http://bugzilla.gdcproject.org/show_bug.cgi?id=214

vlevenfeld@gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vlevenfeld@gmail.com

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

vlevenfeld@gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Hardware|All                         |ARM
                 OS|All                         |Linux

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

--- Comment #1 from vlevenfeld@gmail.com ---
Found a fix: If I go into phobos' epoll.d and change align(1) to align(8),
everything works.

According to this thread (https://github.com/nodejs/node-v0.x-archive/issues/4565) ARM systems require 64-bit data elements to be 8-byte-aligned. Looks like that got broken between 4.9.0 and 4.9.2?

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

--- Comment #2 from Iain Buclaw <ibuclaw@gdcproject.org> ---
Is the library different between 4.9.0 and 4.9.2?

Either the align was added to the library, or the compiler started doing what the code requested.

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

Johannes Pfau <johannespfau@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |johannespfau@gmail.com
             Status|NEW                         |ASSIGNED

--- Comment #3 from Johannes Pfau <johannespfau@gmail.com> ---
I guess the library changed, as that old forum thread mentioned in 2014 there was no align(1) in epoll.d which was wrong on X86.

Whether the struct is packed depends on the target architecture. On ARM the
struct is not packed. The ARM version should just use standard alignment (no
explicit align(8)), the align(1) version should be restricted to X86/(64) and
an additional else static assert(false) should be added. Needs to go into
upstream druntime as well.

(grep __EPOLL_PACKED in glibc). Actually, AFAICS X86 ist the only architecture
using a packed struct.

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

--- Comment #4 from vlevenfeld@gmail.com ---
The library is the same - it is actually not phobos' epoll but libasync's epoll
that is being used in this specific instance (though the two files don't appear
to differ). So the same file is being used with different compilers.
It appears as though, before 4.9.2, GDC ignored the alignment directive (at
least on ARM).
So this is not a bug with GDC, but rather with libasync and phobos? The epoll
implementation should be responsible for setting the correct alignment for the
architecture?

-- 
You are receiving this mail because:
You are watching all bug changes.

http://bugzilla.gdcproject.org/show_bug.cgi?id=214

--- Comment #5 from Johannes Pfau <johannespfau@gmail.com> ---
The align(1) in epoll.d was introduced in 2.066 but IIRC there were also some
changes in align() handling at some point. I think it used to be ignored. But
either way, the compiler is correct and druntime (and probably libasync) need
to be fixed:

https://github.com/D-Programming-Language/druntime/pull/1502

-- 
You are receiving this mail because:
You are watching all bug changes.