mikeash.com pyblog/friday-qa-2009-07-10-type-specifiers-in-c-part-3.html comments

Rob Harris - 2009-07-15 14:17:50

Wed, 15 Jul 2009 14:17:50 GMT

Best write-up on "volatile" I've read. "Easy to fix!" false-starts made me laugh out loud.

mikeash - 2009-07-14 22:38:02

Tue, 14 Jul 2009 22:38:02 GMT

It would make no difference. It's still a different value in memory, even if it's part of the same struct. The "as if" rule remains unchanged. The compiler can still make these assignments happen in any order and the result is still "as if" they happened in the order specified by the program.

Anonymous - 2009-07-14 19:50:39

Tue, 14 Jul 2009 19:50:39 GMT

Excellent series, Mike. One question (that may reveal my ignorance of such things): in the block

gSharedStructure.foo = ...;
gSharedStructure.bar = ...;
gSharedStructure.baz = ...;
gFlag = 1;

would changing the gFlag variable to be part of the structure be enough to make gFlag have the "dependency" on the structure, and thus solve the problem?

mikeash - 2009-07-11 08:14:01

Sat, 11 Jul 2009 08:14:01 GMT

Lockless shared data can be really nice if you do it right. It's just extremely tricky. Certainly it makes a great case for avoiding it if you don't need it, and for minimizing it as much as you can when you do. However, if you can restrict the shared-data code to a single piece of the app (e.g. a reusable data structure) then it's much more sane. You still have to be really really careful in that one section, but once it's up and running you can then use the API it exposes without any worry.

Carl Harris - 2009-07-11 05:57:47

Sat, 11 Jul 2009 05:57:47 GMT

@Carl Harris: s/no much/now much/

Carl Harris - 2009-07-11 05:56:56

Sat, 11 Jul 2009 05:56:56 GMT

Excellent series! The details around interleaving and "as if" in both the compiler and the hardware have always been a bit of mystery to me -- no much less so! Of course, the whole of part 3 makes an outstanding case for avoiding the use of lockless shared data altogether. :-)

natevw - 2009-07-11 01:55:03

Sat, 11 Jul 2009 01:55:03 GMT

@mikeash: Thanks, this series has been quite interesting and will be a helpful reference in the future too!

@Steve Madsen:
I suspect you might be right, the combination of new language syntax and a helpful runtime in GCD may have a profound influence on many programming tasks. There's a "Technology Brief" linked from the bottom right of the Grand Central marketing blurb here, if you hadn't seen it:
http://www.apple.com/macosx/technology/#grandcentral

Vyacheslav Zakovyrya - 2009-07-11 00:52:13

Sat, 11 Jul 2009 00:52:13 GMT

Thanks for this. One more reason to use message passing concurrency to avoid all these troubles

Steve Madsen - 2009-07-10 22:06:16

Fri, 10 Jul 2009 22:06:16 GMT

This was an excellent series. This week's brought a smile to my face, having spent several years writing network kernel code and device drivers, sometimes learning the hard way about volatile, memory barriers and instruction reordering.

I don't have the Developer Preview, but I'd put money down that Grand Central in Snow Leopard is going to turn me into an old man ("In my day...").