Stream: t-compiler/wg-self-profile

Topic: Summarize diffs


Wesley Wiser (Apr 26 2019 at 12:41, on Zulip):

@centril has written up an issue for diff output from the summarize tool. I've added some getting started instructions if anyone is interested in working on this. I'm happy to answer questions or provide help to anyone interested.

Otherwise, I'll probably work on this sometime in the next week or two.

centril (Apr 26 2019 at 12:44, on Zulip):

@Wesley Wiser

We'll need to change the usize fields to be signed. i64 is probably appropriate.

Can't you use impl Sub for QueryData with type Output = QueryDiff; where you use i64 in the QueryDiff?

Wesley Wiser (Apr 26 2019 at 13:27, on Zulip):

Seems unfortunate to have two identical data structures that differ only by the numerical types of a few fields

Wesley Wiser (Apr 26 2019 at 13:27, on Zulip):

I think perf.rlo will also be interested in this feature and they'd probably rather not duplicate the data structures on their side either

centril (Apr 26 2019 at 13:32, on Zulip):

@Wesley Wiser there's always pub struct QueryData<N = usize> { :P

Wesley Wiser (Apr 26 2019 at 13:32, on Zulip):

is that const generics?

Wesley Wiser (Apr 26 2019 at 13:33, on Zulip):

Or whatever we're calling it these days lol

centril (Apr 26 2019 at 13:33, on Zulip):

@Wesley Wiser no it's just a type parameter with a default set to usize

Wesley Wiser (Apr 26 2019 at 13:33, on Zulip):

Oh right

Wesley Wiser (Apr 26 2019 at 13:33, on Zulip):

I read N: usize

Wesley Wiser (Apr 26 2019 at 13:33, on Zulip):

Should probably go have that morning cup of coffee

Alice Ryhl (Apr 26 2019 at 13:43, on Zulip):

I mean isn't a QueryDiff fundamentally different from a QueryData in a semantic way? I would default to two types

Alice Ryhl (Apr 26 2019 at 13:46, on Zulip):

And wouldn't the diff need to keep track of lines only in one of the datasets too in an additional field

Wesley Wiser (Apr 26 2019 at 13:50, on Zulip):

I think naively, any non matching lines would be shown as +\- their full value

Alice Ryhl (Apr 26 2019 at 21:09, on Zulip):

Thinking a bit more about it, I think we need to remember both the old and new value (or ratio I guess), since otherwise we can't detect if, say, +100 cache hits is a 0.1% increase or a 100% increase.

Wesley Wiser (Apr 26 2019 at 21:27, on Zulip):

Yeah, that seems like a reasonable concern. We should test out it though. It may be too much noise to show the ratio for every value (or at all) and the user could always run a regular summarization of the data sets to see the actual values.

Last update: Nov 17 2019 at 07:35UTC