Fixes for DWARF CFI when switching between stacks

[sadiqj]

No description provided.

[kayceesrk]

Why is the offset 8 here? Do you need this? 8 doesn't look right.

[anmolsahoo25]

Since here the address will be determined by the value in memory, I think we should load the value into a register use cfi_def_cfa to set stack pointer to the register value

[kayceesrk]

Can you explain what was the original problem and how the PR fixes it (for posterity)?

And, since we are pushing the OCaml stack pointer information into the C stack, is it necessary to still use rbp at all? Otherwise, can we use rbp, but use other DWARF info to avoid pushing into the stack? I am not comfortable adding 4 memory operations for every fast external call in order to support debugging information. Perhaps this doesn't matter in practice?

[mshinwell]

One general point that it's worth bearing in mind, apologies if this is known already: CFI information isn't just for debugging. It is required to be correct by some language runtimes (Objective-C on macOS is one such) in order to unwind the stack for exception handling.

[kayceesrk]

Agreed. I was a bit sloppy in my use of the term debugging.

[anmolsahoo25]

The following is enough to generate the back-trace correctly -

        I.mov rsp rbp;
        if Config.stats then begin
          I.inc (domain_field Domainstate.Domain_extcall_noalloc);
        end;
        I.mov (domain_field Domainstate.Domain_c_stack) rsp;
        cfi_def_cfa ();
        I.sub (int 8) rsp;
        cfi_def_cfa_offset 8;
        emit_call func;
        I.mov rbp rsp;

Tested it with an external call, seems to work. Also, can remove the space-time related stuff, because we are assuming this function does not alloc.

[kayceesrk]

I.sub (int 8) rsp Will make the stack misaligned. It will fail with functions that expect 16-byte alignment.

…

On Mon, 13 Jan 2020 at 20:34, Anmol Sahoo ***@***.***> wrote: The following is enough to generate the back-trace correctly - I.mov rsp rbp; if Config.stats then begin I.inc (domain_field Domainstate.Domain_extcall_noalloc); end; I.mov (domain_field Domainstate.Domain_c_stack) rsp; cfi_def_cfa (); I.sub (int 8) rsp; cfi_def_cfa_offset 8; emit_call func; I.mov rbp rsp; Tested it with an external call, seems to work. Also, can remove the space-time related stuff, because we are assuming this function does not alloc. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#270?email_source=notifications&email_token=AADEG5HPHMWRJV7BI7LCIALQ5R7GLA5CNFSM4J74QX62YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEIZBCEQ#issuecomment-573706514>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AADEG5BXPLOK5WWIHUIDQHLQ5R7GLANCNFSM4J74QX6Q> .

[sadiqj]

This is superseded by the implementation in: #272