----------------------------------------------------------------------
Sally Tpu Peephole Optimizer version 1,10
Morten Welinder
February-March 1993
----------------------------------------------------------------------
1. Welcome
Welcome to Sally Tpu Peephole Optimizer, Spo. This program lets you
produce faster programs with Borland Pascal version 7.0 by optimizing
the intermidiate code produced by that compiler.
The program and documentation of it (including this file) is copyright
1993 by Morten Welinder. "Sally" as the program outputs is also me.
Units compiled for Windows are currently not handled.
As a special bonus of using this program, the longint shift bug in the
runtime library will be eliminated.
2. Terms of use
I do not currently know what I want to do with this program. Therefore
the following terms apply to all use of Sally Tpu Peephole Optimizer
version 1,10:
1. The program may be used in 1993 only.
2. The program may be used by private individuals only.
3. You don't have to pay for using the program.
4. You may distribute copies of the program, if and if only
you distribute the whole package unmodified.
5. You use the program at your own risc.
3. Bug reports and suggestions
Bug reports, comments et cetera can be snail-mailed to
Morten Welinder
Borups Alle 249B, 3tv
2400 Koebenhavn NV
or emailed to "terra@diku.dk". The latter is my graduate student
account at Department of Computer Science, University of Copenhagen,
Denmark.
Please be specific: use the debugging options to localize the bugs and
include the failing source code.
The mail system at Diku was down some days late in February.
If you sent bug reports (concerning version 1,00) at that time and
didn't get a reply, please resubmit!
4. Usage
The optimizer is a command-line utility and regognizes lots of options.
Entering "Spo" without parameters gives you a list:
Sally Tpu Peephole Optimizer version 1,10
Copyright 1993 by Sally
Usage: Spo Infile(s) [Options]
/? Help
/Ax Assembler Example: Tasmx.Exe
/Tx Target 86, 186, ... , 486
/Rx Range Check (B)ounds, (E)liminate
/Sx Stack Check (U)nfold, (E)liminate
/Ix InOutRes Check (U)nfold, (E)liminate
/Ox Overflow Check (I)ntO, (E)liminate
/Nx 80n87 instructions (S)oft, (H)ard
/Cx Target platform (D)os, (P)rotected, (W)indows
/Dx Debug Keep (L)ist file.
(N)o optimization.
(S)ystem names.
Infile(s) is a list of the units to optimize.
Wildcards are now allowed.
Debug information must be present in the unit.
The optimized versions of the units will be written in the current
directory with an extension of "Ttt".
This may very well change in later versions.
Note that options with parameters (e.g., "/Sx") does not accept any
white space between the "S" and the parameter.
The options are described in greater detail in the following sections.
In order to run Spo you will need 300-600 KB free memory.
The presense of Xms memory (up to 64 KB) may improve capacity very
slightly. See the assembler option for information about how to use an
assembler with larger capacity.
An alternate version for users with 386 processors is available:
Spo386. This version runs under Dpmi and you will need some of the
files that came with Borland Pascal 7.0 (the same files you need to
start Bp.Exe). The alternate version of Spo doesn't have the 64K
limit for the assembly buffer and will thus optimize larger units.
Spo386 has been tested under Borland's dos extender only and since it
uses some dirty tricks to operate in 32 bit mode it may or may not
function under other dos extenders.
Please let me know it you have problems.
5. How?
Spo works by a three phase scheme
1. Disassembly.
2. Optimization of the generated assembly code.
3. Assembly.
This scheme has the side effect that you cannot rely on a particular
coding of some instruction. As an example "Mov Sp,Bp" has two
different encodings. This should not concern you.
In generel the speed-up comes from unfolding system procedures and
functions. A far call and return takes 30 clock cycles plus penalties
for broken instruction queues on a 486.
Other speed-up sources are use of larger instruction sets and less
local view on instructions.
6. Error messages and warnings
There are found kinds of messages from Spo: information, warnings,
errors, and internal errors. This (incomplete) list describes them.
8. New features to come or not
As one might expect this program could be enhanced to input the unit
(or to output the new unit) in older versions of the tpu file format.
It would not be very much work but is it worth the while?
The code generated for the Case statement could be improved
significantly by using jump tables when several braches are present.
The Fpu and Cpu types should be seperated.
Swapping to disk or Xms should be possible in order to prevent the
assembler from running out of memory.
Other procedures than system procedures should be unfolded, perhaps
specialized (producing new procedures when knowing the values of some
of the parameters). This is a technique in its childhood so you should
not expect much. Look up articles on "Partial Evaluation".
Redundant load and save instructions should be eliminated. Much more
should be done with 32-bit instructions, i.e., pushing pointers and
longints. The reason that it has not already been done is that a
concept of "dead" registers and flags has to be formulated.
9. Keeping the lawyers happy
Borland International holds "Tasm", "Turbo Assembler", and "Borland
Pascal" as trademarks; Microsoft Corp. holds "Windows" as trademark;
International Business Machines Corp. holds "Ibm Pc" as trademark;
Intel Corp. holds the cpu names as trademarks.
10. Thanks
Special thanks to William L. Peavy, author of TPU6. Without his work
this program would not have been possibly. If you're out there and
want some minor corrections to the format of tpu files, contact me.
Thanks to Ralf Brown for the interrupt list. There's nothing like
good documentation of technical issues.
Thanks to Duncan Murdoch for keeping the list of known bugs in Borland
Pascal. New versions of programs always seem to follow the line
"old bugs replaced by new and unknown ones."
Thanks to Norbert Juffa for useful comments and for pointing a serious
bugs in the handling of 8087 instructions.
|
Morten Welinder
