[pypy-commit] pypy ppc-jit-backend: merge
hager
noreply at buildbot.pypy.org
Mon Feb 13 18:38:56 CET 2012
Author: hager <sven.hager at uni-duesseldorf.de>
Branch: ppc-jit-backend
Changeset: r52429:68416fec227f
Date: 2012-02-13 18:38 +0100
http://bitbucket.org/pypy/pypy/changeset/68416fec227f/
Log: merge
diff --git a/pypy/jit/backend/ppc/ppc_assembler.py b/pypy/jit/backend/ppc/ppc_assembler.py
--- a/pypy/jit/backend/ppc/ppc_assembler.py
+++ b/pypy/jit/backend/ppc/ppc_assembler.py
@@ -7,7 +7,7 @@
from pypy.jit.backend.ppc.assembler import Assembler
from pypy.jit.backend.ppc.opassembler import OpAssembler
from pypy.jit.backend.ppc.symbol_lookup import lookup
-from pypy.jit.backend.ppc.codebuilder import PPCBuilder
+from pypy.jit.backend.ppc.codebuilder import PPCBuilder, OverwritingBuilder
from pypy.jit.backend.ppc.jump import remap_frame_layout
from pypy.jit.backend.ppc.arch import (IS_PPC_32, IS_PPC_64, WORD,
NONVOLATILES, MAX_REG_PARAMS,
@@ -20,6 +20,7 @@
decode32, decode64,
count_reg_args,
Saved_Volatiles)
+from pypy.jit.backend.ppc.helper.regalloc import _check_imm_arg
import pypy.jit.backend.ppc.register as r
import pypy.jit.backend.ppc.condition as c
from pypy.jit.metainterp.history import (Const, ConstPtr, JitCellToken,
@@ -279,6 +280,28 @@
locs.append(loc)
return locs
+ def _build_malloc_slowpath(self):
+ mc = PPCBuilder()
+ with Saved_Volatiles(mc):
+ # Values to compute size stored in r3 and r4
+ mc.subf(r.r3.value, r.r3.value, r.r4.value)
+ addr = self.cpu.gc_ll_descr.get_malloc_slowpath_addr()
+ mc.call(addr)
+
+ mc.cmp_op(0, r.r3.value, 0, imm=True)
+ jmp_pos = mc.currpos()
+ mc.nop()
+ nursery_free_adr = self.cpu.gc_ll_descr.get_nursery_free_addr()
+ mc.load_imm(r.r4, nursery_free_adr)
+ mc.load(r.r4.value, r.r4.value, 0)
+
+ pmc = OverwritingBuilder(mc, jmp_pos, 1)
+ pmc.bc(4, 2, jmp_pos) # jump if the two values are equal
+ pmc.overwrite()
+ mc.b_abs(self.propagate_exception_path)
+ rawstart = mc.materialize(self.cpu.asmmemmgr, [])
+ self.malloc_slowpath = rawstart
+
def _build_propagate_exception_path(self):
if self.cpu.propagate_exception_v < 0:
return
@@ -383,8 +406,8 @@
gc_ll_descr = self.cpu.gc_ll_descr
gc_ll_descr.initialize()
self._build_propagate_exception_path()
- #if gc_ll_descr.get_malloc_slowpath_addr is not None:
- # self._build_malloc_slowpath()
+ if gc_ll_descr.get_malloc_slowpath_addr is not None:
+ self._build_malloc_slowpath()
if gc_ll_descr.gcrootmap and gc_ll_descr.gcrootmap.is_shadow_stack:
self._build_release_gil(gc_ll_descr.gcrootmap)
self.memcpy_addr = self.cpu.cast_ptr_to_int(memcpy_fn)
@@ -892,6 +915,51 @@
else:
self.mc.extsw(resloc.value, resloc.value)
+ def malloc_cond(self, nursery_free_adr, nursery_top_adr, size):
+ assert size & (WORD-1) == 0 # must be correctly aligned
+ size = max(size, self.cpu.gc_ll_descr.minimal_size_in_nursery)
+ size = (size + WORD - 1) & ~(WORD - 1) # round up
+
+ self.mc.load_imm(r.r3, nursery_free_adr)
+ self.mc.load(r.r3.value, r.r3.value, 0)
+
+ if _check_imm_arg(size):
+ self.mc.addi(r.r4.value, r.r3.value, size)
+ else:
+ self.mc.load_imm(r.r4, size)
+ self.mc.add(r.r4.value, r.r3.value, r.r4.value)
+
+ # XXX maybe use an offset from the value nursery_free_addr
+ self.mc.load_imm(r.r3, nursery_top_adr)
+ self.mc.load(r.r3.value, r.r3.value, 0)
+
+ self.mc.cmp_op(0, r.r4.value, r.r3.value, signed=False)
+
+ fast_jmp_pos = self.mc.currpos()
+ self.mc.nop()
+
+ # XXX update
+ # See comments in _build_malloc_slowpath for the
+ # details of the two helper functions that we are calling below.
