[wasm] Add Vector128 and PackedSimd support to the jiterpreter; add PackedSimd to the interpreter (#82773)

* Add PackedSIMD support to the interpreter (off by default)
* Add SIMD support to the jiterpreter
* Add runtime options governing interpreter vector128 and packedsimd support
* Add some R4 vector128 operations to the interpreter
* Fix jiterpreter MINT_POPCNT_I8 implementation
* Enable compiling the runtime with wasm simd support so that intrinsics can be used
* Add browser-bench measurements for packing vector128

Author: kg

src/mono/CMakeLists.txt

@@ -270,6 +270,7 @@ elseif(CLR_CMAKE_HOST_OS STREQUAL "emscripten")
   add_compile_options(-Wno-strict-prototypes)
   add_compile_options(-Wno-unused-but-set-variable)
   add_compile_options(-Wno-single-bit-bitfield-constant-conversion)
+  add_compile_options(-msimd128)
   set(DISABLE_EXECUTABLES 1)
   # FIXME: Is there a cmake option for this ?
   set(DISABLE_SHARED_LIBS 1)

src/mono/mono/mini/interp/interp-internals.h

@@ -102,7 +102,7 @@ typedef enum {
 
 #define PROFILE_INTERP 0
 
-#if !HOST_BROWSER && __GNUC__
+#if __GNUC__
 #define INTERP_ENABLE_SIMD
 #endif
 
@@ -342,6 +342,12 @@ mono_jiterp_stackval_from_data (MonoType *type, stackval *result, const void *da
 gpointer
 mono_jiterp_frame_data_allocator_alloc (FrameDataAllocator *stack, InterpFrame *frame, int size);
 
+gpointer
+mono_jiterp_get_simd_intrinsic (int arity, int index);
+
+int
+mono_jiterp_get_simd_opcode (int arity, int index);
+
 #endif
 
 static inline int

src/mono/mono/mini/interp/interp-simd-intrins.def

@@ -2,6 +2,10 @@
 #include "interp-internals.h"
 #include "interp-simd.h"
 
+#if HOST_BROWSER
+#include <wasm_simd128.h>
+#endif
+
 #ifdef INTERP_ENABLE_SIMD
 
 typedef gint64 v128_i8 __attribute__ ((vector_size (SIZEOF_V128)));
@@ -12,6 +16,7 @@ typedef gint16 v128_i2 __attribute__ ((vector_size (SIZEOF_V128)));
 typedef guint16 v128_u2 __attribute__ ((vector_size (SIZEOF_V128)));
 typedef gint8 v128_i1 __attribute__ ((vector_size (SIZEOF_V128)));
 typedef guint8 v128_u1 __attribute__ ((vector_size (SIZEOF_V128)));
+typedef float v128_r4 __attribute__ ((vector_size (SIZEOF_V128)));
 
 // get_AllBitsSet
 static void
@@ -39,6 +44,12 @@ interp_v128_i4_op_addition (gpointer res, gpointer v1, gpointer v2)
 	*(v128_i4*)res = *(v128_i4*)v1 + *(v128_i4*)v2;
 }
 
+static void
+interp_v128_r4_op_addition (gpointer res, gpointer v1, gpointer v2)
+{
+	*(v128_r4*)res = *(v128_r4*)v1 + *(v128_r4*)v2;
+}
+
 // op_Subtraction
 static void
 interp_v128_i1_op_subtraction (gpointer res, gpointer v1, gpointer v2)
@@ -58,6 +69,12 @@ interp_v128_i4_op_subtraction (gpointer res, gpointer v1, gpointer v2)
 	*(v128_i4*)res = *(v128_i4*)v1 - *(v128_i4*)v2;
 }
 
