from soc.decoder.power_enums import (spr_dict, spr_byname, XER_bits,
insns, MicrOp, In1Sel, In2Sel, In3Sel,
OutSel, CROutSel)
+
+from soc.decoder.power_enums import SVPtype
+
from soc.decoder.helpers import exts, gtu, ltu, undefined
from soc.consts import PIb, MSRb # big-endian (PowerISA versions)
from soc.decoder.power_svp64 import SVP64RM, decode_extra
+from soc.decoder.isa.radixmmu import RADIX
+from soc.decoder.isa.mem import Mem, swap_order
+
from collections import namedtuple
import math
import sys
'VRSAVE': 256}
-def swap_order(x, nbytes):
- x = x.to_bytes(nbytes, byteorder='little')
- x = int.from_bytes(x, byteorder='big', signed=False)
- return x
-
-
REG_SORT_ORDER = {
# TODO (lkcl): adjust other registers that should be in a particular order
# probably CA, CA32, and CR
"CA": 0,
"CA32": 0,
"MSR": 0,
+ "SVSTATE": 0,
"overflow": 1,
}
return retval
-"""
- Get Root Page
-
- //Accessing 2nd double word of partition table (pate1)
- //Ref: Power ISA Manual v3.0B, Book-III, section 5.7.6.1
- // PTCR Layout
- // ====================================================
- // -----------------------------------------------
- // | /// | PATB | /// | PATS |
- // -----------------------------------------------
- // 0 4 51 52 58 59 63
- // PATB[4:51] holds the base address of the Partition Table,
- // right shifted by 12 bits.
- // This is because the address of the Partition base is
- // 4k aligned. Hence, the lower 12bits, which are always
- // 0 are ommitted from the PTCR.
- //
- // Thus, The Partition Table Base is obtained by (PATB << 12)
- //
- // PATS represents the partition table size right-shifted by 12 bits.
- // The minimal size of the partition table is 4k.
- // Thus partition table size = (1 << PATS + 12).
- //
- // Partition Table
- // ====================================================
- // 0 PATE0 63 PATE1 127
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- // | | | <-- effLPID
- // |----------------------|----------------------|
- // .
- // .
- // .
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- //
- // The effective LPID forms the index into the Partition Table.
- //
- // Each entry in the partition table contains 2 double words, PATE0, PATE1,
- // corresponding to that partition.
- //
- // In case of Radix, The structure of PATE0 and PATE1 is as follows.
- //
- // PATE0 Layout
- // -----------------------------------------------
- // |1|RTS1|/| RPDB | RTS2 | RPDS |
- // -----------------------------------------------
- // 0 1 2 3 4 55 56 58 59 63
- //
- // HR[0] : For Radix Page table, first bit should be 1.
- // RTS1[1:2] : Gives one fragment of the Radix treesize
- // RTS2[56:58] : Gives the second fragment of the Radix Tree size.
- // RTS = (RTS1 << 3 + RTS2) + 31.
- //
- // RPDB[4:55] = Root Page Directory Base.
- // RPDS = Logarithm of Root Page Directory Size right shifted by 3.
- // Thus, Root page directory size = 1 << (RPDS + 3).
- // Note: RPDS >= 5.
- //
- // PATE1 Layout
- // -----------------------------------------------
- // |///| PRTB | // | PRTS |
- // -----------------------------------------------
- // 0 3 4 51 52 58 59 63
- //
- // PRTB[4:51] = Process Table Base. This is aligned to size.
- // PRTS[59: 63] = Process Table Size right shifted by 12.
- // Minimal size of the process table is 4k.
- // Process Table Size = (1 << PRTS + 12).
- // Note: PRTS <= 24.
- //
- // Computing the size aligned Process Table Base:
- // table_base = (PRTB & ~((1 << PRTS) - 1)) << 12
- // Thus, the lower 12+PRTS bits of table_base will
- // be zero.
-
-
- //Ref: Power ISA Manual v3.0B, Book-III, section 5.7.6.2
- //
- // Process Table
- // ==========================
- // 0 PRTE0 63 PRTE1 127
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- // | | | <-- effPID
- // |----------------------|----------------------|
- // .
- // .
- // .
- // |----------------------|----------------------|
- // | | |
- // |----------------------|----------------------|
- //
- // The effective Process id (PID) forms the index into the Process Table.
- //
- // Each entry in the partition table contains 2 double words, PRTE0, PRTE1,
- // corresponding to that process
- //
- // In case of Radix, The structure of PRTE0 and PRTE1 is as follows.
- //
- // PRTE0 Layout
- // -----------------------------------------------
- // |/|RTS1|/| RPDB | RTS2 | RPDS |
- // -----------------------------------------------
- // 0 1 2 3 4 55 56 58 59 63
- //
- // RTS1[1:2] : Gives one fragment of the Radix treesize
- // RTS2[56:58] : Gives the second fragment of the Radix Tree size.
- // RTS = (RTS1 << 3 + RTS2) << 31,
- // since minimal Radix Tree size is 4G.
- //
- // RPDB = Root Page Directory Base.
- // RPDS = Root Page Directory Size right shifted by 3.
- // Thus, Root page directory size = RPDS << 3.
- // Note: RPDS >= 5.
- //
- // PRTE1 Layout
- // -----------------------------------------------
- // | /// |
- // -----------------------------------------------
- // 0 63
- // All bits are reserved.
-
-
-"""
-
-# see qemu/target/ppc/mmu-radix64.c for reference
-class RADIX:
- def __init__(self, mem, caller):
- self.mem = mem
- self.caller = caller
-
- # cached page table stuff
- self.pgtbl0 = 0
- self.pt0_valid = False
- self.pgtbl3 = 0
- self.pt3_valid = False
-
- def ld(self, address, width=8, swap=True, check_in_mem=False):
- print("RADIX: ld from addr 0x{:x} width {:d}".format(address, width))
-
- pte = self._walk_tree()
- # use pte to caclculate phys address
- #mem.ld(address,width,swap,check_in_mem)
-
- # TODO implement
- # def st(self, addr, v, width=8, swap=True):
- # def memassign(self, addr, sz, val):
- def _next_level(self):
- return True
- ## DSISR_R_BADCONFIG
- ## read_entry
- ## DSISR_NOPTE
- ## Prepare for next iteration
-
- def _walk_tree(self):
- """walk tree
-
- // vaddr 64 Bit
- // vaddr |-----------------------------------------------------|
- // | Unused | Used |
- // |-----------|-----------------------------------------|
- // | 0000000 | usefulBits = X bits (typically 52) |
- // |-----------|-----------------------------------------|
- // | |<--Cursize---->| |
- // | | Index | |
- // | | into Page | |
- // | | Directory | |
- // |-----------------------------------------------------|
- // | |
- // V |
- // PDE |---------------------------| |
- // |V|L|//| NLB |///|NLS| |
- // |---------------------------| |
- // PDE = Page Directory Entry |
- // [0] = V = Valid Bit |
- // [1] = L = Leaf bit. If 0, then |
- // [4:55] = NLB = Next Level Base |
- // right shifted by 8 |
- // [59:63] = NLS = Next Level Size |
- // | NLS >= 5 |
- // | V
- // | |--------------------------|
- // | | usfulBits = X-Cursize |
- // | |--------------------------|
- // |---------------------><--NLS-->| |
- // | Index | |
- // | into | |
- // | PDE | |
- // |--------------------------|
- // |
- // If the next PDE obtained by |
- // (NLB << 8 + 8 * index) is a |
- // nonleaf, then repeat the above. |
- // |
- // If the next PDE is a leaf, |
- // then Leaf PDE structure is as |
- // follows |
- // |
- // |
- // Leaf PDE |
- // |------------------------------| |----------------|
- // |V|L|sw|//|RPN|sw|R|C|/|ATT|EAA| | usefulBits |
- // |------------------------------| |----------------|
- // [0] = V = Valid Bit |
- // [1] = L = Leaf Bit = 1 if leaf |
- // PDE |
- // [2] = Sw = Sw bit 0. |
- // [7:51] = RPN = Real Page Number, V
- // real_page = RPN << 12 -------------> Logical OR
- // [52:54] = Sw Bits 1:3 |
- // [55] = R = Reference |
- // [56] = C = Change V
- // [58:59] = Att = Physical Address
- // 0b00 = Normal Memory
- // 0b01 = SAO
- // 0b10 = Non Idenmpotent
- // 0b11 = Tolerant I/O
- // [60:63] = Encoded Access
- // Authority
- //
- """
- # walk tree starts on prtbl
- while True:
- ret = self._next_level()
- if ret: return ret
-
- def _segment_check(self):
- """checks segment valid
- mbits := '0' & r.mask_size;
- v.shift := r.shift + (31 - 12) - mbits;
- nonzero := or(r.addr(61 downto 31) and not finalmask(30 downto 0));
- if r.addr(63) /= r.addr(62) or nonzero = '1' then
- v.state := RADIX_FINISH;
- v.segerror := '1';
- elsif mbits < 5 or mbits > 16 or mbits > (r.shift + (31 - 12)) then
