Fix SSE2/SSSE3 with large blocks

Closes #5

Author: randombit

fecpp.cpp

@@ -173,12 +173,20 @@ void addmul(uint8_t z[], const uint8_t x[], uint8_t y, size_t size)
 #if defined(FECPP_IS_X86)
    if(size >= 16 && has_ssse3())
       {
-      size = addmul_ssse3(z, x, y, size);
+      const size_t consumed = addmul_ssse3(z, x, y, size);
+      z += consumed;
+      x += consumed;
+      y += consumed;
+      size -= consumed;
       }
 
    if(size >= 64 && has_sse2())
       {
-      size = addmul_sse2(z, x, y, size);
+      const size_t consumed = addmul_sse2(z, x, y, size);
+      z += consumed;
+      x += consumed;
+      y += consumed;
+      size -= consumed;
       }
 #endif
 

fecpp_sse2.cpp

@@ -15,6 +15,8 @@ size_t addmul_sse2(uint8_t z[], const uint8_t x[], uint8_t y, size_t size)
 
    const size_t y_bits = 32 - __builtin_clz(y);
 
+   const size_t consumed = size - (size % 64);
+
    // unrolled out to cache line size
    while(size >= 64)
       {
@@ -94,7 +96,7 @@ size_t addmul_sse2(uint8_t z[], const uint8_t x[], uint8_t y, size_t size)
       size -= 64;
       }
 
-   return size;
+   return consumed;
    }
 
 }

fecpp_ssse3.cpp

@@ -541,6 +541,8 @@ size_t addmul_ssse3(uint8_t z[], const uint8_t x[], uint8_t y, size_t size)
    const __m128i t_lo = _mm_load_si128((const __m128i*)(GFTBL + 32*y));
    const __m128i t_hi = _mm_load_si128((const __m128i*)(GFTBL + 32*y + 16));
 
+   const size_t consumed = size - (size % 16);
+
    while(size >= 16)
       {
       const __m128i x_1 = _mm_loadu_si128((const __m128i*)(x));
@@ -565,7 +567,7 @@ size_t addmul_ssse3(uint8_t z[], const uint8_t x[], uint8_t y, size_t size)
       size -= 16;
       }
 
-   return size;
+   return consumed;
    }
 
 }