mirror of
https://github.com/autc04/Retro68.git
synced 2024-11-20 10:42:21 +00:00
609 lines
12 KiB
ArmAsm
609 lines
12 KiB
ArmAsm
|
; SF format is:
|
||
|
;
|
||
|
; [sign] 1.[23bits] E[8bits(n-127)]
|
||
|
;
|
||
|
; SEEEEEEE Emmmmmmm mmmmmmmm mmmmmmmm
|
||
|
;
|
||
|
; [A+0] mmmmmmmm
|
||
|
; [A+1] mmmmmmmm
|
||
|
; [A+2] Emmmmmmm
|
||
|
; [A+3] SEEEEEEE
|
||
|
;
|
||
|
; Special values (xxx != 0):
|
||
|
;
|
||
|
; s1111111 10000000 00000000 00000000 infinity
|
||
|
; s1111111 1xxxxxxx xxxxxxxx xxxxxxxx NaN
|
||
|
; s0000000 00000000 00000000 00000000 zero
|
||
|
; s0000000 0xxxxxxx xxxxxxxx xxxxxxxx denormals
|
||
|
;
|
||
|
; Note that CMPtype is "signed char" for rl78
|
||
|
;
|
||
|
|
||
|
#include "vregs.h"
|
||
|
|
||
|
#define Z PSW.6
|
||
|
|
||
|
START_FUNC ___negsf2
|
||
|
|
||
|
;; Negate the floating point value.
|
||
|
;; Input at [SP+4]..[SP+7].
|
||
|
;; Output to R8..R11.
|
||
|
|
||
|
movw ax, [SP+4]
|
||
|
movw r8, ax
|
||
|
movw ax, [SP+6]
|
||
|
xor a, #0x80
|
||
|
movw r10, ax
|
||
|
ret
|
||
|
|
||
|
END_FUNC ___negsf2
|
||
|
|
||
|
;; ------------------internal functions used by later code --------------
|
||
|
|
||
|
START_FUNC __int_isnan
|
||
|
|
||
|
;; [HL] points to value, returns Z if it's a NaN
|
||
|
|
||
|
mov a, [hl+2]
|
||
|
and a, #0x80
|
||
|
mov x, a
|
||
|
mov a, [hl+3]
|
||
|
and a, #0x7f
|
||
|
cmpw ax, #0x7f80
|
||
|
skz
|
||
|
ret ; return NZ if not NaN
|
||
|
mov a, [hl+2]
|
||
|
and a, #0x7f
|
||
|
or a, [hl+1]
|
||
|
or a, [hl]
|
||
|
bnz $1f
|
||
|
clr1 Z ; Z, normal
|
||
|
ret
|
||
|
1:
|
||
|
set1 Z ; nan
|
||
|
ret
|
||
|
|
||
|
END_FUNC __int_isnan
|
||
|
|
||
|
START_FUNC __int_eithernan
|
||
|
|
||
|
;; call from toplevel functions, returns Z if either number is a NaN,
|
||
|
;; or NZ if both are OK.
|
||
|
|
||
|
movw ax, sp
|
||
|
addw ax, #8
|
||
|
movw hl, ax
|
||
|
call $!__int_isnan
|
||
|
bz $1f
|
||
|
|
||
|
movw ax, sp
|
||
|
addw ax, #12
|
||
|
movw hl, ax
|
||
|
call $!__int_isnan
|
||
|
1:
|
||
|
ret
|
||
|
|
||
|
END_FUNC __int_eithernan
|
||
|
|
||
|
START_FUNC __int_iszero
|
||
|
|
||
|
;; [HL] points to value, returns Z if it's zero
|
||
|
|
||
|
mov a, [hl+3]
|
||
|
and a, #0x7f
|
||
|
or a, [hl+2]
|
||
|
or a, [hl+1]
|
||
|
or a, [hl]
|
||
|
ret
|
||
|
|
||
|
END_FUNC __int_iszero
|
||
|
|
||
|
START_FUNC __int_cmpsf
|
||
|
|
||
|
;; This is always called from some other function here,
|
||
|
;; so the stack offsets are adjusted accordingly.
