#include "llvm/CodeGen/MachineLocation.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
+#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CommandLine.h"
using namespace llvm;
+static cl::opt<bool>
+ReuseFrameIndexVals("arm-reuse-frame-index-vals", cl::Hidden, cl::init(true),
+ cl::desc("Reuse repeated frame index values"));
+
unsigned ARMBaseRegisterInfo::getRegisterNumbering(unsigned RegEnum,
bool *isSPVFP) {
if (isSPVFP)
return false;
}
+const TargetRegisterClass *
+ARMBaseRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
+ const TargetRegisterClass *B,
+ unsigned SubIdx) const {
+ switch (SubIdx) {
+ default: return 0;
+ case 1:
+ case 2:
+ case 3:
+ case 4:
+ // S sub-registers.
+ if (A->getSize() == 8) {
+ if (B == &ARM::SPR_8RegClass)
+ return &ARM::DPR_8RegClass;
+ assert(B == &ARM::SPRRegClass && "Expecting SPR register class!");
+ if (A == &ARM::DPR_8RegClass)
+ return A;
+ return &ARM::DPR_VFP2RegClass;
+ }
+
+ assert(A->getSize() == 16 && "Expecting a Q register class!");
+ if (B == &ARM::SPR_8RegClass)
+ return &ARM::QPR_8RegClass;
+ return &ARM::QPR_VFP2RegClass;
+ case 5:
+ case 6:
+ // D sub-registers.
+ if (B == &ARM::DPR_VFP2RegClass)
+ return &ARM::QPR_VFP2RegClass;
+ if (B == &ARM::DPR_8RegClass)
+ return &ARM::QPR_8RegClass;
+ return A;
+ }
+ return 0;
+}
+
const TargetRegisterClass *
ARMBaseRegisterInfo::getPointerRegClass(unsigned Kind) const {
- return &ARM::GPRRegClass;
+ return ARM::GPRRegisterClass;
}
/// getAllocationOrder - Returns the register allocation order for a specified
}
}
+static unsigned calculateMaxStackAlignment(const MachineFrameInfo *FFI) {
+ unsigned MaxAlign = 0;
+
+ for (int i = FFI->getObjectIndexBegin(),
+ e = FFI->getObjectIndexEnd(); i != e; ++i) {
+ if (FFI->isDeadObjectIndex(i))
+ continue;
+
+ unsigned Align = FFI->getObjectAlignment(i);
+ MaxAlign = std::max(MaxAlign, Align);
+ }
+
+ return MaxAlign;
+}
+
/// hasFP - Return true if the specified function should have a dedicated frame
/// pointer register. This is true if the function has variable sized allocas
/// or if frame pointer elimination is disabled.
bool ARMBaseRegisterInfo::hasFP(const MachineFunction &MF) const {
const MachineFrameInfo *MFI = MF.getFrameInfo();
return (NoFramePointerElim ||
+ needsStackRealignment(MF) ||
MFI->hasVarSizedObjects() ||
MFI->isFrameAddressTaken());
}
+bool ARMBaseRegisterInfo::
+needsStackRealignment(const MachineFunction &MF) const {
+ const MachineFrameInfo *MFI = MF.getFrameInfo();
+ const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned StackAlign = MF.getTarget().getFrameInfo()->getStackAlignment();
+ return (RealignStack &&
+ !AFI->isThumb1OnlyFunction() &&
+ (MFI->getMaxAlignment() > StackAlign) &&
+ !MFI->hasVarSizedObjects());
+}
+
bool ARMBaseRegisterInfo::cannotEliminateFrame(const MachineFunction &MF) const {
const MachineFrameInfo *MFI = MF.getFrameInfo();
if (NoFramePointerElim && MFI->hasCalls())
return true;
- return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken();
+ return MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken()
+ || needsStackRealignment(MF);
}
/// estimateStackSize - Estimate and return the size of the frame.
SmallVector<unsigned, 4> UnspilledCS2GPRs;
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+
+ // Calculate and set max stack object alignment early, so we can decide
+ // whether we will need stack realignment (and thus FP).
