This function was found here. It's an implementation of strcmp:
int strcmp(const char* s1, const char* s2) { while (*s1 && (*s1 == *s2)) s1++, s2++; return *(const unsigned char*)s1 - *(const unsigned char*)s2; } I understand all but the last line, in short what is going on in the last line?
return *(const unsigned char*)s1-*(const unsigned char*)s2; OP: in short what is going on in the last line?
A: The first potential string difference is compared. Both chars are referenced as unsigned char as required by the spec. The 2 are promoted to int and the difference is returned.
Notes:
1 The return value's sign (<0, 0, >0) is the most meaningful part. It is the only part that is specified by the C spec.
2 On some systems char is signed (more common). On others, char is unsigned. Defining the "sign-ness" of the last comparison promotes portability. Note that fgetc() obtains characters as unsigned char.
3 Other than that a string ends with a \0, the character encoding employed (like ASCII - most common), makes no difference at the binary level. If the first chars that differ in 2 strings have values 65 and 97, the first string will be less than the second, even if the character encoding is non-ASCII. OTOH, strcmp("A", "a") will return a negative number when character encoding is ASCII, but may return a positive number in a different character encoding for their underlying value and order are not defined by C.
This implementation is definitely not optimization of the built-in strcmp, it is simply another implementation and I believe it will most probably perform worse than the built-in version.
A comparison function is supposed to return 0 if the values being compared are equal, any negative number if the first value is smaller and any positive number if the first value is greater. And that is what happens on the last line.
The idea of the last line is to cast the characters to unsigned chars and I believe the author meant for this to sort non-standard characters after the standard ones(ASCII codes 0-127).
EDIT: there is no bug in the code and it can and will return negative values if value pointed to by s1 is smaller than the value pointed to by s2 ordering standard characters before characters with code 128 and above.
I'm preffer this code:
int strcmp(const char *str1, const char *str2) { int s1; int s2; do { s1 = *str1++; s2 = *str2++; if (s1 == 0) break; } while (s1 == s2); return (s1 < s2) ? -1 : (s1 > s2); } for ARMv4 it compiled as:
strcmp: ldrsb r3, [r0], #1 ;r3 = *r0++ ldrsb r2, [r1], #1 ;r2 = *r1++ cmp r3, #0 ;compare r3 and 0 beq @1 ;if r3 == 0 goto @1 cmp r3, r2 ;compare r3 and r2 beq strcmp ;if r3 == r2 goto strcmp ;loop is ended @1: cmp r3, r2 ;compare r3 and r2 blt @2 ;if r3 < r2 goto @2 movgt r0, #1 ;if r3 > r2 r0 = 1 movle r0, #0 ;if r3 <= r2 r0 = 0 bx lr ;return r0 @2: mov r0, #-1 ;r0 = -1 bx lr ;return r0 As you can see there is only 6 instructions under the loop + atmost 5 instructions at the end. So complexity is 6 * (strlen+1) + 5.
Moving (s1 == 0) to the while condition causes worse machine code for ARM (I do not know why).
s1 = (unsigned char)*str1++; to implement the exact semantics of strcmp().This implementation can be further optimized, shaving off some comparisons:
int strcmp(const char *s1, const char *s2) { unsigned char c1, c2; while ((c1 = *s1++) == (c2 = *s2++)) { if (c1 == '\0') return 0; } return c1 - c2; } The return value is 0 if the string are identical up to and including the terminating null byte. The sign of the return value is that of the difference between the first differing characters, converted to unsigned char as per the C Standard.
If char is smaller than int, which is true on all but some rare embedded systems, this difference can be computed with a simple subtraction, both c1 and c2 being promoted to int and this difference is guaranteed to fit in the range of type int.
On systems where sizeof(int) == 1, the return value should be computed this way:
return (c1 < c2) ? -1 : 1; strcmp returns which string is greater then the other, not just whether they are equal.
The last line subtracts the first non-matching character to see which is larger. If the whole string matches then it will be 0-0=0 which gives the "equal" result.
This implementation is not really well optimized, as that would take assembly code and knowledge of cache-lines, load sizes etc.
