Change how memory operands are handled in ISA descriptions.

Should enable implementation of split-phase timing loads
with new memory model.
May create slight timing differences under FullCPU, as I
believe we were not handling software prefetches correctly
before when the split MemAcc/Exec model was used.  I haven't
looked into this in any detail though.

arch/alpha/isa/decoder.isa:
    HwLoadStore format split into separate HwLoad and
    HwStore formats.
    Copy instructions now fall under MiscPrefetch format.
    Mem_write_result is now just write_result in store
    conditionals.
arch/alpha/isa/mem.isa:
    Split MemAccExecute and LoadStoreExecute templates
    into separate templates for loads and stores; now
    that memory operands are handled differently from
    registers, it's impossible to have a single template
    serve both.
    Also unified the handling of "regular" prefetches
    (loads to r31) and "misc" prefetches (e.g., wh64)
    under the new scheme.  It looks like SW prefetches
    were not handled correctly in FullCPU up til now,
    since we generated an execute() method for the outer
    instruction but didn't generate a proper method for
    MemAcc::execute() (instead getting a default no-op
    method for that).
arch/alpha/isa/pal.isa:
    Split HwLoadStore into separate HwLoad and HwStore
    formats to select proper template (see change to
    mem.isa in this changeset).
arch/isa_parser.py:
    Stop trying to treat memory operands like register
    operands, since we never used them in a uniform way
    anyway, and it made it impossible to do split-phase
    loads as needed for the new CPU model.  Now there's no
    more 'op_mem_rd', 'op_nonmem_rd', etc.: 'op_rd' just does
    register operands, and the template code is responsible
    for formulating the call to the memory system.  Right now
    the only thing exported by InstObjParams is a new attribute
    'mem_acc_size' which gives the memory access size in bits,
    though more attributes can be added if needed.

    Also moved code in findOperands() method to
    OperandDescriptorList.__init__(), which is where it belongs.

--HG--
extra : convert_revision : 6d53d07e0c5e828455834ded4395fa40f9146a34

1 files changed, 87 insertions, 96 deletions

diff --git a/arch/isa_parser.py b/arch/isa_parser.py
index b12329ebb..fffcc33e5 100755
--- a/arch/isa_parser.py
+++ b/arch/isa_parser.py
@@ -868,16 +868,21 @@ def fixPythonIndentation(s):
     return s
 
 # Error handler.  Just call exit.  Output formatted to work under
-# Emacs compile-mode.
+# Emacs compile-mode.  This function should be called when errors due
+# to user input are detected (as opposed to parser bugs).
 def error(lineno, string):
     spaces = ""
     for (filename, line) in fileNameStack[0:-1]:
         print spaces + "In file included from " + filename
         spaces += "  "
+    # Uncomment the following line to get a Python stack backtrace for
+    # these errors too.  Can be handy when trying to debug the parser.
+    # traceback.print_exc()
     sys.exit(spaces + "%s:%d: %s" % (fileNameStack[-1][0], lineno, string))
 
 # Like error(), but include a Python stack backtrace (for processing
-# Python exceptions).
+# Python exceptions).  This function should be called for errors that
+# appear to be bugs in the parser itself.
 def error_bt(lineno, string):
     traceback.print_exc()
     print >> sys.stderr, "%s:%d: %s" % (input_filename, lineno, string)
@@ -1220,30 +1225,19 @@ class MemOperandTraits(OperandTraits):
         # to avoid 'uninitialized variable' errors from the compiler.
         # Declare memory data variable.
         c = '%s %s = 0;\n' % (type, op_desc.base_name)
-        # Declare var to hold memory access flags.
-        c += 'unsigned %s_flags = memAccessFlags;\n' % op_desc.base_name
-        # If this operand is a dest (i.e., it's a store operation),
-        # then we need to declare a variable for the write result code
-        # as well.
-        if op_desc.is_dest:
-            c += 'uint64_t %s_write_result = 0;\n' % op_desc.base_name
         return c
 
     def makeRead(self, op_desc):
-        (size, type, is_signed) = operandSizeMap[op_desc.eff_ext]
-        eff_type = 'uint%d_t' % size
-        return 'fault = xc->read(EA, (%s&)%s, %s_flags);\n' \
-               % (eff_type, op_desc.base_name, op_desc.base_name)
+        return ''
 
