Optimize computation of CTR input blocks

We don't need to read each block since we know what is there

Improves CTR perf with AES-NI by 5-6%, also helps GCM

GH #969

1 files changed, 13 insertions, 8 deletions

diff --git a/src/lib/stream/ctr/ctr.cpp b/src/lib/stream/ctr/ctr.cpp
index 3608eedf9..22cfade9b 100644
--- a/src/lib/stream/ctr/ctr.cpp
+++ b/src/lib/stream/ctr/ctr.cpp
@@ -142,37 +142,42 @@ void CTR_BE::add_counter(const uint64_t counter)
    if(ctr_size == 4)
       {
       size_t off = (BS - 4);
+      uint32_t low32 = counter + load_be<uint32_t>(&m_counter[off], 0);
+
       for(size_t i = 0; i != ctr_blocks; ++i)
          {
-         uint32_t low32 = load_be<uint32_t>(&m_counter[off], 0);
-         low32 += counter;
          store_be(low32, &m_counter[off]);
          off += BS;
+         low32 += 1;
          }
       }
    else if(ctr_size == 8)
       {
       size_t off = (BS - 8);
+      uint64_t low64 = counter + load_be<uint64_t>(&m_counter[off], 0);
+
       for(size_t i = 0; i != ctr_blocks; ++i)
          {
-         uint64_t low64 = load_be<uint64_t>(&m_counter[off], 0);
-         low64 += counter;
          store_be(low64, &m_counter[off]);
          off += BS;
+         low64 += 1;
          }
       }
    else if(ctr_size == 16)
       {
       size_t off = (BS - 16);
+      uint64_t b0 = load_be<uint64_t>(&m_counter[off], 0);
+      uint64_t b1 = load_be<uint64_t>(&m_counter[off], 1);
+      b1 += counter;
+      b0 += (b1 < counter) ? 1 : 0; // carry
+
       for(size_t i = 0; i != ctr_blocks; ++i)
          {
-         uint64_t b0 = load_be<uint64_t>(&m_counter[off], 0);
-         uint64_t b1 = load_be<uint64_t>(&m_counter[off], 1);
-         b1 += counter;
-         b0 += (b1 < counter) ? 1 : 0; // carry
          store_be(b0, &m_counter[off]);
          store_be(b1, &m_counter[off+8]);
          off += BS;
+         b1 += 1;
+         b0 += (b1 == 0); // carry
          }
       }
    else