- v.state := RADIX_FINISH;
- v.badtree := '1';
- else
- v.state := RADIX_LOOKUP;
- """
-
- def _check_perms(self):
- """check page permissions
- -- test leaf bit
- if data(62) = '1' then
- -- check permissions and RC bits
- perm_ok := '0';
- if r.priv = '1' or data(3) = '0' then
- if r.iside = '0' then
- perm_ok := data(1) or (data(2) and not r.store);
- else
- -- no IAMR, so no KUEP support for now
- -- deny execute permission if cache inhibited
- perm_ok := data(0) and not data(5);
- end if;
- end if;
- rc_ok := data(8) and (data(7) or not r.store);
- if perm_ok = '1' and rc_ok = '1' then
- v.state := RADIX_LOAD_TLB;
- else
- v.state := RADIX_FINISH;
- v.perm_err := not perm_ok;
- -- permission error takes precedence over RC error
- v.rc_error := perm_ok;
- end if;
- """
-
-
-class Mem:
-
- def __init__(self, row_bytes=8, initial_mem=None):
- self.mem = {}
- self.bytes_per_word = row_bytes
- self.word_log2 = math.ceil(math.log2(row_bytes))
- print("Sim-Mem", initial_mem, self.bytes_per_word, self.word_log2)
- if not initial_mem:
- return
-
- # different types of memory data structures recognised (for convenience)
- if isinstance(initial_mem, list):
- initial_mem = (0, initial_mem)
- if isinstance(initial_mem, tuple):
- startaddr, mem = initial_mem
- initial_mem = {}
- for i, val in enumerate(mem):
- initial_mem[startaddr + row_bytes*i] = (val, row_bytes)
-
- for addr, (val, width) in initial_mem.items():
- #val = swap_order(val, width)
- self.st(addr, val, width, swap=False)
-
- def _get_shifter_mask(self, wid, remainder):
- shifter = ((self.bytes_per_word - wid) - remainder) * \
- 8 # bits per byte
- # XXX https://bugs.libre-soc.org/show_bug.cgi?id=377
- # BE/LE mode?
- shifter = remainder * 8
- mask = (1 << (wid * 8)) - 1
- print("width,rem,shift,mask", wid, remainder, hex(shifter), hex(mask))
- return shifter, mask
-
- # TODO: Implement ld/st of lesser width
- def ld(self, address, width=8, swap=True, check_in_mem=False):
- print("ld from addr 0x{:x} width {:d}".format(address, width))
- remainder = address & (self.bytes_per_word - 1)
- address = address >> self.word_log2
- assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
- if address in self.mem:
- val = self.mem[address]
- elif check_in_mem:
- return None
- else:
- val = 0
- print("mem @ 0x{:x} rem {:d} : 0x{:x}".format(address, remainder, val))
-
- if width != self.bytes_per_word:
- shifter, mask = self._get_shifter_mask(width, remainder)
- print("masking", hex(val), hex(mask << shifter), shifter)
- val = val & (mask << shifter)
- val >>= shifter
- if swap:
- val = swap_order(val, width)
- print("Read 0x{:x} from addr 0x{:x}".format(val, address))
- return val
-
- def st(self, addr, v, width=8, swap=True):
- staddr = addr
- remainder = addr & (self.bytes_per_word - 1)
- addr = addr >> self.word_log2
- print("Writing 0x{:x} to ST 0x{:x} "
- "memaddr 0x{:x}/{:x}".format(v, staddr, addr, remainder, swap))
- assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
- if swap:
- v = swap_order(v, width)
- if width != self.bytes_per_word:
- if addr in self.mem:
- val = self.mem[addr]
- else:
- val = 0
- shifter, mask = self._get_shifter_mask(width, remainder)
- val &= ~(mask << shifter)
- val |= v << shifter
- self.mem[addr] = val
- else:
- self.mem[addr] = v
- print("mem @ 0x{:x}: 0x{:x}".format(addr, self.mem[addr]))
-
- def __call__(self, addr, sz):
- val = self.ld(addr.value, sz, swap=False)
- print("memread", addr, sz, val)
- return SelectableInt(val, sz*8)
-
- def memassign(self, addr, sz, val):
- print("memassign", addr, sz, val)
- self.st(addr.value, val.value, sz, swap=False)
-
class GPR(dict):
def __init__(self, decoder, isacaller, svstate, regfile):
def getz(self, rnum):
# rnum = rnum.value # only SelectableInt allowed
- print("GPR getzero", rnum)
+ print("GPR getzero?", rnum)
if rnum == 0:
return SelectableInt(0, 64)
return self[rnum]
