Right now the process for adding in new constants, errors, or syscalls
for Linux is a pain and unreliable. The scripts are designed to be run
on the target architecture and use the header files installed on the
user's system. This makes it hard to generate files for all the
architectures or to have consistency between users. See golang/go#15282.
This CL fixes this issue by making all of the files for the 11 supported
architectures directly from source checkouts of Linux, glibc, and bluez.
This is done using Docker, the gcc cross-compilers, and qemu emulation.
Previously discussed here:
https://go-review.googlesource.com/c/37589/
A README.md file is also added to explain how all the parts of the build
system work.
In order to get the build working for all the architectures, I made
some changes to the other scripts called from mkall_linux.go:
- Files only used for generating linux code, moved to linux/
- linux/mksysnum.pl supports a specified CC compiler.
- The generated C code in mkerrors.sh changed to avoid a warning
- mkerrors.sh headers changed to fix powerpc64 bug in sys/ioctl.h
- linux/types.go no longer needs to export Ptrace structs in lowercase
Build instructions:
- Host system needs to be x86-64 Linux
- Install Docker (https://docs.docker.com/engine/installation/)
- ./mkall.sh (That's it!!!)
Change-Id: I87067c14442ba12f8d51991349a43a9d73f38ae0
Reviewed-on: https://go-review.googlesource.com/37943
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit adds linux/s390x support to the unix package. It is
based on the changes made to the syscall package in
https://golang.org/cl/20961/. It also adds mkpost.go which is
used to cleanup the API generated by cgo -godefs.
The biggest departure that is made with the syscall package is
the use of the -fsigned-char flag to force signed chars. We
couldn't do this in the syscall package because of the need to
maintain compatibility with the gccgo implementation of the syscall
package (gccgo has supported s390x for a longer time than the Go
toolchain). The unix package does not have this constraint.
Using the -fsigned-char flag makes the API look more like the one
generated on amd64 and arm64 and also more consistent with itself
(the syscall package represents chars using both int8 and uint8
types, the sys package will only ever use int8). Unfortunately it
also means that applications transitioning from the syscall package
to the unix package will see a different API on s390x which might
be confusing. I think the tradeoff is worth it though.
Change-Id: I40b90c18ed787e74ba7a2ebd004bd6bd1ba6279a
Reviewed-on: https://go-review.googlesource.com/23045
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Joe Richey