e2k-lcc: Deal with deprecated inefficient functions of lcc

simd-everywhere · Feb 20, 2021 · 9ff51ed · 9ff51ed
1 parent 7fa7150
commit 9ff51ed
Show file tree

Hide file tree

Showing 4 changed files with 91 additions and 14 deletions.
diff --git a/simde/simde-arch.h b/simde/simde-arch.h
@@ -191,6 +191,14 @@
 #define SIMDE_BUG_LCC_XOP_MISSING
 #define SIMDE_BUG_LCC_FMA_WRONG_RESULT
 #define SIMDE_BUG_LCC_AVX_NO_LOAD_STORE_U2
+
+/* 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")
 #endif
 
 /* HP/PA / PA-RISC
@@ -317,7 +325,7 @@
 #  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/x86/clmul.h b/simde/x86/clmul.h
@@ -203,8 +203,21 @@ simde_mm_clmulepi64_si128 (simde__m128i a, simde__m128i b, const int imm8)
 
   return simde__m128i_from_private(r_);
 }
+
+#if defined(SIMDE_X86_PCLMUL_NATIVE) && defined(SIMDE_BUG_PCLMUL_XOP_DEPRECATED)
+  SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+  SIMDE_FUNCTION_ATTRIBUTES
+  simde__m128i
+  simde_undeprecated_mm_clmulepi64_si128 (simde__m128i a, simde__m128i b, const int imm8) {
+    return _mm_clmulepi64_si128(a, b, imm8);
+  }
+  SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
+#else
+  #define simde_undeprecated_mm_clmulepi64_si128 _mm_clmulepi64_si128
+#endif
+
 #if defined(SIMDE_X86_PCLMUL_NATIVE)
-  #define simde_mm_clmulepi64_si128(a, b, imm8) _mm_clmulepi64_si128(a, b, imm8)
+  #define simde_mm_clmulepi64_si128(a, b, imm8) simde_undeprecated_mm_clmulepi64_si128(a, b, imm8)
 #elif defined(SIMDE_ARM_NEON_A64V8_NATIVE) && defined(__ARM_FEATURE_AES)
   #define simde_mm_clmulepi64_si128(a, b, imm8) \
     simde__m128i_from_neon_u64( \
@@ -233,20 +246,20 @@ simde_mm256_clmulepi64_epi128 (simde__m256i a, simde__m256i b, const int imm8)
   #if defined(SIMDE_X86_PCLMUL_NATIVE)
     switch (imm8 & 0x11) {
       case 0x00:
-        r_.m128i[0] = _mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x00);
-        r_.m128i[1] = _mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x00);
+        r_.m128i[0] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x00);
+        r_.m128i[1] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x00);
         break;
       case 0x01:
-        r_.m128i[0] = _mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x01);
-        r_.m128i[1] = _mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x01);
+        r_.m128i[0] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x01);
+        r_.m128i[1] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x01);
         break;
       case 0x10:
-        r_.m128i[0] = _mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x10);
-        r_.m128i[1] = _mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x10);
+        r_.m128i[0] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x10);
+        r_.m128i[1] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x10);
         break;
       case 0x11:
-        r_.m128i[0] = _mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x11);
-        r_.m128i[1] = _mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x11);
+        r_.m128i[0] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[0], b_.m128i[0], 0x11);
+        r_.m128i[1] = simde_undeprecated_mm_clmulepi64_si128(a_.m128i[1], b_.m128i[1], 0x11);
         break;
     }
   #else

diff --git a/simde/x86/f16c.h b/simde/x86/f16c.h
@@ -45,12 +45,14 @@ SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_cvtps_ph(simde__m128 a, const int sae) {
   #if defined(SIMDE_X86_F16C_NATIVE)
+    SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
     switch (sae & SIMDE_MM_FROUND_NO_EXC) {
       case SIMDE_MM_FROUND_NO_EXC:
         return _mm_cvtps_ph(a, SIMDE_MM_FROUND_NO_EXC);
       default:
         return _mm_cvtps_ph(a, 0);
     }
+    SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
   #else
     simde__m128_private a_ = simde__m128_to_private(a);
     simde__m128i_private r_ = simde__m128i_to_private(simde_mm_setzero_si128());
@@ -102,12 +104,14 @@ SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm256_cvtps_ph(simde__m256 a, const int sae) {
   #if defined(SIMDE_X86_F16C_NATIVE) && defined(SIMDE_X86_AVX_NATIVE)
+    SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
     switch (sae & SIMDE_MM_FROUND_NO_EXC) {
       case SIMDE_MM_FROUND_NO_EXC:
         return _mm256_cvtps_ph(a, SIMDE_MM_FROUND_NO_EXC);
       default:
         return _mm256_cvtps_ph(a, 0);
     }
+    SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
   #else
     simde__m256_private a_ = simde__m256_to_private(a);
     simde__m128i_private r_;

diff --git a/simde/x86/xop.h b/simde/x86/xop.h
@@ -3511,8 +3511,21 @@ simde_mm_permute2_ps (simde__m128 a, simde__m128 b, simde__m128i c, const int im
 
   return simde__m128_from_private(r_);
 }
+
+#if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_BUG_PCLMUL_XOP_DEPRECATED)
+  SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+  SIMDE_FUNCTION_ATTRIBUTES
+  simde__m128
+  simde_undeprecated_mm_permute2_ps (simde__m128 a, simde__m128 b, simde__m128i c, const int imm8) {
+    return _mm_permute2_ps(a, b, c, imm8);
+  }
+  SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
+#else
+  #define simde_undeprecated_mm_permute2_ps _mm_permute2_ps
+#endif
+
 #if defined(SIMDE_X86_XOP_NATIVE)
-  #define simde_mm_permute2_ps(a, b, c, imm8) _mm_permute2_ps((a), (b), (c), (imm8))
+  #define simde_mm_permute2_ps(a, b, c, imm8) simde_undeprecated_mm_permute2_ps((a), (b), (c), (imm8))
 #endif
 #if defined(SIMDE_X86_XOP_ENABLE_NATIVE_ALIASES)
   #define _mm_permute2_ps(a, b, c, imm8) simde_mm_permute2_ps((a), (b), (c), (imm8))
@@ -3547,8 +3560,21 @@ simde_mm_permute2_pd (simde__m128d a, simde__m128d b, simde__m128i c, const int
 
   return simde__m128d_from_private(r_);
 }
+
+#if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_BUG_PCLMUL_XOP_DEPRECATED)
+  SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+  SIMDE_FUNCTION_ATTRIBUTES
+  simde__m128d
+  simde_undeprecated_mm_permute2_pd (simde__m128d a, simde__m128d b, simde__m128i c, const int imm8) {
+    return _mm_permute2_pd(a, b, c, imm8);
+  }
+  SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
+#else
+  #define simde_undeprecated_mm_permute2_pd _mm_permute2_pd
+#endif
+
 #if defined(SIMDE_X86_XOP_NATIVE)
-  #define simde_mm_permute2_pd(a, b, c, imm8) _mm_permute2_pd((a), (b), (c), (imm8))
+  #define simde_mm_permute2_pd(a, b, c, imm8) simde_undeprecated_mm_permute2_pd((a), (b), (c), (imm8))
 #endif
 #if defined(SIMDE_X86_XOP_ENABLE_NATIVE_ALIASES)
   #define _mm_permute2_pd(a, b, c, imm8) simde_mm_permute2_pd((a), (b), (c), (imm8))
@@ -3589,8 +3615,21 @@ simde_mm256_permute2_ps (simde__m256 a, simde__m256 b, simde__m256i c, const int
 
   return simde__m256_from_private(r_);
 }
+
+#if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_BUG_PCLMUL_XOP_DEPRECATED)
+  SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+  SIMDE_FUNCTION_ATTRIBUTES
+  simde__m256
+  simde_undeprecated_mm256_permute2_ps (simde__m256 a, simde__m256 b, simde__m256i c, const int imm8) {
+    return _mm256_permute2_ps(a, b, c, imm8);
+  }
+  SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
+#else
+  #define simde_undeprecated_mm256_permute2_ps _mm256_permute2_ps
+#endif
+
 #if defined(SIMDE_X86_XOP_NATIVE)
-  #define simde_mm256_permute2_ps(a, b, c, imm8) _mm256_permute2_ps((a), (b), (c), (imm8))
+  #define simde_mm256_permute2_ps(a, b, c, imm8) simde_undeprecated_mm256_permute2_ps((a), (b), (c), (imm8))
 #endif
 #if defined(SIMDE_X86_XOP_ENABLE_NATIVE_ALIASES)
   #define _mm256_permute2_ps(a, b, c, imm8) simde_mm256_permute2_ps((a), (b), (c), (imm8))
@@ -3631,8 +3670,21 @@ simde_mm256_permute2_pd (simde__m256d a, simde__m256d b, simde__m256i c, const i
 
   return simde__m256d_from_private(r_);
 }
+
+#if defined(SIMDE_X86_XOP_NATIVE) && defined(SIMDE_BUG_PCLMUL_XOP_DEPRECATED)
+  SIMDE_LCC_DISABLE_DEPRECATED_WARNINGS
+  SIMDE_FUNCTION_ATTRIBUTES
+  simde__m256d
+  simde_undeprecated_mm256_permute2_pd (simde__m256d a, simde__m256d b, simde__m256i c, const int imm8) {
+    return _mm256_permute2_pd(a, b, c, imm8);
+  }
+  SIMDE_LCC_REVERT_DEPRECATED_WARNINGS
+#else
+  #define simde_undeprecated_mm256_permute2_pd _mm256_permute2_pd
+#endif
+
 #if defined(SIMDE_X86_XOP_NATIVE)
-  #define simde_mm256_permute2_pd(a, b, c, imm8) _mm256_permute2_pd((a), (b), (c), (imm8))
+  #define simde_mm256_permute2_pd(a, b, c, imm8) simde_undeprecated_mm256_permute2_pd((a), (b), (c), (imm8))
 #endif
 #if defined(SIMDE_X86_XOP_ENABLE_NATIVE_ALIASES)
   #define _mm256_permute2_pd(a, b, c, imm8) simde_mm256_permute2_pd((a), (b), (c), (imm8))