Missing primitive appears without debug flags

[kit-ty-kate]

While compiling cumulus with the latest js_of_ocaml compiler (21ac186), I've been surprised by the following behavior:

$ js_of_ocaml -noruntime +js_of_ocaml/runtime.js +eliom.client/eliom_client.js +js_of_ocaml/weak.js +js_of_ocaml/classlist.js src/client/cumulus.byte -o src/client/cumulus.js
Missing primitives:
  __eliom_request_data
  …
$ js_of_ocaml -noruntime -pretty -debuginfo -noinline +js_of_ocaml/runtime.js +eliom.client/eliom_client.js +js_of_ocaml/weak.js +js_of_ocaml/classlist.js src/client/cumulus.byte -o src/client/cumulus.js
Missing primitives:
   …

Namely, the missing primitive __eliom_request_data appears only without the debug flags -pretty -debuginfo -noinline.

[hhugo]

inlining allows more optimisation.
it the current case, caml_js_var( "__eliom_request_data") translates to __eliom_request_data.

When inlining is disabled, the compiler doesn't manage to resolve caml_js_var(some_variable) and doesn't see __eliom_request_data as a symbol.

latest version of the compiler detect much more missing symbols (done on javacript ast, was done earlier before)
__eliom_request_data symbol being available at runtime only, the compiler warns about missing primitives.

I don't know if anything should be done about this

[kit-ty-kate]

Maybe we can do an exception (no inlining) for caml_js_val ?

[hhugo]

It means to keep caml_js_var("Date"), caml_js_var("Math"), ... instead of Date, Math.

The only update in the compiler, regarding your point, is more warning about missing variables declarations.
We don't want less optimization just to avoid warnings at compile time.

• we could display less warning.
• eliom could provide a default value for __eliom_request_data

[hhugo]

@vouillon, do you have any thought about this ?

[vouillon]

I think we should revert to trigger the linker only for OCaml primitives (declared using the external keyword), and not for all free variables.
For these primitives, calling the linker is reliable and convenient.
On the other hand, using the linker for other variables is not so reliable (this depends on compiler options). And these variables often correspond to Web browser APIs or external libraries, in which cases the linker is not needed.

[hhugo]

#42 should fix the issue.

The linker is triggered on free variables either :

• declared using external
• provided by an external javascript file
• or aliased by the compiler

There is a new option (-debuglinker`) that displays other missing variables.

[hhugo]

#42 merged