Difference between revisions of "Simple graphics language"
(→Just In Time compiler) |
(→Arithmetic and logic peephole optimising) |
||
Line 173: | Line 173: | ||
*No operations | *No operations | ||
− | |||
**not not → nop | **not not → nop | ||
**dup and → nop | **dup and → nop |
Revision as of 17:05, 12 May 2011
Contents
A simple graphics language for microcontrollers
We start with a very simple reverse polish notation interpreter with a preprocessor stage
Sample code:
to circle 360 for 1 forward 1 right next end
Registers:
0 X (nn.ff) 1 Y 2 Angle 3 Scale 4 Colour 5 Pen up/down 6 Local variable pointer 7 Program counter
Stack diagrams
( things read from the stack -- things written on the stack ) x - x position y - y position n - number Examples: ( n n -- n ) This means that two numbers were read from the stack, then one number was written back. For example a simple addition reads two numbers and writes back the result. ( a b c -- c b a ) Reading threee values from the stack and write them back in the opposite order.
Definitions
- Function:
- The compiler/optimiser is free to inline functions for speed reasons.
- Functions that depend on exact knowledge of the return stack should use a Define instead.
- To < function name > < code > End
- Constant:
- Every instance of < constant name > in the source code will be replaced by < costant >
- Define < constant name > < constant > End
- Macro:
- The macro code will be executed at compile time and the output will be compiled (TODO: Figure out if this is wise)
- Hint: Magical constants and look-up tables can be generated by a macro
- Macro < macro name > < code > End
- Comments:
- ( everything here is a comment ( and so is here ) )
Commands
Absolute commands setcolour ( colour -- ) readcolour ( -- colour ) setpos ( x y -- ) readpos ( -- x y ) load ( address -- data ) store ( data address -- ) Variables: var ( n -- )( create n variables of value 0 on the return stack ) unvar ( n -- )( delete n variables from the return stack ) ldl ( index -- )( load local variable ) stl ( n index -- )( store local variable ) Relative commands: penup ( -- ) pendown ( -- ) forward ( length -- ) backward ( length -- ) left ( angle -- ) right ( angle -- ) Loops: for ( n -- ) next ( -- ) do ( -- ) again ( -- ) Conditionals: if ( n -- ) else ( -- ) endif ( -- ) Data processing: + ( n n -- n ) ( to + >r 0 r> - - end ) - ( n n -- n ) ( fundamental ) * ( n n -- n ) ( to * 0 do over 1 and if rot dup >r -rot r> + endif rot 1 lsl rot 1 lsr rot over if again endif >r drop -rot drop drop end ) / ( n n -- n ) lsl ( n n -- n ) ( to lsl for dup + next end ) lsr ( n n -- n ) ( to lsr for 31 for dup 0x80000000 and swap 1 lsl swap if 1 + endif next 0x7fffffff and next end ) and ( n n -- n ) ( fundamental ) or ( n n -- n ) ( to or over not and + end ) eor ( n n -- n ) ( to eor over over or -rot and not and end ) not ( n -- n ) ( to not >r -1 r> - end ) Stack manipulation: drop ( n -- ) ( to drop if endif end ) dup ( n -- n n ) ( fundamental ) swap ( a b -- b a ) ( to swap dup >r >r dup r> eor dup >r eor r> r> eor end ) over ( a b -- a b a ) ( to over >r dup r> swap end ) rot ( a b c -- b c a ) ( to rot >r swap r> swap end ) -rot ( a b c -- c a b ) ( to -rot swap >r swap r> end ) >r ( n -- ) ( fundamental ) r> ( -- n ) ( fundamental )
Compiler
Stage 1 (not implemented in version 1)
- Definitions are expanded
- Macros are executed
Stage 2
- Functions are added to the dictionary
Stage 3
- All words are looked up in the dictionary and the opcode stored in the program array
- The address of each function is stored in the function jump table in the same order as the functions were added to the dictionary
The dictionary
The dictionary used to convert a command into an opcode. It is a zero terminated string like "setcolour,readcolour,setpos,readpos,",0 that is searched linearly from start to end. A comma marks the end of a word. All functions are added to the end of this table in stage 2 of the compiling process.
