If you haven't lived through this era, try one day to download DOSBOX and run Turbo Pascal. You will be impressed with how fast it compiles on an old machine... Often faster than most modern projects on our modern computers.
cowlevel 18 hours ago [-]
Probably no optimization though, right?
joakleaf 16 hours ago [-]
You could enable optimizations. Turbo Pascal generated pretty fast production code in general, and you could also disable run-time checks like range checking etc.
The generated code wasn't perfect, but you could always use assembler for short routines that needed faster code (e.g. loop-unrolling, tighter register usage, or 80386 32-bit register code, which I think TP never supported). There were also tricks that would improve performance (I seem to remember that 'repeat'/'until' was faster than standard 'for' loops).
To be honest, it was a very fast tool to work with (even on a 80286) -- The integrated debugger was a dream to work with for the time too.
I remember switching to C (both Turbo C and later gcc via DJGPP on both a 80386 and 80486). Compiling even standard C was agonizingly slow in comparison to TP, and you couldn't iterate as fast as TP. It was like moving from having fun programming to plain torture.
I moved on to fully embrace C later (and C++). As the computers got more RAM and got faster, it became a lot better experience, but I never fully understood why C was popular as a general purpose language in the 80s and 90s. I doubt C was optimal for development time.
TP was a really balanced compromise in a lot of ways.
switchbak 14 hours ago [-]
For some reason they forced us to use these ancient 8088 machines (in 1996 no less!) at my school. Even there, the speed of Turbo Pascal was just incredible!
Sure things were very spartan, but you could do a heck of a lot on what is basically a trumped up microcontroller. I have mostly felt that modern tech has been a step backwards - because the inner loop has so much latency. Turbo Pascal and Delphi were really leaders in this regard. Python and friends were good too (aside from the runtime errors!), but getting a real compiled and lean binary so quickly was a very special thing. I didn't enjoy moving to Java after that, especially in those early days.
If only I had known about TDD back then, I think ultra low latency compiles and fast running tests would be hard to beat.
jhallenworld 12 hours ago [-]
>I never fully understood why C was popular as a general purpose language in the 80s and 90s
Depends on what you mean by general purpose I guess, but I remember I liked it because of C strings. It's difficult to understand these days, C's strings were better than those in other languages: no 255 byte length limit, 0-based instead of 1-based, escape sequences so that you could have any byte in a string (Turbo Pascal supported control characters, but PASCAL in general did not). Other languages at the time really only supported fixed length space filled strings.
Another big thing was existence of malloc().
joakleaf 11 hours ago [-]
Pascal strings were traditionally indeed maximum 255 bytes. Later versions of TP has PChar which worked as C-string (max 64K chars).
Null-terminated requires that you have sentinel (null) at the end, instead of a length at the beginning.
Having a length (encoded as a byte at the beginning), means they are very easy and efficient to copy and concatenate, because unlike the null-terminated variant, you don't have to search for the terminator which requires a conditional branch on each character until the length.
Conditional branches were really expensive on old CPUs, compared to the 8086's rep movsb instruction, that could just copy a specific length of bytes. On the other hand, I suppose, other string-operations are more tricky, because you need to keep track of your position in the string as you iterate over it.
I've never thought null-terminated strings were a particularly neat design or easy to work with (there are pros and cons) -- But the 255 string limit was annoying, and I wished they would have had pascal strings up-to 64K in length (they could have used a word as length instead of a byte).
Malloc is by no means a 'C' specific thing -- Turbo Pascal had getmem() to allocate memory on the heap.
pjmlp 17 hours ago [-]
There were optimizations already available.
Even if the basic ones that tend to be part of university compiler assignments, .e.g. peephole optimisations, constant propagation, inlining, bounds checking removal, redundant loads removal,....
See chapter 21 from Turbo Pascal 7 Language Guide, to use the last one for MS-DOS as example.
Similar build performance available in other languages, alongside a small set of optimisations.
It is a urban myth that we cannot have both, only because folks experience is tainted with C and C++ compilers.
If you want a good example of fast compilation times with optimisations enabled, check D reference compiler dmd, or to keep on the subject, FreePascal or Delphi 13.
hnlmorg 17 hours ago [-]
Optimisations existed back then. But what Pascal did differently to a lot of modern languages was it only required a single-pass compiler and no macros.
On slow systems of the era, that vastly reduced the amount of work a compiler had to do to turn ASCII into machine code.
These days, compilers run macros, multiple passes of the source code. Often then building their own byte code or an interim machine representation, before turning that into machine code.
Those layers of abstraction aren’t without benefit. You get greater composability of code, greater portability, and so on and so forth. So they exist for a reason.
switchbak 14 hours ago [-]
I often wonder how much benefit we get from most of those optimizations. I rarely felt that Turbo Pascal's runtime was slow or slow enough to cause frustration over something like C.
I think we could do something similar today - very fast compilation, and straightforward but efficient machine code. And perhaps bolt on the heavy duty optimizations where it is needed. Many JIT's work similar to this, but it seems like you could do it in an AOT style too.
Pannoniae 11 hours ago [-]
Yup IMO there's a sorely missing space for a language with very low-level capabilities which doesn't rely on auto-optimisation. C++ is only relatively fast because of the insane amount of optimisations the compilers have + compiler -specific directives.
But I mean more like
- being able to explicitly tell to unroll this loop N times
- native support for builtin SIMD vector types (it can use platform-dependent instructions, whatever)
- specifying aliasing manually, built-in support for range types
- having explicit control over switch statement code generation
- good metaprogramming (doesn't matter if compile-time maths doesn't exactly match runtime behaviour), compile-time code generation/modification
- [[likely]] hints but actually respected i.e. having a guaranteed effect on code layout
- of course no headers, the unit of compilation being a whole library
I don't think it really needs to be single-pass in order to be fast, just cut out most of the heuristic-based optimisations from the backend and make them explicit keywords/attributes in the code...
SleepyMyroslav 5 hours ago [-]
I think this is one of the "paths not taken" already. The economy of making software (by people) voted in favor of "sufficiently smart compiler" with profile guided and link time optimizations. It happened a long time ago and most of the code that I have worked with is structured to require such a smart compiler. The early examples of half codebase in assembly like Photoshop 1.0 in the public archive were clearly more expensive to make for a single platform - not to mention how many platforms people need now to recoup development costs.
What you are describing is what Zig spent years to try to become while running away from LLVM. Sure thing there are other less popular attempts which I can't even bother to list here, not because they are somehow worse, but because they are even further away from gamedev practice that I care about.
I don't know what the economy of generated code would be. It is up for debate right now.
What I can definitely see now is that human-only code made this way is not going to pay off. It can only be an expense for someone who has unlimited budget to spend on a whim.
wahern 8 hours ago [-]
You're describing Go. Only difference was Go settled on GC, to avoid alot of semantic complexity with memory management across threads.
wolvesechoes 13 hours ago [-]
> I often wonder how much benefit we get from most of those optimizations
Well, in a world where many among most used apps are made of NodeJS and Electron frontend, I think we are very far away from having to worry about those optimizations in most contexts. Having those written in TP with zero optimizations would still be more resource efficient and performant.
And when it comes down to squeezing last drop of performance, we forget compilers and still rely on assembly wizards tuning BLAS routines, crypto code etc.
switchbak 13 hours ago [-]
Not to mention all of the superhuman efforts that chip makers take to make existing app patterns (typically driven by C) perform acceptably.
It does make you wonder how much we could simplify and take away while still having things be blazing fast. I would assume some of the old optimizations are actually hurting things on newer architectures.
OskarS 17 hours ago [-]
Optimizations is one reason for modern languages to compile slowly, but it's certainly not the only reason. Even at -O0, C++ and Rust are much slower to compile. The culprits are more likely much more advanced type-systems and monomorphization passes for generics. Slow compile times are much more inherent to the languages (compared to Turbo Pascal, certainly).
pjmlp 17 hours ago [-]
Not even, regarding type systems, see D or Ada compilers.
Or if you prefer to travel back to 1993, here is a developer experience using C++ as if it was Lisp Machine/Smalltalk environment with incremental compilation and hot code reloading.
We got into a culture that accepted slow compilers is normal, it doesn't have to be.
vidarh 16 hours ago [-]
One of the noteworthy papers I remember reading in the early 90's was a demonstration that you could even add SSA form to ETH's Oberon compiler to apply optimisations equivalent to the then-current version of gcc with very little loss in compilation speed.
I guess today if you want fast, optimized codegen you need to look more at JIT research since it seems the AOT compiler research seems as you say to have just accepted that slow compilers is normal.
pjmlp 16 hours ago [-]
Yes, spot on.
A good mix is having a JIT/AOT combo, use JIT for the development workflow, and a full blown AOT with all possible optimisations turned on for production releases.
wolvesechoes 13 hours ago [-]
Not really. In addition to Ada and D mentioned by a sibling comment, Ocaml or F# have even richer type systems than Rust, and compile times are better.
