1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
|
#
# Copyright (C) 2020 Collabora, Ltd.
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice (including the next
# paragraph) shall be included in all copies or substantial portions of the
# Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
import sys
import itertools
from bifrost_isa import parse_instructions, opname_to_c, expand_states
from mako.template import Template
instructions = parse_instructions(sys.argv[1], include_unused = True)
# Constructs a reserved mask for a derived to cull impossible encodings
def reserved_mask(derived):
((pos, width), opts) = derived
reserved = [x is None for x in opts]
mask = sum([(y << x) for x, y in enumerate(reserved)])
return (pos, width, mask)
def reserved_masks(op):
masks = [reserved_mask(m) for m in op[2].get("derived", [])]
return [m for m in masks if m[2] != 0]
# To decode instructions, pattern match based on the rules:
#
# 1. Execution unit (FMA or ADD) must line up.
# 2. All exact bits must match.
# 3. No fields should be reserved in a legal encoding.
# 4. Tiebreaker: Longer exact masks (greater unsigned bitwise inverses) win.
#
# To implement, filter the execution unit and check for exact bits in
# descending order of exact mask length. Check for reserved fields per
# candidate and succeed if it matches.
# found.
def decode_op(instructions, is_fma):
# Filter out the desired execution unit
options = [n for n in instructions.keys() if (n[0] == '*') == is_fma]
# Sort by exact masks, descending
MAX_MASK = (1 << (23 if is_fma else 20)) - 1
options.sort(key = lambda n: (MAX_MASK ^ instructions[n][2]["exact"][0]))
# Map to what we need to template
mapped = [(opname_to_c(op), instructions[op][2]["exact"], reserved_masks(instructions[op])) for op in options]
# Generate checks in order
template = """void
bi_disasm_${unit}(FILE *fp, unsigned bits, struct bifrost_regs *srcs, struct bifrost_regs *next_regs, unsigned staging_register, unsigned branch_offset, struct bi_constants *consts, bool last)
{
fputs(" ", fp);
% for (i, (name, (emask, ebits), derived)) in enumerate(options):
% if len(derived) > 0:
${"else " if i > 0 else ""}if (unlikely(((bits & ${hex(emask)}) == ${hex(ebits)})
% for (pos, width, reserved) in derived:
&& !(${hex(reserved)} & (1 << _BITS(bits, ${pos}, ${width})))
% endfor
))
% else:
${"else " if i > 0 else ""}if (unlikely(((bits & ${hex(emask)}) == ${hex(ebits)})))
% endif
bi_disasm_${name}(fp, bits, srcs, next_regs, staging_register, branch_offset, consts, last);
% endfor
else
fprintf(fp, "INSTR_INVALID_ENC ${unit} %X", bits);
fputs("\\n", fp);
}"""
return Template(template).render(options = mapped, unit = "fma" if is_fma else "add")
