Decoding trace in reverse execution order about libipt

[markus-metzger]

You need to synchronize on a PSB packet. Looks like you'd want pt_<lyr>_sync_backward(). You can use that repeatedly to decode PSB segments from back to front. You'd need to detect when you cross to a PSB segment you already decoded or split the trace to avoid that. You should get status update events when decoding a new PSB+.

[loizos01]

Thank you, I will try that on Monday and close this. Enjoy your weekend!

[loizos01]

Hi,

After the first pt_<lyr>_sync_backward() call , the following calls return pte_eos error. Events drain without error.

Offset: 0000000000000034 pte_eos
Error: synchronising decoder1
Offset: fffffffffffffffe pte_eos
Error: synchronising decoder1
Offset: fffffffffffffffe pte_eos
Error: synchronising decoder1

[markus-metzger]

pte_eos means that you reached the beginning of the trace buffer. How many PSB packets are there in the trace?

[loizos01]

There is only one PSB packet, for some reason I thought each PSB packet corresponded to an executed instruction and by calling pt_<lyr>_sync_backward() would sync onto the next instruction in a backwards direction. What I am trying to optimize is, for instance, now, I sync into the decoder from the beginning and using pt_insn_next I decode the trace successfully until the end. But, I am only interested in the last lets say, 100 executed instructions(out of 20k).

[markus-metzger]

The trace cannot be decoded backwards. You may configure PSB to be as frequent as possible to reduce the amount of trace that needs to be decoded. You may also consider using the block decoder, which is typically faster, but requires decoding the instructions in a block separately.

[loizos01]

Alright, thank you