GNU SUPEROPTIMIZER
The superoptimizer is a function sequence generator that uses an exhaustive
generate-and-test approach to finding the shortest instruction sequence for
a given function. You have to tell the superoptimizer which function and
which CPU you want to generate code for, and how many instructions you can
accept.
The superoptimizer can't generate very long sequences, unless you have a
very fast computer or very much spare time. The time complexity of the used
algorithm is approximately
2n
O(m n )
where m is the number of available instructions on the architecture and n is
the shortest sequence for the goal function. The practical sequence length
limit depends on the target architecture and goal function arity; In most
cases it is about 5, but for a rich instruction set as the HPPA it is just
4. The longest sequence ever generated was for the MC68020 and 7
instructions long. It took several weeks to generate it...
The superoptimizer can't guarantee that it finds the best possible
instruction sequences for all possible goal functions. For example, it
doesn't even try to include immediate constants (other that -1, 0, +1, and
the smallest negative and biggest positive numbers) in the sequences.
Other reasons why not optimal sequences might be found is that not all
instructions are included, not even in their register-only form. Also, some
instructions included might not be correctly simulated. If you encounter
any of these problems, please report them to the address below.
WARNING! The generated sequences might be incorrect with a very small
probability. Always make sure a sequence is correct before using it. So
far, I have never encountered any incorrect sequences. If you find one,
please let me know about it!
Having said this, note that the superoptimizer practically always finds
optimal and correct sequences for functions that depend on registers only.
USAGE INSTRUCTIONS
The superoptimizer supports these CPUs: SPARC v7, Motorola 68000, 68020, and
88000, IBM POWER and PowerPC, AMD 29000, Intel x86 and 960 1.0 and 1.1,
Pyramid, DEC Alpha, HP PA-RISC, and Hitachi SH. SGI Mips is not supported,
since it doesn't have instructions whose use in non-obvious. Some new
instructions, like the Intel P6 and Sparc v9 conditional moves are not
supported.
You need an ANSI C compiler, for example GCC, to compile the superoptimizer.
Type
make CPU=-D<cpuname> superopt
where <cpuname> is one of SPARC, MC68000, MC68020, M88000, POWER, POWERPC,
AM29K, I386, I960 (for i960 1.0), I960B (for I960B 1.1), PYR, ALPHA, HPPA,
or SH. The compilation might take a long time and use up a lot of memory,
especially for HPPA.
You can also build all superoptimizers by typing:
make all
This will create superopt-sparc, superopt-power, etc.
There are also install targets, use `make install' to install a single
superoptimizer and `make install-all' to install all of them.
To run the superoptimizer, type
superopt -f<goal-function> | -all [-assembly] [-max-cost n]
[-shifts] [-extracts] [-no-carry-insns] [-extra-cost n]
and wait until the found instructions sequences are printed. For example,
superopt -flts0 -as
will print all sequences computing the statement
{ r = (signed_word) v0 < 0; }.
See below for some examples of possible goal functions.
By default, the superoptimizer doesn't try all immediate shift counts. To
enable all shift counts, pass -shifts as a command line option. To enable
all bit field extracts, use -extracts.
OPTIONS
The `-f' option has always to be defined to tell the superoptimizer for
which function it should try to to find an instruction sequence. See below
for possible function names.
Option names may be abbreviated.
-assembly
Output assembly suitable to feed the assembler instead of pseudo-
code suitable for humans.
-max-cost n
Limit the `cost' of the instruction sequence to n. May be used to
stop the search if no instruction sequence of that length or
shorter is found. By default this is 4.
-extra-cost n
Search for sequences n more expensive than the cheapest found
sequence. Default is 0 meaning that only the cheapest sequence(s)
are printed.
-no-carry-insns
Don't use instructions that use the carry flag. This might be
desirable on RISCs to simplify instruction scheduling.
-shifts
Include all shift counts supported by the target architecture in
the search. This slows down the search considerably.
-extracts
Include all bit field extracts supported by the target architecture
in the search. This slows down the search considerably.
-f<goal-function-name>
where <goal-function-name> is one of eq, ne, les, ges, lts, gts,
leu, geu, ltu, gtu, eq0, ne0, les0, ges0, lts0, gts0, neq, nne,
nles, nges, nlts, ngts, nleu, ngeu, nltu, ngtu, neq0, nne0, nles0,
nges0, nlts0, ngts0, maxs, mins, maxu, minu, sgn, abs, nabs, gray,
or gray2, etc, etc.
eq, ne, les, etc, computes the C expression "a == b", "a != b", "a
<= b", etc, where the operation codes ending in `s' indicates
signed comparison; `u` indicates unsigned comparison.
eq0,... computes "a == 0", ...
The `n' before the names means that the corresponding function
value is negated, e.g. nlt is the C expression "-(a < b)".
maxs, mins, maxu, minu are binary (i.e. two argument) signed
respectively unsigned max and min.
sgn is the unary sign function; -1 for negative, 0 for zero, and +1
for positive arguments.
abs and nabs are absolute value and negative absolute value,
respectively.
For a complete list of goal function and their definitions, look in
the file goal.def. You can easily add your own goal functions to
that file. After having added a new function, you have to recompile
the superoptimizer.
READING SUPEROPTIMIZER OUTPUT
The superoptimizer by default outputs sequences in high-level language like
syntax. For example, this is the output for M88000/abs:
1: r1:=arith_shift_right(r0,0x1f)
r2:=add_co(r1,r0)
r3:=xor(r2,r1)
2: r1:=arith_shift_right(r0,0x1f)
r2:=add(r1,r0)
r3:=xor(r2,r1)
3: r1:=arith_shift_right(r0,0x1f)
r2:=xor(r1,r0)
r3:=adc_co(r2,r1)
r1:=arith_shift_right(r0,0x1f) means "shift r0 right 31 steps
arithmetically and put the result in r1". add_co is "add and set carry".
adc_co is the subtraction instruction found on most RISCs, i.e. "add with
complement and set carry". This may seem dumb, but there is an important
difference in the way carry is set after an addition-with-complement and a
subtraction. The suffixes "_ci" and "_cio" means respectively that carry
is input but not affected, and that carry is both input and generated.
The interesting value is always the value computed by the last instruction.
*********************************
Please send comments, improvements and new ports to tege@gnu.ai.mit.edu.
The GNU superoptimizer was written by Torbjorn Granlund (currently with
Cygnus Support). Tom Wood (at the time with Data General, now at Motorola)
made major improvements, like the clean way to describe goal functions and
internal instructions. The original superoptimizer idea is due to Henry
Massalin.
The GNU superoptimizer and it's application for tuning GCC are described in
the proceedings of the ACM SIGPLAN conference on Programming Language
Design an Implementation (PLDI), 1992.