gem5/cpu/beta_cpu/rename_impl.hh
Kevin Lim e3fb9afa79 Update to make multiple instruction issue and different latencies work.
Also change to ref counted DynInst.

SConscript:
    Add branch predictor, BTB, load store queue, and storesets.
arch/isa_parser.py:
    Specify the template parameter for AlphaDynInst
base/traceflags.py:
    Add load store queue, store set, and mem dependence unit to the
    list of trace flags.
cpu/base_dyn_inst.cc:
    Change formating, add in debug statement.
cpu/base_dyn_inst.hh:
    Change DynInst to be RefCounted, add flag to clear whether or not this
    instruction can commit.  This is likely to be removed in the future.
cpu/beta_cpu/alpha_dyn_inst.cc:
    AlphaDynInst has been changed to be templated, so now this CC file
    is just used to force instantiations of AlphaDynInst.
cpu/beta_cpu/alpha_dyn_inst.hh:
    Changed AlphaDynInst to be templated on Impl.  Removed some unnecessary
    functions.
cpu/beta_cpu/alpha_full_cpu.cc:
    AlphaFullCPU has been changed to be templated, so this CC file is now
    just used to force instantation of AlphaFullCPU.
cpu/beta_cpu/alpha_full_cpu.hh:
    Change AlphaFullCPU to be templated on Impl.
cpu/beta_cpu/alpha_impl.hh:
    Update it to reflect AlphaDynInst and AlphaFullCPU being templated
    on Impl.  Also removed time buffers from here, as they are really
    a part of the CPU and are thus in the CPU policy now.
cpu/beta_cpu/alpha_params.hh:
    Make AlphaSimpleParams inherit from the BaseFullCPU so that it doesn't
    need to specifically declare any parameters that are already in the
    BaseFullCPU.
cpu/beta_cpu/comm.hh:
    Changed the structure of the time buffer communication structs.  Now
    they include the size of the packet of instructions it is sending.
    Added some parameters to the backwards communication struct, mainly
    for squashing.
cpu/beta_cpu/commit.hh:
    Update typenames to reflect change in location of time buffer structs.
    Update DynInst to DynInstPtr (it is refcounted now).
cpu/beta_cpu/commit_impl.hh:
    Formatting changes mainly.  Also sends back proper information
    on branch mispredicts so that the bpred unit can update itself.
    Updated behavior for non-speculative instructions (stores, any
    other non-spec instructions): once they reach the head of the ROB,
    the ROB signals back to the IQ that it can go ahead and issue the
    non-speculative instruction.  The instruction itself is updated so that
    commit won't try to commit it again until it is done executing.
cpu/beta_cpu/cpu_policy.hh:
    Added branch prediction unit, mem dependence prediction unit, load
    store queue.  Moved time buffer structs from AlphaSimpleImpl to here.
cpu/beta_cpu/decode.hh:
    Changed typedefs to reflect change in location of time buffer structs
    and also the change from DynInst to ref counted DynInstPtr.
cpu/beta_cpu/decode_impl.hh:
    Continues to buffer instructions even while unblocking now.  Changed
    how it loops through groups of instructions so it can properly block
    during the middle of a group of instructions.
cpu/beta_cpu/fetch.hh:
    Changed typedefs to reflect change in location of time buffer structs
    and the change to ref counted DynInsts.  Also added in branch
    brediction unit.
cpu/beta_cpu/fetch_impl.hh:
    Add in branch prediction.  Changed how fetch checks inputs and its
    current state to make for easier logic.
cpu/beta_cpu/free_list.cc:
    Changed int regs and float regs to logically use one flat namespace.
