1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
|
from rpython.jit.backend.llsupport.rewrite import cpu_simplify_scale
from rpython.jit.metainterp.history import (VECTOR, FLOAT, INT,
ConstInt, ConstFloat, TargetToken)
from rpython.jit.metainterp.resoperation import (rop, ResOperation,
GuardResOp, VecOperation, OpHelpers, VecOperationNew,
VectorizationInfo)
from rpython.jit.metainterp.optimizeopt.dependency import (DependencyGraph,
MemoryRef, Node, IndexVar)
from rpython.jit.metainterp.optimizeopt.renamer import Renamer
from rpython.jit.metainterp.resume import AccumInfo
from rpython.rlib.objectmodel import we_are_translated
from rpython.jit.metainterp.jitexc import NotAProfitableLoop
from rpython.rlib.objectmodel import specialize, always_inline
from rpython.jit.metainterp.jitexc import NotAVectorizeableLoop, NotAProfitableLoop
from rpython.rtyper.lltypesystem.lloperation import llop
from rpython.rtyper.lltypesystem import lltype
from rpython.rlib.debug import debug_print
def forwarded_vecinfo(op):
fwd = op.get_forwarded()
if fwd is None or not isinstance(fwd, VectorizationInfo):
# the optimizer clears getforwarded AFTER
# vectorization, it happens that this is not clean
fwd = VectorizationInfo(op)
if not op.is_constant():
op.set_forwarded(fwd)
return fwd
class SchedulerState(object):
def __init__(self, cpu, graph):
self.cpu = cpu
self.renamer = Renamer()
self.graph = graph
self.oplist = []
self.worklist = []
self.invariant_oplist = []
self.invariant_vector_vars = []
self.seen = {}
self.delayed = []
def resolve_delayed(self, needs_resolving, delayed, op):
# recursive solving of all delayed objects
if not delayed:
return
args = op.getarglist()
if op.is_guard():
args = args[:] + op.getfailargs()
for arg in args:
if arg is None or arg.is_constant() or arg.is_inputarg():
continue
if arg not in self.seen:
box = self.renamer.rename_box(arg)
needs_resolving[box] = None
indexvars = self.graph.index_vars
i = len(delayed)-1
while i >= 0:
node = delayed[i]
op = node.getoperation()
if op in needs_resolving:
# either it is a normal operation, or we know that there is a linear combination
del needs_resolving[op]
if op in indexvars:
opindexvar = indexvars[op]
# there might be a variable already, that
# calculated the index variable, thus just reuse it
for var, indexvar in indexvars.items():
if indexvar == opindexvar and var in self.seen:
self.renamer.start_renaming(op, var)
break
else:
if opindexvar.calculated_by(op):
# just append this operation
self.seen[op] = None
self.append_to_oplist(op)
else:
# here is an easier way to calculate just this operation
last = op
for operation in opindexvar.get_operations():
self.append_to_oplist(operation)
last = operation
indexvars[last] = opindexvar
self.renamer.start_renaming(op, last)
self.seen[op] = None
self.seen[last] = None
else:
self.resolve_delayed(needs_resolving, delayed, op)
self.append_to_oplist(op)
self.seen[op] = None
if len(delayed) > i:
del delayed[i]
i -= 1
# some times the recursive call can remove several items from delayed,
# thus we correct the index here
if len(delayed) <= i:
i = len(delayed)-1
def append_to_oplist(self, op):
self.renamer.rename(op)
self.oplist.append(op)
def schedule(self):
self.prepare()
Scheduler().walk_and_emit(self)
self.post_schedule()
def post_schedule(self):
loop = self.graph.loop
jump = loop.jump
if self.delayed:
# some operations can be delayed until the jump instruction,
# handle them here
self.resolve_delayed({}, self.delayed, jump)
self.renamer.rename(jump)
loop.operations = self.oplist
def profitable(self):
return True
def prepare(self):
for node in self.graph.nodes:
if node.depends_count() == 0:
self.worklist.insert(0, node)
def try_emit_or_delay(self, node):
if not node.is_imaginary() and node.is_pure():
# this operation might never be emitted. only if it is really needed
self.delay_emit(node)
return
# emit a now!
self.pre_emit(node, True)
self.mark_emitted(node)
if not node.is_imaginary():
op = node.getoperation()
self.seen[op] = None
self.append_to_oplist(op)
def delay_emit(self, node):
""" it has been decided that the operation might be scheduled later """
delayed = node.delayed or []
if node not in delayed:
delayed.append(node)
node.delayed = None
provides = node.provides()
if len(provides) == 0:
for n in delayed:
self.delayed.append(n)
else:
for to in node.provides():
tnode = to.target_node()
self.delegate_delay(tnode, delayed[:])
self.mark_emitted(node)
def delegate_delay(self, node, delayed):
""" Chain up delays, this can reduce many more of the operations """
if node.delayed is None:
node.delayed = delayed
else:
delayedlist = node.delayed
for d in delayed:
if d not in delayedlist:
delayedlist.append(d)
def mark_emitted(state, node, unpack=True):
""" An operation has been emitted, adds new operations to the worklist
whenever their dependency count drops to zero.
Keeps worklist sorted (see priority) """
worklist = state.worklist
provides = node.provides()[:]
for dep in provides: # COPY
target = dep.to
node.remove_edge_to(target)
if not target.emitted and target.depends_count() == 0:
# sorts them by priority
i = len(worklist)-1
while i >= 0:
cur = worklist[i]
c = (cur.priority - target.priority)
if c < 0: # meaning itnode.priority < target.priority:
worklist.insert(i+1, target)
break
elif c == 0:
# if they have the same priority, sort them
# using the original position in the trace
if target.getindex() < cur.getindex():
worklist.insert(i+1, target)
break
i -= 1
else:
worklist.insert(0, target)
node.clear_dependencies()
node.emitted = True
if not node.is_imaginary():
op = node.getoperation()
state.renamer.rename(op)
if unpack:
state.ensure_args_unpacked(op)
state.post_emit(node)
def delay(self, node):
return False
def has_more(self):
return len(self.worklist) > 0
def ensure_args_unpacked(self, op):
pass
def post_emit(self, node):
pass
def pre_emit(self, orignode, pack_first=True):
delayed = orignode.delayed
if delayed:
# there are some nodes that have been delayed just for this operation
if pack_first:
op = orignode.getoperation()
self.resolve_delayed({}, delayed, op)
for node in delayed:
op = node.getoperation()
if op in self.seen:
continue
if node is not None:
provides = node.provides()
if len(provides) == 0:
# add this node to the final delay list
# might be emitted before jump!
self.delayed.append(node)
else:
for to in node.provides():
tnode = to.target_node()
self.delegate_delay(tnode, [node])
orignode.delayed = None
class Scheduler(object):
""" Create an instance of this class to (re)schedule a vector trace. """
def __init__(self):
pass
def next(self, state):
""" select the next candidate node to be emitted, or None """
worklist = state.worklist
visited = 0
while len(worklist) > 0:
if visited == len(worklist):
return None
node = worklist.pop()
if node.emitted:
continue
if not self.delay(node, state):
return node
worklist.insert(0, node)
visited += 1
return None
def try_to_trash_pack(self, state):
# one element a pack has several dependencies pointing to
# it thus we MUST skip this pack!
if len(state.worklist) > 0:
# break the first!
i = 0
node = state.worklist[i]
i += 1
while i < len(state.worklist) and not node.pack:
node = state.worklist[i]
i += 1
if not node.pack:
return False
pack = node.pack
for n in node.pack.operations:
if n.depends_count() > 0:
pack.clear()
return True
else:
return False
return False
def delay(self, node, state):
""" Delay this operation?