+ # First, we need to call two of them and not just one because we
+ # need to have a mark_gc_roots() in between. Then the calling
+ # convention of slowpath_addr{1,2} are tweaked a lot to allow
+ # the code here to be just two CALLs: slowpath_addr1 gets the
+ # size of the object to allocate from (EDX-EAX) and returns the
+ # result in EAX; self.malloc_slowpath additionally returns in EDX a
+ # copy of heap(nursery_free_adr), so that the final MOV below is
+ # a no-op.
+ self.mark_gc_roots(self.write_new_force_index(),
+ use_copy_area=True)
+ self.mc.call(self.malloc_slowpath)
+
+ offset = self.mc.currpos() - fast_jmp_pos
+ pmc = OverwritingBuilder(self.mc, fast_jmp_pos, 1)
+ pmc.bc(4, 1, offset) # jump if LE (not GT)
+
+ self.mc.load_imm(r.r3, nursery_free_adr)
+ self.mc.store(r.r4.value, r.r3.value, 0)
+
def mark_gc_roots(self, force_index, use_copy_area=False):
if force_index < 0:
return # not needed
diff --git a/pypy/jit/backend/ppc/regalloc.py b/pypy/jit/backend/ppc/regalloc.py
--- a/pypy/jit/backend/ppc/regalloc.py
+++ b/pypy/jit/backend/ppc/regalloc.py
@@ -11,8 +11,9 @@
prepare_binary_int_op,
prepare_binary_int_op_with_imm,
prepare_unary_cmp)
-from pypy.jit.metainterp.history import (INT, REF, FLOAT, Const, ConstInt,
- ConstPtr, Box)
+from pypy.jit.metainterp.history import (Const, ConstInt, ConstFloat, ConstPtr,
+ Box, BoxPtr,
+ INT, REF, FLOAT)
from pypy.jit.metainterp.history import JitCellToken, TargetToken
from pypy.jit.metainterp.resoperation import rop
from pypy.jit.backend.ppc import locations
@@ -507,17 +508,14 @@
gcrootmap = self.cpu.gc_ll_descr.gcrootmap
if gcrootmap:
arglocs = []
- argboxes = []
+ args = op.getarglist()
for i in range(op.numargs()):
- loc, box = self._ensure_value_is_boxed(op.getarg(i), argboxes)
+ loc = self._ensure_value_is_boxed(op.getarg(i), args)
arglocs.append(loc)
- argboxes.append(box)
self.assembler.call_release_gil(gcrootmap, arglocs)
- self.possibly_free_vars(argboxes)
# do the call
faildescr = guard_op.getdescr()
fail_index = self.cpu.get_fail_descr_number(faildescr)
- self.assembler._write_fail_index(fail_index)
args = [imm(rffi.cast(lltype.Signed, op.getarg(0).getint()))]
self.assembler.emit_call(op, args, self, fail_index)
# then reopen the stack
@@ -778,6 +776,43 @@
args = [imm(rffi.cast(lltype.Signed, op.getarg(0).getint()))]
return args
+ def prepare_call_malloc_nursery(self, op):
+ size_box = op.getarg(0)
+ assert isinstance(size_box, ConstInt)
+ size = size_box.getint()
+
+ self.rm.force_allocate_reg(op.result, selected_reg=r.r3)
+ t = TempInt()
+ self.rm.force_allocate_reg(t, selected_reg=r.r4)
+ self.possibly_free_var(op.result)
+ self.possibly_free_var(t)
+
+ gc_ll_descr = self.assembler.cpu.gc_ll_descr
+ self.assembler.malloc_cond(
+ gc_ll_descr.get_nursery_free_addr(),
+ gc_ll_descr.get_nursery_top_addr(),
+ size
+ )
+
+ def get_mark_gc_roots(self, gcrootmap, use_copy_area=False):
+ shape = gcrootmap.get_basic_shape(False)
+ for v, val in self.frame_manager.bindings.items():
+ if (isinstance(v, BoxPtr) and self.rm.stays_alive(v)):
+ assert val.is_stack()
+ gcrootmap.add_frame_offset(shape, val.position * -WORD)
+ for v, reg in self.rm.reg_bindings.items():
+ if reg is r.r3:
+ continue
+ if (isinstance(v, BoxPtr) and self.rm.stays_alive(v)):
+ if use_copy_area:
+ assert reg in self.rm.REGLOC_TO_COPY_AREA_OFS
+ area_offset = self.rm.REGLOC_TO_COPY_AREA_OFS[reg]
+ gcrootmap.add_frame_offset(shape, area_offset)
+ else:
+ assert 0, 'sure??'
+ return gcrootmap.compress_callshape(shape,
+ self.assembler.datablockwrapper)
+
prepare_debug_merge_point = void
prepare_jit_debug = void
@@ -788,11 +823,10 @@
# because it will be needed anyway by the following setfield_gc
# or setarrayitem_gc. It avoids loading it twice from the memory.
arglocs = []
- argboxes = []
+ args = op.getarglist()
for i in range(N):
- loc = self._ensure_value_is_boxed(op.getarg(i), argboxes)
+ loc = self._ensure_value_is_boxed(op.getarg(i), args)
arglocs.append(loc)
- self.rm.possibly_free_vars(argboxes)
return arglocs
prepare_cond_call_gc_wb_array = prepare_cond_call_gc_wb
More information about the pypy-commit
mailing list