This patch only introduces new signals but does not use their value
in scoring a CC candidate. Usage of these signals in CC ranking in both
heiristics and ML model will be introduced in later patches.
Differential Revision: https://reviews.llvm.org/D94473
With every incremental change, one needs to check-in new model upstream.
This also significantly increases the size of the git repo with every
new model.
Testing and comparing the old and previous model is also not possible as
we run only a single model at any point.
One solution is to have a "staging" decision forest which can be
injected into clangd without pushing it to upstream. Compare the
performance of the staging model with the live model. After a couple of
enhancements have been done to staging model, we can then replace the
live model upstream with the staging model. This reduces upstream churn
and also allows us to compare models with current baseline model.
This is done by having a callback in CodeCompleteOptions which is called
only when we want to use a decision forest ranking model. This allows us
to inject different completion model internally.
Differential Revision: https://reviews.llvm.org/D90014
Since we have 2 scoring functions (heuristics and decision forest),
renaming the existing evaluate() function to be more descriptive of the
Heuristics being evaluated in it.
Differential Revision: https://reviews.llvm.org/D88431
By default clangd will score a code completion item using heuristics model.
Scoring can be done by Decision Forest model by passing `--ranking_model=decision_forest` to
clangd.
Features omitted from the model:
- `NameMatch` is excluded because the final score must be multiplicative in `NameMatch` to allow rescoring by the editor.
- `NeedsFixIts` is excluded because the generating dataset that needs 'fixits' is non-trivial.
There are multiple ways (heuristics) to combine the above two features with the prediction of the DF:
- `NeedsFixIts` is used as is with a penalty of `0.5`.
Various alternatives of combining NameMatch `N` and Decision forest Prediction `P`
- N * scale(P, 0, 1): Linearly scale the output of model to range [0, 1]
- N * a^P:
- More natural: Prediction of each Decision Tree can be considered as a multiplicative boost (like NameMatch)
- Ordering is independent of the absolute value of P. Order of two items is proportional to `a^{difference in model prediction score}`. Higher `a` gives higher weightage to model output as compared to NameMatch score.
Baseline MRR = 0.619
MRR for various combinations:
N * P = 0.6346, advantage%=2.5768
N * 1.1^P = 0.6600, advantage%=6.6853
N * **1.2**^P = 0.6669, advantage%=**7.8005**
N * **1.3**^P = 0.6668, advantage%=**7.7795**
N * **1.4**^P = 0.6659, advantage%=**7.6270**
N * 1.5^P = 0.6646, advantage%=7.4200
N * 1.6^P = 0.6636, advantage%=7.2671
N * 1.7^P = 0.6629, advantage%=7.1450
N * 2^P = 0.6612, advantage%=6.8673
N * 2.5^P = 0.6598, advantage%=6.6491
N * 3^P = 0.6590, advantage%=6.5242
N * scaled[0, 1] = 0.6465, advantage%=4.5054
Differential Revision: https://reviews.llvm.org/D88281
Current implementation of heuristic-based scoring function also contains
computation of derived signals (e.g. whether name contains a word from
context, computing file distances, scope distances.)
This is an attempt to separate out the logic for computation of derived
signals from the scoring function.
This will allow us to have a clean API for scoring functions that will
take only concrete code completion signals as input.
Differential Revision: https://reviews.llvm.org/D88146
ContainsActiveParameter is not used anywhere, set incorrectly (see the
removed FIXME) and has no unit tests.
Removing it to simplify the code.
llvm-svn: 362686
Summary:
The hope is this will catch a few patterns with repetition:
SomeClass* S = ^SomeClass::Create()
int getFrobnicator() { return ^frobnicator_; }
// discard the factory, it's no longer valid.
^MyFactory.reset();
Without triggering antipatterns too often:
return Point(x.first, x.^second);
I'm going to gather some data on whether this turns out to be a win overall.
Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, jfb, kadircet, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D61537
llvm-svn: 360030
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Summary:
These are often not expected to be used directly e.g.
```
TEST_F(Fixture, X) {
^ // "Fixture_X_Test" expanded in the macro should be down ranked.
}
```
Only doing this for sema for now, as such symbols are mostly coming from sema
e.g. gtest macros expanded in the main file. We could also add a similar field
for the index symbol.
Reviewers: sammccall
Reviewed By: sammccall
Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, kadircet, cfe-commits
Differential Revision: https://reviews.llvm.org/D53374
llvm-svn: 344736
Summary:
This should make all-scope completion more usable. Scope proximity for
indexes will be added in followup patch.
Reviewers: sammccall
Reviewed By: sammccall
Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, kadircet, cfe-commits
Differential Revision: https://reviews.llvm.org/D53131
llvm-svn: 344688
Summary:
The following are metrics for explicit member access completions. There is no
noticeable impact on other completion types.
