[Clang][OpenMP] Allow `f16` literal suffix when compiling OpenMP target offloading for NVPTX

[tianshilei1992]

Summary

Fix #58087.

[jhuber6]

Seems reasonable, maybe update the comment.

[tianshilei1992]

fix comments

clang/lib/Lex/LiteralSupport.cpp

for (; s != ThisTokEnd; ++s) {
     case 'F':
       if (!isFPConstant) break;  // Error for integer constant.
       if (HasSize)
         break;
       HasSize = true;
 
       // CUDA host and device may have different _Float16 support, therefore
       // allows f16 literals to avoid false alarm.
+      // When we compile for OpenMP target offloading on NVPTX, f16 suffix
+      // should also be supported.
       // ToDo: more precise check for CUDA.
-      if ((Target.hasFloat16Type() || LangOpts.CUDA) && s + 2 < ThisTokEnd &&
-          s[1] == '1' && s[2] == '6') {
+      // TODO: AMDGPU might also support it in the future.
+      if ((Target.hasFloat16Type() || LangOpts.CUDA ||
+           (LangOpts.OpenMPIsDevice && Target.getTriple().isNVPTX())) &&
+          s + 2 < ThisTokEnd && s[1] == '1' && s[2] == '6') {
         s += 2; // success, eat up 2 characters.
         isFloat16 = true;
         continue;
       }
 
       isFloat = true;
       continue;  // Success.
     case 'q':    // FP Suffix for "__float128"

clang/test/OpenMP/float16_sema.cpp

+// RUN: %clang_cc1 -fsyntax-only -x c++ -triple x86_64-unknown-linux-gnu -fopenmp -fopenmp-targets=nvptx64 -verify %s
+// expected-no-diagnostics
+
+int foo() {
+#pragma omp target
+  {
+    __fp16 a = -1.0f16;
+  }
+  return 0;
+}