Add support for caching small reads.

Add benchmarking to verify this is faster.

Test: Ran unit tests.
Change-Id: I1487114331f4581ec2368e56c4f18c6e3e6bcc7d
diff --git a/libunwindstack/Memory.cpp b/libunwindstack/Memory.cpp
index cfa8c6d..a30d65e 100644
--- a/libunwindstack/Memory.cpp
+++ b/libunwindstack/Memory.cpp
@@ -174,6 +174,13 @@
   return std::shared_ptr<Memory>(new MemoryRemote(pid));
 }
 
+std::shared_ptr<Memory> Memory::CreateProcessMemoryCached(pid_t pid) {
+  if (pid == getpid()) {
+    return std::shared_ptr<Memory>(new MemoryCache(new MemoryLocal()));
+  }
+  return std::shared_ptr<Memory>(new MemoryCache(new MemoryRemote(pid)));
+}
+
 size_t MemoryBuffer::Read(uint64_t addr, void* dst, size_t size) {
   if (addr >= raw_.size()) {
     return 0;
@@ -398,4 +405,50 @@
   return 0;
 }
 
+size_t MemoryCache::Read(uint64_t addr, void* dst, size_t size) {
+  // Only bother caching and looking at the cache if this is a small read for now.
+  if (size > 64) {
+    return impl_->Read(addr, dst, size);
+  }
+
+  uint64_t addr_page = addr >> kCacheBits;
+  auto entry = cache_.find(addr_page);
+  uint8_t* cache_dst;
+  if (entry != cache_.end()) {
+    cache_dst = entry->second;
+  } else {
+    cache_dst = cache_[addr_page];
+    if (!impl_->ReadFully(addr_page << kCacheBits, cache_dst, kCacheSize)) {
+      // Erase the entry.
+      cache_.erase(addr_page);
+      return impl_->Read(addr, dst, size);
+    }
+  }
+  size_t max_read = ((addr_page + 1) << kCacheBits) - addr;
+  if (size <= max_read) {
+    memcpy(dst, &cache_dst[addr & kCacheMask], size);
+    return size;
+  }
+
+  // The read crossed into another cached entry, since a read can only cross
+  // into one extra cached page, duplicate the code rather than looping.
+  memcpy(dst, &cache_dst[addr & kCacheMask], max_read);
+  dst = &reinterpret_cast<uint8_t*>(dst)[max_read];
+  addr_page++;
+
+  entry = cache_.find(addr_page);
+  if (entry != cache_.end()) {
+    cache_dst = entry->second;
+  } else {
+    cache_dst = cache_[addr_page];
+    if (!impl_->ReadFully(addr_page << kCacheBits, cache_dst, kCacheSize)) {
+      // Erase the entry.
+      cache_.erase(addr_page);
+      return impl_->Read(addr_page << kCacheBits, dst, size - max_read) + max_read;
+    }
+  }
+  memcpy(dst, cache_dst, size - max_read);
+  return size;
+}
+
 }  // namespace unwindstack