    Future change will be moving them to a single scoreboard to conserve
    space.
cpu/beta_cpu/free_list.hh:
    Mostly debugging statements.  Might be removed for performance in future.
cpu/beta_cpu/full_cpu.cc:
    Added in some debugging statements.  Updated BaseFullCPU to take
    a params object.
cpu/beta_cpu/full_cpu.hh:
    Added params class within BaseCPU that other param classes will be
    able to inherit from.  Updated typedefs to reflect change in location
    of time buffer structs and ref counted DynInst.
cpu/beta_cpu/iew.hh:
    Updated typedefs to reflect change in location of time buffer structs
    and use of ref counted DynInsts.
cpu/beta_cpu/iew_impl.hh:
    Added in load store queue, updated iew to be able to execute non-
    speculative instructions, instead of having them execute in commit.
cpu/beta_cpu/inst_queue.hh:
    Updated change to ref counted DynInsts.  Changed inst queue to hold
    non-speculative instructions as well, which are issued only when
    commit signals backwards that a nonspeculative instruction is at
    the head of the ROB.
cpu/beta_cpu/inst_queue_impl.hh:
    Updated to allow for non-speculative instructions to be in the inst
    queue.  Also added some debug functions.
cpu/beta_cpu/regfile.hh:
    Added debugging statements, changed formatting.
cpu/beta_cpu/rename.hh:
    Updated typedefs, added some functions to clean up code.
cpu/beta_cpu/rename_impl.hh:
    Moved some code into functions to make it easier to read.
cpu/beta_cpu/rename_map.cc:
    Changed int and float reg behavior to use a single flat namespace.  In
    the future, the rename maps can be combined to a single rename map to
    save space.
cpu/beta_cpu/rename_map.hh:
    Added destructor.
cpu/beta_cpu/rob.hh:
    Updated it with change from DynInst to ref counted DynInst.
cpu/beta_cpu/rob_impl.hh:
    Formatting, updated to use ref counted DynInst.
cpu/static_inst.hh:
    Updated forward declaration for AlphaDynInst now that it is templated.

--HG--
extra : convert_revision : 1045f240ee9b6a4bd368e1806aca029ebbdc6dd3
2004-09-23 14:06:03 -04:00

644 lines
20 KiB
C++

#include <list>
#include "cpu/beta_cpu/rename.hh"
template <class Impl>
SimpleRename<Impl>::SimpleRename(Params &params)
: iewToRenameDelay(params.iewToRenameDelay),
decodeToRenameDelay(params.decodeToRenameDelay),
commitToRenameDelay(params.commitToRenameDelay),
renameWidth(params.renameWidth),
commitWidth(params.commitWidth),
numInst(0)
{
_status = Idle;
}
template <class Impl>
void
SimpleRename<Impl>::setCPU(FullCPU *cpu_ptr)
{
DPRINTF(Rename, "Rename: Setting CPU pointer.\n");
cpu = cpu_ptr;
}
template <class Impl>
void
SimpleRename<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr)
{
DPRINTF(Rename, "Rename: Setting time buffer pointer.\n");
timeBuffer = tb_ptr;
// Setup wire to read information from time buffer, from IEW stage.
fromIEW = timeBuffer->getWire(-iewToRenameDelay);
// Setup wire to read infromation from time buffer, from commit stage.
fromCommit = timeBuffer->getWire(-commitToRenameDelay);
// Setup wire to write information to previous stages.
toDecode = timeBuffer->getWire(0);
}
template <class Impl>
void
SimpleRename<Impl>::setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr)
{
DPRINTF(Rename, "Rename: Setting rename queue pointer.\n");
renameQueue = rq_ptr;
// Setup wire to write information to future stages.
toIEW = renameQueue->getWire(0);
}
template <class Impl>
void
SimpleRename<Impl>::setDecodeQueue(TimeBuffer<DecodeStruct> *dq_ptr)
{
DPRINTF(Rename, "Rename: Setting decode queue pointer.\n");
decodeQueue = dq_ptr;
// Setup wire to get information from decode.
fromDecode = decodeQueue->getWire(-decodeToRenameDelay);
}
template <class Impl>
void
SimpleRename<Impl>::setRenameMap(RenameMap *rm_ptr)
{
DPRINTF(Rename, "Rename: Setting rename map pointer.\n");
renameMap = rm_ptr;
}
template <class Impl>
void
SimpleRename<Impl>::setFreeList(FreeList *fl_ptr)
{
DPRINTF(Rename, "Rename: Setting free list pointer.\n");
freeList = fl_ptr;
}
template <class Impl>
void
SimpleRename<Impl>::dumpHistory()
{
typename list<RenameHistory>::iterator buf_it = historyBuffer.begin();
while (buf_it != historyBuffer.end())
{
cprintf("Seq num: %i\nArch reg: %i New phys reg: %i Old phys "
"reg: %i\n", (*buf_it).instSeqNum, (int)(*buf_it).archReg,
(int)(*buf_it).newPhysReg, (int)(*buf_it).prevPhysReg);
buf_it++;
}
}
template <class Impl>
void
SimpleRename<Impl>::block()
{
DPRINTF(Rename, "Rename: Blocking.\n");
// Set status to Blocked.
