[LLVMdev] -time-passes output from llc
641 at rudkx.com
Sat Jul 24 00:04:51 PDT 2010
In looking at the output from -time-passes I notice that there is one set of
data for "Instruction Selection and Scheduling", and then another "Pass
execution timing report".
In looking through the code, it appears that the instruction selection &
scheduling time is already accounted for as part of the target machine
instruction selection pass that is reported in the pass execution report.
Can someone confirm that this is the case?
Is the goal in reporting the instruction selection & scheduling time
separately to keep track of the non-pass, non-target-machine-specific
timings separately to have something that is strictly a measure of the
target independent code generation phases?
Also - I've noticed these reports seem very flaky on Windows. e.g. I often
see 0 time reported for all of the instruction selection/scheduling, and of
the time reported for the passes, it often displays times that are obviously
a multiple of each other, e.g. one phase takes .045s, another .030s, and
another 0.015s. Has anybody else noticed anything like this? I also see
-track-memory reporting negative memory usage.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev