Merge pull request #5437

acb68dba bulletproofs: cut down on keyV allocations (moneromooo-monero)

1 files changed, 33 insertions, 29 deletions

diff --git a/src/ringct/bulletproofs.cc b/src/ringct/bulletproofs.cc
index e394ef088..6270d4d14 100644
--- a/src/ringct/bulletproofs.cc
+++ b/src/ringct/bulletproofs.cc
@@ -48,12 +48,12 @@ extern "C"
 
 //#define DEBUG_BP
 
-#if 1
+#if 0
 #define PERF_TIMER_START_BP(x) PERF_TIMER_START_UNIT(x, 1000000)
 #define PERF_TIMER_STOP_BP(x) PERF_TIMER_STOP(x)
 #else
-#define PERF_TIMER_START_BP(x) ((void*)0)
-#define PERF_TIMER_STOP_BP(x) ((void*)0)
+#define PERF_TIMER_START_BP(x) ((void)0)
+#define PERF_TIMER_STOP_BP(x) ((void)0)
 #endif
 
 #define STRAUS_SIZE_LIMIT 232
@@ -439,35 +439,35 @@ static epee::span<const rct::key> slice(const rct::keyV &a, size_t start, size_t
 
 static rct::key hash_cache_mash(rct::key &hash_cache, const rct::key &mash0, const rct::key &mash1)
 {
-  rct::keyV data;
-  data.reserve(3);
-  data.push_back(hash_cache);
-  data.push_back(mash0);
-  data.push_back(mash1);
-  return hash_cache = rct::hash_to_scalar(data);
+  rct::key data[3];
+  data[0] = hash_cache;
+  data[1] = mash0;
+  data[2] = mash1;
+  rct::hash_to_scalar(hash_cache, data, sizeof(data));
+  return hash_cache;
 }
 
 static rct::key hash_cache_mash(rct::key &hash_cache, const rct::key &mash0, const rct::key &mash1, const rct::key &mash2)
 {
-  rct::keyV data;
-  data.reserve(4);
-  data.push_back(hash_cache);
-  data.push_back(mash0);
-  data.push_back(mash1);
-  data.push_back(mash2);
-  return hash_cache = rct::hash_to_scalar(data);
+  rct::key data[4];
+  data[0] = hash_cache;
+  data[1] = mash0;
+  data[2] = mash1;
+  data[3] = mash2;
+  rct::hash_to_scalar(hash_cache, data, sizeof(data));
+  return hash_cache;
 }
 
 static rct::key hash_cache_mash(rct::key &hash_cache, const rct::key &mash0, const rct::key &mash1, const rct::key &mash2, const rct::key &mash3)
 {
-  rct::keyV data;
-  data.reserve(5);
-  data.push_back(hash_cache);
-  data.push_back(mash0);
-  data.push_back(mash1);
-  data.push_back(mash2);
-  data.push_back(mash3);
-  return hash_cache = rct::hash_to_scalar(data);
+  rct::key data[5];
+  data[0] = hash_cache;
+  data[1] = mash0;
+  data[2] = mash1;
+  data[3] = mash2;
+  data[4] = mash3;
+  rct::hash_to_scalar(hash_cache, data, sizeof(data));
+  return hash_cache;
 }
 
 /* Given a value v (0..2^N-1) and a mask gamma, construct a range proof */
@@ -825,6 +825,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
   size_t inv_offset = 0;
   std::vector<rct::key> to_invert;
   to_invert.reserve(11 * sizeof(proofs));
+  size_t max_logM = 0;
   for (const Bulletproof *p: proofs)
   {
     const Bulletproof &proof = *p;
@@ -861,6 +862,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
     size_t M;
     for (pd.logM = 0; (M = 1<<pd.logM) <= maxM && M < proof.V.size(); ++pd.logM);
     CHECK_AND_ASSERT_MES(proof.L.size() == 6+pd.logM, false, "Proof is not the expected size");
+    max_logM = std::max(pd.logM, max_logM);
 
     const size_t rounds = pd.logM+logN;
     CHECK_AND_ASSERT_MES(rounds > 0, false, "Zero rounds");
@@ -888,7 +890,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
   rct::key tmp;
 
   std::vector<MultiexpData> multiexp_data;
-  multiexp_data.reserve(nV + (2 * (10/*logM*/ + logN) + 4) * proofs.size() + 2 * maxMN);
+  multiexp_data.reserve(nV + (2 * (max_logM + logN) + 4) * proofs.size() + 2 * maxMN);
   multiexp_data.resize(2 * maxMN);
 
   PERF_TIMER_START_BP(VERIFY_line_24_25_invert);
@@ -901,6 +903,8 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
   rct::keyV m_z4(maxMN, rct::zero()), m_z5(maxMN, rct::zero());
   rct::key m_y0 = rct::zero(), y1 = rct::zero();
   int proof_data_index = 0;
+  rct::keyV w_cache;
+  rct::keyV proof8_V, proof8_L, proof8_R;
   for (const Bulletproof *p: proofs)
   {
     const Bulletproof &proof = *p;
@@ -913,9 +917,9 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
     const rct::key weight_z = rct::skGen();
 
     // pre-multiply some points by 8
-    rct::keyV proof8_V = proof.V; for (rct::key &k: proof8_V) k = rct::scalarmult8(k);
-    rct::keyV proof8_L = proof.L; for (rct::key &k: proof8_L) k = rct::scalarmult8(k);
-    rct::keyV proof8_R = proof.R; for (rct::key &k: proof8_R) k = rct::scalarmult8(k);
+    proof8_V.resize(proof.V.size()); for (size_t i = 0; i < proof.V.size(); ++i) proof8_V[i] = rct::scalarmult8(proof.V[i]);
+    proof8_L.resize(proof.L.size()); for (size_t i = 0; i < proof.L.size(); ++i) proof8_L[i] = rct::scalarmult8(proof.L[i]);
+    proof8_R.resize(proof.R.size()); for (size_t i = 0; i < proof.R.size(); ++i) proof8_R[i] = rct::scalarmult8(proof.R[i]);
     rct::key proof8_T1 = rct::scalarmult8(proof.T1);
     rct::key proof8_T2 = rct::scalarmult8(proof.T2);
     rct::key proof8_S = rct::scalarmult8(proof.S);
@@ -976,7 +980,7 @@ bool bulletproof_VERIFY(const std::vector<const Bulletproof*> &proofs)
 
     // precalc
     PERF_TIMER_START_BP(VERIFY_line_24_25_precalc);
-    rct::keyV w_cache(1<<rounds);
+    w_cache.resize(1<<rounds);
     w_cache[0] = winv[0];
     w_cache[1] = pd.w[0];
     for (size_t j = 1; j < rounds; ++j)