Conversation
|
Is a normal runner able to build Debug? I thought the primary reason to use MSFT's runner was to be able to build debug which normal runners were not able to. |
We're going to find out! https://github.com/llvm/circt/actions/runs/24377364005 I thought I was getting debug builds through normal tests ( |
|
Looks like Debug was built successfully! But https://github.com/llvm/circt/blob/main/integration_test/circt-synth/comb-lowering-lec.mlir caused timeout 🙃 Certainly debug caused either Z3 or JIT execution very slow. I'll reduce problem sizes or split files. |
|
#10202 should fix this hopefully. |
Switch the nightly regression from using a `circt/images` Docker container and a custom build/test to using the UBTI build. This is intended to be a one-to-one change where all jobs running previously will run the same as before (just on different infrastructure). This is done for three reasons: (1) the combination of docker images running on hosted runners has empirically been broken for some time (both on the external MSFT-hosted runners as well as in at least one internal hosted GitHub Enterprise instance), (2) moving onto a single build flow for everything greatly simplifies how things are built and tested (there is exactly one build/test flow), and (3) the UBTI flow is more easily modifiable as it is self-documenting as opposed to being split over the `circt/images` and `llvm/circt` repositories. Assisted-by: Claude:claude-sonnet-4-6 Signed-off-by: Schuyler Eldridge <schuyler.eldridge@sifive.com>
seldridge
force-pushed
the
dev/seldridge/ci-ubti-nightly
branch
from
b2295df to
7884b61
Compare
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@sifive.com>
|
8403175 fixes a bug in sccache keys which are case-sensitive. Somewhat interestingly, it seems like GitHub actions conditional syntax is case insensitive which would have made this more obvious. Each Release sccache is 100--200MiB which should enough for us to fir the ~10 of these that a nightly build kicks off. We might even be able to get hits from nightly-to-nightly (as opposed to just from pre-merge-to-nightly). I'm going to do one more end-to-end run to make sure this is good and then merge. |
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@sifive.com>
|
I'm seeing this working for everything except the known valgrind nightly failure: #9914. I'll merge this and treat the valgrind failure as follow-up work. |
3 participants
Switch the nightly regression from using a
circt/imagesDocker containerand a custom build/test to using the UBTI build. This is intended to be a
one-to-one change where all jobs running previously will run the same as
before (just on different infrastructure).
This is done for three reasons: (1) the combination of docker images
running on hosted runners has empirically been broken for some time (both
on the external MSFT-hosted runners as well as in at least one internal
hosted GitHub Enterprise instance), (2) moving onto a single build flow
for everything greatly simplifies how things are built and tested (there
is exactly one build/test flow), and (3) the UBTI flow is more easily
modifiable as it is self-documenting as opposed to being split over the
circt/imagesandllvm/circtrepositories.Assisted-by: Claude:claude-sonnet-4-6
This is stacked on #10188.
I'm not going to try a full nightly until #10188 lands (to get the Verilator and Yosys caching). However, I've started and cancelled a workflow dispatch of this here: https://github.com/llvm/circt/actions/runs/24372135856 This is showing all 17 jobs being launched. A benefit about this is that we now have more flexibility in launching nightlies over different runners. (The integration tests won't run fully for anything other than Linux right now, but this can be made to work.)Full nightly is running here: https://github.com/llvm/circt/actions/runs/24407900984