1
0
mirror of https://github.com/golang/go synced 2024-11-05 22:56:11 -07:00
go/internal/lsp/source/completion.go

2270 lines
64 KiB
Go
Raw Normal View History

// Copyright 2018 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package source
import (
"context"
"fmt"
"go/ast"
"go/constant"
"go/scanner"
"go/token"
"go/types"
"math"
"sort"
"strconv"
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
"strings"
"sync"
"time"
"golang.org/x/tools/go/ast/astutil"
"golang.org/x/tools/internal/event"
"golang.org/x/tools/internal/imports"
"golang.org/x/tools/internal/lsp/fuzzy"
"golang.org/x/tools/internal/lsp/protocol"
"golang.org/x/tools/internal/lsp/snippet"
errors "golang.org/x/xerrors"
)
type CompletionItem struct {
// Label is the primary text the user sees for this completion item.
Label string
// Detail is supplemental information to present to the user.
// This often contains the type or return type of the completion item.
Detail string
// InsertText is the text to insert if this item is selected.
// Any of the prefix that has already been typed is not trimmed.
// The insert text does not contain snippets.
InsertText string
Kind protocol.CompletionItemKind
// An optional array of additional TextEdits that are applied when
// selecting this completion.
//
// Additional text edits should be used to change text unrelated to the current cursor position
// (for example adding an import statement at the top of the file if the completion item will
// insert an unqualified type).
AdditionalTextEdits []protocol.TextEdit
// Depth is how many levels were searched to find this completion.
// For example when completing "foo<>", "fooBar" is depth 0, and
// "fooBar.Baz" is depth 1.
Depth int
// Score is the internal relevance score.
// A higher score indicates that this completion item is more relevant.
Score float64
// snippet is the LSP snippet for the completion item. The LSP
// specification contains details about LSP snippets. For example, a
// snippet for a function with the following signature:
//
// func foo(a, b, c int)
//
// would be:
//
// foo(${1:a int}, ${2: b int}, ${3: c int})
//
// If Placeholders is false in the CompletionOptions, the above
// snippet would instead be:
//
// foo(${1:})
snippet *snippet.Builder
// Documentation is the documentation for the completion item.
Documentation string
}
// Snippet is a convenience returns the snippet if available, otherwise
// the InsertText.
// used for an item, depending on if the callee wants placeholders or not.
func (i *CompletionItem) Snippet() string {
if i.snippet != nil {
return i.snippet.String()
}
return i.InsertText
}
// Scoring constants are used for weighting the relevance of different candidates.
const (
// stdScore is the base score for all completion items.
stdScore float64 = 1.0
// highScore indicates a very relevant completion item.
highScore float64 = 10.0
// lowScore indicates an irrelevant or not useful completion item.
lowScore float64 = 0.01
)
// matcher matches a candidate's label against the user input. The
// returned score reflects the quality of the match. A score of zero
// indicates no match, and a score of one means a perfect match.
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
type matcher interface {
Score(candidateLabel string) (score float32)
}
// prefixMatcher implements case sensitive prefix matching.
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
type prefixMatcher string
func (pm prefixMatcher) Score(candidateLabel string) float32 {
if strings.HasPrefix(candidateLabel, string(pm)) {
return 1
}
return -1
}
// insensitivePrefixMatcher implements case insensitive prefix matching.
type insensitivePrefixMatcher string
func (ipm insensitivePrefixMatcher) Score(candidateLabel string) float32 {
if strings.HasPrefix(strings.ToLower(candidateLabel), string(ipm)) {
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
return 1
}
return -1
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
}
// completer contains the necessary information for a single completion request.
type completer struct {
snapshot Snapshot
pkg Package
qf types.Qualifier
opts *completionOptions
// filename is the name of the file associated with this completion request.
filename string
// file is the AST of the file associated with this completion request.
file *ast.File
// pos is the position at which the request was triggered.
pos token.Pos
// path is the path of AST nodes enclosing the position.
path []ast.Node
// seen is the map that ensures we do not return duplicate results.
seen map[types.Object]bool
// items is the list of completion items returned.
items []CompletionItem
// surrounding describes the identifier surrounding the position.
surrounding *Selection
// inference contains information we've inferred about ideal
// candidates such as the candidate's type.
inference candidateInference
// enclosingFunc contains information about the function enclosing
// the position.
enclosingFunc *funcInfo
// enclosingCompositeLiteral contains information about the composite literal
// enclosing the position.
enclosingCompositeLiteral *compLitInfo
// deepState contains the current state of our deep completion search.
deepState deepCompletionState
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
// matcher matches the candidates against the surrounding prefix.
matcher matcher
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
// methodSetCache caches the types.NewMethodSet call, which is relatively
// expensive and can be called many times for the same type while searching
// for deep completions.
methodSetCache map[methodSetKey]*types.MethodSet
// mapper converts the positions in the file from which the completion originated.
mapper *protocol.ColumnMapper
// startTime is when we started processing this completion request. It does
// not include any time the request spent in the queue.
startTime time.Time
}
// funcInfo holds info about a function object.
type funcInfo struct {
// sig is the function declaration enclosing the position.
sig *types.Signature
// body is the function's body.
body *ast.BlockStmt
}
type compLitInfo struct {
// cl is the *ast.CompositeLit enclosing the position.
cl *ast.CompositeLit
// clType is the type of cl.
clType types.Type
// kv is the *ast.KeyValueExpr enclosing the position, if any.
kv *ast.KeyValueExpr
// inKey is true if we are certain the position is in the key side
// of a key-value pair.
inKey bool
// maybeInFieldName is true if inKey is false and it is possible
// we are completing a struct field name. For example,
// "SomeStruct{<>}" will be inKey=false, but maybeInFieldName=true
// because we _could_ be completing a field name.
maybeInFieldName bool
}
type importInfo struct {
importPath string
name string
pkg Package
}
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
type methodSetKey struct {
typ types.Type
addressable bool
}
// A Selection represents the cursor position and surrounding identifier.
type Selection struct {
content string
cursor token.Pos
mappedRange
}
func (p Selection) Prefix() string {
return p.content[:p.cursor-p.spanRange.Start]
}
func (p Selection) Suffix() string {
return p.content[p.cursor-p.spanRange.Start:]
}
func (c *completer) setSurrounding(ident *ast.Ident) {
if c.surrounding != nil {
return
}
if !(ident.Pos() <= c.pos && c.pos <= ident.End()) {
return
}
c.surrounding = &Selection{
content: ident.Name,
cursor: c.pos,
// Overwrite the prefix only.
mappedRange: newMappedRange(c.snapshot.View().Session().Cache().FileSet(), c.mapper, ident.Pos(), ident.End()),
}
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
switch c.opts.matcher {
case Fuzzy:
c.matcher = fuzzy.NewMatcher(c.surrounding.Prefix())
case CaseSensitive:
c.matcher = prefixMatcher(c.surrounding.Prefix())
default:
c.matcher = insensitivePrefixMatcher(strings.ToLower(c.surrounding.Prefix()))
}
}
func (c *completer) getSurrounding() *Selection {
if c.surrounding == nil {
c.surrounding = &Selection{
content: "",
cursor: c.pos,
mappedRange: newMappedRange(c.snapshot.View().Session().Cache().FileSet(), c.mapper, c.pos, c.pos),
}
}
return c.surrounding
}
// found adds a candidate completion. We will also search through the object's
// members for more candidates.
func (c *completer) found(ctx context.Context, cand candidate) {
obj := cand.obj
if obj.Pkg() != nil && obj.Pkg() != c.pkg.GetTypes() && !obj.Exported() {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
// obj is not accessible because it lives in another package and is not
// exported. Don't treat it as a completion candidate.
return
}
if c.inDeepCompletion() {
// When searching deep, just make sure we don't have a cycle in our chain.
// We don't dedupe by object because we want to allow both "foo.Baz" and
// "bar.Baz" even though "Baz" is represented the same types.Object in both.
for _, seenObj := range c.deepState.chain {
if seenObj == obj {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
return
}
}
} else {
// At the top level, dedupe by object.
if c.seen[obj] {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
return
}
c.seen[obj] = true
}
// If we are running out of budgeted time we must limit our search for deep
// completion candidates.
if c.shouldPrune() {
return
}
// If we know we want a type name, don't offer non-type name
// candidates. However, do offer package names since they can
// contain type names, and do offer any candidate without a type
// since we aren't sure if it is a type name or not (i.e. unimported
// candidate).
if c.wantTypeName() && obj.Type() != nil && !isTypeName(obj) && !isPkgName(obj) {
return
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if c.matchingCandidate(&cand) {
cand.score *= highScore
} else if isTypeName(obj) {
// If obj is a *types.TypeName that didn't otherwise match, check
// if a literal object of this type makes a good candidate.
