On 06/23/2012 07:40 PM, Daniel Santos wrote: > First off, thanks for your lively commentary! > > On 06/23/2012 06:01 PM, Rob Landley wrote: >> On 06/22/2012 11:00 PM, Daniel Santos wrote: >>> Theory of Operation >>> =================== >>> Historically, genericity in C meant function pointers, the overhead of a >>> function call and the inability of the compiler to optimize code across >>> the function call boundary. GCC has been getting better and better at >>> optimization and determining when a value is a compile-time constant and >>> compiling it out. As of gcc 4.6, it has finally reached a point where >>> it's possible to have generic search & insert cores that optimize >>> exactly as well as if they were hand-coded. (see also gcc man page: >>> -findirect-inlining) >> For those of us who stopped upgrading gcc when it went to a non-open >> license, and the people trying to escape to llvm/pcc/open64/tcc/qcc/etc >> and build the kernel with that, this will simply be "less optimized" >> rather than "you're SOL, hail stallman"? > Forgive me, but I'm a little stumped here. When did GCC move to a > "non-open" license? About when it decided anti-tivoization was within its mandate, so delivering the modified code (satisfying "open source") was no longer good enough for the FSF, so they added additional restrictions on how you're allowed to install it on devices. Can 'o worms. The list has been over this at great length. > Either way, yes, one thing that must be considered with this is that > code compiled prior to gcc 4.6 is indeed "less optimized." For gcc 4.2 > through 4.5, the difference is very minor (the compare function is not > inlined). From there back, it starts to grow. In gcc 3.4.6, it's > fairly ugly. Less optimized but still works with old compilers is fine. (Less optimized comes with the territory with old compilers.) > I haven't performed tests on other compilers yet, but I wouldn't be > surprised if llvm/clang does this better. My next phase is to build a > test module so we can quickly get performance data on various platforms > & compilers. Either way, an alternate compiler will have to support at > least some of gcc's extensions to compile Linux. Or patch them out, like tccboot did. Or incorporate them into new standards, as c99 did. Or submit patches to the kernel ala the c99 structure initializers... And who's talking "if"? LLVM/CLANG builds a bootabe linux kernel: http://lwn.net/Articles/441018/ Open64 builds a bootable linux kernel: http://article.gmane.org/gmane.comp.compilers.open64.devel/2498 The PCC guys make an adorable token effort which is largely ignored: http://bsdfund.org/bundle/ Heck, Fabrice bellard did it himself with tinycc back in 2004: http://bellard.org/tcc/tccboot.html (Yes, that was a modified subset of an obsolete version, and then a little side project called "qemu" started taking up all his time, so the tinycc project stagnated in 2005. But it _was_ the first non-gcc compiler to pull this off, not counting Intel's closed-source ICC. And reviving tinycc and gluing qemu's tiny code generator onto it as a back-end so I don't have to handle target support myself is on my todo list: http://landley.net/notes-2012.html#14-06-2012 ) LLVM is the leader this space because Apple's funding it, but I really can't get that enthused about a C compiler written in C++. Of course gcc's been moving that way since 2008 (http://lwn.net/Articles/286539/) so LLVM isn't actually worse there. Highly questionable wither either is really _progress_, though. > Another question that has to be asked is "Is Linux ready for this?" The > answer today may be "yes" or "no", but to say that it can never be ready > is pure folly. I'm saying that adding complexity is not necessarily an improvement, and that over-optimizing at the expense of portability may turn out to be a mistake in the long run. That said, reducing code duplication is a good thing. I'm just not convinced "how well gcc 4.6 optimizes this" is the only relevant test criteria here. > Eventually, you have to decide to stop supporting older > compilers and move to newer paradigms that lower the maintenance burden, You keep using the word "paradigm". Voluntarily. I find this odd. > enabling your developers to work on more important things. C++ offers > lots of these types of paradigms, but the compilers just aren't good > enough yet (of course, there's the laundry list of reasons to not use C++). Um, yes. Yes there is: http://landley.net/notes-2011.html#16-03-2011 http://landley.net/notes-2011.html#19-03-2011 http://landley.net/notes-2011.html#20-03-2011 >>> Layer 2: Type-Safety >>> -------------------- >>> In order to achieve type-safety of a generic interface in C, we must >>> delve deep into the darkened Swamps of The Preprocessor and confront the >>> Prince of Darkness himself: Big Ugly Macro. To be fair, there is an >>> alternative solution (discussed in History & Design Goals), the >>> so-called "x-macro" or "supermacro" where you #define some pre-processor >>> values and include an unguarded header file. With 17 parameters, I >>> choose this solution for its ease of use and brevity, but it's an area >>> worth debate. >> Because this is just _filling_ me with confidence about portability and >> c99 compliance. > This is actually C99 compliant, even though it wont work on all > compilers. Somebody tested it on msvc and told me that it broke > (*faints*). LLVM/CLANG is bsd licensed (alas, with advertising clause, but code from it might actually be incorporatable into the linux kernel, unlke gcc, which is under a license extensively incompatible with the linux kernel). Open64 is GPLv2, same license as the linux kernel. The old tcc code was LGPLv2 but I have email from Fabrice saying he's ok