utils/update_llc_test_checks.py

   1 #!/usr/bin/env python2.7
   2
   3 """A test case update script.
   4
   5 This script is a utility to update LLVM X86 'llc' based test cases with new
   6 FileCheck patterns. It can either update all of the tests in the file or
   7 a single test function.
   8 """
   9
  10 import argparse
  11 import os         # Used to advertise this file's name ("autogenerated_note").
  12 import string
  13 import subprocess
  14 import sys
  15 import re
  16
  17 # Invoke the tool that is being tested.
  18 def llc(args, cmd_args, ir):
  19   with open(ir) as ir_file:
  20     stdout = subprocess.check_output(args.llc_binary + ' ' + cmd_args,
  21                                      shell=True, stdin=ir_file)
  22   # Fix line endings to unix CR style.
  23   stdout = stdout.replace('\r\n', '\n')
  24   return stdout
  25
  26
  27 # RegEx: this is where the magic happens.
  28
  29 ASM_FUNCTION_X86_RE = re.compile(
  30     r'^_?(?P<func>[^:]+):[ \t]*#+[ \t]*@(?P=func)\n[^:]*?'
  31     r'(?P<body>^##?[ \t]+[^:]+:.*?)\s*'
  32     r'^\s*(?:[^:\n]+?:\s*\n\s*\.size|\.cfi_endproc|\.globl|\.comm|\.(?:sub)?section|#+ -- End function)',
  33     flags=(re.M | re.S))
  34
  35 ASM_FUNCTION_ARM_RE = re.compile(
  36         r'^(?P<func>[0-9a-zA-Z_]+):\n' # f: (name of function)
  37         r'\s+\.fnstart\n' # .fnstart
  38         r'(?P<body>.*?)\n' # (body of the function)
  39         r'.Lfunc_end[0-9]+:', # .Lfunc_end0: or # -- End function
  40         flags=(re.M | re.S))
  41
  42 ASM_FUNCTION_AARCH64_RE = re.compile(
  43      r'^_?(?P<func>[^:]+):[ \t]*\/\/[ \t]*@(?P=func)\n'
  44      r'[ \t]+.cfi_startproc\n'
  45      r'(?P<body>.*?)\n'
  46      # This list is incomplete
  47      r'.Lfunc_end[0-9]+:\n',
  48      flags=(re.M | re.S))
  49
  50 ASM_FUNCTION_PPC_RE = re.compile(
  51     r'^_?(?P<func>[^:]+):[ \t]*#+[ \t]*@(?P=func)\n'
  52     r'\.Lfunc_begin[0-9]+:\n'
  53     r'[ \t]+.cfi_startproc\n'
  54     r'(?:\.Lfunc_[gl]ep[0-9]+:\n(?:[ \t]+.*?\n)*)*'
  55     r'(?P<body>.*?)\n'
  56     # This list is incomplete
  57     r'(?:^[ \t]*(?:\.long[ \t]+[^\n]+|\.quad[ \t]+[^\n]+)\n)*'
  58     r'.Lfunc_end[0-9]+:\n',
  59     flags=(re.M | re.S))
  60
  61 ASM_FUNCTION_SYSTEMZ_RE = re.compile(
  62     r'^_?(?P<func>[^:]+):[ \t]*#+[ \t]*@(?P=func)\n'
  63     r'[ \t]+.cfi_startproc\n'
  64     r'(?P<body>.*?)\n'
  65     r'.Lfunc_end[0-9]+:\n',
  66     flags=(re.M | re.S))
  67
  68
  69 SCRUB_WHITESPACE_RE = re.compile(r'(?!^(|  \w))[ \t]+', flags=re.M)
  70 SCRUB_TRAILING_WHITESPACE_RE = re.compile(r'[ \t]+$', flags=re.M)
  71 SCRUB_KILL_COMMENT_RE = re.compile(r'^ *#+ +kill:.*\n')
  72 SCRUB_LOOP_COMMENT_RE = re.compile(
  73     r'# =>This Inner Loop Header:.*|# in Loop:.*', flags=re.M)
  74
  75 SCRUB_X86_SHUFFLES_RE = (
  76     re.compile(
  77         r'^(\s*\w+) [^#\n]+#+ ((?:[xyz]mm\d+|mem)( \{%k\d+\}( \{z\})?)? = .*)$',
  78         flags=re.M))
  79 SCRUB_X86_SP_RE = re.compile(r'\d+\(%(esp|rsp)\)')
  80 SCRUB_X86_RIP_RE = re.compile(r'[.\w]+\(%rip\)')
  81 SCRUB_X86_LCP_RE = re.compile(r'\.LCPI[0-9]+_[0-9]+')
  82
  83 RUN_LINE_RE = re.compile('^\s*;\s*RUN:\s*(.*)$')
  84 TRIPLE_ARG_RE = re.compile(r'-mtriple=([^ ]+)')
  85 TRIPLE_IR_RE = re.compile(r'^target\s+triple\s*=\s*"([^"]+)"$')
  86 IR_FUNCTION_RE = re.compile('^\s*define\s+(?:internal\s+)?[^@]*@(\w+)\s*\(')
  87 CHECK_PREFIX_RE = re.compile('--?check-prefix(?:es)?=(\S+)')
  88 CHECK_RE = re.compile(r'^\s*;\s*([^:]+?)(?:-NEXT|-NOT|-DAG|-LABEL)?:')
  89
  90 def scrub_asm_x86(asm):
  91   # Scrub runs of whitespace out of the assembly, but leave the leading
  92   # whitespace in place.
  93   asm = SCRUB_WHITESPACE_RE.sub(r' ', asm)
  94   # Expand the tabs used for indentation.
  95   asm = string.expandtabs(asm, 2)
  96   # Detect shuffle asm comments and hide the operands in favor of the comments.
  97   asm = SCRUB_X86_SHUFFLES_RE.sub(r'\1 {{.*#+}} \2', asm)
  98   # Generically match the stack offset of a memory operand.
  99   asm = SCRUB_X86_SP_RE.sub(r'{{[0-9]+}}(%\1)', asm)
 100   # Generically match a RIP-relative memory operand.
 101   asm = SCRUB_X86_RIP_RE.sub(r'{{.*}}(%rip)', asm)
 102   # Generically match a LCP symbol.
 103   asm = SCRUB_X86_LCP_RE.sub(r'{{\.LCPI.*}}', asm)
 104   # Strip kill operands inserted into the asm.
 105   asm = SCRUB_KILL_COMMENT_RE.sub('', asm)
 106   # Strip trailing whitespace.
 107   asm = SCRUB_TRAILING_WHITESPACE_RE.sub(r'', asm)
 108   return asm
 109
 110 def scrub_asm_arm_eabi(asm):
 111   # Scrub runs of whitespace out of the assembly, but leave the leading
 112   # whitespace in place.
 113   asm = SCRUB_WHITESPACE_RE.sub(r' ', asm)
 114   # Expand the tabs used for indentation.
 115   asm = string.expandtabs(asm, 2)
 116   # Strip kill operands inserted into the asm.
 117   asm = SCRUB_KILL_COMMENT_RE.sub('', asm)
 118   # Strip trailing whitespace.
 119   asm = SCRUB_TRAILING_WHITESPACE_RE.sub(r'', asm)
 120   return asm
 121
 122 def scrub_asm_powerpc64le(asm):
 123   # Scrub runs of whitespace out of the assembly, but leave the leading
 124   # whitespace in place.
 125   asm = SCRUB_WHITESPACE_RE.sub(r' ', asm)
 126   # Expand the tabs used for indentation.
 127   asm = string.expandtabs(asm, 2)
 128   # Stripe unimportant comments
 129   asm = SCRUB_LOOP_COMMENT_RE.sub(r'', asm)
 130   # Strip trailing whitespace.
 131   asm = SCRUB_TRAILING_WHITESPACE_RE.sub(r'', asm)
 132   return asm
 133
 134 def scrub_asm_systemz(asm):
 135   # Scrub runs of whitespace out of the assembly, but leave the leading
 136   # whitespace in place.
 137   asm = SCRUB_WHITESPACE_RE.sub(r' ', asm)
 138   # Expand the tabs used for indentation.
 139   asm = string.expandtabs(asm, 2)
 140   # Strip trailing whitespace.
 141   asm = SCRUB_TRAILING_WHITESPACE_RE.sub(r'', asm)
 142   return asm
 143
 144
 145 # Build up a dictionary of all the function bodies.
