compiler: Postpone opt_pows until before unevaluation

Author: FabioLuporini

devito/core/cpu.py

@@ -177,7 +177,6 @@ def _specialize_clusters(cls, clusters, **kwargs):
         # Reduce flops
         clusters = cire(clusters, 'sops', sregistry, options, platform)
         clusters = factorize(clusters, **kwargs)
-        clusters = optimize_pows(clusters)
 
         # The previous passes may have created fusion opportunities
         clusters = fuse(clusters)

devito/core/gpu.py

@@ -218,7 +218,6 @@ def _specialize_clusters(cls, clusters, **kwargs):
         # Reduce flops
         clusters = cire(clusters, 'sops', sregistry, options, platform)
         clusters = factorize(clusters, **kwargs)
-        clusters = optimize_pows(clusters)
 
         # The previous passes may have created fusion opportunities
         clusters = fuse(clusters)

devito/operator/operator.py

@@ -29,7 +29,8 @@
 from devito.parameters import configuration
 from devito.passes import (
     Graph, lower_index_derivatives, generate_implicit, generate_macros,
-    minimize_symbols, unevaluate, error_mapper, is_on_device, lower_dtypes
+    minimize_symbols, optimize_pows, unevaluate, error_mapper, is_on_device,
+    lower_dtypes
 )
 from devito.symbolics import estimate_cost, subs_op_args
 from devito.tools import (DAG, OrderedSet, Signer, ReducerMap, as_mapper, as_tuple,
@@ -409,6 +410,10 @@ def _lower_clusters(cls, expressions, profiler=None, **kwargs):
         # Lower all remaining high order symbolic objects
         clusters = lower_index_derivatives(clusters, **kwargs)
 
+        # Turn pows into multiplications. This must happen as late as possible
+        # in the compilation process to maximize the optimization potential
+        clusters = optimize_pows(clusters)
+
         # Make sure no reconstructions can unpick any of the symbolic
         # optimizations performed so far
         clusters = unevaluate(clusters)