The general wisdom about eval() is: do not use it! At least not until you are really out of reasonable options after consulting more than 3 experts on the newsgroups, forums and support@mathworks.com (if your SMS is current)!
Abusing eval() turns it into evil()!
The elves running inside MATLAB needs to be able to track your variables to reason through your code because:
it helps your code run much faster (eval() cannot be pre-compiled)
able to use parallel computing toolbox (it has to know absolutely for sure about any shared writes)
mlint can warn you about potentially pitfalls through code smell.
it keeps you sane while debugging!
This is called ‘transparency’: MATLAB has to see what you are doing every step of the way. According to MATLAB’s parallel computing toolbox’s documentation,
Transparency means that all reference to variables must be visible in the text of the code
which I used as a subtitle of this post.
The 3 major built-in functions that breaks transparency are:
eval(), evalc(): arbitrary code execution resulting in read and write access to its caller workspace.
assignin(): it poofs variables in its caller’s workspace!
evalin(): it breaks open the stack and read the variables in its caller’s workspace!
They should have been replaced by skillful use of dynamic field names, advanced uses of left assignment techniques, and freely passing variables as input arguments (remember MATLAB uses copy-on-write: nothing is copied if you just read it).
There are other frequently used native MATLAB functions (under certain usages) that breaks transparency:
load(): poof variables from the file to the workspace. The best practice is to load the file as a struct, like S=load('file.mat'); , which is fully transparent. Organizing variables into structs actually reduces mental clutter (namespace pollution)!
save(), who(), whos(): basically anything that takes variable names as input and act on the requested variable violates transparency because it’s has the effect of evalin(). I guess the save() function chose to use variable names instead of the actual variables as input because copy-on-write wasn’t available in early days of MATLAB. A example workaround would be:
function save_transparent(filename, varargin)
VN = arrayfun(@inputname, (2:nargin)', 'UniformOutput', false);
if( any( cellfun(@isempty, VN) ) )
error('All variables to save must be named. Cannot be temporaries');
end
S = cell2struct(varargin(:), VN, 1);
save(filename, '-struct', 'S');
end
function save_struct_transparent(filename, S)
save(filename, '-struct', 'S');
end
The good practices to avoid non-transparent load/save also reduces namespace pollution. For example, inputname() peeks into the caller to see what the variable names are, which should not be used lightly. The example above is one of the few uses that I consider justified. I’ve seen novice abusing inputname() because they were not comfortable with cells and structs yet, making it a total mindfuck to reason through.
Unless deleting a known range of elements directly through iterators (no conditions to match), which range–erase() method can be used directly, targeting specific key/value/predicate requires understanding of the container’s underlying data structure.
I’d like to give a summary of Item#9 in “Effective STL” by defining the names and concepts so the complicated rules can be sensibly deduced by a few basic facts.
The words ‘remove‘ and ‘erase‘ has very specific meaning for STL that are not immediately intuitive.
Lives in
Target to match
Purpose
remove_?()
<algorithm>
required
Rearrange wanted elements to front
erase()
container
not accepted
Blindly deleting range/position given
There is a remove() method for lists, which is an old STL naming inconsistency (they should have called it erase() like for associative containers). Treat it as a historical mistake.
The usage is easy to remember once you understand it with the right wording above:
Lists provides a efficient shortcut method (see table below) since linked-lists does not need to be rearranged (just short the pointers).
one-shot methods
contiguous
lists
associative
by content
N/A: use erase-remove idiom
remove(T)
erase(key)
by predicate
N/A: use erase-remove_if idiom
remove_if(pred)
N/A: Use for-loops for now No erase_if() yet as of C++17.
Never try range-based remove_?() for associative containers. It is a data corruption trap if any attempt is made to use anything named remove on associative containers.
The trap used to be possible since <algorithms> and containers were separate, but newer C++ protects you from the trap by checking if the element you are moving is of a MoveAssignable type. Since associative containers’ keys cannot be modified (without a rescan), the elements are not move-assignable.
As for erasing through for-loops (necessary if you want to sneak in an extra step while iterating), C++11 now returns an iterator following the last erased element uniformly across all containers. This helps to preserve the running iterator that gets invalidated immediately after the erase through i=c.erase(i);
* For brevity, I twisted the term unordered here to mean that the native (implementation) data order is dependent on the data involved.
When I said ‘cannot rearrange’, I meant ‘cannot efficiently rearrange’, since there are no cheap O(1) next() or prev() traversal.
It’s a mess to simply copy one element over another (during rearrangement), leaving orphans there, and re-balance a BST or re-hash a hash-map. Nobody wants to go through such pains to remove element when there are tons of more direct ways out there.
a pointer only contains a memory location,
while an array already has memory allocated to hold the data.
