//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "arm-cp-islands"
#include "ARM.h"
#include "ARMMachineFunctionInfo.h"
-#include "Thumb2InstrInfo.h"
#include "MCTargetDesc/ARMAddressingModes.h"
+#include "Thumb2InstrInfo.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetMachine.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/Format.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/Support/CommandLine.h"
+#include "llvm/Target/TargetMachine.h"
#include <algorithm>
using namespace llvm;
+#define DEBUG_TYPE "arm-cp-islands"
+
STATISTIC(NumCPEs, "Number of constpool entries");
STATISTIC(NumSplit, "Number of uncond branches inserted");
STATISTIC(NumCBrFixed, "Number of cond branches fixed");
AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true),
cl::desc("Adjust basic block layout to better use TB[BH]"));
-// FIXME: This option should be removed once it has received sufficient testing.
-static cl::opt<bool>
-AlignConstantIslands("arm-align-constant-islands", cl::Hidden, cl::init(true),
- cl::desc("Align constant islands in code"));
-
/// UnknownPadding - Return the worst case padding that could result from
/// unknown offset bits. This does not include alignment padding caused by
/// known offset bits.
return 0;
}
-/// WorstCaseAlign - Assuming only the low KnownBits bits in Offset are exact,
-/// add padding such that:
-///
-/// 1. The result is aligned to 1 << LogAlign.
-///
-/// 2. No other value of the unknown bits would require more padding.
-///
-/// This may add more padding than is required to satisfy just one of the
-/// constraints. It is necessary to compute alignment this way to guarantee
-/// that we don't underestimate the padding before an aligned block. If the
-/// real padding before a block is larger than we think, constant pool entries
-/// may go out of range.
-static inline unsigned WorstCaseAlign(unsigned Offset, unsigned LogAlign,
- unsigned KnownBits) {
- // Add the worst possible padding that the unknown bits could cause.
- Offset += UnknownPadding(LogAlign, KnownBits);
-
- // Then align the result.
- return RoundUpToAlignment(Offset, 1u << LogAlign);
-}
-
namespace {
/// ARMConstantIslands - Due to limited PC-relative displacements, ARM
/// requires constant pool entries to be scattered among the instructions
/// Offset - Distance from the beginning of the function to the beginning
/// of this basic block.
///
- /// The offset is always aligned as required by the basic block.
+ /// Offsets are computed assuming worst case padding before an aligned
+ /// block. This means that subtracting basic block offsets always gives a
+ /// conservative estimate of the real distance which may be smaller.
+ ///
+ /// Because worst case padding is used, the computed offset of an aligned
+ /// block may not actually be aligned.
unsigned Offset;
/// Size - Size of the basic block in bytes. If the block contains
/// This number should be used to predict worst case padding when
/// splitting the block.
unsigned internalKnownBits() const {
- return Unalign ? Unalign : KnownBits;
+ unsigned Bits = Unalign ? Unalign : KnownBits;
+ // If the block size isn't a multiple of the known bits, assume the
+ // worst case padding.
+ if (Size & ((1u << Bits) - 1))
+ Bits = countTrailingZeros(Size);
+ return Bits;
}
/// Compute the offset immediately following this block. If LogAlign is
if (!LA)
return PO;
// Add alignment padding from the terminator.
