Replace type punning with memcpy. (apache#7415)

The type punning in the existing code is undefined behaviour in C.
In particular, the existing code fails when running on Arm Cortex-M devices.
On Cortex-M, accessing a uint64_t that is not 8-byte aligned generates a hard fault.

Change-Id: I2aecaa220e581af7c91a8bc7886499d70e2aa6f2

src/runtime/crt/common/ndarray.c

@@ -68,22 +68,22 @@ int TVMNDArray_Empty(int32_t ndim, const tvm_index_t* shape, DLDataType dtype, D
 int TVMNDArray_Load(TVMNDArray* ret, const char** strm) {
   int32_t status = 0;
   uint64_t header, reserved;
-  header = ((uint64_t*)*strm)[0];  // NOLINT(*)
+  memcpy(&header, *strm, sizeof(header));
   *strm += sizeof(header);
   if (header != kTVMNDArrayMagic) {
     fprintf(stderr, "Invalid DLTensor file format\n");
     status = -1;
   }
-  reserved = ((uint64_t*)*strm)[0];  // NOLINT(*)
+  memcpy(&reserved, *strm, sizeof(reserved));
   *strm += sizeof(reserved);
   DLContext ctx;
   int ndim;  // sizeof ndim should match dlpack
   DLDataType dtype;
-  ctx = ((DLContext*)*strm)[0];  // NOLINT(*)
+  memcpy(&ctx, *strm, sizeof(ctx));
   *strm += sizeof(ctx);
-  ndim = ((int*)*strm)[0];  // NOLINT(*)
+  memcpy(&ndim, *strm, sizeof(ndim));
   *strm += sizeof(ndim);
-  dtype = ((DLDataType*)*strm)[0];  // NOLINT(*)
+  memcpy(&dtype, *strm, sizeof(dtype));
   *strm += sizeof(dtype);
   if ((ndim < 0) || (ndim > TVM_CRT_MAX_NDIM)) {
     fprintf(stderr, "Invalid ndim=%d: expected to be 0 ~ %d.\n", ndim, TVM_CRT_MAX_NDIM);
@@ -97,7 +97,7 @@ int TVMNDArray_Load(TVMNDArray* ret, const char** strm) {
   int32_t idx;
   if (ndim != 0) {
     for (idx = 0; idx < ndim; idx++) {
-      shape[idx] = ((int64_t*)*strm)[0];  // NOLINT(*)
+      memcpy(&shape[idx], *strm, sizeof(int64_t));
       *strm += sizeof(shape[idx]);
     }
   }
@@ -111,7 +111,7 @@ int TVMNDArray_Load(TVMNDArray* ret, const char** strm) {
     num_elems *= ret->dl_tensor.shape[idx];
   }
   int64_t data_byte_size;
-  data_byte_size = ((int64_t*)*strm)[0];  // NOLINT(*)
+  memcpy(&data_byte_size, *strm, sizeof(data_byte_size));
   *strm += sizeof(data_byte_size);
   if (!(data_byte_size == num_elems * elem_bytes)) {
     fprintf(stderr,

src/runtime/crt/graph_runtime/graph_runtime.c

@@ -777,13 +777,13 @@ int TVMGraphRuntime_LoadParams(TVMGraphRuntime* runtime, const char* param_blob,
   int status = 0;
   const char* bptr = param_blob;
   uint64_t header, reserved;
-  header = ((uint64_t*)bptr)[0];  // NOLINT(*)
+  memcpy(&header, bptr, sizeof(header));
   bptr += sizeof(header);
   if (header != kTVMNDArrayListMagic) {
     fprintf(stderr, "Invalid parameters file format");
     status = -1;
   }
-  reserved = ((uint64_t*)bptr)[0];  // NOLINT(*)
+  memcpy(&reserved, bptr, sizeof(reserved));
   bptr += sizeof(reserved);
 
   // read names
@@ -799,11 +799,11 @@ int TVMGraphRuntime_LoadParams(TVMGraphRuntime* runtime, const char* param_blob,
   memset(names, 0, TVM_CRT_STRLEN_NAME * runtime->nodes_count);
   uint64_t names_count;
   int idx;
-  names_count = ((uint64_t*)bptr)[0];  // NOLINT(*)
+  memcpy(&names_count, bptr, sizeof(names_count));
   bptr += sizeof(names_count);
   for (idx = 0; idx < names_count; idx++) {
     uint64_t name_length;
-    name_length = ((uint64_t*)bptr)[0];  // NOLINT(*)
+    memcpy(&name_length, bptr, sizeof(name_length));
     bptr += sizeof(name_length);
     if (name_length >= TVM_CRT_STRLEN_NAME) {
       fprintf(stderr, "Error: function name longer than expected.\n");
@@ -815,7 +815,7 @@ int TVMGraphRuntime_LoadParams(TVMGraphRuntime* runtime, const char* param_blob,
 
   // read sizes
   uint64_t sz;
-  sz = ((uint64_t*)bptr)[0];  // NOLINT(*)
+  memcpy(&sz, bptr, sizeof(sz));
   bptr += sizeof(sz);
   uint32_t size = sz;
   if (size != names_count) {