-//===-- 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.
//
//===----------------------------------------------------------------------===//
#include "PPC.h"
+#include "MCTargetDesc/PPCPredicates.h"
#include "PPCInstrBuilder.h"
#include "PPCInstrInfo.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetAsmInfo.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/Support/Compiler.h"
-#include <map>
+#include "llvm/Target/TargetSubtargetInfo.h"
using namespace llvm;
-static Statistic<> NumExpanded("ppc-branch-select",
- "Num branches expanded to long format");
+#define DEBUG_TYPE "ppc-branch-select"
+
+STATISTIC(NumExpanded, "Number of branches expanded to long format");
+
+namespace llvm {
+ void initializePPCBSelPass(PassRegistry&);
+}
namespace {
- struct VISIBILITY_HIDDEN PPCBSel : public MachineFunctionPass {
- /// OffsetMap - Mapping between BB and byte offset from start of function.
- /// TODO: replace this with a vector, using the MBB idx as the key.
- std::map<MachineBasicBlock*, unsigned> OffsetMap;
+ struct PPCBSel : public MachineFunctionPass {
+ static char ID;
+ PPCBSel() : MachineFunctionPass(ID) {
+ initializePPCBSelPass(*PassRegistry::getPassRegistry());
+ }
+
+ /// BlockSizes - The sizes of the basic blocks in the function.
+ std::vector<unsigned> BlockSizes;
- virtual bool runOnMachineFunction(MachineFunction &Fn);
+ bool runOnMachineFunction(MachineFunction &Fn) override;
- virtual const char *getPassName() const {
- return "PowerPC Branch Selection";
+ const char *getPassName() const override {
+ return "PowerPC Branch Selector";
}
};
+ char PPCBSel::ID = 0;
}
+INITIALIZE_PASS(PPCBSel, "ppc-branch-select", "PowerPC Branch Selector",
+ false, false)
+
/// createPPCBranchSelectionPass - returns an instance of the Branch Selection
/// Pass
///
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.
- MachineFunction *MF = MI->getParent()->getParent();
- const char *AsmStr = MI->getOperand(0).getSymbolName();
- return MF->getTarget().getTargetAsmInfo()->getInlineAsmLength(AsmStr);
- }
- 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<const PPCInstrInfo *>(Fn.getSubtarget().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();
-
- int Displacement = OffsetMap[trueMBB] - ByteCount;
- unsigned Opcode = MBBI->getOperand(1).getImmedValue();
- unsigned CRReg = MBBI->getOperand(0).getReg();
- unsigned Inverted = PPCInstrInfo::invertPPCBranchOpcode(Opcode);
+ 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) {
+ MachineBasicBlock *Dest = nullptr;
+ if (I->getOpcode() == PPC::BCC && !I->getOperand(2).isImm())
+ Dest = I->getOperand(2).getMBB();
+ else if ((I->getOpcode() == PPC::BC || I->getOpcode() == PPC::BCn) &&
+ !I->getOperand(1).isImm())
+ Dest = I->getOperand(1).getMBB();
+ else if ((I->getOpcode() == PPC::BDNZ8 || I->getOpcode() == PPC::BDNZ ||
+ I->getOpcode() == PPC::BDZ8 || I->getOpcode() == PPC::BDZ) &&
+ !I->getOperand(0).isImm())
+ Dest = I->getOperand(0).getMBB();
+
+ if (!Dest) {
+ MBBStartOffset += TII->GetInstSizeInBytes(I);
+ continue;
+ }
- if (Displacement >= -32768 && Displacement <= 32767) {
- BuildMI(*MBB, MBBJ, Opcode, 2).addReg(CRReg).addMBB(trueMBB);
+ // Determine the offset from the current branch to the destination
+ // block.
+ 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).
- ++NumExpanded;
- 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::BC) {
+ unsigned CRBit = I->getOperand(0).getReg();
+ BuildMI(MBB, I, dl, TII->get(PPC::BCn)).addReg(CRBit).addImm(2);
+ } else if (I->getOpcode() == PPC::BCn) {
+ unsigned CRBit = I->getOperand(0).getReg();
+ BuildMI(MBB, I, dl, TII->get(PPC::BC)).addReg(CRBit).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;
}