[Memory] Fix alignment to 16 bytes

Fix segmentation fault happening on linux when `src` or `dst` is not aligned to
16 bytes.
Use `_mm_storeu_si128` as there is no guarentee that `dest` is propely aligned.
Grab all of words/dwords which are unaligned with 16 bytes.

Author: bwrsandman

src/xenia/base/memory.cc

@@ -34,15 +34,16 @@ void copy_and_swap_16_aligned(void* dest_ptr, const void* src_ptr,
                    0x04, 0x05, 0x02, 0x03, 0x00, 0x01);
 
   size_t i = 0;
-  size_t unaligned_words = (reinterpret_cast<uintptr_t>(src_ptr) & 0xF) / 2;
+  size_t unaligned_words =
+      (0x10 - (reinterpret_cast<uintptr_t>(src_ptr) & 0xF)) / sizeof(uint16_t);
   for (; unaligned_words > 0 && i < count; unaligned_words--, i++) {
     // Copy up to 16 byte alignment.
     dest[i] = byte_swap(src[i]);
   }
   for (; i + 8 <= count; i += 8) {
     __m128i input = _mm_load_si128(reinterpret_cast<const __m128i*>(&src[i]));
     __m128i output = _mm_shuffle_epi8(input, shufmask);
-    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
   }
   for (; i < count; ++i) {  // handle residual elements
     dest[i] = byte_swap(src[i]);
@@ -78,15 +79,16 @@ void copy_and_swap_32_aligned(void* dest_ptr, const void* src_ptr,
                    0x06, 0x07, 0x00, 0x01, 0x02, 0x03);
 
   size_t i = 0;
-  size_t unaligned_dwords = (reinterpret_cast<uintptr_t>(src_ptr) & 0xF) / 4;
+  size_t unaligned_dwords =
+      (0x10 - (reinterpret_cast<uintptr_t>(src_ptr) & 0xF)) / sizeof(uint32_t);
   for (; unaligned_dwords > 0 && i < count; unaligned_dwords--, i++) {
     // Copy up to 16 byte alignment.
     dest[i] = byte_swap(src[i]);
   }
   for (; i + 4 <= count; i += 4) {
     __m128i input = _mm_load_si128(reinterpret_cast<const __m128i*>(&src[i]));
     __m128i output = _mm_shuffle_epi8(input, shufmask);
-    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
   }
   for (; i < count; ++i) {  // handle residual elements
     dest[i] = byte_swap(src[i]);
@@ -122,15 +124,16 @@ void copy_and_swap_64_aligned(void* dest_ptr, const void* src_ptr,
                    0x02, 0x03, 0x04, 0x05, 0x06, 0x07);
 
   size_t i = 0;
-  size_t unaligned_qwords = (reinterpret_cast<uintptr_t>(src_ptr) & 0xF) / 8;
+  size_t unaligned_qwords =
+      (reinterpret_cast<uintptr_t>(src_ptr) & 0xF) / sizeof(uint64_t);
   for (; unaligned_qwords > 0 && i < count; unaligned_qwords--, i++) {
     // Copy up to 16 byte alignment.
     dest[i] = byte_swap(src[i]);
   }
   for (; i + 2 <= count; i += 2) {
     __m128i input = _mm_load_si128(reinterpret_cast<const __m128i*>(&src[i]));
     __m128i output = _mm_shuffle_epi8(input, shufmask);
-    _mm_store_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(&dest[i]), output);
   }
   for (; i < count; ++i) {  // handle residual elements
     dest[i] = byte_swap(src[i]);