xsimd
xsimd copied to clipboard
xtensor-stack
Invalid CodeGen with clang-cl (From visual studio 2019) in release builds (64 bits, ssse3)
Hi,
I've found that the following code snippet does not produce correct CodeGen with clang-cl.
xsimd::batch<float, 4> e{ 0, 1.f, 2.f, 3.f };
auto f = xsimd::exp(e);
for (int i = 0; i < 4; ++i)
CATCH_CHECK(f[i] == Approx(std::exp(e[i])));
Here is the output result:
Additional Info:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
ClangVectorsTests.cpp line 575: Failed
CATCH_CHECK( f[i] == Approx(std::exp(e[i])) )
with expansion:
CATCH_CHECK( 2.0f == Approx( 2.7182817459 ) )
ClangVectorsTests.cpp line 575: Failed
CATCH_CHECK( f[i] == Approx(std::exp(e[i])) )
with expansion:
CATCH_CHECK( 4.0f == Approx( 7.3890562057 ) )
ClangVectorsTests.cpp line 575: Failed
CATCH_CHECK( f[i] == Approx(std::exp(e[i])) )
with expansion:
CATCH_CHECK( 16.0f == Approx( 20.0855369568 ) )
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
After looking at the output, it seems that it is computing pow(2, x); instead of exp(x);
Here is the assembly code:
xsimd::batch<float, 4> e{ 0, 1.f, 2.f, 3.f };
0000000140021982 movaps xmm0,xmmword ptr [__xmm@40400000400000003f80000000000000 (014019C700h)]
0000000140021989 movaps xmmword ptr [e],xmm0
000000014002198D movaps xmm0,xmmword ptr [rbp+70h]
auto f = xsimd::exp(e);
0000000140021991 movaps xmmword ptr [f],xmm0
0000000140021995 xor esi,esi
0000000140021997 jmp ____C_A_T_C_H____T_E_S_T____78+3C5h (01400219B5h)
0000000140021999 nop dword ptr [rax]
for (int i = 0; i < 4; ++i)
00000001400219A0 mov rsi,qword ptr [rbp+1B8h]
00000001400219A7 add rsi,4
00000001400219AB cmp rsi,10h
00000001400219AF je ____C_A_T_C_H____T_E_S_T____78+524h (0140021B14h)
CATCH_CHECK(f[i] == Approx(std::exp(e[i])));
00000001400219B5 lea rdi,[rbp-30h]
00000001400219B9 mov rcx,rdi
00000001400219BC lea rdx,[string "f[i] == Approx(std::exp(e[i]))" (01401C20C2h)]
00000001400219C3 call Catch::StringRef::StringRef (01400B2ED0h)
00000001400219C8 lea rax,[string "C:\\dev\\dev1\\base-cpp\\tests\\src\\m"... (01401BE058h)]
00000001400219CF mov qword ptr [rbp+120h],rax
00000001400219D6 mov qword ptr [rbp+128h],23Fh
00000001400219E1 lea rax,[string "CATCH_CHECK" (01401BF818h)]
00000001400219E8 mov qword ptr [rbp+0B0h],rax
00000001400219EF mov qword ptr [rbp+0B8h],0Bh
00000001400219FA mov dword ptr [rsp+20h],2
0000000140021A02 lea rcx,[catchAssertionHandler]
0000000140021A09 lea rdx,[rbp+0B0h]
0000000140021A10 lea r8,[rbp+120h]
0000000140021A17 mov r9,rdi
0000000140021A1A call Catch::AssertionHandler::AssertionHandler (014008C460h)
0000000140021A1F mov qword ptr [rbp+1B8h],rsi
0000000140021A26 movss xmm0,dword ptr e[rsi]
0000000140021A2C call expf (0140171C60h)
0000000140021A31 xorps xmm1,xmm1
0000000140021A34 cvtss2sd xmm1,xmm0
0000000140021A38 lea rcx,[rbp+0C0h]
0000000140021A3F call Catch::Detail::Approx::Approx (014008B9D0h)
0000000140021A44 mov rax,qword ptr [rbp+1B8h]
0000000140021A4B lea rsi,[rax+rbp]
0000000140021A4F add rsi,30h
0000000140021A53 movss xmm0,dword ptr [rsi]
0000000140021A57 xorps xmm1,xmm1
0000000140021A5A cvtss2sd xmm1,xmm0
