Files
mars-libowfat/fmt/fmt_utf8.c
2025-01-20 19:00:58 +00:00

72 lines
2.0 KiB
C

#include <libowfat/fmt.h>
size_t fmt_utf8(char *dest,uint32_t n) {
size_t i,j;
if (n<=0x7f) {
if (dest) *dest=(char)n;
return 1;
}
for (i=0x3f,j=0; i<=0x7fffffff; i=(i<<5)|0x1f, ++j) {
if (i>=n) {
if (dest) {
size_t k=j*6;
// gcc -fanalyze warns here that j-1 might underflow, leading to
// undefined behavior because right shift by more than integer
// width is undefined. That can't happen because both n>0x7f and
// i>=n would have to be true and we initialize i as 0x3f and
// with each iteration j is incremented. It's a false positive.
*dest++=(char)(((signed char)0xc0 >> (j-1)) | (char)(n >> k));
while (k) {
*dest++=(char)(0x80 | ((n >> (k-6)) & 0x3f));
k-=6;
}
}
return j+1;
}
if (i==0x7fffffff) return 0;
}
/* we were asked to encode a value that cannot be encoded */
return 0;
}
#if defined(__i386__) || defined(__x86_64__)
// on architectures where unaligned writes are ok and don't carry a
// performance pentalty, it may make sense to not write each byte
// individually but write a 32-bit word at the end. The downside is that
// this will write scratch bytes over the bytes following the utf-8. If
// you allocated a few more bytes scratch space, and did not expect
// those bytes to be untouched, you can use this function.
// I did a small benchmark:
// for (i=j=0; i<1000; i+=fmt_utf8(buf+i,i));
// Regular fmt_utf8: 13k cpu cycles, fmt_utf_scratch: 11k cpu cycles.
size_t fmt_utf8_scratch(char *dest,uint32_t n) {
size_t i;
unsigned int j;
if (n<=0x7f) {
if (dest) *dest=(char)n;
return 1;
}
for (i=0x3f,j=0; i<=0x7fffffff; i=(i<<5)|0x1f, ++j) {
if (i>=n) {
if (dest) {
uint64_t buf=0;
while (n>0x3f) {
buf = (buf << 8) + (n & 0x3f) + 0x80;
n >>= 6;
}
buf = (buf << 8) + (unsigned char)((char)0xc0 >> (j-1)) + n;
*(uint64_t*)dest = buf;
}
return j+1;
}
if (i==0x7fffffff) return 0;
}
/* we were asked to encode a value that cannot be encoded */
return 0;
}
#endif
/* unit tested via scan/scan_utf8.c */