diff options
Diffstat (limited to 'Documentation/arm/nwfpe')
| -rw-r--r-- | Documentation/arm/nwfpe/NOTES | 29 | ||||
| -rw-r--r-- | Documentation/arm/nwfpe/README | 70 | ||||
| -rw-r--r-- | Documentation/arm/nwfpe/README.FPE | 156 | ||||
| -rw-r--r-- | Documentation/arm/nwfpe/TODO | 67 |
4 files changed, 322 insertions, 0 deletions
diff --git a/Documentation/arm/nwfpe/NOTES b/Documentation/arm/nwfpe/NOTES new file mode 100644 index 00000000000..40577b5a49d --- /dev/null +++ b/Documentation/arm/nwfpe/NOTES @@ -0,0 +1,29 @@ +There seems to be a problem with exp(double) and our emulator. I haven't +been able to track it down yet. This does not occur with the emulator +supplied by Russell King. + +I also found one oddity in the emulator. I don't think it is serious but +will point it out. The ARM calling conventions require floating point +registers f4-f7 to be preserved over a function call. The compiler quite +often uses an stfe instruction to save f4 on the stack upon entry to a +function, and an ldfe instruction to restore it before returning. + +I was looking at some code, that calculated a double result, stored it in f4 +then made a function call. Upon return from the function call the number in +f4 had been converted to an extended value in the emulator. + +This is a side effect of the stfe instruction. The double in f4 had to be +converted to extended, then stored. If an lfm/sfm combination had been used, +then no conversion would occur. This has performance considerations. The +result from the function call and f4 were used in a multiplication. If the +emulator sees a multiply of a double and extended, it promotes the double to +extended, then does the multiply in extended precision. + +This code will cause this problem: + +double x, y, z; +z = log(x)/log(y); + +The result of log(x) (a double) will be calculated, returned in f0, then +moved to f4 to preserve it over the log(y) call. The division will be done +in extended precision, due to the stfe instruction used to save f4 in log(y). diff --git a/Documentation/arm/nwfpe/README b/Documentation/arm/nwfpe/README new file mode 100644 index 00000000000..771871de0c8 --- /dev/null +++ b/Documentation/arm/nwfpe/README @@ -0,0 +1,70 @@ +This directory contains the version 0.92 test release of the NetWinder +Floating Point Emulator. + +The majority of the code was written by me, Scott Bambrough It is +written in C, with a small number of routines in inline assembler +where required. It was written quickly, with a goal of implementing a +working version of all the floating point instructions the compiler +emits as the first target. I have attempted to be as optimal as +possible, but there remains much room for improvement. + +I have attempted to make the emulator as portable as possible. One of +the problems is with leading underscores on kernel symbols. Elf +kernels have no leading underscores, a.out compiled kernels do. I +have attempted to use the C_SYMBOL_NAME macro wherever this may be +important. + +Another choice I made was in the file structure. I have attempted to +contain all operating system specific code in one module (fpmodule.*). +All the other files contain emulator specific code. This should allow +others to port the emulator to NetBSD for instance relatively easily. + +The floating point operations are based on SoftFloat Release 2, by +John Hauser. SoftFloat is a software implementation of floating-point +that conforms to the IEC/IEEE Standard for Binary Floating-point +Arithmetic. As many as four formats are supported: single precision, +double precision, extended double precision, and quadruple precision. +All operations required by the standard are implemented, except for +conversions to and from decimal. We use only the single precision, +double precision and extended double precision formats. The port of +SoftFloat to the ARM was done by Phil Blundell, based on an earlier +port of SoftFloat version 1 by Neil Carson for NetBSD/arm32. + +The file README.FPE contains a description of what has been implemented +so far in the emulator. The file TODO contains a information on what +remains to be done, and other ideas for the emulator. + +Bug reports, comments, suggestions should be directed to me at +<scottb@netwinder.org>. General reports of "this program doesn't +work correctly when your emulator is installed" are useful for +determining that bugs still exist; but are virtually useless when +attempting to isolate the problem. Please report them, but don't +expect quick action. Bugs still exist. The problem remains in isolating +which instruction contains the bug. Small programs illustrating a specific +problem are a godsend. + +Legal Notices +------------- + +The NetWinder Floating Point Emulator is free software. Everything Rebel.com +has written is provided under the GNU GPL. See the file COPYING for copying +conditions. Excluded from the above is the SoftFloat code. John Hauser's +legal notice for SoftFloat is included below. + +------------------------------------------------------------------------------- +SoftFloat Legal Notice + +SoftFloat was written by John R. Hauser. This work was made possible in +part by the International Computer Science Institute, located at Suite 600, +1947 Center Street, Berkeley, California 94704. Funding was partially +provided by the National Science Foundation under grant MIP-9311980. The +original version of this code was written as part of a project to build +a fixed-point vector processor