0000000140021A5E lea rcx,[rbp+0C0h]
0000000140021A65 call Catch::Detail::Approx::equalityComparisonImpl (014008BEC0h)
0000000140021A6A mov bl,al
0000000140021A6C lea rcx,[rbp+120h]
0000000140021A73 lea rdx,[string "==" (01401BFAA9h)]
0000000140021A7A call Catch::StringRef::StringRef (01400B2ED0h)
0000000140021A7F mov byte ptr [rbp+138h],1
0000000140021A86 mov byte ptr [rbp+139h],bl
0000000140021A8C lea rax,[Catch::BinaryExpr<float const & __ptr64,Catch::Detail::Approx const & __ptr64>::`vftable' (014019CC48h)]
0000000140021A93 mov qword ptr [rbp+130h],rax
0000000140021A9A mov qword ptr [rbp+140h],rsi
0000000140021AA1 movups xmm0,xmmword ptr [rbp+120h]
0000000140021AA8 lea rax,[rbp+148h]
0000000140021AAF movups xmmword ptr [rax],xmm0
0000000140021AB2 lea rax,[rbp+0C0h]
0000000140021AB9 mov qword ptr [rbp+158h],rax
0000000140021AC0 lea rcx,[catchAssertionHandler]
0000000140021AC7 lea rdx,[rbp+130h]
0000000140021ACE call Catch::AssertionHandler::handleExpr (014008C650h)
0000000140021AD3 lea rcx,[rbp+130h]
0000000140021ADA call Catch::ITransientExpression::~ITransientExpression (014008D9A0h)
0000000140021ADF lea rcx,[catchAssertionHandler]
0000000140021AE6 call Catch::AssertionHandler::complete (014008C700h)
0000000140021AEB cmp byte ptr [rbp+1A2h],0
0000000140021AF2 jne ____C_A_T_C_H____T_E_S_T____78+3B0h (01400219A0h)
0000000140021AF8 mov rcx,qword ptr [rbp+1A8h]
0000000140021AFF mov rax,qword ptr [rcx]
0000000140021B02 lea rdx,[catchAssertionHandler]
0000000140021B09 call qword ptr [rax+88h]
0000000140021B0F jmp ____C_A_T_C_H____T_E_S_T____78+3B0h (01400219A0h)
Here is the compiler flags (stripped from include paths):
ClCompile:
All outputs are up-to-date.
C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\Llvm\x64\bin\clang-cl.exe /c /Z7 /nologo /W3 /WX- /diagnostics:column /Ox /Ob2 /Oi /Ot /O
y /D WIN32 /D _WINDOWS /D _USE_MATH_DEFINES=1 /D SMTG_RENAME_ASSERT=1 /D RELEASE=1 /D PP_RELEASE=1 /D NDEBUG=1 /D PP_MACHINE_LITTLE_ENDIAN=1 /D PP_PLATFORM_WINDOWS=1 /D LINK_PLATFORM_WINDOWS=1 /D NOMINMAX /D NOCOMM /D NOMCX /D WIN32_LEAN_AND_MEAN /D WINVER=0x0601 /D _WIN32_WINNT
=0x0601 /D _UNICODE /D UNICODE /D PP_MACHINE_64_BIT=1 /D PP_COMPILER_MSVC /D _CRT_SECURE_NO_WARNINGS /D PP_COMPILER_CLANG_CL /D PP_MACHINE_INTEL=1 /D PP_SSE=1 /D PP_VECSZ=4 /D PP_ALIGN=32 /D "CMAKE_INTDIR=\"Release\"" /D _UNICODE /D UNICODE /EHsc /MT /GS /fp:fast /GR /std:c++17
/Fo"base-tests.dir\Release\\" /Gd /TP -m64 /clang:-fno-math-errno /clang:-ffp-model=fast /clang:-ffp-exception-behavior=ignore /clang:-ffp-contract=fast /clang:-fno-lto /clang:-mssse3 /clang:-mfpmath=sse "C:\dev\dev1\base-cpp\tests\src\main\cpp\tests\SocketTests.cpp"
Can you reproduce the issue? Is there a chance that you program appveyor/trevor to check windows/clang-cl? Many thanks.
Regards, Alexandre
I cross-link this issue with a corresponding one in LLVM bug tracker. It could be an LLVM issue... https://bugs.llvm.org/show_bug.cgi?id=47547
I'm unable to reproduce the issue on Linux with clang. clang and clang-cl share the same codebase so I'd think the bug is fixed, but cannot tell if it was an xsimd issue or not.
