Have you tried a Go 1.13 beta or master? This should have been fixed by https://go-review.googlesource.com/c/go/+/172918.

I plan on tweaking the logic a bit to not use a map if the number of elements is small; I have a patch locally, but I haven't sent it yet. The important bit is not doing the expensive case-insensitive matching in the case where a case-sensitive match exists.

@mvdan I have not tried on Go 1.13 and that patch looks almost identical to mine without the heuristic so I'm confident that that will fix the issue we are seeing.

Cool. In the future, have a look at https://golang.org/doc/contribute.html, which contains information on how to build Go from master. You should use that version to experiment with changes, as the stable release may not contain the latest relevant changes.

@mvdan One difference between your patch and mine is that I iterated backwards through the fields when generating the map to preserve the existing behaviour that the "first" matching struct field would be filled rather than the last.

For example, suppose I have a type like this:

type Foo {
 A string
}

type Bar {
 Foo
 A string
}

If you JSON unmarshal into Bar then it will fill in Bar.A not Foo.A. I believe with that patch the behaviour will be the reverse since the "last" field will be the one that ends up in the map.

It might be worthwhile adding a test for this behaviour as I imagine it might be quite easy to break and will be being relied on.

Hmm, from memory I think there are tests covering this case already, but I might be wrong. I'll take a look and update you here.

Yeah, I can't reproduce a regression, sorry. Here is the code I used: https://play.golang.org/p/Ot_EVbu3lpN

$ go version
go version devel +80cca23b59 Wed Jul 10 21:26:21 2019 +0000 linux/amd64
$ go run f.go
main.Bar{Foo:main.Foo{A:""}, A:"foo"}
$ go1 version
go version go1.12.7 linux/amd64
$ go1 run f.go
main.Bar{Foo:main.Foo{A:""}, A:"foo"}

Note that the decoder has always preferred the last case sensitive match (or the last case insensitive match), so master is correct as far as I can tell. If you do encounter a regression, please file a new issue with a program to reproduce the problem, and we'll look into a fix before the final release.

I'll take a look later and report back, thanks for taking a look.