The trimming optimization deletes parts of the syntax tree
that don't affect the type checking of package-level declarations.
It used to remove unexported struct fields, but this had
observable consequences: it would affect the offset of later
fields, and the size and aligment of structs, causing the
'fieldalignment' analyzer to report incorrect findings.
Also, it required a complex workaround in the UI element
for hovering over a type to account for the missing parts.
This change restores unexported fields.
The logic of recordFieldsUses has been inlined and specialized
for each case (params+results, struct fields, interface
methods) as they are more different than alike.
BenchmarkMemStats on k8s shows +4% HeapAlloc:
a lot, but a small part of the 32% saving of the trimming
optimization as a whole.
Also:
- trimAST: delete func bodies without visiting them.
- minor clarifications.
Updates golang/go#51016
Change-Id: Ifae15564a8fb86af3ea186af351a2a92eb9deb22
Reviewed-on: https://go-review.googlesource.com/c/tools/+/415503
gopls-CI: kokoro <noreply+kokoro@google.com>
Run-TryBot: Alan Donovan <adonovan@google.com>
Auto-Submit: Alan Donovan <adonovan@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
Despite the name, ParseExported only hollowed out declarations -- it
didn't actually drop any from the AST. This leaves a fair amount of
unexported crud behind. Unfortunately, there are a *lot* of ways to
expose an unexported declaration from an exported one, and it can be
done across files. Because of that, discarding unexported declarations
requires a lot of work.
This CL implements a decent attempt at pruning as much as possible from
the AST in ParseExported mode.
First, we analyze the AST of all the files in the package for exported
uses of unexported identifiers, iterating to a fixed point. Then, we
type check those ASTs. If there are missing identifiers (probably due to
a bug in the dependency analysis) we use those errors to re-parse. After
that we give up and fall back to the older, less effective trimming. The
pkg type changes slightly to accomodate the new control flow.
We have to analyze all the files at once because an unexported type
might be exposed in another file. Unfortunately, that means we can't
parse a single file at a time any more -- the result of parsing a file
depends on the result of parsing its siblings. To avoid cache
corruption, we have to do the parsing directly in type checking,
uncached.
This, in turn, required changes to the PosTo* functions. Previously,
they operated just on files, but a file name is no longer sufficient to
get a ParseExported AST. Change them to work on Packages instead. I
squeezed in a bit of refactoring while I was touching them.
Change-Id: I61249144ffa43ad645ed38d79e873e3998b0f38d
Reviewed-on: https://go-review.googlesource.com/c/tools/+/312471
Trust: Heschi Kreinick <heschi@google.com>
Run-TryBot: Heschi Kreinick <heschi@google.com>
gopls-CI: kokoro <noreply+kokoro@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
While looking at Kubernetes I noticed that golang.org/x/text packages
were some of the largest. The problem is the large code-generated
tables, which use ellipsis array literals. Teach gopls to trim the cases
that matter there.
While silly, this trims ~60MB off the live heap, so I think it might be
worth it.
Change-Id: I0cfd80bd5fbc8703ac628312982af9c6ed871758
Reviewed-on: https://go-review.googlesource.com/c/tools/+/248180
Run-TryBot: Heschi Kreinick <heschi@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rebecca Stambler <rstambler@golang.org>
Alan Donovan
Heschi Kreinick