[mono][jit] Adding Vector128.Dot as intrinsic for arm64. (#84245)

* [mono][jit] Adding Vector128.Dot as intrinsic for arm64. * Fixed the (u)int64 case, which now falls back to unoptimized code. * Fixed klass. * Disabled the intrinsic for native ints, enabled for LLVM. * Fixed COMPILE_LLVM condition. * Relaxing non-LLVM restriction for WASM.
dotnet · Apr 17, 2023 · e3f55cb · e3f55cb
1 parent d4f70b1
commit e3f55cb
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 4 deletions.
diff --git a/src/mono/mono/mini/simd-arm64.h b/src/mono/mono/mini/simd-arm64.h
@@ -65,7 +65,7 @@ SIMD_OP  (128, OP_XBINOP,      OP_FMAX,              WTDSS,              _UNDEF,
 SIMD_OP  (128, OP_XBINOP,      OP_IMIN,              WTDSS,              arm_neon_smin,    arm_neon_smin,    arm_neon_smin,   _SKIP,             _UNDEF,           _UNDEF)
 SIMD_OP  (128, OP_XBINOP,      OP_IMIN_UN,           WTDSS,              arm_neon_umin,    arm_neon_umin,    arm_neon_umin,   _SKIP,             _UNDEF,           _UNDEF)
 SIMD_OP  (128, OP_XBINOP,      OP_FMIN,              WTDSS,              _UNDEF,           _UNDEF,           _UNDEF,          _UNDEF,            arm_neon_fmin,    arm_neon_fmin)
-SIMD_OP  (128, OP_XBINOP,      OP_IMUL,              WTDSS,              arm_neon_mul,     arm_neon_mul,     arm_neon_mul,    arm_neon_mul,      _UNDEF,           _UNDEF)
+SIMD_OP  (128, OP_XBINOP,      OP_IMUL,              WTDSS,              arm_neon_mul,     arm_neon_mul,     arm_neon_mul,    _UNDEF,            _UNDEF,           _UNDEF)
 SIMD_OP  (128, OP_XBINOP,      OP_FMUL,              WTDSS,              _UNDEF,           _UNDEF,           _UNDEF,          _UNDEF,            arm_neon_fmul,    arm_neon_fmul)
 SIMD_OP  (128, OP_XBINOP,      OP_FDIV,              WTDSS,              _UNDEF,           _UNDEF,           _UNDEF,          _UNDEF,            arm_neon_fdiv,    arm_neon_fdiv)
 SIMD_OP  (128, OP_XBINOP_FORCEINT,    XBINOP_FORCEINT_AND,    WDSS,      arm_neon_and,     arm_neon_and,     arm_neon_and,    arm_neon_and,      arm_neon_and,     arm_neon_and)

diff --git a/src/mono/mono/mini/simd-intrinsics.c b/src/mono/mono/mini/simd-intrinsics.c
@@ -1324,7 +1324,6 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 		case SN_ConvertToUInt32:
 		case SN_ConvertToUInt64:
 		case SN_Create:
-		case SN_Dot:
 		case SN_GetElement:
 		case SN_GetLower:
 		case SN_GetUpper:
@@ -1339,7 +1338,7 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 	}
 #endif
 
-	MonoClass *klass = cmethod->klass;
+	MonoClass* klass = fsig->param_count > 0 ? args[0]->klass : cmethod->klass;
 	MonoTypeEnum arg0_type = fsig->param_count > 0 ? get_underlying_type (fsig->params [0]) : MONO_TYPE_VOID;
 
 	if (cfg->verbose_level > 1) {
@@ -1554,10 +1553,17 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 	case SN_Dot: {
 		if (!is_element_type_primitive (fsig->params [0]))
 			return NULL;
+#if defined(TARGET_WASM)
+		if (!COMPILE_LLVM (cfg) && (arg0_type == MONO_TYPE_I8 || arg0_type == MONO_TYPE_U8))
+			return NULL;
+#elif defined(TARGET_ARM64)
+		if (!COMPILE_LLVM (cfg) && (arg0_type == MONO_TYPE_I8 || arg0_type == MONO_TYPE_U8 || arg0_type == MONO_TYPE_I || arg0_type == MONO_TYPE_U))
+			return NULL;
+#endif
+
 #if defined(TARGET_ARM64) || defined(TARGET_WASM)
 		int instc0 = type_enum_is_float (arg0_type) ? OP_FMUL : OP_IMUL;
 		MonoInst *pairwise_multiply = emit_simd_ins_for_sig (cfg, klass, OP_XBINOP, instc0, arg0_type, fsig, args);
-
 		return emit_sum_vector (cfg, fsig->params [0], arg0_type, pairwise_multiply);
 #elif defined(TARGET_AMD64)
 		int instc =-1;