+ if (RealignStack) {
+ unsigned MaxAlign = std::max(MFI->getMaxAlignment(),
+ calculateMaxStackAlignment(MFI));
+ MFI->setMaxAlignment(MaxAlign);
+ }
+
// Don't spill FP if the frame can be eliminated. This is determined
// by scanning the callee-save registers to see if any is used.
const unsigned *CSRegs = getCalleeSavedRegs();
}
}
- if (CSRegClasses[i] == &ARM::GPRRegClass) {
+ if (CSRegClasses[i] == ARM::GPRRegisterClass ||
+ CSRegClasses[i] == ARM::tGPRRegisterClass) {
if (Spilled) {
NumGPRSpills++;
// Estimate if we might need to scavenge a register at some point in order
// to materialize a stack offset. If so, either spill one additional
// callee-saved register or reserve a special spill slot to facilitate
- // register scavenging.
- if (RS && !ExtraCSSpill && !AFI->isThumb1OnlyFunction()) {
+ // register scavenging. Thumb1 needs a spill slot for stack pointer
+ // adjustments also, even when the frame itself is small.
+ if (RS && !ExtraCSSpill) {
MachineFrameInfo *MFI = MF.getFrameInfo();
- if (estimateStackSize(MF, MFI) >= estimateRSStackSizeLimit(MF)) {
+ // If any of the stack slot references may be out of range of an
+ // immediate offset, make sure a register (or a spill slot) is
+ // available for the register scavenger. Note that if we're indexing
+ // off the frame pointer, the effective stack size is 4 bytes larger
+ // since the FP points to the stack slot of the previous FP.
+ if (estimateStackSize(MF, MFI) + (hasFP(MF) ? 4 : 0)
+ >= estimateRSStackSizeLimit(MF)) {
// If any non-reserved CS register isn't spilled, just spill one or two
// extra. That should take care of it!
unsigned NumExtras = TargetAlign / 4;
NumExtras--;
}
}
- while (NumExtras && !UnspilledCS2GPRs.empty()) {
- unsigned Reg = UnspilledCS2GPRs.back();
- UnspilledCS2GPRs.pop_back();
- if (!isReservedReg(MF, Reg)) {
- Extras.push_back(Reg);
- NumExtras--;
+ // For non-Thumb1 functions, also check for hi-reg CS registers
+ if (!AFI->isThumb1OnlyFunction()) {
+ while (NumExtras && !UnspilledCS2GPRs.empty()) {
+ unsigned Reg = UnspilledCS2GPRs.back();
+ UnspilledCS2GPRs.pop_back();
+ if (!isReservedReg(MF, Reg)) {
+ Extras.push_back(Reg);
+ NumExtras--;
+ }
}
}
if (Extras.size() && NumExtras == 0) {
MF.getRegInfo().setPhysRegUsed(Extras[i]);
AFI->setCSRegisterIsSpilled(Extras[i]);
}
- } else {
+ } else if (!AFI->isThumb1OnlyFunction()) {
+ // note: Thumb1 functions spill to R12, not the stack.
// Reserve a slot closest to SP or frame pointer.
- const TargetRegisterClass *RC = &ARM::GPRRegClass;
+ const TargetRegisterClass *RC = ARM::GPRRegisterClass;
RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
- RC->getAlignment()));
+ RC->getAlignment(),
+ false));
}
}
}
return ARM::LR;
}
-unsigned ARMBaseRegisterInfo::getFrameRegister(MachineFunction &MF) const {
+unsigned
+ARMBaseRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
if (STI.isTargetDarwin() || hasFP(MF))
return FramePtr;
return ARM::SP;
case ARM::R1:
return ARM::R0;
case ARM::R3:
- // FIXME!
- return STI.isThumb1Only() ? 0 : ARM::R2;
+ return ARM::R2;
case ARM::R5:
return ARM::R4;
case ARM::R7:
case ARM::R0:
return ARM::R1;
case ARM::R2:
- // FIXME!
