[mono][wasm] Make OP_WASM_SIMD_SWIZZLE use the shufflevector LLVM ins… (#75057)

* [mono][wasm] Make OP_WASM_SIMD_SWIZZLE use the shufflevector LLVM instruction.

The wasm.swizzle instrinsic always uses byte sized SIMD lanes, so it cannot represent
all the variants of Vector128.Shuffle ().

* Enable SIMD by default for testing.

* Remove an assert.

* Work around an LLVM problem.

* Revert "Enable SIMD by default for testing."

This reverts commit eebc533d4b90fdd32bbf5ea737bfa29243bdfbfe.

src/mono/mono/mini/llvm-intrinsics.h

@@ -262,7 +262,6 @@ INTRINS_OVR(WASM_BITMASK_V8, wasm_bitmask, Wasm, sse_i1_t)
 INTRINS_OVR(WASM_BITMASK_V4, wasm_bitmask, Wasm, sse_i4_t)
 INTRINS_OVR(WASM_BITMASK_V2, wasm_bitmask, Wasm, sse_i8_t)
 INTRINS(WASM_SHUFFLE, wasm_shuffle, Wasm)
-INTRINS(WASM_SWIZZLE, wasm_swizzle, Wasm)
 #endif
 #if defined(TARGET_ARM64)
 INTRINS_OVR(BITREVERSE_I32, bitreverse, Generic, LLVMInt32Type ())

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

@@ -7713,6 +7713,14 @@ MONO_RESTORE_WARNING
 			default:
 				element_ix = const_int32 (ins->inst_c0);
 			}
+
+#ifdef TARGET_WASM
+			// LLVM seems to return an invalid result when the input is undef (i.e. the result of an invalid shufflevector for example)
+			if (LLVMIsUndef (lhs)) {
+				values [ins->dreg] = LLVMConstNull (LLVMGetElementType (LLVMTypeOf (lhs)));
+				break;
+			}
+#endif
 			LLVMTypeRef lhs_t = LLVMTypeOf (lhs);
 			int vec_width = mono_llvm_get_prim_size_bits (lhs_t);
 			int elem_width = mono_llvm_get_prim_size_bits (elt_t);
@@ -9717,8 +9725,14 @@ MONO_RESTORE_WARNING
 			break;
 		}
 		case OP_WASM_SIMD_SWIZZLE: {
-			LLVMValueRef args [] = { lhs, rhs };
-			values [ins->dreg] = call_intrins (ctx, INTRINS_WASM_SWIZZLE, args, "");
+			int nelems = LLVMGetVectorSize (LLVMTypeOf (lhs));
+			LLVMValueRef indexes [16];
+			for (int i = 0; i < nelems; ++i)
+				indexes [i] = LLVMBuildExtractElement (builder, rhs, const_int32 (i), "");
+			LLVMValueRef shuffle_val = LLVMConstNull (LLVMVectorType (i4_t, nelems));
+			for (int i = 0; i < nelems; ++i)
+				shuffle_val = LLVMBuildInsertElement (builder, shuffle_val, convert (ctx, indexes [i], i4_t), const_int32 (i), "");
+			values [ins->dreg] = LLVMBuildShuffleVector (builder, lhs, LLVMGetUndef (LLVMTypeOf (lhs)), shuffle_val, "");
 			break;
 		}
 #endif