[3/3] gdb: Show type summary for anonymous structures from c_print_typedef

  * Pedro Alves <palves@redhat.com> [2019-07-19 13:38:35 +0100]:

> On 7/12/19 12:37 PM, Andrew Burgess wrote:
> > Currently each language has a la_print_typedef method, this is only
> > used for the "info types" command.
> > 
> > The documentation for "info types" says:
> > 
> >    Print a brief description of all types whose names match the regular
> >    expression @var{regexp} (or all types in your program, if you supply
> >    no argument).
> > 
> > However, if we consider this C code:
> > 
> >    typedef struct {
> >      int a;
> >    } my_type;
> > 
> > Then currently with "info types" this will be printed like this:
> > 
> >    3:      typedef struct {
> >        int a;
> >    } my_type;
> > 
> > I see two problems with this, first the indentation is clearly broken,
> > second, if the struct contained more fields then the it feels like the
> > actual type names could easily get lost in the noise.
> 
> Something odd in "then the it feels"?
> 
> > 
> > Given that "info types" is about discovering type names, I think there
> > is an argument to be made that we should focus on giving _only_ the
> > briefest summary for "info types", and if the user wants to know more
> > they can take the type name and plug it into "ptype".  As such, I
> > propose that a better output would be:
> > 
> >    3:      typedef struct {...} my_type;
> > 
> 
> I think the same rationale applies to enums too?  We currently
> print anonymous enums like:
> 
>  16:     typedef enum {A, B, C} EEE;
> 
> The difference here is that we don't print newline between
> each enumerator.
> 
> It's as if we printed your struct example like this:
> 
>  3:     typedef struct {int a;} my_type;
> 
> which would be a bit more reasonable than the current output.
> 
> I did the 0 -> -1 change locally, and your patch addresses
> enums as well already:
> 
>  16:     typedef enum {...} EEE;
> 
> But I think you should add that to the testcase.
> 
> > The user understands that there is a type called `my_type`, and that
> > it's an alias for an anonymous structure type.
> 
> It's worth explicitly showing (in the commit log, IMO) that this
> only affects anonymous structs/enums.  For named structs/enums, we do
> print an abbreviated form with no fields/enumerators:
> 
>  11:     struct named_struct;
>  18:     enum named_enum;
> 
> So it makes sense to me to be consistent and use an abbreviated
> form for anonymous types too, like in your patch.

I've extended the test case to cover anonymous enums, and also unions
that I'd completely forgotten to cover at all before (though they are
mostly just handled like structs, so no surprises there).

I've extended the commit message to talk about the difference between
named and anonymous structs/enums, to mention that unions are like
structs, and also to mention the differences between C and C++ (with
C++ being less changes by this patch as its output before was more
compact anyway).

Thanks,
Andrew

--

    gdb: Show type summary for anonymous structures from c_print_typedef

    Currently each language has a la_print_typedef method, this is only
    used for the "info types" command.

    The documentation for "info types" says:

       Print a brief description of all types whose names match the regular
       expression @var{regexp} (or all types in your program, if you supply
       no argument).

    However, if we consider this C code:

       typedef struct {
         int a;
       } my_type;

    Then currently with "info types" this will be printed like this:

       3:      typedef struct {
           int a;
       } my_type;

    I see two problems with this, first the indentation is clearly broken,
    second, if the struct contained more fields then it feels like the
    actual type names could easily get lost in the noise.

    Given that "info types" is about discovering type names, I think there
    is an argument to be made that we should focus on giving _only_ the
    briefest summary for "info types", and if the user wants to know more
    they can take the type name and plug it into "ptype".  As such, I
    propose that a better output would be:

       3:      typedef struct {...} my_type;

    The user understands that there is a type called `my_type`, and that
    it's an alias for an anonymous structure type.

    The change to achieve this turns out to be pretty simple, but only
    effects languages that make use of c_print_typedef, which are C, C++,
    asm, minimal, d, go, objc, and opencl.  Other languages will for now
    do whatever they used to do.

    The patch to change how anonymous structs are displayed also changes
    the display of anonymous enums, consider this code sample:

       typedef enum {
         AA, BB, CC
       } anon_enum_t;

    This used to be displayed like this:

       3:      typedef enum {AA, BB, CC} anon_enum_t;

    Which will quickly become cluttered for enums with a large number of
    values.  The modified output looks like this:

       3:      typedef enum {...} anon_enum_t;

    Again, the user can always make use of ptype if they want to see the
    details of the anon_enum_t type.

    It is worth pointing out that this change (to use {...}) only effects
    anonymous structs and enums, named types don't change with this patch,
    consider this code:

       struct struct_t {
         int i;
       };
       enum enum_t {
        AA, BB, CC
       };

    The output from 'info types' remains unchanged, like this:

       4:      enum enum_t;
       1:      struct struct_t;

    An additional area of interest is how C++ handles anonymous types used
    within a typedef; enums are handled basically inline with how C
    handles them, but structs (and classes) are slightly different.  The
    behaviour before the patch is different, and is unchanged by this
    patch.  Consider this code compiled for C++:

       typedef struct {
         int i;
       } struct_t;

    Both before and after this patch, this is show by 'info types' as:

       3:      typedef struct_t struct_t;

    Unions are displayed similarly to structs in both C and C++, the
    handling of anonymous unions changes for C in the same way that
    it changes for anonymous structs.

    I did look at ada, as this is the only language to actually have some
    tests for "info types", however, as I understand it ada doesn't really
    support typedefs, however, by forcing the language we can see what ada
    would print.  So, if we 'set language ada', then originally we printed
    this:

       3:      record
           a: int;
       end record

    Again the indentation is clearly broken, but we also have no mention
    of the type name at all, which is odd, but understandable given the
    lack of typedefs.  If I make a similar change as I'm proposing for C,
    then we now get this output:

       3:      record ... end record

    Which is even less informative I think.  However, the original output
    _is_ tested for in gdb.ada/info_auto_lang.exp, and its not clear to me
    if the change is a good one or not, so for now I have left this out.

    gdb/ChangeLog:

            * c-typeprint.c (c_print_typedef): Pass -1 instead of 0 to
            type_print.

    gdb/testsuite/ChangeLog:

            * gdb.ada/info_auto_lang.exp: Update expected results.
            * gdb.base/info-types.c: Add additional types to check.
            * gdb.base/info-types.exp: Update expected results.

[3/3] gdb: Show type summary for anonymous structures from c_print_typedef

Commit Message

Comments

Patch