simd-everywhere · makise-homura · Feb 20, 2021 · Feb 20, 2021 · Feb 20, 2021 · Feb 20, 2021
diff --git a/README.md b/README.md
@@ -186,13 +186,16 @@ just use search and replace, manual changes are required pretty
 infrequently.
 
 For best performance, in addition to `-O3` (or whatever your compiler's
-equivalent is), you should enable OpenMP 4 SIMD support by defining
-`SIMDE_ENABLE_OPENMP` before including any SIMDe headers, and
-enabling OpenMP support in your compiler.  GCC and ICC both support a
-flag to enable only OpenMP SIMD support instead of full OpenMP (the OpenMP
-SIMD support doesn't require the OpenMP run-time library); for GCC the
-flag is `-fopenmp-simd` (requires GCC version 4.9 or later), for ICC
-the flag is `-qopenmp-simd`.  SIMDe also supports
+equivalent is), you should enable OpenMP 4 SIMD support in your compiler.
+GCC and ICC both support a flag to enable only OpenMP SIMD support instead
+of full OpenMP (the OpenMP SIMD support doesn't require the OpenMP run-time
+library); for GCC the flag is `-fopenmp-simd` (requires GCC version 4.9
+or later), for ICC the flag is `-qopenmp-simd`. Some compilers have this
+support implicitly enabled (the example is MCST Elbrus Compiler, or LCC
+for short). If for some reason you need to disable OpenMP support,
+you need to define `SIMDE_DISABLE_OPENMP` before including any of SIMDe
+headers, and probably disable its support in compiler, if it is implicitly
+enabled (for LCC, there is `-fno-openmp` flag). SIMDe also supports
 using [Cilk Plus](https://www.cilkplus.org/), [GCC loop-specific
 pragmas](https://gcc.gnu.org/onlinedocs/gcc/Loop-Specific-Pragmas.html),
 or [clang pragma loop hint

diff --git a/meson.build b/meson.build
@@ -8,15 +8,6 @@ project('SIMDe', 'c', 'cpp',
 cc = meson.get_compiler('c')
 cxx = meson.get_compiler('cpp')
 
-foreach additional_arg : [ '-Wno-reduced-alignment' ]
-    if cc.has_argument(additional_arg)
-        add_global_arguments(additional_arg, language : 'c')
-    endif
-    if cxx.has_argument(additional_arg)
-        add_global_arguments(additional_arg, language : 'cpp')
-    endif
-endforeach
-
 simde_neon_families = [
   'aba',
   'abd',

diff --git a/simde/arm/neon/types.h b/simde/arm/neon/types.h
@@ -174,7 +174,7 @@ SIMDE_ARM_NEON_TYPE_FLOAT_DEFINE_(64, 2, SIMDE_ALIGN_16_)
     #define SIMDE_ARM_NEON_NEED_PORTABLE_F64X1XN
     #define SIMDE_ARM_NEON_NEED_PORTABLE_F64X2XN
   #endif
-#elif defined(SIMDE_ARCH_X86) || defined(SIMDE_ARCH_AMD64)
+#elif defined(SIMDE_ARCH_X86) || defined(SIMDE_ARCH_AMD64) || defined(SIMDE_ARCH_E2K)
   #define SIMDE_ARM_NEON_NEED_PORTABLE_F32
   #define SIMDE_ARM_NEON_NEED_PORTABLE_F64
 

diff --git a/simde/simde-align.h b/simde/simde-align.h
@@ -116,6 +116,7 @@
   #define SIMDE_ALIGN_OF(Type) alignof(Type)
 #elif \
     HEDLEY_GCC_VERSION_CHECK(2,95,0) || \
+    HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
     HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
     HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
     HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \
@@ -264,6 +265,7 @@
 #if \
     HEDLEY_HAS_ATTRIBUTE(aligned) || \
     HEDLEY_GCC_VERSION_CHECK(2,95,0) || \
+    HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
     HEDLEY_CRAY_VERSION_CHECK(8,4,0) || \
     HEDLEY_IBM_VERSION_CHECK(11,1,0) || \
     HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
@@ -314,7 +316,8 @@
  */
 #if \
     HEDLEY_HAS_BUILTIN(__builtin_assume_aligned) || \
-    HEDLEY_GCC_VERSION_CHECK(4,7,0)
+    HEDLEY_GCC_VERSION_CHECK(4,7,0) || \
+    HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
   #define SIMDE_ALIGN_ASSUME_TO_UNCHECKED(Pointer, Alignment) \
     HEDLEY_REINTERPRET_CAST(__typeof__(Pointer), __builtin_assume_aligned(HEDLEY_CONST_CAST(void*, HEDLEY_REINTERPRET_CAST(const void*, Pointer)), Alignment))
 #elif HEDLEY_INTEL_VERSION_CHECK(13,0,0)
@@ -446,4 +449,18 @@
  */
 #define SIMDE_ALIGN_ASSUME_CAST(Type, Pointer) SIMDE_ALIGN_ASSUME_LIKE(SIMDE_ALIGN_CAST(Type, Pointer), Type)
 