Only if any dependency has not been resolved """
if state.delay(node):
return True
return node.depends_count() != 0
def walk_and_emit(self, state):
""" Emit all the operations into the oplist parameter.
Initiates the scheduling. """
assert isinstance(state, SchedulerState)
while state.has_more():
node = self.next(state)
if node:
state.try_emit_or_delay(node)
continue
# it happens that packs can emit many nodes that have been
# added to the scheuldable_nodes list, in this case it could
# be that no next exists even though the list contains elements
if not state.has_more():
break
if self.try_to_trash_pack(state):
continue
raise AssertionError("schedule failed cannot continue. possible reason: cycle")
if not we_are_translated():
for node in state.graph.nodes:
assert node.emitted
def failnbail_transformation(msg):
msg = '%s\n' % msg
debug_print(msg)
raise NotAVectorizeableLoop
def turn_into_vector(state, pack):
""" Turn a pack into a vector instruction """
check_if_pack_supported(state, pack)
state.costmodel.record_pack_savings(pack, pack.numops())
left = pack.leftmost()
oprestrict = state.cpu.vector_ext.get_operation_restriction(left)
if oprestrict is not None:
newargs = oprestrict.check_operation(state, pack, left)
if newargs:
args = newargs
else:
args = left.getarglist_copy()
else:
args = left.getarglist_copy()
prepare_arguments(state, oprestrict, pack, args)
vecop = VecOperation(left.vector, args, left,
pack.numops(), left.getdescr())
for i,node in enumerate(pack.operations):
op = node.getoperation()
if op.returns_void():
continue
state.setvector_of_box(op,i,vecop)
if pack.is_accumulating() and not op.is_guard():
state.renamer.start_renaming(op, vecop)
if left.is_guard():
prepare_fail_arguments(state, pack, left, vecop)
state.append_to_oplist(vecop)
assert vecop.count >= 1
def prepare_arguments(state, oprestrict, pack, args):
# Transforming one argument to a vector box argument
# The following cases can occur:
# 1) argument is present in the box_to_vbox map.
# a) vector can be reused immediatly (simple case)
# b) the size of the input is mismatching (crop the vector)
# c) values are scattered in differnt registers
# d) the operand is not at the right position in the vector
# 2) argument is not known to reside in a vector
# a) expand vars/consts before the label and add as argument
# b) expand vars created in the loop body
#
if not oprestrict:
return
restrictions = oprestrict.argument_restrictions
for i,arg in enumerate(args):
if i >= len(restrictions) or restrictions[i] is None:
# ignore this argument
continue
restrict = restrictions[i]
if arg.returns_vector():
restrict.check(arg)
continue
pos, vecop = state.getvector_of_box(arg)
if not vecop:
# 2) constant/variable expand this box
expand(state, pack, args, arg, i)
restrict.check(args[i])
continue
# 1)
args[i] = vecop # a)
assemble_scattered_values(state, pack, args, i) # c)
position_values(state, restrict, pack, args, i, pos) # d)
crop_vector(state, oprestrict, restrict, pack, args, i) # b)
restrict.check(args[i])
def prepare_fail_arguments(state, pack, left, vecop):
assert isinstance(left, GuardResOp)
assert isinstance(vecop, GuardResOp)
args = left.getfailargs()[:]
for i, arg in enumerate(args):
pos, newarg = state.getvector_of_box(arg)
if newarg is None:
newarg = arg
if newarg.is_vector(): # can be moved to guard exit!