// We only care about named types (i.e. don't want builtin types).
if _, isNamed := obj.Type().(*types.Named); isNamed {
c.literal(ctx, obj.Type(), cand.imp)
}
}
// Lower score of function calls so we prefer fields and vars over calls.
if cand.expandFuncCall {
cand.score *= 0.9
}
// Prefer private objects over public ones.
if !obj.Exported() && obj.Parent() != types.Universe {
cand.score *= 1.1
}
// Favor shallow matches by lowering score according to depth.
cand.score -= cand.score * c.deepState.scorePenalty()
if cand.score < 0 {
cand.score = 0
}
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
cand.name = c.deepState.chainString(obj.Name())
matchScore := c.matcher.Score(cand.name)
if matchScore > 0 {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
cand.score *= float64(matchScore)
// Avoid calling c.item() for deep candidates that wouldn't be in the top
// MaxDeepCompletions anyway.
if !c.inDeepCompletion() || c.deepState.isHighScore(cand.score) {
if item, err := c.item(ctx, cand); err == nil {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
c.items = append(c.items, item)
}
internal/lsp: add fuzzy completion matching Make use of the existing fuzzy matcher to perform server side fuzzy completion matching. Previously the server did exact prefix matching for completion candidates and left fancy filtering to the client. Having the server do fuzzy matching has two main benefits: - Deep completions now update as you type. The completion candidates returned to the client are marked "incomplete", causing the client to refresh the candidates after every keystroke. This lets the server pick the most relevant set of deep completion candidates. - All editors get fuzzy matching for free. VSCode has fuzzy matching out of the box, but some editors either don't provide it, or it can be difficult to set up. I modified the fuzzy matcher to allow matches where the input doesn't match the final segment of the candidate. For example, previously "ab" would not match "abc.def" because the "b" in "ab" did not match the final segment "def". I can see how this is useful when the text matching happens in a vacuum and candidate's final segment is the most specific part. But, in our case, we have various other methods to order candidates, so we don't want to exclude them just because the final segment doesn't match. For example, if we know our candidate needs to be type "context.Context" and "foo.ctx" is of the right type, we want to suggest "foo.ctx" as soon as the user starts inputting "foo", even though "foo" doesn't match "ctx" at all. Note that fuzzy matching is behind the "useDeepCompletions" config flag for the time being. Change-Id: Ic7674f0cf885af770c30daef472f2e3c5ac4db78 Reviewed-on: https://go-review.googlesource.com/c/tools/+/190099 Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-08-13 14:45:19 -06:00
}
}
c.deepSearch(ctx, cand)
}
// candidate represents a completion candidate.
type candidate struct {
// obj is the types.Object to complete to.
obj types.Object
// score is used to rank candidates.
score float64
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
// name is the deep object name path, e.g. "foo.bar"
name string
// expandFuncCall is true if obj should be invoked in the completion.
// For example, expandFuncCall=true yields "foo()", expandFuncCall=false yields "foo".
expandFuncCall bool
// takeAddress is true if the completion should take a pointer to obj.
// For example, takeAddress=true yields "&foo", takeAddress=false yields "foo".
takeAddress bool
// addressable is true if a pointer can be taken to the candidate.
addressable bool
// makePointer is true if the candidate type name T should be made into *T.
makePointer bool
// dereference is a count of how many times to dereference the candidate obj.
// For example, dereference=2 turns "foo" into "**foo" when formatting.
dereference int
// imp is the import that needs to be added to this package in order
// for this candidate to be valid. nil if no import needed.
imp *importInfo
}
// ErrIsDefinition is an error that informs the user they got no
// completions because they tried to complete the name of a new object
// being defined.
type ErrIsDefinition struct {
objStr string
}
func (e ErrIsDefinition) Error() string {
msg := "this is a definition"
if e.objStr != "" {
msg += " of " + e.objStr
}
return msg
}
// Completion returns a list of possible candidates for completion, given a
// a file and a position.
//
// The selection is computed based on the preceding identifier and can be used by
// the client to score the quality of the completion. For instance, some clients
// may tolerate imperfect matches as valid completion results, since users may make typos.
func Completion(ctx context.Context, snapshot Snapshot, fh FileHandle, protoPos protocol.Position) ([]CompletionItem, *Selection, error) {
ctx, done := event.Start(ctx, "source.Completion")
defer done()
startTime := time.Now()
pkg, pgh, err := getParsedFile(ctx, snapshot, fh, NarrowestPackageHandle)
if err != nil {
return nil, nil, fmt.Errorf("getting file for Completion: %w", err)
}
file, src, m, _, err := pgh.Cached()
if err != nil {
return nil, nil, err
}
spn, err := m.PointSpan(protoPos)
if err != nil {
return nil, nil, err
}
rng, err := spn.Range(m.Converter)
if err != nil {
return nil, nil, err
}
// Completion is based on what precedes the cursor.
// Find the path to the position before pos.
path, _ := astutil.PathEnclosingInterval(file, rng.Start-1, rng.Start-1)
if path == nil {
return nil, nil, errors.Errorf("cannot find node enclosing position")
}
pos := rng.Start
// Check if completion at this position is valid. If not, return early.
switch n := path[0].(type) {
case *ast.BasicLit:
// Skip completion inside any kind of literal.
return nil, nil, nil
case *ast.CallExpr:
if n.Ellipsis.IsValid() && pos > n.Ellipsis && pos <= n.Ellipsis+token.Pos(len("...")) {
// Don't offer completions inside or directly after "...". For
// example, don't offer completions at "<>" in "foo(bar...<>").
return nil, nil, nil
}
case *ast.Ident:
// reject defining identifiers
if obj, ok := pkg.GetTypesInfo().Defs[n]; ok {
if v, ok := obj.(*types.Var); ok && v.IsField() && v.Embedded() {
// An anonymous field is also a reference to a type.
} else {
objStr := ""
if obj != nil {
qual := types.RelativeTo(pkg.GetTypes())
objStr = types.ObjectString(obj, qual)
}
return nil, nil, ErrIsDefinition{objStr: objStr}
}
}
}
opts := snapshot.View().Options()
c := &completer{
pkg: pkg,
snapshot: snapshot,
qf: qualifier(file, pkg.GetTypes(), pkg.GetTypesInfo()),
filename: fh.Identity().URI.Filename(),
file: file,
path: path,
pos: pos,
seen: make(map[types.Object]bool),
enclosingFunc: enclosingFunction(path, pkg.GetTypesInfo()),
enclosingCompositeLiteral: enclosingCompositeLiteral(path, rng.Start, pkg.GetTypesInfo()),
opts: &completionOptions{
matcher: opts.Matcher,
deepCompletion: opts.DeepCompletion,
unimported: opts.UnimportedCompletion,
documentation: opts.CompletionDocumentation,
fullDocumentation: opts.HoverKind == FullDocumentation,
placeholders: opts.Placeholders,
literal: opts.InsertTextFormat == protocol.SnippetTextFormat,
budget: opts.CompletionBudget,
},
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
// default to a matcher that always matches
matcher: prefixMatcher(""),
methodSetCache: make(map[methodSetKey]*types.MethodSet),
mapper: m,
startTime: startTime,
}
if c.opts.deepCompletion {
// Initialize max search depth to unlimited.
c.deepState.maxDepth = -1
}
var cancel context.CancelFunc
if c.opts.budget == 0 {
ctx, cancel = context.WithCancel(ctx)
} else {
ctx, cancel = context.WithDeadline(ctx, c.startTime.Add(c.opts.budget))
}
defer cancel()
if surrounding := c.containingIdent(src); surrounding != nil {
c.setSurrounding(surrounding)
}
c.inference = expectedCandidate(ctx, c)
defer c.sortItems()
// If we're inside a comment return comment completions
for _, comment := range file.Comments {
if comment.Pos() < rng.Start && rng.Start <= comment.End() {
c.populateCommentCompletions(ctx, comment)
return c.items, c.getSurrounding(), nil
}
}
// Struct literals are handled entirely separately.
if c.wantStructFieldCompletions() {
if err := c.structLiteralFieldName(ctx); err != nil {
return nil, nil, err
}
return c.items, c.getSurrounding(), nil
}
if lt := c.wantLabelCompletion(); lt != labelNone {
c.labels(ctx, lt)
return c.items, c.getSurrounding(), nil
}
// Statement candidates offer an entire statement in certain
// contexts, as opposed to a single object.
c.addStatementCandidates()
if c.emptySwitchStmt() {
// Empty switch statements only admit "default" and "case" keywords.
c.addKeywordItems(map[string]bool{}, highScore, CASE, DEFAULT)
return c.items, c.getSurrounding(), nil
}
switch n := path[0].(type) {
case *ast.Ident:
// Is this the Sel part of a selector?
if sel, ok := path[1].(*ast.SelectorExpr); ok && sel.Sel == n {
if err := c.selector(ctx, sel); err != nil {
return nil, nil, err
}
return c.items, c.getSurrounding(), nil
}
if err := c.lexical(ctx); err != nil {
return nil, nil, err
}
// The function name hasn't been typed yet, but the parens are there:
// recv.‸(arg)
case *ast.TypeAssertExpr:
// Create a fake selector expression.
if err := c.selector(ctx, &ast.SelectorExpr{X: n.X}); err != nil {
return nil, nil, err
}
case *ast.SelectorExpr:
if err := c.selector(ctx, n); err != nil {
return nil, nil, err
}
// At the file scope, only keywords are allowed.
case *ast.BadDecl, *ast.File:
c.addKeywordCompletions()
default:
// fallback to lexical completions
if err := c.lexical(ctx); err != nil {
return nil, nil, err
}
}
return c.items, c.getSurrounding(), nil
}
// containingIdent returns the *ast.Ident containing pos, if any. It
// synthesizes an *ast.Ident to allow completion in the face of
// certain syntax errors.
func (c *completer) containingIdent(src []byte) *ast.Ident {
// In the normal case, our leaf AST node is the identifer being completed.
if ident, ok := c.path[0].(*ast.Ident); ok {
return ident
}
pos, tkn, lit := c.scanToken(src)
if !pos.IsValid() {
return nil
}
fakeIdent := &ast.Ident{Name: lit, NamePos: pos}
if _, isBadDecl := c.path[0].(*ast.BadDecl); isBadDecl {
// You don't get *ast.Idents at the file level, so look for bad
// decls and use the manually extracted token.
return fakeIdent
} else if c.emptySwitchStmt() {
// Only keywords are allowed in empty switch statements.