_status = Blocked;
// Add the current inputs onto the skid buffer, so they can be
// reprocessed when this stage unblocks.
skidBuffer.push(*fromDecode);
// Note that this stage only signals previous stages to stall when
// it is the cause of the stall originates at this stage. Otherwise
// the previous stages are expected to check all possible stall signals.
}
template <class Impl>
inline void
SimpleRename<Impl>::unblock()
{
DPRINTF(Rename, "Rename: Read instructions out of skid buffer this "
"cycle.\n");
// Remove the now processed instructions from the skid buffer.
skidBuffer.pop();
// If there's still information in the skid buffer, then
// continue to tell previous stages to stall. They will be
// able to restart once the skid buffer is empty.
if (!skidBuffer.empty()) {
toDecode->renameInfo.stall = true;
} else {
DPRINTF(Rename, "Rename: Done unblocking.\n");
_status = Running;
}
}
template <class Impl>
void
SimpleRename<Impl>::doSquash()
{
typename list<RenameHistory>::iterator hb_it = historyBuffer.begin();
// typename list<RenameHistory>::iterator delete_it;
InstSeqNum squashed_seq_num = fromCommit->commitInfo.doneSeqNum;
#ifdef FULL_SYSTEM
assert(!historyBuffer.empty());
#else
// After a syscall squashes everything, the history buffer may be empty
// but the ROB may still be squashing instructions.
if (historyBuffer.empty()) {
return;
}
#endif // FULL_SYSTEM
// Go through the most recent instructions, undoing the mappings
// they did and freeing up the registers.
while ((*hb_it).instSeqNum > squashed_seq_num)
{
DPRINTF(Rename, "Rename: Removing history entry with sequence "
"number %i.\n", (*hb_it).instSeqNum);
// If it's not simply a place holder, then add the registers.
if (!(*hb_it).placeHolder) {
// Tell the rename map to set the architected register to the
// previous physical register that it was renamed to.
renameMap->setEntry(hb_it->archReg, hb_it->prevPhysReg);
// Put the renamed physical register back on the free list.
freeList->addReg(hb_it->newPhysReg);
}
// delete_it = hb_it;
// hb_it++;
historyBuffer.erase(hb_it++);
assert(hb_it != historyBuffer.end());
}
}
template <class Impl>
void
SimpleRename<Impl>::squash()
{
DPRINTF(Rename, "Rename: Squashing instructions.\n");
// Set the status to Squashing.
_status = Squashing;
numInst = 0;
// Clear the skid buffer in case it has any data in it.
while (!skidBuffer.empty())
{
skidBuffer.pop();
}
doSquash();
}
// In the future, when a SmartPtr is used for DynInst, then this function
// itself can handle returning the instruction's physical registers to
// the free list.
template<class Impl>
void
SimpleRename<Impl>::removeFromHistory(InstSeqNum inst_seq_num)
{
DPRINTF(Rename, "Rename: Removing a committed instruction from the "
"history buffer, until sequence number %lli.\n", inst_seq_num);
typename list<RenameHistory>::iterator hb_it = historyBuffer.end();
--hb_it;
if (hb_it->instSeqNum > inst_seq_num) {
DPRINTF(Rename, "Rename: Old sequence number encountered. Ensure "
"that a syscall happened recently.\n");
return;
}
while ((*hb_it).instSeqNum != inst_seq_num)
{
// Make sure we haven't gone off the end of the list.
assert(hb_it != historyBuffer.end());
// In theory instructions at the end of the history buffer
// should be older than the instruction being removed, which
// means they will have a lower sequence number. Also the
// instruction being removed from the history really should
// be the last instruction in the list, as it is the instruction
// that was just committed that is being removed.
