c++2a: string::reserve still shrinks capacity (P0966)

[rprichard]

Bugzilla Link	45368
Resolution	FIXED
Resolved on	Nov 26, 2020 01:21
Version	unspecified
OS	Linux
CC	@mkurdej,@dwblaikie,@ldionne,@mclow,@tambry
Fixed by commit(s)	841132e

Extended Description

libc++ documents that it has implemented P0966R1[1][2], "string::reserve Should Not Shrink", but as far as I can tell, string::reserve still shrinks the capacity.

#include
#include <stdio.h>
int main() {
std::string foo;
foo.reserve(2000);
printf("%zu\n", foo.capacity());
foo.reserve(1000);
printf("%zu\n", foo.capacity());
return 0;
}

$ /x/llvm-upstream/stage1-install/bin/clang++ -stdlib=libc++ test.cpp && LD_LIBRARY_PATH=/x/llvm-upstream/stage1-install/lib ./a.out

2015
1007

[1] https://libcxx.llvm.org/cxx2a_status.html
[2] http://wg21.link/P0966R1

The P0966R1 change was implemented in D54992[3] / svn commit 347789[4].

[3] https://reviews.llvm.org/D54992
[4] https://reviews.llvm.org/rL347789

P0966R1 allows reserve() and reserve(0) to do different things, so they need to be overloads rather than use a default argument of 0, and the libc++ commit does split the function into two overloaded functions. libc++'s string::reserve(size_type) function still lowers the capacity, though. Its shrink_to_fit() still calls reserve(), which calls reserve(0).

http://eel.is/c++draft/string.capacity#itemdecl:6 reads:

constexpr void reserve(size_type res_arg);

Effects: A directive that informs a basic_­string of a planned change in size, so that the storage allocation can be managed accordingly. After reserve(), capacity() is greater or equal to the argument of reserve if reallocation happens; and equal to the previous value of capacity() otherwise. Reallocation happens at this point if and only if the current capacity is less than the argument of reserve().

The libc++ commit added a test that looks like it verifies that reserve doesn't shrink, but it doesn't really do that. In string.capacity/reserve.pass.cpp:

 template <class S>
 void
 test(S s, typename S::size_type res_arg)
 {
     typename S::size_type old_cap = s.capacity();
     ((void)old_cap); // Prevent unused warning
     S s0 = s;
     if (res_arg <= s.max_size())
     {
         s.reserve(res_arg);
         assert(s == s0);
         assert(s.capacity() >= res_arg);
         assert(s.capacity() >= s.size());
+#if TEST_STD_VER > 17
+        assert(s.capacity() >= old_cap); // resize never shrinks as of P0966
+#endif

(I think the comment meant to say "reserve never shrinks" rather than "resize never shrinks"?)

This call to test() looks like it would catch the issue:

{
typedef std::string S;
...
{
S s(100, 'a');
s.erase(50);
test(s, 5);

We have a string, s with size() == 50 and capacity() >= 100. Calling reserve(5) should leave the capacity() >= 100 as of P0966R1, but libc++ shrinks the string to a little above 50. However, main() passes the string by-value to test(), and the copy constructor makes a new string that's shrunk-to-fit. i.e. AFAICT, the sections in main() that use erase() aren't testing the intended situation.

[rprichard]

assigned to @mkurdej

[mkurdej]

Thanks for a great bug report.
The patch is in a review: https://reviews.llvm.org/D91778.

[mkurdej]

Fixed by commit https://reviews.llvm.org/rG841132efda2157c5f9e07cf31469470a6481ffd9.