|Contents | IMCC|
IMCC - operation
This document describes the principles of IMCC operation.
The main features of imcc are:
See parsing.pod, macros.pod and syntax.pod for more.
Register allocation is done per compilation unit.
IMCC identifiers and temporary variables e.g. $I0 are assigned a physical parrot register depending on the life range of these variables. If the life range of one variable doesn't overlap the range of another variable, they might get the same parrot register. For instance:
$I0 = 10 $I1 = 20
will translate to
set I0, 10 set I0, 20
provided that $I0 is not used after these lines. In this case, the assignment to $I0 is redundant and will be optimized away if imcc is run with optimization level -O2.
PASM registers keep their register. During the usage of a PASM register this register will be not get assigned to. Therefore, they should be used only when absolutely necessary, and you should try to avoid using them within long pieces of code.
To determine the life range of variables, the code gets separated into pieces, called basic blocks. A basic block starts at a label, which can get jumped to, and ends at a branch instruction.
All connections between the basic blocks are calculated. This allows for:
where the range and depth of loops is calculated.
Whenever an operand is marked as an OUT argument, this operand starts with a new value. This means that at this point the life range of the symbol ends and a new life range is started, which allows the allocation of a different register to the same variable or the same register to a different variable.
Variables used as IN parameters must keep their parrot register over their usage range.
imcc detects a register usage, where the first operation is using (reading) a register (and warnings are enabled),
imcc emits an appropriate message.
Consider these two code snippets (block numbers are attached):
.sub main :main 0 $I0 = 0 # initialized 0 if $I0 goto l1 1 $I1 = 1 # init in block 1 1 goto l2 2 l1: 2 $I1 = 2 # init in block 2 3 l2: 3 print $I0 3 print $I1 # all paths leading here do init 3 print "\n" 3 end .end
.sub main @MAIN 0 $I0 = 0 # initialized 0 if $I0 goto l1 # branch to bb 1 or 2 1 $I1 = 1 # init only in block 1 2 l1: 2 print $I0 2 print $I1 # no init in code path from block 0 2 print "\n" 2 end .end
The latter of these emits the warning:
warning:imcc:propagate_need: '$I1' might be used \ uninitialized in _main:7
Once the above information is calculated, the next step is to look at which variables interfere with which others. Non-interfering variables can be given the same parrot register.
imcc then starts allocating registers according to a variable's score. Variables deeply nested inside loops have the highest score and get a parrot register first. Variables with a long life range (i.e. with many interferences) get allocated last.
Optimizations are only done when enabled with the -O switch. Please consult t/imcpasm/*.t for examples. They occur between various stages and may be repeatedly done: e.g. after converting a conditional branch to an absolute one, unreachable code will be removed then, which might cause unused labels ...
Constant arguments to many ops are evaluated. Conditional branches with constant conditions are converted to unconditional branches. Integer arguments to float operations are converted to float operands.
A sequence of code:
if cond, L1 branch L2 L1: ... L2:
will be converted to
unless cond, L2 ... L2:
The same is done for other conditional branches gt, ge, eq and their reverse meanings.
Unconditional branch sequences get optimized to jumps to the final label.
Unreferenced labels are deleted.
Code not reachable after an unconditional branch instruction and basic blocks that are not entered from somewhere get removed.
Note: These are currently experimental and might not do the Right Thing.
For a sequence of code
$I0 = 10 $I1 = 20
where $I0 is not used again, the first assignment will be tossed, resulting in code like:
set I0, 20
Instructions which are invariant to a loop are pulled out of the loop and inserted in front of the loop entry.
imcc either generates PASM or else directly generates a PBC file for running with parrot.
Additionally the generated code can be run immediately inside imcc. All parrot runtime options like -j or -t are available.
imc.c, cfg.c, optimizer.c, pbc.c
Leopold Toetsch <email@example.com>