mirror of
https://github.com/hsoft/collapseos.git
synced 2024-11-30 20:18:06 +11:00
f9ae5ca46b
It always reads in the scratchpad with the same max size. No need for DE-dancing anymore.
1059 lines
27 KiB
NASM
1059 lines
27 KiB
NASM
; *** Consts ***
|
|
; Number of rows in the argspec table
|
|
ARGSPEC_TBL_CNT .equ 31
|
|
; Number of rows in the primary instructions table
|
|
INSTR_TBL_CNT .equ 135
|
|
; size in bytes of each row in the primary instructions table
|
|
INSTR_TBL_ROWSIZE .equ 6
|
|
; Instruction IDs They correspond to the index of the table in instrNames
|
|
I_ADC .equ 0x00
|
|
I_ADD .equ 0x01
|
|
I_AND .equ 0x02
|
|
I_BIT .equ 0x03
|
|
I_CALL .equ 0x04
|
|
I_CCF .equ 0x05
|
|
I_CP .equ 0x06
|
|
I_CPD .equ 0x07
|
|
I_CPDR .equ 0x08
|
|
I_CPI .equ 0x09
|
|
I_CPIR .equ 0x0a
|
|
I_CPL .equ 0x0b
|
|
I_DAA .equ 0x0c
|
|
I_DEC .equ 0x0d
|
|
I_DI .equ 0x0e
|
|
I_DJNZ .equ 0x0f
|
|
I_EI .equ 0x10
|
|
I_EX .equ 0x11
|
|
I_EXX .equ 0x12
|
|
I_HALT .equ 0x13
|
|
I_IM .equ 0x14
|
|
I_IN .equ 0x15
|
|
I_INC .equ 0x16
|
|
I_IND .equ 0x17
|
|
I_INDR .equ 0x18
|
|
I_INI .equ 0x19
|
|
I_INIR .equ 0x1a
|
|
I_JP .equ 0x1b
|
|
I_JR .equ 0x1c
|
|
I_LD .equ 0x1d
|
|
I_LDD .equ 0x1e
|
|
I_LDDR .equ 0x1f
|
|
I_LDI .equ 0x20
|
|
I_LDIR .equ 0x21
|
|
I_NEG .equ 0x22
|
|
I_NOP .equ 0x23
|
|
I_OR .equ 0x24
|
|
I_OTDR .equ 0x25
|
|
I_OTIR .equ 0x26
|
|
I_OUT .equ 0x27
|
|
I_POP .equ 0x28
|
|
I_PUSH .equ 0x29
|
|
I_RET .equ 0x2a
|
|
I_RLA .equ 0x2b
|
|
I_RLCA .equ 0x2c
|
|
I_RRA .equ 0x2d
|
|
I_RRCA .equ 0x2e
|
|
I_SBC .equ 0x2f
|
|
I_SCF .equ 0x30
|
|
I_SUB .equ 0x31
|
|
I_XOR .equ 0x32
|
|
|
|
; Checks whether A is 'N' or 'M'
|
|
checkNOrM:
|
|
cp 'N'
|
|
ret z
|
|
cp 'M'
|
|
ret
|
|
|
|
; Checks whether A is 'n', 'm', 'x' or 'y'
|
|
checknmxy:
|
|
cp 'n'
|
|
ret z
|
|
cp 'm'
|
|
ret z
|
|
cp 'x'
|
|
ret z
|
|
cp 'y'
|
|
ret
|
|
|
|
; Reads string in (HL) and returns the corresponding ID (I_*) in A. Sets Z if
|
|
; there's a match.
|
|
getInstID:
|
|
push bc
|
|
push de
|
|
ld b, I_XOR+1 ; I_XOR is the last
|
|
ld c, 4
|
|
ld de, instrNames
|
|
call findStringInList
|
|
pop de
|
|
pop bc
|
|
ret
|
|
|
|
; Parse the string at (HL) and check if it starts with IX+, IY+, IX- or IY-.
