This isn't quite the same case, but it's a good illustration of the effect: on gcc, if you have an expression f(a(), b()), the order that a and b get evaluated is [1] dependent on the architecture and calling-convention of f. If the calling convention wants you to push arguments from right to left, then b is evaluated first; otherwise, a is evaluated first. If you evaluate arguments in the right order, then after calling the function, you can immediately push the argument on the stack; in the wrong order, the result is now a live variable that needs to be carried over another function call, which is a couple more instructions. I don't have a specific example for increment/decrement instructions, but considering extremely register-poor machines and hardware instruction support for increment/decrement addressing modes, it's not hard to imagine that there are similar cases where forcing the compiler to insert the increment at the 'wrong' point is similarly expensive.
Now, with modern compilers using cross-architecture IRs as their main avenue of optimization, the benefit from this kind of flexibility is very limited, especially since the penalties on modern architectures for the 'wrong' order of things can be reduced to nothing with a bit more cleverness. But compiler developers tend to be loath to change observable behavior, and the standards committee unwilling to mandate that compiler developers have to modify their code, so the fact that some compilers have chosen to implement it in different manners means it's going to remain that way essentially forever. If you were making a new language from scratch, you could easily mandate a particular order of evaluation, and I imagine that every new language in the past several decades has in fact done that.
[1] Or at least was 20 years ago, when I was asked to look into this. GCC may have changed since then.
With modern register allocators and larger register sets, code generation impact from following source evaluation is of course lower than it used to be. Some CPUs can even involve stack slots in register renaming: https://www.agner.org/forum/viewtopic.php?t=41
On the other hand, even modern Scheme leaves evaluation order undefined. It's not just a C issue.
Anyway, yes, this one example has an obvious order it should be applied. But still, something like it shouldn't be allowed.
That would be nice, but don't forget the more general case of pointers and aliasing:
int a = 5;
int *pa = &a;
printf("%d", (a++ + ++*pa));
The compiler cannot statically catch every possible instance of a statement where a variable is updated more than once.Look at the addressing modes for the PDP-11 in https://en.wikipedia.org/wiki/PDP-11_architecture and you'll see you can write (R0)+ to read the contents of the location pointed to by R0, and then increment R0 afterwards (so a post increment).
Back in the day, compilers were simple and optimisations weren't that common, so folding two statements into one and working out that there were no dependencies would have been tough with single pass compilers.
You could argue that without such instructions, C wouldn't have been embraced quite so enthusiastically for systems programming, and the world would have looked rather different.
C just wouldn't be C without things like a[i++]
Both are favorite idioms of C developers. And they are ok if done correctly, clearer than the alternative. They are also unnecessary in modern languages, so those shouldn't copy it (yeah, Python specifically).
The reason that these operators pull their weight in C is because iteration over arrays is achieved by manual incrementation (usually via the leading clauses of the for-loop) followed by direct indexing. Languages with a first-class notion of iteration don't directly index in this way, which overwhelmingly eliminates not only the vast majority of array indexing operations in codebases but also the need to manually futz with the inductive loop variable. Case in point, Rust doesn't have `++` in any form, and it doesn't miss it, because Rust has first-class iteration; on the then relatively rare occasion where you want do want to increment, you can do `+=1`, which doesn't have the footguns of `++` due to assignment being a statement rather than expression, while leading to a simpler language due to leveraging the existing `+=` syntax rather than needing a whole new set of operators.
> which doesn't have the footguns of `++` due to assignment being a statement rather than expression,
So then I implement the local equivalent of inc( v ) and ... same issue, right? Plus with rust macros is there any technical reason you can't trivially implement ++ for yourself? That's the case for most lisps that I touched on earlier.
I'd say that when you're writing a mildly complex loop that involves pointer juggling, one should prefer to be defensive and explicit rather than cleverly trying to compress everything into one-liners.
> So then I implement the local equivalent of inc( v ) and ... same issue, right?
This isn't done in Rust because there's no benefit. It's rare to find an occasion where it's necessary to do something tricky enough to forego using iterators, and when working with raw pointers Rust code just plain doesn't use basic addition for pointer arithmetic; instead it has a variety of pointer arithmetic methods for being explicit about the desired semantics (e.g. ptr::add, ptr::offset, ptr::wrapping_add, etc).
> Plus with rust macros is there any technical reason you can't trivially implement ++ for yourself?
There's not, but people might look at you sideways. Here, I implemented it for you: https://play.rust-lang.org/?version=stable&mode=debug&editio... . It expands to nested blocks with internal assignments, which results in a well-defined semantics following the defined order of evaluation in Rust.
People tend to use FP abstractions for the "x[i++] = f(y[j++])" though, not iteration.
I think if anything people have been leaning more and more into expressions over statements, because when everything is an expression you end up being able to walk the gradient of complexity a bit more nicely than when you end up with a thing that just has to be broken down to a bunch of statements.
