openpower/sv/estimate-compression.py

   1 #! /bin/env python3
   2 # see https://bugs.libre-soc.org/show_bug.cgi?id=532
   3
   4 # Estimate ppc code compression with Libre-SOC encoding attempt v2.
   5
   6
   7 # This script is free software; you can redistribute it and/or modify
   8 # it under the terms of the GNU General Public License as published by
   9 # the Free Software Foundation; either version 3, or (at your option)
  10 # any later version.
  11
  12 # This script is distributed in the hope that it will be useful, but
  13 # WITHOUT ANY WARRANTY; without even the implied warranty of
  14 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15 # General Public License for more details.
  16
  17 # You should have received a copy of the GNU General Public License
  18 # along with this script; see the file COPYING3.  If not see
  19 # <http://www.gnu.org/licenses/>.
  20
  21 # Skeleton originally by Alexandre Oliva <oliva@gnu.org>.
  22
  23
  24 # Feed this script the output of objdump -M raw --no-show-raw-insn ppc-prog
  25
  26 # It will look for insns that can be represented in compressed mode,
  27 # according to the encoding rules in the copcond dictionary below.
  28 # Nothing is assumed as to the actual bit-encoding of the insns, this
  29 # is just to experiment with insn selection and get a quick feedback
  30 # loop for the encoding options in compressed mode.
  31
  32 # In this script, the computations of encoding modes and transitions
  33 # are tuned for the simpler model that uses 1-byte nops for
  34 # transitions in and out of compressed mode, placing compressed-mode
  35 # insns at odd addresses.  At (visible) entry points, mode is forced
  36 # to return to uncompressed mode.
  37
  38 # The entire code stream is printed, without any attempt to modify the
  39 # addresses that go along with or in them; we only insert markers for
  40 # the transition points, and for the compressed instructions.
  41
  42 # The really useful information is printed at the end: a summary of
  43 # transition and compressed-insn counts, and the achieved compression
  44 # rate.
  45
  46 import sys
  47 import re
  48
  49 insn = re.compile('\s+(?P<addr>[0-9a-f]+):\s+(?P<opcode>[^ ]+) *(?P<operands>.*)')
  50
  51 opkind = re.compile('(?P<reg>(?P<regkind>[cf]?r)(?P<regnum>[0-9]+))|(?P<immediate>-?[0-9]+)|(?P<branch>[0-9a-f]+)(?: <.*>)?|(?P<offset>-?[0-9]+)\((?P<basereg>r[0-9]+)\)')
  52
  53 def mapop(op):
  54     match = opkind.fullmatch(op)
  55
  56     if match is None:
  57         op = ('other', op)
  58     elif match['reg'] is not None:
  59         op = (match['regkind'], int(match['regnum']))
  60     elif match['immediate'] is not None:
  61         op = ('imm', int (op).bit_length ())
  62     elif match['branch'] is not None:
  63         op = ('pcoff', (int (match['branch'], 16)
  64                         - int (addr, 16)).bit_length ())
  65     elif match['offset'] is not None:
  66         op = ('ofst', mapop(match['offset']), mapop(match['basereg']))
  67     else:
  68         raise "unrecognized operand kind"
  69
  70     return op
  71
  72 def opclass(mop):
  73     return mop[0]
  74 def regno(mop):
  75     if mop[0] in { 'r', 'fr', 'cr' }:
  76         return mop[1]
  77     else:
  78         raise "operand is not a register"
  79
  80 def immbits(mop):
  81     if mop[0] is 'imm':
  82         return mop[1]
  83     else:
  84         raise "operand is not an immediate"
  85
  86 # Following are predicates to be used in copcond, to tell the mode in
  87 # which opcode with ops as operands is to be represented
  88
  89 # Any occurrence of the opcode can be compressed.
  90 def anyops(opcode, ops):
  91     return 1
  92
  93 # Compress iff first and second operands are the same.
