isa_parser.py

来自「M5,一个功能强大的多处理器系统模拟器.很多针对处理器架构,性能的研究都使用它作」· Python 代码 · 共 1,867 行 · 第 1/5 页

    for cpu in cpu_models:
        result[cpu.name] = t % cpu.strings
    return result

# *If* the template has CPU-specific references, return a single
# string containing a copy of the template for each CPU model with the
# corresponding values substituted in.  If the template has no
# CPU-specific references, it is returned unmodified.
def expand_cpu_symbols_to_string(template):
    if template.find('%(CPU_') != -1:
        return reduce(lambda x,y: x+y,
                      expand_cpu_symbols_to_dict(template).values())
    else:
        return template

# Protect CPU-specific references by doubling the corresponding '%'s
# (in preparation for substituting a different set of references into
# the template).
def protect_cpu_symbols(template):
    return re.sub(r'%(?=\(CPU_)', '%%', template)

###############
# GenCode class
#
# The GenCode class encapsulates generated code destined for various
# output files.  The header_output and decoder_output attributes are
# strings containing code destined for decoder.hh and decoder.cc
# respectively.  The decode_block attribute contains code to be
# incorporated in the decode function itself (that will also end up in
# decoder.cc).  The exec_output attribute is a dictionary with a key
# for each CPU model name; the value associated with a particular key
# is the string of code for that CPU model's exec.cc file.  The
# has_decode_default attribute is used in the decode block to allow
# explicit default clauses to override default default clauses.

class GenCode:
    # Constructor.  At this point we substitute out all CPU-specific
    # symbols.  For the exec output, these go into the per-model
    # dictionary.  For all other output types they get collapsed into
    # a single string.
    def __init__(self,
                 header_output = '', decoder_output = '', exec_output = '',
                 decode_block = '', has_decode_default = False):
        self.header_output = expand_cpu_symbols_to_string(header_output)
        self.decoder_output = expand_cpu_symbols_to_string(decoder_output)
        if isinstance(exec_output, dict):
            self.exec_output = exec_output
        elif isinstance(exec_output, str):
            # If the exec_output arg is a single string, we replicate
            # it for each of the CPU models, substituting and
            # %(CPU_foo)s params appropriately.
            self.exec_output = expand_cpu_symbols_to_dict(exec_output)
        self.decode_block = expand_cpu_symbols_to_string(decode_block)
        self.has_decode_default = has_decode_default

    # Override '+' operator: generate a new GenCode object that
    # concatenates all the individual strings in the operands.
    def __add__(self, other):
        exec_output = {}
        for cpu in cpu_models:
            n = cpu.name
            exec_output[n] = self.exec_output[n] + other.exec_output[n]
        return GenCode(self.header_output + other.header_output,
                       self.decoder_output + other.decoder_output,
                       exec_output,
                       self.decode_block + other.decode_block,
                       self.has_decode_default or other.has_decode_default)

    # Prepend a string (typically a comment) to all the strings.
    def prepend_all(self, pre):
        self.header_output = pre + self.header_output
        self.decoder_output  = pre + self.decoder_output
        self.decode_block = pre + self.decode_block
        for cpu in cpu_models:
            self.exec_output[cpu.name] = pre + self.exec_output[cpu.name]

    # Wrap the decode block in a pair of strings (e.g., 'case foo:'
    # and 'break;').  Used to build the big nested switch statement.
    def wrap_decode_block(self, pre, post = ''):
        self.decode_block = pre + indent(self.decode_block) + post

################
# Format object.
#
# A format object encapsulates an instruction format.  It must provide
# a defineInst() method that generates the code for an instruction
# definition.

exportContextSymbols = ('InstObjParams', 'makeList', 're', 'string')

exportContext = {}

def updateExportContext():
    exportContext.update(exportDict(*exportContextSymbols))
    exportContext.update(templateMap)

def exportDict(*symNames):
    return dict([(s, eval(s)) for s in symNames])


class Format:
    def __init__(self, id, params, code):
        # constructor: just save away arguments
        self.id = id
        self.params = params
        label = 'def format ' + id
        self.user_code = compile(fixPythonIndentation(code), label, 'exec')
        param_list = string.join(params, ", ")
        f = '''def defInst(_code, _context, %s):
                my_locals = vars().copy()
                exec _code in _context, my_locals
                return my_locals\n''' % param_list
        c = compile(f, label + ' wrapper', 'exec')
        exec c
        self.func = defInst

