Skip to content

Latest commit

 

History

History
272 lines (209 loc) · 14.8 KB

Develop.md

File metadata and controls

272 lines (209 loc) · 14.8 KB
Raw

Debug

Visual Studio debug configurations

  1. The Debug mode accept command line arguments argv[] which is set to default values in VS project configurations when running Debug mode, program automaticly loads codes from file /demos/1.txt

  2. The Develop mode invoke the function void debug() which is defined in /src/develop.cpp

  3. The Release is same as the Debug mode except for default values which is not set

Parsing

Currently, CFortranTranslator use /src/grammar/simple_lexer.cpp as tokenizer, /src/grammar/for90.y as parser

Tokenizer(by flex)

#define USE_LEX to enable tokenizer generated by flex from /src/grammar/for90.l

Tokenizer(by simple_lexer)

#undef USE_LEX to enable simple_lexer

simple_lexer is a more flexible tokenizer in order to handle some features of fortran

e.g. Fortran's continuation can exist between a token

        inte
     *ger :: a(10)

Parser

The parser is generated by bison from /src/grammar/for90.y The parser calls int yylex(void) to get one token from tokenizer at a time int yylex(void) is defined to pure_yylex when using flex, and simple_yylex when using simple_lexer

Macros

In /src/parser/parser.h, several helper macros are defined to help managing nodes. These macros are defined to different value according to different memory management strategies. There behaviour is controlled by defining USE_TRIVIAL/USE_POINTER or none of them.

  1. YY2ARG get a ParseNode & from bison's arguments $n $n can be ParseNode or ParseNode *
  2. RETURN_NT generates a bison's result $$(namely ParseNode *) from ParseNode, RETURN_NT does opposite work of YY2ARG $$ can be ParseNode or ParseNode *
  3. CLEAN_RIGHT/CLEAN_DELETE/CLEAN_REUSE clear all bison's arguments
    • CLEAN_RIGHT: a general rule which is deprecated and replaced by the following two cases
    • CLEAN_DELETE: tokens can't be reused in generated NT, usually are NTs
    • CLEAN_REUSE: tokens can be reused in generated NT, usually are Ts. CLEAN_REUSE may be or may not be reused in differenct situations:
      1. USE_TRIVIAL: Reuse
      2. USE_POINTER: Reuse
      3. else: Copy
    • e.g. 
       case_stmt_elem : YY_CASE '(' dimen_slice ')' at_least_one_end_line suite
      CLEAN_DELETE includes YY_CASE, '(', ')', at_least_one_end_line CLEAN_REUSE includes dimen_slice, suite

Extend grammar

This translator supports a subset of Fortran90's grammar. Grammar can be extended by adding new rules into /src/grammar/for90.y

  1. Declare new token by %token in /src/grammar/for90.y
  2. Add pattern of this token in /src/grammar/for90.l
  3. Add rules related to the token in /src/grammar/for90.y
  4. Update bytecodes and grammar tokens in /src/parser/Intent.h
  5. Register keyword in /src/parser/tokenizer.cpp(if this token is keyword)
  6. If this keyword are made up of more than one words, reduction conflicts may be caused between the whole keyword and its prefix, so this keyword must be handled specially by:
    • if this keyword is made up of non-word symbols without spaces between them. e.g. (/ is made up of ( and /. Just add the whole word as a rule to for90.l
    • if rule's keyword is made up words seperated by spaces, like else if add a new item into forward1 in /src/parser/tokenizer.cpp the first part else don't need to be registered as a keyword
  7. Update translation rules in /src/target/gen_config.h

Add new implemetation of Fortran's intrinsic function

This for90std library implements a subset of Fortran90's intrinsic functions.

  1. Implement this function and included it in for90std/for90std.h
    • if a parameter is optional in fortran, wrap it with foroptional, and log all parameters of this function in /gen_config.cpp
    • if the parameter is the only optional parameter, can omit foroptional wrapper
  2. Update funcname_map in /src/target/gen_config.cpp if necessary

Grammar rules explanation

In this section, serveral specific generating rules are discussed

argtable, dimen_slice, pure_paramtable

argtable, dimen_slice, pure_paramtable are a list of different items seprated by ,

Constitution

  • argtable is a list of exp(ref is_exp())
  • dimen_slice is a list of slice(NT_SLICE) or exp
  • pure_paramtable is a list of keyvalue(NT_KEYVALUE/NT_VARIABLE_ENTITY) or slice or exp pure_paramtable will be re-generated in regen_function_array in /src/target/gen_callable.cpp
  • paramtable is argtable or dimen_slice or pure_paramtable

