[pypy-commit] pypy vecopt: remember expansion (reduces register pressure if constant/variable is use more often). but not for heterogeneous expanded vectors, I don't think this happens frequently

[plan_rich]

Author: Richard Plangger <rich at pasra.at>
Branch: vecopt
Changeset: r77799:95db7332c363
Date: 2015-06-03 08:53 +0200
http://bitbucket.org/pypy/pypy/changeset/95db7332c363/

Log:	remember expansion (reduces register pressure if constant/variable
	is use more often). but not for heterogeneous expanded vectors, I
	don't think this happens frequently

diff --git a/rpython/jit/metainterp/optimizeopt/test/test_schedule.py b/rpython/jit/metainterp/optimizeopt/test/test_schedule.py
--- a/rpython/jit/metainterp/optimizeopt/test/test_schedule.py
+++ b/rpython/jit/metainterp/optimizeopt/test/test_schedule.py
@@ -108,6 +108,7 @@
         v2[i64#2] = vec_int_pack(v1[i64#2], i0, 0, 1)
         v3[i64#2] = vec_int_pack(v2[i64#2], i1, 1, 1)
         v4[i64#2] = vec_int_expand(73)
+        #
         v5[i64#2] = vec_int_add(v3[i64#2], v4[i64#2])
         """, False)
         self.assert_equal(loop2, loop3)
@@ -123,6 +124,28 @@
         v2[f64#2] = vec_float_pack(v1[f64#2], f0, 0, 1)
         v3[f64#2] = vec_float_pack(v2[f64#2], f1, 1, 1)
         v4[f64#2] = vec_float_expand(73.0)
+        #
         v5[f64#2] = vec_float_add(v3[f64#2], v4[f64#2])
         """, False)
         self.assert_equal(loop2, loop3)
+
+    def test_scalar_remember_expansion(self):
+        loop1 = self.parse("""
+        f10 = float_add(f0, f5)
+        f11 = float_add(f1, f5)
+        f12 = float_add(f10, f5)
+        f13 = float_add(f11, f5)
+        """)
+        pack1 = self.pack(loop1, 0, 2)
+        pack2 = self.pack(loop1, 2, 4)
+        loop2 = self.schedule(loop1, [pack1, pack2], prepend_invariant=True)
+        loop3 = self.parse("""
+        v1[f64#2] = vec_box(2)
+        v2[f64#2] = vec_float_pack(v1[f64#2], f0, 0, 1)
+        v3[f64#2] = vec_float_pack(v2[f64#2], f1, 1, 1)
+        v4[f64#2] = vec_float_expand(f5) # only expaned once
+        #
+        v5[f64#2] = vec_float_add(v3[f64#2], v4[f64#2])
+        v6[f64#2] = vec_float_add(v5[f64#2], v4[f64#2])
+        """, False)
+        self.assert_equal(loop2, loop3)
diff --git a/rpython/jit/metainterp/optimizeopt/vectorize.py b/rpython/jit/metainterp/optimizeopt/vectorize.py
--- a/rpython/jit/metainterp/optimizeopt/vectorize.py
+++ b/rpython/jit/metainterp/optimizeopt/vectorize.py
@@ -1052,11 +1052,17 @@
     def expand(self, nodes, arg, argidx):
         vbox = self.input_type.new_vector_box(len(nodes))
         box_type = arg.type
+        expanded_map = self.sched_data.expanded_map
         invariant_ops = self.sched_data.invariant_oplist
         invariant_vars = self.sched_data.invariant_vector_vars
         if isinstance(arg, BoxVector):
             box_type = arg.item_type
 
+        # note that heterogenous nodes are not yet tracked
+        already_expanded = expanded_map.get(arg, None)
+        if already_expanded:
+            return already_expanded
+
         for i, node in enumerate(nodes):
             op = node.getoperation()
             if not arg.same_box(op.getarg(argidx)):
@@ -1069,6 +1075,7 @@
             op = ResOperation(expand_opnum, [arg], vbox)
             invariant_ops.append(op)
             invariant_vars.append(vbox)
+            expanded_map[arg] = vbox
             return vbox
 
         op = ResOperation(rop.VEC_BOX, [ConstInt(len(nodes))], vbox)
@@ -1085,6 +1092,7 @@
             op = ResOperation(opnum, [vbox,arg,ci,c1], new_box)
             vbox = new_box
             invariant_ops.append(op)
+
         invariant_vars.append(vbox)
         return vbox
 
@@ -1239,6 +1247,7 @@
         self.vec_reg_size = vec_reg_size
         self.invariant_oplist = []
         self.invariant_vector_vars = []
+        self.expanded_map = {}
 
     def as_vector_operation(self, pack):
         op_count = len(pack.operations)