Fix vector's Narrow intrinsics (#81843)

The BCL Vector classes have non-saturating Narrow methods, while wasm instructions are saturating. AFAIK wasm does not have non-saturating narrow instructions. So instead of

    i8x16.narrow_i16x8_s
    i8x16.narrow_i16x8_u
    i16x8.narrow_i32x4_s
    i16x8.narrow_i32x4_u

use `v8x16.shuffle` instruction to implement the extract narrow operation.

This fixes `System.Numerics.Tests.GenericVectorTests.Narrow[U]Int*` tests.

Author: radekdoulik

src/mono/mono/mini/mini-llvm.c

@@ -9833,6 +9833,37 @@ MONO_RESTORE_WARNING
 			values [ins->dreg] = LLVMBuildShuffleVector (builder, lhs, LLVMGetUndef (LLVMTypeOf (lhs)), shuffle_val, "");
 			break;
 		}
+		case OP_WASM_EXTRACT_NARROW: {
+			int nelems = LLVMGetVectorSize (LLVMTypeOf (lhs));
+			int bytes = 16 / (nelems * 2);
+			LLVMTypeRef itype;
+
+			switch(nelems) {
+				case 2:
+					itype = i4_t;
+					break;
+				case 4:
+					itype = i2_t;
+					break;
+				case 8:
+					itype = i1_t;
+					break;
+				default:
+					g_assert_not_reached();
+			}
+
+			LLVMValueRef mask = LLVMConstNull (LLVMVectorType (i1_t, 16));
+			for (int i = 0; i < nelems; ++i) {
+				for (int j = 0; j < bytes; ++j) {
+					mask = LLVMBuildInsertElement (builder, mask, const_int8 (i * bytes * 2 + j), const_int32 (i * bytes + j), "");
+					mask = LLVMBuildInsertElement (builder, mask, const_int8 (16 + i * bytes * 2 + j), const_int32 (8 + i * bytes + j), "");
+				}
+			}
+
+			LLVMValueRef shuffle = LLVMBuildShuffleVector (builder, LLVMBuildBitCast (builder, lhs, LLVMVectorType (i1_t, 16), ""), LLVMBuildBitCast (builder, rhs, LLVMVectorType (i1_t, 16), ""), mask, "");
+			values [ins->dreg] = LLVMBuildBitCast (builder, shuffle, LLVMVectorType (itype, nelems * 2), "");
+			break;
+		}
 #endif
 #if defined(TARGET_ARM64) || defined(TARGET_X86) || defined(TARGET_AMD64) || defined(TARGET_WASM)
 		case OP_XEQUAL: {

src/mono/mono/mini/mini-ops.h

@@ -853,6 +853,7 @@ MINI_OP(OP_WASM_SIMD_BITMASK, "wasm_bitmask", IREG, XREG, NONE)
 MINI_OP3(OP_WASM_SIMD_SHUFFLE, "wasm_shuffle", XREG, XREG, XREG, XREG)
 MINI_OP(OP_WASM_SIMD_SUM, "wasm_sum", XREG, XREG, NONE)
 MINI_OP(OP_WASM_SIMD_SWIZZLE, "wasm_swizzle", XREG, XREG, XREG)
+MINI_OP(OP_WASM_EXTRACT_NARROW, "wasm_extract_narrow", XREG, XREG, XREG)
 #endif
 
 #if defined(TARGET_X86) || defined(TARGET_AMD64) || defined(TARGET_WASM)

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

@@ -1621,25 +1621,16 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
 		if (size != 16)
 			return NULL;
 
-		int intrins = -1;
 		switch (arg0_type) {
 		case MONO_TYPE_I2:
-			intrins = INTRINS_WASM_NARROW_SIGNED_V16;
-			break;
 		case MONO_TYPE_I4:
-			intrins = INTRINS_WASM_NARROW_SIGNED_V8;
-			break;
+		case MONO_TYPE_I8:
 		case MONO_TYPE_U2:
-			intrins = INTRINS_WASM_NARROW_UNSIGNED_V16;
-			break;
 		case MONO_TYPE_U4:
-			intrins = INTRINS_WASM_NARROW_UNSIGNED_V8;
-			break;
+		case MONO_TYPE_U8:
+			return emit_simd_ins_for_sig (cfg, klass, OP_WASM_EXTRACT_NARROW, -1, -1, fsig, args);
 		}
 
-		if (intrins != -1)
-			return emit_simd_ins_for_sig (cfg, klass, OP_XOP_X_X_X, intrins, arg0_type, fsig, args);
-
 		return NULL;
 #else
 		return NULL;