added documentation about LZ4_FORCE_SW_BITCOUNT

Also : added memory-frugal software byte count for big endian 64-bit cpus.
Disabled by default.
diff --git a/lib/README.md b/lib/README.md
index 707d777..3653c81 100644
--- a/lib/README.md
+++ b/lib/README.md
@@ -46,11 +46,11 @@
 
 #### Build macros
 
-The following build macro can be selected at compilation time :
+The following build macro can be selected to adjust source code behavior at compilation time :
 
-- `LZ4_FAST_DEC_LOOP` : this triggers the optimized decompression loop.
-  This loops works great on x86/x64 cpus, and is automatically enabled on this platform.
-  It's possible to enable or disable it manually, by passing `LZ4_FAST_DEC_LOOP=1` or `0` to the preprocessor.
+- `LZ4_FAST_DEC_LOOP` : this triggers a speed optimized decompression loop, more powerful on modern cpus.
+  This loop works great on x86, x64 and aarch64 cpus, and is automatically enabled for them.
+  It's also possible to enable or disable it manually, by passing `LZ4_FAST_DEC_LOOP=1` or `0` to the preprocessor.
   For example, with `gcc` : `-DLZ4_FAST_DEC_LOOP=1`,
   and with `make` : `CPPFLAGS+=-DLZ4_FAST_DEC_LOOP=1 make lz4`.
 
@@ -66,9 +66,17 @@
   Should this be a problem, it's generally possible to make the compiler ignore these warnings,
   for example with `-Wno-deprecated-declarations` on `gcc`,
   or `_CRT_SECURE_NO_WARNINGS` for Visual Studio.
-  Another method is to define `LZ4_DISABLE_DEPRECATE_WARNINGS`
+  Another project-specific method is to define `LZ4_DISABLE_DEPRECATE_WARNINGS`
   before including the LZ4 header files.
 
+- `LZ4_FORCE_SW_BITCOUNT` : by default, the compression algorithm tries to determine lengths
+  by using bitcount instructions, generally implemented as fast single instructions in many cpus.
+  In case the target cpus doesn't support it, or compiler intrinsic doesn't work, or feature bad performance,
+  it's possible to use an optimized software path instead.
+  This is achieved by setting this build macros .
+  In most cases, it's not expected to be necessary,
+  but it can be legitimately considered for less common platforms.
+
 
 #### Amalgamation
 
@@ -103,7 +111,7 @@
 
 #### Miscellaneous
 
-Other files present in the directory are not source code. There are :
+Other files present in the directory are not source code. They are :
 
  - `LICENSE` : contains the BSD license text
  - `Makefile` : `make` script to compile and install lz4 library (static and dynamic)
diff --git a/lib/lz4.c b/lib/lz4.c
index e10b58e..0ca7b21 100644
--- a/lib/lz4.c
+++ b/lib/lz4.c
@@ -88,6 +88,7 @@
  * Define this parameter if your target system or compiler does not support hardware bit count
  */
 #if defined(_MSC_VER) && defined(_WIN32_WCE)   /* Visual Studio for WinCE doesn't support Hardware bit count */
+#  undef  LZ4_FORCE_SW_BITCOUNT  /* avoid double def */
 #  define LZ4_FORCE_SW_BITCOUNT
 #endif
 
@@ -527,6 +528,9 @@
                                         !defined(LZ4_FORCE_SW_BITCOUNT)
             return (unsigned)__builtin_clzll((U64)val) >> 3;
 #       else
+#if 1
+            /* this method is probably faster,
+             * but adds a 128 bytes lookup table */
             static const unsigned char ctz7_tab[128] = {
                 7, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
                 4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
@@ -537,9 +541,22 @@
                 5, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
                 4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
             };
-            const U64 mask = 0x0101010101010101ULL;
-            U64 t = (((val >> 8) - mask) | val) & mask;
+            U64 const mask = 0x0101010101010101ULL;
+            U64 const t = (((val >> 8) - mask) | val) & mask;
             return ctz7_tab[(t * 0x0080402010080402ULL) >> 57];
+#else
+            /* this method doesn't consume memory space like the previous one,
+             * but it contains several branches,
+             * that may end up slowing execution */
+            static const U32 by32 = sizeof(val)*4;  /* 32 on 64 bits (goal), 16 on 32 bits.
+            Just to avoid some static analyzer complaining about shift by 32 on 32-bits target.
+            Note that this code path is never triggered in 32-bits mode. */
+            unsigned r;
+            if (!(val>>by32)) { r=4; } else { r=0; val>>=by32; }
+            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
+            r += (!val);
+            return r;
+#endif
 #       endif
         } else /* 32 bits */ {
 #       if (defined(__clang__) || (defined(__GNUC__) && ((__GNUC__ > 3) || \