- return WorstCaseAlign(PO, LA, internalKnownBits());
+ return PO + UnknownPadding(LA, internalKnownBits());
}
/// Compute the number of known low bits of postOffset. If this block
static char ID;
ARMConstantIslands() : MachineFunctionPass(ID) {}
- virtual bool runOnMachineFunction(MachineFunction &MF);
+ bool runOnMachineFunction(MachineFunction &MF) override;
- virtual const char *getPassName() const {
+ const char *getPassName() const override {
return "ARM constant island placement and branch shortening pass";
}
private:
void doInitialPlacement(std::vector<MachineInstr*> &CPEMIs);
+ bool BBHasFallthrough(MachineBasicBlock *MBB);
CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI);
unsigned getCPELogAlign(const MachineInstr *CPEMI);
void scanFunctionJumpTables();
for (MachineFunction::iterator MBBI = MF->begin(), E = MF->end();
MBBI != E; ++MBBI) {
MachineBasicBlock *MBB = MBBI;
- unsigned Align = MBB->getAlignment();
unsigned MBBId = MBB->getNumber();
- assert(BBInfo[MBBId].Offset % (1u << Align) == 0);
assert(!MBBId || BBInfo[MBBId - 1].postOffset() <= BBInfo[MBBId].Offset);
}
DEBUG(dbgs() << "Verifying " << CPUsers.size() << " CP users.\n");
<< MCP->getConstants().size() << " CP entries, aligned to "
<< MCP->getConstantPoolAlignment() << " bytes *****\n");
- TII = (const ARMBaseInstrInfo*)MF->getTarget().getInstrInfo();
+ STI = &static_cast<const ARMSubtarget &>(MF->getSubtarget());
+ TII = STI->getInstrInfo();
AFI = MF->getInfo<ARMFunctionInfo>();
- STI = &MF->getTarget().getSubtarget<ARMSubtarget>();
isThumb = AFI->isThumbFunction();
isThumb1 = AFI->isThumb1OnlyFunction();
MF->RenumberBlocks();
}
- // Thumb1 functions containing constant pools get 4-byte alignment.
- // This is so we can keep exact track of where the alignment padding goes.
-
- // ARM and Thumb2 functions need to be 4-byte aligned.
- if (!isThumb1)
- MF->EnsureAlignment(2); // 2 = log2(4)
-
// Perform the initial placement of the constant pool entries. To start with,
// we put them all at the end of the function.
std::vector<MachineInstr*> CPEMIs;
CPEMIs.clear();
DEBUG(dumpBBs());
+ // Functions with jump tables need an alignment of 4 because they use the ADR
+ // instruction, which aligns the PC to 4 bytes before adding an offset.
+ if (!T2JumpTables.empty())
+ MF->ensureAlignment(2);
/// Remove dead constant pool entries.
MadeChange |= removeUnusedCPEntries();
unsigned MaxAlign = Log2_32(MCP->getConstantPoolAlignment());
// Mark the basic block as required by the const-pool.
- // If AlignConstantIslands isn't set, use 4-byte alignment for everything.
- BB->setAlignment(AlignConstantIslands ? MaxAlign : 2);
+ BB->setAlignment(MaxAlign);
// The function needs to be as aligned as the basic blocks. The linker may
// move functions around based on their alignment.
- MF->EnsureAlignment(BB->getAlignment());
+ MF->ensureAlignment(BB->getAlignment());
// Order the entries in BB by descending alignment. That ensures correct
// alignment of all entries as long as BB is sufficiently aligned. Keep
// identity mapping of CPI's to CPE's.
const std::vector<MachineConstantPoolEntry> &CPs = MCP->getConstants();
- const TargetData &TD = *MF->getTarget().getTargetData();
+ const DataLayout &TD = *MF->getTarget().getDataLayout();
for (unsigned i = 0, e = CPs.size(); i != e; ++i) {
unsigned Size = TD.getTypeAllocSize(CPs[i].getType());
assert(Size >= 4 && "Too small constant pool entry");
InsPoint[a] = CPEMI;
// Add a new CPEntry, but no corresponding CPUser yet.
- std::vector<CPEntry> CPEs;
- CPEs.push_back(CPEntry(CPEMI, i));
- CPEntries.push_back(CPEs);
+ CPEntries.emplace_back(1, CPEntry(CPEMI, i));
++NumCPEs;
DEBUG(dbgs() << "Moved CPI#" << i << " to end of function, size = "
<< Size << ", align = " << Align <<'\n');
/// BBHasFallthrough - Return true if the specified basic block can fallthrough
/// into the block immediately after it.
