[pypy-commit] pypy arm64: bridge basics
fijal
pypy.commits at gmail.com
Wed Mar 6 10:05:14 EST 2019
Author: Maciej Fijalkowski <fijall at gmail.com>
Branch: arm64
Changeset: r96219:a50f3a71ef1d
Date: 2019-03-06 14:49 +0000
http://bitbucket.org/pypy/pypy/changeset/a50f3a71ef1d/
Log: bridge basics
diff --git a/rpython/jit/backend/aarch64/assembler.py b/rpython/jit/backend/aarch64/assembler.py
--- a/rpython/jit/backend/aarch64/assembler.py
+++ b/rpython/jit/backend/aarch64/assembler.py
@@ -117,6 +117,89 @@
return AsmInfo(ops_offset, rawstart + loop_head,
size_excluding_failure_stuff - loop_head)
+ def assemble_bridge(self, logger, faildescr, inputargs, operations,
+ original_loop_token, log):
+ if not we_are_translated():
+ # Arguments should be unique
+ assert len(set(inputargs)) == len(inputargs)
+
+ self.setup(original_loop_token)
+ #self.codemap.inherit_code_from_position(faildescr.adr_jump_offset)
+ descr_number = compute_unique_id(faildescr)
+ if log:
+ operations = self._inject_debugging_code(faildescr, operations,
+ 'b', descr_number)
+
+ assert isinstance(faildescr, AbstractFailDescr)
+
+ arglocs = self.rebuild_faillocs_from_descr(faildescr, inputargs)
+
+ regalloc = Regalloc(assembler=self)
+ allgcrefs = []
+ operations = regalloc.prepare_bridge(inputargs, arglocs,
+ operations,
+ allgcrefs,
+ self.current_clt.frame_info)
+ self.reserve_gcref_table(allgcrefs)
+ startpos = self.mc.get_relative_pos()
+
+ self._check_frame_depth(self.mc, regalloc.get_gcmap())
+
+ bridgestartpos = self.mc.get_relative_pos()
+ frame_depth_no_fixed_size = self._assemble(regalloc, inputargs, operations)
+
+ codeendpos = self.mc.get_relative_pos()
+
+ self.write_pending_failure_recoveries()
+
+ fullsize = self.mc.get_relative_pos()
+ rawstart = self.materialize_loop(original_loop_token)
+
+ self.patch_gcref_table(original_loop_token, rawstart)
+ self.process_pending_guards(rawstart)
+
+ debug_start("jit-backend-addr")
+ debug_print("bridge out of Guard 0x%x has address 0x%x to 0x%x" %
+ (r_uint(descr_number), r_uint(rawstart + startpos),
+ r_uint(rawstart + codeendpos)))
+ debug_print(" gc table: 0x%x" % r_uint(rawstart))
+ debug_print(" jump target: 0x%x" % r_uint(rawstart + startpos))
+ debug_print(" resops: 0x%x" % r_uint(rawstart + bridgestartpos))
+ debug_print(" failures: 0x%x" % r_uint(rawstart + codeendpos))
+ debug_print(" end: 0x%x" % r_uint(rawstart + fullsize))
+ debug_stop("jit-backend-addr")
+
+ # patch the jump from original guard
+ self.patch_trace(faildescr, original_loop_token,
+ rawstart + startpos, regalloc)
+
+ self.patch_stack_checks(frame_depth_no_fixed_size + JITFRAME_FIXED_SIZE,
+ rawstart)
+ if not we_are_translated():
+ if log:
+ self.mc._dump_trace(rawstart, 'bridge.asm')
+
+ ops_offset = self.mc.ops_offset
+ frame_depth = max(self.current_clt.frame_info.jfi_frame_depth,
+ frame_depth_no_fixed_size + JITFRAME_FIXED_SIZE)
+ self.fixup_target_tokens(rawstart)
+ self.update_frame_depth(frame_depth)
+
+ if logger:
+ log = logger.log_trace(jl.MARK_TRACE_ASM, None, self.mc)
+ log.write(inputargs, operations, ops_offset)
+ # log that the already written bridge is stitched to a descr!
