Realistic usage of the C99 'restrict' keyword?

Question

I was browsing through some documentation and questions/answers and saw it mentioned. I read a brief description, stating that it would be basically a promise from the programmer that the pointer won't be used to point somewhere else.

Can anyone offer some realistic cases where its worth actually using this?

Solution 1

restrict says that the pointer is the only thing that accesses the underlying object. It eliminates the potential for pointer aliasing, enabling better optimization by the compiler.

For instance, suppose I have a machine with specialized instructions that can multiply vectors of numbers in memory, and I have the following code:

void MultiplyArrays(int* dest, int* src1, int* src2, int n)
{
    for(int i = 0; i < n; i++)
    {
        dest[i] = src1[i]*src2[i];
    }
}

The compiler needs to properly handle if dest, src1, and src2 overlap, meaning it must do one multiplication at a time, from start to the end. By having restrict, the compiler is free to optimize this code by using the vector instructions.

Wikipedia has an entry on restrict, with another example, here.

Solution 2

The Wikipedia example is very illuminating.

It clearly shows how it allows to save one assembly instruction.

Without restrict:

void f(int *a, int *b, int *x) {
  *a += *x;
  *b += *x;
}

Pseudo assembly:

load R1 ← *x    ; Load the value of x pointer
load R2 ← *a    ; Load the value of a pointer
add R2 += R1    ; Perform Addition
set R2 → *a     ; Update the value of a pointer
; Similarly for b, note that x is loaded twice,
; because x may point to a (a aliased by x) thus 
; the value of x will change when the value of a
; changes.
load R1 ← *x
load R2 ← *b
add R2 += R1
set R2 → *b

With restrict:

void fr(int *restrict a, int *restrict b, int *restrict x);

Pseudo assembly:

load R1 ← *x
load R2 ← *a
add R2 += R1
set R2 → *a
; Note that x is not reloaded,
; because the compiler knows it is unchanged
; "load R1 ← *x" is no longer needed.
load R2 ← *b
add R2 += R1
set R2 → *b

Does GCC really do it?

GCC 4.8 Linux x86-64:

gcc -g -std=c99 -O0 -c main.c
objdump -S main.o

With -O0, they are the same.

With -O3:

void f(int *a, int *b, int *x) {
    *a += *x;
   0:   8b 02                   mov    (%rdx),%eax
   2:   01 07                   add    %eax,(%rdi)
    *b += *x;
   4:   8b 02                   mov    (%rdx),%eax
   6:   01 06                   add    %eax,(%rsi)  

void fr(int *restrict a, int *restrict b, int *restrict x) {
    *a += *x;
  10:   8b 02                   mov    (%rdx),%eax
  12:   01 07                   add    %eax,(%rdi)
    *b += *x;
  14:   01 06                   add    %eax,(%rsi) 

For the uninitiated, the calling convention is:

• rdi = first parameter
• rsi = second parameter
• rdx = third parameter

GCC output was even clearer than the wiki article: 4 instructions vs 3 instructions.

Arrays

So far we have single instruction savings, but if pointer represent arrays to be looped over, a common use case, then a bunch of instructions could be saved, as mentioned by supercat.

Consider for example:

void f(char *restrict p1, char *restrict p2) {
    for (int i = 0; i < 50; i++) {
        p1[i] = 4;
        p2[i] = 9;
    }
}

Because of restrict, a smart compiler (or human), could optimize that to:

memset(p1, 4, 50);
memset(p2, 9, 50);

which is potentially much more efficient as it may be assembly optimized on a decent libc implementation (like glibc): Is it better to use std::memcpy() or std::copy() in terms to performance?

Does GCC really do it?

GCC 5.2.1.Linux x86-64 Ubuntu 15.10:

gcc -g -std=c99 -O0 -c main.c
objdump -dr main.o

With -O0, both are the same.

With -O3:

• with restrict:

  3f0:   48 85 d2                test   %rdx,%rdx
  3f3:   74 33                   je     428 <fr+0x38>
  3f5:   55                      push   %rbp
  3f6:   53                      push   %rbx
  3f7:   48 89 f5                mov    %rsi,%rbp
  3fa:   be 04 00 00 00          mov    $0x4,%esi
  3ff:   48 89 d3                mov    %rdx,%rbx
  402:   48 83 ec 08             sub    $0x8,%rsp
  406:   e8 00 00 00 00          callq  40b <fr+0x1b>
                          407: R_X86_64_PC32      memset-0x4
  40b:   48 83 c4 08             add    $0x8,%rsp
  40f:   48 89 da                mov    %rbx,%rdx
  412:   48 89 ef                mov    %rbp,%rdi
  415:   5b                      pop    %rbx
  416:   5d                      pop    %rbp
  417:   be 09 00 00 00          mov    $0x9,%esi
  41c:   e9 00 00 00 00          jmpq   421 <fr+0x31>
                          41d: R_X86_64_PC32      memset-0x4
  421:   0f 1f 80 00 00 00 00    nopl   0x0(%rax)
  428:   f3 c3                   repz retq
  

  Two memset calls as expected.

• without restrict: no stdlib calls, just a 16 iteration wide loop unrolling which I do not intend to reproduce here :-)

I haven't had the patience to benchmark them, but I believe that the restrict version will be faster.

C99

Let's look at the standard for completeness sake.

restrict says that two pointers cannot point to overlapping memory regions. The most common usage is for function arguments.

This restricts how the function can be called, but allows for more compile-time optimizations.

If the caller does not follow the restrict contract, undefined behavior.

The C99 N1256 draft 6.7.3/7 "Type qualifiers" says:

The intended use of the restrict qualifier (like the register storage class) is to promote optimization, and deleting all instances of the qualifier from all preprocessing translation units composing a conforming program does not change its meaning (i.e., observable behavior).

and 6.7.3.1 "Formal definition of restrict" gives the gory details.

Strict aliasing rule

The restrict keyword only affects pointers of compatible types (e.g. two int*) because the strict aliasing rules says that aliasing incompatible types is undefined behavior by default, and so compilers can assume it does not happen and optimize away.

See: What is the strict aliasing rule?

See also

• C++14 does not yet have an analogue for restrict, but GCC has __restrict__ as an extension: What does the restrict keyword mean in C++?
• Many questions that ask: according to the gory details, does this code UB or not?
  • Understanding restrict qualifier by examples
  • Restricted pointer questions
  • Is it legal to assign a restricted pointer to another pointer, and use the second pointer to modify the value?
• A "when to use" question: When to use restrict and when not to
• The related GCC __attribute__((malloc)), which says that the return value of a function is not aliased to anything: GCC: __attribute__((malloc))

Comments

[Alexandre C.]

memcpy vs memmove is one canonical example.

[supercat]

@AlexandreC.: I don't think it's a particularly applicable one, since lack of a "restrict" qualifier does not imply that the program logic will work with overloading source and destination, nor would presence of such a qualifier prevent a called method from determining whether source and destination overlap and, if so, replacing dest with src+(dest-src) which, since it is derived from src, would be allowed to alias it.

[Alexandre C.]

@supercat: That's why I put it as a comment. However, 1) restrict-qualifying arguments to memcpy enables in principle a naïve implementation to be optimized aggressively, and 2) merely calling memcpy enables the compiler to assume that the arguments given to it do not alias, which could allow some optimization around the memcpy call.

[supercat]

@AlexandreC.: It would be very hard for a compiler on most platforms to optimize a naive memcpy--even with "restrict"--to be anywhere near as efficient as version tailored to the target. Call-side optimizations would not require the "restrict" keyword, and in some cases efforts to facilitate those may be counter-productive. For example, many implementations of memcpy could, at zero extra cost, regard memcpy(anything, anything, 0); as a no-op, and ensure that if p is a pointer to at least n writable bytes, memcpy(p,p,n); will have no adverse side-effects. Such cases may arise...