- return STI.isThumb1Only() ? 0 : ARM::R3;
+ return ARM::R3;
case ARM::R4:
return ARM::R5;
case ARM::R6:
return true;
}
+bool ARMBaseRegisterInfo::
+requiresFrameIndexScavenging(const MachineFunction &MF) const {
+ return true;
+}
+
// hasReservedCallFrame - Under normal circumstances, when a frame pointer is
// not required, we reserve argument space for call sites in the function
// immediately on entry to the current function. This eliminates the need for
MBB.erase(I);
}
-/// findScratchRegister - Find a 'free' ARM register. If register scavenger
-/// is not being used, R12 is available. Otherwise, try for a call-clobbered
-/// register first and then a spilled callee-saved register if that fails.
-static
-unsigned findScratchRegister(RegScavenger *RS, const TargetRegisterClass *RC,
- ARMFunctionInfo *AFI) {
- unsigned Reg = RS ? RS->FindUnusedReg(RC) : (unsigned) ARM::R12;
- assert(!AFI->isThumb1OnlyFunction());
- return Reg;
-}
-
-void
+unsigned
ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
- int SPAdj, RegScavenger *RS) const {
+ int SPAdj, int *Value,
+ RegScavenger *RS) const {
unsigned i = 0;
MachineInstr &MI = *II;
MachineBasicBlock &MBB = *MI.getParent();
const MachineFrameInfo *MFI = MF.getFrameInfo();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
assert(!AFI->isThumb1OnlyFunction() &&
- "This eliminateFrameIndex does not suppor Thumb1!");
+ "This eliminateFrameIndex does not support Thumb1!");
while (!MI.getOperand(i).isFI()) {
++i;
unsigned FrameReg = ARM::SP;
int FrameIndex = MI.getOperand(i).getIndex();
int Offset = MFI->getObjectOffset(FrameIndex) + MFI->getStackSize() + SPAdj;
+ bool isFixed = MFI->isFixedObjectIndex(FrameIndex);
+ // When doing dynamic stack realignment, all of these need to change(?)
if (AFI->isGPRCalleeSavedArea1Frame(FrameIndex))
Offset -= AFI->getGPRCalleeSavedArea1Offset();
else if (AFI->isGPRCalleeSavedArea2Frame(FrameIndex))
Offset -= AFI->getGPRCalleeSavedArea2Offset();
else if (AFI->isDPRCalleeSavedAreaFrame(FrameIndex))
Offset -= AFI->getDPRCalleeSavedAreaOffset();
- else if (hasFP(MF) && AFI->hasStackFrame()) {
+ else if (needsStackRealignment(MF)) {
+ // When dynamically realigning the stack, use the frame pointer for
+ // parameters, and the stack pointer for locals.
+ assert (hasFP(MF) && "dynamic stack realignment without a FP!");
+ if (isFixed) {
+ FrameReg = getFrameRegister(MF);
+ Offset -= AFI->getFramePtrSpillOffset();
+ // When referencing from the frame pointer, stack pointer adjustments
+ // don't matter.
+ SPAdj = 0;
+ }
+ } else if (hasFP(MF) && AFI->hasStackFrame()) {
assert(SPAdj == 0 && "Unexpected stack offset!");
- // Use frame pointer to reference fixed objects unless this is a
- // frameless function,
- FrameReg = getFrameRegister(MF);
- Offset -= AFI->getFramePtrSpillOffset();
+ if (isFixed || MFI->hasVarSizedObjects()) {
+ // Use frame pointer to reference fixed objects unless this is a
+ // frameless function.
+ FrameReg = getFrameRegister(MF);
+ Offset -= AFI->getFramePtrSpillOffset();
+ } else if (AFI->isThumb2Function()) {
+ // In Thumb2 mode, the negative offset is very limited.
+ int FPOffset = Offset - AFI->getFramePtrSpillOffset();
+ if (FPOffset >= -255 && FPOffset < 0) {
+ FrameReg = getFrameRegister(MF);
+ Offset = FPOffset;
+ }
+ }
}
- // modify MI as necessary to handle as much of 'Offset' as possible
+ // Modify MI as necessary to handle as much of 'Offset' as possible
bool Done = false;
if (!AFI->isThumbFunction())
Done = rewriteARMFrameIndex(MI, i, FrameReg, Offset, TII);
Done = rewriteT2FrameIndex(MI, i, FrameReg, Offset, TII);
}
if (Done)
- return;
-
- const TargetInstrDesc &Desc = MI.getDesc();
+ return 0;
// If we get here, the immediate doesn't fit into the instruction. We folded
// as much as possible above, handle the rest, providing a register that is
// SP+LargeImm.