+/* In some circumstances we need to define types as packed,
+ * because platform alignment is fixed when variable is placed
+ * on the stack (but this alignment is variable otherwise).
+ * so there's SIMDE_ALIGN_REDUCE_STRUCT macros that allows this
+ * for LCC compiler on Elbrus architecture.
+ */
+#if defined (SIMDE_BUG_LCC_STACK_ALIGNMENT_CAP)
+  #define SIMDE_ALIGN_REDUCE_STRUCT __attribute__((packed, aligned(16)))
+  #define SIMDE_ALIGN_REDUCE_ARRAY  __attribute__((aligned(16)))
+#else
+  #define SIMDE_ALIGN_REDUCE_STRUCT
+  #define SIMDE_ALIGN_REDUCE_ARRAY
+#endif
+
 #endif /* !defined(SIMDE_ALIGN_H) */
diff --git a/simde/simde-arch.h b/simde/simde-arch.h
@@ -182,13 +182,28 @@
    <https://en.wikipedia.org/wiki/Elbrus-8S> */
 #if defined(__e2k__)
 #define SIMDE_ARCH_E2K
-#define SIMDE_SKIP_EXTENDED_E2K_VECTOR_OPS /* Discard features unsupported by compiler */
 #endif
 
 /* Discard features unsupported by Elbrus compiler.
    For lcc > 1.25.10, it may be based on a version. */
 #if defined(__LCC__)
-#define SIMDE_SKIP_EXTENDED_E2K_VECTOR_OPS
+#define SIMDE_BUG_LCC_TOO_STRICT_VECTOR_SHIFTS_AND_COMPARES
+#define SIMDE_BUG_LCC_XOP_MISSING
+#define SIMDE_BUG_LCC_WARNING_ON_SHIFTS
+#define SIMDE_BUG_LCC_FMA_WRONG_RESULT
+#define SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2
+#define SIMDE_BUG_LCC_STACK_ALIGNMENT_CAP
+
+/* Some native functions on E2K with instruction set < v6
+   are declared as deprecated due to inefficiency.
+   Still they are more efficient than SIMDe implementation.
+   So we're using them, and switching off these deprecation warnings. */
+#define SIMDE_BUG_PCLMUL_XOP_DEPRECATED
+#define SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS _Pragma("diag_suppress 1215,1444")
+#define SIMDE_LCC_REVERT_DEPRECATED_WARNINGS _Pragma("diag_default 1215,1444")
+#else
+#define SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+#define SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
 #endif
 
 /* HP/PA / PA-RISC
@@ -267,7 +282,7 @@
 #  if defined(__SSE4_2__)
 #    define SIMDE_ARCH_X86_SSE4_2 1
 #  endif
-#  if defined(__XOP__) && !defined(__LCC__) /* LCC incorrectly defines __XOP__ */
+#  if defined(__XOP__)
 #    define SIMDE_ARCH_X86_XOP 1
 #  endif
 #  if defined(__AVX__)
@@ -285,7 +300,7 @@
 #  if defined(__AVX2__)
 #    define SIMDE_ARCH_X86_AVX2 1
 #  endif
-#  if defined(__FMA__) && !defined(__LCC__) /* LCC incorrectly defines __FMA__ */
+#  if defined(__FMA__)
 #    define SIMDE_ARCH_X86_FMA 1
 #    if !defined(SIMDE_ARCH_X86_AVX)
 #      define SIMDE_ARCH_X86_AVX 1
@@ -315,7 +330,7 @@
 #  if defined(__GFNI__)
 #    define SIMDE_ARCH_X86_GFNI 1
 #  endif
-#  if defined(__PCLMUL__) && !defined(SIMDE_ARCH_E2K) /* E2K has inefficient implementation of PCLMUL */
+#  if defined(__PCLMUL__)
 #    define SIMDE_ARCH_X86_PCLMUL 1
 #  endif
 #  if defined(__VPCLMULQDQ__)

diff --git a/simde/simde-common.h b/simde/simde-common.h
@@ -291,8 +291,14 @@
 #  endif
 #endif
 
