[ARM] Incorrect dependency graph for store dependent on post-indexed load

[dgazzoni]

Consider this non-sensical, but minimal, sequence of instructions:

    ldr x0, [x1], #8
    str x1, [sp]

This is the dependency graph generated by OSACA:

I would expect a dependency from ldr to str due to the update of x1.

Note that replacing the code above by an explicit addition of 8 to x1 generates a dependency from the addition to the store, just as I would expect:

    ldr x0, [x1]
    add x1, x1, #8
    str x1, [sp]

Dependency graph:

Unless I misunderstood how these dependency graphs are supposed to work, I believe that's a bug.

[dgazzoni]

Analyzing the code, I'd like to draw attention to this check:

                    if self.is_read(dst.register, instr_form) and not (
                        dst.get("pre_indexed", False) or dst.get("post_indexed", False)
                    ):
                        yield instr_form, []

I'm not sure if I understand the need for the checks on "pre_indexed" and "post_indexed" there. I tried removing these conditions, i.e. rewriting the code to this:

                    if self.is_read(dst.register, instr_form):
                        yield instr_form, []

This appears to fix the issue. This is the new dependency graph for the first code snippet in the initial bug report (just post-indexed ldr followed by str):

However, given that I don't understand why the condition is there, I have no idea if this breaks something else.