|
||
|
|
||
|
;; X [SP+8] <=> Y [SP+12] : <a> <=> 0
|
||
|
|
||
|
movw ax, sp
|
||
|
addw ax, #8
|
||
|
movw hl, ax
|
||
|
call $!__int_iszero
|
||
|
bnz $1f
|
||
|
|
||
|
movw ax, sp
|
||
|
addw ax, #12
|
||
|
movw hl, ax
|
||
|
call $!__int_iszero
|
||
|
bnz $2f
|
||
|
;; At this point, both args are zero.
|
||
|
mov a, #0
|
||
|
ret
|
||
|
|
||
|
2:
|
||
|
movw ax, sp
|
||
|
addw ax, #8
|
||
|
movw hl, ax
|
||
|
1:
|
||
|
;; At least one arg is non-zero so we can just compare magnitudes.
|
||
|
;; Args are [HL] and [HL+4].
|
||
|
|
||
|
mov a, [HL+3]
|
||
|
xor a, [HL+7]
|
||
|
mov1 cy, a.7
|
||
|
bnc $1f
|
||
|
|
||
|
mov a, [HL+3]
|
||
|
sar a, 7
|
||
|
or a, #1
|
||
|
ret
|
||
|
|
||
|
1: ;; Signs the same, compare magnitude. It's safe to lump
|
||
|
;; the sign bits, exponent, and mantissa together here, since they're
|
||
|
;; stored in the right sequence.
|
||
|
movw ax, [HL+2]
|
||
|
cmpw ax, [HL+6]
|
||
|
bc $ybig_cmpsf ; branch if X < Y
|
||
|
bnz $xbig_cmpsf ; branch if X > Y
|
||
|
|
||
|
movw ax, [HL]
|
||
|
cmpw ax, [HL+4]
|
||
|
bc $ybig_cmpsf ; branch if X < Y
|
||
|
bnz $xbig_cmpsf ; branch if X > Y
|
||
|
|
||
|
mov a, #0
|
||
|
ret
|
||
|
|
||
|
xbig_cmpsf: ; |X| > |Y| so return A = 1 if pos, 0xff if neg
|
||
|
mov a, [HL+3]
|
||
|
sar a, 7
|
||
|
or a, #1
|
||
|
ret
|
||
|
ybig_cmpsf: ; |X| < |Y| so return A = 0xff if pos, 1 if neg
|
||
|
mov a, [HL+3]
|
||
|
xor a, #0x80
|
||
|
sar a, 7
|
||
|
or a, #1
|
||
|
ret
|
||
|
|
||
|
END_FUNC __int_cmpsf
|
||
|
|
||
|
;; ----------------------------------------------------------
|
||
|
|
||
|
START_FUNC ___cmpsf2
|
||
|
;; This functions calculates "A <=> B". That is, if A is less than B
|
||
|
;; they return -1, if A is greater than B, they return 1, and if A
|
||
|
;; and B are equal they return 0. If either argument is NaN the
|
||
|
;; behaviour is undefined.
|
||
|
|
||
|
;; Input at [SP+4]..[SP+7].
|
||
|
;; Output to R8..R9.
|
||
|
|
||
|
call $!__int_eithernan
|
||
|
bnz $1f
|
||
|
movw r8, #1
|
||
|
ret
|
||
|
1:
|
||
|
call $!__int_cmpsf
|
||
|
mov r8, a
|
||
|
sar a, 7
|
||
|
mov r9, a
|
||
|
ret
|
||
|
|
||
|
END_FUNC ___cmpsf2
|
||
|
|
||
|
;; ----------------------------------------------------------
|
||
|
|
||
|
;; These functions are all basically the same as ___cmpsf2
|
||
|
;; except that they define how they handle NaNs.
|
||
|
|
||
|
START_FUNC ___eqsf2
|
||
|
;; Returns zero iff neither argument is NaN
|
||
|
;; and both arguments are equal.
|
||
|
START_ANOTHER_FUNC ___nesf2
|
||
|
;; Returns non-zero iff either argument is NaN or the arguments are
|
||
|
;; unequal. Effectively __nesf2 is the same as __eqsf2
|
||
|
START_ANOTHER_FUNC ___lesf2
|
||
|
;; Returns a value less than or equal to zero if neither
|
||
|
;; argument is NaN, and the first is less than or equal to the second.
|
||
|
START_ANOTHER_FUNC ___ltsf2
|
||
|
;; Returns a value less than zero if neither argument is
|
||
|
;; NaN, and the first is strictly less than the second.
|
||
|
|
||
|
;; Input at [SP+4]..[SP+7].
|
||
|
;; Output to R8.
|
||
|
|
||
|
mov r8, #1
|
||
|
|
||
|
;;; Fall through
|
||
|
|
||
|
START_ANOTHER_FUNC __int_cmp_common
|
||
|
|
||
|
call $!__int_eithernan
|
||
|
sknz
|
||
|
;; return value (pre-filled-in below) for "either is nan"
|
||
|
ret
|
||
|
|
||
|
call $!__int_cmpsf
|
||
|
mov r8, a
|
||
|
ret
|
||
|
|
||
|
END_ANOTHER_FUNC __int_cmp_common
|
||
|
END_ANOTHER_FUNC ___ltsf2
|
||
|
END_ANOTHER_FUNC ___lesf2
|
||
|
END_ANOTHER_FUNC ___nesf2
|
||
|
END_FUNC ___eqsf2
|
||
|
|
||
|
START_FUNC ___gesf2
|
||
|
;; Returns a value greater than or equal to zero if neither argument
|
||
|
;; is a NaN and the first is greater than or equal to the second.
|
||
|
START_ANOTHER_FUNC ___gtsf2
|
||
|
;; Returns a value greater than zero if neither argument
|
||
|
;; is NaN, and the first is strictly greater than the second.
|
||
|
|
||
|
mov r8, #0xffff
|
||
|
br $__int_cmp_common
|
||
|
|
||
|
END_ANOTHER_FUNC ___gtsf2
|
||
|
END_FUNC ___gesf2
|
||
|
|
||
|
;; ----------------------------------------------------------
|
||
|
|
||
|
START_FUNC ___unordsf2
|
||
|
;; Returns a nonzero value if either argument is NaN, otherwise 0.
|
||
|
|
||
|
call $!__int_eithernan
|
||
|
movw r8, #0
|
||
|
sknz ; this is from the call, not the movw
|
||
|
movw r8, #1
|
||
|
ret
|
||
|
|
||
|
END_FUNC ___unordsf2
|
||
|
|
||
|
;; ----------------------------------------------------------
|
||
|
|
||
|
START_FUNC ___fixsfsi
|
||
|
;; Converts its floating point argument into a signed long,
|
||
|
;; rounding toward zero.
|
||
|
;; The behaviour with NaNs and Infinities is not well defined.
|
||
|
;; We choose to return 0 for NaNs, -INTMAX for -inf and INTMAX for +inf.
|
||
|
;; This matches the behaviour of the C function in libgcc2.c.
|
||
|
|
||
|
;; Input at [SP+4]..[SP+7], result is in (lsb) R8..R11 (msb).
|
||
|
|
||
|
;; Special case handling for infinities as __fixunssfsi
|
||
|
;; will not give us the values that we want.
|
||
|
movw ax, sp
|
||
|
addw ax, #4
|
||
|
movw hl, ax
|
||
|
call !!__int_isinf
|
||
|
bnz $1f
|
||
|
mov a, [SP+7]
|
||
|
bt a.7, $2f
|
||
|
;; +inf
|
||
|
movw r8, #-1
|
||
|
movw r10, #0x7fff
|
||
|
ret
|
||
|
;; -inf
|
||
|
2: mov r8, #0
|
||
|
mov r10, #0x8000
|
||
|
ret
|
||
|
|
||
|
;; Load the value into r10:r11:X:A
|
||
|
1: movw ax, [SP+4]
|
||
|
movw r10, ax
|
||
|
movw ax, [SP+6]
|
||
|
|
||
|
;; If the value is positive we can just use __fixunssfsi
|
||
|
bf a.7, $__int_fixunssfsi
|
||
|
|
||
|
;; Otherwise we negate the value, call __fixunssfsi and
|
||
|
;; then negate its result.
|
||
|
clr1 a.7
|
||
|
call $!__int_fixunssfsi
|
||
|
|
||
|
movw ax, #0
|
||
|
subw ax, r8
|
||
|
movw r8, ax
|
||
|
movw ax, #0
|
||
|
sknc
|
||
|
decw ax
|
||
|
subw ax, r10
|
||
|
movw r10, ax
|
||
|
|
||
|
;; Check for a positive result (which should only happen when
|
||
|
;; __fixunssfsi returns UINTMAX or 0). In such cases just return 0.
|
||
|
mov a, r11
|
||
|
bt a.7, $1f
|
||
|
movw r10,#0x0
|
||
|
movw r8, #0x0
|
||
|
|
||
|
1: ret
|
||
|
|
||
|
END_FUNC ___fixsfsi
|
||
|
|
||
|
START_FUNC ___fixunssfsi
|
||
|
;; Converts its floating point argument into an unsigned long
|
||
|
;; rounding towards zero. Negative arguments all become zero.
|
||
|
;; We choose to return 0 for NaNs and -inf, but UINTMAX for +inf.
|
||
|
;; This matches the behaviour of the C function in libgcc2.c.
|
||
|
|
||
|
;; Input at [SP+4]..[SP+7], result is in (lsb) R8..R11 (msb)
|
||
|
|
||
|
;; Get the input value.
|
||
|
movw ax, [SP+4]
|
||
|
movw r10, ax
|
||
|
movw ax, [SP+6]
|
||
|
|
||
|
;; Fall through into the internal function.
|
||
|
|
||
|
.global __int_fixunssfsi
|
||
|
__int_fixunssfsi:
|
||
|
;; Input in (lsb) r10.r11.x.a (msb).
|
||
|
|
||
|
;; Test for a negative input. We shift the other bits at the
|
||
|
;; same time so that A ends up holding the whole exponent:
|
||
|
;;
|
||
|
;; before:
|
||
|
;; SEEEEEEE EMMMMMMM MMMMMMMM MMMMMMMM
|
||
|
;; A X R11 R10
|
||
|
;;
|
||
|
;; after:
|
||
|
;; EEEEEEEE MMMMMMM0 MMMMMMMM MMMMMMMM
|
||
|
;; A X R11 R10
|
||
|
shlw ax, 1
|
||
|
bnc $1f
|
||
|
|
||
|
;; Return zero.
|
||
|
2: movw r8, #0
|
||
|
movw r10, #0
|
||
|
ret
|
||
|
|
||
|
;; An exponent of -1 is either a NaN or infinity.
|
||
|
1: cmp a, #-1
|
||
|
bnz $3f
|
||
|
;; For NaN we return 0. For infinity we return UINTMAX.
|
||
|
mov a, x
|
||
|
or a, r10
|
||
|
or a, r11
|
||
|
cmp0 a
|
||
|
bnz $2b
|
||
|
|
||
|
6: movw r8, #-1 ; -1 => UINT_MAX
|
||
|
movw r10, #-1
|
||
|
ret
|
||
|
|
||
|
;; If the exponent is negative the value is < 1 and so the
|
||
|
;; converted value is 0. Note we must allow for the bias
|
||
|
;; applied to the exponent. Thus a value of 127 in the
|
||
|
;; EEEEEEEE bits actually represents an exponent of 0, whilst
|
||
|
;; a value less than 127 actually represents a negative exponent.
|
||
|
;; Also if the EEEEEEEE bits are all zero then this represents
|
||
|
;; either a denormal value or 0.0. Either way for these values
|
||
|
;; we return 0.
|
||
|
3: sub a, #127
|
||
|
bc $2b
|
||
|
|
||
|
;; A now holds the bias adjusted exponent, which is known to be >= 0.
|
||
|
;; If the exponent is > 31 then the conversion will overflow.
|
||
|
cmp a, #32
|
||
|
bnc $6b
|
||
|
4:
|
||
|
;; Save the exponent in H. We increment it by one because we want
|
||
|
;; to be sure that the loop below will always execute at least once.
|
||
|
inc a
|
||
|
mov h, a
|
||
|
|
||
|
;; Get the top 24 bits of the mantissa into A:X:R10
|
||
|
;; Include the implicit 1-bit that is inherent in the IEEE fp format.
