11 KiB
cc65 floating point support
The current main goal is to implement IEEE754 support in the compiler, using the "float" datatype as the common 32bit float.
*** ANYTHING ELSE COMES LATER ***
You can not use any of this to write software yet. Don't bother. This is for people who want to help pushing the floating point support further.
You can however try the current state of development. You should be able to build small (and slow...) programs that use floats on any supported target.
- Build the compiler/toolchain/libs from this fptest branch.
- Now you can build and run the samples and/or tests.
samples/floattest.c
samples/mandelfloat.c
samples/mathtest.c
samples/tgisincos.c
full math.h is available for C64 when linking agains fp754kernal.o (see below)
Further info
- Right now by default all targets will use the Berkeley Softfloat Library. This solves the annoying "chicken and egg" problem of having to implement the float support both in the compiler and in the target library at the same time, before anything can be tested properly. Fortunately that also means we can use the simulator for running test programs, and test changes in the compiler using our test bench, and that against a library that is known to somewhat work correctly :)
- The default library also contains a collection of math functions (which i dubbed "softmath"), which may fill the gap until more targets have specific wrappers and/or a generic math library (in assembly) was written.
- The default library can be overridden by linking an override file, similar to how you can use the soft80 implementation for conio. Right now such override files are provided for the C64 (c64-fp754kernal.o) and VIC20 (vic20-fp754kernal.o). The samples will automatically use the overrides.
WANTED
-
For the time being, i will not look at writing IEEE754 functions in assembly. Please see below for more info on what is needed to do this, should you be interested in doing this. Please contact me before you are putting work into this, so we can discuss some things and prevent anyone wasting time :)
-
It might be possible to produce a similar kernal- or OS- wrapper override file as the C64 one for other targets (or port the C64 one to other CBM targets).
- If you create a new one, keep in mind that the compiler right now will currently work with IEEE754 floats, which your library calls must also work with (which will involve converting forth and back to whatever other format at runtime), and there is no easy way tp change that.
Roadmap
- Test/Fix using the Softfloat lib some more, fix as much tests as possible
- When all obvious tests have been created and work OK, we can merge
- for the failing tests, create "negative" cases
After the merge, the following things can be done more or less independent from each other (not necessarily by me :)):
-
implement IEEE754 library
- for generic support this will be the best compromise for standard compliance and speed, but will take more RAM than alternatives.
- Once implemented, it will work for all targets.
-
implement support for native FP routines
- Some targets have FP routines in ROM (or OS) that can be used instead of providing our own in the library. This may or may not save RAM, and may or may not be faster than the default library.
- The wrapper library must implement conversion from/to 32bit IEEE754
-
implement support for native FP formats
- for this the wrappers in fp.c must be used in the compiler at all places.
- also we must implement the native format on the host in fp.c
- if the native format uses a different number of bytes than 4 for its native format, we must add support for this in the compiler at various places
- last not least a wrapper library that uses the native format must be created
- it is not unlikely that we will need extra tests for the native format
The Compiler
NOT WORKING YET:
/test/val/float-basic-var-var.c +=, -=
/test/val/float-basic-var-intvar.c +=, -=
/test/val/float-basic-var-intconst.c *, /, +=, -=, *=, /=
/test/val/float-basic-intvar-const.c +=, -=, *=, /=
/test/val/float-basic-intvar-var.c +=, /=
/test/val/float-basic-intconst-var.c *, /
/test/val/float-cmp-const-intvar.c
/test/val/float-cmp-intvar-const.c
/test/val/float-cmp-var-intvar.c
- float values written as "12.34f" work, but "12.34" does not - should it?
TODO:
- more tests are needed
- warnings
- errors
- register variables
- more/all tests should check local/global/register variables
Files & Functions
assignment.c
src/cc65/assignment.c
OpAssignArithmetic Parse an "=" or "op=" operation for arithmetic lhs
OpAddSubAssign Parse a "+=" or "-=" operation
codegen.c
src/cc65/codegen.c
src/cc65/codegen.h
g_getimmed Load a constant into the primary register
g_getstatic Fetch an static memory cell into the primary register
g_getlocal Fetch specified local object (local var) into the primary register
g_putstatic
g_reglong
g_regfloat Make sure, the value in the primary register a float. Convert if necessary
g_typeadjust
g_typecast Cast the value in the primary register to the operand size that is flagged by the lhs value
oper Encode a binary operation.