     def makeWrite(self, op_desc):
+        return ''
+
+    # Return the memory access size *in bits*, suitable for
+    # forming a type via "uint%d_t".  Divide by 8 if you want bytes.
+    def makeAccSize(self, op_desc):
         (size, type, is_signed) = operandSizeMap[op_desc.eff_ext]
-        eff_type = 'uint%d_t' % size
-        wb = 'fault = xc->write((%s&)%s, EA, %s_flags, &%s_write_result);\n' \
-               % (eff_type, op_desc.base_name, op_desc.base_name,
-                  op_desc.base_name)
-        wb += 'if (traceData) { traceData->setData(%s); }' % \
-              op_desc.base_name
-        return wb
+        return size
 
 class NPCOperandTraits(OperandTraits):
     def makeConstructor(self, op_desc):
@@ -1321,6 +1315,11 @@ class OperandDescriptor:
         else:
             self.eff_ext = self.traits.dflt_ext
 
+        # note that mem_acc_size is undefined for non-mem operands...
+        # template must be careful not to use it if it doesn't apply.
+        if self.traits.isMem():
+            self.mem_acc_size = self.traits.makeAccSize(self)
+
     # Finalize additional fields (primarily code fields).  This step
     # is done separately since some of these fields may depend on the
     # register index enumeration that hasn't been performed yet at the
@@ -1340,10 +1339,73 @@ class OperandDescriptor:
         else:
             self.op_wb = ''
 
+
 class OperandDescriptorList:
-    def __init__(self):
+
+    # Find all the operands in the given code block.  Returns an operand
+    # descriptor list (instance of class OperandDescriptorList).
+    def __init__(self, code):
         self.items = []
         self.bases = {}
+        # delete comments so we don't match on reg specifiers inside
+        code = commentRE.sub('', code)
+        # search for operands
+        next_pos = 0
+        while 1:
+            match = operandsRE.search(code, next_pos)
+            if not match:
+                # no more matches: we're done
+                break
+            op = match.groups()
+            # regexp groups are operand full name, base, and extension
+            (op_full, op_base, op_ext) = op
+            # if the token following the operand is an assignment, this is
+            # a destination (LHS), else it's a source (RHS)
+            is_dest = (assignRE.match(code, match.end()) != None)
+            is_src = not is_dest
+            # see if we've already seen this one
+            op_desc = self.find_base(op_base)
+            if op_desc:
+                if op_desc.ext != op_ext:
+                    error(0, 'Inconsistent extensions for operand %s' % \
+                          op_base)
+                op_desc.is_src = op_desc.is_src or is_src
+                op_desc.is_dest = op_desc.is_dest or is_dest
+            else:
+                # new operand: create new descriptor
+                op_desc = OperandDescriptor(op_full, op_base, op_ext,
+                                            is_src, is_dest)
+                self.append(op_desc)
+            # start next search after end of current match
+            next_pos = match.end()
+        self.sort()
+        # enumerate source & dest register operands... used in building
+        # constructor later
+        self.numSrcRegs = 0
+        self.numDestRegs = 0
+        self.numFPDestRegs = 0
+        self.numIntDestRegs = 0
+        self.memOperand = None
+        for op_desc in self.items:
+            if op_desc.traits.isReg():
+                if op_desc.is_src:
+                    op_desc.src_reg_idx = self.numSrcRegs
+                    self.numSrcRegs += 1
+                if op_desc.is_dest:
+                    op_desc.dest_reg_idx = self.numDestRegs
+                    self.numDestRegs += 1
+                    if op_desc.traits.isFloatReg():
+                        self.numFPDestRegs += 1
+                    elif op_desc.traits.isIntReg():
+                        self.numIntDestRegs += 1
+            elif op_desc.traits.isMem():
+                if self.memOperand:
+                    error(0, "Code block has more than one memory operand.")
+                self.memOperand = op_desc
+        # now make a final pass to finalize op_desc fields that may depend
+        # on the register enumeration
+        for op_desc in self.items:
+            op_desc.finalize()
 
     def __len__(self):
         return len(self.items)
@@ -1398,69 +1460,6 @@ commentRE = re.compile(r'//.*\n')
 # (used in findOperands())
 assignRE = re.compile(r'\s*=(?!=)', re.MULTILINE)
 
