Floating Point gets miscompiled.

Mysticial — Sat, 09 Nov 2024 07:23:19 GMT

Compiler: Intel(R) oneAPI DPC++/C++ Compiler 2025.0.0 (2025.0.0.20241008)

OS: Windows (via Visual Studio 17.11.5)

The following code is incorrectly compiled:

#include <iostream> using std::cout; using std::endl; #define NO_INLINE __declspec(noinline) #define FORCE_INLINE inline __forceinline class Object{ public: FORCE_INLINE Object(uint64_t x) : Object(nullptr, (uint32_t)x, (double)(uint32_t)(x >> 32) * 4294967296.) {} FORCE_INLINE Object(void*, double L, double H){ double r0; r0 = (double)(L + H); hi = r0; r0 = (double)(r0 - H); lo = (double)(L - r0); } double lo; double hi; }; NO_INLINE Object func(const Object& a){ return a; } NO_INLINE void test_function(uint32_t k){ uint64_t stop = (uint64_t)1 << k; stop /= 4; for (uint64_t c = 0; c < stop; c++){ Object obj = func(c); cout << "c = " << c << ": obj = " << obj.hi << endl; } } int main(){ test_function(9); }

Compiler Flags:

/GS /W3 /QxAVX /Gy /Zc:wchar_t /Zi /O3 /fp:precise /D "NDEBUG" /D "_CONSOLE" /D "__INTEL_LLVM_COMPILER=20250000" /D "_UNICODE" /D "UNICODE" /Qipo /Zc:forScope /Oi /MD /Fa"x64\Release\" /EHsc /nologo /Fo"x64\Release\" /Fp"x64\Release\ICX2025 Miscompile 1.pch"

Expected Output:

c = 0: obj = 0 c = 1: obj = 1 c = 2: obj = 2 c = 3: obj = 3 c = 4: obj = 4 c = 5: obj = 5 c = 6: obj = 6 c = 7: obj = 7 ...

Actual Output:

c = 0: obj = 0 c = 1: obj = 6.95222e-310 c = 2: obj = 8.04776e-312 c = 3: obj = 8.04776e-312 c = 4: obj = 6.1857e-321 c = 5: obj = 1.7407e-315 c = 6: obj = 1.26481e-321 c = 7: obj = 3.95253e-323 ...

Re:Floating Point gets miscompiled.

yzh_intel — Tue, 12 Nov 2024 21:30:57 GMT

Hi, I reproduced the issue, it seems that the compiler performed some incorrect optimization. I have escalated the issue to compiler engineering team. Thanks for reporting issues to us.

Meanwhile, can you try to set the optimization level to /O2 instead of /O3? In my testing, the case passes with /O2.

Re: Re:Floating Point gets miscompiled.

yzh_intel — Thu, 24 Apr 2025 16:37:15 GMT

Hi @Mysticial I'd like to let you know that the issue has been fixed in compiler 2025.1, please let us know if you still see any issue.

topic Floating Point gets miscompiled. in Intel® oneAPI DPC++/C++ Compiler

Floating Point gets miscompiled.

Re:Floating Point gets miscompiled.

Re: Re:Floating Point gets miscompiled.