NAME
IMCC - operation
VERSION
OVERVIEW
This document describes the principles of imcc operation.
DESCRIPTION
The main features of imcc are:
Source file parsing
See parsing.pod, macros.pod and syntax.pod for more.
Register allocation
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, these might get the same parrot register.
$I0 = 10
$I1 = 20will translate to
set I0, 10
set I0, 20when $I0 is not used after these lines.
The first redundant assignment will be optimized away with -O2.
PASM registers keep their register. During the usage of a PASM register this register will be not get assigned to. So don't use these over long code pieces and only when really needed.
Basic blocks
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.
Special case: subroutine calls (bsr)
Currently, imcc assumes, that a subroutine saves all registers on entry, as P6C does per saveall, and thus a bsr doesn't count as a branch.
But there is one exception to this rule: when the target of the subroutine is known (the label is in the same compilation unit), and when no saveall instruction is found in the first 5 instructions after the label, then the control flow from the bsr call to the label and from the ret to the next basic block is calculated.
Call graph
All connections between the basic blocks are calculated. This allows for:
Loop detection
where the range and depth of loops is calculated.
Life analysis
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.
When 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 to code snippets (block numbers are attached):
.sub _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
.endand:
.sub _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 $I1 # no init in code path from block 0
2 print "\n"
2 end
.endThe latter of these emits the warning:
warning:imcc:propagate_need: '$I1' might be used \
uninitialized in _main:7Interference graph
Once the above information is calculated, the next step is to look at which variables interfere with which other. Non-interfering variables can be given the same parrot register.
Register allocation
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.
Spilling
When there are too many interferences, so that not all variables can be allocated to a parrot register, variables with lowest score are spilled, or stored in a PerlArray when they are not used.
set $I1,1
set $I2,2
...
set $I33, 33
...
print $I1
...
print $I33The usage of all these variables makes it necessary to spill.
new P31, .PerlArray
...
set I0, 1
set P31[0], I0 # store $I1
set I0, 2
set P31[1], I0 # store $I2
...
set I0, P31[0] # fetch $I1
print I0When $I1 or $I2 are accessed, then code is generated to fetch the value from the spill array. The life range of such a spilled variable is then only one fetch and/or store operation and the actual usage. Therefore the variable has minimal interferences with other variables.
Optimization
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 ...
Optimizations with -O1
Constant substitution
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.
If-branch optimization
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.
Branches to branches
Unconditional branch sequences get optimized to jump to the final label.
Unused labels
Labels not referenced somewhere get deleted.
Dead code removal
Code not reachable after an unconditional branch instruction and basic blocks that are not entered from somewhere get removed.
Optimizations with -O2
Note: These are currently experimental and might not do the Right Thing.
Used LHS once
For a sequence of code
$I0 = 10
$I1 = 20where $I0 is not used again, the first assignment will be tossed, resulting in code like:
set I0, 20Loop optimization
Instructions which are invariant to a loop are pulled out of the loop and inserted in front of the loop entry.
Code generation
Imcc either generates PASM or else directly generates a PBC file for running with parrot.
Running code
Additionally the generated code can be run immediately inside imcc. All parrot runtime options like -j or -t are available.
FILES
imc.c, cfg.c, optimizer.c, pbc.c
AUTHOR
Leopold Toetsch <lt@toetsch.at>