-#
-# Find all the operands in the given code block.  Returns an operand
-# descriptor list (instance of class OperandDescriptorList).
-#
-def findOperands(code):
-    operands = OperandDescriptorList()
-    # delete comments so we don't accidentally match on reg specifiers inside
-    code = commentRE.sub('', code)
-    # search for operands
-    next_pos = 0
-    while 1:
-        match = operandsRE.search(code, next_pos)
-        if not match:
-            # no more matches: we're done
-            break
-        op = match.groups()
-        # regexp groups are operand full name, base, and extension
-        (op_full, op_base, op_ext) = op
-        # if the token following the operand is an assignment, this is
-        # a destination (LHS), else it's a source (RHS)
-        is_dest = (assignRE.match(code, match.end()) != None)
-        is_src = not is_dest
-        # see if we've already seen this one
-        op_desc = operands.find_base(op_base)
-        if op_desc:
-            if op_desc.ext != op_ext:
-                error(0, 'Inconsistent extensions for operand %s' % op_base)
-            op_desc.is_src = op_desc.is_src or is_src
-            op_desc.is_dest = op_desc.is_dest or is_dest
-        else:
-            # new operand: create new descriptor
-            op_desc = OperandDescriptor(op_full, op_base, op_ext,
-                                        is_src, is_dest)
-            operands.append(op_desc)
-        # start next search after end of current match
-        next_pos = match.end()
-    operands.sort()
-    # enumerate source & dest register operands... used in building
-    # constructor later
-    srcRegs = 0
-    destRegs = 0
-    operands.numFPDestRegs = 0
-    operands.numIntDestRegs = 0
-    for op_desc in operands:
-        if op_desc.traits.isReg():
-            if op_desc.is_src:
-                op_desc.src_reg_idx = srcRegs
-                srcRegs += 1
-            if op_desc.is_dest:
-                op_desc.dest_reg_idx = destRegs
-                destRegs += 1
-                if op_desc.traits.isFloatReg():
-                    operands.numFPDestRegs += 1
-                elif op_desc.traits.isIntReg():
-                    operands.numIntDestRegs += 1
-    operands.numSrcRegs = srcRegs
-    operands.numDestRegs = destRegs
-    # now make a final pass to finalize op_desc fields that may depend
-    # on the register enumeration
-    for op_desc in operands:
-        op_desc.finalize()
-    return operands
-
 # Munge operand names in code string to make legal C++ variable names.
 # This means getting rid of the type extension if any.
 # (Will match base_name attribute of OperandDescriptor object.)
@@ -1489,7 +1488,7 @@ def makeFlagConstructor(flag_list):
 class CodeBlock:
     def __init__(self, code):
         self.orig_code = code
-        self.operands = findOperands(code)
+        self.operands = OperandDescriptorList(code)
         self.code = substMungedOpNames(substBitOps(code))
         self.constructor = self.operands.concatAttrStrings('constructor')
         self.constructor += \
@@ -1503,22 +1502,14 @@ class CodeBlock:
 
         self.op_decl = self.operands.concatAttrStrings('op_decl')
 
-        is_mem = lambda op: op.traits.isMem()
-        not_mem = lambda op: not op.traits.isMem()
-
         self.op_rd = self.operands.concatAttrStrings('op_rd')
         self.op_wb = self.operands.concatAttrStrings('op_wb')
-        self.op_mem_rd = \
-                 self.operands.concatSomeAttrStrings(is_mem, 'op_rd')
-        self.op_mem_wb = \
-                 self.operands.concatSomeAttrStrings(is_mem, 'op_wb')
-        self.op_nonmem_rd = \
-                 self.operands.concatSomeAttrStrings(not_mem, 'op_rd')
-        self.op_nonmem_wb = \
-                 self.operands.concatSomeAttrStrings(not_mem, 'op_wb')
 
         self.flags = self.operands.concatAttrLists('flags')
 
+        if self.operands.memOperand:
+            self.mem_acc_size = self.operands.memOperand.mem_acc_size
+
         # Make a basic guess on the operand class (function unit type).
         # These are good enough for most cases, and will be overridden
         # later otherwise.