in1_isvec = yield dec2.in1_isvec
in2_isvec = yield dec2.in2_isvec
in3_isvec = yield dec2.in3_isvec
- print ("get_pdecode_idx", in1_sel, In1Sel.RA.value, in1, in1_isvec)
+ print ("get_pdecode_idx_in in1", name, in1_sel, In1Sel.RA.value,
+ in1, in1_isvec)
+ print ("get_pdecode_idx_in in2", name, in2_sel, In2Sel.RB.value,
+ in2, in2_isvec)
+ print ("get_pdecode_idx_in in3", name, in3_sel, In3Sel.RS.value,
+ in3, in3_isvec)
# identify which regnames map to in1/2/3
if name == 'RA':
if (in1_sel == In1Sel.RA.value or
# get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
out = yield dec2.e.write_reg.data
o_isvec = yield dec2.o_isvec
- print ("get_pdecode_idx_out", out_sel, OutSel.RA.value, out, o_isvec)
# identify which regnames map to out / o2
if name == 'RA':
+ print ("get_pdecode_idx_out", out_sel, OutSel.RA.value, out, o_isvec)
if out_sel == OutSel.RA.value:
return out, o_isvec
elif name == 'RT':
+ print ("get_pdecode_idx_out", out_sel, OutSel.RT.value,
+ OutSel.RT_OR_ZERO.value, out, o_isvec)
if out_sel == OutSel.RT.value:
return out, o_isvec
print ("get_pdecode_idx_out not found", name)
initial_insns=None, respect_pc=False,
disassembly=None,
initial_pc=0,
- bigendian=False):
+ bigendian=False,
+ mmu=False):
self.bigendian = bigendian
self.halted = False
# set up registers, instruction memory, data memory, PC, SPRs, MSR
self.svp64rm = SVP64RM()
+ if initial_svstate is None:
+ initial_svstate = 0
if isinstance(initial_svstate, int):
initial_svstate = SVP64State(initial_svstate)
self.svstate = initial_svstate
self.gpr = GPR(decoder2, self, self.svstate, regfile)
+ self.spr = SPR(decoder2, initial_sprs) # initialise SPRs before MMU
self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
+ if mmu:
+ self.mem = RADIX(self.mem, self)
self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
self.pc = PC()
- self.spr = SPR(decoder2, initial_sprs)
self.msr = SelectableInt(initial_msr, 64) # underlying reg
# TODO, needed here:
'memassign': self.memassign,
'NIA': self.pc.NIA,
'CIA': self.pc.CIA,
+ 'SVSTATE': self.svstate.spr,
'CR': self.cr,
'MSR': self.msr,
'undefined': undefined,
yield self.dec2.dec.bigendian.eq(self.bigendian)
yield self.dec2.state.msr.eq(self.msr.value)
yield self.dec2.state.pc.eq(pc)
- yield self.dec2.state.svstate.eq(self.svstate.spr.value)
+ if self.svstate is not None:
+ yield self.dec2.state.svstate.eq(self.svstate.spr.value)
# SVP64. first, check if the opcode is EXT001, and SVP64 id bits set
yield Settle()
pfx.insn[7].value == 0b1 and
pfx.insn[9].value == 0b1)
self.pc.update_nia(self.is_svp64_mode)
+ self.namespace['NIA'] = self.pc.NIA
+ self.namespace['SVSTATE'] = self.svstate.spr
if not self.is_svp64_mode:
return
if name not in ['mtcrf', 'mtocrf']:
illegal = name != asmop
+ # sigh deal with setvl not being supported by binutils (.long)
+ if asmop.startswith('setvl'):
+ illegal = False
+ name = 'setvl'
+
if illegal:
print("illegal", name, asmop)
self.TRAP(0x700, PIb.ILLEG)
dest_cr, src_cr, src_byname, dest_byname = False, False, {}, {}
print ("sv rm", sv_rm, dest_cr, src_cr, src_byname, dest_byname)
- # get SVSTATE srcstep. TODO: dststep (twin predication)
- srcstep = self.svstate.srcstep.asint(msb0=True)
- vl = self.svstate.vl.asint(msb0=True)
- mvl = self.svstate.maxvl.asint(msb0=True)
+ # get SVSTATE VL (oh and print out some debug stuff)
+ if self.is_svp64_mode:
+ vl = self.svstate.vl.asint(msb0=True)
+ srcstep = self.svstate.srcstep.asint(msb0=True)
+ sv_a_nz = yield self.dec2.sv_a_nz
+ in1 = yield self.dec2.e.read_reg1.data
+ print ("SVP64: VL, srcstep, sv_a_nz, in1",
+ vl, srcstep, sv_a_nz, in1)
# VL=0 in SVP64 mode means "do nothing: skip instruction"
if self.is_svp64_mode and vl == 0:
self.pc.update(self.namespace, self.is_svp64_mode)
- print("end of call", self.namespace['CIA'], self.namespace['NIA'])
+ print("SVP64: VL=0, end of call", self.namespace['CIA'],
+ self.namespace['NIA'])