- assert((Offset || (Desc.TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4) &&
+ assert((Offset ||
+ (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode4 ||
+ (MI.getDesc().TSFlags & ARMII::AddrModeMask) == ARMII::AddrMode6) &&
"This code isn't needed if offset already handled!");
- // Insert a set of r12 with the full address: r12 = sp + offset
- // If the offset we have is too large to fit into the instruction, we need
- // to form it with a series of ADDri's. Do this by taking 8-bit chunks
- // out of 'Offset'.
- unsigned ScratchReg = findScratchRegister(RS, &ARM::GPRRegClass, AFI);
- if (ScratchReg == 0)
- // No register is "free". Scavenge a register.
- ScratchReg = RS->scavengeRegister(&ARM::GPRRegClass, II, SPAdj);
+ unsigned ScratchReg = 0;
int PIdx = MI.findFirstPredOperandIdx();
ARMCC::CondCodes Pred = (PIdx == -1)
? ARMCC::AL : (ARMCC::CondCodes)MI.getOperand(PIdx).getImm();
unsigned PredReg = (PIdx == -1) ? 0 : MI.getOperand(PIdx+1).getReg();
if (Offset == 0)
- // Must be addrmode4.
+ // Must be addrmode4/6.
MI.getOperand(i).ChangeToRegister(FrameReg, false, false, false);
else {
+ ScratchReg = MF.getRegInfo().createVirtualRegister(ARM::GPRRegisterClass);
+ if (Value) *Value = Offset;
if (!AFI->isThumbFunction())
emitARMRegPlusImmediate(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg,
Offset, Pred, PredReg, TII);
Offset, Pred, PredReg, TII);
}
MI.getOperand(i).ChangeToRegister(ScratchReg, false, false, true);
+ if (!ReuseFrameIndexVals)
+ ScratchReg = 0;
}
+ return ScratchReg;
}
-/// Move iterator pass the next bunch of callee save load / store ops for
+/// Move iterator past the next bunch of callee save load / store ops for
/// the particular spill area (1: integer area 1, 2: integer area 2,
/// 3: fp area, 0: don't care).
static void movePastCSLoadStoreOps(MachineBasicBlock &MBB,
unsigned GPRCS1Size = 0, GPRCS2Size = 0, DPRCSSize = 0;
int FramePtrSpillFI = 0;
+ // Allocate the vararg register save area. This is not counted in NumBytes.
if (VARegSaveSize)
emitSPUpdate(isARM, MBB, MBBI, dl, TII, -VARegSaveSize);
emitSPUpdate(isARM, MBB, MBBI, dl, TII, -GPRCS1Size);
movePastCSLoadStoreOps(MBB, MBBI, ARM::STR, ARM::t2STRi12, 1, STI);
- // Darwin ABI requires FP to point to the stack slot that contains the
- // previous FP.
+ // Set FP to point to the stack slot that contains the previous FP.
+ // For Darwin, FP is R7, which has now been stored in spill area 1.
+ // Otherwise, if this is not Darwin, all the callee-saved registers go
+ // into spill area 1, including the FP in R11. In either case, it is
+ // now safe to emit this assignment.
if (STI.isTargetDarwin() || hasFP(MF)) {
unsigned ADDriOpc = !AFI->isThumbFunction() ? ARM::ADDri : ARM::t2ADDri;
MachineInstrBuilder MIB =
AFI->setGPRCalleeSavedArea2Offset(GPRCS2Offset);
AFI->setDPRCalleeSavedAreaOffset(DPRCSOffset);
+ movePastCSLoadStoreOps(MBB, MBBI, ARM::VSTRD, 0, 3, STI);
NumBytes = DPRCSOffset;
if (NumBytes) {
- // Insert it after all the callee-save spills.
