'__builtin_fpclassify' and '__builtin_inf' (and probably others) lack sema and codegen support

[Quuxplusone]

Bugzilla Link	PR6083
Status	NEW
Importance	P normal
Reported by	Maurice Gittens (mainmanmauricio@gmail.com)
Reported on	2010-01-19 10:39:42 -0800
Last modified on	2018-10-25 20:11:53 -0700
Version	trunk
Hardware	PC Linux
CC	benny.kra@gmail.com, chris@bubblescope.net, dgregor@apple.com, dsaritz@gmail.com, evan@chromium.org, jyasskin@google.com, llvm-bugs@lists.llvm.org, nlewycky@google.com, richard-llvm@metafoo.co.uk
Fixed by commit(s)
Attachments	builtin_fpclassify_codegen.patch (5495 bytes, text/plain) codegen_builtin_fpclassify.patch (2902 bytes, text/plain)
Blocks
Blocked by
See also

It seems that the gcc builtin '__builtin_fpclassify' is required
by boost while not available in clang++.

[Quuxplusone]

Which Boost header is using __builtin_fpclassify?

[Quuxplusone]

(In reply to comment #1)
> Which Boost header is using __builtin_fpclassify?
>

The include stack at the point of the error testifies:

In file included from boost-test.cpp:112:
In file included from /usr/include/boost/bimap.hpp:13:
In file included from /usr/include/boost/bimap/bimap.hpp:61:
In file included from /usr/include/boost/bimap/detail/bimap_core.hpp:38:
In file included from /usr/include/boost/bimap/relation/mutant_relation.hpp:26:
In file included from /usr/include/boost/functional/hash/hash.hpp:15:
In file included from /usr/include/boost/functional/detail/hash_float.hpp:25:
In file included from
/usr/include/boost/functional/detail/float_functions.hpp:9:
In file included from /usr/include/boost/config/no_tr1/cmath.hpp:21:
/usr/include/c++/4.4.1/cmath:499:14: error: use of undeclared identifier
'__builtin_fpclassify'
return __builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL,

So, I guess boost is including cmath.

Cheers,
Maurice

[Quuxplusone]

The only interesting thing I see about this is here:
http://archives.free.net.ph/message/20080523.164003.a971bed4.ja.html

Apparently this is used like this:

#define FP_NAN 1
#define FP_INF 2
#define FP_INFINITE 2
#define FP_NORMAL 3
#define FP_SUBNORMAL 4
#define FP_ZERO 5
__builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL, FP_SUBNORMAL, FP_ZERO, (X))

[Quuxplusone]

With this patch, we should be able to parse the header:
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20100118/026684.html

However, we'll still reject the code if something actually tries to codegen the
builtin.  Here's the GCC implementation of this feature if anyone is interested:
http://gcc.gnu.org/ml/gcc-patches/2008-05/msg01255.html

[Quuxplusone]

I've added Sema support for __builtin_fpclassify in r96291. I'll post a codegen patch soon.

[Quuxplusone]

Attached builtin_fpclassify_codegen.patch (5495 bytes, text/plain): first shot at codegen support

[Quuxplusone]

Hey Benjamin, it's probably best to email the patch to cfe-commits to get someone to review it, not many people watch bugzilla.

[Quuxplusone]

WebKit uses isfinite(), which also requires builtin support.

[Quuxplusone]

This isn't actually Boost-related, so it doesn't block Boost.

[Quuxplusone]

_Bug 6987 has been marked as a duplicate of this bug._

[Quuxplusone]

Just to confirm, this isn't blocking boost.

It is however blocking including on g++ 4.4.x and 4.5.x.

[Quuxplusone]

isinf is implemented in r103166

[Quuxplusone]

__builtin_isfinite is implemented in r103168

[Quuxplusone]

I added some todo's for isinf_sign and isnormal in r103169.  I don't plan to
implement these, and they are need fpclassify.  Benjamin, your patch looks
basically right, but please use the algorithms and APFloat methods mentioned in
the comment I added.  For fpclassify, I recommend actually emitting a branching
structure, like this:

if (x == 0.0) {
  res = ZERO;
} else {
  xabs = fabs(x);
  if (xabs == infinity)
    res = INFINITY;
  else if (xabs == xabs)
    res = NORMAL;
  else
    res = NAN;
}

Emitting actual branches is good for the branch predictor, codegen can decide
to fold together the blocks if it wants to.

[Quuxplusone]

__builtin_isnormal is implemented in r104118.

[Quuxplusone]

Attached codegen_builtin_fpclassify.patch (2902 bytes, text/plain): CodeGen __builtin_fpclassify implementation

[Quuxplusone]

(In reply to comment #16)
> I created an implementation of __builtin__fpclassify the way Chris Lattner
> described.

Very nice, commited as r105936.

[Quuxplusone]

On linux, linking a trivial program like:

int main(void)
{ return __builtin_fpclassify(1,2,3,4,5,1.0); }

produces the error:

undefined reference to 'fabs'

Which goes away if I add -lm. However, -lm is not required with gcc. This does
not occur on Darwin.

[Quuxplusone]

Just a FYI. gcc rejects calls to __builtin_fpclassify if any of the first 5 parameters (the constant list) are non-constant. Of course there is nothing wrong with allowing more values, I just wanted to check this case had been considered, as gcc purposefully rejects it.

[Quuxplusone]

Comment 18 is caused by the issue in bug 9019.