[update_cc_test_checks] Use lit's shell to run commands #65333
Conversation
While update_test_checks and update_llc_test_checks handle these correctly, it does not currently work in update_cc_test_checks.
I was trying to update a test that redirected stderr to a file and hit errors because the 2>%t.out was being passed as a positional argument to clang. Instead of re-implementing basic shell parsing in this script, we can reuse lit's internal shell to run these commands and handle redirection and pipelines. A more minimal fix could have been to use the shell=True parameter in python's subprocess modules, but this would not work e.g. on Windows. Additionally, I am planning to simplify the common.py infrastructure by reusing the shell helpers from lit and this is the first step in that direction.
|
Looks like only squash and merge is possible here, so I'll commit the baseline test separately before landing this. |
| ret = collections.defaultdict(list) | ||
| # Use clang's JSON AST dump to get the mangled name | ||
| json_dump_args = [args.clang] + clang_args + ["-fsyntax-only", "-o", "-"] | ||
| json_dump_args = clang_cmd.args + ["-fsyntax-only", "-o", "/dev/null"] |
There was a problem hiding this comment.
I don't understand the change from stdout to /dev/null output. How can you get an output from that command when doing execute_pipeline?
There was a problem hiding this comment.
This change is not required, it seems JSON is always printed to stdout. This is just to ensure to other output is produced (although probably -fsyntax-only) is sufficient.
| cmd.redirects[i] = redirect[0], common.applySubstitutions( | ||
| redirect[1], subs | ||
| ) | ||
|
|
||
| # Parse executable args. |
There was a problem hiding this comment.
This comment should be removed.
| ] | ||
| clang_args += ti.args.clang_args | ||
| clang_cmd.args[0:1] = SUBST[clang_cmd.args[0]] | ||
| print(clang_cmd) |
There was a problem hiding this comment.
Is that a forgotten debug print statement?
| # Remove all -verify arguments since they could cause the IR generation to fail | ||
| clang_cmd.args = [x for x in clang_cmd.args if not x.startswith("-verify")] |
There was a problem hiding this comment.
Wouldn't you want the tool to fail then?
There was a problem hiding this comment.
Good catch that is a local diff that should not have been added to this change.
My use-case for that change was updating tests after a merge as a two-stage process, first fix the CHECK lines then the -verify ones, but you're right this can be dropped.
I was trying to update a test that redirected stderr to a file and hit
errors because the
2>%t.outwas being passed as a positional argument toclang. Instead of re-implementing basic shell parsing in this script, we
can reuse lit's internal shell to run these commands and handle
redirection and pipelines.
A more minimal fix could have been to use the shell=True parameter in
python's subprocess modules, but this would not work e.g. on Windows.
Additionally, I am planning to simplify the common.py infrastructure by
reusing the shell helpers from lit and this is the first step in that
direction.