See my suggestion here: https://github.com/ebitengine/purego/issues/71#issuecomment-1332545306
For //go:norace and ABI wrapper functions, there's no instrument emitted. Your problem is that the compiler know there's a native ABI0 implementation for x_cgo_init, so calling it from assembly is force to use the ABI0 one, not the ABI wrapper. On Wednesday, November 30, 2022 at 11:30:38 PM UTC+7 Jarrett Kuklis wrote: > I am one of the main contributors to ebitengine/purego > <https://github.com/ebitengine/purego> a library that aims to call into C > code without invoking the C compiler. The main platform of support is > darwin on arm64 and amd64. It uses //cgo_import_dynamic comments to get the > symbols and calls them using runtime.cgocall. > > That runtime function requires that runtime/cgo be imported which requires > C. I used the work of notti/nocgo <https://github.com/notti/nocgo> as > inspiration to write an implementation of runtime/cgo entirely in Go > (located at internal/fakecgo > <https://github.com/ebitengine/purego/tree/main/internal/fakecgo>). It > has worked for the ebitengine <https://github.com/hajimehoshi/ebiten> > project really well. > > However, I was looking at the draft Go 1.20 release notes and saw that > -race no longer requires Cgo to work on macOS. Which I am all for and > really excited about! This means though that when -race and CGO_ENABLED=0 > our faked runtime/cgo package has race code placed in it. From what I can > tell this is bad because the code in this package is faking being C so it > shouldn't call into the race detector. And it will SIGSEGV jumping to any > code that tries to call it since the runtime isn't fully initialized by the > time these functions are called. > > There is an issue here <https://github.com/ebitengine/purego/issues/71> that > has more information about the issue. > > Obviously, I tried using //go:norace which indeed does remove the race > code from the function but the ABI0 wrapper still has the race detector > code. The functions are called from assembly since we need to move from the > C ABI to Go ABI which forces us into calling the ABI0 wrapper. I can't use > <ABIInternal> since that is only allowed in the runtime. As far as I can > tell there is no way around the wrapper. > > So I was wondering if this is a bug? Shouldn't the //go:norace comment > also apply the the ABI wrappers? If not, is there any other suggestions as > to how purego could keep functioning when CGO_ENABLED=0 and -race are > defined? > > I've had a hard time debugging this since no debugger I've tested links to > the program before it crashes with SIGSEGV. And there is no stack trace. I > may be missing something so I am eager to learn! > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/fd190e4f-8054-497c-820d-3d0de60cd768n%40googlegroups.com.
