Support __float128 on NetBSD libstdc++ x86/x86_64

[krytarowski]

Summary

This adds support for __float128 from GNU libstdc++ with Clang on NetBSD x86/x86_64 targets.

This corrects compilation at least of CMake and probably others like Firefox.

In file included from /tmp/pkgsrc-tmp/devel/cmake/work/cmake-3.8.2/Source/kwsys/Directory.cxx:4:
In file included from /tmp/pkgsrc-tmp/devel/cmake/work/cmake-3.8.2/Source/cmsys/Directory.hxx:8:
In file included from /usr/include/g++/string:40:
In file included from /usr/include/g++/bits/char_traits.h:39:
In file included from /usr/include/g++/bits/stl_algobase.h:64:
In file included from /usr/include/g++/bits/stl_pair.h:59:
In file included from /usr/include/g++/bits/move.h:57:
/usr/include/g++/type_traits:311:39: error: __float128 is not supported on this target

struct __is_floating_point_helper<__float128>
                                  ^

Sponsored by <The NetBSD Foundation>

[joerg]

At the very least, missing test case.

I'm mostly ambivalent about this -- I don't really see the point and without matching soft float support it won't fully work either.

[krytarowski]

Softfloat on the NetBSD of Clang/LLVM part?

[joerg]

Soft-float on the runtime environment part. I.e. non-trivial operations are lowered to library calls.

[krytarowski]

Is this sufficient as a test:

$NetBSD$

--- test/Sema/128bitfloat.cpp.orig	2017-05-29 10:06:54.000000000 +0000
+++ test/Sema/128bitfloat.cpp
@@ -4,6 +4,8 @@
 // RUN: %clang_cc1 -triple i686-windows-gnu -verify -std=c++11 %s
 // RUN: %clang_cc1 -triple x86_64-windows-gnu -verify -std=c++11 %s
 // RUN: %clang_cc1 -triple x86_64-windows-msvc -verify -std=c++11 %s
+// RUN: %clang_cc1 -triple i386--netbsd -verify -std=c++11 %s
+// RUN: %clang_cc1 -triple x86_64--netbsd -verify -std=c++11 %s
 
 #if defined(__FLOAT128__) || defined(__SIZEOF_FLOAT128__)
 __float128 f;
``

[krytarowski]

ping

[joerg]

As I said, I don't see the point in pretending we support float128 when the runtime doesn't contain the necessary pieces.

[krytarowski]

In D34018#783218, @joerg wrote:

As I said, I don't see the point in pretending we support float128 when the runtime doesn't contain the necessary pieces.

On the other hand cmake and other packages break now. What's the alternative?

[krytarowski]

We can patch libstdc++ to conditionally enable float128 on NetBSD, this works too. All other systems I'm aware of including cygwin moved to this->HasFloat128 = true;.

[krytarowski]

We can change NetBSD base libstdc++ in /usr/include/g++/bits/*/c++config.h from

/* Define if __float128 is supported on this host. */
#define _GLIBCXX_USE_FLOAT128 1

to:

#if !defined(__clang__)
/* Define if __float128 is supported on this host. */
#define _GLIBCXX_USE_FLOAT128 1
#endif

[rnk]

I think there's something wrong with clang's implementation of float128. It seems fundamentally incompatible with libstdc++'s is_floating_point_helper mechanism. We can definitely take this patch to unblock things, but we should file a bug about investigating this all the way. This popped up in mingw, I forget elsewhere, and it's wasting peoples' time.

[krytarowski]

Joerg declines to take this patch as it is, as there is no compiler-rt support for float128 on i386 in compiler-rt. GCC knows float128 on i386, amd64, i64, hppa, powerpc [some variations]; currently I focus on i386/amd64.

My proposal was to check if there is used libstdc++, and conditionally enable __float128 for i386/amd64?

[krytarowski]

Rescheduled for future. Maybe I will need to check and improve how good is this i386 __float128 support.

[krytarowski]

This change happened to be required in downstream usage.. knowing its limits can we merge it as it is?

[joerg]

Nothing changed. I don't see how catering to the broken libstdc++ self-configuration helps. So no, I still object to this change.

[krytarowski]

OK.

lib/Basic/Targets.cpp

protected:
   void getOSDefines(const LangOptions &Opts, const llvm::Triple &Triple,
                     MacroBuilder &Builder) const override {
     // NetBSD defines; list based off of gcc output
     Builder.defineMacro("__NetBSD__");
     Builder.defineMacro("__unix__");
     Builder.defineMacro("__ELF__");
     if (Opts.POSIXThreads)
       Builder.defineMacro("_REENTRANT");
+    if (this->HasFloat128)
+      Builder.defineMacro("__FLOAT128__");
 
     switch (Triple.getArch()) {
     default:
       break;
     case llvm::Triple::arm:
     case llvm::Triple::armeb:
     case llvm::Triple::thumb:
     case llvm::Triple::thumbeb:
       Builder.defineMacro("__ARM_DWARF_EH__");
       break;
     }
   }
 public:
   NetBSDTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts)
       : OSTargetInfo<Target>(Triple, Opts) {
     this->MCountName = "_mcount";
+
+    switch (Triple.getArch()) {
+    default:
+      break;
+    case llvm::Triple::x86:
+    case llvm::Triple::x86_64:
+      this->HasFloat128 = true;
+      break;
+    }
   }
 };
 
 // OpenBSD Target
 template<typename Target>
 class OpenBSDTargetInfo : public OSTargetInfo<Target> {
 protected:
   void getOSDefines(const LangOptions &Opts, const llvm::Triple &Triple,