return
# main input registers (RT, RA ...)
# doing this is not part of svp64, it's because output
# registers, to be modified, need to be in the namespace.
regnum, is_vec = yield from get_pdecode_idx_out(self.dec2, name)
- # here's where we go "vector". TODO: zero-testing (RA_IS_ZERO)
- # XXX already done by PowerDecoder2, now
- #if is_vec:
- # regnum += srcstep # TODO, elwidth overrides
# in case getting the register number is needed, _RA, _RB
regname = "_" + name
self.namespace[regname] = regnum
- print('reading reg %s %d' % (name, regnum), is_vec)
+ print('reading reg %s %s' % (name, str(regnum)), is_vec)
reg_val = self.gpr(regnum)
inputs.append(reg_val)
vl = self.svstate.vl.asint(msb0=True)
mvl = self.svstate.maxvl.asint(msb0=True)
srcstep = self.svstate.srcstep.asint(msb0=True)
+ sv_ptype = yield self.dec2.dec.op.SV_Ptype
+ no_out_vec = not (yield self.dec2.no_out_vec)
+ no_in_vec = not (yield self.dec2.no_in_vec)
print (" svstate.vl", vl)
print (" svstate.mvl", mvl)
print (" svstate.srcstep", srcstep)
+ print (" no_out_vec", no_out_vec)
+ print (" no_in_vec", no_in_vec)
+ print (" sv_ptype", sv_ptype, sv_ptype == SVPtype.P2.value)
# check if srcstep needs incrementing by one, stop PC advancing
- # svp64 loop can end early if the dest is scalar
- svp64_dest_vector = not (yield self.dec2.no_out_vec)
- if svp64_dest_vector and srcstep != vl-1:
+ # svp64 loop can end early if the dest is scalar for single-pred
+ # but for 2-pred both src/dest have to be checked.
+ # XXX this might not be true! it may just be LD/ST
+ if sv_ptype == SVPtype.P2.value:
+ svp64_is_vector = (no_out_vec or no_in_vec)
+ else:
+ svp64_is_vector = no_out_vec
+ if svp64_is_vector and srcstep != vl-1:
self.svstate.srcstep += SelectableInt(1, 7)
self.pc.NIA.value = self.pc.CIA.value
self.namespace['NIA'] = self.pc.NIA
+ self.namespace['SVSTATE'] = self.svstate.spr
print("end of sub-pc call", self.namespace['CIA'],
self.namespace['NIA'])
return # DO NOT allow PC to update whilst Sub-PC loop running
print (" svstate.srcstep loop end (PC to update)")
self.pc.update_nia(self.is_svp64_mode)
self.namespace['NIA'] = self.pc.NIA
+ self.namespace['SVSTATE'] = self.svstate.spr
# UPDATE program counter
self.pc.update(self.namespace, self.is_svp64_mode)
- print("end of call", self.namespace['CIA'], self.namespace['NIA'])
+ self.svstate.spr = self.namespace['SVSTATE']
+ print("end of call", self.namespace['CIA'],
+ self.namespace['NIA'],
+ self.namespace['SVSTATE'])
def inject():
result = func(*args, **kwargs)
print("globals after", func_globals['CIA'], func_globals['NIA'])
print("args[0]", args[0].namespace['CIA'],
- args[0].namespace['NIA'])
+ args[0].namespace['NIA'],
+ args[0].namespace['SVSTATE'])
args[0].namespace = func_globals
#exec (func.__code__, func_globals)
return decorator
return variable_injector
+
+
projects / soc.git / blobdiff