 146 def build_function_body_dictionary(raw_tool_output, triple, prefixes, func_dict,
 147                                    verbose):
 148   target_handlers = {
 149       'x86_64': (scrub_asm_x86, ASM_FUNCTION_X86_RE),
 150       'i686': (scrub_asm_x86, ASM_FUNCTION_X86_RE),
 151       'x86': (scrub_asm_x86, ASM_FUNCTION_X86_RE),
 152       'i386': (scrub_asm_x86, ASM_FUNCTION_X86_RE),
 153       'aarch64': (scrub_asm_arm_eabi, ASM_FUNCTION_AARCH64_RE),
 154       'arm-eabi': (scrub_asm_arm_eabi, ASM_FUNCTION_ARM_RE),
 155       'thumb-eabi': (scrub_asm_arm_eabi, ASM_FUNCTION_ARM_RE),
 156       'thumbv8-eabi': (scrub_asm_arm_eabi, ASM_FUNCTION_ARM_RE),
 157       'armeb-eabi': (scrub_asm_arm_eabi, ASM_FUNCTION_ARM_RE),
 158       'powerpc64le': (scrub_asm_powerpc64le, ASM_FUNCTION_PPC_RE),
 159       's390x': (scrub_asm_systemz, ASM_FUNCTION_SYSTEMZ_RE),
 160   }
 161   handlers = None
 162   for prefix, s in target_handlers.items():
 163     if triple.startswith(prefix):
 164       handlers = s
 165       break
 166   else:
 167     raise KeyError('Triple %r is not supported' % (triple))
 168
 169   scrubber, function_re = handlers
 170   for m in function_re.finditer(raw_tool_output):
 171     if not m:
 172       continue
 173     func = m.group('func')
 174     scrubbed_body = scrubber(m.group('body'))
 175     if func.startswith('stress'):
 176       # We only use the last line of the function body for stress tests.
 177       scrubbed_body = '\n'.join(scrubbed_body.splitlines()[-1:])
 178     if verbose:
 179       print >>sys.stderr, 'Processing function: ' + func
 180       for l in scrubbed_body.splitlines():
 181         print >>sys.stderr, '  ' + l
 182     for prefix in prefixes:
 183       if func in func_dict[prefix] and func_dict[prefix][func] != scrubbed_body:
 184         if prefix == prefixes[-1]:
 185           print >>sys.stderr, ('WARNING: Found conflicting asm under the '
 186                                'same prefix: %r!' % (prefix,))
 187         else:
 188           func_dict[prefix][func] = None
 189           continue
 190
 191       func_dict[prefix][func] = scrubbed_body
 192
 193
 194 def add_checks(output_lines, run_list, func_dict, func_name):
 195   printed_prefixes = []
 196   for p in run_list:
 197     checkprefixes = p[0]
 198     for checkprefix in checkprefixes:
 199       if checkprefix in printed_prefixes:
 200         break
 201       if not func_dict[checkprefix][func_name]:
 202         continue
 203       # Add some space between different check prefixes.
 204       if len(printed_prefixes) != 0:
 205         output_lines.append(';')
 206       printed_prefixes.append(checkprefix)
 207       output_lines.append('; %s-LABEL: %s:' % (checkprefix, func_name))
 208       func_body = func_dict[checkprefix][func_name].splitlines()
 209       output_lines.append('; %s:       %s' % (checkprefix, func_body[0]))
 210       for func_line in func_body[1:]:
 211         output_lines.append('; %s-NEXT:  %s' % (checkprefix, func_line))
 212       # Add space between different check prefixes and the first line of code.
 213       # output_lines.append(';')
 214       break
 215   return output_lines
 216
 217
 218 def should_add_line_to_output(input_line, prefix_set):
 219   # Skip any blank comment lines in the IR.
 220   if input_line.strip() == ';':
 221     return False
 222   # Skip any blank lines in the IR.
 223   #if input_line.strip() == '':
 224   #  return False
 225   # And skip any CHECK lines. We're building our own.
 226   m = CHECK_RE.match(input_line)
 227   if m and m.group(1) in prefix_set:
 228     return False
 229
 230   return True
 231
 232
 233 def main():
 234   parser = argparse.ArgumentParser(description=__doc__)
 235   parser.add_argument('-v', '--verbose', action='store_true',
 236                       help='Show verbose output')
 237   parser.add_argument('--llc-binary', default='llc',
 238                       help='The "llc" binary to use to generate the test case')
 239   parser.add_argument(
 240       '--function', help='The function in the test file to update')
 241   parser.add_argument('tests', nargs='+')
 242   args = parser.parse_args()
 243
 244   autogenerated_note = ('; NOTE: Assertions have been autogenerated by '
 245                         'utils/' + os.path.basename(__file__))
 246
 247   for test in args.tests:
 248     if args.verbose:
 249       print >>sys.stderr, 'Scanning for RUN lines in test file: %s' % (test,)
 250     with open(test) as f:
 251       input_lines = [l.rstrip() for l in f]
 252
 253     triple_in_ir = None
 254     for l in input_lines:
 255       m = TRIPLE_IR_RE.match(l)
 256       if m:
 257         triple_in_ir = m.groups()[0]
 258         break