-static bool BBHasFallthrough(MachineBasicBlock *MBB) {
+bool ARMConstantIslands::BBHasFallthrough(MachineBasicBlock *MBB) {
// Get the next machine basic block in the function.
MachineFunction::iterator MBBI = MBB;
// Can't fall off end of function.
- if (llvm::next(MBBI) == MBB->getParent()->end())
+ if (std::next(MBBI) == MBB->getParent()->end())
return false;
- MachineBasicBlock *NextBB = llvm::next(MBBI);
- for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
- E = MBB->succ_end(); I != E; ++I)
- if (*I == NextBB)
- return true;
+ MachineBasicBlock *NextBB = std::next(MBBI);
+ if (std::find(MBB->succ_begin(), MBB->succ_end(), NextBB) == MBB->succ_end())
+ return false;
- return false;
+ // Try to analyze the end of the block. A potential fallthrough may already
+ // have an unconditional branch for whatever reason.
+ MachineBasicBlock *TBB, *FBB;
+ SmallVector<MachineOperand, 4> Cond;
+ bool TooDifficult = TII->AnalyzeBranch(*MBB, TBB, FBB, Cond);
+ return TooDifficult || FBB == nullptr;
}
/// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI,
if (CPEs[i].CPEMI == CPEMI)
return &CPEs[i];
}
- return NULL;
+ return nullptr;
}
/// getCPELogAlign - Returns the required alignment of the constant pool entry
unsigned ARMConstantIslands::getCPELogAlign(const MachineInstr *CPEMI) {
assert(CPEMI && CPEMI->getOpcode() == ARM::CONSTPOOL_ENTRY);
- // Everything is 4-byte aligned unless AlignConstantIslands is set.
- if (!AlignConstantIslands)
- return 2;
-
unsigned CPI = CPEMI->getOperand(1).getIndex();
assert(CPI < MCP->getConstants().size() && "Invalid constant pool index.");
unsigned Align = MCP->getConstants()[CPI].getAlignment();
Scale = 4;
break;
+ case ARM::LDRBi12:
case ARM::LDRi12:
case ARM::LDRcp:
case ARM::t2LDRpci:
// tBR_JTr contains a .align 2 directive.
if (!MBB->empty() && MBB->back().getOpcode() == ARM::tBR_JTr) {
BBI.PostAlign = 2;
- MBB->getParent()->EnsureAlignment(2);
+ MBB->getParent()->ensureAlignment(2);
}
}
CompareMBBNumbers);
MachineBasicBlock* WaterBB = *IP;
if (WaterBB == OrigBB)
- WaterList.insert(llvm::next(IP), NewBB);
+ WaterList.insert(std::next(IP), NewBB);
else
WaterList.insert(IP, OrigBB);
NewWaterList.insert(OrigBB);
MachineInstr *CPEMI, unsigned MaxDisp,
bool NegOk, bool DoDump) {
unsigned CPEOffset = getOffsetOf(CPEMI);
- assert(CPEOffset % 4 == 0 && "Misaligned CPE");
if (DoDump) {
DEBUG({
assert(CPE && "Unexpected!");
if (--CPE->RefCount == 0) {
removeDeadCPEMI(CPEMI);
- CPE->CPEMI = NULL;
+ CPE->CPEMI = nullptr;
--NumCPEs;
return true;
}
if (CPEs[i].CPEMI == CPEMI)
continue;
// Removing CPEs can leave empty entries, skip
- if (CPEs[i].CPEMI == NULL)
+ if (CPEs[i].CPEMI == nullptr)
continue;
if (isCPEntryInRange(UserMI, UserOffset, CPEs[i].CPEMI, U.getMaxDisp(),
U.NegOk)) {
return false;
unsigned BestGrowth = ~0u;
- for (water_iterator IP = prior(WaterList.end()), B = WaterList.begin();;
+ for (water_iterator IP = std::prev(WaterList.end()), B = WaterList.begin();;
--IP) {
MachineBasicBlock* WaterBB = *IP;
// Check if water is in range and is either at a lower address than the
unsigned Growth;
if (isWaterInRange(UserOffset, WaterBB, U, Growth) &&
(WaterBB->getNumber() < U.HighWaterMark->getNumber() ||
- NewWaterList.count(WaterBB)) && Growth < BestGrowth) {
+ NewWaterList.count(WaterBB) || WaterBB == U.MI->getParent()) &&
+ Growth < BestGrowth) {
// This is the least amount of required padding seen so far.
