Reland "[compiler-rt][test] Use packaging.version.Version to compare glibc versions" (#158230)" #158236
Conversation
…glibc versions" (llvm#158230)" This reverts commit d7b7b9c. TODO: with what added?
|
My worry with eating the exception here is that |
|
is the code that looks for the macros.
To get to using |
|
the Fuchsia people also ran into this issue and I was able to print the glibc version number on their bots and it is just "2.36", so I'm not sure what's going on before the change the logs showed
and after the change it started running the test:
and I also confirmed that reverting the previous land did make the CI pass. so the previous patch definitely did something, but I don't understand why based on the "2.36" version number |
|
actually I see both "2.36" and "2.19" in these logs... |
|
never mind all of the above is not what's happening. I believe our bots don't have the new package installed, but the import error is getting swallowed up here. sounds like we'll need to update our bots to have this new python package, but also the error message should be clearer |
|
I'll send out a patch that manually compares version numbers without an external package |
Alternative to llvm#158236 without requiring external packages
|
This reverts commit d7b7b9c.
As the name suggests, packaging's Version is more strict than distutil's LooseVersion.
LooseVersion(".3")is fine butVersion(".3")raises an exception. I think this was happening on the Google bot, which lead us to enable tests that should not have been running.Looking at the way we find the glibc version, it should be
<0 or more numbers>.<0 or more numbers>. So I think the only way we're getting invalid versions is if it comes out asX.or.Y.