 259
 260     raw_lines = [m.group(1)
 261                  for m in [RUN_LINE_RE.match(l) for l in input_lines] if m]
 262     run_lines = [raw_lines[0]] if len(raw_lines) > 0 else []
 263     for l in raw_lines[1:]:
 264       if run_lines[-1].endswith("\\"):
 265         run_lines[-1] = run_lines[-1].rstrip("\\") + " " + l
 266       else:
 267         run_lines.append(l)
 268
 269     if args.verbose:
 270       print >>sys.stderr, 'Found %d RUN lines:' % (len(run_lines),)
 271       for l in run_lines:
 272         print >>sys.stderr, '  RUN: ' + l
 273
 274     run_list = []
 275     for l in run_lines:
 276       commands = [cmd.strip() for cmd in l.split('|', 1)]
 277       llc_cmd = commands[0]
 278
 279       triple_in_cmd = None
 280       m = TRIPLE_ARG_RE.search(llc_cmd)
 281       if m:
 282         triple_in_cmd = m.groups()[0]
 283
 284       filecheck_cmd = ''
 285       if len(commands) > 1:
 286         filecheck_cmd = commands[1]
 287       if not llc_cmd.startswith('llc '):
 288         print >>sys.stderr, 'WARNING: Skipping non-llc RUN line: ' + l
 289         continue
 290
 291       if not filecheck_cmd.startswith('FileCheck '):
 292         print >>sys.stderr, 'WARNING: Skipping non-FileChecked RUN line: ' + l
 293         continue
 294
 295       llc_cmd_args = llc_cmd[len('llc'):].strip()
 296       llc_cmd_args = llc_cmd_args.replace('< %s', '').replace('%s', '').strip()
 297
 298       check_prefixes = [item for m in CHECK_PREFIX_RE.finditer(filecheck_cmd)
 299                                for item in m.group(1).split(',')]
 300       if not check_prefixes:
 301         check_prefixes = ['CHECK']
 302
 303       # FIXME: We should use multiple check prefixes to common check lines. For
 304       # now, we just ignore all but the last.
 305       run_list.append((check_prefixes, llc_cmd_args, triple_in_cmd))
 306
 307     func_dict = {}
 308     for p in run_list:
 309       prefixes = p[0]
 310       for prefix in prefixes:
 311         func_dict.update({prefix: dict()})
 312     for prefixes, llc_args, triple_in_cmd in run_list:
 313       if args.verbose:
 314         print >>sys.stderr, 'Extracted LLC cmd: llc ' + llc_args
 315         print >>sys.stderr, 'Extracted FileCheck prefixes: ' + str(prefixes)
 316
 317       raw_tool_output = llc(args, llc_args, test)
 318       if not (triple_in_cmd or triple_in_ir):
 319         print >>sys.stderr, "Cannot find a triple. Assume 'x86'"
 320
 321       build_function_body_dictionary(raw_tool_output,
 322           triple_in_cmd or triple_in_ir or 'x86', prefixes, func_dict, args.verbose)
 323
 324     is_in_function = False
 325     is_in_function_start = False
 326     func_name = None
 327     prefix_set = set([prefix for p in run_list for prefix in p[0]])
 328     if args.verbose:
 329       print >>sys.stderr, 'Rewriting FileCheck prefixes: %s' % (prefix_set,)
 330     output_lines = []
 331     output_lines.append(autogenerated_note)
 332
 333     for input_line in input_lines:
 334       if is_in_function_start:
 335         if input_line == '':
 336           continue
 337         if input_line.lstrip().startswith(';'):
 338           m = CHECK_RE.match(input_line)
 339           if not m or m.group(1) not in prefix_set:
 340             output_lines.append(input_line)
 341             continue
 342
 343         # Print out the various check lines here.
 344         output_lines = add_checks(output_lines, run_list, func_dict, func_name)
 345         is_in_function_start = False
 346
 347       if is_in_function:
 348         if should_add_line_to_output(input_line, prefix_set) == True:
 349           # This input line of the function body will go as-is into the output.
 350           output_lines.append(input_line)
 351         else:
 352           continue
 353         if input_line.strip() == '}':
 354           is_in_function = False
 355         continue
 356
 357       if input_line == autogenerated_note:
 358         continue
 359
 360       # If it's outside a function, it just gets copied to the output.
 361       output_lines.append(input_line)
 362
 363       m = IR_FUNCTION_RE.match(input_line)
 364       if not m:
 365         continue
 366       func_name = m.group(1)
 367       if args.function is not None and func_name != args.function:
 368         # When filtering on a specific function, skip all others.
 369         continue
 370       is_in_function = is_in_function_start = True
 371
 372     if args.verbose:
 373       print>>sys.stderr, 'Writing %d lines to %s...' % (len(output_lines), test)
 374
 375     with open(test, 'wb') as f:
 376       f.writelines([l + '\n' for l in output_lines])
 377
 378
 379 if __name__ == '__main__':
 380   main()