.EXE Interview with Anders Hejlsberg on Delphi (1995)
February 12, 2020
To commemorate the 25th anniversary of Delphi on Feb 14, 2002 here is a transcript of an interview with Anders Hejlsberg, Chief Architect of Delphi conducted by .EXE Magazine editor Will Watts from 1995. Anders discusses the design and development of Delphi and the then forthcoming 32 bit version for Windows 95. This was the most detailed technical interview published about Delphi at the time.
Q. How did the idea for Delphi evolve from Turbo/Borland Pascal? At what stage did you decide to add the environment, database support etc?
A. The key idea was to design a tool that combines visual development environment, Client/Server database support, and a native code compiler. Before Delphi, you always had to make a choice. Do I go for the performance of a native code compiler, or the ease of use of a visual development environment? Do I go for a powerful object-oriented language, or a proprietary 4GL Client/Server tool? What programmers really want is all of the above, in one package. That's what we set out to do with Delphi.
What it really boils down to is productivity--we wanted to design a tool that would make developers more productive, all the way from prototype to production code. Other products lure you with visual tools, but once you get halfway through your project, they let you down because of sluggish performance, lack of extensibility, or general stability problems. The competition talks about adding extensibility and improving performance. That's a fundamental difference between their products and ours. Extensibility and performance was on the white-board the first day we started designing Delphi, and it permeates the entire product. For example, if you want to design a new component in Visual Basic, you have to write it in another language, such as C or C++ (or Delphi, for that matter). None of your VB skills can be reused, you have to learn a different language, and you can't easily inherit from any of the built-in components. Delphi, on the other hand, allows you to write new components in Delphi, and you can inherit from any of the built-in ones. That's true extensibility, and it translates into a substantial productivity boost.
Another key aspect of Delphi is its versatility. Other tools tend to focus either on Windows application development or on Client/Server development, and one always trades off the other. Delphi is equally adept at both, as is evident from the kinds of applications our customers are building. They range from shrink-wrap Windows utilities and multi-media games, through desktop database applications, and all the way up to multi-user enterprise-wide Client/Server solutions. The point is that almost any Windows application needs some form of database access, and any database application needs some form of Windows specific programming--to be productive, you need a tool that does both.
Delphi really leverages a lot of very mature database technology from Borland including ReportSmith, the Borland Database Engine, SQL Link native drivers for remove servers and the Local InterBase Server. Just the InterBase server alone is a tremendous technology that gives developers the ability to use full ANSI-92 SQL in their applications so they can begin exploring SQL and client/server development all on their local PC.
Q. You emphasise Delphi's versatility as an advantage, but surely it is also a drawback? If one needs to build a client/server application, PowerBuilder offers better CASE/database management facilities than Delphi.
A. There's an inherent advantage to being versatile. Look at the computer on my desktop. Do I need a dedicated word processor, a PC for my spreadsheets and a terminal with access to my customer records? No, I've got one PC that's versatile enough to do all these things.
A very large American retail chain--one of the largest--just standardized on Delphi over PowerBuilder precisely because their engineers can do 85% of all their work using Delphi versus 60% of their work using PowerBuilder. That saves them enormous amounts of money and complexity, including in ways you may not have considered. As an example, skills and techniques learned writing a small utility are directly applicable to client/server projects. A lot of today's programmers started out by writing those little command-line utilities in the good old days. It's a great way to experiment with and master the use of data structures, object-oriented techniques or learning about the Windows API. Consider, also, how using the same tool for a broad range of applications provides a company with a neat training path: Someone can start writing non-database programs and then gradually move onto projects dealing with valuable corporate data.
There's no end to the components and views you can add to Delphi. The population of programmers who can build components in Delphi is much larger than with any other tool on the market. We're back to the days when one programmer in one room can build and test something that can be used by tens of thousands of other people. Can you imagine what the
availability of specialized component sets will be like in six months? In a year?
I think the entire point of combining a component-based visual development environment with an object-oriented compiler and database technology is to make sure you never run out of gas. That's not a bug--it's a feature.
Q. If you want a quick and dirty hack, surely it makes sense to use Visual Basic, because everybody can use it without having to master a scary, complex language like Pascal. If you are doing multi-media or real time work, why mess around with a system which delivers slightly slower performance, and requires you to hand-translate all the header files for any DLLs you may need, when you could just use C++?
A. As we like to say, "It's not your father's Turbo Pascal any more". We made sure that the Object Pascal code you have to write is as easy as BASIC but without limitations.