- movePastCSLoadStoreOps(MBB, MBBI, ARM::FSTD, 0, 3, STI);
+ // Adjust SP after all the callee-save spills.
emitSPUpdate(isARM, MBB, MBBI, dl, TII, -NumBytes);
}
AFI->setGPRCalleeSavedArea1Size(GPRCS1Size);
AFI->setGPRCalleeSavedArea2Size(GPRCS2Size);
AFI->setDPRCalleeSavedAreaSize(DPRCSSize);
+
+ // If we need dynamic stack realignment, do it here.
+ if (needsStackRealignment(MF)) {
+ unsigned Opc;
+ unsigned MaxAlign = MFI->getMaxAlignment();
+ assert (!AFI->isThumb1OnlyFunction());
+ Opc = AFI->isThumbFunction() ? ARM::t2BICri : ARM::BICri;
+
+ AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(Opc), ARM::SP)
+ .addReg(ARM::SP, RegState::Kill)
+ .addImm(MaxAlign-1)));
+ }
}
static bool isCalleeSavedRegister(unsigned Reg, const unsigned *CSRegs) {
static bool isCSRestore(MachineInstr *MI,
const ARMBaseInstrInfo &TII,
const unsigned *CSRegs) {
- return ((MI->getOpcode() == (int)ARM::FLDD ||
+ return ((MI->getOpcode() == (int)ARM::VLDRD ||
MI->getOpcode() == (int)ARM::LDR ||
MI->getOpcode() == (int)ARM::t2LDRi12) &&
MI->getOperand(1).isFI() &&
if (NumBytes != 0)
emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
} else {
- // Unwind MBBI to point to first LDR / FLDD.
+ // Unwind MBBI to point to first LDR / VLDRD.
const unsigned *CSRegs = getCalleeSavedRegs();
if (MBBI != MBB.begin()) {
do
emitSPUpdate(isARM, MBB, MBBI, dl, TII, NumBytes);
// Move SP to start of integer callee save spill area 2.
- movePastCSLoadStoreOps(MBB, MBBI, ARM::FLDD, 0, 3, STI);
+ movePastCSLoadStoreOps(MBB, MBBI, ARM::VLDRD, 0, 3, STI);
emitSPUpdate(isARM, MBB, MBBI, dl, TII, AFI->getDPRCalleeSavedAreaSize());
// Move SP to start of integer callee save spill area 1.
emitSPUpdate(isARM, MBB, MBBI, dl, TII, VARegSaveSize);
}
+namespace {
+ struct MaximalStackAlignmentCalculator : public MachineFunctionPass {
+ static char ID;
+ MaximalStackAlignmentCalculator() : MachineFunctionPass(&ID) {}
+
+ virtual bool runOnMachineFunction(MachineFunction &MF) {
+ MachineFrameInfo *FFI = MF.getFrameInfo();
+ MachineRegisterInfo &RI = MF.getRegInfo();
+
+ // Calculate max stack alignment of all already allocated stack objects.
+ unsigned MaxAlign = calculateMaxStackAlignment(FFI);
+
+ // Be over-conservative: scan over all vreg defs and find, whether vector
+ // registers are used. If yes - there is probability, that vector register
+ // will be spilled and thus stack needs to be aligned properly.
+ for (unsigned RegNum = TargetRegisterInfo::FirstVirtualRegister;
+ RegNum < RI.getLastVirtReg(); ++RegNum)
+ MaxAlign = std::max(MaxAlign, RI.getRegClass(RegNum)->getAlignment());
+
+ if (FFI->getMaxAlignment() == MaxAlign)
+ return false;
+
+ FFI->setMaxAlignment(MaxAlign);
+ return true;
+ }
+
+ virtual const char *getPassName() const {
+ return "ARM Stack Required Alignment Auto-Detector";
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesCFG();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
+ };
+
+ char MaximalStackAlignmentCalculator::ID = 0;
+}
+
+FunctionPass*
+llvm::createARMMaxStackAlignmentCalculatorPass() {
+ return new MaximalStackAlignmentCalculator();
+}
+
#include "ARMGenRegisterInfo.inc"
projects / oota-llvm.git / blobdiff