Replace startup code block with loop to call all entry stubs.

[tmcgilchrist]

This is a small improvement to rewrite the entry_point code generated by cmm_helpers.ml to be data driven.

When starting a native OCaml program the call stack goes through C functions (caml_main, caml_startup_common and caml_startup_common, in startup_nat.c before jumping to the processor specific assembly in runtime/arm64.s (or whatever architecture you have) starting at caml_start_program. Before it jumps to caml_program which calls the module .entry point for the modules included in an executable. At this point currently OCaml generates a block of assembly code for each module entry point that needs to be called, 35 for a minimal program that just relies on the Stdlib (potentially more?). This change re-writes this code to be a loop equivalent to this C code:

 {
   int id = 0;
   while (true) {
     if (id == len_caml_globals_entry_functions) goto out;
     caml_globals_entry_functions[id]();
     caml_globals_inited += 1;
     id += 1;
   }
   out:
   return 1;
 }

What is gained by making this change? More compact assembly code that is data driven using a loop and lookup table of symbols. Possibly faster startup but I have no data to back this up. The work was done original in PR oxcaml/oxcaml#1425 by @poechsel, I have only cherry picked it here after noticing the repeated assembly code blocks.

@poechsel please correct any attribution as you see fit.

[shindere]

Thanks a lot for the nice PR description! Out of curiosity: did you check that the change also works when frame pointers are enabled?

[tmcgilchrist]

Good idea @shindere I have not as yet, but I will do so.

[shindere]

Tim McGilchrist (2024/06/05 00:18 -0700):

Good idea @shindere I have not as yet, but I will do so.

Thanks! I'll wait for you to post here to proceed, although I am likely not the most well suited to review this.

[tmcgilchrist]

@shindere testing with frame-pointers on macOS and Linux is passing.

[shindere]

Tim McGilchrist (2024/06/18 08:08 -0700):

@shindere testing with frame-pointers on macOS and Linux is passing.

Okay, thanks!