X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FPowerPC%2FPPCBranchSelector.cpp;h=21a0fb200f20e84043e4c88baafb0594123402d7;hb=fef904d0e824a2c587f8c1063b6c4fbf47fec898;hp=08c39311fed44e9bffc820bd4cf87aa4e4dd838e;hpb=1822ffbab7b2d9051ff57c0d8d7b9c1c2c086e2b;p=oota-llvm.git diff --git a/lib/Target/PowerPC/PPCBranchSelector.cpp b/lib/Target/PowerPC/PPCBranchSelector.cpp index 08c39311fed..21a0fb200f2 100644 --- a/lib/Target/PowerPC/PPCBranchSelector.cpp +++ b/lib/Target/PowerPC/PPCBranchSelector.cpp @@ -1,39 +1,48 @@ -//===-- PPCBranchSelector.cpp - Emit long conditional branches-----*- C++ -*-=// +//===-- PPCBranchSelector.cpp - Emit long conditional branches ------------===// // // The LLVM Compiler Infrastructure // -// This file was developed by Nate Baegeman and is distributed under the -// University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file contains a pass that scans a machine function to determine which // conditional branches need more than 16 bits of displacement to reach their // target basic block. It does this in two passes; a calculation of basic block -// positions pass, and a branch psuedo op to machine branch opcode pass. This +// positions pass, and a branch pseudo op to machine branch opcode pass. This // pass should be run last, just before the assembly printer. // //===----------------------------------------------------------------------===// +#define DEBUG_TYPE "ppc-branch-select" #include "PPC.h" #include "PPCInstrBuilder.h" #include "PPCInstrInfo.h" +#include "MCTargetDesc/PPCPredicates.h" #include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/Support/Visibility.h" -#include +#include "llvm/Target/TargetMachine.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Support/MathExtras.h" using namespace llvm; +STATISTIC(NumExpanded, "Number of branches expanded to long format"); + namespace { - struct VISIBILITY_HIDDEN PPCBSel : public MachineFunctionPass { - // OffsetMap - Mapping between BB and byte offset from start of function - std::map OffsetMap; + struct PPCBSel : public MachineFunctionPass { + static char ID; + PPCBSel() : MachineFunctionPass(ID) {} + + /// BlockSizes - The sizes of the basic blocks in the function. + std::vector BlockSizes; virtual bool runOnMachineFunction(MachineFunction &Fn); virtual const char *getPassName() const { - return "PowerPC Branch Selection"; + return "PowerPC Branch Selector"; } }; + char PPCBSel::ID = 0; } /// createPPCBranchSelectionPass - returns an instance of the Branch Selection @@ -43,112 +52,135 @@ FunctionPass *llvm::createPPCBranchSelectionPass() { return new PPCBSel(); } -/// getNumBytesForInstruction - Return the number of bytes of code the specified -/// instruction may be. This returns the maximum number of bytes. -/// -static unsigned getNumBytesForInstruction(MachineInstr *MI) { - switch (MI->getOpcode()) { - case PPC::COND_BRANCH: - // while this will be 4 most of the time, if we emit 8 it is just a - // minor pessimization that saves us from having to worry about - // keeping the offsets up to date later when we emit long branch glue. - return 8; - case PPC::IMPLICIT_DEF_GPRC: // no asm emitted - case PPC::IMPLICIT_DEF_G8RC: // no asm emitted - case PPC::IMPLICIT_DEF_F4: // no asm emitted - case PPC::IMPLICIT_DEF_F8: // no asm emitted - return 0; - case PPC::INLINEASM: // Inline Asm: Variable size. - for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) - if (MI->getOperand(i).isExternalSymbol()) { - const char *AsmStr = MI->getOperand(i).getSymbolName(); - // Count the number of newline's in the asm string. - unsigned NumInstrs = 0; - for (; *AsmStr; ++AsmStr) - NumInstrs += *AsmStr == '\n'; - return NumInstrs*4; - } - assert(0 && "INLINEASM didn't have format string??"); - default: - return 4; // PowerPC instructions are all 4 bytes - } -} - - bool PPCBSel::runOnMachineFunction(MachineFunction &Fn) { - // Running total of instructions encountered since beginning of function - unsigned ByteCount = 0; - - // For each MBB, add its offset to the offset map, and count up its - // instructions + const PPCInstrInfo *TII = + static_cast(Fn.getTarget().getInstrInfo()); + // Give the blocks of the function a dense, in-order, numbering. + Fn.RenumberBlocks(); + BlockSizes.resize(Fn.getNumBlockIDs()); + + // Measure each MBB and compute a size for the entire function. + unsigned FuncSize = 0; for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E; ++MFI) { MachineBasicBlock *MBB = MFI; - OffsetMap[MBB] = ByteCount; - + + unsigned BlockSize = 0; for (MachineBasicBlock::iterator MBBI = MBB->begin(), EE = MBB->end(); MBBI != EE; ++MBBI) - ByteCount += getNumBytesForInstruction(MBBI); + BlockSize += TII->GetInstSizeInBytes(MBBI); + + BlockSizes[MBB->getNumber()] = BlockSize; + FuncSize += BlockSize; } - // We're about to run over the MBB's again, so reset the ByteCount - ByteCount = 0; + // If the entire function is smaller than the displacement of a branch field, + // we know we don't need to shrink any branches in this function. This