Add flag to enable native half type

[pirama]

Summary

r235215 enables support in LLVM for legalizing f16 type in the IR. AArch64
already had support for this. r235215 and some backend patches brought support
for ARM, X86, X86-64, Mips and Mips64.

This change exposes the LangOption 'NativeHalfType' in the command line, so the
backend legalization can be used if desired. NativeHalfType is enabled for
OpenCL (current behavior) or if '-fnative-half-type' is set.

[ab]

LGTM, see inline nits. I'll trust you know what you're doing in making this a cc1 option.

-Ahmed

include/clang/Driver/CC1Options.td

550–553

I don't like "half": it should really be about "__fp16". But that's also a problem for -fallow-half-*, so I'm fine with it.

We can have a more explicit description though. How about "Use the native half type for __fp16 instead of promoting to float" ?

lib/Frontend/CompilerInvocation.cpp

1588–1589

How about:

Ops.NativeHalfType |= Arg.hasArg(Opt_fnative_half_type);

1592

Spurious whitespace?

test/CodeGen/fp16-ops.c

5–9

You can split these lines in a separate commit, helps readability.

172–173

Now that it makes a difference, can you add "float" in the ones where it's missing?
You can do that separately as well.

224–225

One thing you could do is separate all the clang-promoted comparisons like this and group them together. You can then share CHECK- and F16TOF32 for these with HALF, if you add a NATIVE-HALF: [[F16TOF32:...]] and yet another, *really* shared CHECK prefix. Not sure it helps though.

[pirama]

Changes based on review comments

[pirama]

A CC1 option is enough for our purposes, and matches the other fallow-half-* argument.

test/CodeGen/fp16-ops.c

224–225

I'll leave this check as-is. Having three different prefixes that are all interdependent is too complicated :)

include/clang/Driver/CC1Options.td

 def fno_deprecated_macro : Flag<["-"], "fno-deprecated-macro">,
   HelpText<"Undefines the __DEPRECATED macro">;
 def fobjc_subscripting_legacy_runtime : Flag<["-"], "fobjc-subscripting-legacy-runtime">,
   HelpText<"Allow Objective-C array and dictionary subscripting in legacy runtime">;
 def vtordisp_mode_EQ : Joined<["-"], "vtordisp-mode=">,
   HelpText<"Control vtordisp placement on win32 targets">;
 def fno_rtti_data : Flag<["-"], "fno-rtti-data">,
   HelpText<"Control emission of RTTI data">;
+def fnative_half_type: Flag<["-"], "fnative-half-type">,
+  HelpText<"Use the native half type for __fp16 instead of promoting to float">;
 def fallow_half_arguments_and_returns : Flag<["-"], "fallow-half-arguments-and-returns">,
   HelpText<"Allow function arguments and returns of type half">;
 
 //===----------------------------------------------------------------------===//
 // Header Search Options
 //===----------------------------------------------------------------------===//
 
 def nostdsysteminc : Flag<["-"], "nostdsysteminc">,
   HelpText<"Disable standard system #include directories">;
 def fdisable_module_hash : Flag<["-"], "fdisable-module-hash">,

lib/Frontend/CompilerInvocation.cpp

#include "clang/Frontend/LangStandards.def"
   Opts.DebuggerCastResultToId = Args.hasArg(OPT_fdebugger_cast_result_to_id);
   Opts.DebuggerObjCLiteral = Args.hasArg(OPT_fdebugger_objc_literal);
   Opts.ApplePragmaPack = Args.hasArg(OPT_fapple_pragma_pack);
   Opts.CurrentModule = Args.getLastArgValue(OPT_fmodule_name);
   Opts.AppExt = Args.hasArg(OPT_fapplication_extension);
   Opts.ImplementationOfModule =
       Args.getLastArgValue(OPT_fmodule_implementation_of);
   Opts.ModuleFeatures = Args.getAllArgValues(OPT_fmodule_feature);
-  Opts.NativeHalfType = Opts.NativeHalfType;
+  Opts.NativeHalfType |= Args.hasArg(OPT_fnative_half_type);
   Opts.HalfArgsAndReturns = Args.hasArg(OPT_fallow_half_arguments_and_returns);
   Opts.GNUAsm = !Args.hasArg(OPT_fno_gnu_inline_asm);
 
   if (!Opts.CurrentModule.empty() && !Opts.ImplementationOfModule.empty() &&
       Opts.CurrentModule != Opts.ImplementationOfModule) {
     Diags.Report(diag::err_conflicting_module_names)
         << Opts.CurrentModule << Opts.ImplementationOfModule;
   }
 
   if (Arg *A = Args.getLastArg(OPT_faddress_space_map_mangling_EQ)) {
     switch (llvm::StringSwitch<unsigned>(A->getValue())
       .Case("target", LangOptions::ASMM_Target)

