mirror of
https://github.com/uffejakobsen/acme.git
synced 2024-11-26 15:49:18 +00:00
ffabea98c1
git-svn-id: https://svn.code.sf.net/p/acme-crossass/code-0/trunk@17 4df02467-bbd4-4a76-a152-e7ce94205b78
88 lines
2.9 KiB
Plaintext
88 lines
2.9 KiB
Plaintext
;ACME 0.94.5
|
|
|
|
!ifdef lib_cbm_mflpt_a !eof
|
|
lib_cbm_mflpt_a = 1
|
|
|
|
; CAUTION! The Commodore BASIC interpreter uses two different formats for
|
|
; handling floating-point values, so do not confuse them:
|
|
; The "float registers" fac1 and fac2 (actually structures in zero page) use a
|
|
; six-byte format commonly known as "flpt" (floating point).
|
|
; When storing values in variables (or reading values from ROM), a compressed
|
|
; five-byte format is used, commonly known as "mflpt" (memory floating point).
|
|
|
|
; This file contains a macro for writing floating point numbers in the five-byte
|
|
; "mflpt" format, where the sign bit is packed into the mantissa.
|
|
; Several interpreter functions use this format (see <cbm/c64/float.a>).
|
|
|
|
; Use the macro like this:
|
|
; +mflpt 3.1415926 ; each use will take up five bytes of memory
|
|
|
|
|
|
; now for the technical stuff (stop reading right now if you value your sanity)
|
|
|
|
; five-byte layout in memory:
|
|
; eeeeeeee smmmmmmm mmmmmmmm mmmmmmmm mmmmmmmm ; eight bits exponent, 32 bits mantissa with sign bit overlay
|
|
|
|
; exponent byte:
|
|
; exponent has a bias of 128 (128 means the decimal point is right before the mantissa's leading digit)
|
|
; if exponent is zero, number value is considered to be zero, regardless of mantissa
|
|
; exponents 1..128 are for values < 1
|
|
; exponents 129..255 are for values >= 1
|
|
|
|
; mantissa:
|
|
; mantissa is stored big-endian(!)
|
|
; the mantissa's mandatory leading '1' is replaced by the sign bit
|
|
|
|
; so logically, this is equivalent to:
|
|
; + .1mmmmmmm mmmmmmmm mmmmmmmm mmmmmmmm * 2^(eeeeeeee - 128) if sign bit is 0
|
|
; - .1mmmmmmm mmmmmmmm mmmmmmmm mmmmmmmm * 2^(eeeeeeee - 128) if sign bit is 1
|
|
|
|
|
|
|
|
; this is ugly, but it gets the job done
|
|
; (if it's stupid, but it works, then it's not stupid)
|
|
!macro mflpt .value {
|
|
!set .float = float(.value) ; make sure to do passes until value is defined
|
|
!ifndef .float {
|
|
!by $ff, $ff, $ff, $ff, $ff ; five place holder bytes
|
|
} else {
|
|
; value is defined, so split up into sign and non-negative value
|
|
!if .float < 0 {
|
|
!set .sign = $80
|
|
!set .float = -.float
|
|
} else {
|
|
!set .sign = $00
|
|
}
|
|
!if .float = 0 {
|
|
!by 0, 0, 0, 0, 0 ; five zeroes (zero is represented by all bits zero)
|
|
} else {
|
|
; split up into exponent and mantissa
|
|
!set .exponent = 128 + 32 ; 128 is cbm's bias, 32 is this algo's bias
|
|
; if mantissa is too large, shift right and adjust exponent
|
|
!do while .float >= (2.0 ^ 32.0) {
|
|
!set .float = .float >> 1
|
|
!set .exponent = .exponent + 1
|
|
}
|
|
; if mantissa is too small, shift left and adjust exponent
|
|
!do while .float < (2.0 ^ 31.0) {
|
|
!set .float = .float << 1
|
|
!set .exponent = .exponent - 1
|
|
}
|
|
!if .exponent < 1 {
|
|
!warn "MFLPT underflow, using zero instead"
|
|
!set .float = 0
|
|
!set .exponent = 0
|
|
!set .sign = 0
|
|
}
|
|
!if .exponent > 255 {
|
|
!error "MFLPT overflow"
|
|
}
|
|
!by .exponent
|
|
!by (127 & int(.float >> 24)) | .sign
|
|
!by 255 & int(.float >> 16)
|
|
!by 255 & int(.float >> 8)
|
|
!by 255 & int(.float)
|
|
}
|
|
}
|
|
}
|