gem5: possible to edit python in tree, factor arm and aarch64 build

readme: stress kills gem5, some gem5 builds much slower than others
This commit is contained in:
Ciro Santilli
2018-05-01 23:26:03 +01:00
parent 533fe808b6
commit b992163d4e
8 changed files with 53 additions and 34 deletions

View File

@@ -2993,6 +2993,8 @@ It just runs forever, so kill it when you get tired:
kill %1
....
`stress -c 1 -t 1` makes gem5 irresponsive for a very long time.
== QEMU
Some QEMU specific features to play with and limitations to cry over.
@@ -4544,19 +4546,22 @@ Unfortunately-we didn't manage to find an ARM analogue: link:kernel_module/pmccn
=== gem5 Python scripts without rebuild
TODO
We have made a crazy setup that allows you to just `cd` into `gem5/gem5`, and edit Python scripts directly there.
Currently, when you modify the python scripts under `gem5/gem5`, you still have to run `./build` to copy the updated scripts to the `out/<arch>/buildroot` directory, which kind of defeats the purpose of having an interpreted language as Python.
This is not normally possible with Buildroot, since normal Buildroot packages first copy files to the output directory (`out/<arch>/buildroot/build/<pkg>`), and then build there.
We would like to solve that, but are blocked on either of the following:
So if you modified the Python scripts with this setup, you would still need to `./build` to copy the modified files over.
* build in-tree and skip the Buildroot sync step. If sync is not skipped, it would copy all builds to all archs, which is huge.
+
But then we have to deal specially with the `m5` tool, which has to be cross compiled
* gem5 out of tree build. Is there support? Looks like SCons does not have anything generic as usual and leaves the burden on project developers:
** https://www.mail-archive.com/gem5-users@gem5.org/msg15421.html
** https://stackoverflow.com/questions/3720142/how-to-force-scons-output-exe-obj-lib-dll-to-specific-build-directory
** https://stackoverflow.com/questions/1762044/how-to-do-an-out-of-source-build-with-scons
For gem5 specifically however, we have hacked up the build so that we `cd` into the `gem5/gem5` tree, and then do an link:https://www.mail-archive.com/gem5-users@gem5.org/msg15421.html[out of tree] build to `out/common/gem5`.
Another advantage of this method is the we factor out the `arm` and `aarch64` gem5 builds which are identical and large, as well as the smaller arch generic pieces.
Using Buildroot for gem5 is still convenient because we use it to:
* to cross build `m5` for us
* check timestamps and skip the gem5 build when it is not requested
The out of build tree is required, because otherwise Buildroot would copy the output build of all archs to each arch directory, resulting in `arch^2` build copies, which is significant.
=== gem5 fs_bigLITTLE
@@ -5125,6 +5130,8 @@ Get results with:
tail -n+1 ../linux-kernel-module-cheat-regression/*/gem5-bench-build-*.txt
....
However, I have noticed that for some builds, with the exact same commands, it just take way longer sometimes, but I haven't been able to pin it down: https://github.com/cirosantilli-work/gem5-issues/issues/10
=== Benchmark machines
==== P51