diff options
author | Steve Reinhardt <stever@eecs.umich.edu> | 2006-12-18 20:47:12 -0800 |
---|---|---|
committer | Steve Reinhardt <stever@eecs.umich.edu> | 2006-12-18 20:47:12 -0800 |
commit | 1428b0de7d3d8903e0674d671829219b554d7aa3 (patch) | |
tree | 4e49d43d690c35eb1691c00ebdd07872f318716f /src | |
parent | 0e78386874f15ad28ebc1a0928256e3cf4301008 (diff) | |
download | gem5-1428b0de7d3d8903e0674d671829219b554d7aa3.tar.xz |
Get rid of generic CacheTags object (fold back into Cache).
--HG--
extra : convert_revision : 8769bd8cc358ab3cbbdbbcd909b2e0f1515e09da
Diffstat (limited to 'src')
-rw-r--r-- | src/SConscript | 1 | ||||
-rw-r--r-- | src/mem/cache/cache.cc | 22 | ||||
-rw-r--r-- | src/mem/cache/cache.hh | 181 | ||||
-rw-r--r-- | src/mem/cache/cache_builder.cc | 91 | ||||
-rw-r--r-- | src/mem/cache/cache_impl.hh | 471 | ||||
-rw-r--r-- | src/mem/cache/prefetch/ghb_prefetcher.cc | 4 | ||||
-rw-r--r-- | src/mem/cache/prefetch/stride_prefetcher.cc | 4 |
7 files changed, 683 insertions, 91 deletions
diff --git a/src/SConscript b/src/SConscript index f54e1de0d..3149f0e59 100644 --- a/src/SConscript +++ b/src/SConscript @@ -118,7 +118,6 @@ base_sources = Split(''' mem/cache/prefetch/stride_prefetcher.cc mem/cache/prefetch/tagged_prefetcher.cc mem/cache/tags/base_tags.cc - mem/cache/tags/cache_tags.cc mem/cache/tags/fa_lru.cc mem/cache/tags/iic.cc mem/cache/tags/lru.cc diff --git a/src/mem/cache/cache.cc b/src/mem/cache/cache.cc index 29ca09060..cb4e7f62e 100644 --- a/src/mem/cache/cache.cc +++ b/src/mem/cache/cache.cc @@ -38,8 +38,6 @@ #include "mem/config/cache.hh" -#include "mem/cache/tags/cache_tags.hh" - #if defined(USE_CACHE_LRU) #include "mem/cache/tags/lru.hh" #endif @@ -73,28 +71,28 @@ #if defined(USE_CACHE_FALRU) -template class Cache<CacheTags<FALRU>, SimpleCoherence>; -template class Cache<CacheTags<FALRU>, UniCoherence>; +template class Cache<FALRU, SimpleCoherence>; +template class Cache<FALRU, UniCoherence>; #endif #if defined(USE_CACHE_IIC) -template class Cache<CacheTags<IIC>, SimpleCoherence>; -template class Cache<CacheTags<IIC>, UniCoherence>; +template class Cache<IIC, SimpleCoherence>; +template class Cache<IIC, UniCoherence>; #endif #if defined(USE_CACHE_LRU) -template class Cache<CacheTags<LRU>, SimpleCoherence>; -template class Cache<CacheTags<LRU>, UniCoherence>; +template class Cache<LRU, SimpleCoherence>; +template class Cache<LRU, UniCoherence>; #endif #if defined(USE_CACHE_SPLIT) -template class Cache<CacheTags<Split>, SimpleCoherence>; -template class Cache<CacheTags<Split>, UniCoherence>; +template class Cache<Split, SimpleCoherence>; +template class Cache<Split, UniCoherence>; #endif #if defined(USE_CACHE_SPLIT_LIFO) -template class Cache<CacheTags<SplitLIFO>, SimpleCoherence>; -template class Cache<CacheTags<SplitLIFO>, UniCoherence>; +template class Cache<SplitLIFO, SimpleCoherence>; +template class Cache<SplitLIFO, UniCoherence>; #endif #endif //DOXYGEN_SHOULD_SKIP_THIS diff --git a/src/mem/cache/cache.hh b/src/mem/cache/cache.hh index 097b0f513..bd88849de 100644 --- a/src/mem/cache/cache.hh +++ b/src/mem/cache/cache.hh @@ -38,10 +38,12 @@ #ifndef __CACHE_HH__ #define __CACHE_HH__ +#include "base/compression/base.hh" #include "base/misc.hh" // fatal, panic, and warn #include "cpu/smt.hh" // SMT_MAX_THREADS #include "mem/cache/base_cache.hh" +#include "mem/cache/cache_blk.hh" #include "mem/cache/miss/miss_buffer.hh" #include "mem/cache/prefetch/prefetcher.hh" @@ -62,6 +64,8 @@ class Cache : public BaseCache public: /** Define the type of cache block to use. */ typedef typename TagStore::BlkType BlkType; + /** A typedef for a list of BlkType pointers. */ + typedef typename TagStore::BlkList BlkList; bool prefetchAccess; @@ -141,6 +145,156 @@ class Cache : public BaseCache PacketPtr invalidatePkt; Request *invalidateReq; + /** + * Policy class for performing compression. + */ + CompressionAlgorithm *compressionAlg; + + /** + * The block size of this cache. Set to value in the Tags object. + */ + const int16_t blkSize; + + /** + * Can this cache should allocate a block on a line-sized write miss. + */ + const bool doFastWrites; + + const bool prefetchMiss; + + /** + * Can the data can be stored in a compressed form. + */ + const bool storeCompressed; + + /** + * Do