assert(hb_it->instSeqNum < inst_seq_num);
DPRINTF(Rename, "Rename: Freeing up older rename of reg %i, sequence"
" number %i.\n",
(*hb_it).prevPhysReg, (*hb_it).instSeqNum);
if (!(*hb_it).placeHolder) {
freeList->addReg((*hb_it).prevPhysReg);
}
historyBuffer.erase(hb_it--);
}
// Finally free up the previous register of the squashed instruction
// itself.
if (!(*hb_it).placeHolder) {
freeList->addReg(hb_it->prevPhysReg);
}
historyBuffer.erase(hb_it);
}
template <class Impl>
inline void
SimpleRename<Impl>::renameSrcRegs(DynInstPtr &inst)
{
unsigned num_src_regs = inst->numSrcRegs();
// Get the architectual register numbers from the source and
// destination operands, and redirect them to the right register.
// Will need to mark dependencies though.
for (int src_idx = 0; src_idx < num_src_regs; src_idx++)
{
RegIndex src_reg = inst->srcRegIdx(src_idx);
// Look up the source registers to get the phys. register they've
// been renamed to, and set the sources to those registers.
RegIndex renamed_reg = renameMap->lookup(src_reg);
DPRINTF(Rename, "Rename: Looking up arch reg %i, got "
"physical reg %i.\n", (int)src_reg, (int)renamed_reg);
inst->renameSrcReg(src_idx, renamed_reg);
// Either incorporate it into the info passed back,
// or make another function call to see if that register is
// ready or not.
if (renameMap->isReady(renamed_reg)) {
DPRINTF(Rename, "Rename: Register is ready.\n");
inst->markSrcRegReady(src_idx);
}
}
}
template <class Impl>
inline void
SimpleRename<Impl>::renameDestRegs(DynInstPtr &inst)
{
typename SimpleRenameMap::RenameInfo rename_result;
unsigned num_dest_regs = inst->numDestRegs();
// Rename the destination registers.
for (int dest_idx = 0; dest_idx < num_dest_regs; dest_idx++)
{
RegIndex dest_reg = inst->destRegIdx(dest_idx);
// Get the physical register that the destination will be
// renamed to.
rename_result = renameMap->rename(dest_reg);
DPRINTF(Rename, "Rename: Renaming arch reg %i to physical "
"reg %i.\n", (int)dest_reg,
(int)rename_result.first);
// Record the rename information so that a history can be kept.
RenameHistory hb_entry(inst->seqNum, dest_reg,
rename_result.first,
rename_result.second);
historyBuffer.push_front(hb_entry);
DPRINTF(Rename, "Rename: Adding instruction to history buffer, "
"sequence number %lli.\n",
(*historyBuffer.begin()).instSeqNum);
// Tell the instruction to rename the appropriate destination
// register (dest_idx) to the new physical register
// (rename_result.first), and record the previous physical
// register that the same logical register was renamed to
// (rename_result.second).
inst->renameDestReg(dest_idx,
rename_result.first,
rename_result.second);
}
// If it's an instruction with no destination registers, then put
// a placeholder within the history buffer. It might be better
// to not put it in the history buffer at all (other than branches,
// which always need at least a place holder), and differentiate
// between instructions with and without destination registers
// when getting from commit the instructions that committed.
if (num_dest_regs == 0) {
RenameHistory hb_entry(inst->seqNum);
historyBuffer.push_front(hb_entry);
DPRINTF(Rename, "Rename: Adding placeholder instruction to "
"history buffer, sequence number %lli.\n",
inst->seqNum);
}
}
template <class Impl>
inline int
SimpleRename<Impl>::calcFreeROBEntries()
{
return fromCommit->commitInfo.freeROBEntries -
renameWidth * iewToRenameDelay;
}
template <class Impl>
inline int
SimpleRename<Impl>::calcFreeIQEntries()
{
return fromIEW->iewInfo.freeIQEntries - renameWidth * iewToRenameDelay;
}
template<class Impl>
void
SimpleRename<Impl>::tick()
{
// Rename will need to try to rename as many instructions as it
// has bandwidth, unless it is blocked.