// *ast.Idents are not parsed, so we must use the manually
// extracted token.
return fakeIdent
} else if tkn.IsKeyword() {
// Otherwise, manually extract the prefix if our containing token
// is a keyword. This improves completion after an "accidental
// keyword", e.g. completing to "variance" in "someFunc(var<>)".
return fakeIdent
}
return nil
}
// scanToken scans pgh's contents for the token containing pos.
func (c *completer) scanToken(contents []byte) (token.Pos, token.Token, string) {
tok := c.snapshot.View().Session().Cache().FileSet().File(c.pos)
var s scanner.Scanner
s.Init(tok, contents, nil, 0)
for {
tknPos, tkn, lit := s.Scan()
if tkn == token.EOF || tknPos >= c.pos {
return token.NoPos, token.ILLEGAL, ""
}
if len(lit) > 0 && tknPos <= c.pos && c.pos <= tknPos+token.Pos(len(lit)) {
return tknPos, tkn, lit
}
}
}
func (c *completer) sortItems() {
sort.SliceStable(c.items, func(i, j int) bool {
// Sort by score first.
if c.items[i].Score != c.items[j].Score {
return c.items[i].Score > c.items[j].Score
}
// Then sort by label so order stays consistent. This also has the
// effect of prefering shorter candidates.
return c.items[i].Label < c.items[j].Label
})
}
// emptySwitchStmt reports whether pos is in an empty switch or select
// statement.
func (c *completer) emptySwitchStmt() bool {
block, ok := c.path[0].(*ast.BlockStmt)
if !ok || len(block.List) > 0 || len(c.path) == 1 {
return false
}
switch c.path[1].(type) {
case *ast.SwitchStmt, *ast.TypeSwitchStmt, *ast.SelectStmt:
return true
default:
return false
}
}
// populateCommentCompletions yields completions for exported
// symbols immediately preceding comment.
func (c *completer) populateCommentCompletions(ctx context.Context, comment *ast.CommentGroup) {
// Using the comment position find the line after
fset := c.snapshot.View().Session().Cache().FileSet()
file := fset.File(comment.Pos())
if file == nil {
return
}
line := file.Line(comment.Pos())
if file.LineCount() < line+1 {
return
}
nextLinePos := file.LineStart(line + 1)
if !nextLinePos.IsValid() {
return
}
// Using the next line pos, grab and parse the exported symbol on that line
for _, n := range c.file.Decls {
if n.Pos() != nextLinePos {
continue
}
switch node := n.(type) {
// handle const, vars, and types
case *ast.GenDecl:
for _, spec := range node.Specs {
switch spec.(type) {
case *ast.ValueSpec:
valueSpec, ok := spec.(*ast.ValueSpec)
if !ok {
continue
}
for _, name := range valueSpec.Names {
if name.String() == "_" || !name.IsExported() {
continue
}
obj := c.pkg.GetTypesInfo().ObjectOf(name)
c.found(ctx, candidate{obj: obj, score: stdScore})
}
case *ast.TypeSpec:
typeSpec, ok := spec.(*ast.TypeSpec)
if !ok {
continue
}
if typeSpec.Name.String() == "_" || !typeSpec.Name.IsExported() {
continue
}
obj := c.pkg.GetTypesInfo().ObjectOf(typeSpec.Name)
c.found(ctx, candidate{obj: obj, score: stdScore})
}
}
// handle functions
case *ast.FuncDecl:
if node.Name.String() == "_" || !node.Name.IsExported() {
continue
}
obj := c.pkg.GetTypesInfo().ObjectOf(node.Name)
// We don't want expandFuncCall inside comments. We add this directly to the
// completions list because using c.found sets expandFuncCall to true by default
item, err := c.item(ctx, candidate{
obj: obj,
name: obj.Name(),
expandFuncCall: false,
score: stdScore,
})
if err != nil {
continue
}
c.items = append(c.items, item)
}
}
}
func (c *completer) wantStructFieldCompletions() bool {
clInfo := c.enclosingCompositeLiteral
if clInfo == nil {
return false
}
return clInfo.isStruct() && (clInfo.inKey || clInfo.maybeInFieldName)
}
func (c *completer) wantTypeName() bool {
return c.inference.typeName.wantTypeName
}
// See https://golang.org/issue/36001. Unimported completions are expensive.
const (
maxUnimportedPackageNames = 5
unimportedMemberTarget = 100
)
// selector finds completions for the specified selector expression.
func (c *completer) selector(ctx context.Context, sel *ast.SelectorExpr) error {
// Is sel a qualified identifier?
if id, ok := sel.X.(*ast.Ident); ok {
if pkgName, ok := c.pkg.GetTypesInfo().Uses[id].(*types.PkgName); ok {
c.packageMembers(ctx, pkgName.Imported(), stdScore, nil)
return nil
}
}
// Invariant: sel is a true selector.
tv, ok := c.pkg.GetTypesInfo().Types[sel.X]
if ok {
return c.methodsAndFields(ctx, tv.Type, tv.Addressable(), nil)
}
// Try unimported packages.
if id, ok := sel.X.(*ast.Ident); ok && c.opts.unimported {
if err := c.unimportedMembers(ctx, id); err != nil {
return err
}
}
return nil
}
func (c *completer) unimportedMembers(ctx context.Context, id *ast.Ident) error {
// Try loaded packages first. They're relevant, fast, and fully typed.
known, err := c.snapshot.CachedImportPaths(ctx)
if err != nil {
return err
}
var paths []string
for path, pkg := range known {
if pkg.GetTypes().Name() != id.Name {
continue
}
paths = append(paths, path)
}
var relevances map[string]int
if len(paths) != 0 {
c.snapshot.View().RunProcessEnvFunc(ctx, func(opts *imports.Options) error {
relevances = imports.ScoreImportPaths(ctx, opts.Env, paths)
return nil
})
}
for path, relevance := range relevances {
pkg := known[path]
if pkg.GetTypes().Name() != id.Name {
continue
}
imp := &importInfo{
importPath: path,
pkg: pkg,
}
if imports.ImportPathToAssumedName(path) != pkg.GetTypes().Name() {
imp.name = pkg.GetTypes().Name()
}
c.packageMembers(ctx, pkg.GetTypes(), stdScore+.1*float64(relevance), imp)
if len(c.items) >= unimportedMemberTarget {
return nil
}
}
ctx, cancel := context.WithCancel(ctx)
defer cancel()
var mu sync.Mutex
add := func(pkgExport imports.PackageExport) {
mu.Lock()
defer mu.Unlock()
if _, ok := known[pkgExport.Fix.StmtInfo.ImportPath]; ok {
return // We got this one above.