The confusion comes from the fact that array names are always seen as pointers anywhere in C, but when an array name is referred in places that the scope happens know the allocated size, namely
Global arrays: everybody knows the size
Local arrays: only the instantiating function knows its size.
, the array name itself has a superpower that pointers lack: report the underlying allocated data size (NOT pointer size) using sizeof(array).
Definition: An array is ‘recognized‘ if the array name is used in the scope that knows the underlying data size.
Corollary: Calling the array name with sizeof() gives the underlying allocated data size.
Examples of consequences that can be derived from the definition above:
Heap allocations always return a pointer type, NOT an array name!
So heap arrays are never recognized.
VLA in C99 are considered local stack arrays, so it’s recognized
x[] is just a cosmetic shorthand for *x: it doesn’t prevent any recognized array from decaying into a pointer across boundary.
The storage duration (static or not) does not matter. e.g.
Heap pointers at global level are not recognized arrays
Static local array still loses the recognition across function boundaries
(unless passed carefully by data type T (&array)[N]).
Most often recognized arrays cannot be aliased without decaying into a pointer. However, we can bind a recognized array to a reference to an array, which is a completely different type. Example:
int v[]{1,2,3,4};
int (&w)[4]=v; // w is a reference to an array of size 4
int* p = v; // Decays v to a pointer. Size information lost.
// int &w[4]=v; // Does not compile: this means an array of 4 references.
Note that the syntax requires a bracket for reference name. Omitting it will lead the compiler to misinterpret it as an array of references, which cannot* be compiled.
This means contrary to common beliefs, you can pass a recognized array across functions through reference, but this is rarely done because of the hassle of explicitly entering the number of elements (4 for the example above) as part of the data type. This can still be done through templates/constexpr, but for such inconvenience, we’re better off using std::vector (or std::array if you want near zero overhead).
However, so far I haven’t found a way to re-recognize an array from a pointer. That means there is no way to keep a local array’s recognition across function boundaries in C since it does not have references like C++.
To summarize with a usage example: this post has described the entire logic needed to decide whether sizeof(x)/sizeof(x[0]) gives you the number of array elements, or how many times your machine pointer type is bigger than the element storage.
* references must be bound on creation. Declaring an array of references means you want to bound references in batches. There are no mechanisms to do so as of C++14.
I was a little confused the first time I saw range-based for-loop in C++11 on “Tour of C++” that it works right out of the box for both recognized array and STL containers and yet the text says it requires begin() and end() to be defined for the object for range-based for-loop to run.
I later learned that despite typical STL usage example writes v.begin(), v.end(), the most bulletproof way is to write begin(v), end(v) instead (Herb Sutter recommends it). Then I started to suspect that C++11 must have defined free-form (non-member) begin(), end() functions that takes in arbitrary recognized arrays. I pulled up my code editor and ran this:
#include <iostream>
int main()
{
int v[4]={1,2,3,4};
std::cout << *(std::crend(v)-1) << std::endl;
return 0;
}
It compiled and ran uneventfully, printing ‘1’ as expected (I’m using crend(), to see if they implemented the more obscure ones). It makes more sense now why range-based for-loop works for arbitrary recognized arrays without making an exception to the begin(), end() requirement.
To confirm that it is the case (since “Tour of C++” didn’t say anything about why arbitrary array works for range-based for loop), I looked up the STL source code from libc++ in LLVM, namely <iterator>, and saw this:
Bingo! There’s a mechanism to do so. But before I close, Stephan T. Lavavej (Mr. STL) mentioned that the template quoted above is no longer required (by the standard) to implement range-based for-loop in C++11.
Now the conclusion becomes that begin(), end() that takes in recognized arrays exist (which completes the logic behind range-based for-loop), but the range-based for loop can (and typically will) handle recognized arrays without these templated functions defined.
Compilers has gotten smarter and smarter nowadays that they’d be able to analyze our code for common patterns (or logically deduce away steps that doesn’t have to be performed at runtime).
Matt Godbolt gave a nice presentation at CppCon 2017 named “What Has My Compiler Done for Me Lately?”. Through observing the emitted assembly code at different optimization levels, he showed that the compiler doesn’t need to be micromanaged (through performance hacks in our code) anymore, as it will emit instructions as the performance-hacked code intended when it is better to do so.
It means the compiler writers already know our bag of performance hack tricks way better than we do. Their efforts spared us from premature optimization and leave us more time to find a better data structure or algorithm to solve the problem.
What I got from the lecture is NOT that we are free to write clumsy code and let the compiler sort it out (though it occasionally can, like factoring a loop doing simple arithmetic series into a one line closed form solution), but we don’t have to make difficult coding choices to accommodate performance anymore.