// Check if _status is BarrierStall. If so, then check if the number
// of free ROB entries is equal to the number of total ROB entries.
// Once equal then wake this stage up. Set status to unblocking maybe.
if (_status != Blocked && _status != Squashing) {
DPRINTF(Rename, "Rename: Status is not blocked, will attempt to "
"run stage.\n");
// Make sure that the skid buffer has something in it if the
// status is unblocking.
assert(_status == Unblocking ? !skidBuffer.empty() : 1);
rename();
// If the status was unblocking, then instructions from the skid
// buffer were used. Remove those instructions and handle
// the rest of unblocking.
if (_status == Unblocking) {
if (fromDecode->size > 0) {
// Add the current inputs onto the skid buffer, so they can be
// reprocessed when this stage unblocks.
skidBuffer.push(*fromDecode);
}
unblock();
}
} else if (_status == Blocked) {
// If stage is blocked and still receiving valid instructions,
// make sure to store them in the skid buffer.
if (fromDecode->size > 0) {
block();
// Continue to tell previous stage to stall.
toDecode->renameInfo.stall = true;
}
if (!fromIEW->iewInfo.stall &&
!fromCommit->commitInfo.stall &&
calcFreeROBEntries() > 0 &&
calcFreeIQEntries() > 0 &&
renameMap->numFreeEntries() > 0) {
// Need to be sure to check all blocking conditions above.
// If they have cleared, then start unblocking.
DPRINTF(Rename, "Rename: Stall signals cleared, going to "
"unblock.\n");
_status = Unblocking;
// Continue to tell previous stage to block until this stage
// is done unblocking.
toDecode->renameInfo.stall = true;
} else {
// Otherwise no conditions have changed. Tell previous
// stage to continue blocking.
toDecode->renameInfo.stall = true;
}
if (fromCommit->commitInfo.squash ||
fromCommit->commitInfo.robSquashing) {
squash();
return;
}
} else if (_status == Squashing) {
if (fromCommit->commitInfo.squash) {
squash();
} else if (!fromCommit->commitInfo.squash &&
!fromCommit->commitInfo.robSquashing) {
DPRINTF(Rename, "Rename: Done squashing, going to running.\n");
_status = Running;
} else {
doSquash();
}
}
// Ugly code, revamp all of the tick() functions eventually.
if (fromCommit->commitInfo.doneSeqNum != 0 && _status != Squashing) {
removeFromHistory(fromCommit->commitInfo.doneSeqNum);
}
// Perhaps put this outside of this function, since this will
// happen regardless of whether or not the stage is blocked or
// squashing.
// Read from the time buffer any necessary data.
// Read registers that are freed, and add them to the freelist.
// This is unnecessary due to the history buffer (assuming the history
// buffer works properly).
/*
while(!fromCommit->commitInfo.freeRegs.empty())
{
PhysRegIndex freed_reg = fromCommit->commitInfo.freeRegs.back();
DPRINTF(Rename, "Rename: Adding freed register %i to freelist.\n",
(int)freed_reg);
freeList->addReg(freed_reg);
fromCommit->commitInfo.freeRegs.pop_back();
}
*/
}
template<class Impl>
void
SimpleRename<Impl>::rename()
{
// Check if any of the stages ahead of rename are telling rename
// to squash. The squash() function will also take care of fixing up
// the rename map and the free list.
if (fromCommit->commitInfo.squash ||
fromCommit->commitInfo.robSquashing) {
DPRINTF(Rename, "Rename: Receiving signal from Commit to squash.\n");
squash();
return;
}
// Check if time buffer is telling this stage to stall.
if (fromIEW->iewInfo.stall ||
fromCommit->commitInfo.stall) {
DPRINTF(Rename, "Rename: Receiving signal from IEW/Commit to "
"stall.\n");
block();
return;
}
// Check if the current status is squashing. If so, set its status
// to running and resume execution the next cycle.
if (_status == Squashing) {
DPRINTF(Rename, "Rename: Done squashing.\n");
_status = Running;
return;
}
// Check the decode queue to see if instructions are available.