|
||
|
;;
|
||
|
;; before:
|
||
|
;; EEEEEEEE MMMMMMM0 MMMMMMMM MMMMMMMM
|
||
|
;; H X R11 R10
|
||
|
;; after:
|
||
|
;; EEEEEEEE 1MMMMMMM MMMMMMMM MMMMMMMM
|
||
|
;; H A X R10
|
||
|
|
||
|
mov a, r11
|
||
|
xch a, x
|
||
|
shr a, 1
|
||
|
set1 a.7
|
||
|
|
||
|
;; Clear B:C:R12:R13
|
||
|
movw bc, #0
|
||
|
movw r12, #0
|
||
|
|
||
|
;; Shift bits from the mantissa (A:X:R10) into (B:C:R12:R13),
|
||
|
;; decrementing the exponent as we go.
|
||
|
|
||
|
;; before:
|
||
|
;; MMMMMMMM MMMMMMMM MMMMMMMM xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx
|
||
|
;; A X R10 B C R12 R13
|
||
|
;; first iter:
|
||
|
;; MMMMMMMM MMMMMMMM MMMMMMM0 xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxM
|
||
|
;; A X R10 B C R12 R13
|
||
|
;; second iter:
|
||
|
;; MMMMMMMM MMMMMMMM MMMMMM00 xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxMM
|
||
|
;; A X R10 B C R12 R13
|
||
|
;; etc.
|
||
|
5:
|
||
|
xch a, r10
|
||
|
shl a, 1
|
||
|
xch a, r10
|
||
|
|
||
|
rolwc ax, 1
|
||
|
|
||
|
xch a, r13
|
||
|
rolc a, 1
|
||
|
xch a, r13
|
||
|
|
||
|
xch a, r12
|
||
|
rolc a, 1
|
||
|
xch a, r12
|
||
|
|
||
|
rolwc bc, 1
|
||
|
|
||
|
dec h
|
||
|
bnz $5b
|
||
|
|
||
|
;; Result is currently in (lsb) r13.r12. c. b. (msb),
|
||
|
;; Move it into (lsb) r8. r9. r10. r11 (msb).
|
||
|
|
||
|
mov a, r13
|
||
|
mov r8, a
|
||
|
|
||
|
mov a, r12
|
||
|
mov r9, a
|
||
|
|
||
|
mov a, c
|
||
|
mov r10, a
|
||
|
|
||
|
mov a, b
|
||
|
mov r11, a
|
||
|
|
||
|
ret
|
||
|
|
||
|
END_FUNC ___fixunssfsi
|
||
|
|
||
|
;; ------------------------------------------------------------------------
|
||
|
|
||
|
START_FUNC ___floatsisf
|
||
|
;; Converts its signed long argument into a floating point.
|
||
|
;; Argument in [SP+4]..[SP+7]. Result in R8..R11.
|
||
|
|
||
|
;; Get the argument.
|
||
|
movw ax, [SP+4]
|
||
|
movw bc, ax
|
||
|
movw ax, [SP+6]
|
||
|
|
||
|
;; Test the sign bit. If the value is positive then drop into
|
||
|
;; the unsigned conversion routine.
|
||
|
bf a.7, $2f
|
||
|
|
||
|
;; If negative convert to positive ...
|
||
|
movw hl, ax
|
||
|
movw ax, #0
|
||
|
subw ax, bc
|
||
|
movw bc, ax
|
||
|
movw ax, #0
|
||
|
sknc
|
||
|
decw ax
|
||
|
subw ax, hl
|
||
|
|
||
|
;; If the result is negative then the input was 0x80000000 and
|
||
|
;; we want to return -0.0, which will not happen if we call
|
||
|
;; __int_floatunsisf.
|
||
|
bt a.7, $1f
|
||
|
|
||
|
;; Call the unsigned conversion routine.
|
||
|
call $!__int_floatunsisf
|
||
|
|
||
|
;; Negate the result.
|
||
|
set1 r11.7
|
||
|
|
||
|
;; Done.
