Re: [PATCH v2 2/7] linkage: add SYM_{ENTRY,START,END}_AT()

[Nick Desaulniers]

On Thu, Feb 10, 2022 at 6:52 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> Both GCC and clang are happy to treat labels as constant expressions:
>
> | [mark@lakrids:~/asm-test]% cat test-label.S
> |         .text
> |
> | start:
> |         nop
> | end:
> |
> |         .if (end - start) == 0
> |         .err
> |         .endif
> |
> | [mark@lakrids:~/asm-test]% usekorg 11.1.0 aarch64-linux-gcc -c test-label.S
> | [mark@lakrids:~/asm-test]% usellvm 13.0.0 clang --target=aarch64-linux -c test-label.S
>
> ... but only GCC is happy to treat symbol definitions as constants:
>
> | [mark@lakrids:~/asm-test]% cat test-symbol.S
> |         .text
> |
> | .set start, .;
> |         nop
> | .set end, .;
> |
> |         .if (end - start) == 0
> |         .err
> |         .endif
> |
> | [mark@lakrids:~/asm-test]% usekorg 11.1.0 aarch64-linux-gcc -c test-symbol.S
> | [mark@lakrids:~/asm-test]% usellvm 13.0.0 clang --target=aarch64-linux -c test-symbol.S
> | test-symbol.S:7:6: error: expected absolute expression
> |  .if (end - start) == 0
> |      ^
> | test-symbol.S:8:2: error: .err encountered
> |  .err
> |  ^
>
> This is obviously a behavioural difference, but I'm not sure whether it's
> intentional, or just an artifact of the differing implementation of GNU as and
> LLVM's integrated assembler. Nich, Nathan, any thoughts on that?
>
> Does clang have any mechanism other than labels to define location constants
> that can be used as absolute expressions? e.g. is there any mechanism to alias
> a label which results in the alias also being a constant?

s/constant/absolute/

Nothing off the top of my head comes to mind as a substitute that will
work as expected today.

I've filed https://github.com/llvm/llvm-project/issues/53728 to
discuss more with folks that understand our AsmParser better.

>From what I can tell, our AsmParser is falling down because the
operands of the binary expression themselves are not considered
absolute.

I doubt we will be able to handle the general case of symbol
differencing; the symbol reference operands could refer to symbols in
different sections that haven't been laid out yet (or whose order
could be shuffled by the linker).  But if they're in the same section
(or "fragment" of a section), we might be able to special case this.

For the expression

> .if (qwerty_fiqin_end - qwerty_fiqin_start) > (0x200 - 0x1c)

can you use local labels (`.L` prefix) rather than symbolic
references? or is there a risk of them not being unique per TU?
-- 
Thanks,
~Nick Desaulniers