test/CodeGen/fp16-ops.c

 // REQUIRES: arm-registered-target
 // RUN: %clang_cc1 -emit-llvm -o - -triple arm-none-linux-gnueabi %s | FileCheck %s --check-prefix=NOHALF --check-prefix=CHECK
 // RUN: %clang_cc1 -emit-llvm -o - -triple aarch64-none-linux-gnueabi %s | FileCheck %s --check-prefix=NOHALF --check-prefix=CHECK
 // RUN: %clang_cc1 -emit-llvm -o - -triple arm-none-linux-gnueabi -fallow-half-arguments-and-returns %s | FileCheck %s --check-prefix=HALF --check-prefix=CHECK
 // RUN: %clang_cc1 -emit-llvm -o - -triple aarch64-none-linux-gnueabi -fallow-half-arguments-and-returns %s | FileCheck %s --check-prefix=HALF --check-prefix=CHECK
+// RUN: %clang_cc1 -emit-llvm -o - -triple arm-none-linux-gnueabi -fnative-half-type %s \
+// RUN:   | FileCheck %s --check-prefix=NATIVE-HALF
+// RUN: %clang_cc1 -emit-llvm -o - -triple aarch64-none-linux-gnueabi -fnative-half-type %s \
+// RUN:   | FileCheck %s --check-prefix=NATIVE-HALF
 typedef unsigned cond_t;
 
 volatile cond_t test;
 volatile __fp16 h0 = 0.0, h1 = 1.0, h2;
 volatile float f0, f1, f2;
 volatile double d0;
 
 void foo(void) {
   // CHECK-LABEL: define void @foo()
 
   // Check unary ops
 
   // NOHALF: [[F16TOF32:call float @llvm.convert.from.fp16.f32]]
   // HALF: [[F16TOF32:fpext half]]
   // CHECK: fptoui float
+  // NATIVE-HALF: fptoui half
   test = (h0);
   // CHECK: uitofp i32
   // NOHALF: [[F32TOF16:call i16 @llvm.convert.to.fp16.f32]]
   // HALF: [[F32TOF16:fptrunc float]]
+  // NATIVE-HALF: uitofp i32 {{.*}} to half
   h0 = (test);
   // CHECK: [[F16TOF32]]
   // CHECK: fcmp une float
+  // NATIVE-HALF: fcmp une half
   test = (!h1);
   // CHECK: [[F16TOF32]]
   // CHECK: fsub float
   // NOHALF: [[F32TOF16]]
   // HALF: [[F32TOF16]]
+  // NATIVE-HALF: fsub half
   h1 = -h1;
   // CHECK: [[F16TOF32]]
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: load volatile half
+  // NATIVE-HALF-NEXT: store volatile half
   h1 = +h1;
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h1++;
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   ++h1;
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   --h1;
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h1--;
 
   // Check binary ops with various operands
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fmul half
   h1 = h0 * h2;
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F32TOF16]]
   // NOHALF: [[F16TOF32]]
   // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fmul half
   h1 = h0 * (__fp16) -2.0f;
   // CHECK: [[F16TOF32]]
   // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fmul float
   h1 = h0 * f2;
   // CHECK: [[F16TOF32]]
   // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fmul float
   h1 = f0 * h2;
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fdiv half
   h1 = (h0 / h2);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
   // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fdiv half
   h1 = (h0 / (__fp16) -2.0f);
   // CHECK: [[F16TOF32]]
   // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fdiv float
   h1 = (h0 / f2);
   // CHECK: [[F16TOF32]]
   // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fdiv float
   h1 = (f0 / h2);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h1 = (h2 + h0);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h1 = ((__fp16)-2.0 + h0);
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fadd float
   h1 = (h2 + f0);
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fadd float
   h1 = (f2 + h0);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fsub half
   h1 = (h2 - h0);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
   // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fsub half
   h1 = ((__fp16)-2.0f - h0);
   // CHECK: [[F16TOF32]]
   // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fsub float
   h1 = (h2 - f0);
   // CHECK: [[F16TOF32]]
   // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fsub float
   h1 = (f2 - h0);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp olt
+  // CHECK: fcmp olt float
+  // NATIVE-HALF: fcmp olt half
   test = (h2 < h0);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp olt
+  // CHECK: fcmp olt float
+  // NATIVE-HALF: fcmp olt half
   test = (h2 < (__fp16)42.0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp olt
+  // CHECK: fcmp olt float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp olt float
   test = (h2 < f0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp olt
+  // CHECK: fcmp olt float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp olt float
   test = (f2 < h0);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ogt
+  // CHECK: fcmp ogt float
+  // NATIVE-HALF: fcmp ogt half
   test = (h0 > h2);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp ogt
+  // CHECK: fcmp ogt float
+  // NATIVE-HALF: fcmp ogt half
   test = ((__fp16)42.0 > h2);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ogt
+  // CHECK: fcmp ogt float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp ogt float
   test = (h0 > f2);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ogt
+  // CHECK: fcmp ogt float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp ogt float
   test = (f0 > h2);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ole
+  // CHECK: fcmp ole float
+  // NATIVE-HALF: fcmp ole half
   test = (h2 <= h0);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp ole
+  // CHECK: fcmp ole float
+  // NATIVE-HALF: fcmp ole half
   test = (h2 <= (__fp16)42.0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ole
+  // CHECK: fcmp ole float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp ole float
   test = (h2 <= f0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp ole
+  // CHECK: fcmp ole float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp ole float
   test = (f2 <= h0);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oge
+  // CHECK: fcmp oge float
+  // NATIVE-HALF: fcmp oge half
   test = (h0 >= h2);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp oge
+  // CHECK: fcmp oge float
+  // NATIVE-HALF: fcmp oge half
   test = (h0 >= (__fp16)-2.0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oge
+  // CHECK: fcmp oge float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp oge float
   test = (h0 >= f2);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oge
+  // CHECK: fcmp oge float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp oge float
   test = (f0 >= h2);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oeq
+  // CHECK: fcmp oeq float
+  // NATIVE-HALF: fcmp oeq half
   test = (h1 == h2);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp oeq
+  // CHECK: fcmp oeq float
+  // NATIVE-HALF: fcmp oeq half
   test = (h1 == (__fp16)1.0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oeq
+  // CHECK: fcmp oeq float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp oeq float
   test = (h1 == f1);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp oeq
+  // CHECK: fcmp oeq float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp oeq float
   test = (f1 == h1);
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp une
+  // CHECK: fcmp une float
+  // NATIVE-HALF: fcmp une half
   test = (h1 != h2);
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fcmp une
+  // CHECK: fcmp une float
+  // NATIVE-HALF: fcmp une half
   test = (h1 != (__fp16)1.0);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp une
+  // CHECK: fcmp une float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp une float
   test = (h1 != f1);
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp une
+  // CHECK: fcmp une float
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fcmp une float
   test = (f1 != h1);
 
