Add drmemtrace replay-as-traced support to record_scheduler_t #6712
Assignees
Labels
Comments
derekbruening
added a commit
that referenced
this issue
Mar 19, 2024
Adds record_scheduler_t (for the record_filter tool) support for replaying as-traced. The key change here is having record_reader_t increment the instruction ordinal on an encoding record instead of an instruction record. This avoids splitting encodings from instructions when operating at instrution boundaries in the scheduler (during replay, but this would also affect skipping for scheduler regions of interest). Adds a test of record_filter on the checked-in threadsig trace in as-traced replay mode. Fixes #6712
|
This is a little more complicated: we have to account for TRACE_MARKER_TYPE_BRANCH_TARGET as well which appears between the encoding and the instruction records. |
derekbruening
added a commit
that referenced
this issue
Mar 21, 2024
Adds record_scheduler_t (for the record_filter tool) support for replaying as-traced. The key change here is having record_reader_t's instruction ordinal, record_scheduler_t's output stream instruction ordinal, and record_scheduler_t's switch boundaries all consider the first in any sequence of encoding records or TRACE_MARKER_TYPE_BRANCH_TARGET markers to start an instruction instead of waiting for the instruction record. (Previously the scheduler switch point did consider encodings, but not branch targets, and the ordinals only considered instructions.) Moving the boundary back avoids splitting encodings from instructions when operating at instrution boundaries in the scheduler (encountered during replay, but this would also affect skipping for scheduler regions of interest). Adds tests of all 3 boundary types to the record_scheduler_t unit test. Adds a test of record_filter on the checked-in threadsig trace in as-traced replay mode. Fixes #6712
derekbruening
added a commit
that referenced
this issue
Apr 2, 2024
Adds two sanity checks developed to identify what was at first believed to be a new bug but turned out to be #6712: + Ensure that last_encoding is empty on an input switch, to avoid recording the wrong encoding in pc2encoding. + Ensure the encoding size matches the instr size when inserting chunk-new encodings. Tested by running record_filter on large proprietary inputs: no checks failed. Issue: #6712
derekbruening
added a commit
that referenced
this issue
Apr 2, 2024
Adds two sanity checks developed to identify what was at first believed to be a new bug but turned out to be #6712: + Ensure that last_encoding is empty on an input switch, to avoid recording the wrong encoding in pc2encoding. + Ensure the encoding size matches the instr size when inserting chunk-new encodings. + Augment the reader check for mismatching encoding vs instr sizes, and abort in release build too. Although #6712 is fixed, these checks are still useful to prevent regressions. Plus, the reader_t check detects other bugs as well and we know at least one other is out there in #6303. Making the reader check abort in release build should help avoid wasted work as already happened here where we didn't notice the printed warnings in release build. Tested by running record_filter on large proprietary inputs: no checks failed. Issue: #6712
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Trying to use record_filter running as-traced on a smallish app based on the threadsig sample fails:
It's the segment endpoint: have to stop before the encoding.
This would also affect skips, incl for ROI.
Better to change record_file_reader to ++ the instr count at the encoding
instead of the instr?
The text was updated successfully, but these errors were encountered: