Borland-Pascal 7.0 Runtime Libary Update - Release 1.4 18-MAR-1994 Welcome to BPL70N15.ZIP, a collection of fast replacement libraries for your Turbo Pascal 7.0 / Borland Pascal 7.0 compiler. There are three libraries in this package, a real mode library (TURBO.TPL), a DOS protected mode library (TPP.TPL), and a Windows library (TPW.TPL). Every file is a complete replacement for the original library bearing the same name that came with your Pascal compiler. Due to the many optimizations in the replacement libraries, many programs compiled with these libraries will run faster. For more detailed information on possible performance improvements, see the file PERFORM.DOC. Only performance information for real mode and DOS protected mode programs can be provided at the moment. Those users already familiar with my previous project, the fast replacement library for Turbo Pascal 6.0 (distributed as TPL60N19.ZIP), may be disappointed that not all the features of that program have been included in BPL70N15.ZIP yet. I don't have much time at the moment, but still wanted to provide a BP 7.0 version of my library as soon as possible. So I decided to port the performance relevant stuff first and work on the other aspects later. The libaries in BPL70N15 maintain 99.9% compatibility with the original libraries. Differences are mostly caused by bug fixes and enhancements. Some bugs from the original libraries supplied by Borland have been eliminated, but there can be no guarantee that new ones have not crept in. Most of the code in the BPL70N15.ZIP libraries was ported from the latest version of my fast replacement library for Turbo Pascal 6.0, TPL60N19.ZIP, which has been proven to be a stable and reliable product. If you discover any bugs, or have other comments, please let me know. My email and snail mail addresses are given below. Although I am under severe time constraints, I will try as hard as possible to fix any bugs reported in as short a time as possible. The legal conditions under which Borland is distributing the source code of the Borland Pascal 7.0 run-time libraries are not entirely clear to me. To stay on the safe side, I assume that they are the same as for the RTL source for the TP 6.0 compiler. Under these conditions, I am not allowed to distribute modified source modules from the library. I may only provide the binaries to third parties. However, some of the modules in the BPL70N15.ZIP libaries do not contain a single line of code written by Borland and are written entirely by me. I am including the source for these modules for your reference. The source of the arithmetic routines can be found in the file ARISOURC.ZIP. The source code of most of the string routines is contained in the file STRSOURC.ZIP. The code of the arithmetic and string routines is hereby released into the public domain. You may use it in your own programs under the condition that you do not include it into a commercial product. Parties interested in commercial use of my code should contact me at my address below. Original library code is Copyright (C) 1983,92 Borland International New / additional library code is Copyright (C) 1988-1994 Norbert Juffa, Wielandtstr. 14, 76137 Karlsruhe, Germany Internet: S_JUFFA@IRAVCL.IRA.UKA.DE Contents of this document: I. Capabilities of RTL replacement II. Revision History III. References I. Capabilities of RTL replacement ================================== General note: BPL70N15 provides you with optimized libraries, it does not enhance the code produced by the Borland Pascal compiler. Thus, only code that uses many library calls can be expected to experience significant performance advantages. Library calls are made by BP 7.0 to operate on LONGINTs, STRINGs, REALs, SETs, perform heap operations such as allocating and deallocating memory (New, Dispose, GetMem, FreeMem), as well as to perform other tasks. One exception where BPL70N15 speeds up your code although no calls to optimized library routines are made is floating-point applications using a 287 or 387 coprocessor. If want to speed up your applications even further than can be accomplished by using BPL70N15, you might want to look at the "Sally TPU peephole optimizer" (SPO for short) written by Morten Welinder (terra@diku.dk). Unlike BPL70N15, this program is not in the public domain, but Morten grants free use of the program for personal, non-commercial use. SPO is a peephole optimizer that aims at optimizing the code produced by the Pascal compiler. Peephole optimizations means that the optimizer looks at a rather small collection of machine instructions at a time and replaces certain sequences it finds with optimized code. A TPU-optimizer speeds up those parts of a program that can't be enhanced by a replacement library and vice versa. So it might be a good idea to combine both tools to get the best performance out of your BP 7.0 programs. The SPO optimizer is currently distributed as a file SPO120.ZIP. It should be available from all ftp-sites that carry BPL70N15 and in particular can be downloaded from garbo.uwasa.fi, which is the upload site for the program. Please note that this is not intended to be an endorsement of the program. Rather, the info provided should be thought of as being a service to those users of BPL70N15 who want to speed up their programs even further than possible by using BPL70N15. Improvements in SYSTEM module -----------------------------o REAL type software arithmetic operations now comply with ANSI/IEEE Standard 754-1985 for Binary Floating Point Arithmetic [1,2] as much as possible. Note that REAL arithmetic by design differs from the standard in many ways, especially available numeric formats, value set, and available operations. The rounding mode implemented here is "round to nearest or even" as specified by the standard. Add, Subtract, Multiply, Squaring, Division, and Square Root deliver exact results with regard to this rounding mode, as demanded by the standard. Conversions from REAL to LONGINT and from EXTENDED to REAL use rounding to nearest or even, as specified in the standard. Correct implementation of above features was tested with the PARANOIA test program [3]. The correctness of basic REAL arithmetic functions has also been tested against the coprocessor/emulator EXTENDED format with the program FUN1_TST. The EXTENDED format carries approximately 19 decimal digits of precision. This description applies to all three libraries in the package.o REAL arithmetic operations have been sped up. Speed-up for SQRT varies between a factor of 11 for a 8086 and 30 for a 486DLC. FRAC now executes at twice the original speed and speed-up is between 50% and 100% for SIN, COS, ARCTAN, LN, EXP and division (2.8x speed up for division on 80386). Overall numeric processing power using REAL arithmetic increases by about 52% for an 8086 and by 85% for an Cyrix 486DLC as measured by the WHETSTONE benchmark [4,5]. This description applies to all three libraries, but the actual values cited are for the real mode library TURBO.TPL and may be different for the other libraries. In general, DOS protected mode and Windows programs tend to be slower than real mode programs by 5-50%.o Overall accuracy of REAL arithmetic transcendental functions has been improved as indicated by Cody&Waite's ELEFUNT tests [6]: DLOG, DEXP, DATAN, DSIN. Correct argument reduction ensures that relative error over the whole argument range does not exceed 1.9e-12 for Exp, 2.8e-12 for Arctan, and 2.7e-12 for Ln. These values have been determined by comparing the function returns of the REAL transcendental functions to the values computed on a Cyrix 83D87 coprocessor for the EXTENDED format. For Sin and Cos, relative error is also in the above range when the argument is reasonably small (e.g. in range -100..100) and not very close to an integer multiple of 0.25*Pi. The error of the transcendental functions expressed in ULPs (units in the last place) over the whole argument range does not exceed 1.6 ULPs for Exp, 1.8 ULPs for Arctan, and 2.2 ULPs for Ln. This description applies to all three libraries in the package.o Execution of coprocessor floating point computations using an 80287 or 80387 has been accelerated. For these coprocessors, NOPs will be inserted before every floating point instruction converted from an emulator interrupt instead of WAITs. As a result of this optimization, an improvement in execution speed of about 10% has been observed running the Lawrence Livermore Loops (LLL) [7] on a Cyrix 83D87, the improvement for the WHETSTONE benchmark on the 83D87 is similar. Maximum performance gain for tight loops (e.g. fractal computation) by this measure is about 22%.o On 80287XL, 80387, 80486DX, or