we need to compress blocks on writebacks (i.e. because + * writeback bus is compressed but storage is not)? + */ + const bool compressOnWriteback; + + /** + * The latency of a compression operation. + */ + const int16_t compLatency; + + /** + * Should we use an adaptive compression scheme. + */ + const bool adaptiveCompression; + + /** + * Do writebacks need to be compressed (i.e. because writeback bus + * is compressed), whether or not they're already compressed for + * storage. + */ + const bool writebackCompressed; + + /** + * Compare the internal block data to the fast access block data. + * @param blk The cache block to check. + * @return True if the data is the same. + */ + bool verifyData(BlkType *blk); + + /** + * Update the internal data of the block. The data to write is assumed to + * be in the fast access data. + * @param blk The block with the data to update. + * @param writebacks A list to store any generated writebacks. + * @param compress_block True if we should compress this block + */ + void updateData(BlkType *blk, PacketList &writebacks, bool compress_block); + + /** + * Handle a replacement for the given request. + * @param blk A pointer to the block, usually NULL + * @param pkt The memory request to satisfy. + * @param new_state The new state of the block. + * @param writebacks A list to store any generated writebacks. + */ + BlkType* doReplacement(BlkType *blk, PacketPtr &pkt, + CacheBlk::State new_state, PacketList &writebacks); + + /** + * Does all the processing necessary to perform the provided request. + * @param pkt The memory request to perform. + * @param lat The latency of the access. + * @param writebacks List for any writebacks that need to be performed. + * @param update True if the replacement data should be updated. + * @return Pointer to the cache block touched by the request. NULL if it + * was a miss. + */ + BlkType* handleAccess(PacketPtr &pkt, int & lat, + PacketList & writebacks, bool update = true); + + /** + * Populates a cache block and handles all outstanding requests for the + * satisfied fill request. This version takes an MSHR pointer and uses its + * request to fill the cache block, while repsonding to its targets. + * @param blk The cache block if it already exists. + * @param mshr The MSHR that contains the fill data and targets to satisfy. + * @param new_state The state of the new cache block. + * @param writebacks List for any writebacks that need to be performed. + * @return Pointer to the new cache block. + */ + BlkType* handleFill(BlkType *blk, MSHR * mshr, CacheBlk::State new_state, + PacketList & writebacks, PacketPtr pkt); + + /** + * Populates a cache block and handles all outstanding requests for the + * satisfied fill request. This version takes two memory requests. One + * contains the fill data, the other is an optional target to satisfy. + * Used for Cache::probe. + * @param blk The cache block if it already exists. + * @param pkt The memory request with the fill data. + * @param new_state The state of the new cache block. + * @param writebacks List for any writebacks that need to be performed. + * @param target The memory request to perform after the fill. + * @return Pointer to the new cache block. + */ + BlkType* handleFill(BlkType *blk, PacketPtr &pkt, + CacheBlk::State new_state, + PacketList & writebacks, PacketPtr target = NULL); + + /** + * Sets the blk to the new state and handles the given request. + * @param blk The cache block being snooped. + * @param new_state The new coherence state for the block. + * @param pkt The request to satisfy + */ + void handleSnoop(BlkType *blk, CacheBlk::State new_state, + PacketPtr &pkt); + + /** + * Sets the blk to the new state. + * @param blk The cache block being snooped. + * @param new_state The new coherence state for the block. + */ + void handleSnoop(BlkType *blk, CacheBlk::State new_state); + + /** + * Create a writeback request for the given block. + * @param blk The block to writeback. + * @return The writeback request for the block. + */ + PacketPtr writebackBlk(BlkType *blk); + + BlkType* findBlock(Addr addr) + { + return tags->findBlock(addr); + } + + BlkType* findBlock(PacketPtr &pkt) + { + return tags->findBlock(pkt->getAddr()); + } + + void invalidateBlk(CacheBlk *blk) + { + tags->invalidateBlk(tags->regenerateBlkAddr(blk->tag, blk->set)); + } + public: class Params @@ -153,15 +307,38 @@ class Cache : public BaseCache Prefetcher<TagStore> *prefetcher; bool prefetchAccess; int hitLatency; + CompressionAlgorithm *compressionAlg; + const int16_t blkSize; + const bool doFastWrites; + const bool prefetchMiss; + const bool storeCompressed; + const bool compressOnWriteback; + const int16_t compLatency; + const bool adaptiveCompression; + const bool writebackCompressed; Params(TagStore *_tags, MissBuffer *mq, Coherence *coh, BaseCache::Params params, Prefetcher<TagStore> *_prefetcher, - bool prefetch_access, int hit_latency) + bool prefetch_access, int hit_latency, + bool do_fast_writes, + bool store_compressed, bool adaptive_compression, + bool writeback_compressed, + CompressionAlgorithm *_compressionAlg, int comp_latency, + bool prefetch_miss) : tags(_tags), missQueue(mq), coherence(coh), baseParams(params), prefetcher(_prefetcher), prefetchAccess(prefetch_access), - hitLatency(hit_latency) + hitLatency(hit_latency), + compressionAlg(_compressionAlg), + blkSize(_tags->getBlockSize()), + doFastWrites(do_fast_writes), + prefetchMiss(prefetch_miss), + storeCompressed(store_compressed), + compressOnWriteback(!store_compressed && writeback_compressed), + compLatency(comp_latency), + adaptiveCompression(adaptive_compression), + writebackCompressed(writeback_compressed) { } }; diff --git a/src/mem/cache/cache_builder.cc b/src/mem/cache/cache_builder.cc index 4b1b19718..e212650f2 100644 --- a/src/mem/cache/cache_builder.cc +++ b/src/mem/cache/cache_builder.cc @@ -70,9 +70,6 @@ #include "base/compression/null_compression.hh" #include "base/compression/lzss_compression.hh" -// CacheTags Templates -#include "mem/cache/tags/cache_tags.hh" - // MissQueue Templates #include "mem/cache/miss/miss_queue.hh" #include "mem/cache/miss/blocking_buffer.hh" @@ -108,8 +105,6 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(BaseCache) Param<int> tgts_per_mshr; Param<int> write_buffers; Param<bool> prioritizeRequests; -// SimObjectParam<Bus *> in_bus; -// SimObjectParam<Bus *> out_bus; SimObjectParam<CoherenceProtocol *> protocol; Param<Addr> trace_addr; Param<int> hash_delay; @@ -122,7 +117,6 @@ BEGIN_DECLARE_SIM_OBJECT_PARAMS(BaseCache) Param<int> compression_latency; Param<int> subblock_size; Param<Counter> max_miss_count; -// SimObjectParam<HierParams *> hier; VectorParam<Range<Addr> > addr_range; // SimObjectParam<MemTraceWriter *> mem_trace; Param<bool> split; @@ -156,9 +150,6 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(BaseCache) INIT_PARAM_DFLT(write_buffers, "number of write buffers", 8), INIT_PARAM_DFLT(prioritizeRequests, "always service demand misses first", false), -/* INIT_PARAM_DFLT(in_bus, "incoming bus object", NULL), - INIT_PARAM(out_bus, "outgoing bus object"), -*/ INIT_PARAM_DFLT(protocol, "coherence protocol to use in the cache", NULL), INIT_PARAM_DFLT(trace_addr, "address to trace", 0), @@ -182,10 +173,6 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(BaseCache) INIT_PARAM_DFLT(max_miss_count, "The number of misses to handle before calling exit", 0), -/* INIT_PARAM_DFLT(hier, - "Hierarchy global variables", - &defaultHierParams), -*/ INIT_PARAM_DFLT(addr_range, "The address range in bytes", vector<Range<Addr> >(1,RangeIn((Addr)0, MaxAddr))), // INIT_PARAM_DFLT(mem_trace, "Memory trace to write accesses to", NULL), @@ -208,47 +195,47 @@ BEGIN_INIT_SIM_OBJECT_PARAMS(BaseCache) END_INIT_SIM_OBJECT_PARAMS(BaseCache) -#define BUILD_CACHE(t, c) do { \ - Prefetcher<CacheTags<t> > *pf; \ - if (pf_policy == "tagged") { \ - BUILD_TAGGED_PREFETCHER(t); \ - } \ - else if (pf_policy == "stride") { \ - BUILD_STRIDED_PREFETCHER(t); \ - } \ - else if (pf_policy == "ghb") { \ - BUILD_GHB_PREFETCHER(t); \ - } \ - else { \ - BUILD_NULL_PREFETCHER(t); \ - } \ - Cache<CacheTags<t>, c>::Params params(tagStore, mq, coh, \ - base_params, \ - pf, \ - prefetch_access, hit_latency); \ - Cache<CacheTags<t>, c> *retval = \ - new Cache<CacheTags<t>, c>(getInstanceName(), params); \ -return retval; \ +#define BUILD_CACHE(TAGS, tags, c) \ + do { \ + Prefetcher<TAGS> *pf; \ + if (pf_policy == "tagged") { \ + BUILD_TAGGED_PREFETCHER(TAGS); \ + } \ + else if (pf_policy == "stride") { \ + BUILD_STRIDED_PREFETCHER(TAGS); \ + } \ + else if (pf_policy == "ghb") { \ + BUILD_GHB_PREFETCHER(TAGS); \ + } \ + else { \ + BUILD_NULL_PREFETCHER(TAGS); \ + } \ + Cache<TAGS, c>::Params params(tags, mq, coh, base_params, \ + pf, prefetch_access, hit_latency, \ + true, \ + store_compressed, \ + adaptive_compression, \ + compressed_bus, \ + compAlg, compression_latency, \ + prefetch_miss); \ + Cache<TAGS, c> *retval = \ + new Cache<TAGS, c>(getInstanceName(), params); \ + return retval; \ } while (0) #define BUILD_CACHE_PANIC(x) do { \ panic("%s not compiled into M5", x); \ } while (0) -#define BUILD_COMPRESSED_CACHE(TAGS, tags, c) \ - do { \ - CompressionAlgorithm *compAlg; \ - if (compressed_bus || store_compressed) { \ - compAlg = new LZSSCompression(); \ - } else { \ - compAlg = new NullCompression(); \ - } \ - CacheTags<TAGS> *tagStore = \ - new CacheTags<TAGS>(tags, compression_latency, true, \ - store_compressed, adaptive_compression, \ - compressed_bus, \ - compAlg, prefetch_miss); \ - BUILD_CACHE(TAGS, c); \ +#define BUILD_COMPRESSED_CACHE(TAGS, tags, c) \ + do { \ + CompressionAlgorithm *compAlg; \ + if (compressed_bus || store_compressed) { \ + compAlg = new LZSSCompression(); \ + } else { \ + compAlg = new NullCompression(); \ + } \ + BUILD_CACHE(TAGS, tags, c); \ } while (0) #if defined(USE_CACHE_FALRU) @@ -328,7 +315,7 @@ return retval; \ #if defined(USE_TAGGED) #define BUILD_TAGGED_PREFETCHER(t) pf = new \ - TaggedPrefetcher<CacheTags<t> >(prefetcher_size, \ + TaggedPrefetcher<t >(prefetcher_size, \ !prefetch_past_page, \ prefetch_serial_squash, \ prefetch_cache_check_push, \ @@ -341,7 +328,7 @@ return retval; \ #if defined(USE_STRIDED) #define BUILD_STRIDED_PREFETCHER(t) pf = new \ - StridePrefetcher<CacheTags<t> >(prefetcher_size, \ + StridePrefetcher<t >(prefetcher_size, \ !prefetch_past_page, \ prefetch_serial_squash, \ prefetch_cache_check_push, \ @@ -355,7 +342,7 @@ return retval; \ #if defined(USE_GHB) #define BUILD_GHB_PREFETCHER(t) pf = new \ - GHBPrefetcher<CacheTags<t> >(prefetcher_size, \ + GHBPrefetcher<t >(prefetcher_size, \ !prefetch_past_page, \ prefetch_serial_squash, \ prefetch_cache_check_push, \ @@ -369,7 +356,7 @@ return retval; \ #if defined(USE_TAGGED) #define BUILD_NULL_PREFETCHER(t) pf = new \ - TaggedPrefetcher<CacheTags<t> >(prefetcher_size, \ + TaggedPrefetcher<t >(prefetcher_size, \ !prefetch_past_page, \ prefetch_serial_squash, \ prefetch_cache_check_push, \ diff --git a/src/mem/cache/cache_impl.hh b/src/mem/cache/cache_impl.hh index 9f48ff1f3..1d78b03c7 100644 --- a/src/mem/cache/cache_impl.hh +++ b/src/mem/cache/cache_impl.hh @@ -72,10 +72,18 @@ Cache(const std::string &_name, prefetchAccess(params.prefetchAccess), tags(params.tags), missQueue(params.missQueue), coherence(params.coherence), prefetcher(params.prefetcher), - hitLatency(params.hitLatency) + hitLatency(params.hitLatency), + compressionAlg(params.compressionAlg), + blkSize(params.blkSize), + doFastWrites(params.doFastWrites), + prefetchMiss(params.prefetchMiss), + storeCompressed(params.storeCompressed), + compressOnWriteback(params.compressOnWriteback), + compLatency(params.compLatency), + adaptiveCompression(params.adaptiveCompression), + writebackCompressed(params.writebackCompressed) { tags->setCache(this); - tags->setPrefetcher(prefetcher); missQueue->setCache(this); missQueue->setPrefetcher(prefetcher); coherence->setCache(this); @@ -98,6 +106,433 @@ Cache<TagStore,Coherence>::regStats() } template<class TagStore, class Coherence> +typename Cache<TagStore,Coherence>::BlkType* +Cache<TagStore,Coherence>::handleAccess(PacketPtr &pkt, int & lat, + PacketList & writebacks, bool update) +{ + // Set the block offset here + int offset = tags->extractBlkOffset(pkt->getAddr()); + + BlkType *blk = NULL; + if (update) { + blk = tags->findBlock(pkt, lat); + } else { + blk = tags->findBlock(pkt->getAddr()); + lat = 0; + } + if (blk != NULL) { + + if (!update) { + if (pkt->isWrite()){ + assert(offset < blkSize); + assert(pkt->getSize() <= blkSize); + assert(offset+pkt->getSize() <= blkSize); + memcpy(blk->data + offset, pkt->getPtr<uint8_t>(), + pkt->getSize()); + } else if (!