The most striking facts I learned from his lecture are
The compiler can emit a one-line CPU instruction that does not have a corresponding native operation C/C++ if your hardware architecture supports it. (e.g. clang can convert a whole loop that counts the number of set bits into just ‘popcnt eax, edi‘)
Through Link-Time Optimization (LTO), we don’t have to pay the performance penalty for language features that are ultimately necessary for the current compilation (e.g. virtuals are automatically dropped if the linker finds that nowhere in the output currently needs it)
With such LTO, why not do away the virtual specifier and make everything unspecified virtual by default anyway (like Java)? For decades, we’ve been making up stories that some classes are not meant to be derived (like STL containers), but the underlying motive is that we don’t want to pay for vtable if we don’t have to.
Instead of confusing new programmers about when should they make a method virtual (plenty of rule-of-thumbs became dogma), focus on telling them whenever they (choose to upcast a reference/pointer to the parent anywhere in their code and) invoke the destructor through the parent reference/pointer, they will pay a ‘hefty’ price of vtable and vptr.
I don’t think anybody (old codebase) will get harmed by turning on virtuals by default and let the linker decide if those virtuals can be dropped. If it changes anything, it might turn buggy code with the wrong destructor called into correct code which runs slower and takes up more space. In terms of correctness, this change might break low-level hacks that expects the objects to be of certain size (e.g. alignment) without vptr.
Even better, add a class specifier that mandates that all uses of its child must not invoke vtable (have the compiler catch that) unless explicitly overridden (the users decide to pay for the vtable). This way the compiler can warn about performance and space issues for the migration.
The old C++’s ideal was “you only pay for the language features you used (written)”, but as compilers gets better, we might be able change it to “you pay extra only for the language features that are actually used (in the finally generated executable) with your permission”.
I’d also like to add Return Value Optimization (RVO) into my list of compiler advances that changes the way we code. C++11 added move semantics, but I think it’s something that the compiler in the future could be able to manage themselves. Even with an old C++ compiler like the one shipped with VisualDSP 5.0, the copy constructor was not called (yes, skipping it is legal even if the copy constructor has side effects) when I do this:
Matrix operator+(const Matrix& a, const Matrix& b)
{
Matrix c(a.dim);
// ... for all element i, c.raw[i] = a.raw[i]+b.raw[i]
return c;
}
Matrix c = a + b;
Actually, the compiler at that time was not that smart about RVO, the actual code I wrote originally had two return branches, which defeats RVO (it’s a defined behavior by the specs):
Matrix operator+(Matrix a, Matrix b)
{
Dims m = a.dims;
if( m == b.dims ) // Both inputs must have same dimensions
{
Matrix c(m); // Construct matrix c with same dimension as a
// ... for all i, c.raw[i] = a.raw[i] + b.raw[i]
return c;
}
else
{
return Matrix::dummy; // A static member, which is a Matrix object
}
}
To take advantage of RVO, I had to reword my code
Matrix operator+(Matrix a, Matrix b)
{
Dims m = a.dims;
if( m == b.dims ) // Both inputs must have same dimensions
{
Matrix c(m); // Construct matrix c with same dimension as a
// ... for all i, c.raw[i] = a.raw[i] + b.raw[i]
}
else
{
Matrix c = Matrix::dummy; // or just "Matrix c";
}
return c;
}
I think days are counting before C++ compilers can do “copy-on-write” like MATLAB does if independent compilation are no longer mandatory!
Given my extensive experience with MATLAB, I’d say it took me a while to get used designing my code with “copy-on write” behavior in mind. Always start with expressive, maintainable, readable and correct code keeping in mind the performance concerns only happens under certain conditions (i.e. passed object gets modified inside the function).
If people start embracing the mentality of letting the compiler do most of the mechanical optimization, we’ll move towards a world that debugging work are gradually displaced by performance-bottleneck hunting. In my view, anything that can be done systematically by programming (like a boilerplate code or idioms) can eventually be automated by better compiler/linker/IDE and language design. It’s the high-level business logic that needs a lot of software designers/engineers to translate fuzzy requirements into concrete steps.
Matt also developed a great website (http://godbolt.org/) that compiles your code repeatedly on the fly and shows you the corresponding assembly code. Here’s an example of how I use it to answer my question of “Should I bother to use std::div() if I want both the quotient and remainder without running the division twice?”:
The website also included a feature to share the pasted code through an URL.
As seen from the emitted assembly code, the answer is NO. The compiler can figure out that I’m repeating the division twice and do only one division and use the quotient (stored in eax) and remainder (stored in edx). Trying to enforce one division through std::div() requires an extra function call, which is strictly worse.
The bottom line: don’t help the compiler! Modern compiler does context free optimizations better than we do. Use the time and energy to rethink about the architecture and data structure instead!