}
// Continue with untyped proposals.
pkg := types.NewPackage(pkgExport.Fix.StmtInfo.ImportPath, pkgExport.Fix.IdentName)
for _, export := range pkgExport.Exports {
score := stdScore + 0.1*float64(pkgExport.Fix.Relevance)
c.found(ctx, candidate{
obj: types.NewVar(0, pkg, export, nil),
score: score,
imp: &importInfo{
importPath: pkgExport.Fix.StmtInfo.ImportPath,
name: pkgExport.Fix.StmtInfo.Name,
},
})
}
if len(c.items) >= unimportedMemberTarget {
cancel()
}
}
return c.snapshot.View().RunProcessEnvFunc(ctx, func(opts *imports.Options) error {
return imports.GetPackageExports(ctx, add, id.Name, c.filename, c.pkg.GetTypes().Name(), opts)
})
}
func (c *completer) packageMembers(ctx context.Context, pkg *types.Package, score float64, imp *importInfo) {
scope := pkg.Scope()
for _, name := range scope.Names() {
obj := scope.Lookup(name)
c.found(ctx, candidate{
obj: obj,
score: score,
imp: imp,
addressable: isVar(obj),
})
}
}
func (c *completer) methodsAndFields(ctx context.Context, typ types.Type, addressable bool, imp *importInfo) error {
internal/lsp: speed up deep completion search 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>
2019-08-16 10:45:09 -06:00
mset := c.methodSetCache[methodSetKey{typ, addressable}]
if mset == nil {
if addressable && !types.IsInterface(typ) && !isPointer(typ) {
// Add methods of *T, which includes methods with receiver T.
mset = types.NewMethodSet(types.NewPointer(typ))
} else {
// Add methods of T.
mset = types.NewMethodSet(typ)
}
c.methodSetCache[methodSetKey{typ, addressable}] = mset
}
for i := 0; i < mset.Len(); i++ {
c.found(ctx, candidate{
obj: mset.At(i).Obj(),
score: stdScore,
imp: imp,
addressable: addressable || isPointer(typ),
})
}
// Add fields of T.
internal/lsp/source: optimize enumeration of a type's fields When searching for deep completions, we can end up enumerating struct types' fields a lot. Optimize fieldSelections to reduce work: - Wait until we see an embedded field before we create the "seen" map. - Use a callback style to iterate over the struct's fields rather than returning a slice of fields. - Change "seen" checking strategy back to track struct types rather than each individual field. Struct with 5 non-embedded fields: name old time/op new time/op delta Fields-16 293ns ± 1% 20ns ± 2% -93.13% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 0B -100.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 0.00 -100.00% (p=0.008 n=5+5) Same struct but add an embedded struct with 2 fields: name old time/op new time/op delta Fields-16 389ns ± 1% 142ns ± 1% -63.53% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 144B ± 0% +20.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.008 n=5+5) I think the alloc/op went up because the "seen" map is no longer allocated on the stack. There is more room for more optimization, but it's probably not worth making things more complicated. Change-Id: I6f9f2124334a8594ef9d6f9b5ac4b3a8aead5f49 Reviewed-on: https://go-review.googlesource.com/c/tools/+/223419 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2020-03-13 13:10:42 -06:00
eachField(typ, func(v *types.Var) {
c.found(ctx, candidate{
internal/lsp/source: optimize enumeration of a type's fields When searching for deep completions, we can end up enumerating struct types' fields a lot. Optimize fieldSelections to reduce work: - Wait until we see an embedded field before we create the "seen" map. - Use a callback style to iterate over the struct's fields rather than returning a slice of fields. - Change "seen" checking strategy back to track struct types rather than each individual field. Struct with 5 non-embedded fields: name old time/op new time/op delta Fields-16 293ns ± 1% 20ns ± 2% -93.13% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 0B -100.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 0.00 -100.00% (p=0.008 n=5+5) Same struct but add an embedded struct with 2 fields: name old time/op new time/op delta Fields-16 389ns ± 1% 142ns ± 1% -63.53% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 144B ± 0% +20.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.008 n=5+5) I think the alloc/op went up because the "seen" map is no longer allocated on the stack. There is more room for more optimization, but it's probably not worth making things more complicated. Change-Id: I6f9f2124334a8594ef9d6f9b5ac4b3a8aead5f49 Reviewed-on: https://go-review.googlesource.com/c/tools/+/223419 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2020-03-13 13:10:42 -06:00
obj: v,
score: stdScore - 0.01,
imp: imp,
addressable: addressable || isPointer(typ),
})
internal/lsp/source: optimize enumeration of a type's fields When searching for deep completions, we can end up enumerating struct types' fields a lot. Optimize fieldSelections to reduce work: - Wait until we see an embedded field before we create the "seen" map. - Use a callback style to iterate over the struct's fields rather than returning a slice of fields. - Change "seen" checking strategy back to track struct types rather than each individual field. Struct with 5 non-embedded fields: name old time/op new time/op delta Fields-16 293ns ± 1% 20ns ± 2% -93.13% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 0B -100.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 0.00 -100.00% (p=0.008 n=5+5) Same struct but add an embedded struct with 2 fields: name old time/op new time/op delta Fields-16 389ns ± 1% 142ns ± 1% -63.53% (p=0.008 n=5+5) name old alloc/op new alloc/op delta Fields-16 120B ± 0% 144B ± 0% +20.00% (p=0.008 n=5+5) name old allocs/op new allocs/op delta Fields-16 4.00 ± 0% 2.00 ± 0% -50.00% (p=0.008 n=5+5) I think the alloc/op went up because the "seen" map is no longer allocated on the stack. There is more room for more optimization, but it's probably not worth making things more complicated. Change-Id: I6f9f2124334a8594ef9d6f9b5ac4b3a8aead5f49 Reviewed-on: https://go-review.googlesource.com/c/tools/+/223419 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2020-03-13 13:10:42 -06:00
})
return nil
}
// lexical finds completions in the lexical environment.
func (c *completer) lexical(ctx context.Context) error {
var scopes []*types.Scope // scopes[i], where i<len(path), is the possibly nil Scope of path[i].
for _, n := range c.path {
// Include *FuncType scope if pos is inside the function body.
switch node := n.(type) {
case *ast.FuncDecl:
if node.Body != nil && nodeContains(node.Body, c.pos) {
n = node.Type
}
case *ast.FuncLit:
if node.Body != nil && nodeContains(node.Body, c.pos) {
n = node.Type
}
}
scopes = append(scopes, c.pkg.GetTypesInfo().Scopes[n])
}
scopes = append(scopes, c.pkg.GetTypes().Scope(), types.Universe)
var (
builtinIota = types.Universe.Lookup("iota")
builtinNil = types.Universe.Lookup("nil")
)
// Track seen variables to avoid showing completions for shadowed variables.
// This works since we look at scopes from innermost to outermost.
seen := make(map[string]struct{})
// Process scopes innermost first.
for i, scope := range scopes {
if scope == nil {
continue
}
Names:
for _, name := range scope.Names() {
declScope, obj := scope.LookupParent(name, c.pos)
if declScope != scope {
continue // Name was declared in some enclosing scope, or not at all.
}
// If obj's type is invalid, find the AST node that defines the lexical block
// containing the declaration of obj. Don't resolve types for packages.
if !isPkgName(obj) && !typeIsValid(obj.Type()) {
// Match the scope to its ast.Node. If the scope is the package scope,
// use the *ast.File as the starting node.
var node ast.Node
if i < len(c.path) {
node = c.path[i]
} else if i == len(c.path) { // use the *ast.File for package scope
node = c.path[i-1]
}
if node != nil {
fset := c.snapshot.View().Session().Cache().FileSet()
if resolved := resolveInvalid(fset, obj, node, c.pkg.GetTypesInfo()); resolved != nil {
obj = resolved
}
}
}
// Don't use LHS of value spec in RHS.
if vs := enclosingValueSpec(c.path); vs != nil {
for _, ident := range vs.Names {
if obj.Pos() == ident.Pos() {
continue Names
}
}
}
// Don't suggest "iota" outside of const decls.
if obj == builtinIota && !c.inConstDecl() {
continue
}
// Rank outer scopes lower than inner.
score := stdScore * math.Pow(.99, float64(i))
// Dowrank "nil" a bit so it is ranked below more interesting candidates.
if obj == builtinNil {
score /= 2
}
// If we haven't already added a candidate for an object with this name.
if _, ok := seen[obj.Name()]; !ok {
seen[obj.Name()] = struct{}{}
c.found(ctx, candidate{
obj: obj,
score: score,
addressable: isVar(obj),
})
}
}
}
if c.inference.objType != nil {
if named, _ := deref(c.inference.objType).(*types.Named); named != nil {
// If we expected a named type, check the type's package for
// completion items. This is useful when the current file hasn't
// imported the type's package yet.
if named.Obj() != nil && named.Obj().Pkg() != nil {
pkg := named.Obj().Pkg()
// Make sure the package name isn't already in use by another
// object, and that this file doesn't import the package yet.
if _, ok := seen[pkg.Name()]; !ok && pkg != c.pkg.GetTypes() && !alreadyImports(c.file, pkg.Path()) {
seen[pkg.Name()] = struct{}{}
obj := types.NewPkgName(0, nil, pkg.Name(), pkg)
imp := &importInfo{
importPath: pkg.Path(),
}
if imports.ImportPathToAssumedName(pkg.Path()) != pkg.Name() {
imp.name = pkg.Name()
}
c.found(ctx, candidate{
obj: obj,
score: stdScore,
imp: imp,
})
}
}
}
}
if c.opts.unimported {
if err := c.unimportedPackages(ctx, seen); err != nil {
return err
}
}
if t := c.inference.objType; t != nil {
// Use variadic element type if we are completing variadic position.
if c.inference.variadicType != nil {
t = c.inference.variadicType
}
t = deref(t)
// If we have an expected type and it is _not_ a named type, see
// if an object literal makes a good candidate. For example, if
// our expected type is "[]int", this will add a candidate of
// "[]int{}".
if _, named := t.(*types.Named); !named {
c.literal(ctx, t, nil)
}
}
// Add keyword completion items appropriate in the current context.
