Author: Richard Plangger <[email protected]>
Branch: s390x-backend
Changeset: r82818:8ab56b6e79d4
Date: 2016-03-06 18:28 +0100
http://bitbucket.org/pypy/pypy/changeset/8ab56b6e79d4/
Log: (untested) fix zero array for ppc
diff --git a/rpython/jit/backend/ppc/opassembler.py
b/rpython/jit/backend/ppc/opassembler.py
--- a/rpython/jit/backend/ppc/opassembler.py
+++ b/rpython/jit/backend/ppc/opassembler.py
@@ -910,79 +910,78 @@
else: self.mc.std(a, b, c)
def emit_zero_array(self, op, arglocs, regalloc):
- base_loc, startindex_loc, length_loc, ofs_loc, itemsize_loc = arglocs
+ base_loc, startindex_loc, length_loc, ofs_loc = arglocs
- # assume that an array where an item size is N:
- # * if N is even, then all items are aligned to a multiple of 2
- # * if N % 4 == 0, then all items are aligned to a multiple of 4
- # * if N % 8 == 0, then all items are aligned to a multiple of 8
- itemsize = itemsize_loc.getint()
- if itemsize & 1: stepsize = 1
- elif itemsize & 2: stepsize = 2
- elif (itemsize & 4) or IS_PPC_32: stepsize = 4
- else: stepsize = WORD
+ stepsize = 8
+ if IS_PPC_32:
+ stepsize = 4
- repeat_factor = itemsize // stepsize
- if repeat_factor != 1:
- # This is only for itemsize not in (1, 2, 4, WORD).
- # Include the repeat_factor inside length_loc if it is a constant
- if length_loc.is_imm():
- length_loc = imm(length_loc.value * repeat_factor)
- repeat_factor = 1 # included
+ if length_loc.is_imm():
+ if length_loc.value <= 0:
+ return # nothing to do
- unroll = -1
+ self.mc.addi(r.SCRATCH2.value, startindex_loc.value, ofs_loc.getint())
+ ofs_loc = r.SCRATCH2
+ # ofs_loc is now the startindex in bytes + the array offset
+
if length_loc.is_imm():
- if length_loc.value <= 8:
- unroll = length_loc.value
- if unroll <= 0:
- return # nothing to do
-
- ofs_loc = self._apply_scale(ofs_loc, startindex_loc, itemsize_loc)
- ofs_loc = self._copy_in_scratch2(ofs_loc)
-
- if unroll > 0:
- assert repeat_factor == 1
- self.mc.li(r.SCRATCH.value, 0)
- self.eza_stXux(r.SCRATCH.value, ofs_loc.value, base_loc.value,
- itemsize)
- for i in range(1, unroll):
- self.eza_stX(r.SCRATCH.value, ofs_loc.value, i * stepsize,
- itemsize)
-
+ self.mc.load_imm(r.SCRATCH, length_loc.value)
+ length_loc = r.SCRATCH
+ jz_location = -1
else:
- if length_loc.is_imm():
- self.mc.load_imm(r.SCRATCH, length_loc.value)
- length_loc = r.SCRATCH
- jz_location = -1
- assert repeat_factor == 1
- else:
- self.mc.cmp_op(0, length_loc.value, 0, imm=True)
- jz_location = self.mc.currpos()
- self.mc.trap()
- length_loc = self._multiply_by_constant(length_loc,
- repeat_factor,
- r.SCRATCH)
- self.mc.mtctr(length_loc.value)
- self.mc.li(r.SCRATCH.value, 0)
-
- self.eza_stXux(r.SCRATCH.value, ofs_loc.value, base_loc.value,
- itemsize)
- bdz_location = self.mc.currpos()
+ # jump to end if length is less than stepsize
+ self.mc.cmp_op(0, length_loc.value, stepsize, imm=True)
+ jz_location = self.mc.currpos()
self.mc.trap()
- loop_location = self.mc.currpos()
- self.eza_stXu(r.SCRATCH.value, ofs_loc.value, stepsize,
- itemsize)
- self.mc.bdnz(loop_location - self.mc.currpos())
+ self.mc.li(r.SCRATCH.value, 0)
- pmc = OverwritingBuilder(self.mc, bdz_location, 1)
