def optimise(self):
"""
Optimise the IR.
Procedure:
1. split in frames
2. convert frames to graphs
3. optimise graphs
4. convert graphs to (flat) frames
5. concatenate frames to get optimised program.
Store result in flat.
"""
# top loop
flat_orig = None
top_loop_counter = 0
while True:
if flat_orig == self.flat:
self.logger.info('optimisation is stable')
break
if top_loop_counter == 10000:
self.logger.warning('top loop limit reached (10000 iterations)')
break
flat_orig = self.flat[:]
top_loop_counter += 1
self.logger.info('top pass %s' % str(top_loop_counter))
# a.
if 'a' in self.enabled_optimisations:
self.logger.info('optimising global control flow graph')
cfg = CFG(self.flat)
#if self.verbosity > 2:
# cfg.cfg_to_diagram("allinstr_graph_before.png")
optimise_tree.optimise(cfg)
#if self.verbosity > 2:
# cfg.cfg_to_diagram("allinstr_graph_after.png")
self.flat = cfg.cfg_to_flat()
# b. jump optimisations
if 'b' in self.enabled_optimisations:
self.logger.info('optimising flat (jumps and branches)')
self.flat = flat_opt.optimise(self.flat)
self.flat = parse_instr.parse(self.flat)
self.logger.info('splitting flat in frames')
frames = split_frames(self.flat)
self.logger.info('creating graph for each frame')
graphs = [CFG(frame) for frame in frames]
self.logger.info('optimising blocks')
for graphnr, graph in enumerate(graphs):
self.logger.info('graph %d of %d' % (graphnr + 1, len(graphs)))
#Dataflow(graph)
if 'f' in self.enabled_optimisations:
l = Liveness(graph,self.verbosity)
self.logger.info('Performing liveness optimalisation on graph')
change = True
while change:
l.analyse()
change = l.optimise()
for blocknr, block in enumerate(graph.blocks):
self.logger.debug('block %d of %d' % (blocknr + 1, len(graph.blocks)))
cf_opt = b_opt.ConstantFold(block)
cp_opt = b_opt.CopyPropagation(block)
dc_opt = b_opt.DeadCode(block)
done = False
subopt_changes = False
i = 0
while (not done):
done = True
i += 1
self.logger.debug('\t pass '+str(i))
# c. constant folding
if 'c' in self.enabled_optimisations:
subopt_changes = cf_opt.optimise()
if subopt_changes: self.stats['cf'] += cf_opt.stats['cf']
done = done & (not subopt_changes)
# d. copy propagation
if 'd' in self.enabled_optimisations:
subopt_changes = cp_opt.optimise()
if subopt_changes:self.stats['cp'] += cp_opt.stats['cp']
done = done & (not subopt_changes)
# e. dead code removal
if 'e' in self.enabled_optimisations:
subopt_changes = dc_opt.optimise()
if subopt_changes:self.stats['dc'] += dc_opt.stats['dc']
done = done & (not subopt_changes)
self.logger.info('basic-block peephole optimisations done:')
self.logger.info('\t constant folds: %d' % (self.stats['cf']))
self.logger.info('\t copy propagations: %d' % (self.stats['cp']))
self.logger.info('\t dead code removes: %d' % (self.stats['dc']))
self.logger.info('joining graphs to frames')
frames = [graph.cfg_to_flat() for graph in graphs]
self.logger.info('joining frames to flat')
self.flat = sum(frames, [])
def optimise(self):
"""
Optimise the IR.
Procedure:
1. split in frames
2. convert frames to graphs
3. optimise graphs
4. convert graphs to (flat) frames
5. concatenate frames to get optimised program.
Store result in flat.
"""
self.logger.info('optimising global control flow graph')
cfg = CFG(self.flat)
if self.verbosity > 2:
cfg.cfg_to_diagram("allinstr_graph_before.png")
optimise_tree.optimise(cfg)
if self.verbosity > 2:
cfg.cfg_to_diagram("allinstr_graph_after.png")
self.flat = cfg.cfg_to_flat()
self.logger.info('optimising flat (jumps and branches)')
self.flat = flat_opt.optimise(self.flat)
self.logger.info('splitting flat in frames')
frames = split_frames(self.flat)
self.logger.info('creating graph for each frame')
graphs = [CFG(frame) for frame in frames]
self.logger.info('optimising blocks')
for graphnr, graph in enumerate(graphs):
self.logger.info('graph %d of %d' % (graphnr + 1, len(graphs)))
Dataflow(graph)
l = Liveness(graph,self.verbosity)
#self.logger.info('Performing liveness optimalisation on graph')
#change = True
#while change:
# l.analyse()
# change = l.optimise()
for blocknr, block in enumerate(graph.blocks):
self.logger.debug('block %d of %d' % (blocknr + 1, len(graph.blocks)))
cf_opt = b_opt.ConstantFold(block)
cp_opt = b_opt.CopyPropagation(block)
dc_opt = b_opt.DeadCode(block)
done = False
subopt_changes = False
i = 0
while (not done):
done = True
i += 1
self.logger.debug('pass '+str(i))
subopt_changes = cf_opt.optimise()
if subopt_changes:self.stats['cf'] += cf_opt.stats['cf']
done = done & (not subopt_changes)
subopt_changes = cp_opt.optimise()
if subopt_changes:self.stats['cp'] += cp_opt.stats['cp']
done = done & (not subopt_changes)
subopt_changes = dc_opt.optimise()
if subopt_changes:self.stats['dc'] += dc_opt.stats['dc']
done = done & (not subopt_changes)
self.logger.info('basic-block peephole optimisations done:')
self.logger.info('\t\tconstant folds: %d' % (self.stats['cf']))
self.logger.info('\t\tcopy propagations: %d' % (self.stats['cp']))
self.logger.info('\t\tdead code removes: %d' % (self.stats['dc']))
self.logger.info('joining graphs to frames')
frames = [graph.cfg_to_flat() for graph in graphs]
self.logger.info('joining frames to flat')
self.flat = sum(frames, [])