with his code being used under BSD, and I'm working on the clearances and/or cleaning out the old code from people other than him to do 2 clause BSD on that. (Just for kicks at the moment. This became harder when the tinycc website bit-rotted so badly 3 years after its last release that the mailing list archive went away, but I'm dinking away at it regardless...) http://pcc.ludd.ltu.se/ is bsd licensed. It seems llvm has taken a bit of the wind out of its' sails (having Apple devote multiple full-time engineers will do that), but development is still chugging along. Somebody glued Linus's Sparse to the LLVM backend last year: http://lwn.net/Articles/456709/ Rather a large number of people reacted to gcc going GPLv3 the same way they reacted to David Dawes changing the XFree86 license. They tend to be quiet about it because the FSF guys scream bloody murder if they find out (a heretic being worse than an atheist, as always). But the "shut up and show me the code" aspects have been chugging along steadily. They all seem to have decided to start over from scratch instead of doing another egcs style fork like x.org did in response to that relicensing. Possibly this is because the one thing GPLv3 _did_ manage to do was undermine GPLv2 to the point that lots of people (myself included) bit the bullet and started releasing BSD code. (Now when you hear "the code is GPL", you can't tell from that whether or not you can incorporate it in a given project. I struggled with that for years before deciding it was untenable in the long run. Copyleft only seems to work when there's a single unified pool whose network effects made a single license a category killer, which GPLv2 was but GPLv3 isn't. Sun tried splitting the community with CDDL, but it took the FSF to have the leverage to truly poison the GPL with a CDDL2 six years _after_ Linus announced he wouldn't switch to such a thing. Section 6 of http://www.dwheeler.com/essays/gpl-compatible.html is downright ironic in retrospect.) Anyway, as I said: can 'o worms. You're welcome to disagree with my assesment, but I'm not alone in it. I'm just usually open about it. Most people who've lost faith in the GPL stay quiet about it, and several mouth the words when everybody else sings from the hymbook to avoid the inquisition. (And yes, the transcript of http://sf.geekitude.com/content/pros-and-cons-gnu-general-public-license-linucon-2005 is also retroactively ironic. Possibly in the Alanis Morisette sense of "actually just unfortunate".) > The "iffy" macro is IFF_EMPTY(), which uses a similar > mechanism to kconfig.h's IS_ENABLED() macro. I kinda doubt that one > would break and not the other. It's fairly easy to test though. Yeah, people kept emailing me that because I implemented ENABLE macros back in 2006 for BusyBox: http://git.busybox.net/busybox/commit/?id=7bfa88f315d71c7f7a1b76fcec3886c7506aca24 At the time I played around with various preprocessor macros but wasn't happy with the readability of any of it, so I modified kconfig and threw some sed in the makefile instead. Alas, when I tried to push some of the config infrastructure upstream, the kconfig maintainer at the time threw up all over it, and it devolved into bikeshedding, and I lost interest (as usual when that happens): http://lkml.indiana.edu/hypermail/linux/kernel/0707.1/1741.html (Also, if you typo a macro name then "build break" instead of "silently always false" is a _feature_, not a bug. Macros that convert #ifdef CONFIG_BLAH to if (ENABLE_BLAH) tend to go the second way because that's the input they work with. Yeah, code review will catch it eventually, but...) So yay standard C99 micro stuff. I'm a little worried about being able to unwrap it all and understand what's going on, but unlike templates we have cc -E to fall back on here, so ok. It's when you say "won't work on all compilers" that I start to worry. C99 is a fairly reasonable standard. My understanding is that our divergences from it are mostly historical (the kernel predates C99). > Aside from that, is the size of the macro. After stripping comments, > the size is roughly 4155 bytes, indeed beyond the C99's minimum required > size of 4096 bytes (even though most compilers use a much larger > buffer). Eh, I'm not worried about that. That _is_ a "fix your compiler" issue. (Hardwired buffer sizes in 2012... tsk.) > So that's another issue that can break on other compilers, > unless it stores the macro with whitespace condensed, which would bring > it down to 3416 bytes. >> (Or I suppose C11!!one! compliance. The new thing that puts asserts in >> the base language and makes u8 a keyword since _that_ won't break >> existing code and putting utf8 string constants within quotes wasn't >> previously possible.) >> >> I'm not saying the standard's perfect, I'm saying a web page that ties >> itself to mozilla at the expense of working on firefox, let alone >> chrome, might be a bit short-sighted these days. XFree86 begat x.org, >> OpenOffice begat libre, etc. The FSF went nuts again and this time >> around EGCS is called LLVM, so talking about gcc 4.6-only features >> thrills some of us less than you might expect. > It's really just a "not broken on gcc 4.6", since -findirect-inlining > was introduced in gcc 4.4 or some such. Ok, so "designed with gcc 4.6's optimizer in mind, regression tested back to 3.4 for at least minimal functionality, and not intentionally breaking other compilers". That's what I wanted to know, and it sounds good to me. Rob -- GNU/Linux isn't: Linux=GPLv2, GNU=GPLv3+, they can't share code. Either it's "mere aggregation", or a license violation. Pick one. -- To unsubscribe from this list: send the line "unsubscribe linux-doc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