+static void
+interp_v128_r4_op_subtraction (gpointer res, gpointer v1, gpointer v2)
+{
+	*(v128_r4*)res = *(v128_r4*)v1 - *(v128_r4*)v2;
+}
+
 // op_BitwiseAnd
 static void
 interp_v128_op_bitwise_and (gpointer res, gpointer v1, gpointer v2)
@@ -124,6 +141,18 @@ interp_v128_i4_op_multiply (gpointer res, gpointer v1, gpointer v2)
 	*(v128_i4*)res = *(v128_i4*)v1 * *(v128_i4*)v2;
 }
 
+static void
+interp_v128_r4_op_multiply (gpointer res, gpointer v1, gpointer v2)
+{
+	*(v128_r4*)res = *(v128_r4*)v1 * *(v128_r4*)v2;
+}
+
+static void
+interp_v128_r4_op_division (gpointer res, gpointer v1, gpointer v2)
+{
+	*(v128_r4*)res = *(v128_r4*)v1 / *(v128_r4*)v2;
+}
+
 // op_UnaryNegation
 static void
 interp_v128_i1_op_negation (gpointer res, gpointer v1)
@@ -535,32 +564,122 @@ interp_v128_i8_shuffle (gpointer res, gpointer v1, gpointer v2)
 	V128_SHUFFLE (gint64, guint64);
 }
 
-#define INTERP_SIMD_INTRINSIC_P_P(a,b)
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b)
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c)
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c)
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c)
+
+// For the wasm packed simd intrinsics we want to automatically generate the C implementations from
+//  their corresponding clang intrinsics. See also:
+// https://github.com/llvm/llvm-project/blob/main/clang/lib/Headers/wasm_simd128.h
+// In this context V means Vector128 and P means void* pointer.
+#ifdef HOST_BROWSER
+
+static v128_t
+_interp_wasm_simd_assert_not_reached (v128_t lhs, v128_t rhs) {
+	g_assert_not_reached ();
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_V_P(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1) { \
+	*((v128_t *)res) = c_intrinsic (v1); \
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_V_V(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1) { \
+	*((v128_t *)res) = c_intrinsic (*((v128_t *)v1)); \
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_I_V(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1) { \
+	*((int32_t *)res) = c_intrinsic (*((v128_t *)v1)); \
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_V_VV(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1, gpointer v2) { \
+	*((v128_t *)res) = c_intrinsic (*((v128_t *)v1), *((v128_t *)v2)); \
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_V_VI(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1, gpointer v2) { \
+	*((v128_t *)res) = c_intrinsic (*((v128_t *)v1), *((int *)v2)); \
+}
+
+#define INTERP_WASM_SIMD_INTRINSIC_V_VVV(id, c_intrinsic, wasm_opcode) \
+static void \
+_mono_interp_simd_ ## id (gpointer res, gpointer v1, gpointer v2, gpointer v3) { \
+	*((v128_t *)res) = c_intrinsic (*((v128_t *)v1), *((v128_t *)v2), *((v128_t *)v3)); \
+}
+
+#include "interp-simd-intrins.def"
+
+#undef INTERP_WASM_SIMD_INTRINSIC_V_P
+#undef INTERP_WASM_SIMD_INTRINSIC_V_V
+#undef INTERP_WASM_SIMD_INTRINSIC_I_V
+#undef INTERP_WASM_SIMD_INTRINSIC_V_VV
+#undef INTERP_WASM_SIMD_INTRINSIC_V_VI
+#undef INTERP_WASM_SIMD_INTRINSIC_V_VVV
+
+// Now generate the wasm opcode tables for the intrinsics
+
+#undef INTERP_SIMD_INTRINSIC_P_P
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c) c,
+
+int interp_simd_p_p_wasm_opcode_table [] = {
+#include "interp-simd-intrins.def"
+};
+
+#undef INTERP_SIMD_INTRINSIC_P_P
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c)
+
+#undef INTERP_SIMD_INTRINSIC_P_PP
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c) c,
+
+int interp_simd_p_pp_wasm_opcode_table [] = {
+#include "interp-simd-intrins.def"
+};
+
+#undef INTERP_SIMD_INTRINSIC_P_PP
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c)
+
+#undef INTERP_SIMD_INTRINSIC_P_PPP
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c) c,
+
+int interp_simd_p_ppp_wasm_opcode_table [] = {
+#include "interp-simd-intrins.def"
+};
+
+#undef INTERP_SIMD_INTRINSIC_P_PPP
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c)
+
+#endif // HOST_BROWSER
 