We've taken great pains to make sure that when you're interacting with components, the code you write is as simple as possible--but no simpler. Many reviewers have remarked that they thought they were coding in Basic when they first started using Delphi. It's that easy. When they want to do something more interesting and start using the richness of the language, they usually start remembering how much they like Pascal.
In fact, I think you miss an essential advantage of Delphi. Anybody who has used a compiler--especially one that supports good type-checking--knows that a compiler is really a programmer's best friend. When it tells you it's probably not a good idea to take the square root of your Window caption, it's showing you a logic error in your code and saving you time. Is it an advantage that BASIC will perform automatic type conversions in that circumstance instead of giving you an error? I wish my spell checker program could complain about the logic of a paragraph I've written in the same way as our compiler warns you about illogical programming statements. Our 32-bit compiler goes even further and offers you all sorts of hints about problems it detects in your program. This kind of help is invaluable and one of the things that makes programming in Delphi very productive.
Q. What is the secret of Delphi's fast compile/link cycle?
A. Borland has over ten years of experience in building the world's fastest compilers, and we've put that knowledge to good use in Delphi- -it compiles at about 350,000 lines per minute on a 90 Mhz Pentium. A number of factors comtribute to this throughput. Delphi units (code modules) compile to .DCU files, which you can think of as a combination of a C++ precompiled header file and an .OBJ file. (It's funny how the hot topic in the C++ community is pre-compiled header files and incremental linking--Borland's Object Pascal technology has had these features for more than eight years.) Delphi units specify what other units they depend on through USES clauses--sort of like C++ #include's of header files. By analyzing the USES clauses of each unit in a project, the compiler can automatically perform minimal builds with no need for a make file. The net result is that the compiler never compiles more than it has to, and it never compiles the same thing more than once. Finally, the clean syntax of Object Pascal allows for very fast parsing.
Q. Is the compiler engine itself written in Delphi? How much does it differ from the Borland Pascal 7 compiler?
A. The compiler is written in assembly language. It is fully backwards compatible with BP7, and we've added lots of object-oriented extensions such as class references, virtual constructors, and the IS and AS operators. We did a lot of work to enable declaring, registering and filing properties and we generate run-time type information that's used to communicate published property, event and method information to the development environment. You'll see some interesting applications of that capability in our 32-bit release. One very unique enhancement was our use of bound method instance pointers to implement event delegation. They're very efficient and fit nicely into the language. And of course we did a lot of work to add structured exception handling. In addition, there are lots of little niceties that people have requested, such as support for C calling conventions.
Q. Delphi implements objects in a manner similar to Apple's Object Pascal, with all objects allocated on the heap. Previous versions of Turbo/Borland Pascal used a more C++ like approach, with the ability to alocate objects on the stack and statically. Can you explain the reasoning behind this change in approach?
A. It really is a question of features vs. complexity. The philosophy of Delphi's Object Pascal language is to deliver the RIGHT set of language features, as opposed to any language feature ever known to mankind. It's the well known 80/20 rule: You can get 80% of the power for 20% of the complexity, but squeezing out that last 20% of power makes the whole thing five times as complex to program. Mixing static and dynamic allocation of objects is one of those features that fall into the latter group. By implementing a pure reference model we were able to simplify the entire Delphi component library, and do away with a lot of the pointer management that plagues other products. Even though Delphi objects are allocated on the heap, in a typical Delphi application you never have to deal with allocating and freeing them.
Q. I find this answer quite surprising and counter-intuitive. You had already implemented mixed static/dynamic allocation, and therefore presumably cracked the problems involved, so why go to the trouble to revert to the Apple Object Pascal approach which you had initially rejected? Is, say, a stack allocated object, with constructors and destructors automatically called as the thing moves in and out of scope, really more complex than a heap allocated object, where you must make special provision to kill the thing off at the end of its life? I would have thought that the fact that the component library *mostly* frees objects automatically but *sometimes* doesn't would tend to add to rather than reduce the application programmer's burden. Also, the change in model must confuse both existing BP programmers and also migrating C++ users.
A. Again, we didn't revert to anything because we really started with a clean slate. Our class reference model is sufficiently powerful and flexible, so having only one sort of class is actually an advantage. Once you give someone two ways to do the same thing, you have made your product less usable and you have to now help them understand when to use a statically allocated class versus a dynamically allocated one. We're quite happy with the choice we've made. It's simple to understand, efficient, and allows us to add garbage collection in some future release. And, of course, if you've got old code from BP7 that uses old style objects, you can still compile it from within Delphi.
