBlock *WaterBB = prior(MachineFunction::iterator(NewMBB));
1220
IP = std::find(WaterList.begin(), WaterList.end(), WaterBB);
1221
if (IP != WaterList.end())
1222
NewWaterList.erase(WaterBB);
1224
// We are adding new water. Update NewWaterList.
1225
NewWaterList.insert(NewIsland);
1228
// Remove the original WaterList entry; we want subsequent insertions in
1229
// this vicinity to go after the one we're about to insert. This
1230
// considerably reduces the number of times we have to move the same CPE
1231
// more than once and is also important to ensure the algorithm terminates.
1232
if (IP != WaterList.end())
1233
WaterList.erase(IP);
1235
// Okay, we know we can put an island before NewMBB now, do it!
1236
MF.insert(NewMBB, NewIsland);
1238
// Update internal data structures to account for the newly inserted MBB.
1239
UpdateForInsertedWaterBlock(NewIsland);
1241
// Decrement the old entry, and remove it if refcount becomes 0.
1242
DecrementOldEntry(CPI, CPEMI);
1244
// Now that we have an island to add the CPE to, clone the original CPE and
1245
// add it to the island.
1246
U.HighWaterMark = NewIsland;
1247
U.CPEMI = BuildMI(NewIsland, DebugLoc::getUnknownLoc(),
1248
TII->get(ARM::CONSTPOOL_ENTRY))
1249
.addImm(ID).addConstantPoolIndex(CPI).addImm(Size);
1250
CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1));
1253
BBOffsets[NewIsland->getNumber()] = BBOffsets[NewMBB->getNumber()];
1254
// Compensate for .align 2 in thumb mode.
1255
if (isThumb && (BBOffsets[NewIsland->getNumber()]%4 != 0 || HasInlineAsm))
1257
// Increase the size of the island block to account for the new entry.
1258
BBSizes[NewIsland->getNumber()] += Size;
1259
AdjustBBOffsetsAfter(NewIsland, Size);
1261
// Finally, change the CPI in the instruction operand to be ID.
1262
for (unsigned i = 0, e = UserMI->getNumOperands(); i != e; ++i)
1263
if (UserMI->getOperand(i).isCPI()) {
1264
UserMI->getOperand(i).setIndex(ID);
1268
DEBUG(errs() << " Moved CPE to #" << ID << " CPI=" << CPI
1269
<< '\t' << *UserMI);
1274
/// RemoveDeadCPEMI - Remove a dead constant pool entry instruction. Update
1275
/// sizes and offsets of impacted basic blocks.
1276
void ARMConstantIslands::RemoveDeadCPEMI(MachineInstr *CPEMI) {
1277
MachineBasicBlock *CPEBB = CPEMI->getParent();
1278
unsigned Size = CPEMI->getOperand(2).getImm();
1279
CPEMI->eraseFromParent();
1280
BBSizes[CPEBB->getNumber()] -= Size;
1281
// All succeeding offsets have the current size value added in, fix this.
1282
if (CPEBB->empty()) {
1283
// In thumb1 mode, the size of island may be padded by two to compensate for
1284
// the alignment requirement. Then it will now be 2 when the block is
1285
// empty, so fix this.
1286
// All succeeding offsets have the current size value added in, fix this.
1287
if (BBSizes[CPEBB->getNumber()] != 0) {
1288
Size += BBSizes[CPEBB->getNumber()];
1289
BBSizes[CPEBB->getNumber()] = 0;
1292
AdjustBBOffsetsAfter(CPEBB, -Size);
1293
// An island has only one predecessor BB and one successor BB. Check if
1294
// this BB's predecessor jumps directly to this BB's successor. This
1295
// shouldn't happen currently.
1296
assert(!BBIsJumpedOver(CPEBB) && "How did this happen?");
1297
// FIXME: remove the empty blocks after all the work is done?
