From:  anatol.php@belski.net ("Anatol Belski")
Date:  26 Nov 2015 07:26:13 Hong Kong Time
Newsgroup:  news.php.net/php.internals.win
Subject:  

RE: [INTERNALS-WIN] Re: [PHP-DEV] Windows (Visual Studio) compiler stuff

NNTP-Posting-Host:  null

Hi Matt,

> -----Original Message-----
> From: Matt Wilmas [mailto:php_lists@realplain.com]
> Sent: Monday, November 23, 2015 8:15 AM
> To: Anatol Belski ; internals@lists.php.net;
internals-
> win@lists.php.net
> Cc: 'Dmitry Stogov' ; 'Pierre Joye'

> Subject: Re: [INTERNALS-WIN] Re: [PHP-DEV] Windows (Visual Studio)
compiler
> stuff
> 
> Hi Anatol, all,
> 
> ----- Original Message -----
> From: "Anatol Belski"
> Sent: Monday, November 16, 2015
> 
> > Hi Matt,
> >
> >> -----Original Message-----
> >> From: Matt Wilmas [mailto:php_lists@realplain.com]
> >> Sent: Monday, November 16, 2015 2:59 PM
> >> To: Anatol Belski ; internals@lists.php.net;
> > internals-
> >> win@lists.php.net
> >> Cc: 'Dmitry Stogov' ; 'Pierre Joye'
> > 
> >> Subject: [INTERNALS-WIN] Re: [PHP-DEV] Windows (Visual Studio)
> >> compiler stuff
> >>
> >> > According to the docs __declspec(noinline) is specific to C++. Also
> >> > with VS it's always much more tedious to inline something than the
> >> > opposite. These are the main two reasons it's disregarded ATM. We
> >> > can add it for compliance with C++, but it'll in best case have no
> >> > effect in the PHP core. Should be tested before, though.
> >>
> >> Yeah, I know what the docs imply ("member function"), which is why I
> > tested it.
> >> I guess you missed my "works as expected" part. :-P
> >>
> >> A test function that just returns a number was automatically inlined
> > (plain C).
> >> Using __declspec(noinline) it was call'ed instead.
> >>
> >> Not sure if any of the "zend_never_inline" PHP stuff is getting
> >> inlined
> > when it's
> >> desired not to be -- I'll compile PHP in a bit and see what it looks
> >> like
> > with
> >> "noinline."
> >>
> > Yeah, I knew it could work, just that it's undocumented so preferred
> > not even to start with it because I haven't expect much gain from it.
> > The functions I've seen with zend_never_inline are rather big and
> > wouldn't get inlined even when forced.
> 
> noinline did have an effect -- 12 KB smaller php7.dll.  So, obviously it's
> preventing those zend_never_inline functions from being inlined when they
> currently are.  Dmitry surely had reason to make them that way --
cache-related,
> I assume.  Any difference, however "minor," is the same as other
compilers, so
> it's nice to know this can be used, with so many of the other GCC/Clang
"tricks"
> missing...
> 
I wasn't telling it wouldn't work. We should check for possible
implications. If there's nothing negative, so we can add this into master.
It always depends, smaller image size vs. function call.

> BTW, something "big" not getting inlined even when forced?  I know the
"rules"
> about what can't be [force] inlined (basically same as GCC) and size isn't
one of
> them. :-)  (I hope not.)  As I've mentioned a bit, to be seen soon, my
"compile-
> time" param parsing optimization will have the "hugest"
> inline function, but it compiles down to literally nothing, which I
finally got to
> work with MSVC as well.  That's why I wasn't liking the idea of a
standalone copy
> of that stuff adding several KB to each module...
> 
Size is one of the factors, the concrete code and usage, too. Despite that,
any compiler doc says that inline is just a suggestion.

> >> > I'd ask you for some concrete case for this, as I'm not sure to
> >> > understand exactly what you mean. The only case where an extra code
> >> > would be generated is with "__declspec(export) inline", but that's
> >> > not the case anywhere within PHP.
> >>
> >> My concrete case is checking tons of generated code! ;-)
> >>
> >> It's simple: useless standalone functions are created for every
> >> "static __forceinline" definition...  Not having static makes it act
> >> like
> > GCC/Clang.
> >>
> > I guess I've understood what you're talking about - abut unreferenced
> > COMDATs (or maybe also duplicated COMDATs). There is a variety of
> > situations for that, not possibly only inlining. Fixing it is done in
> > PHP when building with --enable-debug-pack, that is on in release
> > builds. In your experiments, if you add /Zi CFLAG (or explicitly /Gy)
> > and /OPT:REF,ICF LDFLAG - that will solve it for yur other project.
> > You can read more about COMDAT on MSDN.
> 
> Yeah, I know about the COMDAT stuff.  And I thought I had tried the
/OPT:REF,
> etc. on a standalone test a while ago and it didn't do anything...
> 
> I just now tried --enable-debug-pack, and as I was thinking, it had no
effect.
> 
What do you mean with "no effect"? Don't reduce size? The compiler/linker
options I've mentioned are about removing identical or unreferenced COMDATS,
and they do that. BTW how do you check it? I would like you to be more
precise at this point, please. Did you use link /map or disasm?

> I don't need to solve anything on the other project since I didn't use
static there.
> :-P
> 
> > Hm, probably these options could be revisited, as since 2013 there's
> > also /Gw and /Zc:inline switches which is not implied by /Zi. But have
> > to do more checks, for now the release build options are good enough.
> >
> >> Again, I'll try to compile PHP with those static's removed and report
> >> the
> > effect
> >> later.
> >>
> > Yes, thanks for your effort. I actually didn't check what gcc does for
> > such cases, so curious. But "static" in "static inline" forces every
> > translation unit to have even the same function to have different
> > address, thus eliminating the "one definition" rule for inline. We
> > anyway need "static inline" best compatibility, the compilers handle
> > the rest :)
> 
> First, the report: Removing all the static's with zend_always_inline works
fine
> (since the __forceinline seems to "imply" static, no duplicate symbols).
It makes
> php7.dll 91 KB smaller (NTS --disable-all).
> 
> But then when I tried the /Zc:inline option (really sounds like C++ on
MSDN) the
> other day, I was pleasantly surprised!  "You da man!" :-)
> 
> That saved over 220 KB, without removing static's.  I verified that the
standalone
> functions (from static's) were gone, but obviously it also removed a lot
more.
> Thank you!
> 
> Hopefully that's a switch that can be taken advantage of?
> 
/Zc:inline is documented as C++11 feature. Still it is about enforcing the
definition within the same translation unit, so basically kind of synonymous
to the cl/link options we have. It doesn't enforce C++11, just one that
rule. Whether it'd break some C++ extensions - well, should check. 

But about the "static inline", it is really something that should be kept
everywhere. It is the most convenient option for the compiler/linker
compatibility. A global function is allowed to be defined only once. Since
those functions are in defined in the headers, chances are to see the
duplicated symbol errors which will prevent compilation. VC should actually
should do same. The option using "extern inline" and splitting declaration
and definition are unusable, because those functions have to be usable in
external modules.

In general, if testing goes good, we could add these options
(__declspec(noinline), /Zc:inline and maybe /Gw) to master to release builds
for further observation. But it should be really good tested. We'll check it
in our labs as well. I'll be able to come to this topic either at the end of
this or early next year.

Regards

Anatol