Some Filters can't be reasonably applied if the debuggee is in the wrong
profiling mode. For example, eras profiling requires the debuggee to be
run with `-he`.

We reflect this in ghc-debug-brick by disabling certain filters if their
profiling mode requirements are not met.

Hannes Siebenhandl authored 1 year ago and Hannes Siebenhandl committed 1 year ago

Adds two new requests if the debuggee program has been compiled with
profiling:

* Request CCS_MAIN location
* Dereference IndexTable

Package `ghc-debug-client` adds utility functions to find CCSPtrs given
a cost centre id, either including all transitive CCS children or only
CCS parts that reference a specific cost centre.

Add UI Filter to `ghc-debug-brick` allowing to list closures that have
been allocated under a specific cost centre.

When thunks are evaluated, the thunk closure is replaced with an
`IndClosure`. To make sure we don't miss parts of the string, we also
expand `IndClosure` types.

515eb33d4fcc9382032a9e6b1757b9e2b00a145c changes the order and type of
these two fields. Therefore we have to update the decoding logic.

In particular we make sure to race all the appropiate threads against
each other, so that if any of them fail the tests just fail rather than
hang forever.

`viewport` is unreasonably inefficient for our usecase when the IOTree grows, as it
decides what to display once it rendered to full `IOTree`.
Especially while scrolling, this inefficiency became apparent.

Unfortunately, brick doesn't have any builtin way to improve this, see https://github.com/jtdaugherty/brick/issues/264

We take `drawListElement` as an inspiration, and only render up to
`heightOfBox * 2` elements of the `IOTree`, including expanded tree
nodes.
This speeds up the view noticeably, giving the user a responsive experience.

In particular, we decouple rendering metadata computation from the actual rendering to decide which rows to render.

We now display a "flat" cost centre stack rather than forcing the user
to expand it at each level.

We should never attempt to dereference a null pointer, any null pointers
should be captured at the point of creation.

These are no longer used

Matthew Pickering authored 1 year ago and Zubin committed 1 year ago

Inside an hBox if there are multiple greedy widgets they are allocated
equal space. This meant the commands and number display were allocated
the same number of columns which led to a lot of whitespace.

(cherry picked from commit 7a1fdab5)