From 2d0d87b57ad227324ac59fee0ac58808050cc239 Mon Sep 17 00:00:00 2001
From: Jeffrey Walton <noloader@gmail.com>
Date: Sat, 23 Jun 2018 13:47:59 -0400
Subject: Use pre-splatted key table for LEA NEON and Aarch64

---
 lea-simd.cpp | 12 +++++++++++-
 lea.cpp      | 13 +++++++++----
 2 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/lea-simd.cpp b/lea-simd.cpp
index a43b719a..8dacdf56 100644
--- a/lea-simd.cpp
+++ b/lea-simd.cpp
@@ -140,7 +140,7 @@ uint32x4_t UnpackHigh64(uint32x4_t a, uint32x4_t b)
 template <unsigned int IDX>
 inline uint32x4_t LoadKey(const word32 rkey[])
 {
-    return vdupq_n_u32(rkey[IDX]);
+    return vld1q_u32(&rkey[IDX*4]);
 }
 
 template <unsigned int IDX>
@@ -1012,6 +1012,16 @@ size_t LEA_Dec_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t rounds,
 #endif // CRYPTOPP_SSSE3_AVAILABLE
 
 #if defined(CRYPTOPP_ARM_NEON_AVAILABLE)
+void LEA_SplatKeys_NEON(SecBlock<word32>& rkeys)
+{
+    SecBlock<word32> temp(rkeys.size() * 4);
+    for (size_t i=0, j=0; i<rkeys.size(); i++, j+=4)
+    {
+		vst1q_u32(&temp[j], vdupq_n_u32(rkeys[i]));
+    }
+    std::swap(rkeys, temp);
+}
+
 size_t LEA_Enc_AdvancedProcessBlocks_NEON(const word32* subKeys, size_t rounds,
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
 {
diff --git a/lea.cpp b/lea.cpp
index 35149157..122313da 100644
--- a/lea.cpp
+++ b/lea.cpp
@@ -567,6 +567,8 @@ extern size_t LEA_Dec_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t
 # endif
 
 # if (CRYPTOPP_ARM_NEON_AVAILABLE)
+extern void LEA_SplatKeys_NEON(SecBlock<word32>& rkeys);
+
 extern size_t LEA_Enc_AdvancedProcessBlocks_NEON(const word32* subKeys, size_t rounds,
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);
 
@@ -606,13 +608,16 @@ void LEA::Base::UncheckedSetKey(const byte *userKey, unsigned int keyLength, con
         CRYPTOPP_ASSERT(0);;
     }
 
+    // If we pre-splat the round keys at setup then we avoid a shuffle
+    // at runtime for each subkey used during encryption and decryption.
+    // Pre-splatting saves about 0.7 to 1.0 cpb at the cost of 4x storage.
 #if (CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS) && (CRYPTOPP_SSSE3_AVAILABLE)
     if (HasSSSE3())
-    {
-        // If we pre-splat the round keys at setup then we avoid a shuffle
-        // at runtime for each subkey used during encryption and decryption.
         LEA_SplatKeys_SSSE3(m_rkey);
-    }
+#endif
+#if (CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS) && (CRYPTOPP_ARM_NEON_AVAILABLE)
+    if (HasNEON())
+        LEA_SplatKeys_NEON(m_rkey);
 #endif
 }
 
-- 
cgit v1.2.1