newarg = unpack_from_vector(state, newarg, 0, 1)
args[i] = newarg
vecop.setfailargs(args)
# TODO vecop.rd_snapshot = left.rd_snapshot
@always_inline
def crop_vector(state, oprestrict, restrict, pack, args, i):
# convert size i64 -> i32, i32 -> i64, ...
arg = args[i]
vecinfo = forwarded_vecinfo(arg)
size = vecinfo.bytesize
left = pack.leftmost()
if oprestrict.must_crop_vector(left, i):
newsize = oprestrict.crop_to_size(left, i)
assert arg.type == 'i'
state._prevent_signext(newsize, size)
count = vecinfo.count
vecop = VecOperationNew(rop.VEC_INT_SIGNEXT, [arg, ConstInt(newsize)],
'i', newsize, vecinfo.signed, count)
state.append_to_oplist(vecop)
state.costmodel.record_cast_int(size, newsize, count)
args[i] = vecop
@always_inline
def assemble_scattered_values(state, pack, args, index):
args_at_index = [node.getoperation().getarg(index) for node in pack.operations]
args_at_index[0] = args[index]
vectors = pack.argument_vectors(state, pack, index, args_at_index)
if len(vectors) > 1:
# the argument is scattered along different vector boxes
args[index] = gather(state, vectors, pack.numops())
state.remember_args_in_vector(pack, index, args[index])
@always_inline
def gather(state, vectors, count): # packed < packable and packed < stride:
(_, arg) = vectors[0]
i = 1
while i < len(vectors):
(newarg_pos, newarg) = vectors[i]
vecinfo = forwarded_vecinfo(arg)
newvecinfo = forwarded_vecinfo(newarg)
if vecinfo.count + newvecinfo.count <= count:
arg = pack_into_vector(state, arg, vecinfo.count, newarg, newarg_pos, newvecinfo.count)
i += 1
return arg
@always_inline
def position_values(state, restrict, pack, args, index, position):
arg = args[index]
vecinfo = forwarded_vecinfo(arg)
count = vecinfo.count
newcount = restrict.count
if not restrict.any_count() and newcount != count:
if position == 0:
pass
pass
if position != 0:
# The vector box is at a position != 0 but it
# is required to be at position 0. Unpack it!
arg = args[index]
vecinfo = forwarded_vecinfo(arg)
count = restrict.max_input_count(vecinfo.count)
args[index] = unpack_from_vector(state, arg, position, count)
state.remember_args_in_vector(pack, index, args[index])
def check_if_pack_supported(state, pack):
left = pack.leftmost()
vecinfo = forwarded_vecinfo(left)
insize = vecinfo.bytesize
if left.is_typecast():
# prohibit the packing of signext calls that
# cast to int16/int8.
state._prevent_signext(left.cast_to_bytesize(),
left.cast_from_bytesize())
if left.getopnum() == rop.INT_MUL:
if insize == 8 or insize == 1:
# see assembler for comment why
raise NotAProfitableLoop
def unpack_from_vector(state, arg, index, count):
""" Extract parts of the vector box into another vector box """
assert count > 0
vecinfo = forwarded_vecinfo(arg)
assert index + count <= vecinfo.count
args = [arg, ConstInt(index), ConstInt(count)]
vecop = OpHelpers.create_vec_unpack(arg.type, args, vecinfo.bytesize,
vecinfo.signed, count)
state.costmodel.record_vector_unpack(arg, index, count)
state.append_to_oplist(vecop)
return vecop
def pack_into_vector(state, tgt, tidx, src, sidx, scount):
""" tgt = [1,2,3,4,_,_,_,_]
src = [5,6,_,_]
new_box = [1,2,3,4,5,6,_,_] after the operation, tidx=4, scount=2
"""
assert sidx == 0 # restriction
vecinfo = forwarded_vecinfo(tgt)
newcount = vecinfo.count + scount
args = [tgt, src, ConstInt(tidx), ConstInt(scount)]
vecop = OpHelpers.create_vec_pack(tgt.type, args, vecinfo.bytesize, vecinfo.signed, newcount)
state.append_to_oplist(vecop)
state.costmodel.record_vector_pack(src, sidx, scount)
if not we_are_translated():
_check_vec_pack(vecop)
return vecop
def _check_vec_pack(op):
