[MLGO] Force persistency in tflite buffers.

When training large models, we encounter use-after-free bugs when
writing to the input tensors for various MLGO models. This patch fixes the
issue by marking the tensors as "persistent".

Differential Revision: https://reviews.llvm.org/D135739

llvm/lib/Analysis/TFLiteUtils.cpp

@@ -121,6 +121,15 @@ TFModelEvaluatorImpl::TFModelEvaluatorImpl(
   tflite::InterpreterBuilder Builder(*Model, Resolver);
   Builder(&Interpreter);
 
+  // We assume the input buffers are valid for the lifetime of the interpreter.
+  // By default, tflite allocates memory in an arena and will periodically take
+  // away memory and reallocate it in a different location after evaluations in
+  // order to improve utilization of the buffers owned in the arena. So, we
+  // explicitly mark our input buffers as persistent to avoid this behavior.
+  for (size_t I = 0; I < Interpreter->inputs().size(); ++I)
+    Interpreter->tensor(I)->allocation_type =
+        TfLiteAllocationType::kTfLiteArenaRwPersistent;
+
   if (!Interpreter ||
       Interpreter->AllocateTensors() != TfLiteStatus::kTfLiteOk) {
     invalidate();