1
0
mirror of https://github.com/golang/go synced 2024-11-18 17:24:40 -07:00
go/godoc/analysis/README
Alan Donovan 80c4f06c0f go.tools/godoc: server mode: add support for type and pointer analysis.
See analysis.go for overview of new features.
See README for known bugs and issues.

Much UI polish, testing and optimization work remains, but
this is a starting point.

Flag: we add a new flag -analysis=type,pointer, default "",
for adventurous users only at this stage.
Type analysis takes ~10s for stdlib + go.tools;
Pointer analysis (currently) takes several minutes.

Dependencies: we now include jquery.treeview.js and its GIF
images among the resources.  (bake.go now handles binary.)

LGTM=crawshaw, bgarcia
R=crawshaw, bgarcia
CC=bradfitz, golang-codereviews
https://golang.org/cl/60540044
2014-03-14 18:58:22 -04:00

122 lines
4.2 KiB
Plaintext

Type and Pointer Analysis to-do list
====================================
Alan Donovan <adonovan@google.com>
Overall design
--------------
We should re-run the type and pointer analyses periodically,
as we do with the indexer.
Version skew: how to mitigate the bad effects of stale URLs in old pages?
We could record the file's length/CRC32/mtime in the go/loader, and
refuse to decorate it with links unless they match at serving time.
Use the VFS mechanism when (a) enumerating packages and (b) loading
them. (Requires planned changes to go/loader.)
Future work: shard this using map/reduce for larger corpora.
Testing: how does one test that a web page "looks right"?
Bugs
----
(*go/loader.Program).Load fails if it encounters a single parse error.
Make this more robust.
(*ssa.Program).Create requires transitively error-free packages. We
can make this more robust by making the requirement transitively free
of "hard" errors; soft errors are fine.
Markup of compiler errors is slightly buggy because they overlap with
other selections (e.g. Idents). Fix.
User Interface
--------------
CALLGRAPH:
- Add a search box: given a search node, expand path from each entry
point to it.
- Cause hovering over a given node to highlight that node, and all
nodes that are logically identical to it.
- Initially expand the callgraph trees (but not their toggle divs).
CALLEES:
- The '(' links are not very discoverable. Highlight them?
Type info:
- In the source viewer's lower pane, use a toggle div around the
IMPLEMENTS and METHODSETS lists, like we do in the pacakge view.
Only expand them initially if short.
- Include IMPLEMENTS and METHOD SETS information in search index.
- URLs in IMPLEMENTS/METHOD SETS always link to source, even from the
package docs view. This makes sense for links to non-exported
types, but links to exported types and funcs should probably go to
other package docs.
- Suppress toggle divs for empty method sets.
Misc:
- Add an "analysis help" page explaining the features and UI in more detail.
- The [X] button in the lower pane is subject to scrolling.
- Should the lower pane be floating? An iframe?
When we change document.location by clicking on a link, it will go away.
How do we prevent that (a la Gmail's chat windows)?
- Progress/status: for each file, display its analysis status, one of:
- not in analysis scope
- type analysis running...
- type analysis complete
(+ optionally: there were type errors in this file)
And if PTA requested:
- type analysis complete; PTA not attempted due to type errors
- PTA running...
- PTA complete
- Scroll the selection into view, e.g. the vertical center, or better
still, under the pointer (assuming we have a mouse).
More features
-------------
Display the REFERRERS relation? (Useful but potentially large.)
Display the INSTANTIATIONS relation? i.e. given a type T, show the set of
syntactic constructs that can instantiate it:
var x T
x := T{...}
x = new(T)
x = make([]T, n)
etc
+ all INSTANTIATIONS of all S defined as struct{t T} or [n]T
(Potentially a lot of information.)
(Add this to oracle too.)
Optimisations
-------------
Each call to addLink takes a (per-file) lock. The locking is
fine-grained so server latency isn't terrible, but overall it makes
the link computation quite slow. Batch update might be better.
Memory usage is now about 1.5GB for GOROOT + go.tools. It used to be 700MB.
Optimize for time and space. The main slowdown is the network I/O
time caused by an increase in page size of about 3x: about 2x from
HTML, and 0.7--2.1x from JSON (unindented vs indented). The JSON
contains a lot of filenames (e.g. 820 copies of 16 distinct
filenames). 20% of the HTML is L%d spans (now disabled). The HTML
also contains lots of tooltips for long struct/interface types.
De-dup or just abbreviate? The actual formatting is very fast.
The pointer analysis constraint solver is way too slow. We really
need to implement the constraint optimizer. Also: performance has
been quite unpredictable; I recently optimized it fourfold with
type-based label tracking, but the benefit seems to have evaporated.
TODO(adonovan): take a look at recent CLs for regressions.