Bisecting identifies c39918a as the source of regression. (cc @aclements @randall77 for FYI, since this seems like a somewhat surprising/unintended consequence of that CL)

It's worth noting that

_ = T{1, 1}

used to fail too, as did:

x := T{0, 0}
_ = x

So while failing to compile _ = T{0, 0} is technically a compiler bug and regression, it doesn't seem like a release blocker to me, as it doesn't seem indicative of any more severe correctness issues. Unless anyone objects, I think we can safely punt it to 1.9.

Also, for #17728, I'm interested in moving the logic for decomposing composite-literal assignments into SSA, which I expect will address these issues.

@griesemer and @mdempsky, this issue seems fixed on tip by commit 3e11940 and CL https://go-review.googlesource.com/c/34988/

$ go version
go version devel +11eaf42 Sat Apr 29 04:15:49 2017 +0000 darwin/amd64
$ cat main.go 
package main

type T struct {
	x int
	_ int
}

func main() {
	_ = T{0, 0}
}
$ go run main.go 

PS: It works alright on the playground and was fixed already for Go1.8; that CL got in on January 9th 2017.

I'm puzzled both by how the first commit could have broke this and how the second commit could have fixed it. CL 38006, for example, seems more relevant.

In any case, @odeke-em, want to send a CL adding this code as a test case and mark it as fixing this issue?

My thoughts exactly @josharian: I started by examining before and after
that commit, but it was already fixed.

Sounds like a plan, let me mail the test case.

CL https://golang.org/cl/42173 mentions this issue.