Promotion

  • argtable + slice = dimen_slice, all elements in argtable will be promote to slice(with one child)
  • argtable + keyvalue = pure_paramtable, all elements in argtable will be promote to keyvalue
  • dimen_slice + keyvalue or pure_paramtable + slice is illegal

type_spec, type_name, type_selector

(3) The suffix "- spec" is used consistently for specifiers, such as keyword actual arguments and input / output statement specifiers.It also is used for type declaration attribute specifications(for example, "array - spec" in R512), and in a few other cases.

(4) When reference is made to a type parameter, including the surrounding parentheses, the term "selector" is used.See, for example, "length - selector"(R507) and "kind - selector"(R505).

You can use REAL(x) to get the float copy of x, however, you can also use REAL(kind = 8) to specify a floating number which is same to long double rather than double, so it may cause conflict. To specify, type_name is like INTEGER and a type_spec is like INTEGER(kind = 4), type_nospec can be head of callable, type_spec is not.

array builder

NT_FUCNTIONARRAY and NT_HIDDENDO will NOT be promote to NT_EXPRESSION

stmt, suite

  • stmt is statement end with ';' or '\n'
  • suite is a set of stmt

A partial table of rules

rules left side right side
fortran_program root wrappers
wrappers wrapper +
wrapper / function_decl / program
function_decl NT_FUNCTIONDECLARE
var_def NT_VARIABLEDEFINE/NT_DECLAREDVARIABLE
keyvalue NT_VARIABLE_ENTITY(namely NT_KEYVALUE) variable, NT_EXPRESSION / NT_VARIABLEINITIALDUMMY
NT_VARIABLE_ENTITY variable, exp
suite NT_SUITE stmt
stmt exp / var_def / compound_stmt / output_stmt / input_stmt / dummy_stmt / let_stmt / jump_stmt / interface_decl
NT_ARRAYBUILDER_LIST (NT_HIDDENDO / NT_FUNCTIONARRAY / exp)
type_spec type_name / (type_name, type_selector)

Target code generating

Lazy generating

regen/gen/reused function

When using lazy gen strategy, the node of non-terminal on the left side can change nodes of non-terminals on the right side. which means the AST is not immutable.

  1. regen_ function regen_ functions are declared in /src/target/codegen.h

    A regen_ function will change its input ParseNode &

  2. gen_ function gen_ functions are declared in /src/target/codegen.h

    A gen_ function will not change its input const ParseNode &. The gen_ function uses it input to generate and return a new ParseNode. The function may copy part of it's input as its child nodes.

  3. gen_xxx_reused function gen_xxx_reused function has a form like gen_xxx_reused, are declared in /src/target/codegen.h

    Different from gen_ function, a gen_reused function has input ParseNode &. Instead of copy some of its input like gen_ functions do, a gen_reused reuse some its input, by adding pointers to them directly.

    See parser:macros for more

Order of generating

Upper level Nodes of AST(above stmt level)

Due to fortran's feature of implicit declaration, code above stmt level, including function_decl, program can only be re-generated with correct type after the whole AST is built, by following steps:

  1. gen_fortran_program handles program, function_decl
  2. regen_suite handles suite rule
  3. regen_common generates common statement code

Variables and functions

The implicit declaration feature should also be considered when generating variables and functions, see variable definition/generate functions.

Name mapping

Many type names and function names are mapped in order to avoid possible conflicts, the mapping is defined by pre_map and funcname_map in /src/target/gen_config.h

pre_map and funcname_map have different usages:

  1. Mappings defined in pre_map are checked when doing tokenizing in /src/grammar/for90.l. In this stage, keywords and operators are replaced

  2. Mappings defined in funcname_map are checked when calling regen_function_array in /src/target/gen_callable.cpp. In this stage, only intrinsic functions' names are replaced to avoid possible confliction with other functions.

Variable definition

3-fold strategy

Currently, variable is generated lazily, so the whole process happens after the AST is built.

