63

Consider these two functions:

void foo() {} void bar() {}

is it guaranteed that &foo != &bar?

Similarly,

template<class T> void foo() { }

is it guaranteed that &foo<int> != &foo<double>?

There are two linkers I know of that fold function definitions together.

MSVC aggressively COMDAT folds functions, so two functions with the same implementation can be turned into one function. As a side effect, the two functions share the same address. I was under the impression that this was illegal, but I cannot find where in the standard it is made illegal.

The Gold linker also folds functions, with both a safe and all setting. safe means that if a function address is taken, it is not folded, while all folds even if the address is taken. So gold's fold safe behaves as-if functions have distinct addresses.

While folding might be unexpected, and there is code that relies on distinct (identical implementation) functions having different addresses (so it can be dangerous to fold), is it actually illegal under the current C++ standard? (C++14 at this point) (Naturally as-if safe folding is legal)

Improve this question
19
  • 5
    I'm sure I've seen this question before Commented Oct 23, 2014 at 17:37
  • 5
    Related : Why do two functions have the same address? Commented Oct 23, 2014 at 17:40
  • 1
    @MarcGlisse: Kudos for finding that, just one observation: They never say relying on any two functions having different addresses is allowed. Quote: "ICF can be unsafe, however, as it can change the run-time behaviour of code that relies on each function having a unique address." Commented Oct 24, 2014 at 18:45
  • 5
    @LightnessRacesinOrbit: Yes, you have seen it before: stackoverflow.com/q/14188612 Though with fewer upvotes, answers and controversy. ;-) Commented Nov 12, 2014 at 1:53
  • 1
    @LightnessRacesinOrbit any flag for merging? From my experience with merges this one seems reasonable, the answer to your question definitely fits in very well here. Commented Nov 15, 2014 at 3:24

4 Answers 4

Reset to default
35

It looks like defect report 1400: Function pointer equality deals with this issue and seems to me to say that it is okay to do this optimization but as comments indicate, there is disagreement. It says (emphasis mine):

According to 5.10 [expr.eq] paragraph 2, two function pointers only compare equal if they point to the same function. However, as an optimization, implementations are currently aliasing functions that have identical definitions. It is not clear whether the Standard needs to deal explicitly with this optimization or not.

and the response was:

The Standard is clear on the requirements, and implementations are free to optimize within the constraints of the “as-if” rule.

The question is asking about two issues:

  • Is it okay for these pointers to be considered equal
  • Is it okay to coalesce the functions

Based on comments I see two interpretations of the response:

  1. This optimization is ok, the standard gives the implementation this freedom under the as-if rule. The as-if rule is covered in section 1.9 and means the implementation only has to emulate the observable behavior with respect to the requirements of the standard. This is still my interpretation of the response.

  2. The issue is at hand is completely ignored and the statement merely says no adjustment to the standard is required because clearly the as-if rules covers this but the interpretation is left as an exercise to the reader. Although I acknowledge due to the terseness of the response I can not dismiss this view, it ends up being a totally unhelpful response. It also seems inconsistent with the responses in the other NAD issues which as far as I can tell point out issue if they exist.

What the draft standard says

Since we know we are dealing with the as-if rule, we can start there and note that section 1.8 says:

Unless an object is a bit-field or a base class subobject of zero size, the address of that object is the address of the first byte it occupies. Two objects that are not bit-fields may have the same address if one is a subobject of the other, or if at least one is a base class subobject of zero size and they are of different types; otherwise, they shall have distinct addresses.4

and note 4 says:

Under the “as-if” rule an implementation is allowed to store two objects at the same machine address or not store an object at all if the program cannot observe the difference

but a note from that section says:

A function is not an object, regardless of whether or not it occupies storage in the way that objects do

although it is not normative, the requirements for an object laid out in paragraph 1 do not make sense in the context of a function and so it is consistent with this note. So we are explicitly restricted from aliasing objects with some exceptions but not such restriction applies to functions.

Next we have section 5.10 Equality operators which says (emphasis mine):

[...]Two pointers compare equal if they are both null, both point to the same function, or both represent the same address (3.9.2), otherwise they compare unequal.

which tells us two pointers are equal if they are:

  • Null pointers
  • Point to the same function
  • Represent the same address

The or both represent the same address seems to give enough latitude to allow a compiler to alias two different functions and does not require pointers to different functions to compare unequal.

Observations

Keith Thompson has made some great observations that I feel are worth adding to the answer since they get to core issues involved, he says:

If a program prints the result of &foo == &bar, that's observable behavior; the optimization in question changes the observable behavior.

which I agree with and if we could shows that there is a requirement for the pointers to be unequal that would indeed violate the as-if rule but so far we can not show that.

and:

[...]consider a program that defines empty function and uses their addresses as unique values (think about SIG_DFL, SIG_ERR, and SIG_IGN in <signal.h> / <csignal>). Assigning them the same address would break such a program

As I noted in my comment the C standard requires these macros to generate distinct values, from 7.14 in C11:

[...]which expand to constant expressions with distinct values that have type compatible with the second argument to, and the return value of, the signal function, and whose values compare unequal to the address of any declarable function[...]

So although this case is covered perhaps there are other cases that would make this optimization dangerous.

Update

Jan Hubička a gcc developer wrote a blog post Link time and inter-procedural optimization improvements in GCC 5, code folding was one of many topics he covered.

I asked him to comment on whether folding identical functions to the same address was conforming behavior or not and he says it is not conforming behavior and indeed such an optimization would break gcc itself:

It is not conforming to turn two functions to have same address, so MSVC is quite aggressive here. Doing so, for example, breaks GCC itself because to my surprise address compare is done in the precompiled headers code. It works for many other projects, including Firefox.

