README.txt - Readme file for PIRC compiler, a fresh implementation of the PIR language using Bison and Flex.


PIRC is a fresh implementation of the PIR language. Maintaining the current default implementation (IMCC) is a bit of a pain, and it contains a lot of "XXX" and "TODO" and other kludge alerts. Furthermore, IMCC is not re-entrant, as it has a number of global variables.

Overview ^

The new Bison/Flex based implementation of the PIR compiler is designed as a three-stage compiler:

1. Heredoc preprocessor

2. Macro preprocessor

3. PIR compiler

Heredoc preprocessing ^

The heredoc preprocessor takes the input as written by the PIR programmer, and flattens out all heredoc strings. An example is shown below to illustrate this concept:

The following input:

 .sub main
   $S0 = <<'EOS'
 This is a heredoc string

is transformed into:

   $S0 = "This is a heredoc string\n  divided\n    over\n      five\n        lines.\n"

Macro preprocessing ^

The macro layer basically implements text replacements. The following directives are handled:





The .include directive takes a string argument, which is the name of a file. The contents of this file are inserted at the point where the .include directive is written. To illustrate this, consider the following example:

 .sub main
   print "hi\n"

 .include "lib.pir"

 .sub foo
   print "foo\n"

This will result in the following output:

 .sub main
   print "hi\n"

 .sub foo
   print "foo\n"


The macro directive starts a macro definition. The macro preprocessor implements the expansion of macros. For instance, given the following input:

 .macro say(msg)
   print .msg
   print "\n"

 .sub main
   .say("hi there!")

will result in this output:

 .sub main
   print "hi there!"
   print "\n"


The .macro_const directive is similar to the .macro directive, except that a .macro_const is just a simplified .macro; it merely gives a name to some constant:

 .macro_const PI 3.14

 .sub main
   print "PI is approximately: "
   print .PI
   print "\n"

This will result in the output:

 .sub main
   print "PI is approximately: "
   print 3.14
   print "\n"

PIR Compiler ^

The output of the macro preprocessor is fed to the actual PIR compiler. The PIR compiler builds a data structure during the parsing phase (often referred to as an Abstract Syntax Tree - AST). The rest of this section describes the features of the PIR compiler.

Instruction selection

As Parrot instructions are polymorphic, the PIR compiler is responsible for selecting the right variant of the instruction. The selection is based on the types of the operands. For instance:

 set $I0, 42

will select the set_i_ic instruction: this is the set instruction, taking an integer (i) result operand and an integer constant (ic) operand. Other examples are:

 $P0[1] = 42           --> set_p_kic_ic # kic = key integer constant
 $I0 = $P0["hi"]       --> set_i_p_kc   # kc = key constant from constant table
 $P1 = new "Hash"      --> set_p_sc     # sc = string constant

Constant folding

Expressions that can be evaluated at compile-time are pre-evaluated, saving calculations during runtime. Some constant-folding is required, as Parrot depends on this. For instance:

 add $I0, 1, 2

is not a valid Parrot instruction; there is no add_i_ic_ic instruction. Instead, this will be translated to:

 set $I0, 3

which, as was explained earlier, will select the set_i_ic instruction.

The conditional branch instructions are also pre-evaluated, if possible. For instance, consider the following statement:

 if 1 < 2 goto L1

It is clear during compile time, that 1 is smaller than 2; so instead of evaluating this during runtime, we know for sure that the branch to label L1 will be made, effectively replacing the above statement by:

 goto L1

Likewise, if it's clear that certain instructions don't have any effect, they can be removed altogether:

 if 1 > 2 goto L1        --> nop  # nop is no opcode.
 $I0 = $I0 + 0           --> nop

Another type of optimization is the selection of (slightly) more efficient variants of instructions. For instance, consider the following instruction:

 $I0 = $I0 + $I1

which is actually syntactic sugar for:

 add $I0, $I0, $I1

In C one would write (ignoring the fact that $I0 and $I0 are not a valid C identifiers):

 $I0 += $I1

which is in fact valid PIR as well. When the PIR parser sees an instruction of this form, it will automatically select the variant with 2 operands instead of the 3-operand variant. So:

 add $I0, $I0, $1    # $I0 is an out operand

will be optimized, as if you had written:

 add $I0, $I1        # $I0 is an in/out operand

The PIR parser can do even more improvements, if it sees opportunity to do so. Consider the following statement:

 $I0 = $I0 + 1

or, in Parrot assembly syntax:

 add $I0, $I0, 1

Again, in C one would write (again ignoring the valid identifier issue): $I0++, or in other words, incrementing the given identifier. Parrot has inc and dec instructions built-in as well, so that the above statement $I0 = $I0 + 1 can be optimized to:

 inc $I0

Vanilla Register Allocator

The PIR compiler implements a vanilla register allocator. This means that each declared .local or .param symbol, and each PIR register ($Px, $Sx, $Ix, $Nx) is assigned a unique PASM register, that is associated with the original symbol or PIR register throughout the subroutine.

Any further optimizations on register usage can be implemented by writing a register allocator that takes this initial register allocation as input, and generating a more optimized register usage. Research and benchmarking is needed to decide whether this yields more efficient bytecode. In the end it is a choice between compile-time overhead (register allocation) or runtime memory overhead (more register space needed per sub).

The implementation of the vanilla register allocator is done in the PIR symbol management module (pirsymbol.c).

Status ^

The PIR parser is complete, but should be tested intensively. The back-end creates a data structure representing the input. Currently, only (almost working) PASM output is generated, but eventually a Parrot Byte Code (PBC) file should be generated. In order to do this, we need a proper API to generate the appropriate data structures (such as Parrot PackFile and friends).


The directory compilers/pirc has a number of subdirectories:

doc - contains documentation.

heredoc - contains the implementation of the heredoc preprocessor

macro - contains the implementation of the macro layer

new - contains the Bison/Flex implementation of PIRC

src - contains the hand-written, recursive-descent implementation of PIRC. Note that this is no longer maintained at the moment.

t - for tests.


Usage ^

Currently the different compilers/pre-processors are located in different directories. The different pre-processors are invoked from the main driver in pirc.c. The latter assumes all three processors are compiled, as the following executables:

 heredoc pre-processor: hdocprep
 macro pre-processor:   macroparser

Running a file through the whole PIR compiler is then done as follows:

 $ ./pirc test.pir

When you want to run the heredoc pre-processor only, do this:

 $ ./pirc -H test.pir

When you want to pre-process the file only (heredoc + macro parsing), do this:

 $ ./pirc -E test.pir

Cygwin processable lexer spec. ^

The file pir.l from which the lexer is generated is not processable by Cygwin's default version of Flex. In order to make a reentrant lexer, a newer version is needed, which can be downloaded from the link below.

Just do:

 $ ./configure
 $ make

Then make sure to overwrite the supplied flex binary.


Having a look at this implementation would be greatly appreciated, and any resulting feedback even more :-)


Eventually, either IMCC needs to be fixed rigorously, or, rewritten altogether. PIRC is an attempt to do the latter. The following things need to be considered when replacing IMCC with PIRC:


See also: