YJIT: Allow testing assembler with disasm

[k0kubun]

This PR proposes an alternative way to write assembler tests, rewriting tests of #7453 as a starter.

I don't really memorize how x86_64 instructions are encoded, so it's hard for me to read these test cases. Ideally, I'd like to test the generated code using disasm, assuming libcapstone is not broken. A challenge in it is that addresses in generated code are always different, so those values are replaced with start_addr-origin values.

[maximecb]

This is clever, but I want to be strict about minimizing dependencies, even if they are used only during development. I would rather not merge this at this time.

[k0kubun]

How about this? 57a26a4 I removed new dependencies.

[maximecb]

Still not a huge fan of adding a dev dependency or relying on the output of capstone and string processing 😅

Let's ask @XrXr what his opinion is when he is back next week.

[XrXr]

I thought about doing something like this before and never got around to it. Thanks for looking into this! Being able to see the disassembly make these much easier to read.

In terms of dev dependency, I think we can make it test the disassembly optionally. So we would have something like:

assert_disasm!(
    cb,
    "488d5b08", /* checked when cargo test is built without features=disasm */,
    "lea rbx, [rbx + 8]..." /* additionally checked in `cargo test --all-features`, which we do on CI */
)

Some string manipulation to deindent is probably necessary, but maybe we can do less of it by not testing the addresses; I don't think the addresses are awfully important. I'm less concerned about Capstone because it seems to be reliable WRT always giving the same output.

[maximecb]

I was on the fence, but since Alan is in favor, let's go forward with this 👍

[k0kubun]

Some string manipulation to deindent is probably necessary, but maybe we can do less of it by not testing the addresses; I don't think the addresses are awfully important. I'm less concerned about Capstone because it seems to be reliable WRT always giving the same output.

Ah, I had forgotten that we were passing addresses to disasm_all ourselves. If we pass 0 to it, we wouldn't need address manipulation there. I'll fix it so.

In terms of dev dependency, I think we can make it test the disassembly optionally. So we would have something like
additionally checked in cargo test --all-features, which we do on CI

👍

[k0kubun]

addressed Alan's points in ff6301b and 05c9865.