[AArch64] Support conversion between fp16 and fp128

[zatrazz]

Summary

This is an updated version of https://reviews.llvm.org/D50685 .
The main changes are that

• I have added a CMake test to check if compiler support _Float16 to avoid either to require a more recent gcc to build a stage1 compiler and avoid creating potentially invalid soft-fp symbols (as for ARM fp16 one when built with a compiler configured for hard-float).
• Adjusted the internal tests to use _Float16 as well if compiler supports.

This issue has been brought by some developer from Apache TVM, where is
trigger an compiler ICE with the missing fp conversion.

---

This patch adds both extendhftf2 and trunctfhf2 to support
conversion between half-precision and quad-precision floating-point
values. They are enabled iff the compiler supports _Float16.

It also adjust the extendhfsf2, truncdfhf2 __truncsfhf2 to use
_Float16 when compiler supports it. On AArch64 it allows use the
native FP16 ABI, while on other architectures the expected current
semantic is preserved (arm for instance).

[atrosinenko]

compiler-rt/lib/builtins/extendhftf2.c

1

extendhfsf2.c should probably be extendhftf2.c.

compiler-rt/test/builtins/Unit/fp_test.h

85

Shouldn't it have (TYPE_FP16 result, uint16_t expected) arguments like other compareResultX functions (as if fp_t result, rep_t expected)? The same for test_truncXfhf2(... , uint16_t expected): if I get it right, the common tradition for them is expressing the reference value as corresponding uintXX_t. Meanwhile, this is probably related to your changes to second test case from extendhfsf2_test.c - the original sources look quite suspicious to me.

[zatrazz]

compiler-rt/lib/builtins/extendhftf2.c

1

Indeed, I will update it.

compiler-rt/test/builtins/Unit/fp_test.h

85

Yeah, I agree it should follow the other compare functions and use a uintxx_t as second argument, I will fix it. The extendhfsf2_test.c issue is in fact that it is using a wrong comparison function that is not making explicit what it is currently testing. The test__extendhfsf2 has a 'float' as second argument, however calls compareResultH which expectes a uint16_t (without this change). This makes the compiler to silent cast the result and the expected results could be misleading.

I think it should call compareResultF with toRep32 instead and to make explicit it compares against the uint32_t representation of the float, same as extenddftf2_test.c and extendsftf2_test.c are already doing. I will change it as well.

[zatrazz]

Updated version based on previous comments.

[atrosinenko]

compiler-rt/lib/builtins/extendhftf2.c

14

Technically, linter is used to complain about \ being not at the 80th column for multi-line macroses, but I'm not sure all these rules apply to compiler-rt/builtins library.

compiler-rt/test/builtins/Unit/fp_test.h

6–10

This could probably be moved to fp_lib.h (or to int_types.h - looks like "int" means "internal" there...), possibly switching to typedefs.

This should make other code more idiomatic:

• there would be single-line typedefs for src_t and dst_t in fp_extend.h and fp_trunc.h, just as for other precisions
• the precision converting LibCalls (extend / trunc) would look more idiomatically:

COMPILER_RT_ABI concrete_return_type __libcallXfYf2(concrete_argument_type) { ... }

... with one of those types being TYPE_F16 alias instead of src_t / dst_t.

85

It would probably be even more idiomatically to declare test__extendhfsf2 with the second argument being uint32_t expected, but this may cause quite large changes that are not strictly related to this patch.

[zatrazz]

compiler-rt/lib/builtins/extendhftf2.c

14

Right, I will fix it.

compiler-rt/test/builtins/Unit/fp_test.h

6–10

The fp_lib.h now only defines a single floating-point type (one must define either SINGLE_PRECISION, DOUBLE_PRECISION, or QUAD_PRECISION). For instance on extendsftf2.c, it defines QUAD_PRECISION and includes fp_lib.h. To move the float16 definition to fp_lib.h it would require to add the possibility to define multiple types, similar to what is already done to fp_extend.h/fp_trunc.h. I am not sure if is better, the second fp type is used currently on the extend/trunc builtins.

And I don' t think int_types.h is the correct place, afaiks _Float16 currently not really defined on all architectures so its not really related to an interger type.

Ideally I think compiler-rt should define *all* 'hf' builtins to use _Float16 and built them iff the ABI implements the type (meaning the compiler actually emits libcalls to it). For ABI that support float16 operations without supporting _Float16 type, for instance ARM which supports __fp16, it would be better to move the libcall implementation to the arch-specific folders.

85

I think it should be feabile, it would require to redefine all the inputs in the test below though.

[zatrazz]

Hi atrosinenko, do you think this patch need any more change on the testing side?
The fp_lib.h/int_lib.h change would most likely require in a more complex without
much gain in organization imho.

[atrosinenko]

In D86453#2241702, @zatrazz wrote:

Hi atrosinenko, do you think this patch need any more change on the testing side?
The fp_lib.h/int_lib.h change would most likely require in a more complex without
much gain in organization imho.

Sorry for the delay. Provided the particular testcases are correct, I have found just a single fromRep16 that was probably used instead of toRep16. Other comments are merely random thoughts just in case.

compiler-rt/test/builtins/Unit/extendhfsf2_test.c

10

As previously discussed, I would rather use uint32_t expected instead. But after all, the objective of this patch is not to fix test__extendhfsf2, so even if this would be implemented, then it should probably go to a separate patch and not clutter this one. Commenting this just for completeness.

17

toRep16(a) is probably expected.

compiler-rt/test/builtins/Unit/fp_test.h

6–10

Agree, looks like there is no more suitable place for that define right now.

18

If I get it right, this could be performed unconditionally. On the other hand, the #else branch may be either a good illustration for peculiarities of this function when no native _Float16 is available or some misleading stuff...

[zatrazz]

compiler-rt/test/builtins/Unit/extendhfsf2_test.c

10

Using uint32_t indeed seems a better approach and its aligns somewhat to the change to use _Float16 where applicable. I will send a updated version with this fix.

17

Ack.

compiler-rt/test/builtins/Unit/fp_test.h

18

I don't have a strong preference either, it should be optimized away by the compiler for !COMPILER_RT_HAS_FLOAT16 anyway. I will keep to make it explicit that for COMPILER_RT_HAS_FLOAT16 TYPE_FP16 is expected to be different tha uint16_t.

[zatrazz]

Updated patch based on previous comments.