+ logger.log_patch_guard(descr_number, rawstart)
+
+ # legacy
+ if logger.logger_ops:
+ logger.logger_ops.log_bridge(inputargs, operations, "rewritten",
+ faildescr, ops_offset=ops_offset)
+
+ self.teardown()
+
+ return AsmInfo(ops_offset, startpos + rawstart, codeendpos - startpos)
+
def setup(self, looptoken):
BaseAssembler.setup(self, looptoken)
assert self.memcpy_addr != 0, 'setup_once() not called?'
@@ -220,6 +303,9 @@
def _check_frame_depth_debug(self, mc):
pass
+ def _check_frame_depth(self, mc, gcmap):
+ pass # XXX
+
def update_frame_depth(self, frame_depth):
baseofs = self.cpu.get_baseofs_of_frame_field()
self.current_clt.frame_info.update_frame_depth(baseofs, frame_depth)
@@ -264,6 +350,9 @@
gcreftracers.append(tracer) # keepalive
self.teardown_gcrefs_list()
+ def patch_stack_checks(self, framedepth, rawstart):
+ pass # XXX
+
def load_from_gc_table(self, regnum, index):
address_in_buffer = index * WORD # at the start of the buffer
p_location = self.mc.get_relative_pos(break_basic_block=False)
@@ -281,6 +370,14 @@
# self.codemap.get_final_bytecode(res, size))
return res
+ def patch_trace(self, faildescr, looptoken, bridge_addr, regalloc):
+ b = InstrBuilder()
+ patch_addr = faildescr.adr_jump_offset
+ assert patch_addr != 0
+ b.BL(bridge_addr)
+ b.copy_to_raw_memory(patch_addr)
+ faildescr.adr_jump_offset = 0
+
def process_pending_guards(self, block_start):
clt = self.current_clt
for tok in self.pending_guards:
diff --git a/rpython/jit/backend/aarch64/codebuilder.py b/rpython/jit/backend/aarch64/codebuilder.py
--- a/rpython/jit/backend/aarch64/codebuilder.py
+++ b/rpython/jit/backend/aarch64/codebuilder.py
@@ -137,7 +137,7 @@
def BL(self, target):
target = rffi.cast(lltype.Signed, target)
- self.gen_load_int(r.ip0.value, target)
+ self.gen_load_int_full(r.ip0.value, target)
self.BR(r.ip0.value)
def BR(self, reg):
@@ -147,6 +147,12 @@
def BRK(self):
self.write32(0b11010100001 << 21)
+ def gen_load_int_full(self, r, value):
+ self.MOVZ_r_u16(r, value & 0xFFFF, 0)
+ self.MOVK_r_u16(r, (value >> 16) & 0xFFFF, 16)
+ self.MOVK_r_u16(r, (value >> 32) & 0xFFFF, 32)
+ self.MOVK_r_u16(r, (value >> 48) & 0xFFFF, 48)
+
def gen_load_int(self, r, value):
"""r is the register number, value is the value to be loaded to the
register"""
diff --git a/rpython/jit/backend/aarch64/opassembler.py b/rpython/jit/backend/aarch64/opassembler.py
--- a/rpython/jit/backend/aarch64/opassembler.py
+++ b/rpython/jit/backend/aarch64/opassembler.py
@@ -120,7 +120,7 @@
if target_token in self.target_tokens_currently_compiling:
self.mc.B_ofs(target)
else:
- self.mc.B(target)
+ self.mc.BL(target)
def emit_op_finish(self, op, arglocs):
base_ofs = self.cpu.get_baseofs_of_frame_field()
diff --git a/rpython/jit/backend/aarch64/regalloc.py b/rpython/jit/backend/aarch64/regalloc.py
--- a/rpython/jit/backend/aarch64/regalloc.py
+++ b/rpython/jit/backend/aarch64/regalloc.py
@@ -1,6 +1,7 @@
from rpython.jit.backend.aarch64 import registers as r
from rpython.jit.backend.aarch64 import locations
+from rpython.jit.backend.aarch64.arch import WORD, JITFRAME_FIXED_SIZE
from rpython.jit.metainterp.history import (Const, ConstInt, ConstFloat,
ConstPtr,
@@ -14,6 +15,8 @@
from rpython.jit.backend.aarch64 import registers as r
from rpython.jit.backend.arm.jump import remap_frame_layout_mixed
