In IF decryption, two large powmods are done, one mod p and one mod q. Spawn

one of them off in a new thread and compute the other on the current thread.
Performance on my Core2 shows a 60 to 90% improvement in overall speed in
RSA private key operations. Will probably be even better once std::async
is available (not currently in GCC) since it will probably use a thread pool
which will amortize the thread creation/shutdown cost.

1 files changed, 22 insertions, 1 deletions

diff --git a/src/pubkey/if_algo/if_op.cpp b/src/pubkey/if_algo/if_op.cpp
index 27aef453e..a59c7d5f9 100644
--- a/src/pubkey/if_algo/if_op.cpp
+++ b/src/pubkey/if_algo/if_op.cpp
@@ -7,6 +7,8 @@
 
 #include <botan/if_op.h>
 #include <botan/numthry.h>
+#include <future>
+#include <thread>
 
 namespace Botan {
 
@@ -38,8 +40,27 @@ BigInt Default_IF_Op::private_op(const BigInt& i) const
    if(q == 0)
       throw Internal_Error("Default_IF_Op::private_op: No private key");
 
-   BigInt j1 = powermod_d1_p(i);
+   /*
+   * A simple std::bind(powermod_d1_p, i) would work instead of a
+   * lambda but GCC 4.5's std::result_of doesn't use decltype and gets
+   * confused
+   *
+   * Todo: use std::async() once it is in GCC
+   *    auto future_j1 = std::async(std::bind(powermod_d1_p, i));
+   *    BigInt j2 = powermod_d2_q(i);
+   *    BigInt j1 = future.get();
+   */
+   std::packaged_task<BigInt ()> task_j1([&]() { return powermod_d1_p(i); });
+   auto future_j1 = task_j1.get_future();
+
+   std::thread thr_j1(std::move(task_j1));
+
    BigInt j2 = powermod_d2_q(i);
+
+   BigInt j1 = future_j1.get();
+
+   thr_j1.join();
+
    j1 = reducer.reduce(sub_mul(j1, j2, c));
    return mul_add(j1, q, j2);
    }