[MaskRay]

compiler-rt/lib/builtins/fp_extend.h

43

#ifdef ? ditto below

compiler-rt/test/builtins/Unit/trunctfhf2_test.c

16

Use 2-space indentation. Place { in the end. There is no need following the style of some violating files in this directory.

[zatrazz]

I have adapted the news files using clang-format and fixed minor style issues pointed by previous comments.

[zatrazz]

Ping.

[MaskRay]

compiler-rt/lib/builtins/fp_extend.h

43

You can mark the comment above as "Done".

_Float16 and uint16_t are different types. Do all the *hf* functions change their signatures when _Float16 is supported? This still looks strange.

[zatrazz]

compiler-rt/lib/builtins/fp_extend.h

43

I think using _Float16 for generic support does make more sense, since float16 support is really an extension and with different calling convention on each architecture and even on some different ABIs. While it is uint16_t for armv6 due the soft-fp ABI, it is a complete different type with a different calling convention for aarch64. And I expect that other architectures to use similar strategies now that it is supported on some vector extensions in some new chips (POWER10 for instance).

Also, for ABI which expects *hf* to have 'uint16_t' as calling convetion it would be better to compartmentalize it to the specific arch folder, for instance for arm (and this is what libgcc does for instance).

[MaskRay]

compiler-rt/lib/builtins/fp_extend.h

43

If 'uint16_t' is more of an anomaly, @ab should the arm softfp implementation be moved to the specific arch folder?

[zatrazz]

compiler-rt/lib/builtins/fp_extend.h

43

I would say so, although I am not fully sure if any other architecture does use such routines (which in this case would need to add such routines as well).

[zatrazz]

Ping.

[zatrazz]

Ping (x2).

[zatrazz]

Ping (x3).

compiler-rt/cmake/builtin-config-ix.cmake

 builtin_check_c_compiler_source(COMPILER_RT_HAS_ATOMIC_KEYWORD
 "
 int foo(int x, int y) {
  _Atomic int result = x * y;
  return result;
 }
 ")
 
+builtin_check_c_compiler_source(COMPILER_RT_HAS_FLOAT16
+"
+_Float16 foo(_Float16 x) {
+ return x;
+}
+"
+)
 
 set(ARM64 aarch64)
 set(ARM32 arm armhf armv6m armv7m armv7em armv7 armv7s armv7k)
 set(HEXAGON hexagon)
 set(X86 i386)
 set(X86_64 x86_64)
 set(MIPS32 mips mipsel)
 set(MIPS64 mips64 mips64el)

compiler-rt/lib/builtins/CMakeLists.txt

 # TODO: Several "tf" files (and divtc3.c, but not multc3.c) are in
 # GENERIC_SOURCES instead of here.
 set(GENERIC_TF_SOURCES
   addtf3.c
   comparetf2.c
   divtc3.c
   divtf3.c
   extenddftf2.c
+  extendhftf2.c
   extendsftf2.c
   fixtfdi.c
   fixtfsi.c
   fixtfti.c
   fixunstfdi.c
   fixunstfsi.c
   fixunstfti.c
   floatditf.c
   floatsitf.c
   floattitf.c
   floatunditf.c
   floatunsitf.c
   floatuntitf.c
   multc3.c
   multf3.c
   powitf2.c
   subtf3.c
   trunctfdf2.c
+  trunctfhf2.c
   trunctfsf2.c
 )
 
 option(COMPILER_RT_EXCLUDE_ATOMIC_BUILTIN
   "Skip the atomic builtin (these should normally be provided by a shared library)"
   On)
 
 if(NOT FUCHSIA AND NOT COMPILER_RT_BAREMETAL_BUILD)

 
 if (APPLE)
   add_subdirectory(Darwin-excludes)
   add_subdirectory(macho_embedded)
   darwin_add_builtin_libraries(${BUILTIN_SUPPORTED_OS})
 else ()
   set(BUILTIN_CFLAGS "")
 
+  append_list_if(COMPILER_RT_HAS_FLOAT16 -DCOMPILER_RT_HAS_FLOAT16 BUILTIN_CFLAGS)
+
   append_list_if(COMPILER_RT_HAS_STD_C11_FLAG -std=c11 BUILTIN_CFLAGS)
 
   # These flags would normally be added to CMAKE_C_FLAGS by the llvm
   # cmake step. Add them manually if this is a standalone build.
   if(COMPILER_RT_STANDALONE_BUILD)
     if(COMPILER_RT_BUILTINS_ENABLE_PIC)
       append_list_if(COMPILER_RT_HAS_FPIC_FLAG -fPIC BUILTIN_CFLAGS)
     endif()

compiler-rt/lib/builtins/extendhfsf2.c

 //===-- lib/extendhfsf2.c - half -> single conversion -------------*- C -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #define SRC_HALF
 #define DST_SINGLE
 #include "fp_extend_impl.inc"
 
 // Use a forwarding definition and noinline to implement a poor man's alias,
 // as there isn't a good cross-platform way of defining one.
-COMPILER_RT_ABI NOINLINE float __extendhfsf2(uint16_t a) {
+COMPILER_RT_ABI NOINLINE float __extendhfsf2(src_t a) {
   return __extendXfYf2__(a);
 }
 
-COMPILER_RT_ABI float __gnu_h2f_ieee(uint16_t a) { return __extendhfsf2(a); }
+COMPILER_RT_ABI float __gnu_h2f_ieee(src_t a) { return __extendhfsf2(a); }
 
 #if defined(__ARM_EABI__)
 #if defined(COMPILER_RT_ARMHF_TARGET)
-AEABI_RTABI float __aeabi_h2f(uint16_t a) { return __extendhfsf2(a); }
+AEABI_RTABI float __aeabi_h2f(src_t a) { return __extendhfsf2(a); }
 #else
 COMPILER_RT_ALIAS(__extendhfsf2, __aeabi_h2f)
 #endif
 #endif

compiler-rt/lib/builtins/extendhftf2.c

+//===-- lib/extendhfsf2.c - half -> quad conversion ---------------*- C -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT) \
+    && defined(COMPILER_RT_HAS_FLOAT16)
+#define SRC_HALF
+#define DST_QUAD
+#include "fp_extend_impl.inc"
+
+COMPILER_RT_ABI long double __extendhftf2(src_t a) {
+    return __extendXfYf2__(a);
+}
+
+#endif

compiler-rt/lib/builtins/fp_extend.h

#else
   if (a & REP_C(0xffffffff00000000))
     return __builtin_clz(a >> 32);
   else
     return 32 + __builtin_clz(a & REP_C(0xffffffff));
 #endif
 }
 
