Avoid undefined behavior from memcpy for the portable build.

dec/decode.c

@@ -94,7 +94,16 @@ static uint32_t DecodeWindowBits(BrotliBitReader* br) {
 
 static BROTLI_INLINE BROTLI_NO_ASAN void memmove16(
     uint8_t* dst, uint8_t* src) {
-#ifdef __ARM_NEON__
+#if BROTLI_SAFE_MEMMOVE
+  /* For x86 this compiles to the same binary as signle memcpy.
+     On ARM memcpy is not inlined, so it works slower.
+     This implementation makes decompression 1% slower than regular one,
+     and 2% slower than NEON implementation.
+   */
+  uint32_t buffer[4];
+  memcpy(buffer, src, 16);
+  memcpy(dst, buffer, 16);
+#elif defined(__ARM_NEON__)
   vst1q_u8(dst, vld1q_u8(src));
 #else
   /* memcpy is unsafe for overlapping regions and ASAN detects this.

dec/port.h

@@ -17,7 +17,7 @@
 
    Build options are:
     * BROTLI_BUILD_PORTABLE disables dangerous optimizations, like unaligned
-      read and overlapping memcpy
+      read and overlapping memcpy; this reduces decompression speed by 5%
     * BROTLI_DEBUG dumps file name and line number when decoder detects stream
       or memory error
     * BROTLI_DECODE_DEBUG enables asserts and dumps various state information
@@ -43,8 +43,10 @@
 
 #ifdef BROTLI_BUILD_PORTABLE
 #define BROTLI_ALIGNED_READ 1
+#define BROTLI_SAFE_MEMMOVE 1
 #else
 #define BROTLI_ALIGNED_READ 0
+#define BROTLI_SAFE_MEMMOVE 0
 #endif
 
 #define BROTLI_ASAN_BUILD __has_feature(address_sanitizer)
@@ -136,7 +138,7 @@ OR:
 #define BROTLI_NOINLINE
 #endif
 
-#if BROTLI_ASAN_BUILD
+#if BROTLI_ASAN_BUILD && !defined(BROTLI_BUILD_PORTABLE)
 #define BROTLI_NO_ASAN __attribute__((no_sanitize("address"))) BROTLI_NOINLINE
 #else
 #define BROTLI_NO_ASAN