1300
/// RemoveUnusedCPEntries - Remove constant pool entries whose refcounts
1302
bool ARMConstantIslands::RemoveUnusedCPEntries() {
1303
unsigned MadeChange = false;
1304
for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) {
1305
std::vector<CPEntry> &CPEs = CPEntries[i];
1306
for (unsigned j = 0, ee = CPEs.size(); j != ee; ++j) {
1307
if (CPEs[j].RefCount == 0 && CPEs[j].CPEMI) {
1308
RemoveDeadCPEMI(CPEs[j].CPEMI);
1309
CPEs[j].CPEMI = NULL;
1317
/// BBIsInRange - Returns true if the distance between specific MI and
1318
/// specific BB can fit in MI's displacement field.
1319
bool ARMConstantIslands::BBIsInRange(MachineInstr *MI,MachineBasicBlock *DestBB,
1321
unsigned PCAdj = isThumb ? 4 : 8;
1322
unsigned BrOffset = GetOffsetOf(MI) + PCAdj;
1323
unsigned DestOffset = BBOffsets[DestBB->getNumber()];
1325
DEBUG(errs() << "Branch of destination BB#" << DestBB->getNumber()
1326
<< " from BB#" << MI->getParent()->getNumber()
1327
<< " max delta=" << MaxDisp
1328
<< " from " << GetOffsetOf(MI) << " to " << DestOffset
1329
<< " offset " << int(DestOffset-BrOffset) << "\t" << *MI);
1331
if (BrOffset <= DestOffset) {
1332
// Branch before the Dest.
1333
if (DestOffset-BrOffset <= MaxDisp)
1336
if (BrOffset-DestOffset <= MaxDisp)
1342
/// FixUpImmediateBr - Fix up an immediate branch whose destination is too far
1343
/// away to fit in its displacement field.
1344
bool ARMConstantIslands::FixUpImmediateBr(MachineFunction &MF, ImmBranch &Br) {
1345
MachineInstr *MI = Br.MI;
1346
MachineBasicBlock *DestBB = MI->getOperand(0).getMBB();
1348
// Check to see if the DestBB is already in-range.
1349
if (BBIsInRange(MI, DestBB, Br.MaxDisp))
1353
return FixUpUnconditionalBr(MF, Br);
1354
return FixUpConditionalBr(MF, Br);
1357
/// FixUpUnconditionalBr - Fix up an unconditional branch whose destination is
1358
/// too far away to fit in its displacement field. If the LR register has been
1359
/// spilled in the epilogue, then we can use BL to implement a far jump.
1360
/// Otherwise, add an intermediate branch instruction to a branch.
1362
ARMConstantIslands::FixUpUnconditionalBr(MachineFunction &MF, ImmBranch &Br) {
1363
MachineInstr *MI = Br.MI;
1364
MachineBasicBlock *MBB = MI->getParent();
1366
llvm_unreachable("FixUpUnconditionalBr is Thumb1 only!");
1368
// Use BL to implement far jump.
1369
Br.MaxDisp = (1 << 21) * 2;
1370
MI->setDesc(TII->get(ARM::tBfar));
1371
BBSizes[MBB->getNumber()] += 2;
1372
AdjustBBOffsetsAfter(MBB, 2);
1376
DEBUG(errs() << " Changed B to long jump " << *MI);
1381
/// FixUpConditionalBr - Fix up a conditional branch whose destination is too
1382
/// far away to fit in its displacement field. It is converted to an inverse
1383
/// conditional branch + an unconditional branch to the destination.
1385
ARMConstantIslands::FixUpConditionalBr(MachineFunction &MF, ImmBranch &Br) {
1386
MachineInstr *MI = Br.MI;
1387
MachineBasicBlock *DestBB = MI->getOperand(0).getMBB();
1389
// Add an unconditional branch to the destination and invert the branch
1390
// condition to jump over it:
1396
ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImm();
1397
CC = ARMCC::getOppositeCondition(CC);
1398
unsigned CCReg = MI->getOperand(2).getReg();
1400
// If the branch is at the end of its MBB and that has a fall-through block,
1401
// direct the updated conditional branch to the fall-through block. Otherwise,
1402
// split the MBB before the next instruction.