|
|
; Sets Z if yes, unset if no.
|
|
parseIXY:
|
|
push hl
|
|
ld a, (hl)
|
|
cp 'I'
|
|
jr nz, .end ; Z already unset
|
|
inc hl
|
|
ld a, (hl)
|
|
cp 'X'
|
|
jr z, .match1
|
|
cp 'Y'
|
|
jr z, .match1
|
|
jr .end ; Z already unset
|
|
.match1:
|
|
; Alright, we have IX or IY. Let's see if we have + or - next.
|
|
inc hl
|
|
ld a, (hl)
|
|
cp '+'
|
|
jr z, .end ; Z is already set
|
|
cp '-'
|
|
; The value of Z at this point is our final result
|
|
.end:
|
|
pop hl
|
|
ret
|
|
|
|
; Returns length of string at (HL) in A.
|
|
strlen:
|
|
push bc
|
|
push hl
|
|
ld bc, 0
|
|
ld a, 0 ; look for null char
|
|
.loop:
|
|
cpi
|
|
jp z, .found
|
|
jr .loop
|
|
.found:
|
|
; How many char do we have? the (NEG BC)-1, which started at 0 and
|
|
; decreased at each CPI call. In this routine, we stay in the 8-bit
|
|
; realm, so C only.
|
|
ld a, c
|
|
neg
|
|
dec a
|
|
pop hl
|
|
pop bc
|
|
ret
|
|
|
|
; find argspec for string at (HL). Returns matching argspec in A.
|
|
; Return value 0xff holds a special meaning: arg is not empty, but doesn't match
|
|
; any argspec (A == 0 means arg is empty). A return value of 0xff means an
|
|
; error.
|
|
;
|
|
; If the parsed argument is a number constant, 'N' is returned and IX contains
|
|
; the value of that constant.
|
|
parseArg:
|
|
call strlen
|
|
cp 0
|
|
ret z ; empty string? A already has our result: 0
|
|
|
|
push bc
|
|
push de
|
|
push hl
|
|
|
|
; We always initialize IX to zero so that non-numerical args end up with
|
|
; a clean zero.
|
|
ld ix, 0
|
|
|
|
ld de, argspecTbl
|
|
; DE now points the the "argspec char" part of the entry, but what
|
|
; we're comparing in the loop is the string next to it. Let's offset
|
|
; DE by one so that the loop goes through strings.
|
|
inc de
|
|
ld b, ARGSPEC_TBL_CNT
|
|
.loop1:
|
|
ld a, 4
|
|
call JUMP_STRNCMP
|
|
jr z, .found ; got it!
|
|
ld a, 5
|
|
call JUMP_ADDDE
|
|
djnz .loop1
|
|
|
|
; We exhausted the argspecs. Let's see if we're inside parens.
|
|
call enterParens
|
|
jr z, .withParens
|
|
; (HL) has no parens
|
|
call parseNumber
|
|
jr nz, .nomatch
|
|
; We have a proper number in no parens. Number in IX.
|
|
ld a, 'N'
|
|
jr .end
|
|
.withParens:
|
|
ld c, 'M' ; C holds the argspec type until we reach
|
|
; .numberInParens
|
|
; We have parens. First, let's see if we have a (IX+d) type of arg.
|
|
call parseIXY
|
|
jr nz, .parseNumberInParens ; not I{X,Y}. just parse number.
|
|
; We have IX+/IY+/IX-/IY-.
|
|
; note: the "-" part isn't supported yet.