# Decoding emits a series of function calls to e.g. `fma_fadd_v2f16`. We need to
# emit functions to disassemble a single decoded instruction in a particular
# state. Sync prototypes to avoid moves when calling.
disasm_op_template = Template("""static void
bi_disasm_${c_name}(FILE *fp, unsigned bits, struct bifrost_regs *srcs, struct bifrost_regs *next_regs, unsigned staging_register, unsigned branch_offset, struct bi_constants *consts, bool last)
{
${body.strip()}
}
""")
lut_template_only = Template(""" static const char *${field}[] = {
${", ".join(['"' + x + '"' for x in table])}
};
""")
# Given a lookup table written logically, generate an accessor
lut_template = Template(""" static const char *${field}_table[] = {
${", ".join(['"' + x + '"' for x in table])}
};
const char *${field} = ${field}_table[_BITS(bits, ${pos}, ${width})];
""")
# Helpers for decoding follow. pretty_mods applies dot syntax
def pretty_mods(opts, default):
return [('.' + (opt or 'reserved') if opt != default else '') for opt in opts]
# Recursively searches for the set of free variables required by an expression
def find_context_keys_expr(expr):
if isinstance(expr, list):
return set.union(*[find_context_keys_expr(x) for x in expr[1:]])
elif expr[0] == '#':
return set()
else:
return set([expr])
def find_context_keys(desc, test):
keys = set()
if len(test) > 0:
keys |= find_context_keys_expr(test)
for i, (_, vals) in enumerate(desc.get('derived', [])):
for j, val in enumerate(vals):
if val is not None:
keys |= find_context_keys_expr(val)
return keys
# Compiles a logic expression to Python expression, ctx -> { T, F }
EVALUATORS = {
'and': ' and ',
'or': ' or ',
'eq': ' == ',
'neq': ' != ',
}
def compile_derived_inner(expr, keys):
if expr == []:
return 'True'
elif expr is None or expr[0] == 'alias':
return 'False'
elif isinstance(expr, list):
args = [compile_derived_inner(arg, keys) for arg in expr[1:]]
return '(' + EVALUATORS[expr[0]].join(args) + ')'
elif expr[0] == '#':
return "'{}'".format(expr[1:])
elif expr == 'ordering':
return expr
else:
return "ctx[{}]".format(keys.index(expr))
def compile_derived(expr, keys):
return eval('lambda ctx, ordering: ' + compile_derived_inner(expr, keys))
# Generate all possible combinations of values and evaluate the derived values
# by bruteforce evaluation to generate a forward mapping (values -> deriveds)
def evaluate_forward_derived(vals, ctx, ordering):
for j, expr in enumerate(vals):
if expr(ctx, ordering):
return j
return None
def evaluate_forward(keys, derivf, testf, ctx, ordering):
if not testf(ctx, ordering):
return None
deriv = []
for vals in derivf:
evaled = evaluate_forward_derived(vals, ctx, ordering)
if evaled is None:
return None
deriv.append(evaled)
return deriv
def evaluate_forwards(keys, derivf, testf, mod_vals, ordered):
orderings = ["lt", "gt"] if ordered else [None]
return [[evaluate_forward(keys, derivf, testf, i, order) for i in itertools.product(*mod_vals)] for order in orderings]
# Invert the forward mapping (values -> deriveds) of finite sets to produce a
# backwards mapping (deriveds -> values), suitable for disassembly. This is
# possible since the encoding is unambiguous, so this mapping is a bijection
# (after reserved/impossible encodings)
def invert_lut(value_size, forward, derived, mod_map, keys, mod_vals):
backwards = [None] * (1 << value_size)
for (i, deriveds), ctx in zip(enumerate(forward), itertools.product(*mod_vals)):
# Skip reserved
if deriveds == None:
continue
shift = 0
param = 0
for j, ((x, width), y) in enumerate(derived):
param += (deriveds[j] << shift)
shift += width
assert(param not in backwards)
