On Sun, 2005-03-27 at 15:51 +0300, Denis Vlasenko wrote: > > > It's impossible to be otherwise. A call requires > > > that the return address be written to memory (the stack), > > > using register indirection (the stack-pointer). > > > > and it's a so common pattern that it's optimized to death. Internally a > > call gets transformed to 2 uops or so, one is push eip, the other is the > > jmp (which gets then just absorbed by the "what is the next eip" logic, > > just as a "jmp"s are 0 cycles) > > Arjan, you overlook the fact that kfree() contains 'if(!p) return;' too. > call + test-and-branch can never be faster than test+and+branch
ok so for the non-null case you have test-nbranch-call-test-nbranch vs call-test-nbranch vs the null case where you get test-branch vs call-test-branch (I'm using nbranch here as a non-taken branch; it's also a conditional branch and it has the same misprediction possibility) in the non-null case with if you have *two* chances for the branch predictor to go wrong. (and "wrong" can also mean "cold, eg unknown" here) and always an extra "test-nbranch" sequence, which is probably a cycle at least the offset for that is the null-case-without-if where you have an extra "call", which is also half to a whole cycle. even in the null case it's dubious if there is gain, it depends on how the branch predictor happens to feel that day ;) - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/