We have standard OpenJDK source locations for platform-specific code, as described in http://openjdk.java.net/jeps/201 .
For example: http://hg.openjdk.java.net/jdk9/jdk9/jdk/file/tip/src/java.base/linux/classes/sun/nio/ch/LinuxAsynchronousChannelProvider.java http://hg.openjdk.java.net/jdk9/jdk9/jdk/file/tip/src/java.base/linux/native/libnio/fs/LinuxWatchService.c We have a place for both C and Java sources. Amazingly, there is even one assembly source file: http://hg.openjdk.java.net/jdk9/jdk9/jdk/file/tip/src/java.desktop/unix/native/libawt/sun/awt/medialib/mlib_v_ImageCopy_blk.s But, there is no place at present for (a) ISA-dependent source files, or (b) ISA-and-platform-dependent source files. Definitions: An ISA-dependent source file is coded portably (perhaps in Java, or C with ifdefs) but makes use of specific instructions, such as SPARC VIS or Intel AVX. A ISA-and-platform-dependent source file is coded for a particular platform and ISA; a typical example would be an assembly file whose syntax is platform specific (like linux/x86). Context: The JNR system we are importing to Project Panama will include both types of source files. Straw man proposal: Allow the folder names "cpu.$CPU" and "$OS.$CPU" to occur as a sibling to "share" and $OS in source paths. Here's a BNF style description of JEP 201 source paths: path := 'src/' $MODULE '/' platform_scope '/' ( classes_path | native_path | conf_path ) platform_scope := 'share' | $OS classes_path := 'classes/' $PACKAGE '/' *.java native_path := 'native/' ( 'include/' *.{h,hpp} | $LIBRARY '/' *.{c,cpp} ) conf_path := 'conf/' * The change would be: platform_scope := 'share' | $OS | 'cpu.' $CPU | $OS '.' $CPU Observation 1: This scheme does not entangle the type of source (native vs. classes) with the platform scope. It thus allows for platform-specific assembly files, but encourages shared but ISA-specific code. Observation 2: Less mangled patterns like "share/cpu/$CPU" might be cleaner, but they also perturb the existing pathname lengths, by adding optional pathname components. Controlling pathname depth seems like a desirable goal. In cases where single output artifacts have to be built with support for multiple platforms, we would also add the cpu name somewhere in the $PACKAGE component, and/or the base name of the file (class name), as Graal does. (There is a possible followup question about deploying multiple ISAs and/or platforms in one artifact, kind of like a multi-JAR. But I think we can tackle that later, and it may be that package-name hacking will allow multiple ISAs to co-exist when needed, as is the case with Project Sumatra.) Comments? Does this look like a good starting point for organizing ISA-specific code for Panama? Thanks, — John P.S. As a point of comparison, the hotspot repo uses this grammar, which supports only C and assembly code: path := 'src/' platform_scope '/vm/' $GROUP '/' *.{c,cpp,hpp} platform_scope := 'share' | 'cpu/' $CPU | 'os/' $OS | 'os_cpu/' $OS '_' $CPU where GROUP is something like 'code', 'asm', 'c1', 'classfile', etc., an informal grouping of source files in a language without packages. Also, the base name of the file repeats $OS and $CPU components, if present. Examples: http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/tip/src/cpu/x86/vm/macroAssembler_x86.hpp http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/tip/src/os/linux/vm/osThread_linux.hpp http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/tip/src/os_cpu/linux_x86/vm/thread_linux_x86.hpp