[mlgo] Simplify autogenerated regalloc model

[aidengrossman]

Summary

Currently the autogenerated regalloc model will sometimes
output an incorrect LR index to evict instead of the first LR
with with the mask set to 1. This trips an assertion within
the MLRegallocAdvisor that the evicted LR has a mask of 1. This
patch, made possible by https://reviews.llvm.org/D124565, simplifies
the autogenerated model by taking away all unnecessary features and
getting rid of the functions that were previously to mix in all
the necessary inputs so they wouldn't get pruned by the Tensorflow
XLA AOT compiler. This is no longer necessary after the previously
mentioned patch. This also fixes the nondeterministic behavior
that is sometimes observed where the autogenerated model will
simply output 0 instead of the correct index.

[aidengrossman]

This patch removes all unnecessary features from the autogenerated regalloc model (simplifying it greatly) and gets rid of the nondeterminism that I was experiencing that trips the mask==1 assertion in the MLRegallocAdvisor. Not anything super critical, but slightly annoying during development as it has to be commented out and commented back in after feature extraction development is completed and doesn't do what it it is supposed to during development. Tested locally as working and all the tests in /llvm/test/CodeGen/MLRegalloc pass for me as well.

[yundiqian]

mtrofin@ to take a look to confirm as well.

llvm/lib/Analysis/models/gen-regalloc-eviction-test-model.py

[
             "type": "int64_t",
             "shape": [
                 1
             ]
         }
     }
 ]
 """
-PER_REGISTER_INT64_FEATURE_LIST = [
-    'mask', 'is_hint', 'is_local', 'is_free', 'max_stage', 'min_stage'
-]
-PER_REGISTER_FLOAT32_FEATURE_LIST = ['nr_urgent',
-    'weighed_reads_by_max', 'weighed_writes_by_max',
-    'weighed_read_writes_by_max', 'weighed_indvars_by_max',
-    'hint_weights_by_max', 'start_bb_freq_by_max', 'end_bb_freq_by_max',
-    'hottest_bb_freq_by_max', 'liverange_size', 'use_def_density',
-    'nr_defs_and_uses', 'nr_broken_hints', 'nr_rematerializable'
-]
-PER_REGISTER_FEATURE_LIST = PER_REGISTER_FLOAT32_FEATURE_LIST + \
-    PER_REGISTER_INT64_FEATURE_LIST
-CONTEXT_FEATURE_LIST = ('progress', 'discount', 'reward', 'step_type')
+PER_REGISTER_FEATURE_LIST = ['mask']
 NUM_REGISTERS = 33
 
 
 def get_input_signature():
   """Returns (time_step_spec, action_spec) for LLVM register allocation."""
   inputs = dict(
       (key, tf.TensorSpec(dtype=tf.int64, shape=(NUM_REGISTERS), name=key))
-      for key in PER_REGISTER_INT64_FEATURE_LIST)
-  inputs.update(
-      dict((key,
-            tf.TensorSpec(dtype=tf.float32, shape=(NUM_REGISTERS), name=key))
-           for key in PER_REGISTER_FLOAT32_FEATURE_LIST))
-  inputs['progress'] = tf.TensorSpec(
-      dtype=tf.float32, shape=(), name='progress')
-  inputs.update(
-      dict((key, tf.TensorSpec(dtype=tf.float32, shape=(), name=key))
-           for key in ['discount', 'reward']))
-  inputs.update(
-      dict((key, tf.TensorSpec(dtype=tf.int32, shape=(), name=key))
-           for key in ['step_type']))
+      for key in PER_REGISTER_FEATURE_LIST)
   return inputs
 
 
 def get_output_spec_path(path):
   return os.path.join(path, 'output_spec.json')
 
 
 def build_mock_model(path):
   """Build and save the mock model with the given signature."""
   module = tf.Module()
   # We have to set this useless variable in order for the TF C API to correctly
   # intake it
   module.var = tf.Variable(0, dtype=tf.int64)
 
   def action(*inputs):
-    s1 = tf.reduce_sum([
-        tf.cast(inputs[0][key], tf.float32) for key in PER_REGISTER_FEATURE_LIST
-    ],
-        axis=0)
-    s2 = tf.reduce_sum(
-        [tf.cast(inputs[0][key], tf.float32) for key in CONTEXT_FEATURE_LIST])
-    # Add a large number so s won't be 0.
-    s = s1 + s2 + 123456789.123456789
-    # Equals to mask feature.
-    mask_alias = tf.not_equal(s * tf.cast(inputs[0]['mask'], tf.float32), 0)
-    result = tf.math.argmax(mask_alias, axis=-1) + module.var
+    result = tf.math.argmax(inputs[0]['mask'], axis=-1) + module.var
     return {POLICY_DECISION_LABEL: result}
   module.action = tf.function()(action)
   action = {
       'action': module.action.get_concrete_function(get_input_signature())
   }
   tf.saved_model.save(module, path, signatures=action)
   output_spec_path = get_output_spec_path(path)
   with open(output_spec_path, 'w') as f: