Undefined behavior was, in most cases, intended to encompass a point where different implementations behaved in reasonable ways. In case of overflow, some implementations wrapped, some saturated, some threw an exception, some jumped to an interrupt handler, and some returned 0. They made it undefined intending to encompass all of these. Notice they are all fairly reasonable (except for returning 0).
Later, stupid compiler writers decided that since the wording allows anything to happen, it also allows time travel or memory corruption. These are not reasonable, but they are allowed by the wording of the standard.
The standard should be fixed to say it returns an undefined value or activates an error handling mechanism. Nobody could argue that time travel is an error handling mechanism.
> It also means that the behavior needs to be stable across time
I don't think IB has to be stable over time? For instance, whether and which extended integer types exist is IB and that has definitely changed over time.
> and that tools cannot report this behavior as a bug.
It's absolutely possible to report IB as a bug (or even fully-specified behavior, as demonstrated by -fsanitize=unsigned-integer-overflow or -fsanitize=implicit-integer-conversion).
A good example of C++ bizarre design decisions. The exact handing of float<->int conversions should use implementation-defined behavior instead, defining some cases to be UB just adds more traps with little to no performance benefits.
The same is true for bit-shifts greater than type bit size and negative bit-shifts.
They are not bizarre but a good compromise between consistency and optimization for any possible existing and future CPU architecture.
What the author correctly points out to is that GSL implementation does not remove that undefined behaviour as it lacks the proper checks. And that is the problem as it does not fulfill its own documented guarantees.
tldr; C++ is just fine. The implementation in the library is wrong.
No it isn't. Using undefined behavior for things which can be implementation-defined behavior instead is harmful. It creates more space for bugs and security vulnerabilities.
C++ was designed as an universal way of writing low level system code. It has some abstractions above C, but still allowing for all kinds of optimizations. (Think using a class vs a template, similar abstraction different impact).
To use C++ to code user programs is the actual mistake.
Think about assembly. Should assembly also be safe in all calls? That will make everything slower, and consume way more energy. Will you blame assembly if people used it to create a phone app and it had memory leaks? Or will you blame the developer for using assembly in the first place?
C++ is a tool designed for a range of purposes. There is a push to make it an universal language that makes everything well. But engineering is about making compromises.
For your use, C++ may be the wrong language. I do not disagree with you. But maybe you should be using another language instead. Or we should create a new language that fills the sweet spot that C++ covers, and then evolve C++ to be the safe language that you want. One way or another, it is a compromise.
I thought C++ was designed to fit all sorts of applications. There is many GUI libraries for C++ for example, and the language has all the capabilities to add abstractions for making these sorts of libraries easy to work with. I think you could argue that C was made for low level code, since the working group seems to be focussing on making the core of the language more usable instead of allowing for abstractions that make libraries easier to work with (I think we can all agree that the C interface of gtk is not that great to work with, though I do commend their efforts).
Sorry for rambling so much, I hope I got my point across.
Undefined behavior was, in most cases, intended to encompass a point where different implementations behaved in reasonable ways. In case of overflow, some implementations wrapped, some saturated, some threw an exception, some jumped to an interrupt handler, and some returned 0. They made it undefined intending to encompass all of these. Notice they are all fairly reasonable (except for returning 0).
Later, stupid compiler writers decided that since the wording allows anything to happen, it also allows time travel or memory corruption. These are not reasonable, but they are allowed by the wording of the standard.
The standard should be fixed to say it returns an undefined value or activates an error handling mechanism. Nobody could argue that time travel is an error handling mechanism.
> They made it undefined intending to encompass all of these.
Why wouldn't implementation-defined behavior suffice in this case?
IB means the implementation has to fully document what it does. Evidently they didn't want to impose that requirement.
There's always unspecified behavior if documentation is considered to be too onerous.
It also means that the behavior needs to be stable across time and that tools cannot report this behavior as a bug.
> It also means that the behavior needs to be stable across time
I don't think IB has to be stable over time? For instance, whether and which extended integer types exist is IB and that has definitely changed over time.
> and that tools cannot report this behavior as a bug.
It's absolutely possible to report IB as a bug (or even fully-specified behavior, as demonstrated by -fsanitize=unsigned-integer-overflow or -fsanitize=implicit-integer-conversion).
A good example of C++ bizarre design decisions. The exact handing of float<->int conversions should use implementation-defined behavior instead, defining some cases to be UB just adds more traps with little to no performance benefits.
The same is true for bit-shifts greater than type bit size and negative bit-shifts.
> bizarre design decisions
They are not bizarre but a good compromise between consistency and optimization for any possible existing and future CPU architecture.
What the author correctly points out to is that GSL implementation does not remove that undefined behaviour as it lacks the proper checks. And that is the problem as it does not fulfill its own documented guarantees.
tldr; C++ is just fine. The implementation in the library is wrong.
> C++ is just fine
No it isn't. Using undefined behavior for things which can be implementation-defined behavior instead is harmful. It creates more space for bugs and security vulnerabilities.
C++ was designed as an universal way of writing low level system code. It has some abstractions above C, but still allowing for all kinds of optimizations. (Think using a class vs a template, similar abstraction different impact).
To use C++ to code user programs is the actual mistake.
Think about assembly. Should assembly also be safe in all calls? That will make everything slower, and consume way more energy. Will you blame assembly if people used it to create a phone app and it had memory leaks? Or will you blame the developer for using assembly in the first place?
C++ is a tool designed for a range of purposes. There is a push to make it an universal language that makes everything well. But engineering is about making compromises.
For your use, C++ may be the wrong language. I do not disagree with you. But maybe you should be using another language instead. Or we should create a new language that fills the sweet spot that C++ covers, and then evolve C++ to be the safe language that you want. One way or another, it is a compromise.
I thought C++ was designed to fit all sorts of applications. There is many GUI libraries for C++ for example, and the language has all the capabilities to add abstractions for making these sorts of libraries easy to work with. I think you could argue that C was made for low level code, since the working group seems to be focussing on making the core of the language more usable instead of allowing for abstractions that make libraries easier to work with (I think we can all agree that the C interface of gtk is not that great to work with, though I do commend their efforts).
Sorry for rambling so much, I hope I got my point across.
C was made as more portable assembly, and C++ was made to add lots of features to C.
Apparently you thought wrong and should stop using C++.
What is an appropriate behavior for casting inf or nan to an integer?
No? That would mean implementation defined, not undefined.
They are bizarre, because C developer culture isn't known for doing what is necessary to prove the absence of UB and memory safety problems.
It's all up to the programmer who obviously won't invest the time because that would ruin the development velocity.
If UB was limited to dereferencing freed memory and so on, it would be better.