 #elif defined SRC_HALF
+#if defined COMPILER_RT_HAS_FLOAT16
+typedef _Float16 src_t;
+#else
 typedef uint16_t src_t;
+#endif
 typedef uint16_t src_rep_t;
 #define SRC_REP_C UINT16_C
 static const int srcSigBits = 10;
 #define src_rep_t_clz __builtin_clz
 
 #else
 #error Source should be half, single, or double precision!
 #endif // end source precision
 
 #if defined DST_SINGLE
 typedef float dst_t;
 typedef uint32_t dst_rep_t;
 #define DST_REP_C UINT32_C
 static const int dstSigBits = 23;
 
 #elif defined DST_DOUBLE
 typedef double dst_t;
 typedef uint64_t dst_rep_t;
 #define DST_REP_C UINT64_C
 static const int dstSigBits = 52;
 
 #elif defined DST_QUAD
 typedef long double dst_t;
 typedef __uint128_t dst_rep_t;
 #define DST_REP_C (__uint128_t)
 static const int dstSigBits = 112;
 
 #else
 #error Destination should be single, double, or quad precision!
 #endif // end destination precision
 
 // End of specialization parameters.  Two helper routines for conversion to and
 // from the representation of floating-point data as integer values follow.
 
 static __inline src_rep_t srcToRep(src_t x) {
   const union {
     src_t f;
     src_rep_t i;
   } rep = {.f = x};
   return rep.i;
 }
 
 static __inline dst_t dstFromRep(dst_rep_t x) {
   const union {
     dst_t f;
     dst_rep_t i;
   } rep = {.i = x};
   return rep.f;
 }
 // End helper routines.  Conversion implementation follows.
 
 #endif // FP_EXTEND_HEADER

compiler-rt/lib/builtins/fp_trunc.h

 
 #elif defined SRC_QUAD
 typedef long double src_t;
 typedef __uint128_t src_rep_t;
 #define SRC_REP_C (__uint128_t)
 static const int srcSigBits = 112;
 
 #else
 #error Source should be double precision or quad precision!
 #endif // end source precision
 
 #if defined DST_DOUBLE
 typedef double dst_t;
 typedef uint64_t dst_rep_t;
 #define DST_REP_C UINT64_C
 static const int dstSigBits = 52;
 
 #elif defined DST_SINGLE
 typedef float dst_t;
 typedef uint32_t dst_rep_t;
 #define DST_REP_C UINT32_C
 static const int dstSigBits = 23;
 
 #elif defined DST_HALF
+#if defined COMPILER_RT_HAS_FLOAT16
+typedef _Float16 dst_t;
+#else
 typedef uint16_t dst_t;
+#endif
 typedef uint16_t dst_rep_t;
 #define DST_REP_C UINT16_C
 static const int dstSigBits = 10;
 
 #else
 #error Destination should be single precision or double precision!
 #endif // end destination precision
 
 // End of specialization parameters.  Two helper routines for conversion to and
 // from the representation of floating-point data as integer values follow.
 
 static __inline src_rep_t srcToRep(src_t x) {
   const union {
     src_t f;
     src_rep_t i;
   } rep = {.f = x};
   return rep.i;
 }
 
 static __inline dst_t dstFromRep(dst_rep_t x) {
   const union {
     dst_t f;
     dst_rep_t i;
   } rep = {.i = x};
   return rep.f;
 }
 
 #endif // FP_TRUNC_HEADER

compiler-rt/lib/builtins/truncdfhf2.c

 //===-- lib/truncdfhf2.c - double -> half conversion --------------*- C -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #define SRC_DOUBLE
 #define DST_HALF
 #include "fp_trunc_impl.inc"
 
-COMPILER_RT_ABI uint16_t __truncdfhf2(double a) { return __truncXfYf2__(a); }
+COMPILER_RT_ABI dst_t __truncdfhf2(double a) { return __truncXfYf2__(a); }
 
 #if defined(__ARM_EABI__)
 #if defined(COMPILER_RT_ARMHF_TARGET)
-AEABI_RTABI uint16_t __aeabi_d2h(double a) { return __truncdfhf2(a); }
+AEABI_RTABI dst_t __aeabi_d2h(double a) { return __truncdfhf2(a); }
 #else
 COMPILER_RT_ALIAS(__truncdfhf2, __aeabi_d2h)
 #endif
 #endif

compiler-rt/lib/builtins/truncsfhf2.c

 //===-- lib/truncsfhf2.c - single -> half conversion --------------*- C -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #define SRC_SINGLE
 #define DST_HALF
 #include "fp_trunc_impl.inc"
 
 // Use a forwarding definition and noinline to implement a poor man's alias,
 // as there isn't a good cross-platform way of defining one.
-COMPILER_RT_ABI NOINLINE uint16_t __truncsfhf2(float a) {
+COMPILER_RT_ABI NOINLINE dst_t __truncsfhf2(float a) {
   return __truncXfYf2__(a);
 }
 
-COMPILER_RT_ABI uint16_t __gnu_f2h_ieee(float a) { return __truncsfhf2(a); }
+COMPILER_RT_ABI dst_t __gnu_f2h_ieee(float a) { return __truncsfhf2(a); }
 
 #if defined(__ARM_EABI__)
 #if defined(COMPILER_RT_ARMHF_TARGET)
-AEABI_RTABI uint16_t __aeabi_f2h(float a) { return __truncsfhf2(a); }
+AEABI_RTABI dst_t __aeabi_f2h(float a) { return __truncsfhf2(a); }
 #else
 COMPILER_RT_ALIAS(__truncsfhf2, __aeabi_f2h)
 #endif
 #endif

compiler-rt/lib/builtins/trunctfhf2.c

+//===-- lib/trunctfhf2.c - quad -> half conversion ----------------*- C -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT) \
+    && defined(COMPILER_RT_HAS_FLOAT16)
+#define SRC_QUAD
+#define DST_HALF
+#include "fp_trunc_impl.inc"
+
+COMPILER_RT_ABI dst_t __trunctfhf2(long double a) {
+  return __truncXfYf2__(a);
+}
+
+#endif

compiler-rt/test/builtins/CMakeLists.txt

if (${arch} STREQUAL "armhf")
     string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")
   endif()
 
   if (${arch} STREQUAL "riscv32")
     list(APPEND BUILTINS_TEST_TARGET_CFLAGS -fforce-enable-int128)
     string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")
   endif()
 
