Now a budget of 0 disables mean unlimited and tests no longer set the
budget. Hopefully the deep completion tests will stop flaking.
Updates golang/go#34617
Change-Id: Icdff5e78dcf1cc3d3fcbf0326716b39b00f0a8c1
Reviewed-on: https://go-review.googlesource.com/c/tools/+/203338
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
We now continue deep completion search across function calls. The
function must take no arguments and return a single argument. For
example, when completing "fo<>" you might get candidates such as
"foo.bar().baz()".
Previously we would stop searching for deep completions when we hit a
function call. For example, we would stop at "foo.bar()", never
finding "foo.bar().baz()". At the time I was worried about the search
scope growing too large, but now that we dynamically limit the search
scope there isn't much left to worry about.
Change-Id: I48772c154400662876682503c1f58ef6e3dca688
Reviewed-on: https://go-review.googlesource.com/c/tools/+/201222
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
Tweak a couple things to improve how we reduce our search scope based
on remaining time budget:
- Check our budget on the first candidate rather than waiting for the
1000th candidate. If type checking is slow you can be out of budget
before you even begin.
- Reduce our budget check interval from 1000 candidates to 100
candidates. This just helps us adjust our search scope faster.
The first tweak required me to raise the completion budget for tests
because 100ms is not always enough. I moved the budget into the
completion options so that tests can raise it.
Change-Id: I1aa7909d7baf9c998bc830c960dc579eb33db12a
Reviewed-on: https://go-review.googlesource.com/c/tools/+/195419
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Deep completions can take a long time (500ms+) if there are many
large, deeply nested structs in scope. To make sure we return
completion results in a timely manner we now notice if we have spent
"too long" searching for deep completions and reduce the search scope.
In particular, our overall completion budget is 100ms. This value is
often cited as the longest latency that still feels instantaneous to
most people. As we spend 25%, 50%, and 75% of our budget we limit our
deep completion candidate search depth to 4, 3, and 2,
respectively. If we hit 90% of our budget, we disable deep completions
entirely.
In my testing, limiting the search scope to 4 normally makes even
enormous searches finish in a few milliseconds. Of course, you can
have arbitrarily many objects in scope with arbitrarily many fields,
so to cover our bases we continue to dial down the search depth as
needed.
I replaced the "enabled" field with a "maxDepth" field that disables
deep search when set to 0.
Change-Id: I9b5a07de70709895c065503ae6082d1ea615d1af
Reviewed-on: https://go-review.googlesource.com/c/tools/+/190978
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Cottrell <iancottrell@google.com>
Optimize a few things to speed up deep completions:
- item() is slow, so don't call it unless the candidate's name matches
the input.
- We only end up returning the top 3 deep candidates, so skip deep
candidates early if they are not in the top 3 scores we have seen so
far. This greatly reduces calls to item(), but also avoids a
humongous sort in lsp/completion.go.
- Get rid of error return value from found(). Nothing checked for this
error, and we spent a lot of time allocating the only possible error
"this candidate is not accessible", which is not unexpected to begin
with.
- Cache the call to types.NewMethodSet in methodsAndFields(). This is
relatively expensive and can be called many times for the same type
when searching for deep completions.
- Avoid calling deepState.chainString() twice by calling it once and
storing the result on the candidate.
These optimizations sped up my slow completion from 1.5s to
0.5s. There were around 200k deep candidates examined for this one
completion. The remaining time is dominated by the fuzzy
matcher. Obviously 500ms is still unacceptable under any
circumstances, so there will be subsequent improvements to limit the
deep completion search scope to make sure we always return completions
in a reasonable amount of time.
I also made it so there is always a "matcher" set on the
completer. This makes the matching logic a bit simpler.
Change-Id: Id48ef7031ee1d4ea04515c828277384562b988a8
Reviewed-on: https://go-review.googlesource.com/c/tools/+/190522
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
When it is certain we are completing a struct field name, we don't
want deep completions. The only possible completions are the remaining
field names.
I also silenced the log spam in tests by disabling the go/packages
logger and the lsp logger.
Fixesgolang/go#33614
Change-Id: Icec8d92112b1674fa7a6a21145ab710d054919b4
Reviewed-on: https://go-review.googlesource.com/c/tools/+/190097
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Deep completion refers to searching through an object's fields and
methods for more completion candidates. For example:
func wantsInt(int) { }
var s struct { i int }
wantsInt(<>)
Will now give a candidate for "s.i" since its type matches the
expected type.
We limit to three deep completion results. In some cases there are
many useless deep completion matches. Showing too many options defeats
the purpose of "smart" completions. We also lower a completion item's
score according to its depth so that we favor shallower options. For
now we do not continue searching past function calls to limit our
search scope. In other words, we are not able to suggest results with
any chained fields/methods after the first method call.
Deep completions are behind the "useDeepCompletions" LSP config flag
for now.
Change-Id: I1b888c82e5c4b882f9718177ce07811e2bccbf22
GitHub-Last-Rev: 26522363730036e0b382a7bcd10aa1ed825f6866
GitHub-Pull-Request: golang/tools#100
Reviewed-on: https://go-review.googlesource.com/c/tools/+/177622
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Run-TryBot: Rebecca Stambler <rstambler@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