Q. Exception-handling - what were the major influences on your design?
A. We looked at a number of languages and implementations, and were most influenced by C++ and Modula-3. Delphi is like C++ in that exceptions are classes, but more like Modula-3 in terms of the supporting language constructs.
Exceptions are a quiet revolution--they truly simplify the way you write code. For the most part you can write your code as if errors will never occur, instead of spending the bulk of your time trying to determine if an error occurred, and if so, how best to clean up and back out of what you were doing. Delphi's Visual Component Library was designed from the ground up with exception handling built in, and that is a large part of the reason why Delphi and applications written in Delphi are so fault tolerant. One of my favorite demos is a little two-liner that, on the click of a button, assigns NIL to a pointer, and then dereferences the pointer. Each time you click the button, Delphi reports that a General Protection Fault exception has occurred, but because of the built-in exception handling logic, the app keeps running instead of bringing itself down.
Q. I'd like to draw you out a bit to expand the answer above with a few specifics.
A. As in C++, an exception in Delphi is simply a class, which means you can take advantage of the inheritance mechanism to handle whole sets of exceptions easily. For example, Delphi declares the following classes which deal with floating-point exceptions:
EMathError = class(Exception);
EInvalidOp = class(EMathError);
EZeroDivide = class(EMathError);
EOverflow = class(EMathError);
EUnderflow = class(EMathError);
As you can see, EMathError is the ancestor of the other exceptions. Here's an example of a TRY..EXCEPT statement that handles floating-point exceptions
on EZeroDivide do ...;
on EMathError do ...;
If the PerformCalculations procedure raises an EZeroDivide exception, it is handled by the first handler. If it raises any other EMathError exception, the second handler takes care of it. Since there is no ELSE clause, no other exceptions are handled--they are instead propagated to an enclosing exception handler.
Q. Delphi's ability to handle GP faults is indeed one of it's neatest tricks. Was it difficult to implement?
A. It wasn't too bad, but it did take some nitfy use of TOOLHELP.DLL which implements the Windows low-level system tools interface. We basically register an interrupt callback function which maps processor faults into Delphi exceptions. The reason that it all works, though, is that VCL was engineered from the ground up to be exception aware. Because
of that, when a GP fault occurs and is mapped into an exception, the operation that was in progress will automatically know how to back out and clean itself up.
Q. Can we expect any other major syntax additions/changes, for example Eiffel style assertions?
A. We're always evaluating new language features, and surely there will be some in the upcoming 32-bit version. I'd rather not get into specifics, but as a rule, we don't really think about language extensions in the abstract. Instead we look at the language as part of a bigger picture (class library, component model, visual environment) that must evolve as a whole to support new technologies and improve ease of use.
Q. Can you give Delphi programmer's any guidance on how best to write applications that will be portable to the 32-bit version of Delphi? The new "Cardinal" data type has arrived almost completely unnoticed. Are there any other issues we should be aware of?
A. Delphi's Visual Component Library was designed with portability in mind. As long as you stay away from in-line assembler, 16-bit pointer atrithmetic, and Windows 3.1 API functions which aren't supported in the Win32 API, your apps should port with little or no modification.
The Cardinal and Smallint types were introduced to facilitate portable code. Of the built-in types, Shortint, Smallint, Longint, Byte, and Word have identical representations in 16- and 32-bit code. The Integer and Cardinal types, on the other hand, represent the most efficient signed and unsigned integer types of the particular platform. In the 16-bit version they are 16-bit entities, and in the 32-bit version they are 32-bit entities. In general, you should use Integer and Cardinal whenever possible, and Shortint, Smallint, Longint, Byte, and Word only when the exact storage representation matters.
Any 64K limitations found in the 16-bit version will disappear in the 32-bit version. For example, the 32-bit version allows you to declare arrays and allocate heap blocks of any size up to 4GB!
Q. What is the current state of the 32-bit version? Will it support 16-bit VBXs, like BC++? Delphi 16-bit code runs somewhat slower than C++ - are you doing anything about this for the 32-bit version?
A. Delphi was written to be portable--we've been working on the 16- and 32-bit versions in parallel since day one. The 32-bit version is in field test now, and it will ship shortly after the commercial release of Windows 95. Yes, there is a foundation of 32-bit VBX support technology available in-house, but our primary focus is OCX controls. That's what the competition is working on, and that's where we see the market going. With respect to better code generation, Delphi-32 generates the same high-quality code as Borland C++ 4.5--in fact, they use the same optimizing back-end code generator.
