Other Parts Discussed in Thread: SYSBIOS
TI Compiler Experts-
We're using c6678 and CGT 7.4.2 (command line tools) and we see the following issue. In the code below p1 and p2 should return the same value (pointing to a location in DDR3 mem, where the "session_data" array is located) and most of the time they do. However, intermittently p1 gives a location in an invalid memory area (for example 0x007cxxxx):
p1 = (uint8_t*)&ChanInfo_Core[n].link->term->term_id;
p2 = (uint8_t*)&session_data[n].term.term_id;
The ChanInfo_Core struct "link" pointer points to another ChanInfo_Core struct, and inside that the "term" pointer points to a struct that looks like this:
typedef struct {
uint32_t term_id;
uint32_t media_type : 8;
uint32_t codec_type : 8;
uint32_t vqe_processing_interval : 16;
uint32_t bitrate;
struct ip_addr remote_ip;
struct ip_addr local_ip;
uint32_t remote_port : 16;
uint32_t local_port : 16;
union {
struct voice_attributes voice_attr;
struct video_attributes video_attr;
} attr;
} HOST_TERM_INFO;
and session_data is an array of HOST_TERM_INFO structs. We never see p2 with a bad value. p1 and p2 are calculated in a SYSBIOS task that can be pre-empted, so we're concerned that "pointer to a pointer to a struct containing bit-fields" is generating code that is not atomic and not always reliable. We have a lot of these and similar constructs in our application.
Is there some known advice for this situation ? Avoid bit-fields ? Use padding to adjust struct element alignment ? Alignment of individual structs in the session_data array ? Fewer levels of pointer-to-a-pointer ?
We are testing this on one core, with L2 cache enabled.
Thanks in advance for your help.
-Jeff
Signalogic
PS. We saw this thread:
e2e.ti.com/.../278642
but it seems that was a case of the "pointer is always wrong", not intermittently.
