Re: towlower performance

[Brad Conroy <technosaurus@yahoo.com>]

> Thankfully, though, I think the issue maksis reported is more a matter
> of the towlower/towupper implementation being oriented towards
> size

The current macros in ctype.h evaluate twice and could all really be inlined.
For a test case try building strcasecmp with and without tolower inlined.
It is actually smaller inlined because it doesn't need to setup for a call.

I  try to avoid branching in inlined functions to avoid mis-predictions and to
keep constant time as well as better vectorization (although the unsigned
comparison trick may prevent automatic conversion to SIMD on x86 since
there is no unsigned compare), so here are some branchless static inlined
versions if you want to use them:

static __inline int isalnum(int c){
	return ((unsigned)c-'0' < 10)|(((unsigned)c|32)-'a' < 26);
}
static __inline int isalpha(int c){
	return (((unsigned)c|32)-'a' < 26);
}
static __inline int isascii(int c){
	return (unsigned)c<128;
}
static __inline int isblank(int c){
	return (c==' ')|(c=='\t');
}
static __inline int iscntrl(int c){
	return ((unsigned)c < 0x20) | (c == 0x7f);
}
static __inline int isdigit(int c){
	return (unsigned)c-'0' < 10;
}
static __inline int isgraph(int c){
	return (unsigned)c-0x21 < 0x5e;
}
static __inline int islower(int c){
	return (unsigned)c-'a' < 26;
}
static __inline int isprint(int c){
	return (unsigned)c-0x20 < 0x5f;
}
static __inline int ispunct(int c){
	unsigned y=(unsigned)c;
	return ( ( y-'!' < 95 ) & ( ((y|32)-'a' > 25) | (y-'0' > 9) ) );
}
static __inline int isspace(int c){
	return ((unsigned)c-'\t' < 5)|(c == ' ');
}
static __inline int isupper(int c){
	return (unsigned)c-'A' < 26;
}
static __inline int isxdigit(int c){
	return ((unsigned)c-'0' < 10) | (((unsigned)c|32)-'a' < 6);
}
static __inline int toascii(int c){
	return c & 0x7f;
}
static __inline int tolower(int c){
	return c | ( ((unsigned)c-'A' < 26)<<5 );
}
static __inline int toupper(int c){
	return c & ~(((unsigned)c-'a' < 26)<<5);
}