c.addKeywordCompletions()
return nil
}
func (c *completer) unimportedPackages(ctx context.Context, seen map[string]struct{}) error {
var prefix string
if c.surrounding != nil {
prefix = c.surrounding.Prefix()
}
initialItemCount := len(c.items)
known, err := c.snapshot.CachedImportPaths(ctx)
if err != nil {
return err
}
var paths []string
for path := range known {
if !strings.HasPrefix(path, prefix) {
continue
}
paths = append(paths, path)
}
var relevances map[string]int
if len(paths) != 0 {
c.snapshot.View().RunProcessEnvFunc(ctx, func(opts *imports.Options) error {
relevances = imports.ScoreImportPaths(ctx, opts.Env, paths)
return nil
})
}
for path, relevance := range relevances {
pkg := known[path]
imp := &importInfo{
importPath: path,
pkg: pkg,
}
if imports.ImportPathToAssumedName(path) != pkg.GetTypes().Name() {
imp.name = pkg.GetTypes().Name()
}
score := 0.01 * float64(relevance)
c.found(ctx, candidate{
obj: types.NewPkgName(0, nil, pkg.GetTypes().Name(), pkg.GetTypes()),
score: score,
imp: imp,
})
if len(c.items)-initialItemCount >= maxUnimportedPackageNames {
return nil
}
}
ctx, cancel := context.WithCancel(ctx)
defer cancel()
var mu sync.Mutex
add := func(pkg imports.ImportFix) {
mu.Lock()
defer mu.Unlock()
if _, ok := seen[pkg.IdentName]; ok {
return
}
if len(c.items)-initialItemCount >= maxUnimportedPackageNames {
cancel()
return
}
// Rank unimported packages significantly lower than other results.
score := 0.01 * float64(pkg.Relevance)
// Do not add the unimported packages to seen, since we can have
// multiple packages of the same name as completion suggestions, since
// only one will be chosen.
obj := types.NewPkgName(0, nil, pkg.IdentName, types.NewPackage(pkg.StmtInfo.ImportPath, pkg.IdentName))
c.found(ctx, candidate{
obj: obj,
score: score,
imp: &importInfo{
importPath: pkg.StmtInfo.ImportPath,
name: pkg.StmtInfo.Name,
},
})
}
return c.snapshot.View().RunProcessEnvFunc(ctx, func(opts *imports.Options) error {
return imports.GetAllCandidates(ctx, add, prefix, c.filename, c.pkg.GetTypes().Name(), opts)
})
}
// alreadyImports reports whether f has an import with the specified path.
func alreadyImports(f *ast.File, path string) bool {
for _, s := range f.Imports {
if importPath(s) == path {
return true
}
}
return false
}
// importPath returns the unquoted import path of s,
// or "" if the path is not properly quoted.
func importPath(s *ast.ImportSpec) string {
t, err := strconv.Unquote(s.Path.Value)
if err != nil {
return ""
}
return t
}
func nodeContains(n ast.Node, pos token.Pos) bool {
internal/lsp: improve completion after accidental keywords Sometimes the prefix of the thing you want to complete is a keyword. For example: variance := 123 fmt.Println(var<>) In this case the parser produces an *ast.BadExpr which breaks completion. We now repair this BadExpr by replacing it with an *ast.Ident named "var". We also repair empty decls using a similar approach. This fixes cases like: var typeName string type<> // want to complete to "typeName" We also fix accidental keywords in selectors, such as: foo.var<> The parser produces a phantom "_" in place of the keyword, so we swap it back for an *ast.Ident named "var". In general, though, accidental keywords wreak havoc on the AST so we can only do so much. There are still many cases where a keyword prefix breaks completion. Perhaps in the future the parser can be cursor/in-progress-edit aware and turn accidental keywords into identifiers. Fixes golang/go#34332. PS I tweaked nodeContains() to include n.End() to fix a test failure against tip related to a change to go/parser. When a syntax error is present, an *ast.BlockStmt's End() is now set to the block's final statement's End() (earlier than what it used to be). In order for the cursor pos to test "inside" the block in this case I had to relax the End() comparison. Change-Id: Ib45952cf086cc974f1578298df3dd12829344faa Reviewed-on: https://go-review.googlesource.com/c/tools/+/209438 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Rebecca Stambler <rstambler@golang.org>
2019-11-20 12:15:53 -07:00
return n != nil && n.Pos() <= pos && pos <= n.End()
}
func (c *completer) inConstDecl() bool {
for _, n := range c.path {
if decl, ok := n.(*ast.GenDecl); ok && decl.Tok == token.CONST {
return true
}
}
return false
}
// structLiteralFieldName finds completions for struct field names inside a struct literal.
func (c *completer) structLiteralFieldName(ctx context.Context) error {
clInfo := c.enclosingCompositeLiteral
// Mark fields of the composite literal that have already been set,
// except for the current field.
addedFields := make(map[*types.Var]bool)
for _, el := range clInfo.cl.Elts {
if kvExpr, ok := el.(*ast.KeyValueExpr); ok {
if clInfo.kv == kvExpr {
continue
}
if key, ok := kvExpr.Key.(*ast.Ident); ok {
if used, ok := c.pkg.GetTypesInfo().Uses[key]; ok {
if usedVar, ok := used.(*types.Var); ok {
addedFields[usedVar] = true
}
}
}
}
}
switch t := clInfo.clType.(type) {
case *types.Struct:
for i := 0; i < t.NumFields(); i++ {
field := t.Field(i)
if !addedFields[field] {
c.found(ctx, candidate{
obj: field,
score: highScore,
})
}
}
// Add lexical completions if we aren't certain we are in the key part of a
// key-value pair.
if clInfo.maybeInFieldName {
return c.lexical(ctx)
}
default:
return c.lexical(ctx)
}
return nil
}
func (cl *compLitInfo) isStruct() bool {
_, ok := cl.clType.(*types.Struct)
return ok
}
// enclosingCompositeLiteral returns information about the composite literal enclosing the
// position.
func enclosingCompositeLiteral(path []ast.Node, pos token.Pos, info *types.Info) *compLitInfo {
for _, n := range path {
switch n := n.(type) {
case *ast.CompositeLit:
// The enclosing node will be a composite literal if the user has just
// opened the curly brace (e.g. &x{<>) or the completion request is triggered
// from an already completed composite literal expression (e.g. &x{foo: 1, <>})
//
// The position is not part of the composite literal unless it falls within the
// curly braces (e.g. "foo.Foo<>Struct{}").
if !(n.Lbrace < pos && pos <= n.Rbrace) {
// Keep searching since we may yet be inside a composite literal.
// For example "Foo{B: Ba<>{}}".
break
}
tv, ok := info.Types[n]
if !ok {
return nil
}
clInfo := compLitInfo{
cl: n,
clType: deref(tv.Type).Underlying(),
}
var (
expr ast.Expr
hasKeys bool
)
for _, el := range n.Elts {
// Remember the expression that the position falls in, if any.
if el.Pos() <= pos && pos <= el.End() {
expr = el
}
if kv, ok := el.(*ast.KeyValueExpr); ok {
hasKeys = true
// If expr == el then we know the position falls in this expression,
// so also record kv as the enclosing *ast.KeyValueExpr.
if expr == el {
clInfo.kv = kv
break
}
}
}
if clInfo.kv != nil {
// If in a *ast.KeyValueExpr, we know we are in the key if the position
// is to the left of the colon (e.g. "Foo{F<>: V}".
clInfo.inKey = pos <= clInfo.kv.Colon
} else if hasKeys {
// If we aren't in a *ast.KeyValueExpr but the composite literal has
// other *ast.KeyValueExprs, we must be on the key side of a new
// *ast.KeyValueExpr (e.g. "Foo{F: V, <>}").
clInfo.inKey = true
} else {
switch clInfo.clType.(type) {
case *types.Struct:
if len(n.Elts) == 0 {
// If the struct literal is empty, next could be a struct field
// name or an expression (e.g. "Foo{<>}" could become "Foo{F:}"
// or "Foo{someVar}").
clInfo.maybeInFieldName = true
} else if len(n.Elts) == 1 {
// If there is one expression and the position is in that expression
// and the expression is an identifier, we may be writing a field
// name or an expression (e.g. "Foo{F<>}").