compatible chips the Sin and Cos functions take advantage of the FSIN and FCOS instructions of these coprocessors, speeding up these functions by almost a factor of two. As a side effect, there is also some improvement in accuracy as measured by the DSIN test program from the ELEFUNT test suite. Also, the Arctan function takes advantage of the increased argument range of the FPATAN function. These optimizations result in another 19% increase in WHETSTONE power, so that the total combined speedup over the original library is about 30% for this benchmark using a 387 coprocessor.o STRING operations are faster, especially for longer strings. Most dramatic increase is in the INSERT function, with execution times reduced to up to one fourth compared with the original version of the RTL. Faster string operations cause 7% performance increase for the DHRYSTONE [8,9] benchmark on a 8086.o Improved speed of random number generation. Random for REAL numbers is 10-20% faster, Random for EXTENDED numbers is 5% faster. Due to the improvements in the uniform distribution of integer random numbers, there is a decrease in the speed of integer random number generation of about 5%.o Binary to decimal conversions used in Str and Write procedures have been sped up by up to 70% for integers (BYTE, SHORTINT, INTEGER, WORD, LONGINT), up to 5% for REAL numbers and about 3% for EXTENDED numbers.o Improved speed of LONGINT arithmetic for 8086..80286. Division enjoys a 30% reduction of execution time on 8086. On 386 and 486 type CPUs, the code used in BPL70N15 may be slower than that used by the original library, which uses 32 bit register operations, while BPL70N15 uses only 16 bit operations, however very cleverly. For most applications you will not note any drop in LONGINT performance on 386/486 machines by using BPL70N15.o Several of the functions of the heap manager have been tuned, resulting in 7%-11% faster operation for these routines, depending on the CPU used. This note applies only to the real mode heap manager in TURBO.TPL!o Set functions have been sped up by a few percent, but the add variable range operation may be up to eight times as fast.o UPCASE function has been enhanced to support the complete IBM character set. This means that characters ,,,,,,, are converted to upper case by this function.o Several bugs of the original RTL supplied by Borland have been fixed: The original routines to perform LONGINT shifts provide the wrong results when the program runs on a 386 or 486 type processor and the shift count exceeds 16. This has been fixed by replacing all LONGINT routines with my own code. My code doesn't use 386 specific instructions and foregoes the speed advantage offered by using 32-bit register operations. For all programs but a very few you will not notice any drop in performance on a 386-486 machine, though. GetDir now correctly returns a run-time error 15 (invalid drive) when called with a non existent drive. Differing from the original, it also signals all errors reported by DOS as run-time errors. E.g. when applied to a floppy drive that does not contain a floppy, it will now return run-time error 152 (drive not ready), where previously it would incorrectly signal successful completion of the operation (InOutRes = 0). For programs compiled with $N+, only true INFs are printed out as INF where with the original library some NaNs are also printed as INF. Correct operation can be tested with the INFBUG program. REAL arithmetic EXP functions no longer signals overflow when called with small arguments, but underflows to zero instead as it should. Denormals in EXTENDED computations no longer cause an invalid state on a 8087 coprocessor when being converted to true zeros. Consistency between register contents and tag bits is now asserted. Removal of this bug can be tested with the BUG87 program. Denormals in EXTENDED format are now correctly converted to decimal strings by the Str and Write routines. The original routines print EXTENDED precision denormals as zero. Note that BP 7.0 supports EXTENDED denormals only if your machine has an 80287XL, 80387, 80486 or equivalent. On the 8087 and Intel's original 80287 coprocessor denormals are only supported for the SINGLE and DOUBLE formats. Correct printing of extended precision denormals can be checked with the program DENORMTS. Program initialization