+  append_list_if(COMPILER_RT_HAS_FLOAT16 -DCOMPILER_RT_HAS_FLOAT16 BUILTINS_TEST_TARGET_CFLAGS)
+
   # Compute builtins available in library and add them as lit features.
   if(APPLE)
     # TODO: Support other Apple platforms.
     set(BUILTIN_LIB_TARGET_NAME "clang_rt.builtins_${arch}_osx")
   else()
     set(BUILTIN_LIB_TARGET_NAME "clang_rt.builtins-${arch}")
   endif()
   if (NOT TARGET "${BUILTIN_LIB_TARGET_NAME}")

compiler-rt/test/builtins/Unit/extendhfsf2_test.c

 // RUN: %clang_builtins %s %librt -o %t && %run %t
 // REQUIRES: librt_has_extendhfsf2
 
 #include <stdio.h>
 
 #include "fp_test.h"
 
-float __extendhfsf2(uint16_t a);
+float __extendhfsf2(TYPE_FP16 a);
 
-int test__extendhfsf2(uint16_t a, float expected)
+int test__extendhfsf2(TYPE_FP16 a, float expected)
 {
-    float x = __extendhfsf2(a);
+    TYPE_FP16 x = __extendhfsf2(a);
     int ret = compareResultH(x, expected);
 
     if (ret){
         printf("error in test__extendhfsf2(%#.4x) = %f, "
-               "expected %f\n", a, x, expected);
+               "expected %f\n", fromRep16(a), x, expected);
     }
     return ret;
 }
 
-char assumption_1[sizeof(__fp16) * CHAR_BIT == 16] = {0};
+char assumption_1[sizeof(TYPE_FP16) * CHAR_BIT == 16] = {0};
 
 int main()
 {
     // qNaN
-    if (test__extendhfsf2(UINT16_C(0x7e00),
+    if (test__extendhfsf2(fromRep16(0x7e00),
                           makeQNaN32()))
         return 1;
     // NaN
-    if (test__extendhfsf2(UINT16_C(0x7e00),
-                          makeNaN32(UINT32_C(0x8000))))
+    if (test__extendhfsf2(fromRep16(0x7d00),
+                          makeQNaN32()))
         return 1;
     // inf
-    if (test__extendhfsf2(UINT16_C(0x7c00),
+    if (test__extendhfsf2(fromRep16(0x7c00),
                           makeInf32()))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0xfc00),
+    if (test__extendhfsf2(fromRep16(0xfc00),
                           -makeInf32()))
         return 1;
     // zero
-    if (test__extendhfsf2(UINT16_C(0x0),
+    if (test__extendhfsf2(fromRep16(0x0),
                           0.0f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x8000),
+    if (test__extendhfsf2(fromRep16(0x8000),
                           -0.0f))
         return 1;
-
-    if (test__extendhfsf2(UINT16_C(0x4248),
+    if (test__extendhfsf2(fromRep16(0x4248),
                           3.1415926535f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0xc248),
+    if (test__extendhfsf2(fromRep16(0xc248),
                           -3.1415926535f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x7c00),
+    if (test__extendhfsf2(fromRep16(0x7c00),
                           0x1.987124876876324p+100f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x6e62),
+    if (test__extendhfsf2(fromRep16(0x6e62),
                           0x1.988p+12f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x3c00),
+    if (test__extendhfsf2(fromRep16(0x3c00),
                           0x1.0p+0f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x0400),
+    if (test__extendhfsf2(fromRep16(0x0400),
                           0x1.0p-14f))
         return 1;
     // denormal
-    if (test__extendhfsf2(UINT16_C(0x0010),
+    if (test__extendhfsf2(fromRep16(0x0010),
                           0x1.0p-20f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x0001),
+    if (test__extendhfsf2(fromRep16(0x0001),
                           0x1.0p-24f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x8001),
+    if (test__extendhfsf2(fromRep16(0x8001),
                           -0x1.0p-24f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x0001),
+    if (test__extendhfsf2(fromRep16(0x0001),
                           0x1.5p-25f))
         return 1;
     // and back to zero
-    if (test__extendhfsf2(UINT16_C(0x0000),
+    if (test__extendhfsf2(fromRep16(0x0000),
                           0x1.0p-25f))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0x8000),
+    if (test__extendhfsf2(fromRep16(0x8000),
                           -0x1.0p-25f))
         return 1;
     // max (precise)
-    if (test__extendhfsf2(UINT16_C(0x7bff),
+    if (test__extendhfsf2(fromRep16(0x7bff),
                           65504.0f))
         return 1;
     // max (rounded)
-    if (test__extendhfsf2(UINT16_C(0x7bff),
+    if (test__extendhfsf2(fromRep16(0x7bff),
                           65504.0f))
         return 1;
     // max (to +inf)
-    if (test__extendhfsf2(UINT16_C(0x7c00),
+    if (test__extendhfsf2(fromRep16(0x7c00),
                           makeInf32()))
         return 1;
-    if (test__extendhfsf2(UINT16_C(0xfc00),
+    if (test__extendhfsf2(fromRep16(0xfc00),
                           -makeInf32()))
         return 1;
     return 0;
 }

