tagged records of different formats, the structure of which depend on the
tag. To analyze the structure of the record for some particular tag
we can always ask the compiler to generate as many such records as we need.

The fact, which made it very difficult to start the analysis of DCU32
files, is the extensive use of the data structures, which, by analogy with
OBJ files, I call indices. I.e. all the integer fields, which can take
more than 1 byte value, are encoded in the DCU32 file so, that 1s in the
first bits of the first byte (up to 4 bits) indicate, that additional bytes
are used to represent the value. So, the value of 0x12 is represented by one
byte with the value of 0x24, 0x1243 - two bytes: 0x0D 0x49, and so on,
and for the values which require all 4 bytes (more than 28 bits), the high
4 bits of the 1st byte are not used and filled with zeroes, and the value
is represented directly by the next 4 bytes. Subsequently, in Delphi 4.0
the 64-bit integers were introduced, to encode their values the 1st byte
takes the value of 0xFF, and the next 8 bytes represent the value.

So, the size of any data structure may differ depending on the values of
the indices contained in it, and to successfully parse the DCU32 file it is
required to detect these index fields. To do it I have written and analyzed
some test files which contain something like the following code:

const
  c1=$1;
  c12=$12;
  c123=$123;
  c1234=$1234;
  c12345=$12345;
  c123456=$123456;
  c1234567=$1234567;
  c12345678=$12345678;
  cm1=-$1;
  cm12=-$12;
  cm123=-$123;
  cm1234=-$1234;
  cm12345=-$12345;
  cm123456=-$123456;
  cm1234567=-$1234567;
  cm12345678=-$12345678;

The index field can be detected by the changes in its size and by
the fact that it often takes successive even numbers in the successive
records of the same type.

Detecting the record structure doesn't mean detecting its semantics.
The values of some indices remain uninterpreted. The criterion for
correctness is the successful parsing of the test file, without breaks
in the sequence of tagged records caused by the wrong size detected
for some record.

The results of this analysis are represented in the DCU32.rfi FlexT file.
Unfortunately, not all the information can now be represented in FlexT.
In particular, it can't represent the fact that some tagged records are
enumerated by the two number sequences: the sequence of data types and
the sequence of addresses, with the data type declaration being a member
of both sequences (the address for data type is reserved for its TypeInfo).
A lot of indices in the DCU32 records are references to addresses and data
types using these sequence numbers.

So, to completely restore the structure of DCU32 files, it was anyway
required to write the specialized program DCU32INT, which extracts almost
all the information in readable and close to Pascal syntax form. I call
this program DCU32INT and not DCU32PAS, because it is always possible
to extract an interface part of a DCU32 file, but I doesn't claim yet
that it can extract a Pascal file, which can be immediately compiled
to obtain the same DCU (a lot of work should be done to make it possible).

All the rest of the reconstruction of the DCU32 format had been done using
the DCU32INT program, but the FlexT parse results were still used when
something went wrong. The most complex thing here was to understand the
rules of assignment of the numbers in the sequence of addresses.


The Compiler Information Loss Limitations.
------------------------------------------

There are two kinds of limitations of the DCU32INT program:

1. The ones caused by some disadvantages in its implementation,
  which can be overcome later;
2. The ones caused by information loss in DCU after compiling Pascal
  source, which are inevitable.

Here we'll consider some of the latter limitations.

While converting Pascal source into DCU, Delphi compiler extracts from
source and stores into DCU only the information, which is necessary to
produce later an executable file and also, if required, a debug
information for this file. During this process the compiler performs
some simplifications, which cause information loss.

Examples:

1. Identifiers declared in implementation part and subroutines are discarded
  if the debug info checkbox doesn't checked.

2. Evaluation of expressions. Constant expressions are replaced by their
  values, so one can't determine, e.g. that CDM_FIRST = WM_USER+100,
  it will have the fixed value of $0464.
  
