Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

`three = 1` in the linux sourcecode

github.com

Post reply on HN
Print view View original

81–83 of 83 posts

Re: `three = 1` in the linux sourcecode

#81

Post by jcalvinowens »

userbinator wrote: post #80
jcalvinowens wrote: post #77

Earlier quoted context omitted.

> or use an array That's absolutely horrible: you've now introduced the possibility of accidental out-of-bounds access where there previously was none, not to mention that you just forced the compiler to put those variables on the stack instead of using registers. Contrived example: int array(int lol) { unsigned counter[3] = {1,5,7}; for (int i = 0; i The output is quite different (gcc -c -O3 -std=gnu99): Disassembly…

This is filesystem code. Any miniscule performance gain would be eclipsed by the actual reads/writes. This code should really be -Os'd to save cache.

You're missing the point: it's not that using an array is slower, it's that "int n[3]" can mean something completely different than "int a,b,c" does. A lot of comments on this thread have suggested that the two are interchangeable: they aren't.

Using an array here is wrong: it needlessly complicates the code (now in addition to everything else, I have to remember how big the array is and what each (unnamed) index corresponds to), possibly makes the compiler emit worthless loads/stores, and let's me potentially blow up the world if I mess up and access out-of-bounds.

Re: `three = 1` in the linux sourcecode

#82

Post by jakobe »

jcalvinowens wrote: post #77
jakobe wrote: post #46

Lot's of people suggesting to add explanatory comments. I don't think that's a good idea, it would be better to change the names of the variables: unsigned counter1 = 1; unsigned counter2 = 5; unsigned counter3 = 7; or use an array: unsigned counter[3] = {1,5,7}; No more confusion. That being said, I doubt that the names of these local variables is actually a real source of confusion and is not a problem that needs t…

> or use an array That's absolutely horrible: you've now introduced the possibility of accidental out-of-bounds access where there previously was none, not to mention that you just forced the compiler to put those variables on the stack instead of using registers. Contrived example: int array(int lol) { unsigned counter[3] = {1,5,7}; for (int i = 0; i The output is quite different (gcc -c -O3 -std=gnu99): Disassembly…

Your example has nothing to do with the actual code in the kernel. In the original code, the counters are passed by reference to a function. That means they need to be in memory anyway, right? The compiler can't pass the address of a register.

Re: `three = 1` in the linux sourcecode

#83

Post by jcalvinowens »

jakobe wrote: post #82
jcalvinowens wrote: post #77

Earlier quoted context omitted.

> or use an array That's absolutely horrible: you've now introduced the possibility of accidental out-of-bounds access where there previously was none, not to mention that you just forced the compiler to put those variables on the stack instead of using registers. Contrived example: int array(int lol) { unsigned counter[3] = {1,5,7}; for (int i = 0; i The output is quite different (gcc -c -O3 -std=gnu99): Disassembly…

Your example has nothing to do with the actual code in the kernel. In the original code, the counters are passed by reference to a function. That means they need to be in memory anyway, right? The compiler can't pass the address of a register.

> That means they need to be in memory anyway, right?

Actually no, GCC can optimize out pointer indirection like that if it ends up inlining the function (which it could in the kernel example, since it's static, relatively short, and has only three callers).

Another contrived example:

  static int inlineme(int *lol, int lolol)
  {
        *lol += lolol;
        return *lol * 5;
  }

  int test(int x)
  {
        int tmp = 5;
        tmp += inlineme(&tmp, x);
        return tmp;
  }
... all of which compiles to (with -O2):

  0000000000000000 :
   0:   83 c7 05                add    $0x5,%edi
   3:   8d 04 bf                lea    (%rdi,%rdi,4),%eax
   6:   01 f8                   add    %edi,%eax
   8:   c3                      retq
Note that "tmp" doesn't have to be a constant for that to happen.

As far as the actual kernel function, GCC ends up emitting: (The only uncompressed kernel binary I had laying around was for my beaglebone black, sorry)

  c0136fa0 :
  c0136fa0:       e92d4ff0        push    {r4, r5, r6, r7, r8, r9, sl, fp, lr}
  c0136fa4:       e24dd02c        sub     sp, sp, #44     ; 0x2c
  
  c0136fbc:       e3a03001        mov     r3, #1
  c0136fc0:       e58d301c        str     r3, [sp, #28]
  c0136fc4:       e3a03005        mov     r3, #5
  c0136fc8:       e58d3020        str     r3, [sp, #32]
  c0136fcc:       e3a03007        mov     r3, #7
  c0136fd0:       e58d3024        str     r3, [sp, #36]   ; 0x24
  c0136fd4:       e1a00007        mov     r0, r7
  c0136fd8:       e28d101c        add     r1, sp, #28
  c0136fdc:       e28d2020        add     r2, sp, #32
  c0136fe0:       e28d3024        add     r3, sp, #36     ; 0x24
  c0136fe4:       ebffffd4        bl      c0136f3c
... so in this case it does actually pass by reference.

That doesn't change my argument: an array is the more confusing way to do this, and in similar situations could prevent the compiler from making the optimization it did in my example.

Post reply on HN