compiler-rt/test/builtins/Unit/extendhftf2_test.c

+// RUN: %clang_builtins %s %librt -o %t && %run %t
+// REQUIRES: librt_has_extendhftf2
+
+#include "int_lib.h"
+#include <stdio.h>
+
+#if __LDBL_MANT_DIG__ == 113 && defined (COMPILER_RT_HAS_FLOAT16)
+
+#include "fp_test.h"
+
+COMPILER_RT_ABI long double __extendhftf2(TYPE_FP16 a);
+
+int test__extendhftf2(TYPE_FP16 a, uint64_t expectedHi, uint64_t expectedLo)
+{
+    long double x = __extendhftf2(a);
+    int ret = compareResultLD(x, expectedHi, expectedLo);
+
+    if (ret)
+    {
+        printf("error in test__extendhftf2(%#.4x) = %.20Lf, "
+               "expected %.20Lf\n", toRep16(a), x,
+               fromRep128(expectedHi, expectedLo));
+    }
+    return ret;
+}
+
+char assumption_1[sizeof(TYPE_FP16) * CHAR_BIT == 16] = {0};
+
+#endif
+
+int main()
+{
+#if __LDBL_MANT_DIG__ == 113 && defined (COMPILER_RT_HAS_FLOAT16)
+    // qNaN
+    if (test__extendhftf2(makeQNaN16(),
+                          UINT64_C(0x7fff800000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    // NaN
+    if (test__extendhftf2(makeNaN16(UINT16_C(0x0100)),
+                          UINT64_C(0x7fff400000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    // inf
+    if (test__extendhftf2(makeInf16(),
+                          UINT64_C(0x7fff000000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    if (test__extendhftf2(-makeInf16(),
+                          UINT64_C(0xffff000000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    // zero
+    if (test__extendhftf2(fromRep16(0x0U),
+                          UINT64_C(0x0), UINT64_C(0x0)))
+        return 1;
+    if (test__extendhftf2(fromRep16(0x8000U),
+                          UINT64_C(0x8000000000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    // denormal
+    if (test__extendhftf2(fromRep16(0x0010U),
+                          UINT64_C(0x3feb000000000000),
+                          UINT64_C(0x0000000000000000)))
+        return 1;
+    if (test__extendhftf2(fromRep16(0x0001U),
+                          UINT64_C(0x3fe7000000000000),
+                          UINT64_C(0x0000000000000000)))
+        return 1;
+    if (test__extendhftf2(fromRep16(0x8001U),
+                          UINT64_C(0xbfe7000000000000),
+                          UINT64_C(0x0000000000000000)))
+        return 1;
+
+    // pi
+    if (test__extendhftf2(fromRep16(0x4248U),
+                          UINT64_C(0x4000920000000000),
+                          UINT64_C(0x0000000000000000)))
+        return 1;
+    if (test__extendhftf2(fromRep16(0xc248U),
+                          UINT64_C(0xc000920000000000),
+                          UINT64_C(0x0000000000000000)))
+        return 1;
+
+    if (test__extendhftf2(fromRep16(0x508cU),
+                          UINT64_C(0x4004230000000000),
+                          UINT64_C(0x0)))
+        return 1;
+    if (test__extendhftf2(fromRep16(0x1bb7U),
+                          UINT64_C(0x3ff6edc000000000),
+                          UINT64_C(0x0)))
+        return 1;
+#else
+    printf("skipped\n");
+#endif
+    return 0;
+}

compiler-rt/test/builtins/Unit/fp_test.h

 #include <stdlib.h>
 #include <limits.h>
 #include <string.h>
 #include <stdint.h>
 
+#ifdef COMPILER_RT_HAS_FLOAT16
+#define TYPE_FP16   _Float16
+#else
+#define TYPE_FP16   uint16_t
+#endif
+
 enum EXPECTED_RESULT {
     LESS_0, LESS_EQUAL_0, EQUAL_0, GREATER_0, GREATER_EQUAL_0, NEQUAL_0
 };
 
-static inline uint16_t fromRep16(uint16_t x)
+static inline TYPE_FP16 fromRep16(uint16_t x)
 {
+#ifdef COMPILER_RT_HAS_FLOAT16
+    TYPE_FP16 ret;
+    memcpy (&ret, &x, sizeof (ret));
+    return ret;
+#else
     return x;
+#endif
 }
 
 static inline float fromRep32(uint32_t x)
 {
     float ret;
     memcpy(&ret, &x, 4);
     return ret;
 }

 {
     __uint128_t x = ((__uint128_t)hi << 64) + lo;
     long double ret;
     memcpy(&ret, &x, 16);
     return ret;
 }
 #endif
 
-static inline uint16_t toRep16(uint16_t x)
+static inline uint16_t toRep16(TYPE_FP16 x)
 {
+#ifdef COMPILER_RT_HAS_FLOAT16
+    uint16_t ret;
+    memcpy (&ret, &x, sizeof (ret));
+    return ret;
+#else
     return x;
+#endif
 }
 
 static inline uint32_t toRep32(float x)
 {
     uint32_t ret;
     memcpy(&ret, &x, 4);
     return ret;
 }

 static inline __uint128_t toRep128(long double x)
 {
     __uint128_t ret;
     memcpy(&ret, &x, 16);
     return ret;
 }
 #endif
 
-static inline int compareResultH(uint16_t result,
-                                 uint16_t expected)
+static inline int compareResultH(TYPE_FP16 result,
+                                 TYPE_FP16 expected)
 {
     uint16_t rep = toRep16(result);
+    uint16_t exp = toRep16(expected);
 
-    if (rep == expected){
+    if (rep == exp){
         return 0;
     }
     // test other possible NaN representation(signal NaN)
-    else if (expected == 0x7e00U){
+    else if (exp == 0x7e00U){
         if ((rep & 0x7c00U) == 0x7c00U &&
             (rep & 0x3ffU) > 0){
             return 0;
         }
     }
     return 1;
 }
 
switch(expected){
         case GREATER_0:
             return ">0";
         default:
             return "";
     }
     return "";
 }
 
-static inline uint16_t makeQNaN16(void)
+static inline TYPE_FP16 makeQNaN16(void)
 {
     return fromRep16(0x7e00U);
 }
 
 static inline float makeQNaN32(void)
 {
     return fromRep32(0x7fc00000U);
 }
 
 static inline double makeQNaN64(void)
 {
     return fromRep64(0x7ff8000000000000UL);
 }
 
 #if __LDBL_MANT_DIG__ == 113
 static inline long double makeQNaN128(void)
 {
     return fromRep128(0x7fff800000000000UL, 0x0UL);
 }
 #endif
 
-static inline uint16_t makeNaN16(uint16_t rand)
+static inline TYPE_FP16 makeNaN16(uint16_t rand)
 {
     return fromRep16(0x7c00U | (rand & 0x7fffU));
 }
 
 static inline float makeNaN32(uint32_t rand)
 {
     return fromRep32(0x7f800000U | (rand & 0x7fffffU));
 }
 
 static inline double makeNaN64(uint64_t rand)
 {
     return fromRep64(0x7ff0000000000000UL | (rand & 0xfffffffffffffUL));
 }
 
 #if __LDBL_MANT_DIG__ == 113
 static inline long double makeNaN128(uint64_t rand)
 {
     return fromRep128(0x7fff000000000000UL | (rand & 0xffffffffffffUL), 0x0UL);
 }
 #endif
 