Q. Are there any plans for Borland produced or badged add-ons for Delphi, in addition to the Visual Solutions Pack?
A. We just released the RAD Pack for Delphi, which includes Turbo Debugger for Windows, Resource Workshop, the Resource Expert, Visual Component Library source code, the--much requested--Language Reference Manual, and Visual Solutions Pack 1.1. We did have some quality problems with the initial release of VSP, but those have been resolved, and we now have a Companion Products group to provide Borland-quality add-ons, such as Notes support for Delphi programmers and other often requested components.
Q. Delphi is a terrific tool for rapidly developing state of the art software, but a number of shareware authors have expressed a wish that executables could be made smaller. Is it technically feasible to create a DLL-based version of VCL? Surely this must be possible since COMPLIB is a DLL which is used by the Delphi design environment?
A. It's something we're looking at, and certainly some of the 16-bit complexities with respect to multiple DLL clients are gone in 32-bit land. At this point I can't really comment on specific solutions, other than to say that we're actively looking at ways to make our executables even smaller.
Q. A long-standing and major criticism of Borland Pascal is the proprietary nature of the object file format. It's appreciated that going to the OBJ file format would be a retrograde step, but why won't Borland at least document the file format? That way, developers can create their own tools such as disassemblers, C to Pascal linkers and so forth. Again, it's understood that the file format changes with each release of the compiler, but documenting the changes with each new version would enable other developers to create conversion tools even if Borland don't want to do this. At the moment, if you don't have the source code, all your units become useless each time the compiler is updated.
A. We're well aware of these issues, and the 32-bit version will address them in a number of ways. What I can tell you at this point is that the 32-bit compiler has an option to produce
.OBJ files, which can be linked with .OBJ files produced by other compilers.
Q. A related issue: the move to Windows has diluted the importance of the OBJ issue, because you can now call DLLs. But the Delphi user has still to translate the (typically) C/C++ headers into Delphi import units, an exercise which is at best tedious and time consuming and, if you happen not to have had C++ experience, quite hard. It's the sort of job best left to a machine. Given that Borland has a lot of C++ parsing expertise lying around on the ground, have there ever been any plans to create such a tool?
A. Well, I'm not sure which C/C++ headers you're talking about.
We've already translated all the Windows and OLE 2 API header files, and corresponding interface units are included with Delphi. But you're right, if you have a 3rd party DLL that was previously only interfaced to C/C++, somebody will have to do the translation. Usually, it's not that bad and I think you'll see an increasing number of vendors providing Delphi interface files for their DLLs. Also, I think you'll see more and more products take advantage of the OLE 2 ITypeLib and ITypeInfo interfaces, and we'll provide a tool that takes that information and produces a Delphi interface unit.
Q. The ability to create a single EXE for redistribution is very attractive, but somewhat spoilt by the need to include the BDE with database applications, even if they only want to access the odd DBF. Any plans to clean this up?
A. We're working with several third parties, including SAX Software, Eschelon Development, Sterling Software, Great Lakes Software, and Shoreline Software. They have, or will soon have, products to help you deploy your Delphi database applications. In addition, we're making a deployment kit available, via CompuServe and Internet.
Q. Also on data access: is it possible to modify/inherit from the data access controls to provide, for example, 'native' access to FoxPro/Clipper databases? If so, are any such products being developed by Borland or Third Parties?
A. I know of several Third Parties working on native access to FoxPro/Clipper as well as B-Trieve. Some of them are in beta at this point. You can contact them for additional information on the DELPHI CompuServe forum or find out about them in the Delphi "Power Tools" catalog.
Just to clarify: Sax, Eschalon, Sterling etc do install/setup tools that work with Delphi and the Borland Database Engine. There are other companies creating components that provide direct access to BTrieve, Fox, and other database formats.
Q. What is Delphi's main niche in the developer tools market? Compared with, say, Visual Basic, PowerBuilder and C++?
A. Delphi is a general purpose Windows RAD (Rapid Application Development) tool. The point is that Delphi is NOT a niche tool. From the onset, we've designed Delphi to be able to take you from prototype to production, be that whether you're targetting a Client/Server environment or just writing a Windows application. I hear our competitors say that you can use their tool for rapid prototyping, and then port your app to C++ for production. But you know, rapid application development isn't really rapid unless you can go from prototype to production, all using the same tool! I also hear how competing products will address performance issues by generating C or C++ source code. This idea of building the application with one tool, and then having it generate C or C++ source files that have to be run through another tool, is ludicrous. How are you expected to debug the final code? Requiring users to find bugs in machine generated C++ code, and understand how that maps to the original 4GL code, just doesn't make sense. We've been shipping development environments with integrated compilers for 12 years--I think the time is gone when programmers would accept anything else.