Yes, compiling Rust in the release mode could take long time and it would be worth all this baggage - monomorphization, trait resolution etc. But there is no reason, except not giving shit, to no have a fast and simple pipeline for debug builds.
whobre 14 hours ago [-]
If you are interested in an optimizing Pascal compiler from that era, check DRI Pascal MT/+
muyuu 17 hours ago [-]
they are pretty minimal
the game back then was to use inline assembly for the critical bits, this really worked a treat for the meta of the 90s, with computers that were being asked for multimedia processing with slow CPUs, little RAM and typically no hardware acceleration
summa_tech 17 hours ago [-]
As a positive side of that tradeoff, inline assembly in Turbo Pascal was extremely pleasant to write. Because the optimizations were so light, it did not require complex constraint specifications on the values passed in and out, so you could just refer to local variables directly.
vidarh 16 hours ago [-]
My first compiler was for a Pascal-like language that started with just reading assembler in and spitting assembler out, and then I layered in blocks and expressions step by step and slowly rewrite the compiler itself from a basic text processor in assembler to a proper compiler, making use of each new feature as soon as I'd made sure it worked.
I wish I still had the code for it - because it had a very simple Wirth-style code generator it was predictable enough that you could even use registers directly in expressions. (M68k, so e.g. you could do "x = D7.w" to set the variable x equal to the low 16 bit word of register D7)
EDIT: Worth mentioning Randall Hyde's "High Level Assembly"[1] here as an actually published language that did something similar.
Good to see. Delphi was the able successor and thriving till date!
Held the top for being one of the fastest compiler (it still crazy fast sans optimization) on the planet for a good period!
On a side note,
Good ol'days! Enjoyed the Borland offerings to the core. Things were too advanced for the time and on DOS!
Brings back some good memories though!
plinkplonk 16 hours ago [-]
you have to pay to get the 32 bit compiler source code.
(nothing wrong with this, this information is way down the page, and it might help someone else wondering if this were open source)
"Most of the source code is published with comments on Turbo Pascal website. Here you will be able to get all the internals of the compiler. However, if you would like to use this compiler for your own project and you need the complete TPC16 package with all source files then please continue reading. You can get TPC16 Turbo Pascal compiler source code for some small compensation"
TPC 32 - 150 Euros for two downloads in the next 24 hours.
In the old days, it was quite a common trick to write a sketchy compiler in assembler (or whatever you had) then rewrite it in itself. Iterate two or three times and you had something.
Slapshot_gd 15 hours ago [-]
Wow, the memories from this!! Using inline assembly for mouse support was amazing at the time... Heck, I remember writing a UI/shell for Home Management (included basic apps, and a neat feature to estimate calculate your monthly power bills), that I ran under OS/2 (with win32 support). Ended-up winning a science fair with it in high school in 1993, prize being a copy of OS/2 Warp. I still have that shrinked-wrapped at home :)
As a nerd of a certain age, at one time I had a mighty shelf of programming references from Waite Group and Swan and (of course) O'Reilly. Over time, I let nearly all of them go because it's faster to have stuff available digitally now (and has been for 20 years or more), but I kept a few as keepsakes.
The survivors live on a shelf about six feet from where I'm typing, and there you'll find my K&R book, Booch's work on OO, my college FORTRAN textbook, an O'Reilly twofer (a camel and a baboon; extra points if you can identify those by animal alone), and a fat blue tome called Mastering Turbo Pascal 5.5. It has a rendering of a 911 on the cover. My pal Mike gave it to me for my birthday in probably 1990, back when we were both in college and writing code for a computer lab at the business school there.
Years turned into decades, and we remained close. Last summer, my wife and I relocated to North Carolina, in part because Mike and his wife live here. I'm having lunch with him today, just as I do every week.
Sometimes, good languages make good friends.
dolmen 17 hours ago [-]
I started programming with Turbo Pascal at high school. I made a few games in DOS text mode.
Nowadays my favorite language is Go. It appears that among the programming languages that I did use, Go is the one that feels the nearest to Turbo Pascal.
jan_m_savage 16 hours ago [-]
that's cuz one of the designers of Go was a student of Wirth's.
ferguess_k 15 hours ago [-]
Do you happen to work as a first responder in the past?
ubermonkey 13 hours ago [-]
Not me, no.
000000000001 16 hours ago [-]
Baboon - JUnit pocket guide?
Camel - Perl Programming :)
ubermonkey 13 hours ago [-]
Camel is, of course, Learning Perl.
Baboon is Embedding Perl in HTML with Mason.
myth_drannon 12 hours ago [-]
Tom Swan has a bunch of huge Mastering Turbo X books, I have Turbo Assembler one.