1403
MachineBasicBlock *MBB = MI->getParent();
1404
MachineInstr *BMI = &MBB->back();
1405
bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB);
1409
if (llvm::next(MachineBasicBlock::iterator(MI)) == prior(MBB->end()) &&
1410
BMI->getOpcode() == Br.UncondBr) {
1411
// Last MI in the BB is an unconditional branch. Can we simply invert the
1412
// condition and swap destinations:
1418
MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB();
1419
if (BBIsInRange(MI, NewDest, Br.MaxDisp)) {
1420
DEBUG(errs() << " Invert Bcc condition and swap its destination with "
1422
BMI->getOperand(0).setMBB(DestBB);
1423
MI->getOperand(0).setMBB(NewDest);
1424
MI->getOperand(1).setImm(CC);
1431
SplitBlockBeforeInstr(MI);
1432
// No need for the branch to the next block. We're adding an unconditional
1433
// branch to the destination.
1434
int delta = TII->GetInstSizeInBytes(&MBB->back());
1435
BBSizes[MBB->getNumber()] -= delta;
1436
MachineBasicBlock* SplitBB = llvm::next(MachineFunction::iterator(MBB));
1437
AdjustBBOffsetsAfter(SplitBB, -delta);
1438
MBB->back().eraseFromParent();
1439
// BBOffsets[SplitBB] is wrong temporarily, fixed below
1441
MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
1443
DEBUG(errs() << " Insert B to BB#" << DestBB->getNumber()
1444
<< " also invert condition and change dest. to BB#"
1445
<< NextBB->getNumber() << "\n");
1447
// Insert a new conditional branch and a new unconditional branch.
1448
// Also update the ImmBranch as well as adding a new entry for the new branch.
1449
BuildMI(MBB, DebugLoc::getUnknownLoc(),
1450
TII->get(MI->getOpcode()))
1451
.addMBB(NextBB).addImm(CC).addReg(CCReg);
1452
Br.MI = &MBB->back();
1453
BBSizes[MBB->getNumber()] += TII->GetInstSizeInBytes(&MBB->back());
1454
BuildMI(MBB, DebugLoc::getUnknownLoc(), TII->get(Br.UncondBr)).addMBB(DestBB);
1455
BBSizes[MBB->getNumber()] += TII->GetInstSizeInBytes(&MBB->back());
1456
unsigned MaxDisp = getUnconditionalBrDisp(Br.UncondBr);
1457
ImmBranches.push_back(ImmBranch(&MBB->back(), MaxDisp, false, Br.UncondBr));
1459
// Remove the old conditional branch. It may or may not still be in MBB.
1460
BBSizes[MI->getParent()->getNumber()] -= TII->GetInstSizeInBytes(MI);
1461
MI->eraseFromParent();
1463
// The net size change is an addition of one unconditional branch.
1464
int delta = TII->GetInstSizeInBytes(&MBB->back());
1465
AdjustBBOffsetsAfter(MBB, delta);
1469
/// UndoLRSpillRestore - Remove Thumb push / pop instructions that only spills
1470
/// LR / restores LR to pc. FIXME: This is done here because it's only possible
1471
/// to do this if tBfar is not used.
1472
bool ARMConstantIslands::UndoLRSpillRestore() {
1473
bool MadeChange = false;
1474
for (unsigned i = 0, e = PushPopMIs.size(); i != e; ++i) {
1475
MachineInstr *MI = PushPopMIs[i];
1476
// First two operands are predicates, the third is a zero since there
1478
if (MI->getOpcode() == ARM::tPOP_RET &&
1479
MI->getOperand(3).getReg() == ARM::PC &&
1480
MI->getNumExplicitOperands() == 4) {
1481
BuildMI(MI->getParent(), MI->getDebugLoc(), TII->get(ARM::tBX_RET));
1482
MI->eraseFromParent();
1489
bool ARMConstantIslands::OptimizeThumb2Instructions(MachineFunction &MF) {
1490
bool MadeChange = false;
1492
// Shrink ADR and LDR from constantpool.