|
|
inc hl ; (HL) now points to X or Y
|
|
ld a, (hl)
|
|
inc hl ; advance HL to the number part
|
|
inc hl ; this is the number
|
|
cp 'Y'
|
|
jr nz, .notY
|
|
ld c, 'y'
|
|
jr .parseNumberInParens
|
|
.notY:
|
|
ld c, 'x'
|
|
.parseNumberInParens:
|
|
call parseNumber
|
|
jr nz, .nomatch
|
|
; We have a proper number in parens. Number in IX
|
|
ld a, c ; M, x, or y
|
|
jr .end
|
|
.nomatch:
|
|
; We get no match
|
|
ld a, 0xff
|
|
jr .end
|
|
.found:
|
|
; found the matching argspec row. Our result is one byte left of DE.
|
|
dec de
|
|
ld a, (de)
|
|
.end:
|
|
pop hl
|
|
pop de
|
|
pop bc
|
|
ret
|
|
|
|
; Returns, with Z, whether A is a groupId
|
|
isGroupId:
|
|
cp 0xc ; max group id + 1
|
|
jr nc, .notgroup ; >= 0xc? not a group
|
|
cp 0
|
|
jr z, .notgroup ; 0? not supposed to happen. something's wrong.
|
|
; A is a group. ensure Z is set
|
|
cp a
|
|
ret
|
|
.notgroup:
|
|
call JUMP_UNSETZ
|
|
ret
|
|
|
|
; Find argspec A in group id H.
|
|
; Set Z according to whether we found the argspec
|
|
; If found, the value in A is the argspec value in the group (its index).
|
|
findInGroup:
|
|
push bc
|
|
push hl
|
|
|
|
cp 0 ; is our arg empty? If yes, we have nothing to do
|
|
jr z, .notfound
|
|
|
|
push af
|
|
ld a, h
|
|
cp 0xa
|
|
jr z, .specialGroupCC
|
|
cp 0xb
|
|
jr z, .specialGroupABCDEHL
|
|
jr nc, .notfound ; > 0xb? not a group
|
|
pop af
|
|
; regular group
|
|
push de
|
|
ld de, argGrpTbl
|
|
; group ids start at 1. decrease it, then multiply by 4 to have a
|
|
; proper offset in argGrpTbl
|
|
dec h
|
|
push af
|
|
ld a, h
|
|
rla
|
|
rla
|
|
call JUMP_ADDDE ; At this point, DE points to our group
|
|
pop af
|
|
ex hl, de ; And now, HL points to the group
|
|
pop de
|
|
|
|
ld bc, 4
|
|
jr .find
|
|
|
|
.specialGroupCC:
|
|
ld hl, argGrpCC
|
|
jr .specialGroupEnd
|
|
.specialGroupABCDEHL:
|
|
ld hl, argGrpABCDEHL
|
|
.specialGroupEnd:
|
|
pop af ; from the push af just before the special group check
|
|
ld bc, 8
|
|
|
|
.find:
|
|
; This part is common to regular and special group. We expect HL to
|
|
; point to the group and BC to contain its length.
|
|
push bc ; save the start value loop index so we can sub
|
|
.loop:
|
|
cpi
|
|
jr z, .found
|
|
jp po, .notfound
|
|
jr .loop
|
|
.found:
|
|
; we found our result! Now, what we want to put in A is the index of
|
|
; the found argspec.
|
|
pop hl ; we pop from the "push bc" above. L is now 4 or 8
|
|
ld a, l
|
|
sub c
|
|
dec a ; cpi DECs BC even when there's a match, so C == the
|
|
; number of iterations we've made. But our index is
|
|
; zero-based (1 iteration == 0 index).
|
|
cp a ; ensure Z is set
|
|
jr .end
|
|
.notfound:
|
|
pop bc ; from the push bc in .find
|
|
call JUMP_UNSETZ
|
|
.end:
|
|
pop hl
|
|
pop bc
|
|
ret
|
|
|
|
; Compare argspec from instruction table in A with argument in (HL).
|
|
; For constant args, it's easy: if A == (HL), it's a success.
|
|
; If it's not this, then we check if it's a numerical arg.
