From d9797c414a95403364db9f834a2e59a7ac41e5d8 Mon Sep 17 00:00:00 2001
From: Jeffrey Walton <noloader@gmail.com>
Date: Sat, 23 Jun 2018 14:19:18 -0400
Subject: [PATCH] Add CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS for LEA

---
 lea-simd.cpp | 8 +++++++-
 lea.cpp      | 4 ++++
 lea.h        | 8 ++++++++
 3 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/lea-simd.cpp b/lea-simd.cpp
index 8dacdf56..703930c9 100644
--- a/lea-simd.cpp
+++ b/lea-simd.cpp
@@ -140,7 +140,11 @@ uint32x4_t UnpackHigh64(uint32x4_t a, uint32x4_t b)
 template <unsigned int IDX>
 inline uint32x4_t LoadKey(const word32 rkey[])
 {
+#if (CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS)
     return vld1q_u32(&rkey[IDX*4]);
+#else
+    return vdupq_n_u32(rkey[IDX]);
+#endif
 }
 
 template <unsigned int IDX>
@@ -1012,15 +1016,17 @@ size_t LEA_Dec_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t rounds,
 #endif // CRYPTOPP_SSSE3_AVAILABLE
 
 #if defined(CRYPTOPP_ARM_NEON_AVAILABLE)
+# if (CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS)
 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]));
+        vst1q_u32(&temp[j], vdupq_n_u32(rkeys[i]));
     }
     std::swap(rkeys, temp);
 }
+# endif  // CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS
 
 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 122313da..75de3556 100644
--- a/lea.cpp
+++ b/lea.cpp
@@ -567,7 +567,9 @@ extern size_t LEA_Dec_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t
 # endif
 
 # if (CRYPTOPP_ARM_NEON_AVAILABLE)
+#  if (CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS)
 extern void LEA_SplatKeys_NEON(SecBlock<word32>& rkeys);
+#  endif  // CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS
 
 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);
@@ -616,8 +618,10 @@ void LEA::Base::UncheckedSetKey(const byte *userKey, unsigned int keyLength, con
         LEA_SplatKeys_SSSE3(m_rkey);
 #endif
 #if (CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS) && (CRYPTOPP_ARM_NEON_AVAILABLE)
+# if (CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS)
     if (HasNEON())
         LEA_SplatKeys_NEON(m_rkey);
+# endif  // CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS
 #endif
 }
 
diff --git a/lea.h b/lea.h
index 02fccd45..5b9f74b0 100644
--- a/lea.h
+++ b/lea.h
@@ -19,6 +19,14 @@
 # define CRYPTOPP_LEA_ADVANCED_PROCESS_BLOCKS 1
 #endif
 
+// Define this if you want to pre-splat the round key table
+// for NEON and Aarch64. Pre-splatting the round key increases
+// performance by about 0.7 cpb on ARM server boards like an
+// AMD Opteron A1100. However, it crushes performance on ARM
+// dev-boards like LeMaker HiKey and Pine64. HiKey and Pine64
+// run about 8 cpb slower when pre-splatting the round keys.
+// # define CRYPTOPP_LEA_ARM_SPLAT_ROUNDKEYS 1
+
 NAMESPACE_BEGIN(CryptoPP)
 
 /// \brief LEA block cipher information