is a + // common case. + if (FuncSize < (1 << 15)) { + BlockSizes.clear(); + return false; + } - // For each MBB, find the conditional branch pseudo instructions, and - // calculate the difference between the target MBB and the current ICount - // to decide whether or not to emit a short or long branch. + // For each conditional branch, if the offset to its destination is larger + // than the offset field allows, transform it into a long branch sequence + // like this: + // short branch: + // bCC MBB + // long branch: + // b!CC $PC+8 + // b MBB // - // short branch: - // bCC .L_TARGET_MBB - // - // long branch: - // bInverseCC $PC+8 - // b .L_TARGET_MBB - for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E; - ++MFI) { - MachineBasicBlock *MBB = MFI; - - for (MachineBasicBlock::iterator MBBI = MBB->begin(), EE = MBB->end(); - MBBI != EE; ++MBBI) { - // We may end up deleting the MachineInstr that MBBI points to, so - // remember its opcode now so we can refer to it after calling erase() - unsigned ByteSize = getNumBytesForInstruction(MBBI); - if (MBBI->getOpcode() == PPC::COND_BRANCH) { - MachineBasicBlock::iterator MBBJ = MBBI; - ++MBBJ; - - // condbranch operands: - // 0. CR0 register - // 1. bc opcode - // 2. target MBB - // 3. fallthrough MBB - MachineBasicBlock *trueMBB = - MBBI->getOperand(2).getMachineBasicBlock(); + bool MadeChange = true; + bool EverMadeChange = false; + while (MadeChange) { + // Iteratively expand branches until we reach a fixed point. + MadeChange = false; + + for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E; + ++MFI) { + MachineBasicBlock &MBB = *MFI; + unsigned MBBStartOffset = 0; + for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); + I != E; ++I) { + if (I->getOpcode() != PPC::BCC || I->getOperand(2).isImm()) { + MBBStartOffset += TII->GetInstSizeInBytes(I); + continue; + } - int Displacement = OffsetMap[trueMBB] - ByteCount; - unsigned Opcode = MBBI->getOperand(1).getImmedValue(); - unsigned CRReg = MBBI->getOperand(0).getReg(); - unsigned Inverted = PPCInstrInfo::invertPPCBranchOpcode(Opcode); + // Determine the offset from the current branch to the destination + // block. + MachineBasicBlock *Dest = I->getOperand(2).getMBB(); - if (Displacement >= -32768 && Displacement <= 32767) { - BuildMI(*MBB, MBBJ, Opcode, 2).addReg(CRReg).addMBB(trueMBB); + int BranchSize; + if (Dest->getNumber() <= MBB.getNumber()) { + // If this is a backwards branch, the delta is the offset from the + // start of this block to this branch, plus the sizes of all blocks + // from this block to the dest. + BranchSize = MBBStartOffset; + + for (unsigned i = Dest->getNumber(), e = MBB.getNumber(); i != e; ++i) + BranchSize += BlockSizes[i]; } else { - // Long branch, skip next branch instruction (i.e. $PC+8). - BuildMI(*MBB, MBBJ, Inverted, 2).addReg(CRReg).addImm(2); - BuildMI(*MBB, MBBJ, PPC::B, 1).addMBB(trueMBB); + // Otherwise, add the size of the blocks between this block and the + // dest to the number of bytes left in this block. + BranchSize = -MBBStartOffset; + + for (unsigned i = MBB.getNumber(), e = Dest->getNumber(); i != e; ++i) + BranchSize += BlockSizes[i]; } + + // If this branch is in range, ignore it. + if (isInt<16>(BranchSize)) { + MBBStartOffset += 4; + continue; + } + + // Otherwise, we have to expand it to a long branch. + MachineInstr *OldBranch = I; + DebugLoc dl = OldBranch->getDebugLoc(); + + if (I->getOpcode() == PPC::BCC) { + // The BCC operands are: + // 0. PPC branch predicate + // 1. CR register + // 2. Target MBB + PPC::Predicate Pred = (PPC::Predicate)I->getOperand(0).getImm(); + unsigned CRReg = I->getOperand(1).getReg(); + + // Jump over the uncond branch inst (i.e. $PC+8) on opposite condition. + BuildMI(MBB, I, dl, TII->get(PPC::BCC)) + .addImm(PPC::InvertPredicate(Pred)).addReg(CRReg).addImm(2); + } else if (I->getOpcode() == PPC::BDNZ) { + BuildMI(MBB, I, dl, TII->get(PPC::BDZ)).addImm(2); + } else if (I->getOpcode() == PPC::BDNZ8) { + BuildMI(MBB, I, dl, TII->get(PPC::BDZ8)).addImm(2); + } else if (I->getOpcode() == PPC::BDZ) { + BuildMI(MBB, I, dl, TII->get(PPC::BDNZ)).addImm(2); + } else if (I->getOpcode() == PPC::BDZ8) { + BuildMI(MBB, I, dl, TII->get(PPC::BDNZ8)).addImm(2); + } else { + llvm_unreachable("Unhandled branch type!"); + } + + // Uncond branch to the real destination. + I = BuildMI(MBB, I, dl, TII->get(PPC::B)).addMBB(Dest); + + // Remove the old branch from the function. + OldBranch->eraseFromParent(); - // Erase the psuedo COND_BRANCH instruction, and then back up the - // iterator so that when the for loop increments it, we end up in - // the correct place rather than iterating off the end. - MBB->erase(MBBI); - MBBI = --MBBJ; + // Remember that this instruction is 8-bytes, increase the size of the + // block by 4, remember to iterate. + BlockSizes[MBB.getNumber()] += 4; + MBBStartOffset += 8; + ++NumExpanded; + MadeChange = true; } - ByteCount += ByteSize; } + EverMadeChange |= MadeChange; } - OffsetMap.clear(); + BlockSizes.clear(); return true; }