How do you specify a label reference using .intel_syntax?

Question

I'm having some really big problems trying to get certain x86 instructions assembled properly using .intel_syntax, compiling with -m32 (i.e. 32-bit mode). If I have some code like this:

    #define STACK_SIZE                      0x4000

    .att_syntax
    movl    $(stack + STACK_SIZE), %esp

    .comm   stack, STACK_SIZE

...all is fine. The decompiled output from this looks like this (AT&T syntax first, Intel syntax shown afterwards).

100010:       bc 70 5a 10 00          mov    $0x105a70,%esp
100010:       bc 70 5a 10 00          mov    esp,0x105a70

However, if I change my code like this (which I feel "should" work):

    .intel_syntax noprefix
    mov     esp, stack + STACK_SIZE

...I get this output instead:

100010:       8b 25 70 5a 10 00       mov    0x105a70,%esp
100010:       8b 25 70 5a 10 00       mov    esp,DWORD PTR ds:0x105a70

Obviously, this is wrong; I'm not trying to dereference the stack label but instead create a reference to it.

For now, my workaround is to use AT&T syntax for some parts of my file and Intel syntax for the rest. This feels like a kludge. Unfortunately, the information about GAS Intel mode feels a bit sparse; this page gave some hints but nothing that really helped.

Many thanks in advance. If I had reputation enough to give you a bounty, I would. :)

Answer 1

That seems right. You are correct that there isn't too much information on .intel_syntax. I wrote to a gal whose name I found in the source code, asking if there was any documentation. She replied that there wasn't, that she'd just reverse engineered some stuff they got from Intel. I poked around the source and found that offset flat: was required (at that time) - the colon was mandatory, too! Now, just plain offset seems to work... and there may be more documentation available, too(?).

I should think that Nasm would work with the same toolchain... no?

Answer 2

Found it!

I turned out that I must write my code like this:

    mov     esp, offset stack + STACK_SIZE

This feels silly; I believe NASM wouldn't have forced me to do this. However, I prefer to use GAS for this project to avoid the toolchain needed to build it. If anyone feels like elaborating on the question further (e.g. to confirm or reject this suggestion of mine), please do so, it would be interesting to read your thoughts also.

(Egor Skriptunoff's comment at this question made me understand how this should be done with GAS.)