arg0 = op.getarg(0)
arg1 = op.getarg(1)
index = op.getarg(2)
count = op.getarg(3)
assert op.is_vector()
assert arg0.is_vector()
assert index.is_constant()
assert isinstance(count, ConstInt)
vecinfo = forwarded_vecinfo(op)
argvecinfo = forwarded_vecinfo(arg0)
assert argvecinfo.bytesize == vecinfo.bytesize
if arg1.is_vector():
assert argvecinfo.bytesize == vecinfo.bytesize
else:
assert count.value == 1
assert index.value < vecinfo.count
assert index.value + count.value <= vecinfo.count
assert vecinfo.count > argvecinfo.count
def expand(state, pack, args, arg, index):
""" Expand a value into a vector box. useful for arithmetic
of one vector with a scalar (either constant/varialbe)
"""
left = pack.leftmost()
box_type = arg.type
expanded_map = state.expanded_map
ops = state.invariant_oplist
variables = state.invariant_vector_vars
if not arg.is_constant() and arg not in state.inputargs:
# cannot be created before the loop, expand inline
ops = state.oplist
variables = None
for i, node in enumerate(pack.operations):
op = node.getoperation()
if not arg.same_box(op.getarg(index)):
break
i += 1
else:
# note that heterogenous nodes are not yet tracked
vecop = state.find_expanded([arg])
if vecop:
args[index] = vecop
return vecop
left = pack.leftmost()
vecinfo = forwarded_vecinfo(left)
vecop = OpHelpers.create_vec_expand(arg, vecinfo.bytesize, vecinfo.signed, pack.numops())
ops.append(vecop)
if variables is not None:
variables.append(vecop)
state.expand([arg], vecop)
args[index] = vecop
return vecop
# quick search if it has already been expanded
expandargs = [op.getoperation().getarg(index) for op in pack.operations]
vecop = state.find_expanded(expandargs)
if vecop:
args[index] = vecop
return vecop
arg_vecinfo = forwarded_vecinfo(arg)
vecop = OpHelpers.create_vec(arg.type, arg_vecinfo.bytesize, arg_vecinfo.signed, pack.opnum())
ops.append(vecop)
for i,node in enumerate(pack.operations):
op = node.getoperation()
arg = op.getarg(index)
arguments = [vecop, arg, ConstInt(i), ConstInt(1)]
vecinfo = forwarded_vecinfo(vecop)
vecop = OpHelpers.create_vec_pack(arg.type, arguments, vecinfo.bytesize,
vecinfo.signed, vecinfo.count+1)
ops.append(vecop)
state.expand(expandargs, vecop)
if variables is not None:
variables.append(vecop)
args[index] = vecop
class VecScheduleState(SchedulerState):
def __init__(self, graph, packset, cpu, costmodel):
SchedulerState.__init__(self, cpu, graph)
self.box_to_vbox = {}
self.vec_reg_size = cpu.vector_ext.vec_size()
self.expanded_map = {}
self.costmodel = costmodel
self.inputargs = {}
self.packset = packset
for arg in graph.loop.inputargs:
self.inputargs[arg] = None
self.accumulation = {}
def expand(self, args, vecop):
index = 0
if len(args) == 1:
# loop is executed once, thus sets -1 as index
index = -1
for arg in args:
self.expanded_map.setdefault(arg, []).append((vecop, index))
index += 1
def find_expanded(self, args):
if len(args) == 1:
candidates = self.expanded_map.get(args[0], [])
for (vecop, index) in candidates:
if index == -1:
# found an expanded variable/constant
return vecop
return None
possible = {}
for i, arg in enumerate(args):
expansions = self.expanded_map.get(arg, [])
candidates = [vecop for (vecop, index) in expansions \
if i == index and possible.get(vecop,True)]
for vecop in candidates:
for key in possible.keys():
if key not in candidates:
# delete every not possible key,value
possible[key] = False
# found a candidate, append it if not yet present
possible[vecop] = True
if not possible:
# no possibility left, this combination is not expanded
return None
for vecop,valid in possible.items():
if valid:
return vecop
