llvm-project/llvm/lib/Target/R600/AMDGPUIndirectAddressing.cpp

327 lines
13 KiB
C++

//===-- AMDGPUIndirectAddressing.cpp - Indirect Adressing Support ---------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
///
/// Instructions can use indirect addressing to index the register file as if it
/// were memory. This pass lowers RegisterLoad and RegisterStore instructions
/// to either a COPY or a MOV that uses indirect addressing.
//
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "R600InstrInfo.h"
#include "R600MachineFunctionInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/Debug.h"
using namespace llvm;
namespace {
class AMDGPUIndirectAddressingPass : public MachineFunctionPass {
private:
static char ID;
const AMDGPUInstrInfo *TII;
bool regHasExplicitDef(MachineRegisterInfo &MRI, unsigned Reg) const;
public:
AMDGPUIndirectAddressingPass(TargetMachine &tm) :
MachineFunctionPass(ID),
TII(static_cast<const AMDGPUInstrInfo*>(tm.getInstrInfo()))
{ }
virtual bool runOnMachineFunction(MachineFunction &MF);
const char *getPassName() const { return "R600 Handle indirect addressing"; }
};
} // End anonymous namespace
char AMDGPUIndirectAddressingPass::ID = 0;
FunctionPass *llvm::createAMDGPUIndirectAddressingPass(TargetMachine &tm) {
return new AMDGPUIndirectAddressingPass(tm);
}
bool AMDGPUIndirectAddressingPass::runOnMachineFunction(MachineFunction &MF) {
MachineRegisterInfo &MRI = MF.getRegInfo();
int IndirectBegin = TII->getIndirectIndexBegin(MF);
int IndirectEnd = TII->getIndirectIndexEnd(MF);
if (IndirectBegin == -1) {
// No indirect addressing, we can skip this pass
assert(IndirectEnd == -1);
return false;
}
// The map keeps track of the indirect address that is represented by
// each virtual register. The key is the register and the value is the
// indirect address it uses.
std::map<unsigned, unsigned> RegisterAddressMap;
// First pass - Lower all of the RegisterStore instructions and track which
// registers are live.
for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
BB != BB_E; ++BB) {
// This map keeps track of the current live indirect registers.
// The key is the address and the value is the register
std::map<unsigned, unsigned> LiveAddressRegisterMap;
MachineBasicBlock &MBB = *BB;
for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
I != MBB.end(); I = Next) {
Next = llvm::next(I);
MachineInstr &MI = *I;
if (!TII->isRegisterStore(MI)) {
continue;
}
// Lower RegisterStore
unsigned RegIndex = MI.getOperand(2).getImm();
unsigned Channel = MI.getOperand(3).getImm();
unsigned Address = TII->calculateIndirectAddress(RegIndex, Channel);
const TargetRegisterClass *IndirectStoreRegClass =
TII->getIndirectAddrStoreRegClass(MI.getOperand(0).getReg());
if (MI.getOperand(1).getReg() == AMDGPU::INDIRECT_BASE_ADDR) {
// Direct register access.
unsigned DstReg = MRI.createVirtualRegister(IndirectStoreRegClass);
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY), DstReg)
.addOperand(MI.getOperand(0));
RegisterAddressMap[DstReg] = Address;
LiveAddressRegisterMap[Address] = DstReg;
} else {
// Indirect register access.
MachineInstrBuilder MOV = TII->buildIndirectWrite(BB, I,
MI.getOperand(0).getReg(), // Value
Address,
MI.getOperand(1).getReg()); // Offset
for (int i = IndirectBegin; i <= IndirectEnd; ++i) {
unsigned Addr = TII->calculateIndirectAddress(i, Channel);
unsigned DstReg = MRI.createVirtualRegister(IndirectStoreRegClass);
MOV.addReg(DstReg, RegState::Define | RegState::Implicit);
RegisterAddressMap[DstReg] = Addr;
LiveAddressRegisterMap[Addr] = DstReg;
}
}
MI.eraseFromParent();
}
// Update the live-ins of the succesor blocks
for (MachineBasicBlock::succ_iterator Succ = MBB.succ_begin(),
SuccEnd = MBB.succ_end();
SuccEnd != Succ; ++Succ) {
std::map<unsigned, unsigned>::const_iterator Key, KeyEnd;
for (Key = LiveAddressRegisterMap.begin(),
KeyEnd = LiveAddressRegisterMap.end(); KeyEnd != Key; ++Key) {
(*Succ)->addLiveIn(Key->second);
}
}
}
// Second pass - Lower the RegisterLoad instructions
for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
BB != BB_E; ++BB) {
// Key is the address and the value is the register
std::map<unsigned, unsigned> LiveAddressRegisterMap;
MachineBasicBlock &MBB = *BB;
MachineBasicBlock::livein_iterator LI = MBB.livein_begin();
while (LI != MBB.livein_end()) {
std::vector<unsigned> PhiRegisters;
// Make sure this live in is used for indirect addressing
if (RegisterAddressMap.find(*LI) == RegisterAddressMap.end()) {
++LI;
continue;
}
unsigned Address = RegisterAddressMap[*LI];
LiveAddressRegisterMap[Address] = *LI;
PhiRegisters.push_back(*LI);
// Check if there are other live in registers which map to the same
// indirect address.
for (MachineBasicBlock::livein_iterator LJ = llvm::next(LI),
LE = MBB.livein_end();
LJ != LE; ++LJ) {
unsigned Reg = *LJ;
if (RegisterAddressMap.find(Reg) == RegisterAddressMap.end()) {
continue;
}
if (RegisterAddressMap[Reg] == Address) {
if (!regHasExplicitDef(MRI, Reg)) {
continue;
}
PhiRegisters.push_back(Reg);
}
}
if (PhiRegisters.size() == 1) {
// We don't need to insert a Phi instruction, so we can just add the
// registers to the live list for the block.
LiveAddressRegisterMap[Address] = *LI;
MBB.removeLiveIn(*LI);
} else {
// We need to insert a PHI, because we have the same address being
// written in multiple predecessor blocks.
const TargetRegisterClass *PhiDstClass =
TII->getIndirectAddrStoreRegClass(*(PhiRegisters.begin()));
unsigned PhiDstReg = MRI.createVirtualRegister(PhiDstClass);
MachineInstrBuilder Phi = BuildMI(MBB, MBB.begin(),
MBB.findDebugLoc(MBB.begin()),
TII->get(AMDGPU::PHI), PhiDstReg);
for (std::vector<unsigned>::const_iterator RI = PhiRegisters.begin(),
RE = PhiRegisters.end();
RI != RE; ++RI) {
unsigned Reg = *RI;
MachineInstr *DefInst = MRI.getVRegDef(Reg);
assert(DefInst);
MachineBasicBlock *RegBlock = DefInst->getParent();
Phi.addReg(Reg);
Phi.addMBB(RegBlock);
MBB.removeLiveIn(Reg);
}
RegisterAddressMap[PhiDstReg] = Address;
LiveAddressRegisterMap[Address] = PhiDstReg;
}
LI = MBB.livein_begin();
}
for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
I != MBB.end(); I = Next) {
Next = llvm::next(I);
MachineInstr &MI = *I;
if (!TII->isRegisterLoad(MI)) {
if (MI.getOpcode() == AMDGPU::PHI) {
continue;
}
// Check for indirect register defs
for (unsigned OpIdx = 0, NumOperands = MI.getNumOperands();
OpIdx < NumOperands; ++OpIdx) {
MachineOperand &MO = MI.getOperand(OpIdx);
if (MO.isReg() && MO.isDef() &&
RegisterAddressMap.find(MO.getReg()) != RegisterAddressMap.end()) {
unsigned Reg = MO.getReg();
unsigned LiveAddress = RegisterAddressMap[Reg];
// Chain the live-ins
if (LiveAddressRegisterMap.find(LiveAddress) !=
RegisterAddressMap.end()) {
MI.addOperand(MachineOperand::CreateReg(
LiveAddressRegisterMap[LiveAddress],
false, // isDef
true, // isImp
true)); // isKill
}
LiveAddressRegisterMap[LiveAddress] = Reg;
}
}
continue;
}
const TargetRegisterClass *SuperIndirectRegClass =
TII->getSuperIndirectRegClass();
const TargetRegisterClass *IndirectLoadRegClass =
TII->getIndirectAddrLoadRegClass();
unsigned IndirectReg = MRI.createVirtualRegister(SuperIndirectRegClass);
unsigned RegIndex = MI.getOperand(2).getImm();
unsigned Channel = MI.getOperand(3).getImm();
unsigned Address = TII->calculateIndirectAddress(RegIndex, Channel);
if (MI.getOperand(1).getReg() == AMDGPU::INDIRECT_BASE_ADDR) {
// Direct register access
unsigned Reg = LiveAddressRegisterMap[Address];
unsigned AddrReg = IndirectLoadRegClass->getRegister(Address);
if (regHasExplicitDef(MRI, Reg)) {
// If the register we are reading from has an explicit def, then that
// means it was written via a direct register access (i.e. COPY
// or other instruction that doesn't use indirect addressing). In
// this case we know where the value has been stored, so we can just
// issue a copy.
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY),
MI.getOperand(0).getReg())
.addReg(Reg);
} else {
// If the register we are reading has an implicit def, then that
// means it was written by an indirect register access (i.e. An
// instruction that uses indirect addressing.
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY),
MI.getOperand(0).getReg())
.addReg(AddrReg);
}
} else {
// Indirect register access
// Note on REQ_SEQUENCE instructons: You can't actually use the register
// it defines unless you have an instruction that takes the defined
// register class as an operand.
MachineInstrBuilder Sequence = BuildMI(MBB, I, MBB.findDebugLoc(I),
TII->get(AMDGPU::REG_SEQUENCE),
IndirectReg);
for (int i = IndirectBegin; i <= IndirectEnd; ++i) {
unsigned Addr = TII->calculateIndirectAddress(i, Channel);
if (LiveAddressRegisterMap.find(Addr) == LiveAddressRegisterMap.end()) {
continue;
}
unsigned Reg = LiveAddressRegisterMap[Addr];
// We only need to use REG_SEQUENCE for explicit defs, since the
// register coalescer won't do anything with the implicit defs.
MachineInstr *DefInstr = MRI.getVRegDef(Reg);
if (!DefInstr->getOperand(0).isReg() ||
DefInstr->getOperand(0).getReg() != Reg) {
continue;
}
// Insert a REQ_SEQUENCE instruction to force the register allocator
// to allocate the virtual register to the correct physical register.
Sequence.addReg(LiveAddressRegisterMap[Addr]);
Sequence.addImm(TII->getRegisterInfo().getIndirectSubReg(Addr));
}
MachineInstrBuilder Mov = TII->buildIndirectRead(BB, I,
MI.getOperand(0).getReg(), // Value
Address,
MI.getOperand(1).getReg()); // Offset
Mov.addReg(IndirectReg, RegState::Implicit | RegState::Kill);
}
MI.eraseFromParent();
}
}
return false;
}
bool AMDGPUIndirectAddressingPass::regHasExplicitDef(MachineRegisterInfo &MRI,
unsigned Reg) const {
MachineInstr *DefInstr = MRI.getVRegDef(Reg);
return DefInstr && DefInstr->getOperand(0).isReg() &&
DefInstr->getOperand(0).getReg() == Reg;
}