RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-26 22:42:46 +02:00
Code Issues Projects Releases Wiki Activity
6e3e66375a
llvm-mirror/lib/Target/ARM/Thumb2InstrInfo.cpp
Evan Cheng 6e3e66375a - Add pseudo instructions tLDRpci_pic and t2LDRpci_pic which does a pc-relative 
load of a GV from constantpool and then add pc. It allows the code sequence to
  be rematerializable so it would be hoisted by machine licm.
- Add a late pass to break these pseudo instructions into a number of real
  instructions. Also move the code in Thumb2 IT pass that breaks up t2MOVi32imm
  to this pass. This is done before post regalloc scheduling to allow the
  scheduler to proper schedule these instructions. It also allow them to be
  if-converted and shrunk by later passes.

llvm-svn: 86304
2009-11-06 23:52:48 +00:00

546 lines
18 KiB
C++
Raw Blame History

//===- Thumb2InstrInfo.cpp - Thumb-2 Instruction Information --------*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains the Thumb-2 implementation of the TargetInstrInfo class.
//
//===----------------------------------------------------------------------===//
#include "Thumb2InstrInfo.h"
#include "ARM.h"
#include "ARMConstantPoolValue.h"
#include "ARMAddressingModes.h"
#include "ARMGenInstrInfo.inc"
#include "ARMMachineFunctionInfo.h"
#include "llvm/GlobalValue.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/ADT/SmallVector.h"
#include "Thumb2InstrInfo.h"
using namespace llvm;
Thumb2InstrInfo::Thumb2InstrInfo(const ARMSubtarget &STI)
: ARMBaseInstrInfo(STI), RI(*this, STI) {
}
unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const {
// FIXME
return 0;
}
bool
Thumb2InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
if (MBB.empty()) return false;
switch (MBB.back().getOpcode()) {
case ARM::t2LDM_RET:
case ARM::t2B: // Uncond branch.
case ARM::t2BR_JT: // Jumptable branch.
case ARM::t2TBB: // Table branch byte.
case ARM::t2TBH: // Table branch halfword.
case ARM::tBR_JTr: // Jumptable branch (16-bit version).
case ARM::tBX_RET:
case ARM::tBX_RET_vararg:
case ARM::tPOP_RET:
case ARM::tB:
case ARM::tBRIND:
return true;
default:
break;
}
return false;
}
bool
Thumb2InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
unsigned DestReg, unsigned SrcReg,
const TargetRegisterClass *DestRC,
const TargetRegisterClass *SrcRC) const {
DebugLoc DL = DebugLoc::getUnknownLoc();
if (I != MBB.end()) DL = I->getDebugLoc();
if (DestRC == ARM::GPRRegisterClass &&
SrcRC == ARM::GPRRegisterClass) {
BuildMI(MBB, I, DL, get(ARM::tMOVgpr2gpr), DestReg).addReg(SrcReg);
return true;
} else if (DestRC == ARM::GPRRegisterClass &&
SrcRC == ARM::tGPRRegisterClass) {
BuildMI(MBB, I, DL, get(ARM::tMOVtgpr2gpr), DestReg).addReg(SrcReg);
return true;
} else if (DestRC == ARM::tGPRRegisterClass &&
SrcRC == ARM::GPRRegisterClass) {
BuildMI(MBB, I, DL, get(ARM::tMOVgpr2tgpr), DestReg).addReg(SrcReg);
return true;
}
// Handle SPR, DPR, and QPR copies.
