diff options
Diffstat (limited to 'src/cpu')
-rw-r--r-- | src/cpu/base_dyn_inst.hh | 5 | ||||
-rw-r--r-- | src/cpu/o3/lsq_unit.hh | 33 | ||||
-rw-r--r-- | src/cpu/o3/lsq_unit_impl.hh | 22 | ||||
-rw-r--r-- | src/cpu/o3/regfile.hh | 14 | ||||
-rw-r--r-- | src/cpu/o3/rename_impl.hh | 10 |
5 files changed, 43 insertions, 41 deletions
diff --git a/src/cpu/base_dyn_inst.hh b/src/cpu/base_dyn_inst.hh index 6c6d90076..eed05c2f1 100644 --- a/src/cpu/base_dyn_inst.hh +++ b/src/cpu/base_dyn_inst.hh @@ -877,6 +877,11 @@ BaseDynInst<Impl>::write(T data, Addr addr, unsigned flags, uint64_t *res) effAddrValid = true; physEffAddr = req->getPaddr(); memReqFlags = req->getFlags(); + + if (req->isCondSwap()) { + assert(res); + req->setExtraData(*res); + } #if 0 if (cpu->system->memctrl->badaddr(physEffAddr)) { fault = TheISA::genMachineCheckFault(); diff --git a/src/cpu/o3/lsq_unit.hh b/src/cpu/o3/lsq_unit.hh index e1b27048d..f24de20d9 100644 --- a/src/cpu/o3/lsq_unit.hh +++ b/src/cpu/o3/lsq_unit.hh @@ -289,15 +289,19 @@ class LSQUnit { struct SQEntry { /** Constructs an empty store queue entry. */ SQEntry() - : inst(NULL), req(NULL), size(0), data(0), + : inst(NULL), req(NULL), size(0), canWB(0), committed(0), completed(0) - { } + { + bzero(data, sizeof(data)); + } /** Constructs a store queue entry for a given instruction. */ SQEntry(DynInstPtr &_inst) - : inst(_inst), req(NULL), size(0), data(0), + : inst(_inst), req(NULL), size(0), canWB(0), committed(0), completed(0) - { } + { + bzero(data, sizeof(data)); + } /** The store instruction. */ DynInstPtr inst; @@ -306,7 +310,7 @@ class LSQUnit { /** The size of the store. */ int size; /** The store data. */ - IntReg data; + char data[sizeof(IntReg)]; /** Whether or not the store can writeback. */ bool canWB; /** Whether or not the store is committed. */ @@ -554,22 +558,14 @@ LSQUnit<Impl>::read(Request *req, T &data, int load_idx) if ((store_has_lower_limit && store_has_upper_limit)) { // Get shift amount for offset into the store's data. int shift_amt = req->getVaddr() & (store_size - 1); - // @todo: Magic number, assumes byte addressing - shift_amt = shift_amt << 3; - - // Cast this to type T? - data = storeQueue[store_idx].data >> shift_amt; - // When the data comes from the store queue entry, it's in host - // order. When it gets sent to the load, it needs to be in guest - // order so when the load converts it again, it ends up back - // in host order like the inst expects. - data = TheISA::htog(data); + memcpy(&data, storeQueue[store_idx].data + shift_amt, sizeof(T)); assert(!load_inst->memData); load_inst->memData = new uint8_t[64]; - memcpy(load_inst->memData, &data, req->getSize()); + memcpy(load_inst->memData, + storeQueue[store_idx].data + shift_amt, req->getSize()); DPRINTF(LSQUnit, "Forwarding from store idx %i to load to " "addr %#x, data %#x\n", @@ -716,7 +712,10 @@ LSQUnit<Impl>::write(Request *req, T &data, int store_idx) storeQueue[store_idx].req = req; storeQueue[store_idx].size = sizeof(T); - storeQueue[store_idx].data = data; + assert(sizeof(T) <= sizeof(storeQueue[store_idx].data)); + + T gData = htog(data); + memcpy(storeQueue[store_idx].data, &gData, sizeof(T)); // This function only writes the data to the store queue, so no fault // can happen here. diff --git a/src/cpu/o3/lsq_unit_impl.hh b/src/cpu/o3/lsq_unit_impl.hh index 2aa0d6b6a..44e2cea76 100644 --- a/src/cpu/o3/lsq_unit_impl.hh +++ b/src/cpu/o3/lsq_unit_impl.hh @@ -645,22 +645,10 @@ LSQUnit<Impl>::writebackStores() assert(!inst->memData); inst->memData = new uint8_t[64]; - TheISA::IntReg convertedData = - TheISA::htog(storeQueue[storeWBIdx].data); - - //FIXME This is a hack to get SPARC working. It, along with endianness - //in the memory system in general, need to be straightened out more - //formally. The problem is that the data's endianness is swapped when - //it's in the 64 bit data field in the store queue. The data that you - //want won't start at the beginning of the field anymore unless it was - //a 64 bit access. - memcpy(inst->memData, - (uint8_t *)&convertedData + - (TheISA::ByteOrderDiffers ? - (sizeof(TheISA::IntReg) - req->getSize()) : 0), - req->getSize()); - - PacketPtr data_pkt = new Packet(req, MemCmd::WriteReq, + memcpy(inst->memData, storeQueue[storeWBIdx].data, req->getSize()); + + MemCmd command = req->isSwap() ? MemCmd::SwapReq : MemCmd::WriteReq; + PacketPtr data_pkt = new Packet(req, command, Packet::Broadcast); data_pkt->dataStatic(inst->memData); @@ -677,7 +665,7 @@ LSQUnit<Impl>::writebackStores() inst->seqNum); // @todo: Remove this SC hack once the memory system handles it. - if (req->isLocked()) { + if (inst->isStoreConditional()) { // Disable recording the result temporarily. Writing to // misc regs normally updates the result, but this is not // the desired behavior when handling store conditionals. diff --git a/src/cpu/o3/regfile.hh b/src/cpu/o3/regfile.hh index bbc69fc96..b5b1cd021 100644 --- a/src/cpu/o3/regfile.hh +++ b/src/cpu/o3/regfile.hh @@ -174,7 +174,7 @@ class PhysRegFile // Remove the base Float reg dependency. reg_idx = reg_idx - numPhysicalIntRegs; - assert(reg_idx < numPhysicalFloatRegs + numPhysicalIntRegs); + assert(reg_idx < numPhysicalFloatRegs); DPRINTF(IEW, "RegFile: Setting float register %i to %#x\n", int(reg_idx), (uint64_t)val); @@ -189,7 +189,7 @@ class PhysRegFile // Remove the base Float reg dependency. reg_idx = reg_idx - numPhysicalIntRegs; - assert(reg_idx < numPhysicalFloatRegs + numPhysicalIntRegs); + assert(reg_idx < numPhysicalFloatRegs); DPRINTF(IEW, "RegFile: Setting float register %i to %#x\n", int(reg_idx), (uint64_t)val); @@ -204,7 +204,7 @@ class PhysRegFile // Remove the base Float reg dependency. reg_idx = reg_idx - numPhysicalIntRegs; - assert(reg_idx < numPhysicalFloatRegs + numPhysicalIntRegs); + assert(reg_idx < numPhysicalFloatRegs); DPRINTF(IEW, "RegFile: Setting float register %i to %#x\n", int(reg_idx), (uint64_t)val); @@ -217,7 +217,7 @@ class PhysRegFile // Remove the base Float reg dependency. reg_idx = reg_idx - numPhysicalIntRegs; - assert(reg_idx < numPhysicalFloatRegs + numPhysicalIntRegs); + assert(reg_idx < numPhysicalFloatRegs); DPRINTF(IEW, "RegFile: Setting float register %i to %#x\n", int(reg_idx), (uint64_t)val); @@ -232,11 +232,11 @@ class PhysRegFile MiscReg readMiscReg(int misc_reg, unsigned thread_id) { - return miscRegs[thread_id].readReg(misc_reg, - cpu->tcBase(thread_id)); + return miscRegs[thread_id].readReg(misc_reg, cpu->tcBase(thread_id)); } - void setMiscRegNoEffect(int misc_reg, const MiscReg &val, unsigned thread_id) + void setMiscRegNoEffect(int misc_reg, + const MiscReg &val, unsigned thread_id) { miscRegs[thread_id].setRegNoEffect(misc_reg, val); } diff --git a/src/cpu/o3/rename_impl.hh b/src/cpu/o3/rename_impl.hh index ec630b31e..431705e19 100644 --- a/src/cpu/o3/rename_impl.hh +++ b/src/cpu/o3/rename_impl.hh @@ -996,7 +996,12 @@ DefaultRename<Impl>::renameSrcRegs(DynInstPtr &inst,unsigned tid) if (src_reg < TheISA::FP_Base_DepTag) { flat_src_reg = TheISA::flattenIntIndex(inst->tcBase(), src_reg); DPRINTF(Rename, "Flattening index %d to %d.\n", (int)src_reg, (int)flat_src_reg); + } else { + // Floating point and Miscellaneous registers need their indexes + // adjusted to account for the expanded number of flattened int regs. + flat_src_reg = src_reg - TheISA::FP_Base_DepTag + TheISA::NumIntRegs; } + inst->flattenSrcReg(src_idx, flat_src_reg); // Look up the source registers to get the phys. register they've @@ -1033,8 +1038,13 @@ DefaultRename<Impl>::renameDestRegs(DynInstPtr &inst,unsigned tid) RegIndex dest_reg = inst->destRegIdx(dest_idx); RegIndex flat_dest_reg = dest_reg; if (dest_reg < TheISA::FP_Base_DepTag) { + // Integer registers are flattened. flat_dest_reg = TheISA::flattenIntIndex(inst->tcBase(), dest_reg); DPRINTF(Rename, "Flattening index %d to %d.\n", (int)dest_reg, (int)flat_dest_reg); + } else { + // Floating point and Miscellaneous registers need their indexes + // adjusted to account for the expanded number of flattened int regs. + flat_dest_reg = dest_reg - TheISA::FP_Base_DepTag + TheISA::NumIntRegs; } inst->flattenDestReg(dest_idx, flat_dest_reg); |