Bug 1637

More suggestions from Carlos

-----------------------------------------------------------------
char * Test::Hello::get_string (
    ACE_ENV_SINGLE_ARG_DECL
  )
  ACE_THROW_SPEC ((
    CORBA::SystemException
  ))
{
  if (!this->is_evaluated ())
    {
      ACE_NESTED_CLASS (CORBA, Object)::tao_object_initialize (this);
    }

  if (this->the_TAO_Hello_Proxy_Broker_ == 0)
  Test_Hello_setup_collocation (
          this->ACE_NESTED_CLASS (CORBA, Object)::_is_collocated ()
        );
    }
 ....
 ...
 ..
 .
----------------

        you may want to change that for a single call, like:

----------------
/* private */ /* inline ?? */ void Test::Hello::__tao_initialize()
{
  if (!this->is_evaluated ())
    {
      ACE_NESTED_CLASS (CORBA, Object)::tao_object_initialize (this);
    }

  if (this->the_TAO_Hello_Proxy_Broker_ == 0)
    {
      Test_Hello_setup_collocation (
          this->ACE_NESTED_CLASS (CORBA, Object)::_is_collocated ()
);
    }
}

char * Test::Hello::get_string (
    ACE_ENV_SINGLE_ARG_DECL
  )
  ACE_THROW_SPEC ((
    CORBA::SystemException
  ))
{
  __tao_initialize();
 ....
 ...
 ..
 .
----------------

        Not much difference if you only have one operation, but much
better if you have a lot.

 The next change is more interesting.  What can we do to reduce
the cost of the Typecodes and the Anys?  First notice the following
line counts (in TAO/tests/Hello):

$ /home/coryan/UCI/ACE_wrappers/build/Linux/TAO/TAO_IDL/tao_idl -Ge 1 -
DBORG_EVENT_HEADER  -St  Test.idl
$ wc -l TestC.cpp
    303 TestC.cpp

$ /home/coryan/UCI/ACE_wrappers/build/Linux/TAO/TAO_IDL/tao_idl -Ge 1 -
DBORG_EVENT_HEADER  Test.idl
$ wc -l TestC.cpp
    399 TestC.cpp

        Nearly 25% of the code in the C.cpp file is just Typecode
madness.  In terms of build time this is even more interesting:

-- Without -St --
$ time g++ -Wwrite-strings -W -Wall -Wpointer-arith -pipe -O3 -g    -D_LARGEFILE64_SOURCE -
D_REENTRANT -DACE_HA\
S_AIO_CALLS -D_GNU_SOURCE -DBORG_EVENT_HEADER  -I/home/coryan/UCI/ACE_wrappers/
build/Linux -I/home/coryan/UCI/A\
CE_wrappers/build/Linux/TAO -DACE_HAS_EXCEPTIONS -DACE_NO_INLINE    -DTAO_BUILD_DLL  -
c -o foo.o TestC.cpp

real    0m2.730s
user    0m2.550s
sys     0m0.210s

-- With -St --
$ /home/coryan/UCI/ACE_wrappers/build/Linux/TAO/TAO_IDL/tao_idl -Ge 1 -
DBORG_EVENT_HEADER  -St  Test.idl
[coryan@glamdring Hello]$ time g++ -Wwrite-strings -W -Wall -Wpointer-arith -pipe -O3 -g    -
D_LARGEFILE64_SOUR\
CE -D_REENTRANT -DACE_HAS_AIO_CALLS -D_GNU_SOURCE -DBORG_EVENT_HEADER  -I/home/
coryan/UCI/ACE_wrappers/build/Li\
nux -I/home/coryan/UCI/ACE_wrappers/build/Linux/TAO -DACE_HAS_EXCEPTIONS -
DACE_NO_INLINE    -DTAO_BUILD_DLL  -cCE -D_REENTRANT -DACE_HAS_AIO_CALLS -
D_GNU_SOURCE -DBORG_EVENT_HEADER  -I/home/coryan/UCI/ACE_wrappers/build/Li\
nux -I/home/coryan/UCI/ACE_wrappers/build/Linux/TAO -DACE_HAS_EXCEPTIONS -
DACE_NO_INLINE    -DTAO_BUILD_DLL  -c\
 -o foo.o TestC.cpp

real    0m1.924s
user    0m1.810s
sys     0m0.150s

        That's nearly a 40% improvement!  I'll admit this is a small
file, anything with lots of operations and no types will not look this
good, but something with lots of types and no operations will look
much better :-)

        Furthermore, the Typecode.h files in the ORB are bloated and
slow to compile.

        What I will propose is that we completely change the way we
generate Typecode information.  First we need to remember that a
Typecode, after it is completely extracted, is simply a hierarchal
data structure that represents the elements in an IDL construct.  What
we could do is create this hierarchy directly, skipping the
intermediate binary representation.

        The solution below should work, I can only find one problem
with it: it does not work if B.idl includes (and uses types from)
A.idl but A.idl was compiled without -St....  Frankly, this
ill-defined at best.  I have some solutions to offer there (read [*])

        Before we go into abstractions, let me give you an example,
suppose we have IDL like this:

typedef long Token;

struct Quz {
  Token first_token;
  Token second_token;
  string the_string;
};

union Foo switch(short) {
  case 0:
  case 1:  short bar;
  case 2:  Quz the_quz;
  default: Token baz;
};

        I think if we generate code like this:

 // C++ (*C.cpp file)
 TAO::Alias_TypeCode _tao_tc_Token("IDL:Token/1.0", "Token", &CORBA::_tc_long);
 CORBA::TypeCode_ptr _tc_Token = &_tao_tc_Token;

       first notice that the *value* of CORBA::_tc_long is not know
at compile time, but its *address* is known (or at least well-defined)
at compile time.  Naturally the TAO::Alias_TypeCode class will have to
derive from CORBA::TypeCode, and it will need to know how to use the
parameters effectively.  I'll show how to implement such classes in a
second.

