[Clang][FP16] Add 4 builtins for _Float16

[pengfei]

Summary

We are lacking builtins support for _Float16. In most cases, we can use other floating-type builtins and truncate them to _Float16.
But it's a problem to SNaN, e.g., https://gcc.godbolt.org/z/cqr5nG1jh
This patch adds __builtin_nansf16 support as well as other 3 ones since they are usually used together.

[LuoYuanke]

clang/include/clang/Basic/Builtins.def

145

Is the builtin sorted in alphabet order?

[pengfei]

clang/include/clang/Basic/Builtins.def

145

No, I think there's no strict order requirment here. The xxxf128 builtins have already not met it.

[LuoYuanke]

clang/test/CodeGen/builtin_Float16.c

8

Is _Float16 a legal type for target armv7a and aarch64?

[pengfei]

clang/test/CodeGen/builtin_Float16.c

8

Yes, see https://clang.llvm.org/docs/LanguageExtensions.html#half-precision-floating-point

[LuoYuanke]

clang/test/CodeGen/builtin_Float16.c

8

Maybe use __fp16 because it is supported on every target.

[pengfei]

clang/test/CodeGen/builtin_Float16.c

8

No, I intentionally use _Float16 here because I want to make it correspond to the type define x in Builtins.def
I think we should always use x (represents _Float16) rather than h (represents __fp16) when defining FP16 builtins. It doesn't make sense to to add new buildins for storage type, because all arithmetic operations should be promoted to float.

clang/include/clang/Basic/Builtins.def

 BUILTIN(__builtin_fmodf128, "LLdLLdLLd", "Fne")
 BUILTIN(__builtin_frexp , "ddi*"  , "Fn")
 BUILTIN(__builtin_frexpf, "ffi*"  , "Fn")
 BUILTIN(__builtin_frexpl, "LdLdi*", "Fn")
 BUILTIN(__builtin_frexpf128, "LLdLLdi*", "Fn")
 BUILTIN(__builtin_huge_val, "d", "nc")
 BUILTIN(__builtin_huge_valf, "f", "nc")
 BUILTIN(__builtin_huge_vall, "Ld", "nc")
+BUILTIN(__builtin_huge_valf16, "x", "nc")
 BUILTIN(__builtin_huge_valf128, "LLd", "nc")
 BUILTIN(__builtin_inf  , "d"   , "nc")
 BUILTIN(__builtin_inff , "f"   , "nc")
 BUILTIN(__builtin_infl , "Ld"  , "nc")
+BUILTIN(__builtin_inff16 , "x"  , "nc")
 BUILTIN(__builtin_inff128 , "LLd"  , "nc")
 BUILTIN(__builtin_labs , "LiLi"  , "Fnc")
 BUILTIN(__builtin_llabs, "LLiLLi", "Fnc")
 BUILTIN(__builtin_ldexp , "ddi"  , "Fne")
 BUILTIN(__builtin_ldexpf, "ffi"  , "Fne")
 BUILTIN(__builtin_ldexpl, "LdLdi", "Fne")
 BUILTIN(__builtin_ldexpf128, "LLdLLdi", "Fne")
 BUILTIN(__builtin_modf , "ddd*"  , "Fn")
 BUILTIN(__builtin_modff, "fff*"  , "Fn")
 BUILTIN(__builtin_modfl, "LdLdLd*", "Fn")
 BUILTIN(__builtin_modff128, "LLdLLdLLd*", "Fn")
 BUILTIN(__builtin_nan,  "dcC*" , "FnU")
 BUILTIN(__builtin_nanf, "fcC*" , "FnU")
 BUILTIN(__builtin_nanl, "LdcC*", "FnU")
+BUILTIN(__builtin_nanf16, "xcC*", "FnU")
 BUILTIN(__builtin_nanf128, "LLdcC*", "FnU")
 BUILTIN(__builtin_nans,  "dcC*" , "FnU")
 BUILTIN(__builtin_nansf, "fcC*" , "FnU")
 BUILTIN(__builtin_nansl, "LdcC*", "FnU")
+BUILTIN(__builtin_nansf16, "xcC*", "FnU")
 BUILTIN(__builtin_nansf128, "LLdcC*", "FnU")
 BUILTIN(__builtin_powi , "ddi"  , "Fnc")
 BUILTIN(__builtin_powif, "ffi"  , "Fnc")
 BUILTIN(__builtin_powil, "LdLdi", "Fnc")
 BUILTIN(__builtin_pow , "ddd"  , "Fne")
 BUILTIN(__builtin_powf, "fff"  , "Fne")
 BUILTIN(__builtin_powf16, "hhh"  , "Fne")
 BUILTIN(__builtin_powl, "LdLdLd", "Fne")

clang/lib/AST/ExprConstant.cpp

 bool FloatExprEvaluator::VisitCallExpr(const CallExpr *E) {
   switch (E->getBuiltinCallee()) {
   default:
     return ExprEvaluatorBaseTy::VisitCallExpr(E);
 
   case Builtin::BI__builtin_huge_val:
   case Builtin::BI__builtin_huge_valf:
   case Builtin::BI__builtin_huge_vall:
+  case Builtin::BI__builtin_huge_valf16:
   case Builtin::BI__builtin_huge_valf128:
   case Builtin::BI__builtin_inf:
   case Builtin::BI__builtin_inff:
   case Builtin::BI__builtin_infl:
+  case Builtin::BI__builtin_inff16:
   case Builtin::BI__builtin_inff128: {
     const llvm::fltSemantics &Sem =
       Info.Ctx.getFloatTypeSemantics(E->getType());
     Result = llvm::APFloat::getInf(Sem);
     return true;
   }
 
   case Builtin::BI__builtin_nans:
   case Builtin::BI__builtin_nansf:
   case Builtin::BI__builtin_nansl:
+  case Builtin::BI__builtin_nansf16:
   case Builtin::BI__builtin_nansf128:
     if (!TryEvaluateBuiltinNaN(Info.Ctx, E->getType(), E->getArg(0),
                                true, Result))
       return Error(E);
     return true;
 
   case Builtin::BI__builtin_nan:
   case Builtin::BI__builtin_nanf:
   case Builtin::BI__builtin_nanl:
+  case Builtin::BI__builtin_nanf16:
   case Builtin::BI__builtin_nanf128:
     // If this is __builtin_nan() turn this into a nan, otherwise we
     // can't constant fold it.
     if (!TryEvaluateBuiltinNaN(Info.Ctx, E->getType(), E->getArg(0),
                                false, Result))
       return Error(E);
     return true;
 

clang/test/CodeGen/builtin_Float16.c

+// RUN: %clang_cc1 -emit-llvm -o - -triple x86_64-linux-pc -target-feature +avx512fp16 %s | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -o - -triple spir-unknown-unknown %s | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -o - -triple armv7a--none-eabi %s | FileCheck %s
+// RUN: %clang_cc1 -emit-llvm -o - -triple aarch64-linux-gnu %s | FileCheck %s
+
+void test_float16_builtins(void) {
+  volatile _Float16 res;
+
+  // CHECK: store volatile half 0xH7C00, ptr %res, align 2
+  res = __builtin_huge_valf16();
+  // CHECK: store volatile half 0xH7C00, ptr %res, align 2
+  res = __builtin_inff16();
+  // CHECK: store volatile half 0xH7E00, ptr %res, align 2
+  res = __builtin_nanf16("");
+  // CHECK: store volatile half 0xH7D00, ptr %res, align 2
+  res = __builtin_nansf16("");
+}