|
|
; If A is a group ID, we do something else: we check that (HL) exists in the
|
|
; groupspec (argGrpTbl). Moreover, we go and write the group's "value" (index)
|
|
; in (HL+1). This will save us significant processing later in getUpcode.
|
|
; Set Z according to whether we match or not.
|
|
matchArg:
|
|
cp a, (hl)
|
|
ret z
|
|
; not an exact match. Before we continue: is A zero? Because if it is,
|
|
; we have to stop right here: no match possible.
|
|
cp 0
|
|
jr nz, .checkIfNumber ; not a zero, we can continue
|
|
; zero, stop here
|
|
call JUMP_UNSETZ
|
|
ret
|
|
.checkIfNumber:
|
|
; not an exact match, let's check for numerical constants.
|
|
call JUMP_UPCASE
|
|
call checkNOrM
|
|
jr z, .expectsNumber
|
|
jr .notNumber
|
|
.expectsNumber:
|
|
; Our argument is a number N or M. Never a lower-case version. At this
|
|
; point in the processing, we don't care about whether N or M is upper,
|
|
; we do truncation tests later. So, let's just perform the same == test
|
|
; but in a case-insensitive way instead
|
|
cp a, (hl)
|
|
ret ; whether we match or not, the result of Z is
|
|
; the good one.
|
|
.notNumber:
|
|
; A bit of a delicate situation here: we want A to go in H but also
|
|
; (HL) to go in A. If not careful, we overwrite each other. EXX is
|
|
; necessary to avoid invoving other registers.
|
|
push hl
|
|
exx
|
|
ld h, a
|
|
push hl
|
|
exx
|
|
ld a, (hl)
|
|
pop hl
|
|
call findInGroup
|
|
pop hl
|
|
ret nz
|
|
; we found our group? let's write down its "value" in (HL+1). We hold
|
|
; this value in A at the moment.
|
|
inc hl
|
|
ld (hl), a
|
|
dec hl
|
|
ret
|
|
|
|
; Compare primary row at (DE) with ID in A. Sets Z flag if there's a match.
|
|
matchPrimaryRow:
|
|
push hl
|
|
push ix
|
|
ld ixh, d
|
|
ld ixl, e
|
|
cp (ix)
|
|
jr nz, .end
|
|
; name matches, let's see the rest
|
|
ld hl, curArg1
|
|
ld a, (ix+1)
|
|
call matchArg
|
|
jr nz, .end
|
|
ld hl, curArg2
|
|
ld a, (ix+2)
|
|
call matchArg
|
|
.end:
|
|
pop ix
|
|
pop hl
|
|
ret
|
|
|
|
; *** Special opcodes ***
|
|
; The special upcode handling routines below all have the same signature.
|
|
; Instruction row is at IX and we're expected to perform the same task as
|
|
; getUpcode. The number of bytes, however, must go in C instead of A
|
|
; No need to preserve HL, DE, BC and IX: it's handled by getUpcode already.
|
|
|
|
; Handle like a regular "JP (IX+d)" except that we refuse any displacement: if
|
|
; a displacement is specified, we error out.
|
|
handleJPIX:
|
|
ld a, 0xdd
|
|
jr handleJPIXY
|
|
handleJPIY:
|
|
ld a, 0xfd
|
|
handleJPIXY:
|
|
ld (instrUpcode), a
|
|
ld a, (curArg1+1)
|
|
cp 0 ; numerical argument *must* be zero
|
|
jr nz, .error
|
|
; ok, we're good
|
|
ld a, 0xe9 ; second upcode
|
|
ld (instrUpcode+1), a
|
|
ld c, 2
|
|
ret
|
|
.error:
|
|
xor c
|
|
ret
|
|
|
|
; Handle the first argument of BIT. Sets Z if first argument is valid, unset it
|
|
; if there's an error.