- pmc.bdz(self.mc.currpos() - bdz_location)
+ # NOTE the following assumes that bytes have been passed to both
startindex
+ # and length. Thus we zero 4/8 bytes in a loop in 1) and every
remaining
+ # byte is zeroed in another loop in 2)
+
+ # first store of case 1)
+ self.eza_stXux(r.SCRATCH.value, ofs_loc.value, base_loc.value,
stepsize)
+ self.mc.subi(length_loc.value, length_loc.value, stepsize)
+ self.mc.cmp_op(0, length_loc.value, stepsize, imm=True)
+ lt_location = self.mc.currpos()
+ self.mc.trap() # jump over the loop if we are already done with 1)
+
+ # 1) The next loop copies WORDS into the memory chunk starting at
startindex
+ # ending at startindex + length. These are bytes
+ loop_location = self.mc.currpos()
+ self.eza_stXu(r.SCRATCH.value, ofs_loc.value, stepsize, stepsize)
+ self.mc.subi(length_loc.value, length_loc.value, stepsize)
+ self.mc.cmp_op(0, length_loc.value, stepsize, imm=True)
+ self.mc.bge(loop_location - self.mc.currpos())
+
+ pmc = OverwritingBuilder(self.mc, lt_location, 1)
+ pmc.blt(self.mc.currpos() - lt_location)
+ pmc.overwrite()
+
+ if jz_location != -1:
+ pmc = OverwritingBuilder(self.mc, jz_location, 1)
+ pmc.ble(self.mc.currpos() - jz_location) # !GT
pmc.overwrite()
- if jz_location != -1:
- pmc = OverwritingBuilder(self.mc, jz_location, 1)
- pmc.ble(self.mc.currpos() - jz_location) # !GT
- pmc.overwrite()
+ # 2) There might be some bytes left to be written.
+ # following scenario: length_loc == 3 bytes, stepsize == 4!
+ # need to write the last bytes.
+ self.mc.cmp_op(0, length_loc.value, 0, imm=True)
+ jle_location = self.mc.curpos()
+ self.mc.trap()
+
+ self.mc.mtctr(length_loc.value)
+
+ loop_position = self.mc.curpos()
+ self.eza_stXu(r.SCRATCH.value, ofs_loc.value, 1, 1)
+ self.mc.bdnz(self.mc.currpos() - loop_location)
+
+ pmc = OverwritingBuilder(self.mc, jle_location, 1)
+ pmc.ble(self.mc.currpos() - jle_location) # !GT
+ pmc.overwrite()
+
+
class StrOpAssembler(object):
diff --git a/rpython/jit/backend/ppc/regalloc.py
b/rpython/jit/backend/ppc/regalloc.py
--- a/rpython/jit/backend/ppc/regalloc.py
+++ b/rpython/jit/backend/ppc/regalloc.py
@@ -955,12 +955,19 @@
return arglocs
def prepare_zero_array(self, op):
- itemsize, ofs, _ = unpack_arraydescr(op.getdescr())
+ _, ofs, _ = unpack_arraydescr(op.getdescr())
base_loc = self.ensure_reg(op.getarg(0))
startindex_loc = self.ensure_reg_or_16bit_imm(op.getarg(1))
length_loc = self.ensure_reg_or_16bit_imm(op.getarg(2))
+ # startindex and length are bytes, not array items anymore.
+ # rewrite already applied the scale!
+ startindex_scale_box = op.getarg(3)
+ assert startindex_scale_box.getint() == 1
+ length_scale_box = op.getarg(4)
+ assert length_scale_box.getint() == 1
+ #
ofs_loc = self.ensure_reg_or_16bit_imm(ConstInt(ofs))
- return [base_loc, startindex_loc, length_loc, ofs_loc, imm(itemsize)]
+ return [base_loc, startindex_loc, length_loc, ofs_loc]
def prepare_cond_call(self, op):
self.load_condition_into_cc(op.getarg(0))
_______________________________________________
pypy-commit mailing list
[email protected]
https://mail.python.org/mailman/listinfo/pypy-commit