backwards[param] = ctx
return backwards
# Compute the value of all indirectly specified modifiers by using the
# backwards mapping (deriveds -> values) as a run-time lookup table.
def build_lut(mnemonic, desc, test):
# Construct the system
facts = []
mod_map = {}
for ((name, pos, width), default, values) in desc.get('modifiers', []):
mod_map[name] = (width, values, pos, default)
derived = desc.get('derived', [])
# Find the keys and impose an order
key_set = find_context_keys(desc, test)
ordered = 'ordering' in key_set
key_set.discard('ordering')
keys = list(key_set)
# Evaluate the deriveds for every possible state, forming a (state -> deriveds) map
testf = compile_derived(test, keys)
derivf = [[compile_derived(expr, keys) for expr in v] for (_, v) in derived]
mod_vals = [mod_map[k][1] for k in keys]
forward = evaluate_forwards(keys, derivf, testf, mod_vals, ordered)
# Now invert that map to get a (deriveds -> state) map
value_size = sum([width for ((x, width), y) in derived])
backwards = [invert_lut(value_size, f, derived, mod_map, keys, mod_vals) for f in forward]
# From that map, we can generate LUTs
output = ""
if ordered:
output += "bool ordering = (_BITS(bits, {}, 3) > _BITS(bits, {}, 3));\n".format(desc["srcs"][0][0], desc["srcs"][1][0])
for j, key in enumerate(keys):
# Only generate tables for indirect specifiers
if mod_map[key][2] is not None:
continue
idx_parts = []
shift = 0
for ((pos, width), _) in derived:
idx_parts.append("(_BITS(bits, {}, {}) << {})".format(pos, width, shift))
shift += width
built_idx = (" | ".join(idx_parts)) if len(idx_parts) > 0 else "0"
default = mod_map[key][3]
if ordered:
for i, order in enumerate(backwards):
options = [ctx[j] if ctx is not None and ctx[j] is not None else "reserved" for ctx in order]
output += lut_template_only.render(field = key + "_" + str(i), table = pretty_mods(options, default))
output += " const char *{} = ordering ? {}_1[{}] : {}_0[{}];\n".format(key, key, built_idx, key, built_idx)
else:
options = [ctx[j] if ctx is not None and ctx[j] is not None else "reserved" for ctx in backwards[0]]
output += lut_template_only.render(field = key + "_table", table = pretty_mods(options, default))
output += " const char *{} = {}_table[{}];\n".format(key, key, built_idx)
return output
def disasm_mod(mod, skip_mods):
if mod[0][0] in skip_mods:
return ''
else:
return ' fputs({}, fp);\n'.format(mod[0][0])
def disasm_op(name, op):
(mnemonic, test, desc) = op
is_fma = mnemonic[0] == '*'
# Modifiers may be either direct (pos is not None) or indirect (pos is
# None). If direct, we just do the bit lookup. If indirect, we use a LUT.
body = ""
skip_mods = []
body += build_lut(mnemonic, desc, test)
for ((mod, pos, width), default, opts) in desc.get('modifiers', []):
if pos is not None:
body += lut_template.render(field = mod, table = pretty_mods(opts, default), pos = pos, width = width) + "\n"
# Mnemonic, followed by modifiers
body += ' fputs("{}", fp);\n'.format(mnemonic)
srcs = desc.get('srcs', [])
for mod in desc.get('modifiers', []):
# Skip per-source until next block
if mod[0][0][-1] in "0123" and int(mod[0][0][-1]) < len(srcs):
continue
body += disasm_mod(mod, skip_mods)
body += ' fputs(" ", fp);\n'
body += ' bi_disasm_dest_{}(fp, next_regs, last);\n'.format('fma' if is_fma else 'add')
# Next up, each source. Source modifiers are inserterd here
for i, (pos, mask) in enumerate(srcs):
body += ' fputs(", ", fp);\n'
body += ' dump_src(fp, _BITS(bits, {}, 3), *srcs, consts, {});\n'.format(pos, "true" if is_fma else "false")
# Error check if needed
if (mask != 0xFF):
body += ' if (!({} & (1 << _BITS(bits, {}, 3)))) fputs("(INVALID)", fp);\n'.format(hex(mask), pos, 3)
# Print modifiers suffixed with this src number (e.g. abs0 for src0)
for mod in desc.get('modifiers', []):
if mod[0][0][-1] == str(i):
body += disasm_mod(mod, skip_mods)
# And each immediate
for (imm, pos, width) in desc.get('immediates', []):
body += ' fprintf(fp, ", {}:%u", _BITS(bits, {}, {}));\n'.format(imm, pos, width)
# Attach a staging register if one is used
if desc.get('staging'):
body += ' fprintf(fp, ", @r%u", staging_register);\n'
return disasm_op_template.render(c_name = opname_to_c(name), body = body)
print('#include "util/macros.h"')
print('#include "disassemble.h"')
states = expand_states(instructions)
print('#define _BITS(bits, pos, width) (((bits) >> (pos)) & ((1 << (width)) - 1))')
for st in states:
print(disasm_op(st, states[st]))
print(decode_op(states, True))
print(decode_op(states, False))
|