wip(fsst): support big-endian systems

Author: mhx

fsst/libfsst.cpp

@@ -529,11 +529,11 @@ extern "C" u32 fsst_export(fsst_encoder_t *encoder, u8 *buf) {
    // The version field is now there just for future-proofness, but not used yet
 
    // version allows keeping track of fsst versions, track endianness, and encoder reconstruction
-   u64 version = (FSST_VERSION << 32) |  // version is 24 bits, most significant byte is 0 
+   dwarfs::uint64le_t version((FSST_VERSION << 32) |  // version is 24 bits, most significant byte is 0 
                  (((u64) e->symbolTable->suffixLim) << 24) | 
                  (((u64) e->symbolTable->terminator) << 16) | 
                  (((u64) e->symbolTable->nSymbols) << 8) | 
-                 FSST_ENDIAN_MARKER; // least significant byte is nonzero
+                 FSST_ENDIAN_MARKER); // least significant byte is nonzero
 
    /* do not assume unaligned reads here */
    memcpy(buf, &version, 8);
@@ -553,7 +553,7 @@ extern "C" u32 fsst_export(fsst_encoder_t *encoder, u8 *buf) {
 #define FSST_CORRUPT 32774747032022883 /* 7-byte number in little endian containing "corrupt" */
 
 extern "C" u32 fsst_import(fsst_decoder_t *decoder, u8 const *buf) {
-   u64 version = 0;
+   dwarfs::uint64le_t version;
    u32 code, pos = 17;
    u8 lenHisto[8];
 

fsst/libfsst.hpp

@@ -32,6 +32,8 @@
 #include <fcntl.h>
 #include <stddef.h>
 
+#include <dwarfs/endian.h>
+
 using namespace std;
 
 #include "fsst.h" // the official FSST API -- also usable by C mortals
@@ -61,7 +63,7 @@ typedef uint64_t u64;
 
 namespace libfsst {
 inline uint64_t fsst_unaligned_load(u8 const* V) {
-    uint64_t Ret;
+    dwarfs::uint64le_t Ret;
     memcpy(&Ret, V, sizeof(uint64_t)); // compiler will generate efficient code (unaligned load, where possible)
     return Ret;
 }
@@ -70,7 +72,7 @@ struct Symbol {
    static const unsigned maxLength = 8;
 
    // the byte sequence that this symbol stands for
-   union { char str[maxLength]; u64 num; } val; // usually we process it as a num(ber), as this is fast
+   union { char str[maxLength]; dwarfs::uint64le_t num{}; } val; // usually we process it as a num(ber), as this is fast
 
    // icl = u64 ignoredBits:16,code:12,length:4,unused:32 -- but we avoid exposing this bit-field notation
    u64 icl;  // use a single u64 to be sure "code" is accessed with one load and can be compared with one comparison