[supercat]

...naturally in certain kinds of application code (e.g. a sort routine swapping an item with itself), and in implementations where they have no adverse side effects, letting those cases be handled by the general-case code may be more efficient than having to add special-case tests. Unfortunately, some compiler writers seem to think it's better to require that programmers add code the compiler may not be able to optimize out, so as to facilitate "optimization opportunities" which compilers would very seldom exploit anyway.

[ysap]

@Michael - If I am not mistaking, then the problem would be only when dest is overlapping either of the source vectors. Why would there be a problem if src1 and src2 overlap?

[supercat]

The "restrict" qualifier can actually allow much larger savings. For example, given void zap(char *restrict p1, char *restrict p2) { for (int i=0; i<50; i++) { p1[i] = 4; p2[i] = 9; } }, the restrict qualifiers would let the compiler rewrite the code as "memset(p1,4,50); memset(p2,9,50);". Restrict is vastly superior to type-based aliasing; it's a shame compilers focus more on the latter.

[Ciro Santilli OurBigBook.com]

@supercat great example, added to answer.

[supercat]

Glad you like it. Optimizations like that can be made by humans, but it's good to give the compiler freedom since a compiler may know things a human can't (e.g. combining things in a loop sometimes reduces loop overhead, but splitting things may allow use of parallelism or specialized instructions). The restrict keyword makes that possible in ways that type-based aliasing does not.

[tim18]

restrict normally has an effect only when pointing to an object which is modified, in which case it asserts no hidden side effects need be taken into account. Most compilers use it to facilitate vectorization. Msvc uses run time check for data overlap for that purpose.

[tim18]

Optimization based on type based aliasing is important to avoid concerns about pointers being modified as a side effect of storing into other data types. Reluctance to standardize restrict beyond c99 is extreme so it seems the widely accepted __restrict will never become a standard.

[Ciro Santilli OurBigBook.com]

@tim18 "Reluctance to standardize restrict beyond c99 is extreme so it seems the widely accepted __restrict will never become a standard": do you mean in C++? What is __restrict?

[supercat]

@tim18: The "restrict" keyword can enable many optimizations that even aggressive type-based optimization cannot. Further, the existence of "restrict" in the language--unlike aggressive type-based aliasing--never makes it impossible to accomplish tasks as efficiently as would be possible in their absence (since code which would be broken by "restrict" can simply not use it, while code which is broken by aggressive TBAA must be often be rewritten in less-efficient fashion).

[tim18]

restrict is an alternate spelling of restrict which is supported by several C++ compilers, including g++, MSVC++, Intel C++. Each of those compilers has a slightly different flavor, but only Intel C++ has an option to support restrict without the underscores in C++. All support __restrict in C89 as well as C99... Some support __restrict as an alternate spelling.

[supercat]

@tim18: Surround things that contain double-underlines in backticks, as in __restrict. Otherwise the double-underlines may be misinterpreted as an indication that you're shouting.

[Admin]

Adding the register keyword to the for loop variable also makes it faster in addition to adding restrict.

[Mark Fischler]

Actually, the register keyword is only advisory. And in compilers since about the year 2000, the i (and the n for the comparison) in the example will be optimized into a register whether or not you use the register keyword.

[remcycles]

More important than not shouting is that the underscores have meaning directly relevant to the point you are trying to make.

[elcuco]

GCC 7 and Clang 9 - at any optimization (-O1, -Os) __restrict__ as no more meaning, as the compiler will optimize the same. Reminds me of the register keyword. Still upvote for the epic answer. Nice job!

[Ciro Santilli OurBigBook.com]

@elcuco hmmm, I'm surprised that restrict has no meaning, since the compiler simply cannot know if the memories will be aliased or not, since restrict actually changes the function interface and forbids aliased callers.

[Gabriel Staples]

For anyone else like me who is wondering what in the world "pointer aliasing" means, Cornell University writes about it here: cvw.cac.cornell.edu/vector/coding_aliasing.

[supercat]

@MarkFischler: At least when targeting ARM targets, GCC still makes use of the register keyword when using -O0. In some cases that keyword will allow the compiler to generate code at -O0 which would be as efficient (sometimes more efficient!) as at higher optimization level if things like stack frame setup/teardown were omitted on leaf functions that don't use the stack.