Which exact imperative languages are you taking as benchmarks of performance here? The lower-level ones like C or Fortran where you can squeeze out amazing performance if the compiler is any good, or the fancier ones with garbage collection, dynamic typing, bounds checking, and other features with large overheads?
To give one example pair, Java (imperative) and Scala (functional) share a significant amount of compiler infrastructure and end up being pretty much the same speed.
In practice, the answer to how fast languages are is pretty complex, and sometimes counterintuitive. For example, one of the “large overheads” you mentioned, bounds checking, isn’t a significant overhead at all. Modern compilers are very aggressive about removing bounds checks when it can be proven that they’ll never fail (which is true for almost all good code), and moving them out of loops. C programs, for security reasons, are generally run past a theorem prover (splint) to ensure that they have bounds checks where they’re needed, which means that C programs and Scala programs end up with the same set of bounds checks in the compiled output, with differences due to different features of the respective theorem provers/optimizers, which operate on intermediate representations that look nothing like the original languages anyways. Similarly, garbage collection has a reputation as an expensive feature, because of bad garbage collectors, but a good compiler can prove when most objects will leave scope, and take them out of the garbage collector’s domain; it doesn’t have to be expensive. Again for dynamic typing; most dynamically typed programs can have type inference performed on them to make statically-typed programs, it’s just that the compilers aren’t always good enough to do that. (But sometimes they are. And there are plenty of dynamically typed imperative languages around, so it’s not a functional/imperative thing.)
Fortran’s reputation for speed, as far as I can tell, is mainly due to the LINPACK linear algebra library, which started in the Fortran world but is actually written in assembly to use funky vector instruction sets these days anyways. C is great for interfacing with hardware and writing schedulers and memory allocators, because it maps so closely to assembly language; and more work has been put into C compilers than compilers for any other language. But on the sort of code that compilers are good at, it’s the same speed as OCaml or Scala, because they all end up compiling to the same thing.
(My knowledge of this subject is mainly academic; I wrote a compiler in school, and I try to keep tabs on research in the field, but I’m not an expert by any means. Take this for what it’s worth.)
You make the situation with optimizing compilers sound really optimistic! Unfortunately, it seems to me that things don’t work anywhere so well in practice. Yes, the practical handling of fancy language features has come a long way from naive straightforward implementations, but I’d say you exaggerate how good it is.
For example, I haven’t followed closely the work in bounds checking elimination, but I know that a lot of papers are still being published about it, indicating that there are still large enough overheads to make the problem interesting. (Which is not surprising, considering that the problem is after all undecidable in general. Also, as far as I know, bounds checks are normally not added by C compilers, and there are depressing limits to what theorem provers can figure out about the usual C where you pass pointers around liberally.)
It’s similar with garbage collection, dynamic type checks, and other fancy features. Their overheads can certainly be reduced greatly by smart compilers and efficient run-time operations, sometimes to the point where there is no difference from C, but definitely not always, and often not reliably and predictably.
(Fortran, by the way, has traditionally had the advantage of being highly amenable to automatic optimization, including automatic parallelization, especially when used for typical array-based numerical computations. This has in turn led to a lot of fruitful effort put into optimizing and parallelizing numerical Fortran, leading to its unprecedented performance and acceptance in these areas, and its lead has been eroded only in recent years. You can’t possibly say that this is just due to a single popular library.)
To give one example pair, Java (imperative) and Scala (functional) share a significant amount of compiler infrastructure and end up being pretty much the same speed.
In practice, the answer to how fast languages are is pretty complex, and sometimes counterintuitive. For example, one of the “large overheads” you mentioned, bounds checking, isn’t a significant overhead at all. Modern compilers are very aggressive about removing bounds checks when it can be proven that they’ll never fail (which is true for almost all good code), and moving them out of loops. C programs, for security reasons, are generally run past a theorem prover (splint) to ensure that they have bounds checks where they’re needed, which means that C programs and Scala programs end up with the same set of bounds checks in the compiled output, with differences due to different features of the respective theorem provers/optimizers, which operate on intermediate representations that look nothing like the original languages anyways. Similarly, garbage collection has a reputation as an expensive feature, because of bad garbage collectors, but a good compiler can prove when most objects will leave scope, and take them out of the garbage collector’s domain; it doesn’t have to be expensive. Again for dynamic typing; most dynamically typed programs can have type inference performed on them to make statically-typed programs, it’s just that the compilers aren’t always good enough to do that. (But sometimes they are. And there are plenty of dynamically typed imperative languages around, so it’s not a functional/imperative thing.)
Fortran’s reputation for speed, as far as I can tell, is mainly due to the LINPACK linear algebra library, which started in the Fortran world but is actually written in assembly to use funky vector instruction sets these days anyways. C is great for interfacing with hardware and writing schedulers and memory allocators, because it maps so closely to assembly language; and more work has been put into C compilers than compilers for any other language. But on the sort of code that compilers are good at, it’s the same speed as OCaml or Scala, because they all end up compiling to the same thing.
(My knowledge of this subject is mainly academic; I wrote a compiler in school, and I try to keep tabs on research in the field, but I’m not an expert by any means. Take this for what it’s worth.)
You make the situation with optimizing compilers sound really optimistic! Unfortunately, it seems to me that things don’t work anywhere so well in practice. Yes, the practical handling of fancy language features has come a long way from naive straightforward implementations, but I’d say you exaggerate how good it is.
For example, I haven’t followed closely the work in bounds checking elimination, but I know that a lot of papers are still being published about it, indicating that there are still large enough overheads to make the problem interesting. (Which is not surprising, considering that the problem is after all undecidable in general. Also, as far as I know, bounds checks are normally not added by C compilers, and there are depressing limits to what theorem provers can figure out about the usual C where you pass pointers around liberally.)
It’s similar with garbage collection, dynamic type checks, and other fancy features. Their overheads can certainly be reduced greatly by smart compilers and efficient run-time operations, sometimes to the point where there is no difference from C, but definitely not always, and often not reliably and predictably.
(Fortran, by the way, has traditionally had the advantage of being highly amenable to automatic optimization, including automatic parallelization, especially when used for typical array-based numerical computations. This has in turn led to a lot of fruitful effort put into optimizing and parallelizing numerical Fortran, leading to its unprecedented performance and acceptance in these areas, and its lead has been eroded only in recent years. You can’t possibly say that this is just due to a single popular library.)