Everybody loops pprint; -Z unstable-options #19964
Conversation
pnkfelix
changed the title
|
Note: This series of commits is not quite complete: I still need to migrate the |
|
see also PR #19900, which attempted to stabilize some CLI for |
|
(also, there is a reasonable argument against |
|
r=me once tests are passing on this and you're ready to merge |
|
Thanks for addressing this @pnkfelix. I like this idea in principle though I haven't looked at the patch. |
extend the `rustc` command line interface with options that we do not want to commit to making part of the long-term public interface for `rustc`.
This prints out a transformed version of the input source code where
every function body is replaced with `loop { }`.
All such bodies are (1.) trivial and (2.) guaranteed to pass the
type-checker in a valid compiler; therefore they make very nice input
to start with when narrowing down a bug exposed by a large test input
(such as librustc itself).
and `everybody_loops` options to `--xpretty`.
pnkfelix
force-pushed
the
everybody-loops-pprint
branch
from
3657207 to
bf2f84b
Compare
|
cc #19051 |
…ichton
NOTE: Not yet ready for merge; see my first comment below. I will remove this note after I post the appropriate follow-on commit to get `run-make` working too.
This PR adds a new pretty printing mode, `everybody_loops`, which replaces every function body in the input with `loop { }`. I have found this to be useful for building narrow test cases starting from bugs in static analyses like typeck and borrowck that are only exposed initially from complex input source crates like `librustc`.
The process to follow is:
1. You first generate the new input with every function body replaced with `loop { }`, then
2. you cut-and-paste the original function body that exposes the error,
3. re-run the compiler on the modified output, to confirm you still see the bug -- but now you should get results faster and with much narrower debug output, since all the other function bodies are trivial.
4., optional you iteratively delete all the other items that are now found to be no longer necessary since all of the function bodies are now just `loop { }`.
(Strictly speaking, there are bugs in the pretty printer and/or parser that make the process above not as seamless as described. For example, it seems that `use super::{A, B};` can in some contexts be pretty-printed with whitespace separating `super` and `::` and `{A, B};`, which the parser then errors on. But that is not an argument against the overall technique, which still appears pretty effective, though perhaps it would be even better if combined with macro-expansion in some way.)
----
Up front: I do believe this sort of local development hack should not want this option to be part of the stable public interface for `rustc`. So while I could make a `-Z` flag just for this, I am worried that our `-Z` infrastructure is generally too weak to express the kinds of options I want to add (see e.g. #19892 where I use the hack of employing environment variables, when I really would have preferred to leverage `libgetopts`).
Thus, this PR incorporates what I believe to be a reasonable compromise: Add a single new `-Z` flag, `-Z unstable-options`, which expands the set of valid options to `rustc`. This way, we still leverage all of the `getopts` infrastructure: for example, `rustc -Z unstable-options --help` prints out information about both the stable and unstable options (while `rustc --help` without the `-Z unstable-options` flag just prints out information about the stable options. The main drawback that I can see is that processing such options is a little bit awkward, since you need to make sure that you check `sess.debugging_opt(config::UNSTABLE_OPTIONS)` before querying the `getopts` matches for the option in question. But it really is not that bad; all of the ugliest stuff is paid for up front in this PR (e.g. the way I try to ensure that we do not waste time parsing the options twice in the common case where all the provided options are stable).
----
With `-Z unstable-options` in place, it is clear that pretty-print modes like `everybody_loops` belongs under an unstable option, as does the control flow graph visualizing pretty printer. To deal with that, I added a new unstable-option, `--xpretty` and moved the latter two pretty print modes there.
|
(heh, I forgot to take care of the exact thing I said I would do, namely fixing the |
…`--xpretty` option.
Conflicts: src/librustc/session/config.rs src/librustc_driver/lib.rs src/librustc_driver/pretty.rs
This PR adds a new pretty printing mode,
everybody_loops, which replaces every function body in the input withloop { }. I have found this to be useful for building narrow test cases starting from bugs in static analyses like typeck and borrowck that are only exposed initially from complex input source crates likelibrustc.The process to follow is:
loop { }, then4., optional you iteratively delete all the other items that are now found to be no longer necessary since all of the function bodies are now just
loop { }.(Strictly speaking, there are bugs in the pretty printer and/or parser that make the process above not as seamless as described. For example, it seems that
use super::{A, B};can in some contexts be pretty-printed with whitespace separatingsuperand::and{A, B};, which the parser then errors on. But that is not an argument against the overall technique, which still appears pretty effective, though perhaps it would be even better if combined with macro-expansion in some way.)Up front: I do believe this sort of local development hack should not want this option to be part of the stable public interface for
rustc. So while I could make a-Zflag just for this, I am worried that our-Zinfrastructure is generally too weak to express the kinds of options I want to add (see e.g. #19892 where I use the hack of employing environment variables, when I really would have preferred to leveragelibgetopts).Thus, this PR incorporates what I believe to be a reasonable compromise: Add a single new
-Zflag,-Z unstable-options, which expands the set of valid options torustc. This way, we still leverage all of thegetoptsinfrastructure: for example,rustc -Z unstable-options --helpprints out information about both the stable and unstable options (whilerustc --helpwithout the-Z unstable-optionsflag just prints out information about the stable options. The main drawback that I can see is that processing such options is a little bit awkward, since you need to make sure that you checksess.debugging_opt(config::UNSTABLE_OPTIONS)before querying thegetoptsmatches for the option in question. But it really is not that bad; all of the ugliest stuff is paid for up front in this PR (e.g. the way I try to ensure that we do not waste time parsing the options twice in the common case where all the provided options are stable).With
-Z unstable-optionsin place, it is clear that pretty-print modes likeeverybody_loopsbelongs under an unstable option, as does the control flow graph visualizing pretty printer. To deal with that, I added a new unstable-option,--xprettyand moved the latter two pretty print modes there.