1493
for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) {
1494
CPUser &U = CPUsers[i];
1495
unsigned Opcode = U.MI->getOpcode();
1496
unsigned NewOpc = 0;
1501
case ARM::t2LEApcrel:
1502
if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
1503
NewOpc = ARM::tLEApcrel;
1509
if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
1510
NewOpc = ARM::tLDRpci;
1520
unsigned UserOffset = GetOffsetOf(U.MI) + 4;
1521
unsigned MaxOffs = ((1 << Bits) - 1) * Scale;
1522
// FIXME: Check if offset is multiple of scale if scale is not 4.
1523
if (CPEIsInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) {
1524
U.MI->setDesc(TII->get(NewOpc));
1525
MachineBasicBlock *MBB = U.MI->getParent();
1526
BBSizes[MBB->getNumber()] -= 2;
1527
AdjustBBOffsetsAfter(MBB, -2);
1533
MadeChange |= OptimizeThumb2Branches(MF);
1534
MadeChange |= OptimizeThumb2JumpTables(MF);
1538
bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) {
1539
bool MadeChange = false;
1541
for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) {
1542
ImmBranch &Br = ImmBranches[i];
1543
unsigned Opcode = Br.MI->getOpcode();
1544
unsigned NewOpc = 0;
1562
unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale;
1563
MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB();
1564
if (BBIsInRange(Br.MI, DestBB, MaxOffs)) {
1565
Br.MI->setDesc(TII->get(NewOpc));
1566
MachineBasicBlock *MBB = Br.MI->getParent();
1567
BBSizes[MBB->getNumber()] -= 2;
1568
AdjustBBOffsetsAfter(MBB, -2);
1574
Opcode = Br.MI->getOpcode();
1575
if (Opcode != ARM::tBcc)
1579
unsigned PredReg = 0;
1580
ARMCC::CondCodes Pred = llvm::getInstrPredicate(Br.MI, PredReg);
1581
if (Pred == ARMCC::EQ)
1583
else if (Pred == ARMCC::NE)
1584
NewOpc = ARM::tCBNZ;
1587
MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB();
1588
// Check if the distance is within 126. Subtract starting offset by 2
1589
// because the cmp will be eliminated.
1590
unsigned BrOffset = GetOffsetOf(Br.MI) + 4 - 2;
1591
unsigned DestOffset = BBOffsets[DestBB->getNumber()];
1592
if (BrOffset < DestOffset && (DestOffset - BrOffset) <= 126) {
1593
MachineBasicBlock::iterator CmpMI = Br.MI; --CmpMI;
1594
if (CmpMI->getOpcode() == ARM::tCMPzi8) {
1595
unsigned Reg = CmpMI->getOperand(0).getReg();
1596
Pred = llvm::getInstrPredicate(CmpMI, PredReg);
1597
if (Pred == ARMCC::AL &&
1598
CmpMI->getOperand(1).getImm() == 0 &&
1599
isARMLowRegister(Reg)) {
1600
MachineBasicBlock *MBB = Br.MI->getParent();
1601
MachineInstr *NewBR =
1602
BuildMI(*MBB, CmpMI, Br.MI->getDebugLoc(), TII->get(NewOpc))
1603
.addReg(Reg).addMBB(DestBB, Br.MI->getOperand(0).getTargetFlags());
1604
CmpMI->eraseFromParent();
1605
Br.MI->eraseFromParent();
1607
BBSizes[MBB->getNumber()] -= 2;
1608
AdjustBBOffsetsAfter(MBB, -2);
1619
/// OptimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller
1620
/// jumptables when it's possible.
