Optimize DBIter::Prev() by reducing stack overhead

Summary:
It looks like we are spending significant amount of time creating std::deque<std::string> every time we do Iterator::Prev()

{F921567}

By using merge_operands_ as a DBIter data member w create it once and reduce this overhead and see ~30% performance improvement when using Iterator::Prev() on hot data

Orignal performance

```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readreverse" --db="/dev/shm/bench_prev_opt/" --use_existing_db --disable_auto_compactions
readreverse  :       0.713 micros/op 1402219 ops/sec;  155.1 MB/s
readreverse  :       0.609 micros/op 1641386 ops/sec;  181.6 MB/s
readreverse  :       0.684 micros/op 1461150 ops/sec;  161.6 MB/s
readreverse  :       0.629 micros/op 1589842 ops/sec;  175.9 MB/s
readreverse  :       0.647 micros/op 1544530 ops/sec;  170.9 MB/s
```

After optimization

```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readreverse" --db="/dev/shm/bench_prev_opt/" --use_existing_db --disable_auto_compactions
readreverse  :       0.488 micros/op 2051189 ops/sec;  226.9 MB/s
readreverse  :       0.505 micros/op 1980892 ops/sec;  219.1 MB/s
readreverse  :       0.541 micros/op 1846971 ops/sec;  204.3 MB/s
readreverse  :       0.497 micros/op 2013612 ops/sec;  222.8 MB/s
readreverse  :       0.480 micros/op 2082665 ops/sec;  230.4 MB/s
```

Test Plan: make check -j64

Reviewers: sdong, anthony, rven, igor, yhchiang

Reviewed By: yhchiang

Subscribers: jkedgar, dhruba

Differential Revision: https://reviews.facebook.net/D52563

db/db_iter.cc

@@ -190,6 +190,8 @@ class DBIter: public Iterator {
   IterKey prefix_start_;
   bool prefix_same_as_start_;
   bool iter_pinned_;
+  // List of operands for merge operator.
+  std::deque<std::string> merge_operands_;
 
   // No copying allowed
   DBIter(const DBIter&);
@@ -493,8 +495,7 @@ void DBIter::PrevInternal() {
 // saved_value_
 bool DBIter::FindValueForCurrentKey() {
   assert(iter_->Valid());
-  // Contains operands for merge operator.
+  merge_operands_.clear();
-  std::deque<std::string> operands;
   // last entry before merge (could be kTypeDeletion, kTypeSingleDeletion or
   // kTypeValue)
   ValueType last_not_merge_type = kTypeDeletion;
@@ -514,19 +515,19 @@ bool DBIter::FindValueForCurrentKey() {
     last_key_entry_type = ikey.type;
     switch (last_key_entry_type) {
       case kTypeValue:
-        operands.clear();
+        merge_operands_.clear();
         saved_value_ = iter_->value().ToString();
         last_not_merge_type = kTypeValue;
         break;
       case kTypeDeletion:
       case kTypeSingleDeletion:
-        operands.clear();
+        merge_operands_.clear();
         last_not_merge_type = last_key_entry_type;
         PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
         break;
       case kTypeMerge:
         assert(user_merge_operator_ != nullptr);
-        operands.push_back(iter_->value().ToString());
+        merge_operands_.push_back(iter_->value().ToString());
         break;
       default:
         assert(false);
@@ -548,8 +549,9 @@ bool DBIter::FindValueForCurrentKey() {
       if (last_not_merge_type == kTypeDeletion) {
         StopWatchNano timer(env_, statistics_ != nullptr);
         PERF_TIMER_GUARD(merge_operator_time_nanos);
-        user_merge_operator_->FullMerge(saved_key_.GetKey(), nullptr, operands,
+        user_merge_operator_->FullMerge(saved_key_.GetKey(), nullptr,
-                                        &saved_value_, logger_);
+                                        merge_operands_, &saved_value_,
+                                        logger_);
         RecordTick(statistics_, MERGE_OPERATION_TOTAL_TIME,
                    timer.ElapsedNanos());
       } else {
@@ -560,7 +562,8 @@ bool DBIter::FindValueForCurrentKey() {
           StopWatchNano timer(env_, statistics_ != nullptr);
           PERF_TIMER_GUARD(merge_operator_time_nanos);
           user_merge_operator_->FullMerge(saved_key_.GetKey(), &temp_slice,
-                                          operands, &saved_value_, logger_);
+                                          merge_operands_, &saved_value_,
+                                          logger_);
           RecordTick(statistics_, MERGE_OPERATION_TOTAL_TIME,
                      timer.ElapsedNanos());
         }