18.2. ARM trace hardware

Non-invasive debug, enables observation of the core behavior while it is executing. Although there are different kinds of non-invasive debug, this section describes trace and trace hardware in particular. It is possible to record memory accesses performed (including address and data values) and generate a real-time trace of the program, seeing peripheral accesses, stack and heap accesses and changes to variables. For many real-time systems, it is not possible to use invasive debug methods. Consider, for example, an engine management system, where although you can stop the core at a particular point, the engine keeps moving and you will not be able to do useful debug. Even in systems with less onerous real-time requirements, trace can be very useful.

Trace is typically provided by an internal hardware block connected to the core. This is known as an Embedded Trace Macrocell (ETM) and is part of most ARM processor-based systems. In some cases, there is one ETM per core. System-on-Chip designers can omit this block from their silicon to reduce costs. These blocks observe, but do not affect, core behavior and are able to monitor instruction execution and data accesses.

There are two main problems with capturing trace. The first is that with current very high core clock speeds, even a few seconds of operation can mean trillions of cycles of execution. Clearly, to make sense of this volume of information would be extremely difficult.

The second, related problem is that current cores can potentially perform one or more 64-bit cache accesses per cycle, and to record both the data address and data values can require a large bandwidth. This presents a problem in that typically, only a few pins might be provided on the chip and these outputs can be switched at significantly lower rates than the core can be clocked. If the core generates 100 bits of information every cycle at a speed of 1GHz, but the chip can only output four bits of trace at a speed of 200MHz, then there is a problem.

To solve this latter problem, the trace macrocell tries to compress information to reduce the bandwidth required. However, the main method to deal with these issues is to control the trace block so that only selected trace information is gathered. For example, you might trace only execution, without recording data values.

In addition, it is common to store trace information in an on-chip memory buffer (the Embedded Trace Buffer (ETB)). This alleviates the problem of getting information off-chip at speed, but has an additional cost in terms of silicon area (and therefore price of the chip) and also provides a fixed limit on the amount of trace that can be captured.

The ETB stores the compressed trace information in a circular fashion, continuously capturing trace information until stopped. The size of the ETB varies between chip implementations, but a buffer of 8 or 16KB is typically enough to hold a few thousand lines of program trace. When a program fails, if the trace buffer is enabled, you can see a portion of program history. With this program history, it is easier to walk back through your program to see what happened before the point of failure. This is particularly useful for investigating intermittent and real-time failures that can be difficult to identify through traditional debug methods that require stopping and starting the core. The use of hardware tracing can significantly reduce the amount of time required to find these failures, as the trace shows exactly what was executed, what the timing was and what data accesses occurred.

Copyright © 2015 ARM. All rights reserved.ARM DEN0024A