in collaboration with the University of +California at Berkeley, overseen by Profs. Nelson Morgan and John Wawrzynek. + +THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort +has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT +TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO +PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY +AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE. +------------------------------------------------------------------------------- diff --git a/Documentation/arm/nwfpe/README.FPE b/Documentation/arm/nwfpe/README.FPE new file mode 100644 index 00000000000..26f5d7bb9a4 --- /dev/null +++ b/Documentation/arm/nwfpe/README.FPE @@ -0,0 +1,156 @@ +The following describes the current state of the NetWinder's floating point +emulator. + +In the following nomenclature is used to describe the floating point +instructions. It follows the conventions in the ARM manual. + +<S|D|E> = <single|double|extended>, no default +{P|M|Z} = {round to +infinity,round to -infinity,round to zero}, + default = round to nearest + +Note: items enclosed in {} are optional. + +Floating Point Coprocessor Data Transfer Instructions (CPDT) +------------------------------------------------------------ + +LDF/STF - load and store floating + +<LDF|STF>{cond}<S|D|E> Fd, Rn +<LDF|STF>{cond}<S|D|E> Fd, [Rn, #<expression>]{!} +<LDF|STF>{cond}<S|D|E> Fd, [Rn], #<expression> + +These instructions are fully implemented. + +LFM/SFM - load and store multiple floating + +Form 1 syntax: +<LFM|SFM>{cond}<S|D|E> Fd, <count>, [Rn] +<LFM|SFM>{cond}<S|D|E> Fd, <count>, [Rn, #<expression>]{!} +<LFM|SFM>{cond}<S|D|E> Fd, <count>, [Rn], #<expression> + +Form 2 syntax: +<LFM|SFM>{cond}<FD,EA> Fd, <count>, [Rn]{!} + +These instructions are fully implemented. They store/load three words +for each floating point register into the memory location given in the +instruction. The format in memory is unlikely to be compatible with +other implementations, in particular the actual hardware. Specific +mention of this is made in the ARM manuals. + +Floating Point Coprocessor Register Transfer Instructions (CPRT) +---------------------------------------------------------------- + +Conversions, read/write status/control register instructions + +FLT{cond}<S,D,E>{P,M,Z} Fn, Rd Convert integer to floating point +FIX{cond}{P,M,Z} Rd, Fn Convert floating point to integer +WFS{cond} Rd Write floating point status register +RFS{cond} Rd Read floating point status register +WFC{cond} Rd Write floating point control register +RFC{cond} Rd Read floating point control register + +FLT/FIX are fully implemented. + +RFS/WFS are fully implemented. + +RFC/WFC are fully implemented. RFC/WFC are supervisor only instructions, and +presently check the CPU mode, and do an invalid instruction trap if not called +from supervisor mode. + +Compare instructions + +CMF{cond} Fn, Fm Compare floating +CMFE{cond} Fn, Fm Compare floating with exception +CNF{cond} Fn, Fm Compare negated floating +CNFE{cond} Fn, Fm Compare negated floating with exception + +These are fully implemented. + +Floating Point Coprocessor Data Instructions (CPDT) +--------------------------------------------------- + +Dyadic operations: + +ADF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - add +SUF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - subtract +RSF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - reverse subtract +MUF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - multiply +DVF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - divide +RDV{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - reverse divide + +These are fully implemented. + +FML{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - fast multiply +FDV{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - fast divide +FRD{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - fast reverse divide + +These are fully implemented as well. They use the same algorithm as the +non-fast versions. Hence, in this implementation their performance is +equivalent to the MUF/DVF/RDV instructions. This is acceptable according +to the ARM manual. The manual notes these are defined only for single +operands, on the actual FPA11 hardware they do not work for double or +extended precision operands. The emulator currently does not check +the requested permissions conditions, and performs the requested operation. + +RMF{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - IEEE remainder + +This is fully implemented. + +Monadic operations: + +MVF{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - move +MNF{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - move negated + +These are fully implemented. + +ABS{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - absolute value +SQT{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - square root +RND{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - round + +These are fully implemented. + +URD{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - unnormalized round +NRM{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - normalize + +These are implemented. URD is implemented using the same code as the RND +instruction. Since URD cannot return a unnormalized number, NRM becomes +a NOP. + +Library calls: + +POW{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - power +RPW{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - reverse power +POL{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - polar