-#if !defined(SIMDE_ENABLE_OPENMP) && ((defined(_OPENMP) && (_OPENMP >= 201307L)) || (defined(_OPENMP_SIMD) && (_OPENMP_SIMD >= 201307L)))
-#  define SIMDE_ENABLE_OPENMP
+#if !defined(SIMDE_DISABLE_OPENMP)
+  #if !defined(SIMDE_ENABLE_OPENMP) && ((defined(_OPENMP) && (_OPENMP >= 201307L)) || (defined(_OPENMP_SIMD) && (_OPENMP_SIMD >= 201307L)))
+  #  define SIMDE_ENABLE_OPENMP
+  #endif
+#else
+  #if defined(SIMDE_ENABLE_OPENMP)
+  #  undef SIMDE_ENABLE_OPENMP
+  #endif
 #endif
 
 #if !defined(SIMDE_ENABLE_CILKPLUS) && (defined(__cilk) || defined(HEDLEY_INTEL_VERSION))
@@ -322,7 +328,7 @@
 #  define SIMDE_VECTORIZE_SAFELEN(l) HEDLEY_PRAGMA(clang loop vectorize_width(l))
 #  define SIMDE_VECTORIZE_REDUCTION(r) SIMDE_VECTORIZE
 #  define SIMDE_VECTORIZE_ALIGNED(a)
-#elif HEDLEY_GCC_VERSION_CHECK(4,9,0) && !defined(__LCC__)
+#elif HEDLEY_GCC_VERSION_CHECK(4,9,0)
 #  define SIMDE_VECTORIZE HEDLEY_PRAGMA(GCC ivdep)
 #  define SIMDE_VECTORIZE_SAFELEN(l) SIMDE_VECTORIZE
 #  define SIMDE_VECTORIZE_REDUCTION(r) SIMDE_VECTORIZE

diff --git a/simde/x86/avx.h b/simde/x86/avx.h
@@ -91,7 +91,7 @@ typedef union {
       SIMDE_ALIGN_TO_16 SIMDE_POWER_ALTIVEC_VECTOR(double)             altivec_f64[2];
     #endif
   #endif
-} simde__m256_private;
+} SIMDE_ALIGN_REDUCE_STRUCT simde__m256_private;
 
 typedef union {
   #if defined(SIMDE_VECTOR_SUBSCRIPT)
@@ -149,7 +149,7 @@ typedef union {
       SIMDE_ALIGN_TO_16 SIMDE_POWER_ALTIVEC_VECTOR(double)             altivec_f64[2];
     #endif
   #endif
-} simde__m256d_private;
+} SIMDE_ALIGN_REDUCE_STRUCT simde__m256d_private;
 
 typedef union {
   #if defined(SIMDE_VECTOR_SUBSCRIPT)
@@ -207,7 +207,7 @@ typedef union {
       SIMDE_ALIGN_TO_16 SIMDE_POWER_ALTIVEC_VECTOR(double)             altivec_f64[2];
     #endif
   #endif
-} simde__m256i_private;
+} SIMDE_ALIGN_REDUCE_STRUCT simde__m256i_private;
 