Blazing fast compiler (on that era's hardware), glorious TUI framework (Turbo Vision) and a nice help system. Anders Hejlsberg is a genius.
After using this and Delphi and then having to develop UIs on the web it was like losing an arm.
If I recall correctly it already has a few years behind it, thus it was actually written by someone that can use the donation.
> This is not the "original" Borland source code of the Turbo Pascal since it was probably not written in Pascal.
Indeed, it was written in Assembly during the CP/M days, and slowly got to have more Turbo Pascal into it as the language evolved.
The generated code wasn't perfect, but you could always use assembler for short routines that needed faster code (e.g. loop-unrolling, tighter register usage, or 80386 32-bit register code, which I think TP never supported). There were also tricks that would improve performance (I seem to remember that 'repeat'/'until' was faster than standard 'for' loops).
To be honest, it was a very fast tool to work with (even on a 80286) -- The integrated debugger was a dream to work with for the time too.
I remember switching to C (both Turbo C and later gcc via DJGPP on both a 80386 and 80486). Compiling even standard C was agonizingly slow in comparison to TP, and you couldn't iterate as fast as TP. It was like moving from having fun programming to plain torture.
I moved on to fully embrace C later (and C++). As the computers got more RAM and got faster, it became a lot better experience, but I never fully understood why C was popular as a general purpose language in the 80s and 90s. I doubt C was optimal for development time.
TP was a really balanced compromise in a lot of ways.
Sure things were very spartan, but you could do a heck of a lot on what is basically a trumped up microcontroller. I have mostly felt that modern tech has been a step backwards - because the inner loop has so much latency. Turbo Pascal and Delphi were really leaders in this regard. Python and friends were good too (aside from the runtime errors!), but getting a real compiled and lean binary so quickly was a very special thing. I didn't enjoy moving to Java after that, especially in those early days.
If only I had known about TDD back then, I think ultra low latency compiles and fast running tests would be hard to beat.
Depends on what you mean by general purpose I guess, but I remember I liked it because of C strings. It's difficult to understand these days, C's strings were better than those in other languages: no 255 byte length limit, 0-based instead of 1-based, escape sequences so that you could have any byte in a string (Turbo Pascal supported control characters, but PASCAL in general did not). Other languages at the time really only supported fixed length space filled strings.
Another big thing was existence of malloc().
Null-terminated requires that you have sentinel (null) at the end, instead of a length at the beginning.
Having a length (encoded as a byte at the beginning), means they are very easy and efficient to copy and concatenate, because unlike the null-terminated variant, you don't have to search for the terminator which requires a conditional branch on each character until the length.
Conditional branches were really expensive on old CPUs, compared to the 8086's rep movsb instruction, that could just copy a specific length of bytes. On the other hand, I suppose, other string-operations are more tricky, because you need to keep track of your position in the string as you iterate over it.
I've never thought null-terminated strings were a particularly neat design or easy to work with (there are pros and cons) -- But the 255 string limit was annoying, and I wished they would have had pascal strings up-to 64K in length (they could have used a word as length instead of a byte).
Malloc is by no means a 'C' specific thing -- Turbo Pascal had getmem() to allocate memory on the heap.
Even if the basic ones that tend to be part of university compiler assignments, .e.g. peephole optimisations, constant propagation, inlining, bounds checking removal, redundant loads removal,....
See chapter 21 from Turbo Pascal 7 Language Guide, to use the last one for MS-DOS as example.
Similar build performance available in other languages, alongside a small set of optimisations.
It is a urban myth that we cannot have both, only because folks experience is tainted with C and C++ compilers.
If you want a good example of fast compilation times with optimisations enabled, check D reference compiler dmd, or to keep on the subject, FreePascal or Delphi 13.
On slow systems of the era, that vastly reduced the amount of work a compiler had to do to turn ASCII into machine code.
These days, compilers run macros, multiple passes of the source code. Often then building their own byte code or an interim machine representation, before turning that into machine code.
Those layers of abstraction aren’t without benefit. You get greater composability of code, greater portability, and so on and so forth. So they exist for a reason.
I think we could do something similar today - very fast compilation, and straightforward but efficient machine code. And perhaps bolt on the heavy duty optimizations where it is needed. Many JIT's work similar to this, but it seems like you could do it in an AOT style too.