3. Resolution of rename types. The rename types (types, which are defined
  by declarations like THandle = integer), are replaced by their reference
  type, so all the references to the THandle type in the source code are
  replaced by the System.Integer type.
  
4. Merge of fields in the records with variants. The declaration like

  TVarRec = record
    case Byte of
      vtInteger:    (VInteger: Integer; VType: Byte);
      vtBoolean:    (VBoolean: Boolean);
    -----------------------------------------
  end;

  Is stored as

  TVarRec{88,7F9FF4C2}=record
    VInteger: Integer{F:2 Ofs:0};
    VType: Byte{F:2 Ofs:4};
    VBoolean: Boolean{F:2 Ofs:0};
    ----------------------------------
  end;

  where Ofs:_ is a field offset information. Of course, we can group
  the fields into cases according to their order and offsets (this
  capability is now supported by DCU32INT), and TVarRec is parsed as

  TVarRec=record 
    case Integer of
     0: (
       VInteger: Integer;
       VType: Byte);
     1: (
       VBoolean: Boolean);
    ----------------------------------
  end;

  But the information about the case labels is lost here completely, 
  and it can't be used, e.g. to display safely (Delphi version 
  independent) the Variant type value using the TVarData definition.

All the above mentioned limitations can be demonstrated by Delphi
browser and evaluator (those utilities are also limited by them).

So, the extracted Pascal code can cause some problems, if used in other
version of Delphi, than that, which produced the DCU.

Home page.
----------

The latest version of this program and all the related news will be available
at http://hmelnov.icc.ru/DCU/

Please, send me (e-mail: alex@icc.ru) bug reports (including the units 
which were not parsed correctly), but first check:
  1. that you have the latest version of DCU32INT,
  2. that this bug was not already reported at 
     http://hmelnov.icc.ru/DCU/FAQ.htm.


Collaboration.
--------------

If you'll create something useful using my program or information contained
in its source, or will substantially improve this program, please, send me 
your results. I'll publish all such programs, which I'll consider to be 
useful, at my site (including the links to their home pages, if available,
and/or other author information).

I can propose the following lines of improvements, which I'm not going to
develop myself in the nearest future:

1. The DCU32INT is a console application, but it would be interesting to 
  create some kind of DCU Browser as a GUI application.

2. The disassembler, which is used in DCU32INT, is rather simple, it would be 
  useful to improve it using some data flow analysis techniques.

3. The ideal final result for this kind of program would be to restore 
  completely the Pascal source from DCU. Of course this problem 
  is VERY hard. More simple problem, but still VERY hard, is to produce
  the Pascal source, where all the procedures are ASSEMBLER. The easier 
  approach is to produce assembler procedures which could have incorrect
  semantics (e.g. DB instead of some opcodes), but still could be compiled 
  to produce correct executable. Note, that the result of this kind of 
  program would be Delphi version dependent.

4. Compiler discards some names, and other information. We could create
  additional input file for DCU32INT, which could contain some additional 
  guess information for the DCU being parsed, e.g. names of unnamed local 
  variables or types, entry points in the procedure code or even additional 
  type declarations (which could be compiled using DCC32 into additional DCU
  and then extracted from it).

5. Personally, I prefer to use Delphi and not C++ Builder, but it could be 
  interesting to modify this program to enable it generate its output in 
  the close to C++ Builder syntax form (C++ syntax + Borland's extensions) 
  instead of Pascal. This program could be used to partially transfer
  programs frop Pascal to C through DCU.
  
------------------------------------------------------------------------

                             IMPORTANT NOTE:

This software is provided 'as-is', without any expressed or implied warranty.
In no event will the author be held liable for any damages arising from the
use of this software.
Permission is granted to anyone to use this software for any purpose,
including commercial applications, and to alter it and redistribute it
freely, subject to the following restrictions:
1. The origin of this software must not be misrepresented, you must not
   claim that you wrote the original software.
2. Altered source versions must be plainly marked as such, and must not
   be misrepresented as being the original software.
3. This notice may not be removed or altered from any source
   distribution.