RTL Representation

1
RTL Representation

• Kevin Klues
• Department of Computer Science and Engineering
• Fall 2005 Programming Language Seminar

2
Structure of Presentation

• Introduction
• RTL Example
• RTL Structure and Use in GCC
• RTL Compiler Pass
• Reading RTL
• Interfacing with GCC using RTL
• Exercise Suggestions

3
Introduction

• What is RTL?
• RTL Register Transfer Language
• Produced directly from Tree representation
• Breakdown of program into individual instructions
• Algebraic representation of each instruction
  describing exactly what it does
• Abstract low level intermediate representation
  that will eventually be mapped into Assembler
  code
• Has both internal and textual form
• Why is it important to learn about RTL?
• MOST of work done by compiler done on RTL
  representation
• Implementation of a back end tightly coupled with
  RTL

4
Example Textual RTL
int add(int x, int y) return
xy int sub(int x, int y) return
x-y int main(int argc, char argv)
int x, y, z y 5 x 6
z add(x,y) x sub(z,y)
return 0
(call_insn 21 20 22 (nil) (set (regSI 0 eax)
(call (memQI (symbol_refSI ("add")) 0 S1
A8) (const_int 8 0x8))) -1 (nil)
(expr_listREG_EH_REGION (const_int 0 0x0)
(nil)) (nil))
(call_insn 29 28 30 (nil) (set (regSI 0 eax)
(call (memQI (symbol_refSI ("sub")) 0 S1
A8) (const_int 8 0x8))) -1 (nil)
(expr_listREG_EH_REGION (const_int 0 0x0)
(nil)) (nil))
5
RTL Structure and Use in GCC

• RTL Object Types
• RTL Classes and Formats
• Access to Operands
• Flags in RTL Expressions
• Machine Modes
• RTL Representation of Functions

6
RTL Object Types

• Integers
• Wide Integers
• Strings
• Expressions (RTX)
• Classified by expression codes (rtl.def)
• Code put into RTX format with PUT_CODE(x, code)
• Code of an RTX extracted with GET_CODE(x)
• Vectors
• Arbitrary number of pointers to expressions

7
RTL Classes and Formats

• All written code associated with an expression
  code
• Expression code in turn associated with given
  class
• Can determine class of RTX with
  GET_RTX_CLASS(code)
• o,lt,1,c,3,2,b,I,m,g,a,x (see rtl.def file)
• Each expression code also has specified format
• Specifies number of contained objects and their
  type
• e,i, w, s, E, u, n, S, V, 0, , T, t, B, b (see
  rtl.c file)
• Defining RTX giving format and associating class
• DEF_RTL_EXPR(CONST, "const", "e", 'o')
• DEF_RTL_EXPR(JUMP_INSN, "jump_insn", "iuueiee0",
  'i')

8
Access to RTX operands

• XEXP, XINT, XWINT, XSTR
• Can see how used in rtl.c
• Any operand can be accessed as any type, so must
  be careful
• Type of each operand based on how defined in
  rtl.def
• Three ways to access Vector operands
• XVEC(exp, idx) Access to vector pointer at idx
  in exp
• XVECLEN(exp, idx) Access length of vector at
  idx in exp
• XVEC(exp, idx, elnum) Access RTX at elnum in
  vector at idx in exp
• Access to Special Operands
• MEM, REG, SYMBOL_REF
• Flags in RTL Expressions
• One-bit bit-fields used in certain types of
  expressions
• Can see how used in rtlanal.c, varasm.c

9
Machine Modes

• Describes size of data object and representation
  used for it
• Declared in enum machine_mode (machmode.h file)
• Very few explicit references to an RTXs machine
  mode
• Divided into classes
• Declared in enum mode_class (machmode.h file)
• Macros exist for getting the mode, setting a
  mode, etc.
• All RTXs have room for machine mode declaration
• Some tree expressions did too (declarations,
  types)
• Machine modes exist that a given machine MUST
  support
• QImode Quarter Integer, represents single
  byte as integer
• Modes corresponding to BITS_PER_WORD,
  FLOAT_TYPE_SIZE, DOUBLE_TYPE_SIZE

10
Function Representation

• Doubly-linked chain of insn objects
• Insn is nothing more than RTX with special
  expression code
• Every insn contains at least three extra fields
• unique id-num, pointer to preceding insn, pointer
  to next insn
• Occupy same position in all insns, independent of
  expression code
• Accessed using INSN_UID, PREV_INSN, NEXT_INSN
  macros
• First insn in function obtained by calling
  get_insns
• Certain expression codes contain even more extra
  fields
• insn, jump_insn, call_insn (see rtl.def file)
• Special expression codes for functions calling
  subroutines
• call_insn caller of a subroutine
• call subroutine callee
• Machine mode of insn normally VOIDmode
• May change during different passes of compiler

11
RTL Compiler Passes

• Source files for rtl generation are
• Stmt.c, calls.c, expr.c, explow.c, expmed.c,
  function.c, optabs.c, emit-rtl.c, insn-emit.c,
  expr.h, insn-flags.h, insn-codes.h
• Passes
• Generate exception handling landing pads
  (except.c)
• Cleanup control flow graph (cfgcleanup.c,
  cfgrtl.c, jump.c)
• Common subexpression elimination (cse.c)
• Global common subexpression elimination (gcse.c,
  lcm.c)
• 2 Loop Optimization passes (loop.c, dependence.c,
  loop-.c)
• Jump bypassing (gcse.c)
• If conversion (ifcvt.c)
• Web construction (web.c)
• Life Analysis (flow.c)
• Instruction Combination (combine.c)

12
RTL Compiler Passes (cont)

• Passes
• Register Movement (regmove.c)
• Optimize mode switching (mode-switching.c)
• Modulo-scheduling (modulo-sched.c)
• Instruction scheduling (haifa-sched.c, sched-.c)
• 4 Register allocation passes (regclass.c,
  local-alloc.c, global.c, reload.c, reload1.c,
  reload.h)
• Basic Block Reordering (bb-reorder.c, predict.c)
• Variable Tracking
• Delayed branch scheduling (reorg.c)
• Branch Shortening
• Register-to-stack conversion (reg-stack.c)
• Final (final.c, insn-output.c)
• Debug information output (dbxout.c, sbout.c,
  dwarf.c, vmsdbgout.c)

13
Reading RTL

• To dump an rtl file when compiling your program,
  add dr flags
• gcc dr foo.c
• Creates foo.c.rtl file with textual rtl dump
• If egypt and graphviz installed, can get
  graphical representation
• egypt foo.c.rtl dotty
• egypt foo.c.rtl dot -Grotate90 -Gsize8.5,11
  -Tps -o callgraph.ps
• egypt foo.c.rtl dot -Gpage8.5,11 -Tps -o
  callgraph.ps
• To read RTL object from file within C code, call
  read_rtx()
• Takes single stdio stream argument
• Defined in read-rtl.c file
• Not available in compiler itself, but in programs
  generating back-end

14
Interfacing with GCC using RTL

• Common misconception about RTL
• People want to store RTL representation in text
  file and use as interface between language front
  end and gcc
• INFEASIBLE because.
• GCC designed to use RTL internally only
• Correct RTL VERY dependant on particular target
  machine
• RTL does not contain complete description of all
  information contained within a program
• Proper way to interface GCC to new language is
  through tree data structure already discussed
• RTL for a particular machine then generated
  automatically from this representation
• See tree.h, tree.def, and notes from previous
  classes

15
Exercise Suggestions

• Morgan........