        Structures are easy:

 // C++ (*C.cpp file)
 TAO::TypeCode_Field _tao_fields_tc_Quz[] = {
   { "first_token", &_tc_Token },
   { "second_token", &_tc_Token },
   { "the_string", &CORBA::_tc_string },
 };
 TAO::Struct_TypeCode _tao_tc_Quz(
    "IDL:Quz/1.0", "Quz",
    _tao_fields_tc_Quz, sizeof(_tao_fields_tc_Quz));

        here I can show some of the implementation:

namespace TAO {
class Struct_TypeCode
  : public CORBA::TypeCode // This should be abstract
{
public:
  Struct_TypeCode(
    char const * id,
    char const * name,
    TAO::TypeCode_Field * fields,
    size_t nfields)
    : id_(id)
    , name_(name)
    , fields_(fields)
    , nfields_(nfields)
  { }

  virtual CORBA::TCKind() const { return _tk_struct; }
  char const * id() const { return id_; }
  char const * name() const { return name_; }
  CORBA::ULong member_count() const { return nfields_; }
  char const * member_name(CORBA::ULong i) const
  {
    if(i >= nfields_) {
      throw ....;

    return fields_[i].name_;
  }
  CORBA::TypeCode_ptr member_type(CORBA::ULong i) const
  {
 if(i >= nfields_) {
      throw ....;

    return *(fields_[i].type_); // notice the indirection
  }

  // operations that do not make sense are *supposed* to throw...
  CORBA::TypeCode_ptr discriminator_type() const {
    throw BadKind();
  }
};

        as you can see very is straightforward.  I think
arrays, sequences, interfaces and stuff like that are simple too.
Unions are a little more tricky:

 // C++ (*C.cpp file)
 TAO::TypeCode_Branch _tao_branches_tc_Foo[] = {
   { 0, "bar", &CORBA::_tc_short },
   { 1, "bar", &CORBA::_tc_short },
   { 2, "the_quz", &_tc_Quz },
 };
 TAO::Struct_TypeCode _tao_tc_Foo(
    "IDL:Foo/1.0", "Foo",
    &CORBA::_tc_short, // discriminator
_tao_branches_tc_Foo, sizeof(_tao_branches_tc_Foo),
    "baz", &_tc_Token // default (could be NULL if there is none)
 );

        I think recursive types simply work too, as in:

// IDL
struct Recursive;
typedef sequence<Recusive> Recursion;
struct Recursive {
  long x;
  Recursion y;
};

// C++ (Header)
CORBA::TypeCode_ptr _tc_Recursive;
CORBA::TypeCode_ptr _tc_Recursion;

// C++ (CPP)
TAO::Alias_TypeCode _tao_tc_Recursion(.. &_tc_Recursive ...); // legal!
CORBA::TypeCode_ptr _tc_Recursion = &_tao_tc_Recursion;

TAO::Struct_TypeCode _tao_tc_Recursive(.. &_tc_Recursion ...); // OK!
CORBA::TypeCode_ptr _tc_Recursive = &_tao_tc_Recursive;

_tao_branches_tc_Foo, sizeof(_tao_branches_tc_Foo),
    "baz", &_tc_Token // default (could be NULL if there is none)
 );

        I think recursive types simply work too, as in:

// IDL
struct Recursive;
typedef sequence<Recusive> Recursion;
struct Recursive {
  long x;
  Recursion y;
};

// C++ (Header)
CORBA::TypeCode_ptr _tc_Recursive;
CORBA::TypeCode_ptr _tc_Recursion;

// C++ (CPP)
TAO::Alias_TypeCode _tao_tc_Recursion(.. &_tc_Recursive ...); // legal!
CORBA::TypeCode_ptr _tc_Recursion = &_tao_tc_Recursion;

TAO::Struct_TypeCode _tao_tc_Recursive(.. &_tc_Recursion ...); // OK!
CORBA::TypeCode_ptr _tc_Recursive = &_tao_tc_Recursive;

     The demarshaling code to receive typecodes from the wire will
have to get smarter, but that's hidden in the library.


                                Let me know what you think.




[*]  So what if B.idl did not define _tc_Foo but A.idl needs it?
Several solutions:
- If we successfully eliminate the *motivation* for the -St option,
  then we can remove the option and avoid the problem completely!

- The application is invoking undefined behavior, i.e. as soon as the
  use -St they are outside the CORBA spec and we can tell them what
  the rules are.  In this case "-St everywhere or nowhere."

- Use macros to detect (at build time) if A.idl was compiled without
  -St.  In that case fallback on some other mechanism, like the old
  arrays.  For example, the generated code could include some macros
  like:

// AC.h
#define _tao_compile_trait__A_IDL__with_St

        which are used like this:

// BC.cpp

#if defined(_tao_compile_trait__A_IDL__with_St)
// Generate _tc_Foo here
namespace {
CORBA::TypeCode _tc_Foo = ....;
}
#endif

        Using macros will probably fail when the filenames or the
type names are ambiguous.  One could dream of template
meta-programming techniques to work around that, but it is overkill.