[Regression] Missing type declarations in C code if members are used in {.emit.}

[yglukhov]

Introduced in 70b28a3.

type Foo {.exportc.} = object
  a: cint
proc f(foo: ptr Foo): cint =
  {.emit: "`result` = `foo`->a;".}
echo f(nil)

Error: execution of an external compiler program 'clang -c  -w  -I/Users/yglukhov/Projects/nim/lib -o /Users/yglukhov/nimcache/test.o /Users/yglukhov/nimcache/test.c' failed with exit code: 1

/Users/yglukhov/nimcache/test.c:82:14: error: incomplete definition of type 'struct Foo'
        result = foo->a;
                 ~~~^
/Users/yglukhov/nimcache/test.c:24:16: note: forward declaration of 'struct Foo'
typedef struct Foo Foo;
               ^
1 error generated.

@cooldome, can we fix it so that {.exportc.} types are always declared completely?

[cooldome]

Hi Yuriy,
I think exportc is irrelevant here, what is actually needs to be patched is genEmit. I will have a look

[cooldome]

I have a patch, I will prepare PR later today

[Zdreni]

I think this may be related to that issue. Seems that ptr or var reference to structure type doesn't cause generation of C code for that type. For example:

type stbtt_hheap_chunk {.exportc.} = object
    next: ptr stbtt_hheap_chunk

proc test(x: ptr stbtt_hheap_chunk) =
   {.emit: """
   if (`x`->next) {} else {}
   """.}

However adding next line:

var testHeapChunk: stbtt_hheap_chunk

causes code to be generated

[cooldome]

Hi Zdreni,
Your example compiles without an issue. Exclusion of stbtt_hheap_chunk is intensional in this case, you are not using it

[cooldome]

Zdreni, maybe you have more comprehensive example

[yglukhov]

The talk is probably about ttf lib which is exactly the reason why I filed this issue.

[Zdreni]

Sorry, example was fixed. This was a bit shortened extraction of code from ttf (0.2.4) nimble package.
The error should be: "incomplete definition of type 'struct stbtt_hheap_chunk'"

[Zdreni]

After fix, the issue still remains in other form. Just use additional level of indirection:

type stbtt_hheap_chunk {.exportc.} = object
    next: ptr stbtt_hheap_chunk

type stbtt_hheap {.exportc.} = object
    head: ptr stbtt_hheap_chunk

proc test(x: ptr stbtt_hheap) =
   {.emit: """
   if (`x`->head->next) {} else {}
   """.}

Causes:
error: incomplete definition of type 'struct stbtt_hheap_chunk'

[Zdreni]

I also need to mention that simple tracking of field names mentioned in emit possibly may not help in corner cases where name of field is generated by macro/concatenating two strings. TBH, I doubt there is a reason for optimizing out structures, that may be possibly used in emit blocks.

[cooldome]

checking...

[cooldome]

I suggest I'll add support for expression backticking in emit pragma:
Then the following is going to work:

proc test(x: ptr stbtt_hheap) =
   {.emit: """
   if (`x.head.next`) {} else {}
   """.}

and also

proc test(x: ptr stbtt_hheap) =
   {.emit: """
   if (`x.head`->next) {} else {}
   """.}

It is useful feature on its own plus it will allow Nim to trace type usage properly.

Let me know if you are ok with the proposal.

[yglukhov]

@cooldome, I'm ok with your proposal, although I think exportc is a nice indicator for the type to be defined. I mean why else would you want to place exportc on it?

[Zdreni]

Proposed solution with expression backticking, however, will not work on lot of existing code, leaving packages broken. And yes, I agree with @yglukhov. I see no harm in generating exportc structures, because being not used they will be wiped by C compiler.

[cooldome]

@yglukhov

Yuriy,
What you are asking is not unreasonable, but it has nothing to do with this regression. Compiler did ignore not used types before and it ignores them now, no changes to exportc treatment were introduced either.
I suggest raising a separate feature request. The implementation would require codegen pass to walk the type definitions (currently type sections completely ignored) and check their flags.

[cooldome]

I have submittted the PR with backticking of expressions

[yglukhov]

Seems like this issue was closed by irrelevant PR. Also note, that this can be easily worked around by adding

(void)sizeof(`MyType`);

to the emit pragma.

[cooldome]

I have raised separate feature request #7448
Mean while I suggest using proposed workaround and close this issue.

[Araq]

Too much of a fringe case.