The function jump table
This table contains the address of each function, in the same order as the functions appear in the dictionary. It is filled in during stage 3 when the actual code position is known.
Interpreter
The opcode pointer always points to the byte after the one being processed.
do opcode = program(opcodepointer) opcodepointer += 1 switch opcode case 0 push 0 case 1-4 read byte(s) value = byte0 or (byte1<<8) or (byte2<<16) or (byte3<<24) push value case command execute command case function pushr localpointer pushr opcodepointer opcodepointer = jumptable(function) end switch again
Opcodes
enum operands {op_ldc0, op_ldc8, op_ldc16, op_ldc24, op_ldc32, op_lds, op_end, op_setcolour, op_readcolour, op_setpos, op_readpos, op_load8, op_load16, op_load32, op_store8, op_store16, op_store32, op_var, op_unvar, op_ldl, op_stl, op_penup, op_pendown, op_forward, op_backward, op_left, op_right, op_call, op_for, op_next, op_do, op_again, op_if, op_else, op_endif, op_add, op_sub, op_mul, op_umul, op_div, op_udiv, op_lsl, op_lsr, op_asr, op_and, op_or, op_eor, op_not, op_drop, op_dup, op_swap, op_over, op_rot, op_nrot, op_dtor, op_rtod, op_nop};
- DIV, MUL and LSR are unsigned.
- /, *, asr are signed.
- Opcodes larger than op_nop are local function calls.
Constants
op_ldc0, op_ldc8, op_ldc16, op_ldc24, op_ldc32
- If the two or three most significant bytes are 0xff then shorter code is generated by:
- op_ldc16 0xnn 0xnn op_not
- op_ldc8 0xnn op_not
Optimisations
The bytecode is optimised in several separate steps after the compilation. Unused instruction slots are padded with nops. After each stage the code is compacted and the nops removed.
Arithmetic and logic peephole optimising
- Constant folding
- 1 8 lsl → 256 (for all arithmetic and logic operations)
- Strength reduction
- 0 div → -1 (TODO: check if this is what the interpreter would normally do...)
- 1 div → nop
- pow2(n) div → (n - 1) asr
- pow2(n) udiv → (n - 1) lsr
- 1 mul → nop
- 2 mul → dup +
- No operations
- not not → nop
- dup and → nop
- dup or → nop
Loading of constants
- Repeating numbers are shortened by using op_dup and op_over, 1 3 3 1 3 can be shortened to:
- op_ldc8 0x01 op_ldc8 0x03 op_dup op_ldc8 0x01 op_over
- Applying an eight bit constant and an operation on a copy of one of the two previous numbers.
- Applying an operation on copies of the two previous numbers.
Stack reordering
Optimisation is done by a brute force search generating optimal fragments of up to 7 instructions.
Common Subexpression Elimination
Identify common sub-expressions in the code, and use the result for each instance, rather than re-evaluating them each time.
- Not done
Loop Invariant Motion
This is the 'lifting of expressions' from loops. The compiler can identify that a particular expression inside a loop actually does not change as the loop is running.
- Not done
Tail Call Optimization and Tail Recursion
A tailcall is a call immediately before a return. Normally this function will be called, and when it returns to the caller, the caller returns again. Tailcall optimization avoids this by restoring the saved registers before jumping to the tailcall. The called function will now return directly to the caller's caller, saving a return sequence.
- Not done, will require a jump op-code
Just In Time compiler
The JIT compiler can be made very fast if the code quality is not very important. So loops can be compiled without considering how many times they will be executed. If they are executed more than 3 times there will be a benefit.
Register allocation
There are three register pools, stack, return stack and free. The pools are implemented as stacks with special functions.
Stack manipulation
All stack manipulations are done by register renaming so no code is generated.