Cache p.sig_words() in curve object

Avoid using Barett reduction in core operations; seems to help perf.

1 files changed, 9 insertions, 0 deletions

diff --git a/src/math/numbertheory/curve_gfp.h b/src/math/numbertheory/curve_gfp.h
index a7be8987c..0a91fc52d 100644
--- a/src/math/numbertheory/curve_gfp.h
+++ b/src/math/numbertheory/curve_gfp.h
@@ -44,6 +44,8 @@ class BOTAN_DLL CurveGFp
          p_dash = (((r * r_inv) - 1) / p).word_at(0);
 
          a_r = reducer_p.multiply(a, r);
+
+         p_words = p.sig_words();
          }
 
       // CurveGFp(const CurveGFp& other) = default;
@@ -87,6 +89,11 @@ class BOTAN_DLL CurveGFp
       */
       word get_p_dash() const { return p_dash; }
 
+      /**
+      * @return p.sig_words()
+      */
+      u32bit get_p_words() const { return p_words; }
+
       const Modular_Reducer& mod_p() const { return reducer_p; }
 
       /**
@@ -114,6 +121,8 @@ class BOTAN_DLL CurveGFp
       // Curve parameters
       BigInt p, a, b;
 
+      u32bit p_words; // cache of p.sig_words()
+
       // Montgomery parameters
       BigInt r, r_inv, a_r;
       word p_dash;