And when I checked 3 different compilers, each of them chose a different way to use FMAs.
Even with integer math, you can get different numerical results via UB (e.g. expressions with signed overflow one way and not another).
Just switching between left to right or right to left wouldn't be that useful but it also permits to interleave the subexpression evaluation. Grouping memory fetches/writes, taking into account how many execution units and registers of different kinds a CPU has can have some performance benefits.
For example if you have something like `++a[0] + ++a[1] + ++a[2] + ++a[3]` instead of evaluating each increment one by one both GCC and Clang will vectorize it loading all 4 values from memory using single simd instruction, incrementing and then writing result back to memory. And if you add fifth one (but not 8) which needs to be handled using regular instruction, that will be done after the first 4. If standard defined that left subexpression of addition is fully evaluated before the right expression that wouldn't be allowed.
I'm no expert, but surely this would still be allowed so long as the compiler can prove that incrementing a[0] has no effect on the value of a[1]?
I didn't open TFA but my first thought was "Is this even defined?".
It kinda make sense that suck fucktardedness could be not defined.
At least according to this: https://en.wikipedia.org/wiki/Operators_in_C_and_C%2B%2B#Exp...
I think the main confusion here comes from the fact that "a" is just a value, not a pointer, where it matters when the value/address which the pointer points at is accessed (before of after the increment of the pointer's own 'value').
Anyway… my C skills are rusty. Maybe I get it wrong. :) In any case I always would use brackets to avoid any ambiguity in constructs like this.
Actually the compiler (at least clang) warns about this:
$ gcc -W -Wall test.c -o test
test.c:8:7: warning: multiple unsequenced modifications to 'a' [-Wunsequenced]
a = a++ + ++a;
^ ~~
1 warning generated.
The undefined behaviour stems from the fact that "a" is modified multiple times between the "sequence points" (so it's irrelevant to the actual problem that this happens with ++, --, pre-, or, post-, or in which order) We only can modify the variable safely once on the right side without entering bizarro world.A construct like this certainly can be confusing.
The OP article provides experimental details but annoyingly does not give the big picture w.r.t. C language specifications (provided in the links though).
There are three concepts at interplay here which is at the root of the problem; 1) Expressions (evaluates to a single value) 2) Statements (tells computer to perform an action) and 3) Sequence Points (specific moment during execution when all previous side-effects are guaranteed to be complete).
It is the sequence points during the evaluation of expressions which is important to understand here. From https://en.wikipedia.org/wiki/Sequence_point;
In C and C++, a sequence point defines any point in a computer program’s execution at which it is guaranteed that all side effects of previous evaluations will have been performed, and no side effects from subsequent evaluations have yet been performed. They are a core concept for determining the validity of and, if valid, the possible results of expressions...
1) An expression's evaluation can be "sequenced before" the evaluation of another expression. (Equivalently, the other expression's evaluation can be "sequenced after" that of the first.)
2) The expression's evaluation is "indeterminately sequenced", meaning that one is "sequenced before" the other, but which is unspecified.
3) The expression's evaluation is "unsequenced", meaning the operations in each expression may be interleaved.
The "Order of Evaluation" states; (from https://en.cppreference.com/c/language/eval_order)
"Order of evaluation of the operands of any C operator, including the order of evaluation of function arguments in a function-call expression, and the order of evaluation of the subexpressions within any expression is unspecified (except where noted below)."
The "Single Update Rule" states; (from https://www.accellera.org/images/eda/sv-bc/0282.html)
Between consecutive "sequence points" an object's value can be modified only once by an expression. The C language defines the following sequence points:
Left operand of the logical-AND operator (&&).
Left operand of the logical-OR operator (||).
Left operand of the comma operator.
Function-call operator.
First operand of the conditional operator.
The end of a full initialization expression.
The expression in an expression statement.
The controlling expression in a selection (if or switch) statement.
The controlling expression of a while or do statement.
Each of the three expressions of a for statement.
The expression in a return statement.
Putting all of the above together in OP's code snippet; The "single update rule" fails for the expression since the variable a is modified multiple times between two consecutive sequence points and hence the result is UB.For more detailed explanations, see Angelika Langer's Sequence Points and Expression Evaluation in C++ - https://angelikalanger.com/Articles/VSJ/SequencePoints/Seque...
The horrible undefined behavior of signed integer overflow at least can be explained by the fact that multiple CPU architectures handling those differently existed (though the fact that C even 'attracts' its ill-defined signed integers when you're using unsigned ones by returning a signed int when left shifting an uint16_t by an uint16_t for example is not as forgivable imho)
But this here is something that could be completely defined at the language level, there's nothing CPU dependent here, they could have simply stated in the language specification that e.g. the order of execution of statements is from left to right (and/or other rules like post increment happens after the full statement is finished for example, my point is not whether the rule I type here is complete enough or not but that the language designers could have made it completely defined).