from rpython.jit.backend.aarch64.locations import imm
+from rpython.jit.backend.llsupport.gcmap import allocate_gcmap
+
class TempInt(TempVar):
@@ -450,6 +453,64 @@
self.rm._check_invariants()
self.vfprm._check_invariants()
+ def prepare_bridge(self, inputargs, arglocs, operations, allgcrefs,
+ frame_info):
+ operations = self._prepare(inputargs, operations, allgcrefs)
+ self._update_bindings(arglocs, inputargs)
+ return operations
+
+ def _update_bindings(self, locs, inputargs):
+ used = {}
+ i = 0
+ for loc in locs:
+ if loc is None:
+ loc = r.fp
+ arg = inputargs[i]
+ i += 1
+ if loc.is_core_reg():
+ self.rm.reg_bindings[arg] = loc
+ used[loc] = None
+ elif loc.is_vfp_reg():
+ self.vfprm.reg_bindings[arg] = loc
+ used[loc] = None
+ else:
+ assert loc.is_stack()
+ self.frame_manager.bind(arg, loc)
+
+ # XXX combine with x86 code and move to llsupport
+ self.rm.free_regs = []
+ for reg in self.rm.all_regs:
+ if reg not in used:
+ self.rm.free_regs.append(reg)
+ self.vfprm.free_regs = []
+ for reg in self.vfprm.all_regs:
+ if reg not in used:
+ self.vfprm.free_regs.append(reg)
+ # note: we need to make a copy of inputargs because possibly_free_vars
+ # is also used on op args, which is a non-resizable list
+ self.possibly_free_vars(list(inputargs))
+ self.fm.finish_binding()
+ self._check_invariants()
+
+ def get_gcmap(self, forbidden_regs=[], noregs=False):
+ frame_depth = self.fm.get_frame_depth()
+ gcmap = allocate_gcmap(self.assembler,
+ frame_depth, JITFRAME_FIXED_SIZE)
+ for box, loc in self.rm.reg_bindings.iteritems():
+ if loc in forbidden_regs:
+ continue
+ if box.type == REF and self.rm.is_still_alive(box):
+ assert not noregs
+ assert loc.is_core_reg()
+ val = loc.value
+ gcmap[val // WORD // 8] |= r_uint(1) << (val % (WORD * 8))
+ for box, loc in self.fm.bindings.iteritems():
+ if box.type == REF and self.rm.is_still_alive(box):
+ assert loc.is_stack()
+ val = loc.position + JITFRAME_FIXED_SIZE
+ gcmap[val // WORD // 8] |= r_uint(1) << (val % (WORD * 8))
+ return gcmap
+
def get_final_frame_depth(self):
return self.frame_manager.get_frame_depth()
diff --git a/rpython/jit/backend/aarch64/runner.py b/rpython/jit/backend/aarch64/runner.py
--- a/rpython/jit/backend/aarch64/runner.py
+++ b/rpython/jit/backend/aarch64/runner.py
@@ -2,6 +2,7 @@
from rpython.rtyper.lltypesystem import llmemory, lltype
from rpython.jit.backend.aarch64.assembler import AssemblerARM64
from rpython.jit.backend.aarch64 import registers as r
+from rpython.jit.backend.aarch64.regalloc import VFPRegisterManager
from rpython.jit.backend.llsupport.llmodel import AbstractLLCPU
class CPU_ARM64(AbstractLLCPU):
@@ -9,6 +10,9 @@
backend_name = "aarch64"
frame_reg = r.fp
all_reg_indexes = range(len(r.all_regs))
+ gen_regs = r.all_regs
+ float_regs = VFPRegisterManager.all_regs
+
IS_64_BIT = True
@@ -23,6 +27,14 @@
def setup_once(self):
self.assembler.setup_once()
+ def compile_bridge(self, faildescr, inputargs, operations,
+ original_loop_token, log=True, logger=None):
+ clt = original_loop_token.compiled_loop_token
+ clt.compiling_a_bridge()
+ return self.assembler.assemble_bridge(logger, faildescr, inputargs,
+ operations,
+ original_loop_token, log=log)
+
def cast_ptr_to_int(x):
adr = llmemory.cast_ptr_to_adr(x)
return CPU_ARM64.cast_adr_to_int(adr)
More information about the pypy-commit
mailing list