The point is that calling a function is much more expensive than comparing a single byte. You'll have to fix your stack, jump to another location, check the first byte, unwind the stack, and continue your program.

EDIT: A quick test shows adding comparison of the first byte reducing execution time here from 2.6 seconds to 0.6 seconds.

(this is in reply to tptacek, I can't reply to his post?)

You're right, but this is such a "meh" microoptimization. Compared to the cost of a real memcmp(), a simple function call is below the noise floor.

I, and my data, disagree. Most memcmp() calls look at a single byte and return.

Let us consider five test cases:

memcmp0 - Calling memcmp() at each byte in the file looking ofr a match, as the original article.

memcmp1 - As above, but optimize by checking the first character and only calling if it matches.

memmem0 - A single call to the glibc memmem() function. Not portable.

kmp0 - The Knuth-Morris-Pratt algorithm as implemented at "Exact String Matching Algorithms", http://www-igm.univ-mlv.fr/~lecroq/string/index.html

bm0 - The Boyer Moore algorithm, also from above

Looking for a 1k needle which lies at the end of a 100MB haystack of random bytes, with the data in file system buffers on a 1.6GHz Atom processor running Linux 2.6.29:

  algorithm  milliseconds
    memcmp0      2400     ==================================
    memcmp1       530     ========
    memmem0       460     =======
    kmp0          970     ==============
    bm0            70     = 

On the whole codesink1 wins with the Boyer Moore suggestion. In broader terms, algorithm selection beats optimization and all is right with the CS universe. (KMP's poor showing surprises me, but I've not used it before and don't know what to expect.)

If we assume the data is out on a hard drive, then we get these numbers (same machine, a WD Green Power HD which is an evil thing to benchmark with because it has variable spindle and seek speeds, but suffice it to say it is not terribly fast). All OS caches flushed before each test.

  algorithm  milliseconds
    memcmp0      3500     ===================================
    memcmp1      1300     =============
    memmem0      1200     ============
    kmp0         1700     =================
    bm0          1100     =========== 

Here disk read time sets a lower bound, algorithm is still important, but not so much as before. I have a sneaky suspicion that mmap() is a poor choice here. There doesn't seem to be much overlap between the IO and compute when looking at the memcmp0 times.

Testing on a beefier machine, with a Core 2 Duo 2.8GHz, OS X gives this:

  algorithm  milliseconds
    memcmp0      1100     =========================
    memcmp1       180     ====
    memmem0       not portable
    kmp0          360     ========
    bm0            88     == 

The processor is several times faster, but bringing pages in from the OS is slightly slower. I don't know how to flush OS buffers on this machine so I can't do the disk reading tests.

Update: well crap. this fell of the front page before I finished testing. No one will see these results, but I enjoyed the tests anyway. It is useful to update my rules of thumb and preconceived notions.

I'm impressed with the benchmark, but it seems to me like using the right algorithm for the job (in this case, Boyer-Moore or memmem) beats hacking memcmp to do a job it shouldn't be employed to do.

I see it, and I'm impressed. I'll link to it for you!

very good analysis, thanks for sharing that

I'm having a hard time believing that a benchmark that produces a 300% speedup just from eliminating CALL and RET could be realistic. The cost of accessing cached memory on the program stack is close to the cost of accessing a register.