From 7fba2a1bccc7674794eb832720d6c9e568039192 Mon Sep 17 00:00:00 2001
From: Oleg Andreev <oleganza@gmail.com>
Date: Wed, 22 May 2019 11:14:26 -0700
Subject: [PATCH] avoid unnecessary allocation

---
 src/backend/serial/scalar_mul/pippenger.rs | 23 ++++++++++-----------
 src/backend/vector/scalar_mul/pippenger.rs | 24 ++++++++--------------
 2 files changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/backend/serial/scalar_mul/pippenger.rs b/src/backend/serial/scalar_mul/pippenger.rs
index 82cf1bd..2425ee1 100644
--- a/src/backend/serial/scalar_mul/pippenger.rs
+++ b/src/backend/serial/scalar_mul/pippenger.rs
@@ -112,7 +112,7 @@ impl VartimeMultiscalarMul for Pippenger {
             .map(|_| EdwardsPoint::identity())
             .collect();
 
-        let columns: Vec<_> = (0..digits_count).map(|digit_index| {
+        let mut columns = (0..digits_count).rev().map(|digit_index| {
 
             // Clear the buckets when processing another digit.
             for i in 0..buckets_count {
@@ -150,18 +150,17 @@ impl VartimeMultiscalarMul for Pippenger {
             }
 
             buckets_sum
-        })
-        .collect();
-        // ^ Note: we collect points because if we chain .rev().fold()
-        // then the .map() will run in reversed order, producing incorrect digit values
-        // (they can only be produced in lo->hi order).
+        });
 
-        // Add the intermediate per-digit results in hi->lo order
-        // so that we can minimize doublings.
-        Some(columns[0..(digits_count - 1)].iter().rev().fold(
-            columns[digits_count - 1],
-            |total, &p| total.mul_by_pow_2(w as u32) + p,
-        ))
+        // Take the high column as an initial value to avoid wasting time doubling the identity element in `fold()`.
+        // `unwrap()` always succeeds because we know we have more than zero digits.
+        let hi_column = columns.next().unwrap();
+
+        Some(
+            columns.fold(hi_column, |total, p| {
+                total.mul_by_pow_2(w as u32) + p
+            }).into()
+        )
     }
 }
 
diff --git a/src/backend/vector/scalar_mul/pippenger.rs b/src/backend/vector/scalar_mul/pippenger.rs
index 070273b..039ea56 100644
--- a/src/backend/vector/scalar_mul/pippenger.rs
+++ b/src/backend/vector/scalar_mul/pippenger.rs
@@ -69,7 +69,7 @@ impl VartimeMultiscalarMul for Pippenger {
             .map(|_| ExtendedPoint::identity())
             .collect();
 
-        let columns: Vec<ExtendedPoint> = (0..digits_count).map(|digit_index| {
+        let mut columns = (0..digits_count).rev().map(|digit_index| {
 
             // Clear the buckets when processing another digit.
             for i in 0..buckets_count {
@@ -107,22 +107,16 @@ impl VartimeMultiscalarMul for Pippenger {
             }
 
             buckets_sum
-        })
-        .collect();
-        // ^ Note: we collect points because if we chain .rev().fold()
-        // then the .map() will run in reversed order, producing incorrect digit values
-        // (they can only be produced in lo->hi order).
+        });
+
+        // Take the high column as an initial value to avoid wasting time doubling the identity element in `fold()`.
+        // `unwrap()` always succeeds because we know we have more than zero digits.
+        let hi_column = columns.next().unwrap();
 
-        // Add the intermediate per-digit results in hi->lo order
-        // so that we can minimize doublings.
         Some(
-            columns[0..(digits_count - 1)]
-                .iter()
-                .rev()
-                .fold(columns[digits_count - 1], |total, &p| {
-                    &total.mul_by_pow_2(w as u32) + &CachedPoint::from(p)
-                })
-                .into(),
+            columns.fold(hi_column, |total, p| {
+                &total.mul_by_pow_2(w as u32) + &CachedPoint::from(p)
+            }).into()
         )
     }
 }