Before:
EXPLICIT_MEMBER_ACCESS
Total measurements: 24382
All measurements: MRR: 62.27 Top10: 80.21% Top-100: 94.48%
Full identifiers: MRR: 98.81 Top10: 99.89% Top-100: 99.95%
0-5 filter len:
MRR: 13.25 46.31 62.47 67.77 70.40 81.91
Top-10: 29% 74% 84% 91% 91% 97%
Top-100: 67% 99% 99% 99% 99% 100%
After:
EXPLICIT_MEMBER_ACCESS
Total measurements: 24382
All measurements: MRR: 63.18 Top10: 80.58% Top-100: 95.07%
Full identifiers: MRR: 98.79 Top10: 99.89% Top-100: 99.95%
0-5 filter len:
MRR: 13.84 48.39 63.55 68.83 71.28 82.64
Top-10: 30% 75% 84% 91% 91% 97%
Top-100: 70% 99% 99% 99% 99% 100%
* Top-N: wanted result is found in the first N completion results.
* MRR: Mean reciprocal rank.
Remark: the change seems to have minor positive impact. Although the improvement
is relatively small, down-ranking non-instance members in instance member access
should reduce noise in the completion results.
Reviewers: sammccall
Reviewed By: sammccall
Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, cfe-commits
Differential Revision: https://reviews.llvm.org/D49543
llvm-svn: 337681
Summary:
We now compute a distance from the main file to the symbol header, which
is a weighted count of:
- some number of #include traversals from source file --> included file
- some number of FS traversals from file --> parent directory
- some number of FS traversals from parent directory --> child file/dir
This calculation is performed in the appropriate URI scheme.
This means we'll get some proximity boost from header files in main-file
contexts, even when these are in different directory trees.
This extended file proximity model is not yet incorporated in the index
interface/implementation.
Reviewers: ioeric
Subscribers: mgorny, ilya-biryukov, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D48441
llvm-svn: 336177
Summary:
Also move unittest: URI scheme to TestFS so that it can be shared by
different tests.
Reviewers: sammccall
Reviewed By: sammccall
Subscribers: ilya-biryukov, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D47935
llvm-svn: 334810
Summary: These have few signals other than being keywords, so the boost is high.
Reviewers: ilya-biryukov
Subscribers: ioeric, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D48083
llvm-svn: 334711
Summary:
Now we have most of Sema's code completion signals incorporated in Quality,
which will allow us to give consistent ranking to sema/index results.
Therefore we can/should stop using Sema priority as an explicit signal.
This fixes some issues like namespaces always having a terrible score.
The most important missing signals are:
- Really dumb/rarely useful completions like:
SomeStruct().^SomeStruct
SomeStruct().^operator=
SomeStruct().~SomeStruct()
We already filter out destructors, this patch adds injected names and
operators to that list.
- type matching the expression context.
Ilya has a plan to add this in a way that's compatible with indexes
(design doc should be shared real soon now!)
Reviewers: ioeric
Subscribers: ilya-biryukov, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D47871
llvm-svn: 334192
Summary: Fix a couple of bugs in tests an in Quality to keep tests passing.
Reviewers: ioeric
Subscribers: ilya-biryukov, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D47815
llvm-svn: 334089
Summary:
This signal is considered a relevance rather than a quality signal because it's
dependent on the query (the fact that it's completion, and implicitly the query
context).
This is part of the effort to reduce reliance on Sema priority, so we can have
consistent ranking between Index and Sema results.
Reviewers: ioeric
Subscribers: klimek, ilya-biryukov, MaskRay, jkorous, cfe-commits
Differential Revision: https://reviews.llvm.org/D47762
llvm-svn: 334026
Summary:
Code completion scoring was embedded in CodeComplete.cpp, which is bad:
- awkward to test. The mechanisms (extracting info from index/sema) can be
unit-tested well, the policy (scoring) should be quantitatively measured.
Neither was easily possible, and debugging was hard.
The intermediate signal struct makes this easier.
- hard to reuse. This is a bug in workspaceSymbols: it just presents the
results in the index order, which is not sorted in practice, it needs to rank
them!
Also, index implementations care about scoring (both query-dependent and
independent) in order to truncate result lists appropriately.
The main yak shaved here is the build() function that had 3 variants across
unit tests is unified in TestTU.h (rather than adding a 4th variant).
Reviewers: ilya-biryukov
Subscribers: klimek, mgorny, ioeric, MaskRay, jkorous, mgrang, cfe-commits
Differential Revision: https://reviews.llvm.org/D46524
llvm-svn: 332378
Utkarsh Saxena