Hi XS people,
This seems the proper venue to share the solution to a bug that has
been bedeviling my CPAN module IPC::MMA on some threaded
systems since its introduction.
The problem affected three array operations that take a Perl list in
their operands: push, unshift, and splice. The XS interface routines
for all three included code that took the variable pointers from the
Perl stack and put them into a C array:
mm_array_push (array, ...)
...
PREINIT:
int add_count = items - 1;
SV *addSVs[add_count];
int i=0;
CODE:
while (i < add_count) addSVs[i++] = ST(i+1);
The C spec is not definitive on when some operations should happen.
The && and || operators are defined to be time-sequential, in that
everything on the left side happens (is evaluated) before anything
on the right side (which doesn't happen if the result of the && or ||
can be determined from the left result).
I thought assignment operators were like that too, in that everything
on the right side should happen before anything on the left side. But
I found that {addSVs[i]=ST(i+1); i++;} fixed my problems.
So I went on a web search to determine whether the C spec said that
assignment was a time-sequential operator, in which case this would
be a bug in the compiler. It was a hard search. Finally I found a
page that settled the issue:
http://www.mail-archive.com/tinycc-de...@nongnu.org/msg02398.html
I was wrong. The right side can end up using either the original or
incremented value of i.
Why this only showed up on some threaded systems is a puzzler. In
theory the compiler could generate code in which one thread/core
evaluates the right-side expression while another thread/core computes
the left-side address. But the overhead of dispatching and converging
two threads would seem greater than just doing the whole statement
in one thread.
Don't know: I've never written a code generator for threads, and my
compiler-writing days are behind me.
Or maybe the correlation with threads was false. Maybe it was a matter
of different versions of gcc on different systems. I'll never know but
I'll wonder for quite some time.
Maybe this will help someone someday,
Craig MacKenna