BestGrowth = Growth;
WaterIter = IP;
if (BBHasFallthrough(UserMBB)) {
// Size of branch to insert.
unsigned Delta = isThumb1 ? 2 : 4;
- // End of UserBlock after adding a branch.
- unsigned UserBlockEnd = UserBBI.postOffset() + Delta;
// Compute the offset where the CPE will begin.
- unsigned CPEOffset = WorstCaseAlign(UserBlockEnd, CPELogAlign,
- UserBBI.postKnownBits());
+ unsigned CPEOffset = UserBBI.postOffset(CPELogAlign) + Delta;
if (isOffsetInRange(UserOffset, CPEOffset, U)) {
DEBUG(dbgs() << "Split at end of BB#" << UserMBB->getNumber()
<< format(", expected CPE offset %#x\n", CPEOffset));
- NewMBB = llvm::next(MachineFunction::iterator(UserMBB));
+ NewMBB = std::next(MachineFunction::iterator(UserMBB));
// Add an unconditional branch from UserMBB to fallthrough block. Record
// it for branch lengthening; this new branch will not get out of range,
// but if the preceding conditional branch is out of range, the targets
unsigned MaxDisp = getUnconditionalBrDisp(UncondBr);
ImmBranches.push_back(ImmBranch(&UserMBB->back(),
MaxDisp, false, UncondBr));
- BBInfo[UserMBB->getNumber()].Size += Delta;
+ computeBlockSize(UserMBB);
adjustBBOffsetsAfter(UserMBB);
return;
}
// up the insertion point.
// Try to split the block so it's fully aligned. Compute the latest split
- // point where we can add a 4-byte branch instruction, and then
- // WorstCaseAlign to LogAlign.
+ // point where we can add a 4-byte branch instruction, and then align to
+ // LogAlign which is the largest possible alignment in the function.
unsigned LogAlign = MF->getAlignment();
assert(LogAlign >= CPELogAlign && "Over-aligned constant pool entry");
unsigned KnownBits = UserBBI.internalKnownBits();
unsigned UPad = UnknownPadding(LogAlign, KnownBits);
- unsigned BaseInsertOffset = UserOffset + U.getMaxDisp();
+ unsigned BaseInsertOffset = UserOffset + U.getMaxDisp() - UPad;
DEBUG(dbgs() << format("Split in middle of big block before %#x",
BaseInsertOffset));
- // Account for alignment and unknown padding.
- BaseInsertOffset &= ~((1u << LogAlign) - 1);
- BaseInsertOffset -= UPad;
-
// The 4 in the following is for the unconditional branch we'll be inserting
// (allows for long branch on Thumb1). Alignment of the island is handled
// inside isOffsetInRange.
// pool entries following this block; only the last one is in the water list.
// Back past any possible branches (allow for a conditional and a maximally
// long unconditional).
- if (BaseInsertOffset >= BBInfo[UserMBB->getNumber()+1].Offset)
- BaseInsertOffset = BBInfo[UserMBB->getNumber()+1].Offset -
- (isThumb1 ? 6 : 8);
- unsigned EndInsertOffset =
- WorstCaseAlign(BaseInsertOffset + 4, LogAlign, KnownBits) +
+ if (BaseInsertOffset + 8 >= UserBBI.postOffset()) {
+ // Ensure BaseInsertOffset is larger than the offset of the instruction
+ // following UserMI so that the loop which searches for the split point
+ // iterates at least once.
