mikeash.com pyblog/friday-qa-2011-12-23-disassembling-the-assembly-part-2.html comments

Owen Shepherd - 2012-01-01 13:30:20

Sun, 01 Jan 2012 13:30:20 GMT

That is most certainly not a far jump/call. In AT&T syntax, a far jump/call is encoded using the "ljmp" or "lcall" mnemonic, and is used for cross-segment calls and jumps. Now, since cross-segment calls and jumps are nigh-on obsolete, you'll only very rarely see them..

Chris Suter - 2011-12-31 02:04:57

Sat, 31 Dec 2011 02:04:57 GMT

> If you peek at the generated machine code with a disassembler, it turns out to not be a mov instruction at all, but rather a lea!

Are you sure? I don’t see this.

> GOTPCREL is a directive which allows the rip-relative address of a function to be inserted at link time so a direct call can be made

I don’t think that’s quite right. See below.

> "far jump" (a branch over a long distance of code, which, by necessity, is much slower).

Again, I don’t think a “far jump” is necessarily much slower because it is over a long distance of code. It might be marginally slower because you might have to use more bytes to encode the instruction (but that won’t be *much* slower), and it will obviously be slower if the address you’re jumping to causes a page fault or a bit slower if it isn’t in the cache, but it’s not slower simply because it’s a branch over a long distance of code. Things might be a bit slower in your example because an indirect jump is being used (but you’d need to check the processor documentation to see by how much and I wouldn’t be surprised to find that if it’s cached, it’s nil).

> Note: I'm not 100% sure of my facts on this one; I'd appreciate any insight anyone has on the specifics of @GOTPCREL

@GOTPCREL allows you to load an address in a global offset table in a single instruction (using instruction relative addressing). The Global Offset Table (GOT) stores the address of objc_msgSend and any other global addresses that might be required. It is fixed up by dyld at runtime. The GOT is always at a fixed offset relative to code so you can use instruction relative addressing to load an address from it.

All of the above said, I’m no expert on this, so please don’t take my word on the above; I might be wrong.

Gwynne Raskind - 2011-12-23 23:10:20

Fri, 23 Dec 2011 23:10:20 GMT

Jens: I spent something like an hour trying to Goggle and otherwise look up that particular function attribute! I guess my Google-fu needs some work :). Thanks for the code listing!

Jens Ayton - 2011-12-23 22:55:08

Fri, 23 Dec 2011 22:55:08 GMT

it turns out to be extremely difficult to get Clang to actually emit such assembly under optimizing compilation without just inlining the function, and the unoptimized version is different.

__attribute__((noinline)) float MyFPFunction(float parameter)

The call and print sequence is:



    movss    LCPI1_0(%rip), %xmm0  # Load argument

    callq    _MyFPFunction

    cvtss2sd    %xmm0, %xmm0  # Promote to double

    leaq    L__unnamed_cfstring_(%rip), %rdi

    movb    $1, %al

    callq    _NSLog