_, clInfo.maybeInFieldName = expr.(*ast.Ident)
}
case *types.Map:
// If we aren't in a *ast.KeyValueExpr we must be adding a new key
// to the map.
clInfo.inKey = true
}
}
return &clInfo
default:
if breaksExpectedTypeInference(n) {
return nil
}
}
}
return nil
}
// enclosingFunction returns the signature and body of the function
// enclosing the given position.
func enclosingFunction(path []ast.Node, info *types.Info) *funcInfo {
for _, node := range path {
switch t := node.(type) {
case *ast.FuncDecl:
if obj, ok := info.Defs[t.Name]; ok {
return &funcInfo{
sig: obj.Type().(*types.Signature),
body: t.Body,
}
}
case *ast.FuncLit:
if typ, ok := info.Types[t]; ok {
return &funcInfo{
sig: typ.Type.(*types.Signature),
body: t.Body,
}
}
}
}
return nil
}
func (c *completer) expectedCompositeLiteralType() types.Type {
clInfo := c.enclosingCompositeLiteral
switch t := clInfo.clType.(type) {
case *types.Slice:
if clInfo.inKey {
return types.Typ[types.Int]
}
return t.Elem()
case *types.Array:
if clInfo.inKey {
return types.Typ[types.Int]
}
return t.Elem()
case *types.Map:
if clInfo.inKey {
return t.Key()
}
return t.Elem()
case *types.Struct:
// If we are completing a key (i.e. field name), there is no expected type.
if clInfo.inKey {
return nil
}
// If we are in a key-value pair, but not in the key, then we must be on the
// value side. The expected type of the value will be determined from the key.
if clInfo.kv != nil {
if key, ok := clInfo.kv.Key.(*ast.Ident); ok {
for i := 0; i < t.NumFields(); i++ {
if field := t.Field(i); field.Name() == key.Name {
return field.Type()
}
}
}
} else {
// If we aren't in a key-value pair and aren't in the key, we must be using
// implicit field names.
// The order of the literal fields must match the order in the struct definition.
// Find the element that the position belongs to and suggest that field's type.
if i := exprAtPos(c.pos, clInfo.cl.Elts); i < t.NumFields() {
return t.Field(i).Type()
}
}
}
return nil
}
// typeModifier represents an operator that changes the expected type.
type typeModifier struct {
mod typeMod
arrayLen int64
}
type typeMod int
const (
star typeMod = iota // pointer indirection for expressions, pointer indicator for types
address // address operator ("&")
chanRead // channel read operator ("<-")
slice // make a slice type ("[]" in "[]int")
array // make an array type ("[2]" in "[2]int")
)
type objKind int
const (
kindArray objKind = 1 << iota
kindSlice
kindChan
kindMap
kindStruct
kindString
)
// candidateInference holds information we have inferred about a type that can be
// used at the current position.
type candidateInference struct {
// objType is the desired type of an object used at the query position.
objType types.Type
// objKind is a mask of expected kinds of types such as "map", "slice", etc.
objKind objKind
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// variadicType is the scalar variadic element type. For example,
// when completing "append([]T{}, <>)" objType is []T and
// variadicType is T.
variadicType types.Type
// modifiers are prefixes such as "*", "&" or "<-" that influence how
// a candidate type relates to the expected type.
modifiers []typeModifier
// convertibleTo is a type our candidate type must be convertible to.
convertibleTo types.Type
// typeName holds information about the expected type name at
// position, if any.
typeName typeNameInference
// assignees are the types that would receive a function call's
// results at the position. For example:
//
// foo := 123
// foo, bar := <>
//
// at "<>", the assignees are [int, <invalid>].
assignees []types.Type
// variadicAssignees is true if we could be completing an inner
// function call that fills out an outer function call's variadic
// params. For example:
//
// func foo(int, ...string) {}
//
// foo(<>) // variadicAssignees=true
// foo(bar<>) // variadicAssignees=true
// foo(bar, baz<>) // variadicAssignees=false
variadicAssignees bool
}
// typeNameInference holds information about the expected type name at
// position.
type typeNameInference struct {
// wantTypeName is true if we expect the name of a type.
wantTypeName bool
// modifiers are prefixes such as "*", "&" or "<-" that influence how
// a candidate type relates to the expected type.
modifiers []typeModifier
// assertableFrom is a type that must be assertable to our candidate type.
assertableFrom types.Type
// wantComparable is true if we want a comparable type.
wantComparable bool
}
// expectedCandidate returns information about the expected candidate
// for an expression at the query position.
func expectedCandidate(ctx context.Context, c *completer) (inf candidateInference) {
inf.typeName = expectTypeName(c)
if c.enclosingCompositeLiteral != nil {
inf.objType = c.expectedCompositeLiteralType()
}
Nodes:
for i, node := range c.path {
switch node := node.(type) {
case *ast.BinaryExpr:
// Determine if query position comes from left or right of op.
e := node.X
if c.pos < node.OpPos {
e = node.Y
}
if tv, ok := c.pkg.GetTypesInfo().Types[e]; ok {
inf.objType = tv.Type
break Nodes
}
case *ast.AssignStmt:
// Only rank completions if you are on the right side of the token.
if c.pos > node.TokPos {
i := exprAtPos(c.pos, node.Rhs)
if i >= len(node.Lhs) {
i = len(node.Lhs) - 1
}
if tv, ok := c.pkg.GetTypesInfo().Types[node.Lhs[i]]; ok {
inf.objType = tv.Type
}
// If we have a single expression on the RHS, record the LHS
// assignees so we can favor multi-return function calls with
// matching result values.
if len(node.Rhs) <= 1 {
for _, lhs := range node.Lhs {
inf.assignees = append(inf.assignees, c.pkg.GetTypesInfo().TypeOf(lhs))
}
} else {
// Otherwse, record our single assignee, even if its type is
// not available. We use this info to downrank functions
// with the wrong number of result values.
inf.assignees = append(inf.assignees, c.pkg.GetTypesInfo().TypeOf(node.Lhs[i]))
}
}
return inf
case *ast.ValueSpec:
if node.Type != nil && c.pos > node.Type.End() {
inf.objType = c.pkg.GetTypesInfo().TypeOf(node.Type)
}
return inf
case *ast.CallExpr:
// Only consider CallExpr args if position falls between parens.
if node.Lparen <= c.pos && c.pos <= node.Rparen {
// For type conversions like "int64(foo)" we can only infer our
// desired type is convertible to int64.
if typ := typeConversion(node, c.pkg.GetTypesInfo()); typ != nil {
inf.convertibleTo = typ
break Nodes
}
if tv, ok := c.pkg.GetTypesInfo().Types[node.Fun]; ok {
if sig, ok := tv.Type.(*types.Signature); ok {
numParams := sig.Params().Len()
if numParams == 0 {
return inf
}
var (
exprIdx = exprAtPos(c.pos, node.Args)
isLastParam = exprIdx == numParams-1
beyondLastParam = exprIdx >= numParams
)
// If we have one or zero arg expressions, we may be
// completing to a function call that returns multiple
// values, in turn getting passed in to the surrounding
// call. Record the assignees so we can favor function
// calls that return matching values.
if len(node.Args) <= 1 {
for i := 0; i < sig.Params().Len(); i++ {
inf.assignees = append(inf.assignees, sig.Params().At(i).Type())
}
// Record that we may be completing into variadic parameters.
inf.variadicAssignees = sig.Variadic()
}
if sig.Variadic() {
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
variadicType := deslice(sig.Params().At(numParams - 1).Type())
// If we are beyond the last param or we are the last
// param w/ further expressions, we expect a single
// variadic item.
if beyondLastParam || isLastParam && len(node.Args) > numParams {
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
inf.objType = variadicType
break Nodes
}
// Otherwise if we are at the last param then we are
// completing the variadic positition (i.e. we expect a
// slice type []T or an individual item T).
if isLastParam {
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
inf.variadicType = variadicType
}
}
// Make sure not to run past the end of expected parameters.
if beyondLastParam {
inf.objType = sig.Params().At(numParams - 1).Type()
} else {
inf.objType = sig.Params().At(exprIdx).Type()
}
}
}
if funIdent, ok := node.Fun.(*ast.Ident); ok {
obj := c.pkg.GetTypesInfo().ObjectOf(funIdent)
if obj != nil && obj.Parent() == types.Universe {
// Defer call to builtinArgType so we can provide it the
// inferred type from its parent node.
defer func() {
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
inf = c.builtinArgType(obj, node, inf)
inf.objKind = c.builtinArgKind(ctx, obj, node)
}()
// The expected type of builtin arguments like append() is
// the expected type of the builtin call itself. For
// example:
//
// var foo []int = append(<>)
//
// To find the expected type at <> we "skip" the append()
// node and get the expected type one level up, which is
// []int.
continue Nodes
}
}
}
return inf
case *ast.ReturnStmt:
if c.enclosingFunc != nil {
sig := c.enclosingFunc.sig
// Find signature result that corresponds to our return statement.
if resultIdx := exprAtPos(c.pos, node.Results); resultIdx < len(node.Results) {
if resultIdx < sig.Results().Len() {
inf.objType = sig.Results().At(resultIdx).Type()
}
}
}
return inf
case *ast.CaseClause:
if swtch, ok := findSwitchStmt(c.path[i+1:], c.pos, node).(*ast.SwitchStmt); ok {
if tv, ok := c.pkg.GetTypesInfo().Types[swtch.Tag]; ok {
inf.objType = tv.Type
}
}
return inf
case *ast.SliceExpr:
// Make sure position falls within the brackets (e.g. "foo[a:<>]").
if node.Lbrack < c.pos && c.pos <= node.Rbrack {
inf.objType = types.Typ[types.Int]
}
return inf
case *ast.IndexExpr:
// Make sure position falls within the brackets (e.g. "foo[<>]").