// If there are no available instructions to rename, then do nothing.
// Or, if the stage is currently unblocking, then go ahead and run it.
if (fromDecode->size == 0 && _status != Unblocking) {
DPRINTF(Rename, "Rename: Nothing to do, breaking out early.\n");
// Should I change status to idle?
return;
}
////////////////////////////////////
// Actual rename part.
////////////////////////////////////
DynInstPtr inst;
// If we're unblocking, then we may be in the middle of an instruction
// group. Subtract off numInst to get the proper number of instructions
// left.
int insts_available = _status == Unblocking ?
skidBuffer.front().size - numInst :
fromDecode->size;
bool block_this_cycle = false;
// Will have to do a different calculation for the number of free
// entries. Number of free entries recorded on this cycle -
// renameWidth * renameToDecodeDelay
int free_rob_entries = calcFreeROBEntries();
int free_iq_entries = calcFreeIQEntries();
int min_iq_rob = min(free_rob_entries, free_iq_entries);
unsigned to_iew_index = 0;
// Check if there's any space left.
if (min_iq_rob <= 0) {
DPRINTF(Rename, "Rename: Blocking due to no free ROB or IQ "
"entries.\n"
"Rename: ROB has %d free entries.\n"
"Rename: IQ has %d free entries.\n",
free_rob_entries,
free_iq_entries);
block();
// Tell previous stage to stall.
toDecode->renameInfo.stall = true;
return;
} else if (min_iq_rob < insts_available) {
DPRINTF(Rename, "Rename: Will have to block this cycle. Only "
"%i insts can be renamed due to IQ/ROB limits.\n",
min_iq_rob);
insts_available = min_iq_rob;
block_this_cycle = true;
}
while (insts_available > 0) {
DPRINTF(Rename, "Rename: Sending instructions to iew.\n");
// Get the next instruction either from the skid buffer or the
// decode queue.
inst = _status == Unblocking ? skidBuffer.front().insts[numInst] :
fromDecode->insts[numInst];
if (inst->isSquashed()) {
DPRINTF(Rename, "Rename: instruction %i with PC %#x is "
"squashed, skipping.\n",
inst->seqNum, inst->readPC());
// Go to the next instruction.
++numInst;
// Decrement how many instructions are available.
--insts_available;
continue;
}
DPRINTF(Rename, "Rename: Processing instruction %i with PC %#x.\n",
inst->seqNum, inst->readPC());
// If it's a trap instruction, then it needs to wait here within
// rename until the ROB is empty. Needs a way to detect that the
// ROB is empty. Maybe an event?
// Would be nice if it could be avoided putting this into a
// specific stage and instead just put it into the AlphaFullCPU.
// Might not really be feasible though...
// (EXCB, TRAPB)
if (inst->isSerializing()) {
panic("Rename: Serializing instruction encountered.\n");
DPRINTF(Rename, "Rename: Serializing instruction "
"encountered.\n");
// Change status over to BarrierStall so that other stages know
// what this is blocked on.
_status = BarrierStall;
block_this_cycle = true;
break;
}
// Check here to make sure there are enough destination registers
// to rename to. Otherwise block.
if (renameMap->numFreeEntries() < inst->numDestRegs())
{
DPRINTF(Rename, "Rename: Blocking due to lack of free "
"physical registers to rename to.\n");
// Need some sort of event based on a register being freed.
block_this_cycle = true;
break;
}
renameSrcRegs(inst);
renameDestRegs(inst);
// Put instruction in rename queue.
toIEW->insts[to_iew_index] = inst;
++(toIEW->size);
// Decrease the number of free ROB and IQ entries.
--free_rob_entries;
--free_iq_entries;
// Increment which instruction we're on.
++to_iew_index;
++numInst;
// Decrement how many instructions are available.
--insts_available;
}
// Check if there's any instructions left that haven't yet been renamed.
// If so then block.
if (block_this_cycle) {
block();
toDecode->renameInfo.stall = true;
} else {
// If we had a successful rename and didn't have to exit early, then
// reset numInst so it will refer to the correct instruction on next
// run.
numInst = 0;
}
}