-static inline uint16_t makeInf16(void)
+static inline TYPE_FP16 makeInf16(void)
 {
     return fromRep16(0x7c00U);
 }
 
 static inline float makeInf32(void)
 {
     return fromRep32(0x7f800000U);
 }

compiler-rt/test/builtins/Unit/truncdfhf2_test.c

 // RUN: %clang_builtins %s %librt -o %t && %run %t
 // REQUIRES: librt_has_truncdfhf2
 
 #include <stdio.h>
 
 #include "fp_test.h"
 
-uint16_t __truncdfhf2(double a);
+TYPE_FP16 __truncdfhf2(double a);
 
-int test__truncdfhf2(double a, uint16_t expected)
+int test__truncdfhf2(double a, TYPE_FP16 expected)
 {
-    uint16_t x = __truncdfhf2(a);
+    TYPE_FP16 x = __truncdfhf2(a);
     int ret = compareResultH(x, expected);
 
     if (ret){
-        printf("error in test__truncdfhf2(%f) = %#.4x, "
-               "expected %#.4x\n", a, x, fromRep16(expected));
+        printf("error in test__truncdfhf2(%lf) = %#.4x, "
+               "expected %#.4x\n", a, fromRep16(x), fromRep16(expected));
     }
     return ret;
 }
 
 char assumption_1[sizeof(__fp16) * CHAR_BIT == 16] = {0};
 
 int main()
 {
     // qNaN
     if (test__truncdfhf2(makeQNaN64(),
-                         UINT16_C(0x7e00)))
+                         fromRep16(0x7e00)))
         return 1;
     // NaN
     if (test__truncdfhf2(makeNaN64(UINT64_C(0x8000)),
-                         UINT16_C(0x7e00)))
+                         fromRep16(0x7e00)))
         return 1;
     // inf
     if (test__truncdfhf2(makeInf64(),
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncdfhf2(-makeInf64(),
-                         UINT16_C(0xfc00)))
+                         fromRep16(0xfc00)))
         return 1;
     // zero
-    if (test__truncdfhf2(0.0, UINT16_C(0x0)))
+    if (test__truncdfhf2(0.0, fromRep16(0x0)))
         return 1;
-    if (test__truncdfhf2(-0.0, UINT16_C(0x8000)))
+    if (test__truncdfhf2(-0.0, fromRep16(0x8000)))
         return 1;
 
     if (test__truncdfhf2(3.1415926535,
-                         UINT16_C(0x4248)))
+                         fromRep16(0x4248)))
         return 1;
     if (test__truncdfhf2(-3.1415926535,
-                         UINT16_C(0xc248)))
+                         fromRep16(0xc248)))
         return 1;
     if (test__truncdfhf2(0x1.987124876876324p+1000,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncdfhf2(0x1.987124876876324p+12,
-                         UINT16_C(0x6e62)))
+                         fromRep16(0x6e62)))
         return 1;
     if (test__truncdfhf2(0x1.0p+0,
-                         UINT16_C(0x3c00)))
+                         fromRep16(0x3c00)))
         return 1;
     if (test__truncdfhf2(0x1.0p-14,
-                         UINT16_C(0x0400)))
+                         fromRep16(0x0400)))
         return 1;
     // denormal
     if (test__truncdfhf2(0x1.0p-20,
-                         UINT16_C(0x0010)))
+                         fromRep16(0x0010)))
         return 1;
     if (test__truncdfhf2(0x1.0p-24,
-                         UINT16_C(0x0001)))
+                         fromRep16(0x0001)))
         return 1;
     if (test__truncdfhf2(-0x1.0p-24,
-                         UINT16_C(0x8001)))
+                         fromRep16(0x8001)))
         return 1;
     if (test__truncdfhf2(0x1.5p-25,
-                         UINT16_C(0x0001)))
+                         fromRep16(0x0001)))
         return 1;
     // and back to zero
     if (test__truncdfhf2(0x1.0p-25,
-                         UINT16_C(0x0000)))
+                         fromRep16(0x0000)))
         return 1;
     if (test__truncdfhf2(-0x1.0p-25,
-                         UINT16_C(0x8000)))
+                         fromRep16(0x8000)))
         return 1;
     // max (precise)
     if (test__truncdfhf2(65504.0,
-                         UINT16_C(0x7bff)))
+                         fromRep16(0x7bff)))
         return 1;
     // max (rounded)
     if (test__truncdfhf2(65519.0,
-                         UINT16_C(0x7bff)))
+                         fromRep16(0x7bff)))
         return 1;
     // max (to +inf)
     if (test__truncdfhf2(65520.0,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncdfhf2(-65520.0,
-                         UINT16_C(0xfc00)))
+                         fromRep16(0xfc00)))
         return 1;
     if (test__truncdfhf2(65536.0,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     return 0;
 }

compiler-rt/test/builtins/Unit/truncsfhf2_test.c

 // RUN: %clang_builtins %s %librt -o %t && %run %t
 // REQUIRES: librt_has_truncsfhf2
 
 #include <stdio.h>
 
 #include "fp_test.h"
 
-uint16_t __truncsfhf2(float a);
+TYPE_FP16 __truncsfhf2(float a);
 
-int test__truncsfhf2(float a, uint16_t expected)
+int test__truncsfhf2(float a, TYPE_FP16 expected)
 {
-    uint16_t x = __truncsfhf2(a);
+    TYPE_FP16 x = __truncsfhf2(a);
     int ret = compareResultH(x, expected);
 
     if (ret){
         printf("error in test__truncsfhf2(%f) = %#.4x, "
-               "expected %#.4x\n", a, x, fromRep16(expected));
+               "expected %#.4x\n", a, fromRep16(x), fromRep16(expected));
     }
     return ret;
 }
 
 char assumption_1[sizeof(__fp16) * CHAR_BIT == 16] = {0};
 
 int main()
 {
     // qNaN
     if (test__truncsfhf2(makeQNaN32(),
-                         UINT16_C(0x7e00)))
+                         fromRep16(0x7e00)))
         return 1;
     // NaN
     if (test__truncsfhf2(makeNaN32(UINT32_C(0x8000)),
-                         UINT16_C(0x7e00)))
+                         fromRep16(0x7e00)))
         return 1;
     // inf
     if (test__truncsfhf2(makeInf32(),
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncsfhf2(-makeInf32(),
-                         UINT16_C(0xfc00)))
+                         fromRep16(0xfc00)))
         return 1;
     // zero
-    if (test__truncsfhf2(0.0f, UINT16_C(0x0)))
+    if (test__truncsfhf2(0.0f, fromRep16(0x0)))
         return 1;
-    if (test__truncsfhf2(-0.0f, UINT16_C(0x8000)))
+    if (test__truncsfhf2(-0.0f, fromRep16(0x8000)))
         return 1;
 