+ BaseInsertOffset =
+ std::max(UserBBI.postOffset() - UPad - 8,
+ UserOffset + TII->GetInstSizeInBytes(UserMI) + 1);
+ DEBUG(dbgs() << format("Move inside block: %#x\n", BaseInsertOffset));
+ }
+ unsigned EndInsertOffset = BaseInsertOffset + 4 + UPad +
CPEMI->getOperand(2).getImm();
MachineBasicBlock::iterator MI = UserMI;
++MI;
unsigned CPUIndex = CPUserIndex+1;
unsigned NumCPUsers = CPUsers.size();
- MachineInstr *LastIT = 0;
+ MachineInstr *LastIT = nullptr;
for (unsigned Offset = UserOffset+TII->GetInstSizeInBytes(UserMI);
Offset < BaseInsertOffset;
- Offset += TII->GetInstSizeInBytes(MI),
- MI = llvm::next(MI)) {
+ Offset += TII->GetInstSizeInBytes(MI), MI = std::next(MI)) {
+ assert(MI != UserMBB->end() && "Fell off end of block");
if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == MI) {
CPUser &U = CPUsers[CPUIndex];
if (!isOffsetInRange(Offset, EndInsertOffset, U)) {
// reused within the block, but it doesn't matter much. Also assume CPEs
// are added in order with alignment padding. We may eventually be able
// to pack the aligned CPEs better.
- EndInsertOffset = RoundUpToAlignment(EndInsertOffset,
- 1u << getCPELogAlign(U.CPEMI)) +
- U.CPEMI->getOperand(2).getImm();
+ EndInsertOffset += U.CPEMI->getOperand(2).getImm();
CPUIndex++;
}
if (CC != ARMCC::AL)
MI = LastIT;
}
+
+ // We really must not split an IT block.
+ DEBUG(unsigned PredReg;
+ assert(!isThumb || getITInstrPredicate(MI, PredReg) == ARMCC::AL));
+
NewMBB = splitBlockBeforeInstr(MI);
}
// If the original WaterList entry was "new water" on this iteration,
// propagate that to the new island. This is just keeping NewWaterList
// updated to match the WaterList, which will be updated below.
- if (NewWaterList.count(WaterBB)) {
- NewWaterList.erase(WaterBB);
+ if (NewWaterList.erase(WaterBB))
NewWaterList.insert(NewIsland);
- }
+
// The new CPE goes before the following block (NewMBB).
- NewMBB = llvm::next(MachineFunction::iterator(WaterBB));
+ NewMBB = std::next(MachineFunction::iterator(WaterBB));
} else {
// No water found.
// next iteration for constant pools, but in this context, we don't want
// it. Check for this so it will be removed from the WaterList.
// Also remove any entry from NewWaterList.
- MachineBasicBlock *WaterBB = prior(MachineFunction::iterator(NewMBB));
+ MachineBasicBlock *WaterBB = std::prev(MachineFunction::iterator(NewMBB));
IP = std::find(WaterList.begin(), WaterList.end(), WaterBB);
if (IP != WaterList.end())
NewWaterList.erase(WaterBB);
// Increase the size of the island block to account for the new entry.
BBInfo[NewIsland->getNumber()].Size += Size;
- adjustBBOffsetsAfter(llvm::prior(MachineFunction::iterator(NewIsland)));
+ adjustBBOffsetsAfter(std::prev(MachineFunction::iterator(NewIsland)));
// Finally, change the CPI in the instruction operand to be ID.
for (unsigned i = 0, e = UserMI->getNumOperands(); i != e; ++i)
if (CPEBB->empty()) {
BBInfo[CPEBB->getNumber()].Size = 0;
- // This block no longer needs to be aligned. <rdar://problem/10534709>.