return ARMBaseInstrInfo::copyRegToReg(MBB, I, DestReg, SrcReg, DestRC, SrcRC);
}
void Thumb2InstrInfo::
storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
unsigned SrcReg, bool isKill, int FI,
const TargetRegisterClass *RC) const {
DebugLoc DL = DebugLoc::getUnknownLoc();
if (I != MBB.end()) DL = I->getDebugLoc();
if (RC == ARM::GPRRegisterClass) {
MachineFunction &MF = *MBB.getParent();
MachineFrameInfo &MFI = *MF.getFrameInfo();
MachineMemOperand *MMO =
MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI),
MachineMemOperand::MOStore, 0,
MFI.getObjectSize(FI),
MFI.getObjectAlignment(FI));
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::t2STRi12))
.addReg(SrcReg, getKillRegState(isKill))
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
return;
}
ARMBaseInstrInfo::storeRegToStackSlot(MBB, I, SrcReg, isKill, FI, RC);
}
void Thumb2InstrInfo::
loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
unsigned DestReg, int FI,
const TargetRegisterClass *RC) const {
DebugLoc DL = DebugLoc::getUnknownLoc();
if (I != MBB.end()) DL = I->getDebugLoc();
if (RC == ARM::GPRRegisterClass) {
MachineFunction &MF = *MBB.getParent();
MachineFrameInfo &MFI = *MF.getFrameInfo();
MachineMemOperand *MMO =
MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI),
MachineMemOperand::MOLoad, 0,
MFI.getObjectSize(FI),
MFI.getObjectAlignment(FI));
AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::t2LDRi12), DestReg)
.addFrameIndex(FI).addImm(0).addMemOperand(MMO));
return;
}
ARMBaseInstrInfo::loadRegFromStackSlot(MBB, I, DestReg, FI, RC);
}
void Thumb2InstrInfo::reMaterialize(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
unsigned DestReg, unsigned SubIdx,
const MachineInstr *Orig) const {
DebugLoc dl = Orig->getDebugLoc();
unsigned Opcode = Orig->getOpcode();
switch (Opcode) {
default: {
MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
MI->getOperand(0).setReg(DestReg);
MBB.insert(I, MI);
break;
}
case ARM::t2LDRpci_pic: {
MachineFunction &MF = *MBB.getParent();
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
MachineConstantPool *MCP = MF.getConstantPool();
unsigned CPI = Orig->getOperand(1).getIndex();
const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI];
assert(MCPE.isMachineConstantPoolEntry() &&
"Expecting a machine constantpool entry!");
ARMConstantPoolValue *ACPV =
static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal);
assert(ACPV->isGlobalValue() && "Expecting a GV!");
unsigned PCLabelId = AFI->createConstPoolEntryUId();
ARMConstantPoolValue *NewCPV =
new ARMConstantPoolValue(ACPV->getGV(), PCLabelId, ARMCP::CPValue, 4);
CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment());
MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode),
DestReg)
.addConstantPoolIndex(CPI).addImm(PCLabelId);
(*MIB).setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end());
break;
}
}
MachineInstr *NewMI = prior(I);
NewMI->getOperand(0).setSubReg(SubIdx);
}
void llvm::emitT2RegPlusImmediate(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI, DebugLoc dl,
unsigned DestReg, unsigned BaseReg, int NumBytes,
ARMCC::CondCodes Pred, unsigned PredReg,
const ARMBaseInstrInfo &TII) {
bool isSub = NumBytes < 0;
if (isSub) NumBytes = -NumBytes;
// If profitable, use a movw or movt to materialize the offset.
// FIXME: Use the scavenger to grab a scratch register.
if (DestReg != ARM::SP && DestReg != BaseReg &&
NumBytes >= 4096 &&
ARM_AM::getT2SOImmVal(NumBytes) == -1) {
bool Fits = false;
if (NumBytes < 65536) {
// Use a movw to materialize the 16-bit constant.
BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), DestReg)
.addImm(NumBytes)
.addImm((unsigned)Pred).addReg(PredReg).addReg(0);
Fits = true;
} else if ((NumBytes & 0xffff) == 0) {
// Use a movt to materialize the 32-bit constant.
BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVTi16), DestReg)
.addReg(DestReg)
.addImm(NumBytes >> 16)
.addImm((unsigned)Pred).addReg(PredReg).addReg(0);
Fits = true;
}
if (Fits) {
if (isSub) {
BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr), DestReg)
.addReg(BaseReg, RegState::Kill)
.addReg(DestReg, RegState::Kill)
.addImm((unsigned)Pred).addReg(PredReg).addReg(0);
} else {
BuildMI(MBB, MBBI, dl, TII.get(ARM::t2ADDrr), DestReg)
.addReg(DestReg, RegState::Kill)
.addReg(BaseReg, RegState::Kill)
.addImm((unsigned)Pred).addReg(PredReg).addReg(0);
}
return;
}
}
while (NumBytes) {
unsigned ThisVal = NumBytes;
unsigned Opc = 0;
if (DestReg == ARM::SP && BaseReg != ARM::SP) {
// mov sp, rn. Note t2MOVr cannot be used.
BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVgpr2gpr),DestReg).addReg(BaseReg);
BaseReg = ARM::SP;
continue;
}
if (BaseReg == ARM::SP) {
// sub sp, sp, #imm7
if (DestReg == ARM::SP && (ThisVal < ((1 << 7)-1) * 4)) {
assert((ThisVal & 3) == 0 && "Stack update is not multiple of 4?");
Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
// FIXME: Fix Thumb1 immediate encoding.
BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
.addReg(BaseReg).addImm(ThisVal/4);
NumBytes = 0;
continue;
}
// sub rd, sp, so_imm
Opc = isSub ? ARM::t2SUBrSPi : ARM::t2ADDrSPi;
if (ARM_AM::getT2SOImmVal(NumBytes) != -1) {
NumBytes = 0;
} else {
// FIXME: Move this to ARMAddressingModes.h?
unsigned RotAmt = CountLeadingZeros_32(ThisVal);
ThisVal = ThisVal & ARM_AM::rotr32(0xff000000U, RotAmt);
NumBytes &= ~ThisVal;
assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 &&
"Bit extraction didn't work?");
}
} else {
assert(DestReg != ARM::SP && BaseReg != ARM::SP);
Opc = isSub ? ARM::t2SUBri : ARM::t2ADDri;
if (ARM_AM::getT2SOImmVal(NumBytes) != -1) {
NumBytes = 0;
} else if (ThisVal < 4096) {
Opc = isSub ? ARM::t2SUBri12 : ARM::t2ADDri12;
NumBytes = 0;
} else {
// FIXME: Move this to ARMAddressingModes.h?
unsigned RotAmt = CountLeadingZeros_32(ThisVal);
ThisVal = ThisVal & ARM_AM::rotr32(0xff000000U, RotAmt);
NumBytes &= ~ThisVal;
assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 &&
"Bit extraction didn't work?");
}
}
// Build the new ADD / SUB.
