e2k: Allow natively and correctly implemented XOP and FMA functions

Author: makise-homura

simde/simde-arch.h

@@ -189,6 +189,8 @@
 #if defined(__LCC__)
 #define SIMDE_SKIP_EXTENDED_E2K_VECTOR_OPS
 #define SIMDE_BUG_LCC_WARNING_ON_SHIFTS
+#define SIMDE_BUG_LCC_XOP_MISSING
+#define SIMDE_BUG_LCC_FMA_WRONG_RESULT
 #endif
 
 /* HP/PA / PA-RISC
@@ -267,7 +269,7 @@
 #  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 +287,7 @@
 #  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

simde/x86/fma.h

@@ -156,7 +156,7 @@ simde_mm256_fmadd_ps (simde__m256 a, simde__m256 b, simde__m256 c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128d
 simde_mm_fmadd_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fmadd_sd(a, b, c);
   #else
     return simde_mm_add_sd(simde_mm_mul_sd(a, b), c);
@@ -170,7 +170,7 @@ simde_mm_fmadd_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128
 simde_mm_fmadd_ss (simde__m128 a, simde__m128 b, simde__m128 c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fmadd_ss(a, b, c);
   #else
     return simde_mm_add_ss(simde_mm_mul_ss(a, b), c);
@@ -296,7 +296,7 @@ simde_mm256_fmsub_ps (simde__m256 a, simde__m256 b, simde__m256 c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128d
 simde_mm_fmsub_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fmsub_sd(a, b, c);
   #else
     return simde_mm_sub_sd(simde_mm_mul_sd(a, b), c);
@@ -310,7 +310,7 @@ simde_mm_fmsub_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128
 simde_mm_fmsub_ss (simde__m128 a, simde__m128 b, simde__m128 c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fmsub_ss(a, b, c);
   #else
     return simde_mm_sub_ss(simde_mm_mul_ss(a, b), c);
@@ -528,7 +528,7 @@ simde_mm256_fnmadd_ps (simde__m256 a, simde__m256 b, simde__m256 c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128d
 simde_mm_fnmadd_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fnmadd_sd(a, b, c);
   #else
     simde__m128d_private
@@ -551,7 +551,7 @@ simde_mm_fnmadd_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128
 simde_mm_fnmadd_ss (simde__m128 a, simde__m128 b, simde__m128 c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fnmadd_ss(a, b, c);
   #else
     simde__m128_private
@@ -674,7 +674,7 @@ simde_mm256_fnmsub_ps (simde__m256 a, simde__m256 b, simde__m256 c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128d
 simde_mm_fnmsub_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fnmsub_sd(a, b, c);
   #else
     simde__m128d_private
@@ -697,7 +697,7 @@ simde_mm_fnmsub_sd (simde__m128d a, simde__m128d b, simde__m128d c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128
 simde_mm_fnmsub_ss (simde__m128 a, simde__m128 b, simde__m128 c) {
-  #if defined(SIMDE_X86_FMA_NATIVE)
+  #if defined(SIMDE_X86_FMA_NATIVE) && !defined(SIMDE_BUG_LCC_FMA_WRONG_RESULT)
     return _mm_fnmsub_ss(a, b, c);
   #else
     simde__m128_private

simde/x86/xop.h

@@ -78,7 +78,7 @@ simde_mm_cmov_si128 (simde__m128i a, simde__m128i b, simde__m128i c) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m256i
 simde_mm256_cmov_si256 (simde__m256i a, simde__m256i b, simde__m256i c) {
-  #if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_98521)
+  #if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_X86_AVX_NATIVE) && !defined(SIMDE_BUG_GCC_98521) && !defined(SIMDE_BUG_LCC_XOP_MISSING)
     return _mm256_cmov_si256(a, b, c);
   #elif defined(SIMDE_X86_AVX512VL_NATIVE)
     return _mm256_ternarylogic_epi32(a, b, c, 0xe4);