1621
bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) {
1622
bool MadeChange = false;
1624
// FIXME: After the tables are shrunk, can we get rid some of the
1625
// constantpool tables?
1626
MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
1627
if (MJTI == 0) return false;
1629
const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
1630
for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
1631
MachineInstr *MI = T2JumpTables[i];
1632
const TargetInstrDesc &TID = MI->getDesc();
1633
unsigned NumOps = TID.getNumOperands();
1634
unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
1635
MachineOperand JTOP = MI->getOperand(JTOpIdx);
1636
unsigned JTI = JTOP.getIndex();
1637
assert(JTI < JT.size());
1640
bool HalfWordOk = true;
1641
unsigned JTOffset = GetOffsetOf(MI) + 4;
1642
const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
1643
for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
1644
MachineBasicBlock *MBB = JTBBs[j];
1645
unsigned DstOffset = BBOffsets[MBB->getNumber()];
1646
// Negative offset is not ok. FIXME: We should change BB layout to make
1647
// sure all the branches are forward.
1648
if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2)
1650
unsigned TBHLimit = ((1<<16)-1)*2;
1651
if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit)
1653
if (!ByteOk && !HalfWordOk)
1657
if (ByteOk || HalfWordOk) {
1658
MachineBasicBlock *MBB = MI->getParent();
1659
unsigned BaseReg = MI->getOperand(0).getReg();
1660
bool BaseRegKill = MI->getOperand(0).isKill();
1663
unsigned IdxReg = MI->getOperand(1).getReg();
1664
bool IdxRegKill = MI->getOperand(1).isKill();
1665
MachineBasicBlock::iterator PrevI = MI;
1666
if (PrevI == MBB->begin())
1669
MachineInstr *AddrMI = --PrevI;
1671
// Examine the instruction that calculate the jumptable entry address.
1672
// If it's not the one just before the t2BR_JT, we won't delete it, then
1673
// it's not worth doing the optimization.
1674
for (unsigned k = 0, eee = AddrMI->getNumOperands(); k != eee; ++k) {
1675
const MachineOperand &MO = AddrMI->getOperand(k);
1676
if (!MO.isReg() || !MO.getReg())
1678
if (MO.isDef() && MO.getReg() != BaseReg) {
1682
if (MO.isUse() && !MO.isKill() && MO.getReg() != IdxReg) {
1690
// The previous instruction should be a tLEApcrel or t2LEApcrelJT, we want
1691
// to delete it as well.
1692
MachineInstr *LeaMI = --PrevI;
1693
if ((LeaMI->getOpcode() != ARM::tLEApcrelJT &&
1694
LeaMI->getOpcode() != ARM::t2LEApcrelJT) ||
1695
LeaMI->getOperand(0).getReg() != BaseReg)
1701
unsigned Opc = ByteOk ? ARM::t2TBB : ARM::t2TBH;
1702
MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
1703
.addReg(IdxReg, getKillRegState(IdxRegKill))
1704
.addJumpTableIndex(JTI, JTOP.getTargetFlags())
1705
.addImm(MI->getOperand(JTOpIdx+1).getImm());
1706
// FIXME: Insert an "ALIGN" instruction to ensure the next instruction
1707
// is 2-byte aligned. For now, asm printer will fix it up.
1708
unsigned NewSize = TII->GetInstSizeInBytes(NewJTMI);
1709
unsigned OrigSize = TII->GetInstSizeInBytes(AddrMI);
1710
OrigSize += TII->GetInstSizeInBytes(LeaMI);
1711
OrigSize += TII->GetInstSizeInBytes(MI);
1713
AddrMI->eraseFromParent();
1714
LeaMI->eraseFromParent();
1715
MI->eraseFromParent();
1717
int delta = OrigSize - NewSize;
1718
BBSizes[MBB->getNumber()] -= delta;
1719
AdjustBBOffsetsAfter(MBB, -delta);
1729
/// ReorderThumb2JumpTables - Adjust the function's block layout to ensure that
1730
/// jump tables always branch forwards, since that's what tbb and tbh need.