(pkt->flags & SATISFIED)) { + pkt->flags |= SATISFIED; + pkt->result = Packet::Success; + assert(offset < blkSize); + assert(pkt->getSize() <= blkSize); + assert(offset + pkt->getSize() <=blkSize); + memcpy(pkt->getPtr<uint8_t>(), blk->data + offset, + pkt->getSize()); + } + return blk; + } + + // Hit + if (blk->isPrefetch()) { + //Signal that this was a hit under prefetch (no need for + //use prefetch (only can get here if true) + DPRINTF(HWPrefetch, "Hit a block that was prefetched\n"); + blk->status &= ~BlkHWPrefetched; + if (prefetchMiss) { + //If we are using the miss stream, signal the + //prefetcher otherwise the access stream would have + //already signaled this hit + prefetcher->handleMiss(pkt, curTick); + } + } + + if ((pkt->isWrite() && blk->isWritable()) || + (pkt->isRead() && blk->isValid())) { + + // We are satisfying the request + pkt->flags |= SATISFIED; + + if (blk->isCompressed()) { + // If the data is compressed, need to increase the latency + lat += (compLatency/4); + } + + bool write_data = false; + + assert(verifyData(blk)); + + assert(offset < blkSize); + assert(pkt->getSize() <= blkSize); + assert(offset+pkt->getSize() <= blkSize); + + if (pkt->isWrite()) { + if (blk->checkWrite(pkt->req)) { + write_data = true; + blk->status |= BlkDirty; + memcpy(blk->data + offset, pkt->getPtr<uint8_t>(), + pkt->getSize()); + } + } else { + assert(pkt->isRead()); + if (pkt->req->isLocked()) { + blk->trackLoadLocked(pkt->req); + } + memcpy(pkt->getPtr<uint8_t>(), blk->data + offset, + pkt->getSize()); + } + + if (write_data || + (adaptiveCompression && blk->isCompressed())) + { + // If we wrote data, need to update the internal block + // data. + updateData(blk, writebacks, + !(adaptiveCompression && + blk->isReferenced())); + } + } else { + // permission violation, treat it as a miss + blk = NULL; + } + } else { + // complete miss (no matching block) + if (pkt->req->isLocked() && pkt->isWrite()) { + // miss on store conditional... just give up now + pkt->req->setScResult(0); + pkt->flags |= SATISFIED; + } + } + + return blk; +} + +template<class TagStore, class Coherence> +typename Cache<TagStore,Coherence>::BlkType* +Cache<TagStore,Coherence>::handleFill(BlkType *blk, PacketPtr &pkt, + CacheBlk::State new_state, + PacketList & writebacks, + PacketPtr target) +{ +#ifndef NDEBUG + BlkType *tmp_blk = findBlock(pkt->getAddr()); + assert(tmp_blk == blk); +#endif + blk = doReplacement(blk, pkt, new_state, writebacks); + + + if (pkt->isRead()) { + memcpy(blk->data, pkt->getPtr<uint8_t>(), blkSize); + } + + blk->whenReady = pkt->finishTime; + + // Respond to target, if any + if (target) { + + target->flags |= SATISFIED; + + if (target->cmd == Packet::InvalidateReq) { + invalidateBlk(blk); + blk = NULL; + } + + if (blk && (target->isWrite() ? blk->isWritable() : blk->isValid())) { + assert(target->isWrite() || target->isRead()); + assert(target->getOffset(blkSize) + target->getSize() <= blkSize); + if (target->isWrite()) { + if (blk->checkWrite(pkt->req)) { + blk->status |= BlkDirty; + memcpy(blk->data + target->getOffset(blkSize), + target->getPtr<uint8_t>(), target->getSize()); + } + } else { + if (pkt->req->isLocked()) { + blk->trackLoadLocked(pkt->req); + } + memcpy(target->getPtr<uint8_t>(), + blk->data + target->getOffset(blkSize), + target->getSize()); + } + } + } + + if (blk) { + // Need to write the data into the block + updateData(blk, writebacks, !adaptiveCompression || true); + } + return blk; +} + +template<class TagStore, class Coherence> +typename Cache<TagStore,Coherence>::BlkType* +Cache<TagStore,Coherence>::handleFill(BlkType *blk, MSHR * mshr, + CacheBlk::State new_state, + PacketList & writebacks, PacketPtr pkt) +{ +/* +#ifndef NDEBUG + BlkType *tmp_blk = findBlock(mshr->pkt->getAddr()); + assert(tmp_blk == blk); +#endif + PacketPtr pkt = mshr->pkt;*/ + blk = doReplacement(blk, pkt, new_state, writebacks); + + if (pkt->isRead()) { + memcpy(blk->data, pkt->getPtr<uint8_t>(), blkSize); + } + + blk->whenReady = pkt->finishTime; + + + // respond to MSHR targets, if any + + // First offset for critical word first calculations + int initial_offset = 0; + + if (mshr->hasTargets()) { + initial_offset = mshr->getTarget()->getOffset(blkSize); + } + + while (mshr->hasTargets()) { + PacketPtr target = mshr->getTarget(); + + target->flags |= SATISFIED; + + // How