added XXH_REROLL consistency test

to catch issues such as #240.
Also added XXH64 test for len==4

.travis.yml

@@ -22,6 +22,8 @@ matrix:
         - CPPFLAGS=-DXXH_VECTOR=1 make check   # SSE2 code path
         - make clean
         - CPPFLAGS="-mavx2 -DXXH_VECTOR=2" make check   # AVX2 code path
+        - make clean
+        - CPPFLAGS=-DXXH_REROLL=1 make check   # reroll code path (#240)
         - make -C tests/bench
 
     - name: ARM + aarch64 compilation and consistency checks

README.md

@@ -72,7 +72,7 @@ they modify libxxhash behavior. They are all disabled by default.
                      with performance improvements observed in the +200% range .
                      See [this article](https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html) for details.
                      Note: there is no need for an `xxhash.o` object file in this case.
-- `XXH_REROLL` : reduce size of generated code
+- `XXH_REROLL` : reduce size of generated code. Impact on performance vary, depending on platform and algorithm.
 - `XXH_ACCEPT_NULL_INPUT_POINTER` : if set to `1`, when input is a `NULL` pointer,
                                     xxhash result is the same as a zero-length input
                                     (instead of a dereference segfault).

xxhsum.c

@@ -739,6 +739,7 @@ static void BMK_sanityCheck(void)
     BMK_testXXH64(NULL        ,  0, prime, 0xAC75FDA2929B17EFULL);
     BMK_testXXH64(sanityBuffer,  1, 0,     0xE934A84ADB052768ULL);
     BMK_testXXH64(sanityBuffer,  1, prime, 0x5014607643A9B4C3ULL);
+    BMK_testXXH64(sanityBuffer,  4, 0,     0x9136A0DCA57457EEULL);
     BMK_testXXH64(sanityBuffer, 14, 0,     0x8282DCC4994E35C8ULL);
     BMK_testXXH64(sanityBuffer, 14, prime, 0xC3BD6BF63DEB6DF0ULL);
     BMK_testXXH64(sanityBuffer,222, 0,     0xB641AE8CB691C174ULL);