[clang] [clang][AMDGPU] Enable module splitting by default (PR #128509)

[Shilei Tian via cfe-commits]

================
@@ -708,6 +712,34 @@ void amdgpu::getAMDGPUTargetFeatures(const Driver &D,
                             options::OPT_m_amdgpu_Features_Group);
 }
 
+static unsigned GetFullLTOPartitions(const Driver &D, const ArgList &Args) {
+  const Arg *A = Args.getLastArg(options::OPT_flto_partitions_EQ);
+  // In the absence of an option, use the number of available threads with a cap
+  // at 16 partitions. More than 16 partitions rarely benefits code splitting
+  // and can lead to more empty/small modules each with their own overhead.
+  if (!A)
+    return std::min(16u, llvm::hardware_concurrency().compute_thread_count());
+  int Value;
+  if (StringRef(A->getValue()).getAsInteger(10, Value) || (Value < 1)) {
+    D.Diag(diag::err_drv_invalid_int_value)
+        << A->getAsString(Args) << A->getValue();
+    return 1;
+  }
+
+  return Value;
+}
+
+void amdgpu::addFullLTOPartitionOption(const Driver &D,
+                                       const llvm::opt::ArgList &Args,
+                                       llvm::opt::ArgStringList &CmdArgs) {
+  // TODO: restrict to gpu-rdc only?
----------------
shiltian wrote:

Do we support no-gpu-rdc and full LTO? I thought gpu-rdc turns on LTO.

https://github.com/llvm/llvm-project/pull/128509