Split optimisation passes.

numba/core/codegen.py

@@ -520,8 +520,13 @@ def _optimize_final_module(self):
         """
         Internal: optimize this library's final module.
         """
-        self._codegen._mpm.run(self._final_module)
+        # A cheaper optimisation pass is run first to try and get as many
+        # refops into the same function as possible via inlining
+        self._codegen._mpm_cheap.run(self._final_module)
+        # Refop pruning is then run on the heavily inlined function
         self._final_module = remove_redundant_nrt_refct(self._final_module)
+        # The full optimisation suite is then run on the refop pruned IR
+        self._codegen._mpm_full.run(self._final_module)
 
     def _get_module_for_linking(self):
         """
@@ -1037,7 +1042,9 @@ def _init(self, llvm_module):
         self._engine = JitEngine(engine)
         self._target_data = engine.target_data
         self._data_layout = str(self._target_data)
-        self._mpm = self._module_pass_manager()
+        self._mpm_cheap = self._module_pass_manager(loop_vectorize=False,
+                                                    opt=2)
+        self._mpm_full = self._module_pass_manager()
 
         self._engine.set_object_cache(self._library_class._object_compiled_hook,
                                       self._library_class._object_getbuffer_hook)
@@ -1066,21 +1073,21 @@ def create_library(self, name):
     def unserialize_library(self, serialized):
         return self._library_class._unserialize(self, serialized)
 
-    def _module_pass_manager(self):
+    def _module_pass_manager(self, **kwargs):
         pm = ll.create_module_pass_manager()
         self._tm.add_analysis_passes(pm)
-        with self._pass_manager_builder() as pmb:
+        with self._pass_manager_builder(**kwargs) as pmb:
             pmb.populate(pm)
         return pm
 
-    def _function_pass_manager(self, llvm_module):
+    def _function_pass_manager(self, llvm_module, **kwargs):
         pm = ll.create_function_pass_manager(llvm_module)
         self._tm.add_analysis_passes(pm)
-        with self._pass_manager_builder() as pmb:
+        with self._pass_manager_builder(**kwargs) as pmb:
             pmb.populate(pm)
         return pm
 
-    def _pass_manager_builder(self):
+    def _pass_manager_builder(self, **kwargs):
         """
         Create a PassManagerBuilder.
 
@@ -1089,8 +1096,18 @@ def _pass_manager_builder(self):
         or function pass manager.  Otherwise some optimizations will be
         missed...
         """
+        if 'opt' in kwargs:
+            opt_level = kwargs.pop('opt')
+        else:
+            opt_level = config.OPT
+
+        if 'loop_vectorize' in kwargs:
+            loop_vectorize = kwargs.pop('loop_vectorize')
+        else:
+            loop_vectorize = config.LOOP_VECTORIZE
+
         pmb = lp.create_pass_manager_builder(
-            opt=config.OPT, loop_vectorize=config.LOOP_VECTORIZE)
+            opt=opt_level, loop_vectorize=loop_vectorize, **kwargs)
         return pmb
 
     def _check_llvm_bugs(self):

numba/tests/test_vectorization.py

@@ -0,0 +1,36 @@
+import numpy as np
+from numba import njit, types
+from unittest import TestCase
+from numba.tests.support import override_env_config
+
+_DEBUG = False
+if _DEBUG:
+    from llvmlite import binding as llvm
+    # Prints debug info from the LLVMs vectorizer
+    llvm.set_option("", "--debug-only=loop-vectorize")
+
+
+class TestVectorization(TestCase):
+    """
+    Tests to assert that code which should vectorize does indeed vectorize
+    """
+    def gen_ir(self, func, args_tuple, **flags):
+        with override_env_config(
+            "NUMBA_CPU_NAME", "skylake-avx512"
+        ), override_env_config("NUMBA_CPU_FEATURES", ""):
+            jobj = njit(**flags)(func)
+            jobj.compile(args_tuple)
+            ol = jobj.overloads[jobj.signatures[0]]
+            return ol.library.get_llvm_str()
+
+    def test_nditer_loop(self):
+        # see https://github.com/numba/numba/issues/5033
+        def do_sum(x):
+            acc = 0
+            for v in np.nditer(x):
+                acc += v.item()
+            return acc
+
+        llvm_ir = self.gen_ir(do_sum, (types.float64[::1],), fastmath=True)
+        self.assertIn("vector.body", llvm_ir)
+        self.assertIn("llvm.loop.isvectorized", llvm_ir)