if node.Lbrack < c.pos && c.pos <= node.Rbrack {
if tv, ok := c.pkg.GetTypesInfo().Types[node.X]; ok {
switch t := tv.Type.Underlying().(type) {
case *types.Map:
inf.objType = t.Key()
case *types.Slice, *types.Array:
inf.objType = types.Typ[types.Int]
}
}
}
return inf
case *ast.SendStmt:
// Make sure we are on right side of arrow (e.g. "foo <- <>").
if c.pos > node.Arrow+1 {
if tv, ok := c.pkg.GetTypesInfo().Types[node.Chan]; ok {
if ch, ok := tv.Type.Underlying().(*types.Chan); ok {
inf.objType = ch.Elem()
}
}
}
return inf
case *ast.RangeStmt:
if nodeContains(node.X, c.pos) {
inf.objKind |= kindSlice | kindArray | kindMap | kindString
if node.Value == nil {
inf.objKind |= kindChan
}
}
return inf
case *ast.StarExpr:
inf.modifiers = append(inf.modifiers, typeModifier{mod: star})
case *ast.UnaryExpr:
switch node.Op {
case token.AND:
inf.modifiers = append(inf.modifiers, typeModifier{mod: address})
case token.ARROW:
inf.modifiers = append(inf.modifiers, typeModifier{mod: chanRead})
}
default:
if breaksExpectedTypeInference(node) {
return inf
}
}
}
return inf
}
// applyTypeModifiers applies the list of type modifiers to a type.
// It returns nil if the modifiers could not be applied.
func (ci candidateInference) applyTypeModifiers(typ types.Type, addressable bool) types.Type {
for _, mod := range ci.modifiers {
switch mod.mod {
case star:
// For every "*" indirection operator, remove a pointer layer
// from candidate type.
if ptr, ok := typ.Underlying().(*types.Pointer); ok {
typ = ptr.Elem()
} else {
return nil
}
case address:
// For every "&" address operator, add another pointer layer to
// candidate type, if the candidate is addressable.
if addressable {
typ = types.NewPointer(typ)
} else {
return nil
}
case chanRead:
// For every "<-" operator, remove a layer of channelness.
if ch, ok := typ.(*types.Chan); ok {
typ = ch.Elem()
} else {
return nil
}
}
}
return typ
}
// applyTypeNameModifiers applies the list of type modifiers to a type name.
func (ci candidateInference) applyTypeNameModifiers(typ types.Type) types.Type {
for _, mod := range ci.typeName.modifiers {
switch mod.mod {
case star:
// For every "*" indicator, add a pointer layer to type name.
typ = types.NewPointer(typ)
case array:
typ = types.NewArray(typ, mod.arrayLen)
case slice:
typ = types.NewSlice(typ)
}
}
return typ
}
// matchesVariadic returns true if we are completing a variadic
// parameter and candType is a compatible slice type.
func (ci candidateInference) matchesVariadic(candType types.Type) bool {
return ci.variadicType != nil && types.AssignableTo(candType, ci.objType)
}
// findSwitchStmt returns an *ast.CaseClause's corresponding *ast.SwitchStmt or
// *ast.TypeSwitchStmt. path should start from the case clause's first ancestor.
func findSwitchStmt(path []ast.Node, pos token.Pos, c *ast.CaseClause) ast.Stmt {
// Make sure position falls within a "case <>:" clause.
if exprAtPos(pos, c.List) >= len(c.List) {
return nil
}
// A case clause is always nested within a block statement in a switch statement.
if len(path) < 2 {
return nil
}
if _, ok := path[0].(*ast.BlockStmt); !ok {
return nil
}
switch s := path[1].(type) {
case *ast.SwitchStmt:
return s
case *ast.TypeSwitchStmt:
return s
default:
return nil
}
}
// breaksExpectedTypeInference reports if an expression node's type is unrelated
// to its child expression node types. For example, "Foo{Bar: x.Baz(<>)}" should
// expect a function argument, not a composite literal value.
func breaksExpectedTypeInference(n ast.Node) bool {
switch n.(type) {
case *ast.FuncLit, *ast.CallExpr, *ast.IndexExpr, *ast.SliceExpr, *ast.CompositeLit:
return true
default:
return false
}
}
// expectTypeName returns information about the expected type name at position.
func expectTypeName(c *completer) typeNameInference {
var (
wantTypeName bool
wantComparable bool
modifiers []typeModifier
assertableFrom types.Type
)
Nodes:
for i, p := range c.path {
switch n := p.(type) {
case *ast.FieldList:
// Expect a type name if pos is in a FieldList. This applies to
// FuncType params/results, FuncDecl receiver, StructType, and
// InterfaceType. We don't need to worry about the field name
// because completion bails out early if pos is in an *ast.Ident
// that defines an object.
wantTypeName = true
break Nodes
case *ast.CaseClause:
// Expect type names in type switch case clauses.
if swtch, ok := findSwitchStmt(c.path[i+1:], c.pos, n).(*ast.TypeSwitchStmt); ok {
// The case clause types must be assertable from the type switch parameter.
ast.Inspect(swtch.Assign, func(n ast.Node) bool {
if ta, ok := n.(*ast.TypeAssertExpr); ok {
assertableFrom = c.pkg.GetTypesInfo().TypeOf(ta.X)
return false
}
return true
})
wantTypeName = true
break Nodes
}
return typeNameInference{}
case *ast.TypeAssertExpr:
// Expect type names in type assert expressions.
if n.Lparen < c.pos && c.pos <= n.Rparen {
// The type in parens must be assertable from the expression type.
assertableFrom = c.pkg.GetTypesInfo().TypeOf(n.X)
wantTypeName = true
break Nodes
}
return typeNameInference{}
case *ast.StarExpr:
modifiers = append(modifiers, typeModifier{mod: star})
case *ast.CompositeLit:
// We want a type name if position is in the "Type" part of a
// composite literal (e.g. "Foo<>{}").
if n.Type != nil && n.Type.Pos() <= c.pos && c.pos <= n.Type.End() {
wantTypeName = true
}
break Nodes
case *ast.ArrayType:
// If we are inside the "Elt" part of an array type, we want a type name.
if n.Elt.Pos() <= c.pos && c.pos <= n.Elt.End() {
wantTypeName = true
if n.Len == nil {
// No "Len" expression means a slice type.
modifiers = append(modifiers, typeModifier{mod: slice})
} else {
// Try to get the array type using the constant value of "Len".
tv, ok := c.pkg.GetTypesInfo().Types[n.Len]
if ok && tv.Value != nil && tv.Value.Kind() == constant.Int {
if arrayLen, ok := constant.Int64Val(tv.Value); ok {
modifiers = append(modifiers, typeModifier{mod: array, arrayLen: arrayLen})
}
}
}
// ArrayTypes can be nested, so keep going if our parent is an
// ArrayType.
if i < len(c.path)-1 {
if _, ok := c.path[i+1].(*ast.ArrayType); ok {
continue Nodes
}
}
break Nodes
}
case *ast.MapType:
wantTypeName = true
if n.Key != nil {
wantComparable = nodeContains(n.Key, c.pos)
} else {
// If the key is empty, assume we are completing the key if
// pos is directly after the "map[".
wantComparable = c.pos == n.Pos()+token.Pos(len("map["))
}
break Nodes
case *ast.ValueSpec:
wantTypeName = nodeContains(n.Type, c.pos)
break Nodes
case *ast.TypeSpec:
wantTypeName = nodeContains(n.Type, c.pos)
default:
if breaksExpectedTypeInference(p) {
return typeNameInference{}
}
}
}
return typeNameInference{
wantTypeName: wantTypeName,
wantComparable: wantComparable,
modifiers: modifiers,
assertableFrom: assertableFrom,
}
}
func (c *completer) fakeObj(T types.Type) *types.Var {
return types.NewVar(token.NoPos, c.pkg.GetTypes(), "", T)
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// anyCandType reports whether f returns true for any candidate type
// derivable from c. For example, from "foo" we might derive "&foo",
// and "foo()".
func (c *candidate) anyCandType(f func(t types.Type, addressable bool) bool) bool {
if c.obj == nil || c.obj.Type() == nil {
return false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
objType := c.obj.Type()
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if f(objType, c.addressable) {
return true
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// If c is a func type with a single result, offer the result type.
if sig, ok := objType.Underlying().(*types.Signature); ok {
if sig.Results().Len() == 1 && f(sig.Results().At(0).Type(), false) {
// Mark the candidate so we know to append "()" when formatting.
c.expandFuncCall = true
return true
}
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
var (
seenPtrTypes map[types.Type]bool
ptrType = objType
ptrDepth int
)
// Check if dereferencing c would match our type inference. We loop
// since c could have arbitrary levels of pointerness.
for {
ptr, ok := ptrType.Underlying().(*types.Pointer)
if !ok {
break
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
ptrDepth++
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// Avoid pointer type cycles.
if seenPtrTypes[ptrType] {
break
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if _, named := ptrType.(*types.Named); named {
// Lazily allocate "seen" since it isn't used normally.
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if seenPtrTypes == nil {
seenPtrTypes = make(map[types.Type]bool)
}
// Track named pointer types we have seen to detect cycles.
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
seenPtrTypes[ptrType] = true
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if f(ptr.Elem(), false) {
// Mark the candidate so we know to prepend "*" when formatting.
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
c.dereference = ptrDepth
return true
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
ptrType = ptr.Elem()
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// Check if c is addressable and a pointer to c matches our type inference.
if c.addressable && f(types.NewPointer(objType), false) {
// Mark the candidate so we know to prepend "&" when formatting.
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
c.takeAddress = true
return true
}
return false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// matchingCandidate reports whether cand matches our type inferences.
func (c *completer) matchingCandidate(cand *candidate) bool {
if isTypeName(cand.obj) {
return c.matchingTypeName(cand)
} else if c.wantTypeName() {
// If we want a type, a non-type object never matches.
return false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if c.inference.candTypeMatches(cand) {
return true
}
candType := cand.obj.Type()
if candType == nil {
return false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if sig, ok := candType.Underlying().(*types.Signature); ok {
if c.inference.assigneesMatch(cand, sig) {
// Invoke the candidate if its results are multi-assignable.
cand.expandFuncCall = true
return true
}
}
// Default to invoking *types.Func candidates. This is so function
// completions in an empty statement (or other cases with no expected type)
// are invoked by default.
cand.expandFuncCall = isFunc(cand.obj)
return false
}
// candTypeMatches reports whether cand makes a good completion
// candidate given the candidate inference. cand's score may be
// mutated to downrank the candidate in certain situations.
func (ci *candidateInference) candTypeMatches(cand *candidate) bool {
expTypes := make([]types.Type, 0, 2)
if ci.objType != nil {
expTypes = append(expTypes, ci.objType)
}
if ci.variadicType != nil {
expTypes = append(expTypes, ci.variadicType)
}
return cand.anyCandType(func(candType types.Type, addressable bool) bool {
// Take into account any type modifiers on the expected type.
candType = ci.applyTypeModifiers(candType, addressable)
if candType == nil {
return false
}
if ci.convertibleTo != nil && types.ConvertibleTo(candType, ci.convertibleTo) {
return true
}
if len(expTypes) == 0 {
// If we have no expected type but were able to apply type
// modifiers to our candidate type, count that as a match. This
// handles cases like:
//
// var foo chan int
// <-fo<>
//
// There is no exected type at "<>", but we were able to apply
// the "<-" type modifier to "foo", so it matches.
if len(ci.modifiers) > 0 {
return true
}
// If we have no expected type, fall back to checking the
// expected "kind" of object, if available.
return ci.kindMatches(candType)
}
for _, expType := range expTypes {
matches, untyped := ci.typeMatches(expType, candType)
if !matches {
continue
}
// Lower candidate score for untyped conversions. This avoids
// ranking untyped constants above candidates with an exact type
// match. Don't lower score of builtin constants, e.g. "true".
if untyped && !types.Identical(candType, expType) && cand.obj.Parent() != types.Universe {
cand.score /= 2
}
return true
}
return false
})
}
// typeMatches reports whether an object of candType makes a good
// completion candidate given the expected type expType. It also
// returns a second bool which is true if both types are basic types
// of the same kind, and at least one is untyped.
func (ci *candidateInference) typeMatches(expType, candType types.Type) (bool, bool) {
// Handle untyped values specially since AssignableTo gives false negatives
// for them (see https://golang.org/issue/32146).
if candBasic, ok := candType.Underlying().(*types.Basic); ok {
if wantBasic, ok := expType.Underlying().(*types.Basic); ok {
// Make sure at least one of them is untyped.
if isUntyped(candType) || isUntyped(expType) {
// Check that their constant kind (bool|int|float|complex|string) matches.
// This doesn't take into account the constant value, so there will be some
// false positives due to integer sign and overflow.
if candBasic.Info()&types.IsConstType == wantBasic.Info()&types.IsConstType {
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
return true, true
}
}
}
}
// AssignableTo covers the case where the types are equal, but also handles
// cases like assigning a concrete type to an interface type.
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
return types.AssignableTo(candType, expType), false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// kindMatches reports whether candType's kind matches our expected
// kind (e.g. slice, map, etc.).
func (ci *candidateInference) kindMatches(candType types.Type) bool {
return ci.objKind&candKind(candType) > 0
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// assigneesMatch reports whether an invocation of sig matches the
// number and type of any assignees.
func (ci *candidateInference) assigneesMatch(cand *candidate, sig *types.Signature) bool {
if len(ci.assignees) == 0 {
return false
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
// Uniresult functions are always usable and are handled by the
// normal, non-assignees type matching logic.
if sig.Results().Len() == 1 {
return false
}
var numberOfResultsCouldMatch bool
if ci.variadicAssignees {
numberOfResultsCouldMatch = sig.Results().Len() >= len(ci.assignees)-1
} else {
numberOfResultsCouldMatch = sig.Results().Len() == len(ci.assignees)
}
// If our signature doesn't return the right number of values, it's
// not a match, so downrank it. For example:
//
// var foo func() (int, int)
// a, b, c := <> // downrank "foo()" since it only returns two values
if !numberOfResultsCouldMatch {
cand.score /= 2
return false
}
// If at least one assignee has a valid type, and all valid
// assignees match the corresponding sig result value, the signature
// is a match.
allMatch := false
for i := 0; i < sig.Results().Len(); i++ {
var assignee types.Type
// If we are completing into variadic parameters, deslice the
// expected variadic type.
if ci.variadicAssignees && i >= len(ci.assignees)-1 {
assignee = ci.assignees[len(ci.assignees)-1]
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
if elem := deslice(assignee); elem != nil {
assignee = elem
}
} else {
assignee = ci.assignees[i]
}
if assignee == nil {
continue
}
internal/lsp/source: untangle completion type comparison The code to check if a candidate object matches our candidate inference had become complicated, messy, and in some cases incorrect. The main source of the complexity is the "derived" expected and candidate types. When considering a candidate object "foo", we also consider "&foo", "foo()", and "*foo", as appropriate. On the expected side of things, when completing the a variadic function parameter we expect either the variadic slice type and the scalar element type. The code had grown organically to handle the expanding concerns, but that resulted in confused code that didn't handle the interplay between the various facets of candidate inference. For example, we were inappropriately invoking func candidates when completing variadic args: func foo(...func()) func bar() {} foo(bar<>) // oops - expanded to "bar()" and we weren't type matching functions properly as builtin args: func myMap() map[string]int { ... } delete(myM<>) // we weren't preferring (or invoking) "myMap()" We also had methods like "typeMatches" which took both a "candidate" object and a "candType" type, which doesn't make sense because the candidate contains the type already. Now instead we explicitly iterate over all the derived candidate and expected types so they are treated the same. There are still some warts left but I think this is a step in the right direction. Change-Id: If84a84b34a8fb771a32231f7ab64ca192f611b3d Reviewed-on: https://go-review.googlesource.com/c/tools/+/218877 Run-TryBot: Muir Manders <muir@mnd.rs> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>
2020-02-08 20:59:28 -07:00
allMatch, _ = ci.typeMatches(assignee, sig.Results().At(i).Type())
if !allMatch {
break
}
}
return allMatch
}
func (c *completer) matchingTypeName(cand *candidate) bool {
if !c.wantTypeName() {
return false
}
typeMatches := func(candType types.Type) bool {
// Take into account any type name modifier prefixes.
candType = c.inference.applyTypeNameModifiers(candType)
if from := c.inference.typeName.assertableFrom; from != nil {
// Don't suggest the starting type in type assertions. For example,
// if "foo" is an io.Writer, don't suggest "foo.(io.Writer)".
if types.Identical(from, candType) {
return false
}
if intf, ok := from.Underlying().(*types.Interface); ok {
if !types.AssertableTo(intf, candType) {
return false
}
}
}
if c.inference.typeName.wantComparable && !types.Comparable(candType) {
return false
}
// We can expect a type name and have an expected type in cases like:
//
// var foo []int
// foo = []i<>
//
// Where our expected type is "[]int", and we expect a type name.
if c.inference.objType != nil {
return types.AssignableTo(candType, c.inference.objType)
}
// Default to saying any type name is a match.
return true
}
if typeMatches(cand.obj.Type()) {
return true
}
if typeMatches(types.NewPointer(cand.obj.Type())) {
cand.makePointer = true
return true
}
return false
}
// candKind returns the objKind of candType, if any.
func candKind(candType types.Type) objKind {
switch t := candType.Underlying().(type) {
case *types.Array:
return kindArray
case *types.Slice:
return kindSlice
case *types.Chan:
return kindChan
case *types.Map:
return kindMap
case *types.Pointer:
// Some builtins handle array pointers as arrays, so just report a pointer
// to an array as an array.
if _, isArray := t.Elem().Underlying().(*types.Array); isArray {
return kindArray
}
case *types.Basic:
if t.Info()&types.IsString > 0 {
return kindString
}
}
return 0
}