PIE mode tests with fibers #387
Conversation
|
huh, interesting... I cannot reproduce this. haven't gotten a single failure. \o/ I got nothin. |
|
oh wait, these tests are different. |
|
Looks like this might be a good work-around for non-alpine 8.3 builds. |
|
It looks like it still fails on Debian too. |
| @@ -48,11 +48,11 @@ jobs: | |||
| run: go build | |||
| - | |||
| name: Run library tests | |||
| run: go test -race -v ./... | |||
| run: CGO_CXXFLAGS=-fPIE CGO_CFLAGS=-fPIE CGO_LDFLAGS=-pie go test -buildmode=pie -v ./... | |||
There was a problem hiding this comment.
You need to preserve the previous values of these env vars (set line 37). Same in the Docker images.
|
It looks like this way of linking just so happens to lay out the memory in such a way that sometimes (if not most of the time), the fiber stack stays within the bounds of Go's stack. So, instead of a guaranteed explosion, you just randomly get explosions. In practice, I'm not seeing any explosions, but I think the tests sometimes passing/failing illustrates that the possibility is there. |
Running
CGO_CXXFLAGS=-fPIE CGO_CFLAGS=-fPIE CGO_LDFLAGS=-pie go test -buildmode=pie -v ./...(-raceis not compatible withpie) appears to be passing tests on x64 (local). Opening this PR to have it test all other architectures as well.I have no idea why this works and it might be an excellent question for golang/go/#62130 to see if it is actually working or just pretending to work by somehow making a larger stack that just so happens to pass the test.
Continuation of #99, #171. Closes #46, closes #99, closes #171.