    def defineInst(self, name, args, lineno):
        context = {}
        updateExportContext()
        context.update(exportContext)
        if len(name):
            Name = name[0].upper()
            if len(name) > 1:
                Name += name[1:]
        context.update({ 'name': name, 'Name': Name })
        try:
            vars = self.func(self.user_code, context, *args[0], **args[1])
        except Exception, exc:
            error(lineno, 'error defining "%s": %s.' % (name, exc))
        for k in vars.keys():
            if k not in ('header_output', 'decoder_output',
                         'exec_output', 'decode_block'):
                del vars[k]
        return GenCode(**vars)

# Special null format to catch an implicit-format instruction
# definition outside of any format block.
class NoFormat:
    def __init__(self):
        self.defaultInst = ''

    def defineInst(self, name, args, lineno):
        error(lineno,
              'instruction definition "%s" with no active format!' % name)

# This dictionary maps format name strings to Format objects.
formatMap = {}

# Define a new format
def defFormat(id, params, code, lineno):
    # make sure we haven't already defined this one
    if formatMap.get(id, None) != None:
        error(lineno, 'format %s redefined.' % id)
    # create new object and store in global map
    formatMap[id] = Format(id, params, code)


##############
# Stack: a simple stack object.  Used for both formats (formatStack)
# and default cases (defaultStack).  Simply wraps a list to give more
# stack-like syntax and enable initialization with an argument list
# (as opposed to an argument that's a list).

class Stack(list):
    def __init__(self, *items):
        list.__init__(self, items)

    def push(self, item):
        self.append(item);

    def top(self):
        return self[-1]

# The global format stack.
formatStack = Stack(NoFormat())

# The global default case stack.
defaultStack = Stack( None )

# Global stack that tracks current file and line number.
# Each element is a tuple (filename, lineno) that records the
# *current* filename and the line number in the *previous* file where
# it was included.
fileNameStack = Stack()

###################
# Utility functions

#
# Indent every line in string 's' by two spaces
# (except preprocessor directives).
# Used to make nested code blocks look pretty.
#
def indent(s):
    return re.sub(r'(?m)^(?!#)', '  ', s)

#
# Munge a somewhat arbitrarily formatted piece of Python code
# (e.g. from a format 'let' block) into something whose indentation
# will get by the Python parser.
#
# The two keys here are that Python will give a syntax error if
# there's any whitespace at the beginning of the first line, and that
# all lines at the same lexical nesting level must have identical
# indentation.  Unfortunately the way code literals work, an entire
# let block tends to have some initial indentation.  Rather than
# trying to figure out what that is and strip it off, we prepend 'if
# 1:' to make the let code the nested block inside the if (and have
# the parser automatically deal with the indentation for us).
#
# We don't want to do this if (1) the code block is empty or (2) the
# first line of the block doesn't have any whitespace at the front.

def fixPythonIndentation(s):
    # get rid of blank lines first
    s = re.sub(r'(?m)^\s*\n', '', s);
    if (s != '' and re.match(r'[ \t]', s[0])):
        s = 'if 1:\n' + s
    return s

# Error handler.  Just call exit.  Output formatted to work under
# Emacs compile-mode.  Optional 'print_traceback' arg, if set to True,
# prints a Python stack backtrace too (can be handy when trying to
# debug the parser itself).
def error(lineno, string, print_traceback = False):
    spaces = ""
    for (filename, line) in fileNameStack[0:-1]:
        print spaces + "In file included from " + filename + ":"
        spaces += "  "
    # Print a Python stack backtrace if requested.
    if (print_traceback):
        traceback.print_exc()
    if lineno != 0:
        line_str = "%d:" % lineno
    else:
        line_str = ""
    sys.exit(spaces + "%s:%s %s" % (fileNameStack[-1][0], line_str, string))


#####################################################################
#
#                      Bitfield Operator Support
#
#####################################################################

bitOp1ArgRE = re.compile(r'<\s*(\w+)\s*:\s*>')

bitOpWordRE = re.compile(r'(?<![\w\.])([\w\.]+)<\s*(\w+)\s*:\s*(\w+)\s*>')
bitOpExprRE = re.compile(r'\)<\s*(\w+)\s*:\s*(\w+)\s*>')

def substBitOps(code):
    # first convert single-bit selectors to two-index form
    # i.e., <n> --> <n:n>
    code = bitOp1ArgRE.sub(r'<\1:\1>', code)
    # simple case: selector applied to ID (name)
    # i.e., foo<a:b> --> bits(foo, a, b)
    code = bitOpWordRE.sub(r'bits(\1, \2, \3)', code)
    # if selector is applied to expression (ending in ')'),
    # we need to search backward for matching '('
    match = bitOpExprRE.search(code)
    while match:
        exprEnd = match.start()
        here = exprEnd - 1
        nestLevel = 1
        while nestLevel > 0:
            if code[here] == '(':
                nestLevel -= 1
            elif code[here] == ')':
                nestLevel += 1
            here -= 1
            if here < 0:
                sys.exit("Didn't find '('!")
        exprStart = here+1
        newExpr = r'bits(%s, %s, %s)' % (code[exprStart:exprEnd+1],
                                         match.group(1), match.group(2))
        code = code[:exprStart] + newExpr + code[match.end():]
        match = bitOpExprRE.search(code)
    return code


####################
# Template objects.
#
# Template objects are format strings that allow substitution from
# the attribute spaces of other objects (e.g. InstObjParams instances).

labelRE = re.compile(r'(?<!%)%\(([^\)]+)\)[sd]')

class Template:
    def __init__(self, t):
        self.template = t

    def subst(self, d):
        myDict = None

        # Protect non-Python-dict substitutions (e.g. if there's a printf
        # in the templated C++ code)
        template = protect_non_subst_percents(self.template)
        # CPU-model-specific substitutions are handled later (in GenCode).
        template = protect_cpu_symbols(template)

        # Build a dict ('myDict') to use for the template substitution.
        # Start with the template namespace.  Make a copy since we're
        # going to modify it.
        myDict = templateMap.copy()

        if isinstance(d, InstObjParams):
            # If we're dealing with an InstObjParams object, we need
            # to be a little more sophisticated.  The instruction-wide
            # parameters are already formed, but the parameters which
            # are only function wide still need to be generated.
            compositeCode = ''

            myDict.update(d.__dict__)
            # The "operands" and "snippets" attributes of the InstObjParams
            # objects are for internal use and not substitution.
            del myDict['operands']
            del myDict['snippets']

            snippetLabels = [l for l in labelRE.findall(template)
                             if d.snippets.has_key(l)]

            snippets = dict([(s, mungeSnippet(d.snippets[s]))
                             for s in snippetLabels])

            myDict.update(snippets)

            compositeCode = ' '.join(map(str, snippets.values()))

            # Add in template itself in case it references any
            # operands explicitly (like Mem)
            compositeCode += ' ' + template

            operands = SubOperandList(compositeCode, d.operands)

            myDict['op_decl'] = operands.concatAttrStrings('op_decl')

            is_src = lambda op: op.is_src
            is_dest = lambda op: op.is_dest

            myDict['op_src_decl'] = \
                      operands.concatSomeAttrStrings(is_src, 'op_src_decl')
            myDict['op_dest_decl'] = \
                      operands.concatSomeAttrStrings(is_dest, 'op_dest_decl')

            myDict['op_rd'] = operands.concatAttrStrings('op_rd')
            myDict['op_wb'] = operands.concatAttrStrings('op_wb')

            if d.operands.memOperand:
                myDict['mem_acc_size'] = d.operands.memOperand.mem_acc_size
                myDict['mem_acc_type'] = d.operands.memOperand.mem_acc_type

        elif isinstance(d, dict):
            # if the argument is a dictionary, we just use it.
            myDict.update(d)
        elif hasattr(d, '__dict__'):
            # if the argument is an object, we use its attribute map.
            myDict.update(d.__dict__)
        else:
            raise TypeError, "Template.subst() arg must be or have dictionary"
        return template % myDict

    # Convert to string.  This handles the case when a template with a
    # CPU-specific term gets interpolated into another template or into
    # an output block.
    def __str__(self):
        return expand_cpu_symbols_to_string(self.template)

#####################################################################
#
#                             Code Parser
#
# The remaining code is the support for automatically extracting
# instruction characteristics from pseudocode.