 #undef INTERP_SIMD_INTRINSIC_P_P
-#define INTERP_SIMD_INTRINSIC_P_P(a,b) b,
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c) b,
 PP_SIMD_Method interp_simd_p_p_table [] = {
 #include "interp-simd-intrins.def"
 };
 #undef INTERP_SIMD_INTRINSIC_P_P
-#define INTERP_SIMD_INTRINSIC_P_P(a,b)
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c)
 
 #undef INTERP_SIMD_INTRINSIC_P_PP
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b) b,
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c) b,
 PPP_SIMD_Method interp_simd_p_pp_table [] = {
 #include "interp-simd-intrins.def"
 };
 #undef INTERP_SIMD_INTRINSIC_P_PP
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c)
 
 #undef INTERP_SIMD_INTRINSIC_P_PPP
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b) b,
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c) b,
 PPPP_SIMD_Method interp_simd_p_ppp_table [] = {
 #include "interp-simd-intrins.def"
 };
 #undef INTERP_SIMD_INTRINSIC_P_PPP
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c)
 
 #endif // INTERP_ENABLE_SIMD

src/mono/mono/mini/interp/interp-simd.c

@@ -11,6 +11,12 @@ extern PP_SIMD_Method interp_simd_p_p_table [];
 extern PPP_SIMD_Method interp_simd_p_pp_table [];
 extern PPPP_SIMD_Method interp_simd_p_ppp_table [];
 
+#if HOST_BROWSER
+extern int interp_simd_p_p_wasm_opcode_table [];
+extern int interp_simd_p_pp_wasm_opcode_table [];
+extern int interp_simd_p_ppp_wasm_opcode_table [];
+#endif
+
 #endif /* __MONO_MINI_INTERP_SIMD_H__ */
 
 

src/mono/mono/mini/interp/interp-simd.h

@@ -8907,4 +8907,42 @@ mono_jiterp_enum_hasflag (MonoClass *klass, gint32 *dest, stackval *sp1, stackva
 	*dest = mono_interp_enum_hasflag (sp1, sp2, klass);
 }
 
+EMSCRIPTEN_KEEPALIVE gpointer
+mono_jiterp_get_simd_intrinsic (int arity, int index)
+{
+#ifdef INTERP_ENABLE_SIMD
+	switch (arity) {
+		case 1:
+			return interp_simd_p_p_table [index];
+		case 2:
+			return interp_simd_p_pp_table [index];
+		case 3:
+			return interp_simd_p_ppp_table [index];
+		default:
+			g_assert_not_reached();
+	}
+#else
+	g_assert_not_reached();
+#endif
+}
+
+EMSCRIPTEN_KEEPALIVE int
+mono_jiterp_get_simd_opcode (int arity, int index)
+{
+#ifdef INTERP_ENABLE_SIMD
+	switch (arity) {
+		case 1:
+			return interp_simd_p_p_wasm_opcode_table [index];
+		case 2:
+			return interp_simd_p_pp_wasm_opcode_table [index];
+		case 3:
+			return interp_simd_p_ppp_wasm_opcode_table [index];
+		default:
+			g_assert_not_reached();
+	}
+#else
+	g_assert_not_reached();
+#endif
+}
+
 #endif

src/mono/mono/mini/interp/interp.c

@@ -41,35 +41,35 @@ typedef enum {
 
 /* SIMD opcodes, grouped by signature */
 
-#define INTERP_SIMD_INTRINSIC_P_P(a,b)
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b)
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c)
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c)
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c)
 
 #undef INTERP_SIMD_INTRINSIC_P_P
-#define INTERP_SIMD_INTRINSIC_P_P(a,b) a,
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c) a,
 typedef enum {
 #include "interp-simd-intrins.def"
 } MintSIMDOpsPP;
 #undef INTERP_SIMD_INTRINSIC_P_P
-#define INTERP_SIMD_INTRINSIC_P_P(a,b)
+#define INTERP_SIMD_INTRINSIC_P_P(a,b,c)
 
 #undef INTERP_SIMD_INTRINSIC_P_PP
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b) a,
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c) a,
 typedef enum {
 #include "interp-simd-intrins.def"
 	INTERP_SIMD_INTRINSIC_P_PP_LAST
 } MintSIMDOpsPPP;
 #undef INTERP_SIMD_INTRINSIC_P_PP
-#define INTERP_SIMD_INTRINSIC_P_PP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_PP(a,b,c)
 
 #undef INTERP_SIMD_INTRINSIC_P_PPP
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b) a,
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c) a,
 typedef enum {
 #include "interp-simd-intrins.def"
 	INTERP_SIMD_INTRINSIC_P_PPP_LAST
 } MintSIMDOpsPPPP;
 #undef INTERP_SIMD_INTRINSIC_P_PPP
-#define INTERP_SIMD_INTRINSIC_P_PPP(a,b)
+#define INTERP_SIMD_INTRINSIC_P_PPP(a,b,c)
 
 #if NO_UNALIGNED_ACCESS
 #  if G_BYTE_ORDER == G_LITTLE_ENDIAN

src/mono/mono/mini/interp/mintops.h

@@ -1,12 +1,14 @@
 SIMD_METHOD(get_Count)
 SIMD_METHOD(get_AllBitsSet)
 SIMD_METHOD(get_IsHardwareAccelerated)
+SIMD_METHOD(get_IsSupported)
 SIMD_METHOD(get_Item)
 SIMD_METHOD(get_One)
 SIMD_METHOD(get_Zero)
 SIMD_METHOD(op_Addition)
 SIMD_METHOD(op_BitwiseAnd)
 SIMD_METHOD(op_BitwiseOr)
+SIMD_METHOD(op_Division)
 SIMD_METHOD(op_Equality)
 SIMD_METHOD(op_ExclusiveOr)
 SIMD_METHOD(op_Explicit)
@@ -24,6 +26,7 @@ SIMD_METHOD(ConditionalSelect)
 SIMD_METHOD(Create)
 SIMD_METHOD(CreateScalar)
 SIMD_METHOD(CreateScalarUnsafe)
+
 SIMD_METHOD(Equals)
 SIMD_METHOD(ExtractMostSignificantBits)
 SIMD_METHOD(GreaterThan)
@@ -36,3 +39,20 @@ SIMD_METHOD(ShiftRightLogical)
 SIMD_METHOD(Shuffle)
 SIMD_METHOD(WidenLower)
 SIMD_METHOD(WidenUpper)
+
+// PackedSimd
+SIMD_METHOD(Splat)
+SIMD_METHOD(ExtractLane)
+SIMD_METHOD(ReplaceLane)
+SIMD_METHOD(Swizzle)
+SIMD_METHOD(Add)
+SIMD_METHOD(Subtract)
+SIMD_METHOD(Multiply)
+SIMD_METHOD(Dot)
+SIMD_METHOD(Negate)
+SIMD_METHOD(And)
+SIMD_METHOD(Bitmask)
+SIMD_METHOD(CompareEqual)
+SIMD_METHOD(CompareNotEqual)
+SIMD_METHOD(ConvertNarrowingSignedSaturate)
+SIMD_METHOD(ConvertNarrowingUnsignedSaturate)