/// Implements shrink-wrapping of the stack frame. By default, stack frame
/// is created in the function entry block, and is cleaned up in every block
/// that returns. This function finds alternate blocks: one for the frame
/// setup (prolog) and one for the cleanup (epilog).
void HexagonFrameLowering::findShrunkPrologEpilog(MachineFunction &MF,
MachineBasicBlock *&PrologB, MachineBasicBlock *&EpilogB) const {
static unsigned ShrinkCounter = 0;
if (ShrinkLimit.getPosition()) {
if (ShrinkCounter >= ShrinkLimit)
return;
ShrinkCounter++;
}
auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget());
auto &HRI = *HST.getRegisterInfo();
MachineDominatorTree MDT;
MDT.runOnMachineFunction(MF);
MachinePostDominatorTree MPT;
MPT.runOnMachineFunction(MF);
typedef DenseMap<unsigned,unsigned> UnsignedMap;
UnsignedMap RPO;
typedef ReversePostOrderTraversal<const MachineFunction*> RPOTType;
RPOTType RPOT(&MF);
unsigned RPON = 0;
for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I)
RPO[(*I)->getNumber()] = RPON++;
// Don't process functions that have loops, at least for now. Placement
// of prolog and epilog must take loop structure into account. For simpli-
// city don't do it right now.
for (auto &I : MF) {
unsigned BN = RPO[I.getNumber()];
for (auto SI = I.succ_begin(), SE = I.succ_end(); SI != SE; ++SI) {
// If found a back-edge, return.
if (RPO[(*SI)->getNumber()] <= BN)
return;
}
}
// Collect the set of blocks that need a stack frame to execute. Scan
// each block for uses/defs of callee-saved registers, calls, etc.
SmallVector<MachineBasicBlock*,16> SFBlocks;
BitVector CSR(Hexagon::NUM_TARGET_REGS);
for (const MCPhysReg *P = HRI.getCalleeSavedRegs(&MF); *P; ++P)
CSR[*P] = true;
for (auto &I : MF)
if (needsStackFrame(I, CSR))
SFBlocks.push_back(&I);
DEBUG({
dbgs() << "Blocks needing SF: {";
for (auto &B : SFBlocks)
dbgs() << " BB#" << B->getNumber();
dbgs() << " }\n";
});