return None
def post_emit(self, node):
pass
def pre_emit(self, node, pack_first=True):
op = node.getoperation()
if op.is_guard():
# add accumulation info to the descriptor
failargs = op.getfailargs()[:]
descr = op.getdescr()
# note: stitching a guard must resemble the order of the label
# otherwise a wrong mapping is handed to the register allocator
for i,arg in enumerate(failargs):
if arg is None:
continue
accum = self.accumulation.get(arg, None)
if accum:
from rpython.jit.metainterp.compile import AbstractResumeGuardDescr
assert isinstance(accum, AccumPack)
assert isinstance(descr, AbstractResumeGuardDescr)
info = AccumInfo(i, arg, accum.operator)
descr.attach_vector_info(info)
seed = accum.getleftmostseed()
failargs[i] = self.renamer.rename_map.get(seed, seed)
op.setfailargs(failargs)
SchedulerState.pre_emit(self, node, pack_first)
def profitable(self):
return self.costmodel.profitable()
def prepare(self):
SchedulerState.prepare(self)
self.packset.accumulate_prepare(self)
for arg in self.graph.loop.label.getarglist():
self.seen[arg] = None
def try_emit_or_delay(self, node):
# emission might be blocked by other nodes if this node has a pack!
if node.pack:
assert node.pack.numops() > 1
for i,node in enumerate(node.pack.operations):
self.pre_emit(node, i==0)
self.mark_emitted(node, unpack=False)
turn_into_vector(self, node.pack)
elif not node.emitted:
SchedulerState.try_emit_or_delay(self, node)
def delay(self, node):
if node.pack:
pack = node.pack
if pack.is_accumulating():
for node in pack.operations:
for dep in node.depends():
if dep.to.pack is not pack:
return True
else:
for node in pack.operations:
if node.depends_count() > 0:
return True
return False
def ensure_args_unpacked(self, op):
""" If a box is needed that is currently stored within a vector
box, this utility creates a unpacking instruction.
"""
# unpack for an immediate use
for i, argument in enumerate(op.getarglist()):
if not argument.is_constant():
arg = self.ensure_unpacked(i, argument)
if argument is not arg:
op.setarg(i, arg)
# unpack for a guard exit
if op.is_guard():
# could be moved to the guard exit
fail_args = op.getfailargs()
for i, argument in enumerate(fail_args):
if argument and not argument.is_constant():
arg = self.ensure_unpacked(i, argument)
if argument is not arg:
fail_args[i] = arg
op.setfailargs(fail_args)
def ensure_unpacked(self, index, arg):
if arg in self.seen or arg.is_vector():
return arg
(pos, var) = self.getvector_of_box(arg)
if var:
if var in self.invariant_vector_vars:
return arg
if arg in self.accumulation:
return arg
args = [var, ConstInt(pos), ConstInt(1)]
vecinfo = forwarded_vecinfo(var)
vecop = OpHelpers.create_vec_unpack(var.type, args, vecinfo.bytesize,
vecinfo.signed, 1)
self.renamer.start_renaming(arg, vecop)
self.seen[vecop] = None
self.costmodel.record_vector_unpack(var, pos, 1)
self.append_to_oplist(vecop)
return vecop
return arg
def _prevent_signext(self, outsize, insize):
if insize != outsize:
if outsize < 4 or insize < 4:
raise NotAProfitableLoop
def getvector_of_box(self, arg):
return self.box_to_vbox.get(arg, (-1, None))
def setvector_of_box(self, var, off, vector):
if var.returns_void():
assert 0, "not allowed to rename void resop"
vecinfo = forwarded_vecinfo(vector)
assert off < vecinfo.count
assert not var.is_vector()
self.box_to_vbox[var] = (off, vector)
def remember_args_in_vector(self, pack, index, box):
arguments = [op.getoperation().getarg(index) for op in pack.operations]
for i,arg in enumerate(arguments):
vecinfo = forwarded_vecinfo(arg)
if i >= vecinfo.count:
break
self.setvector_of_box(arg, i, box)
def post_schedule(self):