     if (test__truncsfhf2(3.1415926535f,
-                         UINT16_C(0x4248)))
+                         fromRep16(0x4248)))
         return 1;
     if (test__truncsfhf2(-3.1415926535f,
-                         UINT16_C(0xc248)))
+                         fromRep16(0xc248)))
         return 1;
     if (test__truncsfhf2(0x1.987124876876324p+100f,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncsfhf2(0x1.987124876876324p+12f,
-                         UINT16_C(0x6e62)))
+                         fromRep16(0x6e62)))
         return 1;
     if (test__truncsfhf2(0x1.0p+0f,
-                         UINT16_C(0x3c00)))
+                         fromRep16(0x3c00)))
         return 1;
     if (test__truncsfhf2(0x1.0p-14f,
-                         UINT16_C(0x0400)))
+                         fromRep16(0x0400)))
         return 1;
     // denormal
     if (test__truncsfhf2(0x1.0p-20f,
-                         UINT16_C(0x0010)))
+                         fromRep16(0x0010)))
         return 1;
     if (test__truncsfhf2(0x1.0p-24f,
-                         UINT16_C(0x0001)))
+                         fromRep16(0x0001)))
         return 1;
     if (test__truncsfhf2(-0x1.0p-24f,
-                         UINT16_C(0x8001)))
+                         fromRep16(0x8001)))
         return 1;
     if (test__truncsfhf2(0x1.5p-25f,
-                         UINT16_C(0x0001)))
+                         fromRep16(0x0001)))
         return 1;
     // and back to zero
     if (test__truncsfhf2(0x1.0p-25f,
-                         UINT16_C(0x0000)))
+                         fromRep16(0x0000)))
         return 1;
     if (test__truncsfhf2(-0x1.0p-25f,
-                         UINT16_C(0x8000)))
+                         fromRep16(0x8000)))
         return 1;
     // max (precise)
     if (test__truncsfhf2(65504.0f,
-                         UINT16_C(0x7bff)))
+                         fromRep16(0x7bff)))
         return 1;
     // max (rounded)
     if (test__truncsfhf2(65519.0f,
-                         UINT16_C(0x7bff)))
+                         fromRep16(0x7bff)))
         return 1;
     // max (to +inf)
     if (test__truncsfhf2(65520.0f,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncsfhf2(65536.0f,
-                         UINT16_C(0x7c00)))
+                         fromRep16(0x7c00)))
         return 1;
     if (test__truncsfhf2(-65520.0f,
-                         UINT16_C(0xfc00)))
+                         fromRep16(0xfc00)))
         return 1;
     return 0;
 }

compiler-rt/test/builtins/Unit/trunctfhf2_test.c

+// RUN: %clang_builtins %s %librt -o %t && %run %t
+// REQUIRES: librt_has_trunctfhf2
+
+#include "int_lib.h"
+#include <stdio.h>
+
+#if __LDBL_MANT_DIG__ == 113 && defined (COMPILER_RT_HAS_FLOAT16)
+
+#include "fp_test.h"
+
+TYPE_FP16 __trunctfhf2(long double a);
+
+int test__trunctfhf2(long double a, TYPE_FP16 expected)
+{
+    TYPE_FP16 x = __trunctfhf2(a);
+    int ret = compareResultH(x, expected);
+
+    if (ret)
+    {
+        printf("error in test__trunctfhf2(%.20Lf) = %#.4x, "
+               "expected %#.4x\n", a, fromRep16(x), fromRep16(expected));
+    }
+    return ret;
+}
+
+char assumption_1[sizeof(TYPE_FP16) * CHAR_BIT == 16] = {0};
+
+#endif
+
+int main()
+{
+#if __LDBL_MANT_DIG__ == 113 && defined (COMPILER_RT_HAS_FLOAT16)
+    // qNaN
+    if (test__trunctfhf2(makeQNaN128(),
+                         fromRep16(0x7e00)))
+        return 1;
+    // NaN
+    if (test__trunctfhf2(makeNaN128(UINT64_C(0x810000000000)),
+                         fromRep16(0x7e00)))
+        return 1;
+    // inf
+    if (test__trunctfhf2(makeInf128(),
+                         fromRep16(0x7c00)))
+        return 1;
+    if (test__trunctfhf2(-makeInf128(),
+                         fromRep16(0xfc00)))
+        return 1;
+    // zero
+    if (test__trunctfhf2(0.0L, fromRep16(0x0)))
+        return 1;
+    if (test__trunctfhf2(-0.0L, fromRep16(0x8000)))
+        return 1;
+
+    if (test__trunctfhf2(3.1415926535L,
+                         fromRep16(0x4248)))
+        return 1;
+    if (test__trunctfhf2(-3.1415926535L,
+                         fromRep16(0xc248)))
+        return 1;
+    if (test__trunctfhf2(0x1.987124876876324p+100L,
+                         fromRep16(0x7c00)))
+        return 1;
+    if (test__trunctfhf2(0x1.987124876876324p+12L,
+                         fromRep16(0x6e62)))
+        return 1;
+    if (test__trunctfhf2(0x1.0p+0L,
+                         fromRep16(0x3c00)))
+        return 1;
+    if (test__trunctfhf2(0x1.0p-14L,
+                         fromRep16(0x0400)))
+        return 1;
+    // denormal
+    if (test__trunctfhf2(0x1.0p-20L,
+                         fromRep16(0x0010)))
+        return 1;
+    if (test__trunctfhf2(0x1.0p-24L,
+                         fromRep16(0x0001)))
+        return 1;
+    if (test__trunctfhf2(-0x1.0p-24L,
+                         fromRep16(0x8001)))
+        return 1;
+    if (test__trunctfhf2(0x1.5p-25L,
+                         fromRep16(0x0001)))
+        return 1;
+    // and back to zero
+    if (test__trunctfhf2(0x1.0p-25L,
+                         fromRep16(0x0000)))
+        return 1;
+    if (test__trunctfhf2(-0x1.0p-25L,
+                         fromRep16(0x8000)))
+        return 1;
+    // max (precise)
+    if (test__trunctfhf2(65504.0L,
+                         fromRep16(0x7bff)))
+        return 1;
+    // max (rounded)
+    if (test__trunctfhf2(65519.0L,
+                         fromRep16(0x7bff)))
+        return 1;
+    // max (to +inf)
+    if (test__trunctfhf2(65520.0L,
+                         fromRep16(0x7c00)))
+        return 1;
+    if (test__trunctfhf2(65536.0L,
+                         fromRep16(0x7c00)))
+        return 1;
+    if (test__trunctfhf2(-65520.0L,
+                         fromRep16(0xfc00)))
+        return 1;
+
+    if (test__trunctfhf2(0x1.23a2abb4a2ddee355f36789abcdep+5L,
+                         fromRep16(0x508f)))
+        return 1;
+    if (test__trunctfhf2(0x1.e3d3c45bd3abfd98b76a54cc321fp-9L,
+                         fromRep16(0x1b8f)))
+        return 1;
+    if (test__trunctfhf2(0x1.234eebb5faa678f4488693abcdefp+453L,
+                         fromRep16(0x7c00)))
+        return 1;
+    if (test__trunctfhf2(0x1.edcba9bb8c76a5a43dd21f334634p-43L,
+                         fromRep16(0x0)))
+        return 1;
+#else
+    printf("skipped\n");
+#endif
+    return 0;
+}