AddDefaultCC(AddDefaultPred(BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
.addReg(BaseReg, RegState::Kill)
.addImm(ThisVal)));
BaseReg = DestReg;
}
}
static unsigned
negativeOffsetOpcode(unsigned opcode)
{
switch (opcode) {
case ARM::t2LDRi12: return ARM::t2LDRi8;
case ARM::t2LDRHi12: return ARM::t2LDRHi8;
case ARM::t2LDRBi12: return ARM::t2LDRBi8;
case ARM::t2LDRSHi12: return ARM::t2LDRSHi8;
case ARM::t2LDRSBi12: return ARM::t2LDRSBi8;
case ARM::t2STRi12: return ARM::t2STRi8;
case ARM::t2STRBi12: return ARM::t2STRBi8;
case ARM::t2STRHi12: return ARM::t2STRHi8;
case ARM::t2LDRi8:
case ARM::t2LDRHi8:
case ARM::t2LDRBi8:
case ARM::t2LDRSHi8:
case ARM::t2LDRSBi8:
case ARM::t2STRi8:
case ARM::t2STRBi8:
case ARM::t2STRHi8:
return opcode;
default:
break;
}
return 0;
}
static unsigned
positiveOffsetOpcode(unsigned opcode)
{
switch (opcode) {
case ARM::t2LDRi8: return ARM::t2LDRi12;
case ARM::t2LDRHi8: return ARM::t2LDRHi12;
case ARM::t2LDRBi8: return ARM::t2LDRBi12;
case ARM::t2LDRSHi8: return ARM::t2LDRSHi12;
case ARM::t2LDRSBi8: return ARM::t2LDRSBi12;
case ARM::t2STRi8: return ARM::t2STRi12;
case ARM::t2STRBi8: return ARM::t2STRBi12;
case ARM::t2STRHi8: return ARM::t2STRHi12;
case ARM::t2LDRi12:
case ARM::t2LDRHi12:
case ARM::t2LDRBi12:
case ARM::t2LDRSHi12:
case ARM::t2LDRSBi12:
case ARM::t2STRi12:
case ARM::t2STRBi12:
case ARM::t2STRHi12:
return opcode;
default:
break;
}
return 0;
}
static unsigned
immediateOffsetOpcode(unsigned opcode)
{
switch (opcode) {
case ARM::t2LDRs: return ARM::t2LDRi12;
case ARM::t2LDRHs: return ARM::t2LDRHi12;
case ARM::t2LDRBs: return ARM::t2LDRBi12;
case ARM::t2LDRSHs: return ARM::t2LDRSHi12;
case ARM::t2LDRSBs: return ARM::t2LDRSBi12;
case ARM::t2STRs: return ARM::t2STRi12;
case ARM::t2STRBs: return ARM::t2STRBi12;
case ARM::t2STRHs: return ARM::t2STRHi12;
case ARM::t2LDRi12:
case ARM::t2LDRHi12:
case ARM::t2LDRBi12:
case ARM::t2LDRSHi12:
case ARM::t2LDRSBi12:
case ARM::t2STRi12:
case ARM::t2STRBi12:
case ARM::t2STRHi12:
case ARM::t2LDRi8:
case ARM::t2LDRHi8:
case ARM::t2LDRBi8:
case ARM::t2LDRSHi8:
case ARM::t2LDRSBi8:
case ARM::t2STRi8:
case ARM::t2STRBi8:
case ARM::t2STRHi8:
return opcode;
default:
break;
}
return 0;
}
bool llvm::rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
unsigned FrameReg, int &Offset,
const ARMBaseInstrInfo &TII) {
unsigned Opcode = MI.getOpcode();
const TargetInstrDesc &Desc = MI.getDesc();
unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
bool isSub = false;
// Memory operands in inline assembly always use AddrModeT2_i12.
if (Opcode == ARM::INLINEASM)
AddrMode = ARMII::AddrModeT2_i12; // FIXME. mode for thumb2?
if (Opcode == ARM::t2ADDri || Opcode == ARM::t2ADDri12) {
Offset += MI.getOperand(FrameRegIdx+1).getImm();
bool isSP = FrameReg == ARM::SP;
if (Offset == 0) {
// Turn it into a move.
MI.setDesc(TII.get(ARM::tMOVgpr2gpr));
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
MI.RemoveOperand(FrameRegIdx+1);
Offset = 0;
return true;
}
if (Offset < 0) {
Offset = -Offset;
isSub = true;
MI.setDesc(TII.get(isSP ? ARM::t2SUBrSPi : ARM::t2SUBri));
} else {
MI.setDesc(TII.get(isSP ? ARM::t2ADDrSPi : ARM::t2ADDri));
}
// Common case: small offset, fits into instruction.
if (ARM_AM::getT2SOImmVal(Offset) != -1) {
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
Offset = 0;
return true;
}
// Another common case: imm12.
if (Offset < 4096) {
unsigned NewOpc = isSP
? (isSub ? ARM::t2SUBrSPi12 : ARM::t2ADDrSPi12)
: (isSub ? ARM::t2SUBri12 : ARM::t2ADDri12);
MI.setDesc(TII.get(NewOpc));
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
Offset = 0;
return true;
}
// Otherwise, extract 8 adjacent bits from the immediate into this
// t2ADDri/t2SUBri.
unsigned RotAmt = CountLeadingZeros_32(Offset);
unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xff000000U, RotAmt);
// We will handle these bits from offset, clear them.