routine now tries to prevent that programs compiled with the $G+ (286 code generation) switch are run on 8086 and 8088. The checks done are not 100% safe, but catch most of these cases, displaying the message "CPU > 8086 required" and aborting the program with a return code of 254 ($FE) instead of letting it crash. Note that this check lets programs compiled with $G+ run on 80186 and V20/V30 processors, since they have the ability to execute all 80286 real mode instructions produced by Turbo Pascal. This note applies only to real mode programs, as DOS protected mode and Windows programs will not run on anything less than a 286 anyhow. Improvements in CRT module --------------------------o Bug fix in routine DirectWrite. The method used to prevent "snow" when writing directly to a CGA graphics card was not entirely safe. When used in a heavily interrupted program (e.g. serial communication as a background task), it would not always write during the time when scanning was in the invisible parts of the screen. The method used now is 100% save and is even faster, since it takes advantage of the horizontal and vertical retrace periods, as opposed to the old method which only used the horizontal retrace time. The new routine has been tested successfully on an original IBM-CGA card. II. Revision History ====================o Changes since version 1.4, dated 01-07-1994 When fixing the bug in the Str function in the Windows library TPW.TPL for machines without a math coprocessor in version 1.4, I accidentally broke the code for machines that do have a math coprocessor. Thanks to Siegfried Voessner (VOESSNERQFTUG.AC.AT) for reporting this bug.o Changes since version 1.3, dated 10-24-1993 Fixed bug in the Str function in the Windows library TPW.TPL that caused the Str function to abort with run time error 207 when passed a non-zero argument on a PC without a math coprocessor. The bug was caused by the inability of Windows' coprocessor emulator to correctly emulate the FBSTP instruction. Many thanks to Eduardo Mauro (eduardo.mauro@bbs816.mandic.onsp.br) for reporting this bug and helping me to track it down.o Changes since version 1.2, dated 05-20-1993 Fixed bug in the code of the Instr function. The previously provided code did not work correctly when compiled with the $G+ option. Thanks to Flurin Honegger (honegger@urz.unibas.ch) for reporting this bug. Changes since version 1.1, dated 03-15-1993o Fixed bug in the Write and WriteLn routines of the CRT unit. Due to this bug, these routines could not function properly on a Hercules Graphics Card or other monochrome display. Thanks to Miha Vitorovic (Miha.Vitorovic@f108.n380.z2.fidonet.org) for reporting this bug. Changes since version 1.0, dated 03-10-1993o Fixed bug in LONGINT and REAL Val routine. Val erroneously returned the wrong value or an error code for syntactically correct strings. Thanks to Dennis J. Basiaga (dennisb@dancer.cc.bellcore.com), who was the first to report this bug. Version 1.0, original release. III. References ===============[1] IEEE: IEEE Standard for Binary Floating-Point Arithmetic. SIGPLAN Notices, Vol. 22, No. 2, 1985, pp. 9-25[2] IEEE Standard for Binary Floating-Point Arithmetic. ANSI/IEEE Std 754-1985. New York, NY: Institute of Electrical and Electronics Engineers 1985[3] Karpinski, R.: Paranoia: A Floating-Point Benchmark. Byte, February 1985, pp. 223-235[4] Curnow, H.J.; Wichmann, B.A.: A synthetic benchmark. Computer Journal, Vol. 19, No. 1, 1976, pp. 43-49[5] Wichmannn, B.A.: Validation code for the Whetstone benchmark. NPL Report DITC 107/88, National Physics Laboratory, UK, March 1988[6] Cody, W.J.; Waite, W.: Software Manual for the Elementary Functions. Englewood Cliffs, NJ: Prentice Hall 1980[7] McMahon, H.H.: The Livermore Fortran Kernels: A Test of the Numerical Performance Range. Technical Report UCRL-53745, Lawrence Livermore National Laboratory, December 1986, p. 179[8] Weicker, R.P.: Dhrystone: A Synthetic Systems Programming Benchmark. Communications of the ACM, Vol. 27, No. 10, October 1984, pp. 1013-1030[9] Weicker, R.P.: Dhrystone Benchmark: Rationale for Version 2 and Measurement Rules. SIGPLAN Notices, Vol. 23, No. 8, August 1988, pp. 49-62[10] 387DX User's Manual, Programmer's Reference. Intel 1989 Note: PARANOIA, DHRYSTONE, WHETSTONE, LLL, and ELEFUNT source code is available from NETLIB@ORNL.GOV