many bytes pass the first request is this one + int transfer_offset = target->getOffset(blkSize) - initial_offset; + if (transfer_offset < 0) { + transfer_offset += blkSize; + } + + // If critical word (no offset) return first word time + Tick completion_time = tags->getHitLatency() + + transfer_offset ? pkt->finishTime : pkt->firstWordTime; + + if (target->cmd == Packet::InvalidateReq) { + //Mark the blk as invalid now, if it hasn't been already + if (blk) { + invalidateBlk(blk); + blk = NULL; + } + + //Also get rid of the invalidate + mshr->popTarget(); + + DPRINTF(Cache, "Popping off a Invalidate for addr %x\n", + pkt->getAddr()); + + continue; + } + + if (blk && (target->isWrite() ? blk->isWritable() : blk->isValid())) { + assert(target->isWrite() || target->isRead()); + assert(target->getOffset(blkSize) + target->getSize() <= blkSize); + if (target->isWrite()) { + if (blk->checkWrite(pkt->req)) { + blk->status |= BlkDirty; + memcpy(blk->data + target->getOffset(blkSize), + target->getPtr<uint8_t>(), target->getSize()); + } + } else { + if (pkt->req->isLocked()) { + blk->trackLoadLocked(pkt->req); + } + memcpy(target->getPtr<uint8_t>(), + blk->data + target->getOffset(blkSize), + target->getSize()); + } + } else { + // Invalid access, need to do another request + // can occur if block is invalidated, or not correct + // permissions +// mshr->pkt = pkt; + break; + } + respondToMiss(target, completion_time); + mshr->popTarget(); + } + + if (blk) { + // Need to write the data into the block + updateData(blk, writebacks, !adaptiveCompression || true); + } + + return blk; +} + + +template<class TagStore, class Coherence> +void +Cache<TagStore,Coherence>::handleSnoop(BlkType *blk, + CacheBlk::State new_state, + PacketPtr &pkt) +{ + //Must have the block to supply + assert(blk); + // Can only supply data, and if it hasn't already been supllied + assert(pkt->isRead()); + assert(!(pkt->flags & SATISFIED)); + pkt->flags |= SATISFIED; + Addr offset = pkt->getOffset(blkSize); + assert(offset < blkSize); + assert(pkt->getSize() <= blkSize); + assert(offset + pkt->getSize() <=blkSize); + memcpy(pkt->getPtr<uint8_t>(), blk->data + offset, pkt->getSize()); + + handleSnoop(blk, new_state); +} + +template<class TagStore, class Coherence> +void +Cache<TagStore,Coherence>::handleSnoop(BlkType *blk, + CacheBlk::State new_state) +{ + if (blk && blk->status != new_state) { + if ((new_state && BlkValid) == 0) { + invalidateBlk(blk); + } else { + assert(new_state >= 0 && new_state < 128); + blk->status = new_state; + } + } +} + +template<class TagStore, class Coherence> +PacketPtr +Cache<TagStore,Coherence>::writebackBlk(BlkType *blk) +{ + assert(blk && blk->isValid() && blk->isModified()); + int data_size = blkSize; + data_size = blk->size; + if (compressOnWriteback) { + // not already compressed + // need to compress to ship it + assert(data_size == blkSize); + uint8_t *tmp_data = new uint8_t[blkSize]; + data_size = compressionAlg->compress(tmp_data,blk->data, + data_size); + delete [] tmp_data; + } + +/* PacketPtr writeback = + buildWritebackReq(tags->regenerateBlkAddr(blk->tag, blk->set), + blk->asid, blkSize, + blk->data, data_size); +*/ + + Request *writebackReq = + new Request(tags->regenerateBlkAddr(blk->tag, blk->set), blkSize, 0); + PacketPtr writeback = new Packet(writebackReq, Packet::Writeback, -1); + writeback->allocate(); + memcpy(writeback->getPtr<uint8_t>(),blk->data,blkSize); + + blk->status &= ~BlkDirty; + return writeback; +} + + +template<class TagStore, class Coherence> +bool +Cache<TagStore,Coherence>::verifyData(BlkType *blk) +{ + bool retval; + // The data stored in the blk + uint8_t *blk_data = new uint8_t[blkSize]; + tags->readData(blk, blk_data); + // Pointer for uncompressed data, assumed uncompressed + uint8_t *tmp_data = blk_data; + // The size of the data being stored, assumed uncompressed + int data_size = blkSize; + + // If the block is compressed need to uncompress to access + if (blk->isCompressed()){ + // Allocate new storage for the data + tmp_data = new uint8_t[blkSize]; + data_size = compressionAlg->uncompress(tmp_data,blk_data, blk->size); + assert(data_size == blkSize); + // Don't need to keep blk_data around + delete [] blk_data; + } else { + assert(blkSize == blk->size); + } + + retval = memcmp(tmp_data, blk->data, blkSize) == 0; + delete [] tmp_data; + return retval; +} + +template<class TagStore, class Coherence> +void +Cache<TagStore,Coherence>::updateData(BlkType *blk, PacketList &writebacks, + bool compress_block) +{ + if (storeCompressed && compress_block) { + uint8_t *comp_data = new uint8_t[blkSize]; + int new_size = compressionAlg->compress(comp_data, blk->data, blkSize); + if (new_size > (blkSize - tags->getSubBlockSize())){ + // no benefit to storing it compressed + blk->status &= ~BlkCompressed; + tags->writeData(blk, blk->data, blkSize, + writebacks); + } else { + // Store the data compressed + blk->status |= BlkCompressed; + tags->writeData(blk, comp_data, new_size, + writebacks); + } + delete [] comp_data; + } else { + blk->status &= ~BlkCompressed; + tags->writeData(blk, blk->data, blkSize, writebacks); + } +} + +template<class TagStore, class Coherence> +typename Cache<TagStore,Coherence>::BlkType* +Cache<TagStore,Coherence>::doReplacement(BlkType *blk, PacketPtr &pkt, + CacheBlk::State new_state, + PacketList &writebacks) +{ + if (blk == NULL) { + // need to do a replacement + BlkList compress_list; + blk = tags->findReplacement(pkt, writebacks, compress_list); + while (adaptiveCompression && !compress_list.empty()) { + updateData(compress_list.front(), writebacks, true); + compress_list.pop_front(); + } + if (blk->isValid()) { + DPRINTF(Cache, "replacement: replacing %x with %x: %s\n", + tags->regenerateBlkAddr(blk->tag,blk->set), pkt->getAddr(), + (blk->isModified()) ? "writeback" : "clean"); + + if (blk->isModified()) { + // Need to write the data back + writebacks.push_back(writebackBlk(blk)); + } + } + blk->tag = tags->extractTag(pkt->getAddr(), blk); + } else { + // must be a status change + // assert(blk->status != new_state); + if (blk->status == new_state) warn("Changing state to same value\n"); + } + + blk->status = new_state; + return blk; +} + + +template<class TagStore, class Coherence> bool Cache<TagStore,Coherence>::access(PacketPtr &pkt) { @@ -112,7 +547,7 @@ Cache<TagStore,Coherence>::access(PacketPtr &pkt) prefetcher->handleMiss(pkt, curTick); } if (!pkt->req->isUncacheable()) { - blk = tags->handleAccess(pkt, lat, writebacks); + blk = handleAccess(pkt, lat, writebacks); } else { size = pkt->getSize(); } @@ -130,7 +565,7 @@ Cache<TagStore,Coherence>::access(PacketPtr &pkt) warn("WriteInv doing a fastallocate" "with an outstanding miss to the same address\n"); } - blk = tags->handleFill(NULL, pkt, BlkValid | BlkWritable, + blk = handleFill(NULL, pkt, BlkValid | BlkWritable, writebacks); ++fastWrites; } @@ -195,7 +630,7 @@ Cache<TagStore,Coherence>::getPacket() if (!pkt->req->isUncacheable()) { if (pkt->cmd == Packet::HardPFReq) misses[Packet::HardPFReq][0/*pkt->req->getThreadNum()*/]++; - BlkType *blk = tags->findBlock(pkt); + BlkType *blk = findBlock(pkt); Packet::Command cmd = coherence->getBusCmd(pkt->cmd, (blk)? blk->status : 0); missQueue->setBusCmd(pkt, cmd); @@ -224,7 +659,7 @@ Cache<TagStore,Coherence>::sendResult(PacketPtr &pkt, MSHR* mshr, if (upgrade) { assert(pkt); //Upgrades need to be fixed pkt->flags &= ~CACHE_LINE_FILL; - BlkType *blk = tags->findBlock(pkt); + BlkType *blk = findBlock(pkt); CacheBlk::State old_state = (blk) ? blk->status : 0; CacheBlk::State new_state = coherence->getNewState(pkt,old_state); if (old_state != new_state) @@ -233,7 +668,7 @@ Cache<TagStore,Coherence>::sendResult(PacketPtr &pkt, MSHR* mshr, //Set the state on the upgrade memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize); PacketList writebacks; - tags->handleFill(blk, mshr, new_state, writebacks, pkt); + handleFill(blk, mshr, new_state, writebacks, pkt); assert(writebacks.empty()); missQueue->handleResponse(pkt, curTick + hitLatency); } @@ -275,7 +710,7 @@ Cache<TagStore,Coherence>::handleResponse(PacketPtr &pkt) if (pkt->isCacheFill() && !pkt->isNoAllocate()) { DPRINTF(Cache, "Block for addr %x being updated in Cache\n", pkt->getAddr()); - blk = tags->findBlock(pkt); + blk = findBlock(pkt); CacheBlk::State old_state = (blk) ? blk->status : 0; PacketList writebacks; CacheBlk::State new_state = coherence->getNewState(pkt,old_state); @@ -284,7 +719,7 @@ Cache<TagStore,Coherence>::handleResponse(PacketPtr &pkt) "state %i to %i\n", pkt->getAddr(), old_state, new_state); - blk = tags->handleFill(blk, (MSHR*)pkt->senderState, + blk = handleFill(blk, (MSHR*)pkt->senderState, new_state, writebacks, pkt); while (!writebacks.empty()) { missQueue->doWriteback(writebacks.front()); @@ -332,7 +767,7 @@ Cache<TagStore,Coherence>::snoop(PacketPtr &pkt) } Addr blk_addr = pkt->getAddr() & ~(Addr(blkSize-1)); - BlkType *blk = tags->findBlock(pkt); + BlkType *blk = findBlock(pkt); MSHR *mshr = missQueue->findMSHR(blk_addr); if (coherence->hasProtocol() || pkt->isInvalidate()) { //@todo Move this into handle bus req @@ -435,7 +870,7 @@ Cache<TagStore,Coherence>::snoop(PacketPtr &pkt) "now supplying data, new state is %i\n", pkt->cmdString(), blk_addr, new_state); - tags->handleSnoop(blk, new_state, pkt); + handleSnoop(blk, new_state, pkt); respondToSnoop(pkt, curTick + hitLatency); return; } @@ -443,7 +878,7 @@ Cache<TagStore,Coherence>::snoop(PacketPtr &pkt) DPRINTF(Cache, "Cache snooped a %s request for addr %x, " "new state is %i\n", pkt->cmdString(), blk_addr, new_state); - tags->handleSnoop(blk, new_state); + handleSnoop(blk, new_state); } template<class TagStore, class Coherence> @@ -501,7 +936,7 @@ Cache<TagStore,Coherence>::probe(PacketPtr &pkt, bool update, PacketList writebacks; int lat; - BlkType *blk = tags->handleAccess(pkt, lat, writebacks, update); + BlkType *blk = handleAccess(pkt, lat, writebacks, update); DPRINTF(Cache, "%s %x %s\n", pkt->cmdString(), pkt->getAddr(), (blk) ? "hit" : "miss"); @@ -557,7 +992,7 @@ Cache<TagStore,Coherence>::probe(PacketPtr &pkt, bool update, if (!pkt->req->isUncacheable() /*Uncacheables just go through*/ && (pkt->cmd != Packet::Writeback)/*Writebacks on miss fall through*/) { // Fetch the cache block to fill - BlkType *blk = tags->findBlock(pkt); + BlkType *blk = findBlock(pkt); Packet::Command temp_cmd = coherence->getBusCmd(pkt->cmd, (blk)? blk->status : 0); @@ -593,7 +1028,7 @@ return 0; DPRINTF(Cache, "Block for blk addr %x moving from state " "%i to %i\n", busPkt->getAddr(), old_state, new_state); - tags->handleFill(blk, busPkt, new_state, writebacks, pkt); + handleFill(blk, busPkt, new_state, writebacks, pkt); //Free the packet delete busPkt; @@ -639,7 +1074,7 @@ Cache<TagStore,Coherence>::snoopProbe(PacketPtr &pkt) } Addr blk_addr = pkt->getAddr() & ~(Addr(blkSize-1)); - BlkType *blk = tags->findBlock(pkt); + BlkType *blk = findBlock(pkt); MSHR *mshr = missQueue->findMSHR(blk_addr); CacheBlk::State new_state = 0; bool satisfy = coherence->handleBusRequest(pkt,blk,mshr, new_state); @@ -648,14 +1083,14 @@ Cache<TagStore,Coherence>::snoopProbe(PacketPtr &pkt) "now supplying data, new state is %i\n", pkt->cmdString(), blk_addr, new_state); - tags->handleSnoop(blk, new_state, pkt); + handleSnoop(blk, new_state, pkt); return hitLatency; } if (blk) DPRINTF(Cache, "Cache snooped a %s request for addr %x, " "new state is %i\n", pkt->cmdString(), blk_addr, new_state); - tags->handleSnoop(blk, new_state); + handleSnoop(blk, new_state); return 0; } diff --git a/src/mem/cache/prefetch/ghb_prefetcher.cc b/src/mem/cache/prefetch/ghb_prefetcher.cc index a6c419113..7d537641e 100644 --- a/src/mem/cache/prefetch/ghb_prefetcher.cc +++ b/src/mem/cache/prefetch/ghb_prefetcher.cc @@ -34,8 +34,6 @@ * GHB Prefetcher template instantiations. */ -#include "mem/cache/tags/cache_tags.hh" - #include "mem/cache/tags/lru.hh" #include "mem/cache/prefetch/ghb_prefetcher.hh" @@ -43,6 +41,6 @@ // Template Instantiations #ifndef DOXYGEN_SHOULD_SKIP_THIS -template class GHBPrefetcher<CacheTags<LRU> >; +template class GHBPrefetcher<LRU >; #endif //DOXYGEN_SHOULD_SKIP_THIS diff --git a/src/mem/cache/prefetch/stride_prefetcher.cc b/src/mem/cache/prefetch/stride_prefetcher.cc index 2204871cc..847f2979e 100644 --- a/src/mem/cache/prefetch/stride_prefetcher.cc +++ b/src/mem/cache/prefetch/stride_prefetcher.cc @@ -34,8 +34,6 @@ * Stride Prefetcher template instantiations. */ -#include "mem/cache/tags/cache_tags.hh" - #include "mem/cache/tags/lru.hh" #include "mem/cache/prefetch/stride_prefetcher.hh" @@ -43,6 +41,6 @@ // Template Instantiations #ifndef DOXYGEN_SHOULD_SKIP_THIS -template class StridePrefetcher<CacheTags<LRU> >; +template class StridePrefetcher<LRU >; #endif //DOXYGEN_SHOULD_SKIP_THIS |