Use carryless multiplies for NIST b233 and k233 curves (GH #783, PR #784)

Use carryless multiplies for NIST b233 and k233 curves.

1 files changed, 16 insertions, 19 deletions

diff --git a/sha_simd.cpp b/sha_simd.cpp
index 7184f3e1..0b31ea37 100644
--- a/sha_simd.cpp
+++ b/sha_simd.cpp
@@ -46,17 +46,6 @@
 # define EXCEPTION_EXECUTE_HANDLER 1

 #endif

 

-// Thanks to Peter Cordes, https://stackoverflow.com/q/54016821/608639

-#if (CRYPTOPP_ARM_NEON_AVAILABLE)

-# ifndef PACK32x4

-#  if defined(_MSC_VER)

-#   define PACK32x4(w,x,y,z) { ((w) + (word64(x) << 32)), ((y) + (word64(z) << 32)) }

-#  else

-#   define PACK32x4(w,x,y,z) { (w), (x), (y), (z) }

-#  endif

-# endif  // PACK32x4

-#endif // Microsoft workaround

-

 // Clang __m128i casts

 #define M128_CAST(x) ((__m128i *)(void *)(x))

 #define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))

@@ -95,9 +84,10 @@ bool CPU_ProbeSHA1()
     volatile bool result = true;

     __try

     {

-        uint32x4_t data1 = PACK32x4(1,2,3,4);

-        uint32x4_t data2 = PACK32x4(5,6,7,8);

-        uint32x4_t data3 = PACK32x4(9,10,11,12);

+        unsigned int w[] = {1,2,3,4, 5,6,7,8, 9,10,11,12};

+        uint32x4_t data1 = vld1q_u32(w+0);

+        uint32x4_t data2 = vld1q_u32(w+4);

+        uint32x4_t data3 = vld1q_u32(w+8);

 

         uint32x4_t r1 = vsha1cq_u32 (data1, 0, data2);

         uint32x4_t r2 = vsha1mq_u32 (data1, 0, data2);

@@ -130,7 +120,10 @@ bool CPU_ProbeSHA1()
         result = false;

     else

     {

-        uint32x4_t data1 = {1,2,3,4}, data2 = {5,6,7,8}, data3 = {9,10,11,12};

+        unsigned int w[] = {1,2,3,4, 5,6,7,8, 9,10,11,12};

+        uint32x4_t data1 = vld1q_u32(w+0);

+        uint32x4_t data2 = vld1q_u32(w+4);

+        uint32x4_t data3 = vld1q_u32(w+8);

 

         uint32x4_t r1 = vsha1cq_u32 (data1, 0, data2);

         uint32x4_t r2 = vsha1mq_u32 (data1, 0, data2);

@@ -159,9 +152,10 @@ bool CPU_ProbeSHA2()
     volatile bool result = true;

     __try

     {

-        uint32x4_t data1 = PACK32x4(1,2,3,4);

-        uint32x4_t data2 = PACK32x4(5,6,7,8);

-        uint32x4_t data3 = PACK32x4(9,10,11,12);

+        unsigned int w[] = {1,2,3,4, 5,6,7,8, 9,10,11,12};

+        uint32x4_t data1 = vld1q_u32(w+0);

+        uint32x4_t data2 = vld1q_u32(w+4);

+        uint32x4_t data3 = vld1q_u32(w+8);

 

         uint32x4_t r1 = vsha256hq_u32 (data1, data2, data3);

         uint32x4_t r2 = vsha256h2q_u32 (data1, data2, data3);

@@ -193,7 +187,10 @@ bool CPU_ProbeSHA2()
         result = false;

     else

     {

-        uint32x4_t data1 = {1,2,3,4}, data2 = {5,6,7,8}, data3 = {9,10,11,12};

+        unsigned int w[] = {1,2,3,4, 5,6,7,8, 9,10,11,12};

+        uint32x4_t data1 = vld1q_u32(w+0);

+        uint32x4_t data2 = vld1q_u32(w+4);

+        uint32x4_t data3 = vld1q_u32(w+8);

 

         uint32x4_t r1 = vsha256hq_u32 (data1, data2, data3);

         uint32x4_t r2 = vsha256h2q_u32 (data1, data2, data3);