  94 def same01(opcode, ops):
  95     if ops[0] == ops[1]:
  96         return 1
  97     else:
  98         return 0
  99
 100 # Registers representable in a made-up 3-bit mapping.
 101 cregs2 = { 1, 2, 3, 4, 5, 6, 7, 31 }
 102
 103 # Return true iff mop is a regular register present in cregs2
 104 def bin2regs3(mop):
 105     return opclass(mop) is 'r' and regno(mop) in cregs2
 106
 107 # Return true iff mop is an immediate of at most 8 bits.
 108 def imm8(mop):
 109     return opclass(mop) is 'imm' and immbits(mop) <= 8
 110
 111 # Compress binary opcodes iff the first two operands (output and first
 112 # input operand) are registers representable in 3 bits in compressed
 113 # mode, and the immediate operand can be represented in 8 bits.
 114 def bin2regs3imm8(opcode, ops):
 115     if bin2regs3(ops[0]) and bin2regs3(ops[1]) and imm8(ops[2]):
 116         return 1
 117     else:
 118         return 0
 119
 120 # Map opcodes that might be compressed to a function that returns the
 121 # mode (index into mode_list below) in which the insn is to be
 122 # represented.  Those not mentioned in copcond are assumed
 123 # Uncomopressed.
 124 copcond = {
 125     # Pretending anything goes, just for demonstration purposes.
 126     'mr': anyops,
 127     'ld': anyops,
 128     'std': anyops,
 129     # Output and first input operand must coincide for these.
 130     'add': same01,
 131     'sub': same01,
 132     # Limiting register and operand range:
 133     'addi': bin2regs3imm8
 134     # Anything else is uncompressed.
 135 }
 136
 137 enter_compressed = 0
 138 leave_compressed = 0
 139 count_compressed = 0
 140 count_uncompressed = 0
 141 current_mode = 0
 142 mode_list = ['Uncompressed', 'Compressed'] # for documentation purposes only
 143
 144 for line in sys.stdin:
 145     if line[-1] is '\n':
 146         line = line[:-1]
 147
 148     match = insn.fullmatch(line)
 149     if match is None:
 150         print(line)
 151         # Switch to uncompressed mode at function boundaries
 152         if current_mode is not 0:
 153             print('<leave compressed mode>')
 154             current_mode = 0
 155             leave_compressed += 1
 156         continue
 157
 158     addr = match['addr']
 159     opcode = match['opcode']
 160     operands = match['operands']
 161
 162     if opcode in copcond:
 163         this_mode = copcond[opcode](opcode,
 164                                     [mapop(op) for op in operands.split(',')])
 165     else:
 166         this_mode = 0
 167
 168     if this_mode is 1:
 169         if current_mode is not 1:
 170             print('\t\tcin.nop')
 171             current_mode = 1
 172             enter_compressed += 1
 173         print(line + ' (c)')
 174         count_compressed += 1
 175     else:
 176         if current_mode is not 0:
 177             print('\t\tcout.nop')
 178             current_mode = 0
 179             leave_compressed += 1
 180         print(line)
 181         count_uncompressed += 1
 182
 183 transition_bytes = 1 * enter_compressed + 1 * leave_compressed
 184 compressed_bytes = 2 * count_compressed
 185 uncompressed_bytes = 4 * count_uncompressed
 186 total_bytes = transition_bytes + compressed_bytes + uncompressed_bytes
 187 original_bytes = 2 * compressed_bytes + uncompressed_bytes
 188
 189 print()
 190 print('Summary')
 191 print('Compressed instructions: %i' % count_compressed)
 192 print('Uncompressed instructions: %i' % count_uncompressed)
 193 print('Transitions into compressed mode: %i' % enter_compressed)
 194 print('Transitions out of compressed mode: %i' % leave_compressed)
 195 print('Compressed size estimate: %i' % total_bytes)
 196 print('Original size: %i' % original_bytes)
 197 print('Compressed/original ratio: %f' % (total_bytes / original_bytes))