1731
bool ARMConstantIslands::ReorderThumb2JumpTables(MachineFunction &MF) {
1732
bool MadeChange = false;
1734
MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
1735
if (MJTI == 0) return false;
1737
const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
1738
for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
1739
MachineInstr *MI = T2JumpTables[i];
1740
const TargetInstrDesc &TID = MI->getDesc();
1741
unsigned NumOps = TID.getNumOperands();
1742
unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
1743
MachineOperand JTOP = MI->getOperand(JTOpIdx);
1744
unsigned JTI = JTOP.getIndex();
1745
assert(JTI < JT.size());
1747
// We prefer if target blocks for the jump table come after the jump
1748
// instruction so we can use TB[BH]. Loop through the target blocks
1749
// and try to adjust them such that that's true.
1750
int JTNumber = MI->getParent()->getNumber();
1751
const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
1752
for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
1753
MachineBasicBlock *MBB = JTBBs[j];
1754
int DTNumber = MBB->getNumber();
1756
if (DTNumber < JTNumber) {
1757
// The destination precedes the switch. Try to move the block forward
1758
// so we have a positive offset.
1759
MachineBasicBlock *NewBB =
1760
AdjustJTTargetBlockForward(MBB, MI->getParent());
1762
MJTI->ReplaceMBBInJumpTable(JTI, JTBBs[j], NewBB);
1771
MachineBasicBlock *ARMConstantIslands::
1772
AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB)
1774
MachineFunction &MF = *BB->getParent();
1776
// If it's the destination block is terminated by an unconditional branch,
1777
// try to move it; otherwise, create a new block following the jump
1778
// table that branches back to the actual target. This is a very simple
1779
// heuristic. FIXME: We can definitely improve it.
1780
MachineBasicBlock *TBB = 0, *FBB = 0;
1781
SmallVector<MachineOperand, 4> Cond;
1782
SmallVector<MachineOperand, 4> CondPrior;
1783
MachineFunction::iterator BBi = BB;
1784
MachineFunction::iterator OldPrior = prior(BBi);
1786
// If the block terminator isn't analyzable, don't try to move the block
1787
bool B = TII->AnalyzeBranch(*BB, TBB, FBB, Cond);
1789
// If the block ends in an unconditional branch, move it. The prior block
1790
// has to have an analyzable terminator for us to move this one. Be paranoid
1791
// and make sure we're not trying to move the entry block of the function.
1792
if (!B && Cond.empty() && BB != MF.begin() &&
1793
!TII->AnalyzeBranch(*OldPrior, TBB, FBB, CondPrior)) {
1794
BB->moveAfter(JTBB);
1795
OldPrior->updateTerminator();
1796
BB->updateTerminator();
1797
// Update numbering to account for the block being moved.
1798
MF.RenumberBlocks();
1803
// Create a new MBB for the code after the jump BB.
1804
MachineBasicBlock *NewBB =
1805
MF.CreateMachineBasicBlock(JTBB->getBasicBlock());
1806
MachineFunction::iterator MBBI = JTBB; ++MBBI;
1807
MF.insert(MBBI, NewBB);
1809
// Add an unconditional branch from NewBB to BB.
1810
// There doesn't seem to be meaningful DebugInfo available; this doesn't
1811
// correspond directly to anything in the source.
1812
assert (isThumb2 && "Adjusting for TB[BH] but not in Thumb2?");
1813
BuildMI(NewBB, DebugLoc::getUnknownLoc(), TII->get(ARM::t2B)).addMBB(BB);
1815
// Update internal data structures to account for the newly inserted MBB.
1816
MF.RenumberBlocks(NewBB);
1819
NewBB->addSuccessor(BB);
1820
JTBB->removeSuccessor(BB);
1821
JTBB->addSuccessor(NewBB);
added
removed
libclamav/c++/llvm/lib/Target/ARM/ARMConstantIslandPass.cpp