  1. Step 1:

    1. Case 1: when encountering a UnknownVariant during parsing (regen_exp, after the AST is built):

      ** Variables are registered to symbol table mostly in the condition** All variables, once reached by the parser will be registered to symbol table, by calling check_implicit_variable.

      check_implicit_variable checks whether this variable has been registered to gen_context().variables by the otehr cases already. If this variable hasn't been registed, check_implicit_variable will register it to gen_context().variables by:

      1. Add VariableInfo node
      2. .type is deduced by its name in function gen_implicit_type
      3. Set .implicit_defined = true, if an explicit declaration of this variable is found later, .implicit_defined will be set back to false automatically.
      4. .vardef is pointer(!= nullptr) to a ParseNode node.
      5. Set commonblock_name = "", commonblock_index = 0, if this variable is found belong to a common block, this two fields will be set.

    2. Case 2: when encounter NT_COMMONBLOCK(in regen_stmt, after the AST is built):

      This is an explicit definition

      Call regen_common to register VariableInfo into gen_context().variables, mark commonblock_name and commonblock_index

    3. Case 3: when encounter NT_VARIABLEDEFINESET and NT_VARIABLEDEFINE(in regen_stmt, after the AST is built):

      This is an explicit definition

      These two nodes are generated into NT_VARIABLEDEFINESET and NT_VARIABLEDEFINE, in /src/grammar/for90.y. Register(add a new VariableInfo only when the variable is never used, or modify the exsiting VariableInfo, for the most cases) corresponding VariableInfo to symbol table.

    Step 1 works simultanously with regen_ functions below suite level. After Step 1, all variables, whether explicit or implicit, are registered. However there is an exception that all implicit variables (e.g. A) used to initialize an variable (e.g. B) will not be registed until regen_vardef is called to B, in Step 2

    integer A = B + 1 ! B is not registered

  2. Step 2:

    This procedure is defined in the function regen_all_variables, which includes a while-loop in which regen_vardef is called to every variable. This while-loop can't be replaced by an 1-pass for-loop, because regen_vardef may introduce new variables, according to Step 1. After the while-loop, all VariableInfo of this suite are generated(and their .generate field will all set to true).

  3. Step 3:

    This procedure is defined in the function regen_all_variables_decl_str, according to function's 2-phase generating strategy. regen_all_variables_decl_str is called in regen_function_2, it will generate code for variables generated in Step 2. The generated codes depends on whether this variable is common block.

Common block

Common block is in global name space(finfo = get_function("", ""), vinfo = get_commonblock(commonblock_name)->variables[commonblock_index] = get_variable("", "BLOCK_" + commonblock_name, local_varname))

Interface

  1. All items in interface are firstly variables, so it will be

    1. registered by add_variable under finfo->local_name
    2. its local_name will be exactly the item's name

  2. All items in interface can also considered to be function(with no function body), so it will be

    1. registered by add_function under get_context().current_module
    2. its local_name will be finfo->local_name + "@" + the item's name

Generate functions

Functions in fortran are strongly connected:

  1. Functions shares common blocks, common blocks have relationship with body of function it belongs to. so variable decl part of a function must generated after all information of common block is gathered, which requires more than 1-pass scan of all functions.

  2. Call function with keyword arguments needs the body of callee function parsed

2-fold strategy

ParseNode(AST)

All parse tree nodes are defined in /src/Intent.h with an NT_ prefix

struct ParseNode

  1. fs:
    • fs.CurrentTerm.what: immediate-generated code, generated from child's fs.CurrentTerm.what, or from other infomations
    • fs.CurrentTerm.token: refer /src/Intent.h
  2. child
  3. attr: attrs including
    • FunctionAttr
    • VariableDescAttr
  4. father: pointer to parent node

Nodes

Child ParseNode may also be referred when generating upper level ParseNode, so do not change child index of:

  1. NT_VARIABLE_ENTITY: referred in function_decl
  2. NT_FUNCTIONDECLARE: can represent interface, referred in paramtable and function_decl

Symbols

All variables(including commom block) and functions is now logged in /src/Variable.h and /src/Function.h by VariableInfo and FunctionInfo

VariableDesc

Item Rule
kind typecast_spec
len typecast_spec
dimension variable_desc_elem
intent variable_desc_elem
optional variable_desc_elem
parameter variable_desc_elem

Attributes

-> means a ParseNode has this ParseAttr

ParseAttr Usage
VariableDescAttr NT_DECLAREDVARIABLE or NT_VARIABLEDEFINE or NT_VARIABLEINTIAL nodes of NT_VARIABLEDEFINE.NT_PARAMTABLE_PURE
FunctionAttr NT_FUNCTIONDECLARE
VarialbeAttr NT_FUNCTIONDECLARE