SchedulerState.post_schedule(self)
loop = self.graph.loop
self.ensure_args_unpacked(loop.jump)
loop.prefix = self.invariant_oplist
if len(self.invariant_vector_vars) + len(self.invariant_oplist) > 0:
# label
args = loop.label.getarglist_copy() + self.invariant_vector_vars
opnum = loop.label.getopnum()
op = loop.label.copy_and_change(opnum, args)
self.renamer.rename(op)
loop.prefix_label = op
# jump
args = loop.jump.getarglist_copy() + self.invariant_vector_vars
opnum = loop.jump.getopnum()
op = loop.jump.copy_and_change(opnum, args)
self.renamer.rename(op)
loop.jump = op
class Pack(object):
""" A pack is a set of n statements that are:
* isomorphic
* independent
"""
FULL = 0
_attrs_ = ('operations', 'accumulator', 'operator', 'position')
operator = '\x00'
position = -1
accumulator = None
def __init__(self, ops):
self.operations = ops
self.update_pack_of_nodes()
def numops(self):
return len(self.operations)
@specialize.arg(1)
def leftmost(self, node=False):
if node:
return self.operations[0]
return self.operations[0].getoperation()
@specialize.arg(1)
def rightmost(self, node=False):
if node:
return self.operations[-1]
return self.operations[-1].getoperation()
def pack_type(self):
ptype = self.input_type
if self.input_type is None:
# load does not have an input type, but only an output type
ptype = self.output_type
return ptype
def input_byte_size(self):
""" The amount of bytes the operations need with the current
entries in self.operations. E.g. cast_singlefloat_to_float
takes only #2 operations.
"""
return self._byte_size(self.input_type)
def output_byte_size(self):
""" The amount of bytes the operations need with the current
entries in self.operations. E.g. vec_load(..., descr=short)
with 10 operations returns 20
"""
return self._byte_size(self.output_type)
def pack_load(self, vec_reg_size):
""" Returns the load of the pack a vector register would hold
just after executing the operation.
returns: < 0 - empty, nearly empty
= 0 - full
> 0 - overloaded
"""
left = self.leftmost()
if left.returns_void():
if rop.is_primitive_store(left.opnum):
# make this case more general if it turns out this is
# not the only case where packs need to be trashed
descr = left.getdescr()
bytesize = descr.get_item_size_in_bytes()
return bytesize * self.numops() - vec_reg_size
else:
assert left.is_guard() and left.getopnum() in \
(rop.GUARD_TRUE, rop.GUARD_FALSE)
vecinfo = forwarded_vecinfo(left.getarg(0))
bytesize = vecinfo.bytesize
return bytesize * self.numops() - vec_reg_size
return 0
if self.numops() == 0:
return -1
if left.is_typecast():
# casting is special, often only takes a half full vector
if left.casts_down():
# size is reduced
size = left.cast_input_bytesize(vec_reg_size)
return left.cast_from_bytesize() * self.numops() - size
else:
# size is increased
#size = left.cast_input_bytesize(vec_reg_size)
return left.cast_to_bytesize() * self.numops() - vec_reg_size
vecinfo = forwarded_vecinfo(left)
return vecinfo.bytesize * self.numops() - vec_reg_size
def is_full(self, vec_reg_size):
""" If one input element times the opcount is equal
to the vector register size, we are full!
"""
return self.pack_load(vec_reg_size) == Pack.FULL
def opnum(self):
assert len(self.operations) > 0
return self.operations[0].getoperation().getopnum()
def clear(self):
for node in self.operations:
node.pack = None
node.pack_position = -1
def update_pack_of_nodes(self):
for i,node in enumerate(self.operations):
node.pack = self
node.pack_position = i
def split(self, packlist, vec_reg_size, vector_ext):
""" Combination phase creates the biggest packs that are possible.