+ // This block no longer needs to be aligned.
CPEBB->setAlignment(0);
} else
// Entries are sorted by descending alignment, so realign from the front.
for (unsigned j = 0, ee = CPEs.size(); j != ee; ++j) {
if (CPEs[j].RefCount == 0 && CPEs[j].CPEMI) {
removeDeadCPEMI(CPEs[j].CPEMI);
- CPEs[j].CPEMI = NULL;
+ CPEs[j].CPEMI = nullptr;
MadeChange = true;
}
}
++NumCBrFixed;
if (BMI != MI) {
- if (llvm::next(MachineBasicBlock::iterator(MI)) == prior(MBB->end()) &&
+ if (std::next(MachineBasicBlock::iterator(MI)) == std::prev(MBB->end()) &&
BMI->getOpcode() == Br.UncondBr) {
// Last MI in the BB is an unconditional branch. Can we simply invert the
// condition and swap destinations:
MBB->back().eraseFromParent();
// BBInfo[SplitBB].Offset is wrong temporarily, fixed below
}
- MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
+ MachineBasicBlock *NextBB = std::next(MachineFunction::iterator(MBB));
DEBUG(dbgs() << " Insert B to BB#" << DestBB->getNumber()
<< " also invert condition and change dest. to BB#"
bool ARMConstantIslands::optimizeThumb2Branches() {
bool MadeChange = false;
- for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) {
- ImmBranch &Br = ImmBranches[i];
+ // The order in which branches appear in ImmBranches is approximately their
+ // order within the function body. By visiting later branches first, we reduce
+ // the distance between earlier forward branches and their targets, making it
+ // more likely that the cbn?z optimization, which can only apply to forward
+ // branches, will succeed.
+ for (unsigned i = ImmBranches.size(); i != 0; --i) {
+ ImmBranch &Br = ImmBranches[i-1];
unsigned Opcode = Br.MI->getOpcode();
unsigned NewOpc = 0;
unsigned Scale = 1;
// FIXME: After the tables are shrunk, can we get rid some of the
// constantpool tables?
MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
- if (MJTI == 0) return false;
+ if (!MJTI) return false;
const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
DEBUG(dbgs() << "Shrink JT: " << *MI << " addr: " << *AddrMI
<< " lea: " << *LeaMI);
unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT;
- MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
+ MachineBasicBlock::iterator MI_JT = MI;
+ MachineInstr *NewJTMI =
+ BuildMI(*MBB, MI_JT, MI->getDebugLoc(), TII->get(Opc))
.addReg(IdxReg, getKillRegState(IdxRegKill))
.addJumpTableIndex(JTI, JTOP.getTargetFlags())
.addImm(MI->getOperand(JTOpIdx+1).getImm());
bool MadeChange = false;
MachineJumpTableInfo *MJTI = MF->getJumpTableInfo();
- if (MJTI == 0) return false;
+ if (!MJTI) return false;
const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
// try to move it; otherwise, create a new block following the jump
// table that branches back to the actual target. This is a very simple
// heuristic. FIXME: We can definitely improve it.
- MachineBasicBlock *TBB = 0, *FBB = 0;
+ MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
SmallVector<MachineOperand, 4> Cond;
SmallVector<MachineOperand, 4> CondPrior;
MachineFunction::iterator BBi = BB;
- MachineFunction::iterator OldPrior = prior(BBi);
+ MachineFunction::iterator OldPrior = std::prev(BBi);
// If the block terminator isn't analyzable, don't try to move the block
bool B = TII->AnalyzeBranch(*BB, TBB, FBB, Cond);
// Update numbering to account for the block being moved.
MF->RenumberBlocks();
++NumJTMoved;
- return NULL;
+ return nullptr;
}
// Create a new MBB for the code after the jump BB.