Mirrors/asmjit
1
0
Fork 0
mirror of https://github.com/asmjit/asmjit.git synced 2025-12-17 20:44:37 +03:00
Code Issues Packages Projects Releases Wiki Activity

Codebase update and improvements, instruction DB update

Browse Source 
* 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
This commit is contained in:
kobalicek
2025-11-02 22:31:14 +01:00
parent 5134d396bd
commit b56f4176cb
209 changed files with 1687 additions and 1857 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
asmjit-testing/tests/asmjit_test_unicompiler_avx2fma.cpp Normal file
View File

@@ -0,0 +1,49 @@
// 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