From 740984b30be923e0c171a52fe357a05016fe08c0 Mon Sep 17 00:00:00 2001 From: Nilay Vaish Date: Sat, 5 Sep 2015 09:35:39 -0500 Subject: ruby: call setMRU from L1 controllers, not from sequencer Currently the sequencer calls the function setMRU that updates the replacement policy structures with the first level caches. While functionally this is correct, the problem is that this requires calling findTagInSet() which is an expensive function. This patch removes the calls to setMRU from the sequencer. All controllers should now update the replacement policy on their own. The set and the way index for a given cache entry can be found within the AbstractCacheEntry structure. Use these indicies to update the replacement policy structures. --- src/mem/protocol/MESI_Three_Level-L0cache.sm | 28 ++++++++++++++++---- src/mem/protocol/MESI_Two_Level-L1cache.sm | 29 +++++++++++++------- src/mem/protocol/MI_example-cache.sm | 4 +++ src/mem/protocol/MOESI_CMP_directory-L1cache.sm | 17 ++++++++++-- src/mem/protocol/MOESI_CMP_token-L1cache.sm | 25 ++++++++++++++---- src/mem/protocol/MOESI_hammer-cache.sm | 35 +++++++++++++++++-------- src/mem/protocol/RubySlicc_Types.sm | 1 + src/mem/ruby/structures/CacheMemory.cc | 8 ++++++ src/mem/ruby/structures/CacheMemory.hh | 2 ++ src/mem/ruby/system/Sequencer.cc | 12 +++------ 10 files changed, 121 insertions(+), 40 deletions(-) (limited to 'src') diff --git a/src/mem/protocol/MESI_Three_Level-L0cache.sm b/src/mem/protocol/MESI_Three_Level-L0cache.sm index 4a0766ce8..7e8626dc9 100644 --- a/src/mem/protocol/MESI_Three_Level-L0cache.sm +++ b/src/mem/protocol/MESI_Three_Level-L0cache.sm @@ -459,21 +459,38 @@ machine(L0Cache, "MESI Directory L0 Cache") } } - action(h_load_hit, "h", desc="If not prefetch, notify sequencer the load completed.") { + action(h_load_hit, "hd", desc="If not prefetch, notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Dcache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk); } - action(hx_load_hit, "hx", desc="If not prefetch, notify sequencer the load completed.") { + action(h_ifetch_hit, "hi", desc="If not prefetch, notify sequencer the ifetch completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Icache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk); + } + + action(hx_load_hit, "hxd", desc="notify sequencer the load completed.") { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Dcache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk, true); + } + + action(hx_ifetch_hit, "hxi", desc="notify sequencer the ifetch completed.") { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Icache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk, true); } action(hh_store_hit, "\h", desc="If not prefetch, notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk); cache_entry.Dirty := true; } @@ -481,6 +498,7 @@ machine(L0Cache, "MESI Directory L0 Cache") action(hhx_store_hit, "\hx", desc="If not prefetch, notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk, true); cache_entry.Dirty := true; } @@ -625,7 +643,7 @@ machine(L0Cache, "MESI Directory L0 Cache") } transition({S,E,M}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } @@ -712,7 +730,7 @@ machine(L0Cache, "MESI Directory L0 Cache") transition(Inst_IS, Data, S) { u_writeInstToCache; - hx_load_hit; + hx_ifetch_hit; s_deallocateTBE; o_popIncomingResponseQueue; kd_wakeUpDependents; @@ -720,7 +738,7 @@ machine(L0Cache, "MESI Directory L0 Cache") transition(Inst_IS, Data_Exclusive, E) { u_writeInstToCache; - hx_load_hit; + hx_ifetch_hit; s_deallocateTBE; o_popIncomingResponseQueue; kd_wakeUpDependents; diff --git a/src/mem/protocol/MESI_Two_Level-L1cache.sm b/src/mem/protocol/MESI_Two_Level-L1cache.sm index 8033e5983..b9be4663f 100644 --- a/src/mem/protocol/MESI_Two_Level-L1cache.sm +++ b/src/mem/protocol/MESI_Two_Level-L1cache.sm @@ -810,36 +810,47 @@ machine(L1Cache, "MESI Directory L1 Cache CMP") sequencer.invalidateSC(address); } - action(h_load_hit, "h", - desc="If not prefetch, notify sequencer the load completed.") + action(h_load_hit, "hd", + desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk); } - action(hx_load_hit, "hx", - desc="If not prefetch, notify sequencer the load