Offset &= ~ThisImmVal;
assert(ARM_AM::getT2SOImmVal(ThisImmVal) != -1 &&
"Bit extraction didn't work?");
MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
} else {
// AddrMode4 cannot handle any offset.
if (AddrMode == ARMII::AddrMode4)
return false;
// AddrModeT2_so cannot handle any offset. If there is no offset
// register then we change to an immediate version.
unsigned NewOpc = Opcode;
if (AddrMode == ARMII::AddrModeT2_so) {
unsigned OffsetReg = MI.getOperand(FrameRegIdx+1).getReg();
if (OffsetReg != 0) {
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
return Offset == 0;
}
MI.RemoveOperand(FrameRegIdx+1);
MI.getOperand(FrameRegIdx+1).ChangeToImmediate(0);
NewOpc = immediateOffsetOpcode(Opcode);
AddrMode = ARMII::AddrModeT2_i12;
}
unsigned NumBits = 0;
unsigned Scale = 1;
if (AddrMode == ARMII::AddrModeT2_i8 || AddrMode == ARMII::AddrModeT2_i12) {
// i8 supports only negative, and i12 supports only positive, so
// based on Offset sign convert Opcode to the appropriate
// instruction
Offset += MI.getOperand(FrameRegIdx+1).getImm();
if (Offset < 0) {
NewOpc = negativeOffsetOpcode(Opcode);
NumBits = 8;
isSub = true;
Offset = -Offset;
} else {
NewOpc = positiveOffsetOpcode(Opcode);
NumBits = 12;
}
} else {
// VFP and NEON address modes.
int InstrOffs = 0;
if (AddrMode == ARMII::AddrMode5) {
const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1);
InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
InstrOffs *= -1;
}
NumBits = 8;
Scale = 4;
Offset += InstrOffs * 4;
assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
if (Offset < 0) {
Offset = -Offset;
isSub = true;
}
}
if (NewOpc != Opcode)
MI.setDesc(TII.get(NewOpc));
MachineOperand &ImmOp = MI.getOperand(FrameRegIdx+1);
// Attempt to fold address computation
// Common case: small offset, fits into instruction.
int ImmedOffset = Offset / Scale;
unsigned Mask = (1 << NumBits) - 1;
if ((unsigned)Offset <= Mask * Scale) {
// Replace the FrameIndex with fp/sp
MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
if (isSub) {
if (AddrMode == ARMII::AddrMode5)
// FIXME: Not consistent.
ImmedOffset |= 1 << NumBits;
else
ImmedOffset = -ImmedOffset;
}
ImmOp.ChangeToImmediate(ImmedOffset);
Offset = 0;
return true;
}
// Otherwise, offset doesn't fit. Pull in what we can to simplify
ImmedOffset = ImmedOffset & Mask;
if (isSub) {
if (AddrMode == ARMII::AddrMode5)
// FIXME: Not consistent.
ImmedOffset |= 1 << NumBits;
else {
ImmedOffset = -ImmedOffset;
if (ImmedOffset == 0)
// Change the opcode back if the encoded offset is zero.
MI.setDesc(TII.get(positiveOffsetOpcode(NewOpc)));
}
}
ImmOp.ChangeToImmediate(ImmedOffset);
Offset &= ~(Mask*Scale);
}
Offset = (isSub) ? -Offset : Offset;
return Offset == 0;
}
Reference in New Issue View Git Blame Copy Permalink