|
||
|
ret
|
||
|
|
||
|
1: ;; Return -0.0 aka 0xcf000000
|
||
|
|
||
|
clrb a
|
||
|
mov r8, a
|
||
|
mov r9, a
|
||
|
mov r10, a
|
||
|
mov a, #0xcf
|
||
|
mov r11, a
|
||
|
ret
|
||
|
|
||
|
START_ANOTHER_FUNC ___floatunsisf
|
||
|
;; Converts its unsigned long argument into a floating point.
|
||
|
;; Argument in [SP+4]..[SP+7]. Result in R8..R11.
|
||
|
|
||
|
;; Get the argument.
|
||
|
movw ax, [SP+4]
|
||
|
movw bc, ax
|
||
|
movw ax, [SP+6]
|
||
|
|
||
|
2: ;; Internal entry point from __floatsisf
|
||
|
;; Input in AX (high) and BC (low)
|
||
|
.global __int_floatunsisf
|
||
|
__int_floatunsisf:
|
||
|
|
||
|
;; Special case handling for zero.
|
||
|
cmpw ax, #0
|
||
|
bnz $1f
|
||
|
movw ax, bc
|
||
|
cmpw ax, #0
|
||
|
movw ax, #0
|
||
|
bnz $1f
|
||
|
|
||
|
;; Return 0.0
|
||
|
movw r8, ax
|
||
|
movw r10, ax
|
||
|
ret
|
||
|
|
||
|
1: ;; Pre-load the loop count/exponent.
|
||
|
;; Exponents are biased by 0x80 and we start the loop knowing that
|
||
|
;; we are going to skip the highest set bit. Hence the highest value
|
||
|
;; that we can get for the exponent is 0x1e (bits from input) + 0x80 = 0x9e.
|
||
|
mov h, #0x9e
|
||
|
|
||
|
;; Move bits off the top of AX:BC until we hit a 1 bit.
|
||
|
;; Decrement the count of remaining bits as we go.
|
||
|
|
||
|
2: shlw bc, 1
|
||
|
rolwc ax, 1
|
||
|
bc $3f
|
||
|
dec h
|
||
|
br $2b
|
||
|
|
||
|
;; Ignore the first one bit - it is implicit in the IEEE format.
|
||
|
;; The count of remaining bits is the exponent.
|
||
|
|
||
|
;; Assemble the final floating point value. We have...
|
||
|
;; before:
|
||
|
;; EEEEEEEE MMMMMMMM MMMMMMMM MMMMMMMM xxxxxxxx
|
||
|
;; H A X B C
|
||
|
;; after:
|
||
|
;; 0EEEEEEE EMMMMMMM MMMMMMMM MMMMMMMM
|
||
|
;; R11 R10 R9 R8
|
||
|
|
||
|
|
||
|
3: shrw ax, 1
|
||
|
mov r10, a
|
||
|
mov a, x
|
||
|
mov r9, a
|
||
|
|
||
|
mov a, b
|
||
|
rorc a, 1
|
||
|
|
||
|
;; If the bottom bit of B was set before we shifted it out then we
|
||
|
;; need to round the result up. Unless none of the bits in C are set.
|
||
|
;; In this case we are exactly half-way between two values, and we
|
||
|
;; round towards an even value. We round up by increasing the
|
||
|
;; mantissa by 1. If this results in a zero mantissa we have to
|
||
|
;; increment the exponent. We round down by ignoring the dropped bits.
|
||
|
|
||
|
bnc $4f
|
||
|
cmp0 c
|
||
|
sknz
|
||
|
bf a.0, $4f
|
||
|
|
||
|
5: ;; Round the mantissa up by 1.
|
||
|
add a, #1
|
||
|
addc r9, #0
|
||
|
addc r10, #0
|
||
|
bf r10.7, $4f
|
||
|
inc h
|
||
|
clr1 r10.7
|
||
|
|
||
|
4: mov r8, a
|
||
|
mov a, h
|
||
|
shr a, 1
|
||
|
mov r11, a
|
||
|
sknc
|
||
|
set1 r10.7
|
||
|
ret
|
||
|
|
||
|
END_ANOTHER_FUNC ___floatunsisf
|
||
|
END_FUNC ___floatsisf
|