427661 - Treehydra: Expose struct lang_type. Support anonymous tagged unions automatically

Status: RESOLVED FIXED

(Developer Infrastructure :: Source Code Analysis, defect)

Description

[David Mandelin [:dmandelin]]

For bug 427536, I need struct lang_type from cp-tree.h exposed to Treehydra. (We need this data to tell the different between a struct and a class.) This is a union type.

We also kind of need it for pretty-printing pointer-to-member types, but that's less important.

Comment 1

[(dormant account)]

Attachment: I hate unions and ad-hoc ways to describe them

This should work.

Comment 2

[David Mandelin [:dmandelin]]

Thanks. That works.

Unfortunately, there is more where that came from (as I just discovered after reenabling the union-y code). Dehydra uses DECL_CLONED_FUNCTION_P, which ultimately requires this line of code to work in Treehydra:

  DECL_LANG_SPECIFIC(node).u.f.u5.cloned_function;

This refers to struct lang_decl in cp-tree.h, going through union field u, then struct field f, then union field u5. 

Comment 3

[David Mandelin [:dmandelin]]

Just another quick note, the DECL_LANG_SPECIFIC/lang_decl union stuff also seems to be required in order to do anything much with templates.

Comment 4

[(dormant account)]

This is going to take a pretty serious switcheroo in convert_tree.js and the generated files.

I'm posting a plan of attack for own notes here:
1) Make convert_* functions work like dehydra_defineProperty such that they return void
This also helps avoid rooting for location_t conversion
2) Implement correct support for %1 and %0 gengtype crap. This means that structs and unions defined within other structs will no longer get a toplevel func, instead they will get "inlined" into the conversion func for the toplevel struct
3) Enable all nested tagged unions, later might enable remaining toplevel ones.

Comment 5

[(dormant account)]

I've done 1) so far, the rest should be finished by end of the week. 

This bug blocks further Treehydra changes since rearranges the code a fair bit.

Comment 6

[(dormant account)]

Attachment: lang_type for dmandelin

the convert_tree.js code makes sense, but i can't decypher how node.type.lang_specific.u.h is supposed to work in C given that there is no h in lang_decl_u4

Any ideas? I'm done with this for today

Comment 7

[(dormant account)]

figured it out, turned i didn't infer enough about GCC tagged unions. In a tagged union if there is no tag on a field, means it's always present.
GCC also marked the .h part was being a "skip", evil thing.

Comment 8

[(dormant account)]

Committed fix + testcase