AMDGPU: Add fmed3 half builtin

[kzhuravl]

LGTM.

[arsenm]

r295874

include/clang/Basic/BuiltinsAMDGPU.def

 TARGET_BUILTIN(__builtin_amdgcn_ldexph, "hhi", "nc", "16-bit-insts")
 TARGET_BUILTIN(__builtin_amdgcn_frexp_manth, "hh", "nc", "16-bit-insts")
 TARGET_BUILTIN(__builtin_amdgcn_frexp_exph, "sh", "nc", "16-bit-insts")
 TARGET_BUILTIN(__builtin_amdgcn_fracth, "hh", "nc", "16-bit-insts")
 TARGET_BUILTIN(__builtin_amdgcn_classh, "bhi", "nc", "16-bit-insts")
 TARGET_BUILTIN(__builtin_amdgcn_s_memrealtime, "LUi", "n", "s-memrealtime")
 
 //===----------------------------------------------------------------------===//
+// GFX9+ only builtins.
+//===----------------------------------------------------------------------===//
+
+TARGET_BUILTIN(__builtin_amdgcn_fmed3h, "hhhh", "nc", "gfx9-insts")
+
+//===----------------------------------------------------------------------===//
 // Special builtins.
 //===----------------------------------------------------------------------===//
 BUILTIN(__builtin_amdgcn_read_exec, "LUi", "nc")
 
 //===----------------------------------------------------------------------===//
 // R600-NI only builtins.
 //===----------------------------------------------------------------------===//
 

lib/Basic/Targets.cpp

if (getTriple().getArch() == llvm::Triple::amdgcn) {
     if (CPU.empty())
       CPU = "tahiti";
 
     switch (parseAMDGCNName(CPU)) {
     case GK_GFX6:
     case GK_GFX7:
       break;
 
+    case GK_GFX9:
+      Features["gfx9-insts"] = true;
+      LLVM_FALLTHROUGH;
     case GK_GFX8:
       Features["s-memrealtime"] = true;
       Features["16-bit-insts"] = true;
       break;
 
     case GK_NONE:
       return false;
     default:

lib/CodeGen/CGBuiltin.cpp

Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
   case AMDGPU::BI__builtin_amdgcn_fcmp:
   case AMDGPU::BI__builtin_amdgcn_fcmpf:
     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fcmp);
   case AMDGPU::BI__builtin_amdgcn_class:
   case AMDGPU::BI__builtin_amdgcn_classf:
   case AMDGPU::BI__builtin_amdgcn_classh:
     return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class);
   case AMDGPU::BI__builtin_amdgcn_fmed3f:
+  case AMDGPU::BI__builtin_amdgcn_fmed3h:
     return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3);
   case AMDGPU::BI__builtin_amdgcn_read_exec: {
     CallInst *CI = cast<CallInst>(
       EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, true, "exec"));
     CI->setConvergent();
     return CI;
   }
 

test/CodeGenOpenCL/builtins-amdgcn-gfx9.cl

+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx900 -S -emit-llvm -o - %s | FileCheck %s
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// CHECK-LABEL: @test_fmed3_f16
+// CHECK: call half @llvm.amdgcn.fmed3.f16(half %a, half %b, half %c)
+void test_fmed3_f16(global half* out, half a, half b, half c)
+{
+  *out = __builtin_amdgcn_fmed3h(a, b, c);
+}

test/SemaOpenCL/builtins-amdgcn-error-f16.cl

 // REQUIRES: amdgpu-registered-target
-// RUN: %clang_cc1 -triple amdgcn-- -target-cpu tahiti -verify -S -o - %s
+// RUN: %clang_cc1 -triple amdgcn-- -verify -S -o - %s
 
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
-void test_f16(global half *out, half a, half b, half c)
+__attribute__((target("arch=tahiti")))
+void test_f16_tahiti(global half *out, half a, half b, half c)
 {
   *out = __builtin_amdgcn_div_fixuph(a, b, c); // expected-error {{'__builtin_amdgcn_div_fixuph' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_rcph(a); // expected-error {{'__builtin_amdgcn_rcph' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_rsqh(a); // expected-error {{'__builtin_amdgcn_rsqh' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_sinh(a); // expected-error {{'__builtin_amdgcn_sinh' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_cosh(a); // expected-error {{'__builtin_amdgcn_cosh' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_ldexph(a, b); // expected-error {{'__builtin_amdgcn_ldexph' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_frexp_manth(a); // expected-error {{'__builtin_amdgcn_frexp_manth' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_frexp_exph(a); // expected-error {{'__builtin_amdgcn_frexp_exph' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_fracth(a); // expected-error {{'__builtin_amdgcn_fracth' needs target feature 16-bit-insts}}
   *out = __builtin_amdgcn_classh(a, b); // expected-error {{'__builtin_amdgcn_classh' needs target feature 16-bit-insts}}
+  *out = __builtin_amdgcn_fmed3h(a, b, c); // expected-error {{'__builtin_amdgcn_fmed3h' needs target feature gfx9-insts}}
 }

test/SemaOpenCL/builtins-amdgcn-error-gfx9.cl

+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-- -target-cpu fiji -verify -S -o - %s
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+void test_gfx9_fmed3h(global half *out, half a, half b, half c)
+{
+  *out = __builtin_amdgcn_fmed3h(a, b, c); // expected-error {{'__builtin_amdgcn_fmed3h' needs target feature gfx9-insts}}
+}