mirror of
https://github.com/asmjit/asmjit.git
synced 2025-12-16 20:17:05 +03:00
b56f4176cb
* Denested src folder to root, renamed testing to asmjit-testing
* Refactored how headers are included into <asmjit/...> form. This
is necessary as compilers would never simplify a path once a ..
appears in include directory - then paths such as ../core/../core
appeared in asserts, which was ugly
* Moved support utilities into asmjit/support/... (still included
by asmjit/core.h for convenience and compatibility)
* Added CMakePresets.json for making it easy to develop AsmJit
* Reworked CMakeLists to be shorter and use CMake option(),
etc... This simplifies it and makes it using more standard
features
* ASMJIT_EMBED now creates asmjit_embed INTERFACE library,
which is accessible via asmjit::asmjit target - this simplifies
embedding and makes it the same as library targets from a CMake
perspective
* Removed ASMJIT_DEPS - this is now provided by cmake target
aliases - 'asmjit::asmjit' so users should not need this variable
* Changed meaning of ASMJIT_LIBS - this now contains only AsmJit
dependencies without asmjit::asmjit target alias. Don't rely on
ASMJIT_LIBS anymore as it's only used internally
* Removed ASMJIT_NO_DEPRECATED option - AsmJit is not going
to provide controllable deprecations in the future
* Removed ASMJIT_NO_VALIDATION in favor of ASMJIT_NO_INTROSPECTION,
which now controls query, features, and validation API presence
* Removed ASMJIT_DIR option - it was never really needed
* Removed AMX_TRANSPOSE feature from instruction database (X86).
Intel has removed it as well, so it's a feature that won't
be siliconized
50 lines
1.3 KiB
C++
50 lines
1.3 KiB
C++
// This file is part of AsmJit project <https://asmjit.com>
|
|
//
|
|
// See <asmjit/core.h> or LICENSE.md for license and copyright information
|
|
// SPDX-License-Identifier: Zlib
|
|
|
|
#include <asmjit/ujit.h>
|
|
|
|
#if defined(ASMJIT_UJIT_X86)
|
|
|
|
#if defined(_MSC_VER)
|
|
#include <intrin.h>
|
|
#else
|
|
#include <immintrin.h>
|
|
#endif
|
|
|
|
namespace UniCompilerTests {
|
|
|
|
// A reference implementation of MUL+ADD with the use of FMA. This has to be provided otherwise the
|
|
// compiler may use FPU registers in 32-bit x86 case, which would make the result different than when
|
|
// compiled by JIT compiler that would use XMM registers (32/64-bit SSE/AVX operations).
|
|
|
|
float fmadd_fma_ref(float a, float b, float c) noexcept {
|
|
__m128 av = _mm_set1_ps(a);
|
|
__m128 bv = _mm_set1_ps(b);
|
|
__m128 cv = _mm_set1_ps(c);
|
|
|
|
return _mm_cvtss_f32(_mm_fmadd_ss(av, bv, cv));
|
|
}
|
|
|
|
double fmadd_fma_ref(double a, double b, double c) noexcept {
|
|
__m128d av = _mm_set1_pd(a);
|
|
__m128d bv = _mm_set1_pd(b);
|
|
__m128d cv = _mm_set1_pd(c);
|
|
|
|
return _mm_cvtsd_f64(_mm_fmadd_sd(av, bv, cv));
|
|
}
|
|
|
|
void madd_fma_check_valgrind_bug(const float a[4], const float b[4], const float c[4], float dst[4]) noexcept {
|
|
__m128 av = _mm_loadu_ps(a);
|
|
__m128 bv = _mm_loadu_ps(b);
|
|
__m128 cv = _mm_loadu_ps(c);
|
|
|
|
__m128 dv = _mm_fmadd_ss(av, bv, cv);
|
|
_mm_storeu_ps(dst, dv);
|
|
}
|
|
|
|
} // {UniCompilerTests}
|
|
|
|
#endif // ASMJIT_UJIT_X86
|