   // CHECK: [[F16TOF32]]
-  // CHECK: fcmp une
+  // CHECK: fcmp une float
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fcmp une half {{.*}}, 0xH0000
   h1 = (h1 ? h2 : h0);
   // Check assignments (inc. compound)
   h0 = h1;
   // NOHALF: [[F32TOF16]]
   // HALF: store {{.*}} half 0xHC000
+  // NATIVE-HALF: store {{.*}} half 0xHC000
   h0 = (__fp16)-2.0f;
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fptrunc float
   h0 = f0;
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
   // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h0 += h1;
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fadd
+  // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fadd half
   h0 += (__fp16)1.0f;
   // CHECK: [[F16TOF32]]
-  // CHECK: fadd
+  // CHECK: fadd float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fadd float
+  // NATIVE-HALF: fptrunc float
   h0 += f2;
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fsub
+  // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fsub half
   h0 -= h1;
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fsub
+  // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fsub half
   h0 -= (__fp16)1.0;
   // CHECK: [[F16TOF32]]
-  // CHECK: fsub
+  // CHECK: fsub float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fsub float
+  // NATIVE-HALF: fptrunc float
   h0 -= f2;
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fmul
+  // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fmul half
   h0 *= h1;
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fmul
+  // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fmul half
   h0 *= (__fp16)1.0;
   // CHECK: [[F16TOF32]]
-  // CHECK: fmul
+  // CHECK: fmul float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fmul float
+  // NATIVE-HALF: fptrunc float
   h0 *= f2;
 
   // CHECK: [[F16TOF32]]
   // CHECK: [[F16TOF32]]
-  // CHECK: fdiv
+  // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fdiv half
   h0 /= h1;
   // CHECK: [[F16TOF32]]
   // NOHALF: [[F16TOF32]]
-  // CHECK: fdiv
+  // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fdiv half
   h0 /= (__fp16)1.0;
   // CHECK: [[F16TOF32]]
-  // CHECK: fdiv
+  // CHECK: fdiv float
   // CHECK: [[F32TOF16]]
+  // NATIVE-HALF: fpext half
+  // NATIVE-HALF: fdiv float
+  // NATIVE-HALF: fptrunc float
   h0 /= f2;
 
   // Check conversions to/from double
   // NOHALF: call i16 @llvm.convert.to.fp16.f64(
   // HALF: fptrunc double {{.*}} to half
+  // NATIVE-HALF: fptrunc double {{.*}} to half
   h0 = d0;
 
   // CHECK: [[MID:%.*]] = fptrunc double {{%.*}} to float
   // NOHALF: call i16 @llvm.convert.to.fp16.f32(float [[MID]])
   // HALF: fptrunc float [[MID]] to half
+  // NATIVE-HALF: [[MID:%.*]] = fptrunc double {{%.*}} to float
+  // NATIVE-HALF: fptrunc float {{.*}} to half
   h0 = (float)d0;
 
   // NOHALF: call double @llvm.convert.from.fp16.f64(
   // HALF: fpext half {{.*}} to double
+  // NATIVE-HALF: fpext half {{.*}} to double
   d0 = h0;
 
   // NOHALF: [[MID:%.*]] = call float @llvm.convert.from.fp16.f32(
   // HALF: [[MID:%.*]] = fpext half {{.*}} to float
   // CHECK: fpext float [[MID]] to double
+  // NATIVE-HALF: [[MID:%.*]] = fpext half {{.*}} to float
+  // NATIVE-HALF: fpext float [[MID]] to double
   d0 = (float)h0;
 }