[Bug]: new flat_hash_map to empty string could cause memory corruption in c++20

[JuniorHsu]

Describe the issue

while we migrate to c++20, we see regression for flat_hash_map<absl::string_view, absl::string_view> of empty c-string. see STR for detail

Steps to reproduce the problem

while we do

using StringViewMap = absl::flat_hash_map<absl::string_view, absl::string_view>;

const StringViewMap& replacements() {
  CONSTRUCT_ON_FIRST_USE(StringViewMap, {{"xxxxxxxxx", ""}, {".", "_"}});
}

auto& r = replacements();

std::cout << "r.size: " << r.size() << std::endl;
for (const auto& [k, v] : r) {
   std::cout << "k: " << k << " v: " << v << std::endl;
}

where the definition is in https://github.com/envoyproxy/envoy/blob/4af52310675fbaf7d1e05cf89134332421c8ba2e/source/common/common/macros.h#L34

this will output
r.size: 1 from my side and see the element in the flat_hash_map contains lots of code section in the key of map like [Alignment = 8UL, Alloc = std::allocator<std::pair<const std::string, Envoy::Regex::EngineFactory *> with size 2^64 -1

replacing "" with string() is one workaround

What version of Abseil are you using?

version = "20230802.1",

What operating system and version are you using?

ubuntu

What compiler and version are you using?

clang 12.0.1

What build system are you using?

bazel 6.1.0

Additional context

No response

[derekmauro]

The example code provided contains syntax errors. Sure, I could try to fix them, but then I would just be guessing what your code actually says and possibly hiding the error. Please provide a working reproduction.

[JuniorHsu]

The example code provided contains syntax errors. Sure, I could try to fix them, but then I would just be guessing what your code actually says and possibly hiding the error. Please provide a working reproduction.

const StringViewMap& replacements() const {

remove second const should work, sorry i try to have a minimal viable test but didn't compile again

I modify the initial STR too

[derekmauro]

I will not look at this until I have a complete reproduction.

[JuniorHsu]

I'll try to have a unit test

[JuniorHsu]

it's reproducible in my macOS from this commit
JuniorHsu@432bd53

~/code/abseil-cpp $ bazel test --sandbox_debug -c dbg  --test_output=streamed --verbose_failures --cxxopt=-std=c++20 absl/container:flat_hash_map_test --test_filter=FlatHashMap.EmptyStringViewMap
WARNING: Streamed test output requested. All tests will be run locally, without sharding, one at a time
INFO: Analyzed target //absl/container:flat_hash_map_test (0 packages loaded, 3 targets configured).
Running main() from gmock_main.cc
Note: Google Test filter = FlatHashMap.EmptyStringViewMap
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from FlatHashMap
[ RUN      ] FlatHashMap.EmptyStringViewMap
absl/container/flat_hash_map_test.cc:91: Failure
Expected equality of these values:
  r.size()
    Which is: 1
  2

[  FAILED  ] FlatHashMap.EmptyStringViewMap (0 ms)
[----------] 1 test from FlatHashMap (0 ms total)

[derekmauro]

Thank you for the reproduction.

This isn't an absl::flat_hash_map issue. std::unordered_map demonstrates the same behavior.

If you replace new type{__VA_ARGS__} with new type(__VA_ARGS__) in CONSTRUCT_ON_FIRST_USE then it will work, but I don't think you want to do that, because it will change the meaning of other code.

Basically this is a more complicated form of

std::vector<int> v(5, 0);  // Vector v holds 5 zeros -> {0, 0, 0, 0, 0}
std::vector<int> v{5, 0};  // Vector v holds a 5 and a 0 -> {5, 0}

C++20 added a new string_view constructor which is why you are seeing this issue in C++20 only. A different constructor is getting called.

My only advice is not to use CONSTRUCT_ON_FIRST_USE at all. absl::NoDestructor might be helpful here.
See also https://abseil.io/tips/88.

[JuniorHsu]

Thanks for digging on this. I filed envoyproxy/envoy#33804
Could be a design defect on c++ 20 if we could reproduce in std::unordered_map