Hi! Is there a reason why trace's *-profile tables (time-profile, cpu-profile, counters-profile) always use an instruction pointer value that is 1 byte larger than what should be a true instruction pointer value?
Odd valued IPs on Apple M2 are definitely incorrect as instructions have to be word-aligned.
It's also worth mentioning that addresses in "source" tables ("time-sample" for "time-profile", "kdebug-counters-with-pmi-sample" for "cpu-profile" and "counters-profile") are correct (or, at least, are correctly aligned aligned).
Here's an example:
% xctrace version
xctrace version 15.2 (15C500b)
# run recording
% xctrace record --template "Time Profile" --output TP.trace --launch /bin/dd if=/dev/random of=/dev/null bs=1024 count=1048576
# extract "source" table
% xctrace export --input TP.trace --xpath='/trace-toc/run[1]/data/table[@schema="time-sample"]' > tp.time-sample.xml
# extract "derived" table
% xctrace export --input TP.trace --xpath='/trace-toc/run[1]/data/table[@schema="time-profile"]' > tp.time-profile.xml
% xmllint --xpath '//row[1]' tp.time-sample.xml
<row><sample-time id="1" fmt="00:00.040.502">40502000</sample-time><thread id="2" fmt="Main Thread 0x37c2d0a (dd, pid: 32471)"><tid id="3" fmt="0x37c2d0a">58469642</tid><process id="4" fmt="dd (32471)"><pid id="5" fmt="32471">32471</pid><device-session id="6" fmt="TODO">TODO</device-session></process></thread><core id="7" fmt="CPU 4 (P Core)">4</core><thread-state id="8" fmt="Running">Running</thread-state><sentinel/><kperf-bt id="9" fmt="PC:0x1863149fc, 3 frames, 1 regs, pid: 32471"><text-addresses id="10" fmt="frag 1717">6546645708 6546360308 0</text-addresses><text-address id="11" fmt="0x1863149fc">6546344444</text-address><process ref="4"/><register-content id="12" fmt="0x2e4c00018635e2cc">3336041430521340620</register-content></kperf-bt><time-sample-kind id="13" fmt="Timer Fired">0</time-sample-kind></row>
% xmllint --xpath '//row[1]' tp.time-profile.xml
<row><sample-time id="1" fmt="00:00.040.502">40502000</sample-time><thread id="2" fmt="Main Thread 0x37c2d0a (dd, pid: 32471)"><tid id="3" fmt="0x37c2d0a">58469642</tid><process id="4" fmt="dd (32471)"><pid id="5" fmt="32471">32471</pid><device-session id="6" fmt="TODO">TODO</device-session></process></thread><process ref="4"/><core id="7" fmt="CPU 4 (P Core)">4</core><thread-state id="8" fmt="Running">Running</thread-state><weight id="9" fmt="1.00 ms">1000000</weight><backtrace id="10"><frame id="11" name="0x1863149fd" addr="0x1863149fd"><binary id="12" name="dyld" UUID="324E4AD9-E01F-3183-B09F-3E20B326643A" arch="arm64e" load-addr="0x186313000" path="/usr/lib/dyld"/></frame><frame id="13" name="0x18635e2cc" addr="0x18635e2cc"><binary ref="12"/></frame><frame id="14" name="start" addr="0x1863187f4"><binary ref="12"/></frame></backtrace></row>
As you can see, <kperf-bt id="9" fmt="PC:0x1863149fc, 3 frames, 1 regs, pid: 32471"> refers to a properly aligned address, but the address from <backtrace id="10"><frame id="11" name="0x1863149fd" addr="0x1863149fd">... is off by one.
It seems like only an address from the top frame is incorrect, as others are aligned properly.
The same issue exists for "CPU Profile" and "CPU Counters" instruments and could be reproduced on macOs running on both x86-64 and Apple-Silicon CPUs.