completed.") + action(h_ifetch_hit, "hi", desc="Notify sequencer the instruction fetch completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk); + } + + action(hx_load_hit, "hx", desc="Notify sequencer the load completed.") + { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.readCallback(address, cache_entry.DataBlk, true); } - action(hh_store_hit, "\h", - desc="If not prefetch, notify sequencer that store completed.") + action(hh_store_hit, "\h", desc="Notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk); cache_entry.Dirty := true; } - action(hhx_store_hit, "\hx", - desc="If not prefetch, notify sequencer that store completed.") + action(hhx_store_hit, "\hx", desc="Notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.writeCallback(address, cache_entry.DataBlk, true); cache_entry.Dirty := true; } @@ -1081,7 +1092,7 @@ machine(L1Cache, "MESI Directory L1 Cache CMP") } transition({S,E,M}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } diff --git a/src/mem/protocol/MI_example-cache.sm b/src/mem/protocol/MI_example-cache.sm index f3b1600f9..334106615 100644 --- a/src/mem/protocol/MI_example-cache.sm +++ b/src/mem/protocol/MI_example-cache.sm @@ -352,6 +352,7 @@ machine(L1Cache, "MI Example L1 Cache") action(r_load_hit, "r", desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc,"%s\n", cache_entry.DataBlk); + cacheMemory.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk, false); } @@ -359,6 +360,7 @@ machine(L1Cache, "MI Example L1 Cache") peek(responseNetwork_in, ResponseMsg) { assert(is_valid(cache_entry)); DPRINTF(RubySlicc,"%s\n", cache_entry.DataBlk); + cacheMemory.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk, true, machineIDToMachineType(in_msg.Sender)); } @@ -367,6 +369,7 @@ machine(L1Cache, "MI Example L1 Cache") action(s_store_hit, "s", desc="Notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc,"%s\n", cache_entry.DataBlk); + cacheMemory.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk, false); } @@ -374,6 +377,7 @@ machine(L1Cache, "MI Example L1 Cache") peek(responseNetwork_in, ResponseMsg) { assert(is_valid(cache_entry)); DPRINTF(RubySlicc,"%s\n", cache_entry.DataBlk); + cacheMemory.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk, true, machineIDToMachineType(in_msg.Sender)); } diff --git a/src/mem/protocol/MOESI_CMP_directory-L1cache.sm b/src/mem/protocol/MOESI_CMP_directory-L1cache.sm index 6c5d3a20f..2ef80efd2 100644 --- a/src/mem/protocol/MOESI_CMP_directory-L1cache.sm +++ b/src/mem/protocol/MOESI_CMP_directory-L1cache.sm @@ -635,21 +635,32 @@ machine(L1Cache, "Directory protocol") } } - action(h_load_hit, "h", desc="Notify sequencer the load completed.") { + action(h_load_hit, "hd", desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk); + } + + action(h_ifetch_hit, "hi", desc="Notify the sequencer about ifetch completion.") { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk); } action(hx_load_hit, "hx", desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.readCallback(address, cache_entry.DataBlk, true); } action(hh_store_hit, "\h", desc="Notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk); cache_entry.Dirty := true; } @@ -657,6 +668,8 @@ machine(L1Cache, "Directory protocol") action(xx_store_hit, "\xx", desc="Notify sequencer that store completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.writeCallback(address, cache_entry.DataBlk, true); cache_entry.Dirty := true; } @@ -964,7 +977,7 @@ machine(L1Cache, "Directory protocol") } transition({S, SM, O, OM, MM, MM_W, M, M_W}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } diff --git a/src/mem/protocol/MOESI_CMP_token-L1cache.sm b/src/mem/protocol/MOESI_CMP_token-L1cache.sm index c5a7cd940..230adfc4b 100644 --- a/src/mem/protocol/MOESI_CMP_token-L1cache.sm +++ b/src/mem/protocol/MOESI_CMP_token-L1cache.sm @@ -1282,12 +1282,22 @@ machine(L1Cache, "Token protocol") } } + action(h_load_hit, "hd", desc="Notify sequencer the load completed.") { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "Address: %s, Data Block: %s\n", + address, cache_entry.DataBlk); - action(h_load_hit, "h", desc="Notify sequencer the load completed.") { + L1Dcache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk, false, + MachineType:L1Cache); + } + + action(h_ifetch_hit, "hi", desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "Address: %s, Data Block: %s\n", address, cache_entry.DataBlk); + L1Icache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk, false, MachineType:L1Cache); } @@ -1297,6 +1307,8 @@ machine(L1Cache, "Token protocol") DPRINTF(RubySlicc, "Address: %s, Data Block: %s\n", address, cache_entry.DataBlk); peek(responseNetwork_in, ResponseMsg) { + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.readCallback(address, cache_entry.DataBlk, isExternalHit(address, in_msg.Sender), machineIDToMachineType(in_msg.Sender)); @@ -1308,6 +1320,7 @@ machine(L1Cache, "Token protocol") DPRINTF(RubySlicc, "Address: %s, Data Block: %s\n", address, cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk, false, MachineType:L1Cache); cache_entry.Dirty := true; @@ -1319,6 +1332,8 @@ machine(L1Cache, "Token protocol") DPRINTF(RubySlicc, "Address: %s, Data Block: %s\n", address, cache_entry.DataBlk); peek(responseNetwork_in, ResponseMsg) { + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.writeCallback(address, cache_entry.DataBlk, isExternalHit(address, in_msg.Sender), machineIDToMachineType(in_msg.Sender)); @@ -1700,7 +1715,7 @@ machine(L1Cache, "Token protocol") } transition({S, SM, S_L, SM_L}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } @@ -1782,7 +1797,7 @@ machine(L1Cache, "Token protocol") // Transitions from Owned transition({O, OM}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } @@ -1872,7 +1887,7 @@ machine(L1Cache, "Token protocol") // Transitions from Modified transition({MM, MM_W}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } @@ -1947,7 +1962,7 @@ machine(L1Cache, "Token protocol") // Transitions from Dirty Exclusive transition({M, M_W}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileInstHit; k_popMandatoryQueue; } diff --git a/src/mem/protocol/MOESI_hammer-cache.sm b/src/mem/protocol/MOESI_hammer-cache.sm index 303bf1784..88b7308ed 100644 --- a/src/mem/protocol/MOESI_hammer-cache.sm +++ b/src/mem/protocol/MOESI_hammer-cache.sm @@ -853,9 +853,18 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } } - action(h_load_hit, "h", desc="Notify sequencer the load completed.") { + action(h_load_hit, "hd", desc="Notify sequencer the load completed.") { assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Dcache.setMRU(cache_entry); + sequencer.readCallback(address, cache_entry.DataBlk, false, + testAndClearLocalHit(cache_entry)); + } + + action(h_ifetch_hit, "hi", desc="Notify sequencer the ifetch completed.") { + assert(is_valid(cache_entry)); + DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); + L1Icache.setMRU(cache_entry); sequencer.readCallback(address, cache_entry.DataBlk, false, testAndClearLocalHit(cache_entry)); } @@ -865,7 +874,8 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") assert(is_valid(tbe)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); peek(responseToCache_in, ResponseMsg) { - + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.readCallback(address, cache_entry.DataBlk, true, machineIDToMachineType(in_msg.Sender), tbe.InitialRequestTime, tbe.ForwardRequestTime, tbe.FirstResponseTime); @@ -876,6 +886,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") assert(is_valid(cache_entry)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); peek(mandatoryQueue_in, RubyRequest) { + L1Dcache.setMRU(cache_entry); sequencer.writeCallback(address, cache_entry.DataBlk, false, testAndClearLocalHit(cache_entry)); @@ -897,7 +908,8 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") assert(is_valid(tbe)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); peek(responseToCache_in, ResponseMsg) { - + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.writeCallback(address, cache_entry.DataBlk, true, machineIDToMachineType(in_msg.Sender), tbe.InitialRequestTime, tbe.ForwardRequestTime, tbe.FirstResponseTime); @@ -910,7 +922,8 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") assert(is_valid(cache_entry)); assert(is_valid(tbe)); DPRINTF(RubySlicc, "%s\n", cache_entry.DataBlk); - + L1Icache.setMRU(address); + L1Dcache.setMRU(address); sequencer.writeCallback(address, cache_entry.DataBlk, true, machineIDToMachineType(tbe.LastResponder), tbe.InitialRequestTime, tbe.ForwardRequestTime, tbe.FirstResponseTime); @@ -1504,7 +1517,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition({S, SM, ISM}, Ifetch) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstHit; k_popMandatoryQueue; } @@ -1518,7 +1531,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition(SR, Ifetch, S) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstMiss; uu_profileL2Hit; k_popMandatoryQueue; @@ -1566,7 +1579,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition({O, OM, SS, MM_W, M_W}, {Ifetch}) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstHit; k_popMandatoryQueue; } @@ -1580,7 +1593,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition(OR, Ifetch, O) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstMiss; uu_profileL2Hit; k_popMandatoryQueue; @@ -1631,7 +1644,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") // Transitions from Modified transition({MM, M}, {Ifetch}) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstHit; k_popMandatoryQueue; } @@ -1657,7 +1670,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition(MMR, Ifetch, MM) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstMiss; uu_profileL2Hit; k_popMandatoryQueue; @@ -1738,7 +1751,7 @@ machine({L1Cache, L2Cache}, "AMD Hammer-like protocol") } transition(MR, Ifetch, M) { - h_load_hit; + h_ifetch_hit; uu_profileL1InstMiss; uu_profileL2Hit; k_popMandatoryQueue; diff --git a/src/mem/protocol/RubySlicc_Types.sm b/src/mem/protocol/RubySlicc_Types.sm index d032adfd8..f464b3c7d 100644 --- a/src/mem/protocol/RubySlicc_Types.sm +++ b/src/mem/protocol/RubySlicc_Types.sm @@ -156,6 +156,7 @@ structure (CacheMemory, external = "yes") { Cycles getTagLatency(); Cycles getDataLatency(); void setMRU(Addr); + void setMRU(AbstractCacheEntry); void recordRequestType(CacheRequestType, Addr); bool checkResourceAvailable(CacheResourceType, Addr); diff --git a/src/mem/ruby/structures/CacheMemory.cc b/src/mem/ruby/structures/CacheMemory.cc index ab2647759..931f58a8e 100644 --- a/src/mem/ruby/structures/CacheMemory.cc +++ b/src/mem/ruby/structures/CacheMemory.cc @@ -343,6 +343,14 @@ CacheMemory::setMRU(Addr address) m_replacementPolicy_ptr->touch(cacheSet, loc, curTick()); } +void +CacheMemory::setMRU(const AbstractCacheEntry *e) +{ + uint32_t cacheSet = e->getSetIndex(); + uint32_t loc = e->getWayIndex(); + m_replacementPolicy_ptr->touch(cacheSet, loc, curTick()); +} + void CacheMemory::recordCacheContents(int cntrl, CacheRecorder* tr) const { diff --git a/src/mem/ruby/structures/CacheMemory.hh b/src/mem/ruby/structures/CacheMemory.hh index 1af446950..7ce674e61 100644 --- a/src/mem/ruby/structures/CacheMemory.hh +++ b/src/mem/ruby/structures/CacheMemory.hh @@ -106,6 +106,8 @@ class CacheMemory : public SimObject // Set this address to most recently used void setMRU(Addr address); + // Set this entry to most recently used + void setMRU(const AbstractCacheEntry *e); // Functions for locking and unlocking cache lines corresponding to the // provided address. These are required for supporting atomic memory diff --git a/src/mem/ruby/system/Sequencer.cc b/src/mem/ruby/system/Sequencer.cc index b21c70743..740db7d8d 100644 --- a/src/mem/ruby/system/Sequencer.cc +++ b/src/mem/ruby/system/Sequencer.cc @@ -496,19 +496,15 @@ Sequencer::hitCallback(SequencerRequest* srequest, DataBlock& data, const Cycles forwardRequestTime, const Cycles firstResponseTime) { + warn_once("Replacement policy updates recently became the responsibility " + "of SLICC state machines. Make sure to setMRU() near callbacks " + "in .sm files!"); + PacketPtr pkt = srequest->pkt; Addr request_address(pkt->getAddr()); - Addr request_line_address = makeLineAddress(pkt->getAddr()); RubyRequestType type = srequest->m_type; Cycles issued_time = srequest->issue_time; - // Set this cache entry to the most recently used - if (type == RubyRequestType_IFETCH) { - m_instCache_ptr->setMRU(request_line_address); - } else { - m_dataCache_ptr->setMRU(request_line_address); - } - assert(curCycle() >= issued_time); Cycles total_latency = curCycle() - issued_time; -- cgit v1.2.3