But I mean more like
- being able to explicitly tell to unroll this loop N times
- native support for builtin SIMD vector types (it can use platform-dependent instructions, whatever)
- specifying aliasing manually, built-in support for range types
- having explicit control over switch statement code generation
- good metaprogramming (doesn't matter if compile-time maths doesn't exactly match runtime behaviour), compile-time code generation/modification
- [[likely]] hints but actually respected i.e. having a guaranteed effect on code layout
- of course no headers, the unit of compilation being a whole library
I don't think it really needs to be single-pass in order to be fast, just cut out most of the heuristic-based optimisations from the backend and make them explicit keywords/attributes in the code...
What you are describing is what Zig spent years to try to become while running away from LLVM. Sure thing there are other less popular attempts which I can't even bother to list here, not because they are somehow worse, but because they are even further away from gamedev practice that I care about.
I don't know what the economy of generated code would be. It is up for debate right now.
What I can definitely see now is that human-only code made this way is not going to pay off. It can only be an expense for someone who has unlimited budget to spend on a whim.
Well, in a world where many among most used apps are made of NodeJS and Electron frontend, I think we are very far away from having to worry about those optimizations in most contexts. Having those written in TP with zero optimizations would still be more resource efficient and performant.
And when it comes down to squeezing last drop of performance, we forget compilers and still rely on assembly wizards tuning BLAS routines, crypto code etc.
It does make you wonder how much we could simplify and take away while still having things be blazing fast. I would assume some of the old optimizations are actually hurting things on newer architectures.
Or if you prefer to travel back to 1993, here is a developer experience using C++ as if it was Lisp Machine/Smalltalk environment with incremental compilation and hot code reloading.
"Lucid Energize Demo"
https://www.youtube.com/watch?v=pQQTScuApWk
With one of the very first ideas of a LSP could be like,
https://dreamsongs.com/Files/Energize.pdf
We got into a culture that accepted slow compilers is normal, it doesn't have to be.
I guess today if you want fast, optimized codegen you need to look more at JIT research since it seems the AOT compiler research seems as you say to have just accepted that slow compilers is normal.
A good mix is having a JIT/AOT combo, use JIT for the development workflow, and a full blown AOT with all possible optimisations turned on for production releases.
Yes, compiling Rust in the release mode could take long time and it would be worth all this baggage - monomorphization, trait resolution etc. But there is no reason, except not giving shit, to no have a fast and simple pipeline for debug builds.
the game back then was to use inline assembly for the critical bits, this really worked a treat for the meta of the 90s, with computers that were being asked for multimedia processing with slow CPUs, little RAM and typically no hardware acceleration
I wish I still had the code for it - because it had a very simple Wirth-style code generator it was predictable enough that you could even use registers directly in expressions. (M68k, so e.g. you could do "x = D7.w" to set the variable x equal to the low 16 bit word of register D7)
EDIT: Worth mentioning Randall Hyde's "High Level Assembly"[1] here as an actually published language that did something similar.
[1] https://en.wikipedia.org/wiki/High_Level_Assembly
Held the top for being one of the fastest compiler (it still crazy fast sans optimization) on the planet for a good period!
On a side note,
Good ol'days! Enjoyed the Borland offerings to the core. Things were too advanced for the time and on DOS!
Brings back some good memories though!
"Most of the source code is published with comments on Turbo Pascal website. Here you will be able to get all the internals of the compiler. However, if you would like to use this compiler for your own project and you need the complete TPC16 package with all source files then please continue reading. You can get TPC16 Turbo Pascal compiler source code for some small compensation"
TPC 32 - 150 Euros for two downloads in the next 24 hours.
https://turbopascal.org/turbo-pascal-internals
As a nerd of a certain age, at one time I had a mighty shelf of programming references from Waite Group and Swan and (of course) O'Reilly. Over time, I let nearly all of them go because it's faster to have stuff available digitally now (and has been for 20 years or more), but I kept a few as keepsakes.
The survivors live on a shelf about six feet from where I'm typing, and there you'll find my K&R book, Booch's work on OO, my college FORTRAN textbook, an O'Reilly twofer (a camel and a baboon; extra points if you can identify those by animal alone), and a fat blue tome called Mastering Turbo Pascal 5.5. It has a rendering of a 911 on the cover. My pal Mike gave it to me for my birthday in probably 1990, back when we were both in college and writing code for a computer lab at the business school there.
Years turned into decades, and we remained close. Last summer, my wife and I relocated to North Carolina, in part because Mike and his wife live here. I'm having lunch with him today, just as I do every week.
Sometimes, good languages make good friends.
Nowadays my favorite language is Go. It appears that among the programming languages that I did use, Go is the one that feels the nearest to Turbo Pascal.
Camel - Perl Programming :)
Baboon is Embedding Perl in HTML with Mason.
* https://news.ycombinator.com/item?id=49967652