FYI, cause of the problem has been identified: UB due to signed integer overflow clang++ -fsanitize=signed-integer-overflow overflow.cpp ./a.out -2147483648 overflow.cpp:24:52: runtime error: signed integer overflow: 2147483647 - -2147483648 cannot be represented in type 'int' Thanks and credits to Jan Korous (and the awesome UndefinedBehaviorSanitizer :-)) Mateusz On 20 May 2018 at 11:47, Mateusz Loskot wrote:

Hi,

I posted this to clang mailing list a http://lists.llvm.org/pipermail/cfe-users/2018-May/001335.html but since it is related to Boost.GIL bugs I'm investigating, I'd like to reach for feedback here too.

I'll appreciate any comments regarding the issue explained below.

ML

---------- Forwarded message ---------- From: Mateusz Loskot Date: 19 May 2018 at 21:38 Subject: Boost.GIL test failing with clang 5.x while pass with 15 gcc/clang versions To: cfe-users@lists.llvm.org

Hi,

While testing Boost.GIL library [1] with gcc and clang, I noticed a peculiar issue. One particular test is failing with lang 5.x while passing with total of 15-17 other versions of clang and GCC (total workflow of CircleCI with at [2]).

Below is extracted minimal program equivalent to Boost.GIL channel_invert algorithm. It includes two variants: plain expression and the same expression wrapped with a function template:

#include <limits> #include <iostream> #include <typeinfo>

template <typename C> inline C channel_invert1(C x) { return std::numeric_limits<C>::max() - x + std::numeric_limits<C>::min(); }

template <typename C> inline C channel_invert2(C x) { return (x - std::numeric_limits<C>::max()) * (-1) + std::numeric_limits<C>::min(); }

int main() { int x = std::numeric_limits<int>::min(); std::cout << x << std::endl;

// plain expressions int x_invert1 = std::numeric_limits<int>::max() - x + std::numeric_limits<int>::min(); int x_invert2 = (x - std::numeric_limits<int>::max()) * (-1) + std::numeric_limits<int>::min(); std::cout << x_invert1 << std::endl; std::cout << x_invert2 << std::endl;

// the same expressions wrapped in function template std::cout << channel_invert1<int>(x) << std::endl; std::cout << channel_invert2<int>(x) << std::endl; }

If compiled as optimised variant (-O2 or -O3) with clang 5.x outputs the following:

-2147483648 2147483647 2147483647 -1 -1

The last two negative one is not expected.

If compiled with clang 3.9, 4.0 or gcc from 5.1 to 7.3 it outputs

-2147483648 2147483647 2147483647 2147483647 2147483647

Could anyone help me to understand what is going on in the clang 5 case? Or, what UB is this hitting?

[1] https://github.com/boostorg/gil/issues/89 [2] https://circleci.com/workflow-run/3a14dd64-6c38-46b2-a6da-678c0075ca27

Best regards, -- Mateusz Loskot, http://mateusz.loskot.net

-- Mateusz Loskot, http://mateusz.loskot.net