angle (arctan2) + +LOG{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - logarithm to base 10 +LGN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - logarithm to base e +EXP{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - exponent +SIN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - sine +COS{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - cosine +TAN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - tangent +ASN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arcsine +ACS{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arccosine +ATN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arctangent + +These are not implemented. They are not currently issued by the compiler, +and are handled by routines in libc. These are not implemented by the FPA11 +hardware, but are handled by the floating point support code. They should +be implemented in future versions. + +Signalling: + +Signals are implemented. However current ELF kernels produced by Rebel.com +have a bug in them that prevents the module from generating a SIGFPE. This +is caused by a failure to alias fp_current to the kernel variable +current_set[0] correctly. + +The kernel provided with this distribution (vmlinux-nwfpe-0.93) contains +a fix for this problem and also incorporates the current version of the +emulator directly. It is possible to run with no floating point module +loaded with this kernel. It is provided as a demonstration of the +technology and for those who want to do floating point work that depends +on signals. It is not strictly necessary to use the module. + +A module (either the one provided by Russell King, or the one in this +distribution) can be loaded to replace the functionality of the emulator +built into the kernel. diff --git a/Documentation/arm/nwfpe/TODO b/Documentation/arm/nwfpe/TODO new file mode 100644 index 00000000000..8027061b60e --- /dev/null +++ b/Documentation/arm/nwfpe/TODO @@ -0,0 +1,67 @@ +TODO LIST +--------- + +POW{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - power +RPW{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - reverse power +POL{cond}<S|D|E>{P,M,Z} Fd, Fn, <Fm,#value> - polar angle (arctan2) + +LOG{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - logarithm to base 10 +LGN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - logarithm to base e +EXP{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - exponent +SIN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - sine +COS{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - cosine +TAN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - tangent +ASN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arcsine +ACS{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arccosine +ATN{cond}<S|D|E>{P,M,Z} Fd, <Fm,#value> - arctangent + +These are not implemented. They are not currently issued by the compiler, +and are handled by routines in libc. These are not implemented by the FPA11 +hardware, but are handled by the floating point support code. They should +be implemented in future versions. + +There are a couple of ways to approach the implementation of these. One +method would be to use accurate table methods for these routines. I have +a couple of papers by S. Gal from IBM's research labs in Haifa, Israel that +seem to promise extreme accuracy (in the order of 99.8%) and reasonable speed. +These methods are used in GLIBC for some of the transcendental functions. + +Another approach, which I know little about is CORDIC. This stands for +Coordinate Rotation Digital Computer, and is a method of computing +transcendental functions using mostly shifts and adds and a few +multiplications and divisions. The ARM excels at shifts and adds, +so such a method could be promising, but requires more research to +determine if it is feasible. + +Rounding Methods + +The IEEE standard defines 4 rounding modes. Round to nearest is the +default, but rounding to + or - infinity or round to zero are also allowed. +Many architectures allow the rounding mode to be specified by modifying bits +in a control register. Not so with the ARM FPA11 architecture. To change +the rounding mode one must specify it with each instruction. + +This has made porting some benchmarks difficult. It is possible to +introduce such a capability into the emulator. The FPCR contains +bits describing the rounding mode. The emulator could be altered to +examine a flag, which if set forced it to ignore the rounding mode in +the instruction, and use the mode specified in the bits in the FPCR. + +This would require a method of getting/setting the flag, and the bits +in the FPCR. This requires a kernel call in ArmLinux, as WFC/RFC are +supervisor only instructions. If anyone has any ideas or comments I +would like to hear them. + +[NOTE: pulled out from some docs on ARM floating point, specifically + for the Acorn FPE, but not limited to it: + + The floating point control register (FPCR) may only be present in some + implementations: it is there to control the hardware in an implementation- + specific manner, for example to disable the floating point system. The user + mode of the ARM is not permitted to use this register (since the right is + reserved to alter it between implementations) and the WFC and RFC + instructions will trap if tried in user mode. + + Hence, the answer is yes, you could do this, but then you will run a high + risk of becoming isolated if and when hardware FP emulation comes out + -- Russell]. |