In hindsight, after months more of reading defect reports and thinking about optimization issues, I am biased towards a more conservative reading of the committee's response. Taking the address of a function is observable behavior and therefore folding identical functions would violate the as-if rule.

Update 2

Also see this llvm-dev discussion: Zero length function pointer equality:

This is a well-known conformance-violating bug in link.exe; LLVM should not be making things worse by introducing a similar bug itself. Smarter linkers (for example, I think both lld and gold) will do identical function combining only if all but one of the function symbols is only used as the target of calls (and not to actually observe the address). And yes, this non-conforming behavior (rarely) breaks things in practice. See this research paper.

Improve this answer
Sign up to request clarification or add additional context in comments.

21 Comments

close, but not quite: it is allowed for two different functions to share the same binary code (under as-if), but it may not be ok for two different function's addresses to be compared and be equal. One way to approach that is to require "stub" noops or jmps at the front of the function, so each gets a different address, but the same body (as I believe some non-MSVC COMDAT folding equivalents do)
@Yakk: That would round counter to the common definition of address, which holds sway if there is no different definition in the standard.
I find the committee response not merely terse, but opaque. A footnote in the standard says that "an implementation is free to disregard any requirement of this International Standard as long as the result is as if the requirement had been obeyed, as far as can be determined from the observable behavior of the program". If a program prints the result of &foo == &bar, that's observable behavior; the optimization in question changes the observable behavior. My reading of the response is that the pointers must compare unequal, but I suspect that's not what was meant.
@ShafikYaghmour: It's not explicit, but I see it as common sense. Obviously others disagree. The question is whether two distinctly defined functions can be "the same function". Can an optimizing compiler that shares code for two functions cause them to be "the same function", in a manner visible to user code? Can it do the same for distinct objects? Practically, consider a program that defines empty function and uses their addresses as unique values (think about SIG_DFL, SIG_ERR, and SIG_IGN in <signal.h> / <csignal>). Assigning them the same address would break such a program.
@KeithThompson: It would seem your interpretation would allow a function to be merged with a distinct-but-identical functions whose address is not taken, but every function whose address is taken may only be merged if their addresses compare as distinct (on an architecture where pointers have more bits than hardware addresses, it may be possible for two pointers to represent the same physical address but still appear as distinct). If the intention is that functions whose address is taken must yield distinct addresses, I wonder why it doesn't say so?
|
11

Yes. From the standard (§5.10/1): "Two pointers of the same type compare equal if and only if they are both null, both point to the same function, or both represent the same address"

Once they have been instantiated, foo<int> and foo<double> are two different functions, so the above applies to them as well.

Improve this answer

22 Comments

That doesn't prevent them from representing the same address but pointing to different functions.
@Nawaz "the same address" is relevant for objects, where the object representation requires them to have an address. And it is weasel wording to avoid the issue where a one past the end of the array ends up pointing to another object. It's the "both point to the same function" which is relevant here.
@MikeSeymour Functions don't have addresses (in the standard).
@LightnessRacesinOrbit You won't find a statement that functions don't have addresses. However, only objects occupy storage, and only things which occupy storage can have addresses. See too §5.3.1/3; the unary & operator on a function returns a "pointer", not its address (unlike the case for objects).
@Deduplicator The standard defines all of the terms it uses. As far as the standard is concerned, despite the careless wording of one section title (which uses the word address in its everyday sense, and not in the sense the standard uses it), functions simply don't have addresses.
|
11

So the problematic part is clearly the phrase or both represent the same address (3.9.2).

IMO this part is clearly there to define the semantics for object pointer types. And only for object pointer types.

The phrase references section 3.9.2, which means we should look there. 3.9.2 talks (among others) about the addresses that object pointers represent. It does not talk about the addresses that function pointers represent. Which, IMO, leaves just two possible interpretations:

1) The phrase simply does not apply to function pointers. Which leaves just the two null pointers and two pointers to the same function comparing equal, which is what probably most of us expected.

2) The phrase does apply. Since it's referring to 3.9.2, which says nothing about the addresses that function pointers represent, we may make any two function pointers compare equal. Which is very unexpected and of course renders comparing function pointers utterly useless.

So, while technically an argument could be made that (2) is a valid interpretation, IMO it's not a meaningful interpretation and thus should be disregarded. And since not everyone seems to agree on this, I also think that a clarification in the standard is needed.

Improve this answer

Comments

3

5.10 Equality operators [expr.eq]

1 The == (equal to) and the != (not equal to) operators group left-to-right. The operands shall have arithmetic, enumeration, pointer, or pointer to member type, or type std::nullptr_t. The operators == and != both yield true or false, i.e., a result of type bool. In each case below, the operands shall have the same type after the specified conversions have been applied.
2 If at least one of the operands is a pointer, pointer conversions (4.10) and qualification conversions (4.4) are performed on both operands to bring them to their composite pointer type (Clause 5). Comparing pointers is defined as follows: Two pointers compare equal if they are both null, both point to the same function, or both represent the same address (3.9.2), otherwise they compare unequal.

Let's take the last bit-for-bit:

  1. Two null pointers compare equal.
    Good for your sanity.
  2. Two pointers to the same function compare equal.
    Anything else would be extremely surprising.
    It also means that only one out-of-line version of any inline-function may ever have its address taken, unless you want to make function-pointer comparisons prohibitively complicated and expensive.
  3. Both represent the same address.
    Now that one is what it's all about. Dropping this and reducing if and only if to a simple if would leave it to interpretation, but that's a clear mandate to make any two functions identical, as long as it does not otherwise change observable behavior of a conformant program.
Improve this answer

Comments

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.