g_push
g_push_float Push the primary register or a constant value onto the stack
g_inc
g_dec
g_defdata
g_defdata_float
(see below) g_add, g_sub, g_mul, g_div, g_neg, g_bneg
(see below) g_eq, g_ne, g_lt, g_gt, g_le, g_ge
(invalid) g_mod, g_or, g_xor, g_and, g_asr, g_asl
datatype.c
src/cc65/datatype.c
ArithmeticConvert Perform the usual arithmetic conversions for binary operators.
expr.c
src/cc65/expr.c
LimitExprValue
parseadd
parsesub
fp.c
src/common/fp.c (OK)
src/common/fp.h (OK)
wrapper for doing floating point operations on target floats
FP_D_As32bitRaw converts double into 32bit (float) and then returns its raw content as a 32bit int
initdata.c
src/cc65/initdata.c
DefineData Output a data definition for the given expression
loadexpr.c
src/cc65/loadexpr.c
LoadExpr Load an expression into the primary register if it is not already there.
locals.c
src/cc65/locals.c (OK?)
ParseAutoDecl Parse the declarator of an auto variable.
scanner.c
src/cc65/scanner.c (OK?)
NumericConst Parse a numeric constant
Note: Scanner fixes should be directly promoted to upstream. Any differences in this branch should be related to debugging/logging.
typeconv.c
src/cc65/typeconv.c (OK?)
DoConversion Emit code to convert the given expression to a new type
The Library
variants
The floating point support calls library functions for any operations on non constant values.
softfloat
This is a Port of "Berkeley SoftFloat Release 2c". It is currently used by default for all targets.
softmath
Contains a collection of math functions, (hopefully) enough to completely implement math.h in C. This is currently used by default for all targets.
cbmkernal
This is a wrapper to the CBM kernal functions. This is fairly complete, including math functions. To use this, link against c64-fp754kernal.o. The c64 samples will do this by default.
ieee754
This should become a freestanding IEEE754 library, which can completely replace the softfloat (and softmath) library.
- basically everything missing except addition/substraction
- compare functions are missing
- mul, div functions are missing
- type conversion functions are missing
runtime functions
These must be available in the runtime library.
func description softfloat cbmfp 754 codegen.c
aufloat Primary 8bit unsigned -> float - * - g_regfloat
afloat Primary 8bit signed -> float - * - g_regfloat
axufloat Primary 16bit unsigned -> float * * - g_regfloat
axfloat Primary 16bit signed -> float * * - g_regfloat
eaxufloat Primary 32bit unsigned -> float * * - g_regfloat
eaxfloat Primary 32bit signed -> float * * - g_regfloat
feaxint Primary float -> 16bit int - * - g_regint
feaxlong Primary float -> 32bit long - * - g_reglong
ftosaddeax Primary = TOS + Primary * * ? g_add
ftossubeax Primary = TOS - Primary * * ? g_sub
ftosrsubeax Primary = Primary - TOS * * - g_rsub
ftosmuleax Primary = TOS * Primary * * - g_mul
ftosdiveax Primary = TOS / Primary * * - g_div
fnegeax Primary = -Primary * * - g_neg
fbnegeax Primary = !Primary (return bool!) * * - g_bneg
ftosgeeax Test for greater than or equal to * * - g_ge
ftosgteax Test for greater than * * - g_gt
ftosleeax Test for less than or equal to * * - g_le
ftoslteax Test for less than * * - g_lt
ftosneeax Test for not equal * * - g_ne
ftoseqeax Test for equal * * - g_eq
math.h functions
These are optional, required for standard libm.
func description softmath cbmfp 754
float powf(float f, float a) * * -
float sinf(float s) * * -
float cosf(float s) * * -
float logf(float x) * * -
float expf(float x) * * -
float sqrtf(float x) * * -
float tanf(float x) * * -
float atanf(float x) * * -
float fabsf(float x) * * -
float roundf(float x) * * -
float truncf(float x) * * -
float ceilf(float x) * - -
float floorf(float x) * - -
float fmodf(float x, float y) * - -
extra functions
Optional utility functions.
func description softfloat cbmfp 754
char *_ftostr(char *d, float s) * * ? for printf family
float _strtof(char *d) - * - for scanf family