Q. How have Borland's recent troubles affected Delphi's development and its take-up? Was the absence of a Language Reference Manual in the initial product a consequence of these troubles? (You had to mention the Language Reference Manual? My mistake, I'm sorry! --Zack)
A. In the two years we were developing Delphi, the company did go through some difficult times. That was all resolved before we shipped. Now the entire company is focused on development tools, we've won the Lotus lawsuit, we've launched Delphi and Delphi Client/Server worldwide, and both products continue to sell well above expectations. In fact, I understand that Gray Matter reports that Delphi is the #1 selling development tool in the UK. The Delphi development team is 100% intact, and focused on Delphi for Windows 95.
We really underestimated the demand for the Language Reference Manual. (You're telling me! --Zack) It will be included in Delphi for Windows 95. Meanwhile, we've made good by uploading an Adobe Acrobat version to CompuServe and our WWW page, and a printed version is now also available from Borland.
Q. We are all pleased you resisted the opportunity to christen a product Power Visual Turbo Pascal Objects for Windows - but how did it come to be called `Delphi'?
A. We actually tried to call it "Power Visual Turbo Pascal Objects for Windows", but that name was already trademarked :-). One of the senior guys in QA (Danny Thorpe) dreamed up Delphi as a code name quite early on, and everytime we did a market survey of product name candidates, everyone said "well, those are ok, but we really like `Delphi'". So in the end, we kept it.
Q. Which part of Delphi are you most proud of? ... and which part least?
A. The thing I'm the least proud of is probably the initial lack of a Language Reference Manual. But that's taken care of now. (Dammit, I said I was sorry already! --Zack)
What I'm most proud of is the fact that the energy we invested in foundation technologies like extensibility and exception handling enabled us to build Delphi in itself. Can
you imagine VB or PowerBuilder written in themselves? By building Delphi in Delphi, we really got to feel on our own bodies what was right about the product, and what needed fixing. I sometimes hear frustrated users comment "The programmers that wrote this %&##$ thing should be forced to use it themselves!". Well, we did, and we're really proud of the result.
-------- CHRONOLOGY --------
- 1960 Anders born in Copenhagen, Denmark.
- 1979 Enrolls at the Danish Engineering Academy. Co-founds PolyData, one of the first Danish microcomputer companies.
- 1980 Releases his first Pascal compiler--a 12K Pascal subset in ROM for the British NASCOM Z-80 based kit computer. Eventually sells the rights to this product to Lucas Logic.
- 1982 PolyPascal for CP/M-80 released. Product is now a complete implementation of the Pascal language.
- 1983 Sells the Borland founders (Niels Alex Jensen, Ole Henriksen, Mogens Glad, and Philippe Kahn) on the idea of a Pascal compiler with an integrated editor. In November releases Turbo Pascal 1.0 for CP/M-80, CP/M-86, and MS-DOS. The newly formed Borland company, essentially penniless, places an advert for Turbo Pascal in Byte, bluffing the Byte Ad executives into giving them credit. The compiler is priced at $49.95 and is an instant smash hit.
- 1986 Turbo Pascal 4.0 released, featuring an Integrated Devlopment Environment--the first of its kind for the PC environment--and introducing modular compilation (previously Turbo Pascal programs had to be compiled all in one go and could be no larger than 64K unless overlays were used). CP/M support is dropped.
- 1988 Turbo Pascal 5.0 released, featuring integrated debugging and VROOMM (Virtual Runtime Object Oriented Memory Manager) overlay management technology.
- 1989 In response to Microsoft's object oriented QuickPascal, Borland releases Turbo Pascal 5.5, which has its own OOP extensions. Microsoft later drops QuickPascal from its product line.
- 1990 Turbo Pascal 6.0 features a new, much improved Integrated Development Environment, and includes the Turbo Vision object-oriented application framework.
- 1991 First release of Turbo Pascal for Windows. Features a Windows hosted IDE and the ObjectWindows Library (also known as OWL).
- 1992 Borland Pascal 7.0 includes both a DOS and a Windows hosted IDE, and allows developers to target DOS, DOS Protected Mode, and Windows.
- 1995 Delphi and Delphi Client/Server released on schedule on February 14th.
This interview originally appeared in EXE, the UK's leading programming magazine. If someone can send me a .EXE cover or logo, that is appreciated!
Thanks to Ben Riga for sending it to me. Thanks for also not mentioning the Language Reference Manual.