In this step the pack is reduced in size to fit into an
vector register.
"""
before_count = len(packlist)
pack = self
while pack.pack_load(vec_reg_size) > Pack.FULL:
pack.clear()
oplist, newoplist = pack.slice_operations(vec_reg_size, vector_ext)
pack.operations = oplist
pack.update_pack_of_nodes()
if not pack.leftmost().is_typecast():
assert pack.is_full(vec_reg_size)
#
newpack = pack.clone(newoplist)
load = newpack.pack_load(vec_reg_size)
if load >= Pack.FULL:
pack.update_pack_of_nodes()
pack = newpack
packlist.append(newpack)
else:
newpack.clear()
newpack.operations = []
break
pack.update_pack_of_nodes()
def opcount_filling_vector_register(self, vec_reg_size, vector_ext):
left = self.leftmost()
oprestrict = vector_ext.get_operation_restriction(left)
return oprestrict.opcount_filling_vector_register(left, vec_reg_size)
def slice_operations(self, vec_reg_size, vector_ext):
count = self.opcount_filling_vector_register(vec_reg_size, vector_ext)
assert count > 0
newoplist = self.operations[count:]
oplist = self.operations[:count]
assert len(newoplist) + len(oplist) == len(self.operations)
assert len(newoplist) != 0
return oplist, newoplist
def rightmost_match_leftmost(self, other):
""" Check if pack A can be combined with pack B """
assert isinstance(other, Pack)
rightmost = self.operations[-1]
leftmost = other.operations[0]
# if it is not accumulating it is valid
if self.is_accumulating():
if not other.is_accumulating():
return False
elif self.position != other.position:
return False
return rightmost is leftmost
def argument_vectors(self, state, pack, index, pack_args_index):
vectors = []
last = None
for arg in pack_args_index:
pos, vecop = state.getvector_of_box(arg)
if vecop is not last and vecop is not None:
vectors.append((pos, vecop))
last = vecop
return vectors
def __repr__(self):
if len(self.operations) == 0:
return "Pack(empty)"
packs = self.operations[0].op.getopname() + '[' + ','.join(['%2d' % (o.opidx) for o in self.operations]) + ']'
if self.operations[0].op.getdescr():
packs += 'descr=' + str(self.operations[0].op.getdescr())
return "Pack(%dx %s)" % (self.numops(), packs)
def is_accumulating(self):
return False
def clone(self, oplist):
return Pack(oplist)
class Pair(Pack):
""" A special Pack object with only two statements. """
def __init__(self, left, right):
assert isinstance(left, Node)
assert isinstance(right, Node)
Pack.__init__(self, [left, right])
def __eq__(self, other):
if isinstance(other, Pair):
return self.left is other.left and \
self.right is other.right
class AccumPack(Pack):
SUPPORTED = { rop.FLOAT_ADD: '+',
rop.INT_ADD: '+',
rop.FLOAT_MUL: '*',
}
def __init__(self, nodes, operator, position):
Pack.__init__(self, nodes)
self.operator = operator
self.position = position
def getdatatype(self):
accum = self.leftmost().getarg(self.position)
vecinfo = forwarded_vecinfo(accum)
return vecinfo.datatype
def getbytesize(self):
accum = self.leftmost().getarg(self.position)
vecinfo = forwarded_vecinfo(accum)
return vecinfo.bytesize
def getleftmostseed(self):
return self.leftmost().getarg(self.position)
def getseeds(self):
""" The accumulatoriable holding the seed value """
return [op.getoperation().getarg(self.position) for op in self.operations]
def reduce_init(self):
if self.operator == '*':
return 1
return 0
def is_accumulating(self):
return True
def clone(self, oplist):
return AccumPack(oplist, self.operator, self.position)
|