 #if defined(SIMDE_X86_AVX_NATIVE)
   typedef __m256 simde__m256;
@@ -3935,7 +3935,7 @@ simde_mm256_loadu_si256 (void const * mem_addr) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m256
 simde_mm256_loadu2_m128 (const float hiaddr[HEDLEY_ARRAY_PARAM(4)], const float loaddr[HEDLEY_ARRAY_PARAM(4)]) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     return _mm256_loadu2_m128(hiaddr, loaddr);
   #else
     return
@@ -3951,7 +3951,7 @@ simde_mm256_loadu2_m128 (const float hiaddr[HEDLEY_ARRAY_PARAM(4)], const float
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m256d
 simde_mm256_loadu2_m128d (const double hiaddr[HEDLEY_ARRAY_PARAM(2)], const double loaddr[HEDLEY_ARRAY_PARAM(2)]) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     return _mm256_loadu2_m128d(hiaddr, loaddr);
   #else
     return
@@ -3967,7 +3967,7 @@ simde_mm256_loadu2_m128d (const double hiaddr[HEDLEY_ARRAY_PARAM(2)], const doub
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m256i
 simde_mm256_loadu2_m128i (const simde__m128i* hiaddr, const simde__m128i* loaddr) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     return _mm256_loadu2_m128i(hiaddr, loaddr);
   #else
     return
@@ -4547,8 +4547,21 @@ simde_mm256_permute_ps (simde__m256 a, const int imm8)
   return simde__m256_from_private(r_);
 }
 #if defined(SIMDE_X86_AVX_NATIVE)
+#if defined(SIMDE_BUG_LCC_STACK_ALIGNMENT_CAP)
+/* Patched implementation from e2kbuiltin.h */
+#  define simde_mm256_permute_ps(a, imm8) ({                        \
+    type_union_256 __attribute__((aligned(16))) __s, __dst;         \
+    __s.__v8sf = (__v8sf)(a);                                       \
+    SELECT_CONST_32F (__s.l.l1, __s.l.l0, __dst.l.l0, (imm8));      \
+    SELECT_CONST_32F (__s.l.l1, __s.l.l0, __dst.l.l1, (imm8) >> 4); \
+    SELECT_CONST_32F (__s.l.l3, __s.l.l2, __dst.l.l2, (imm8));      \
+    SELECT_CONST_32F (__s.l.l3, __s.l.l2, __dst.l.l3, (imm8) >> 4); \
+    (__m256)(__dst.__v8sf);                                         \
+})
+#else
 #  define simde_mm256_permute_ps(a, imm8) _mm256_permute_ps(a, imm8)
 #endif
+#endif
 #if defined(SIMDE_X86_AVX_ENABLE_NATIVE_ALIASES)
   #undef _mm256_permute_ps
   #define _mm256_permute_ps(a, imm8) simde_mm256_permute_ps(a, imm8)
@@ -5023,9 +5036,9 @@ simde_mm256_shuffle_ps (simde__m256 a, simde__m256 b, const int imm8)
 
   return simde__m256_from_private(r_);
 }
-#if defined(SIMDE_X86_AVX_NATIVE)
+#if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_LCC_STACK_ALIGNMENT_CAP)
   #define simde_mm256_shuffle_ps(a, b, imm8) _mm256_shuffle_ps(a, b, imm8)
-#elif SIMDE_NATURAL_VECTOR_SIZE_LE(128)
+#elif SIMDE_NATURAL_VECTOR_SIZE_LE(128) || defined(SIMDE_BUG_LCC_STACK_ALIGNMENT_CAP)
   #define simde_mm256_shuffle_ps(a, b, imm8) \
       simde_mm256_set_m128( \
           simde_mm_shuffle_ps(simde_mm256_extractf128_ps(a, 1), simde_mm256_extractf128_ps(b, 1), (imm8)), \
@@ -5230,7 +5243,7 @@ simde_mm256_storeu_si256 (void* mem_addr, simde__m256i a) {
 SIMDE_FUNCTION_ATTRIBUTES
 void
 simde_mm256_storeu2_m128 (simde_float32 hi_addr[4], simde_float32 lo_addr[4], simde__m256 a) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     _mm256_storeu2_m128(hi_addr, lo_addr, a);
   #else
     simde_mm_storeu_ps(lo_addr, simde_mm256_castps256_ps128(a));
@@ -5245,7 +5258,7 @@ simde_mm256_storeu2_m128 (simde_float32 hi_addr[4], simde_float32 lo_addr[4], si
 SIMDE_FUNCTION_ATTRIBUTES
 void
 simde_mm256_storeu2_m128d (simde_float64 hi_addr[2], simde_float64 lo_addr[2], simde__m256d a) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     _mm256_storeu2_m128d(hi_addr, lo_addr, a);
   #else
     simde_mm_storeu_pd(lo_addr, simde_mm256_castpd256_pd128(a));
@@ -5260,7 +5273,7 @@ simde_mm256_storeu2_m128d (simde_float64 hi_addr[2], simde_float64 lo_addr[2], s
 SIMDE_FUNCTION_ATTRIBUTES
 void
 simde_mm256_storeu2_m128i (simde__m128i* hi_addr, simde__m128i* lo_addr, simde__m256i a) {
-  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341)
+  #if defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_91341) && !defined(SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2)
     _mm256_storeu2_m128i(hi_addr, lo_addr, a);
   #else
     simde_mm_storeu_si128(lo_addr, simde_mm256_castsi256_si128(a));