Title
Strange specification of max_size() for an allocator
Status
new
Section
[allocator.requirements]
Submitter
Jon Cohen

Created on 2017-12-06.00:00:00, last changed 2018-08-21.18:33:48.

Messages

Date: 2018-08-21.18:33:48

Proposed resolution:

This wording is relative to N4762.

  1. Change [allocator.requirements], Table 32 — "Descriptive variable definitions", as indicated:

    Table 32 — Descriptive variable definitions
    Variable Definition
    T, U, C any cv-unqualified object type (3.9)
    a, a1, a2 lvalues of type X
    a3 an lvalue of type const X
  2. Change [allocator.requirements], Table 33 — "Cpp17Allocator requirements", as indicated:

    Table 33 — Cpp17Allocator requirements
    Expression Return type Assertion/note
    pre-/post-condition
    Default
    a3.max_size() X::size_type the largest value that can
    meaningfully be passed to
    X::allocate().
    [Note: Larger values might cause
    an exception to be thrown. — end note]
    Throws: Nothing.
    numeric_limits<size_type>::max()
    / sizeof(value_type)
Date: 2018-08-15.00:00:00

[ 2018-08-21, Jonathan comments and provides wording ]

The phrase "the largest value that can meaningfully be passed to X::allocate()" is meaningless. Is it a requirement on the caller, so that larger values must not be passed? Or a hint from the allocator implementor that larger values will produce a bad_alloc exception? Can the return value change dynamically, based on the free memory available to the allocator?! — LWG 197 says it can't change.

As noted in the LWG 2162 comments, we don't currently guarantee it can be called on a const object (so allocator_traits will not use the allocator's max_size() if it's non-const, although that was unclear before DR 2284). In addition to adding "Throws: nothing" we should ensure it's callable on const lvalues, and clarify what "meaningfully" means and who is supposed to care about it. My proposed resolution doesn't achieve all of this, but is a start.

Date: 2018-01-29.17:21:16

[ 2018-01; Priority set to 3 after mailing list discussion ]

Date: 2017-12-11.20:43:01

Table 31 in the C++17 standard specifies X::max_size() (where X is an allocator type) as "The largest value that can meaningfully be passed to X::allocate()". Noticeably missing is the statement "Throws: Nothing".

As an example of why this is an issue, note that vector::max_size() and allocator_traits::max_size() are both marked noexcept. We must then interpret max_size() as being allowed to sometimes call std::terminate, or else {vector, allocator_traits, ...}::max_size() must be allowed to directly calculate numeric_limits<size_type>::max() / sizeof(value_type) instead of querying the allocator, even if Alloc::max_size() exists. This seems like a bug in the wording for the requirements of max_size() in an allocator type. I think an issue should be opened on this subject to add Throws: Nothing or similar to the requirements of max_size() for an allocator.

As an example consider writing up a framework to test the exception-safety of types in a given framework, since they were all written in an exception-free environment. One of the types in the framework is an allocator which, in a controlled way, can throw an exception at any point where it is allowed by the standard. It's important that the test framework be as pedantic as possible, so the allocator type throws on max_size(), since it is currently allowed to by the standard. When a reasonable vector implementation (at least those in libstdc++ and msvc) is, for example, asked to construct a vector from an initializer_list, it will call allocator_traits<Alloc>::max_size(), which will terminate the program because the exception thrown in Alloc::max_size() propagated through the noexcept traits function. Although this is conformant behavior, I think it's a bug in the standard that a function as benign as max_size() can terminate the program in this manner, and I think the fix is that a conformant allocator should be required to supply a non-throwing max_size() member function.

Daniel:

This problem was shortly discussed during review of LWG 2162 (see comment 2012-08-05). At that time the more drastic but also more consistent requirement that an allocator's max_size function shall not throw exceptions has not been added. IMO this position should be reconsidered to follow the spirit of the new issue LWG 3044.

History
Date User Action Args
2018-08-21 18:33:48adminsetmessages: + msg10053
2018-08-21 18:33:48adminsetmessages: + msg10052
2018-01-29 17:21:16adminsetmessages: + msg9628
2017-12-06 00:00:00admincreate