llvm/include/llvm/IR/RuntimeLibcalls.def

 HANDLE_LIBCALL(LLRINT_PPCF128, "llrintl")
 
 // Conversion
 HANDLE_LIBCALL(FPEXT_F32_PPCF128, "__gcc_stoq")
 HANDLE_LIBCALL(FPEXT_F64_PPCF128, "__gcc_dtoq")
 HANDLE_LIBCALL(FPEXT_F80_F128, "__extendxftf2")
 HANDLE_LIBCALL(FPEXT_F64_F128, "__extenddftf2")
 HANDLE_LIBCALL(FPEXT_F32_F128, "__extendsftf2")
+HANDLE_LIBCALL(FPEXT_F16_F128, "__extendhftf2")
 HANDLE_LIBCALL(FPEXT_F32_F64, "__extendsfdf2")
 HANDLE_LIBCALL(FPEXT_F16_F32, "__gnu_h2f_ieee")
 HANDLE_LIBCALL(FPROUND_F32_F16, "__gnu_f2h_ieee")
 HANDLE_LIBCALL(FPROUND_F64_F16, "__truncdfhf2")
 HANDLE_LIBCALL(FPROUND_F80_F16, "__truncxfhf2")
 HANDLE_LIBCALL(FPROUND_F128_F16, "__trunctfhf2")
 HANDLE_LIBCALL(FPROUND_PPCF128_F16, "__trunctfhf2")
 HANDLE_LIBCALL(FPROUND_F64_F32, "__truncdfsf2")

llvm/lib/CodeGen/TargetLoweringBase.cpp

 }
 
 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
 /// UNKNOWN_LIBCALL if there is none.
 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
   if (OpVT == MVT::f16) {
     if (RetVT == MVT::f32)
       return FPEXT_F16_F32;
+    if (RetVT == MVT::f128)
+      return FPEXT_F16_F128;
   } else if (OpVT == MVT::f32) {
     if (RetVT == MVT::f64)
       return FPEXT_F32_F64;
     if (RetVT == MVT::f128)
       return FPEXT_F32_F128;
     if (RetVT == MVT::ppcf128)
       return FPEXT_F32_PPCF128;
   } else if (OpVT == MVT::f64) {

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom);
   setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i128, Custom);
   setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
   setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
   setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
   setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom);
   setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom);
   setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i128, Custom);
+  setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
   setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
   setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
+  setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Custom);
   setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom);
   setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Custom);
 
   // Variable arguments.
   setOperationAction(ISD::VASTART, MVT::Other, Custom);
   setOperationAction(ISD::VAARG, MVT::Other, Custom);
   setOperationAction(ISD::VACOPY, MVT::Other, Custom);
   setOperationAction(ISD::VAEND, MVT::Other, Expand);

llvm/test/CodeGen/AArch64/arm64-fp128.ll

 ; CHECK-NEXT: %bb.
 ; CHECK-NEXT: mov v[[VAL:[0-9]+]].16b, v0.16b
 ; CHECK-NEXT: [[IFFALSE]]:
 ; CHECK: str q[[VAL]], [{{x[0-9]+}}, :lo12:lhs]
   ret void
 ; CHECK: ret
 }
 
+@varhalf = global half 0.0, align 2
 @varfloat = global float 0.0, align 4
 @vardouble = global double 0.0, align 8
 
 define void @test_round() {
 ; CHECK-LABEL: test_round:
 
   %val = load fp128, fp128* @lhs, align 16
 
+  %half = fptrunc fp128 %val to half
+  store half %half, half* @varhalf, align 2
+; CHECK: ldr q0, [{{x[0-9]+}}, :lo12:lhs]
+; CHECK: bl __trunctfhf2
+; CHECK: str h0, [{{x[0-9]+}}, :lo12:varhalf]
+
   %float = fptrunc fp128 %val to float
   store float %float, float* @varfloat, align 4
 ; CHECK: bl __trunctfsf2
 ; CHECK: str s0, [{{x[0-9]+}}, :lo12:varfloat]
 
   %double = fptrunc fp128 %val to double
   store double %double, double* @vardouble, align 8
 ; CHECK: bl __trunctfdf2
 ; CHECK: str d0, [{{x[0-9]+}}, :lo12:vardouble]
 
   ret void
 }
 
 define void @test_extend() {
 ; CHECK-LABEL: test_extend:
 
   %val = load fp128, fp128* @lhs, align 16
 
+  %half = load half, half* @varhalf
+  %fromhalf = fpext half %half to fp128
+  store volatile fp128 %fromhalf, fp128* @lhs, align 16
+; CHECK: ldr h0, [{{x[0-9]+}}, :lo12:varhalf]
+; CHECK: bl __extendhftf2
+; CHECK: str q0, [{{x[0-9]+}}, :lo12:lhs]
+
   %float = load float, float* @varfloat
   %fromfloat = fpext float %float to fp128
   store volatile fp128 %fromfloat, fp128* @lhs, align 16
 ; CHECK: bl __extendsftf2
 ; CHECK: str q0, [{{x[0-9]+}}, :lo12:lhs]
 
   %double = load double, double* @vardouble
   %fromdouble = fpext double %double to fp128