|
|
handleBIT:
|
|
ld a, (curArg1+1)
|
|
cp 8
|
|
jr nc, .error ; >= 8? error
|
|
; We're good
|
|
cp a ; ensure Z
|
|
ret
|
|
.error:
|
|
xor c
|
|
call JUMP_UNSETZ
|
|
ret
|
|
|
|
handleBITHL:
|
|
call handleBIT
|
|
ret nz ; error
|
|
ld a, 0xcb ; first upcode
|
|
ld (instrUpcode), a
|
|
ld a, (curArg1+1) ; 0-7
|
|
ld b, 3 ; displacement
|
|
call rlaX
|
|
or 0b01000110 ; 2nd upcode
|
|
ld (instrUpcode+1), a
|
|
ld c, 2
|
|
ret
|
|
|
|
handleBITIX:
|
|
ld a, 0xdd
|
|
jr handleBITIXY
|
|
handleBITIY:
|
|
ld a, 0xfd
|
|
handleBITIXY:
|
|
ld (instrUpcode), a ; first upcode
|
|
call handleBIT
|
|
ret nz ; error
|
|
ld a, 0xcb ; 2nd upcode
|
|
ld (instrUpcode+1), a
|
|
ld a, (curArg2+1) ; IXY displacement
|
|
ld (instrUpcode+2), a
|
|
ld a, (curArg1+1) ; 0-7
|
|
ld b, 3 ; displacement
|
|
call rlaX
|
|
or 0b01000110 ; 4th upcode
|
|
ld (instrUpcode+3), a
|
|
ld c, 4
|
|
ret
|
|
|
|
handleBITR:
|
|
call handleBIT
|
|
ret nz ; error
|
|
; get group value
|
|
ld a, (curArg2+1) ; group value
|
|
ld c, a
|
|
; write first upcode
|
|
ld a, 0xcb ; first upcode
|
|
ld (instrUpcode), a
|
|
; get bit value
|
|
ld a, (curArg1+1) ; 0-7
|
|
ld b, 3 ; displacement
|
|
call rlaX
|
|
; Now we have group value in stack, bit value in A (properly shifted)
|
|
; and we want to OR them together
|
|
or c ; Now we have our ORed value
|
|
or 0b01000000 ; and with the constant value for that byte...
|
|
; we're good!
|
|
ld (instrUpcode+1), a
|
|
ld c, 2
|
|
ret
|
|
|
|
handleIM:
|
|
ld a, (curArg1+1)
|
|
cp 0
|
|
jr z, .im0
|
|
cp 1
|
|
jr z, .im1
|
|
cp 2
|
|
jr z, .im2
|
|
; error
|
|
ld c, 0
|
|
ret
|
|
.im0:
|
|
ld a, 0x46
|
|
jr .proceed
|
|
.im1:
|
|
ld a, 0x56
|
|
jr .proceed
|
|
.im2:
|
|
ld a, 0x5e
|
|
.proceed:
|
|
ld (instrUpcode+1), a
|
|
ld a, 0xed
|
|
ld (instrUpcode), a
|
|
ld c, 2
|
|
ret
|
|
|
|
handleLDIXn:
|
|
ld a, 0xdd
|
|
jr handleLDIXYn
|
|
handleLDIYn:
|
|
ld a, 0xfd
|
|
handleLDIXYn:
|
|
ld (instrUpcode), a
|
|
ld a, 0x36 ; second upcode
|
|
ld (instrUpcode+1), a
|
|
ld a, (curArg1+1) ; IXY displacement
|
|
ld (instrUpcode+2), a
|
|
ld a, (curArg2+1) ; N
|
|
ld (instrUpcode+3), a
|
|
ld c, 4
|
|
ret
|
|
.error:
|
|
xor c
|
|
ret
|
|
|
|
handleLDIXr:
|
|
ld a, 0xdd
|
|
jr handleLDIXYr
|
|
handleLDIYr:
|
|
ld a, 0xfd
|
|
handleLDIXYr:
|
|
ld (instrUpcode), a
|
|
ld a, (curArg2+1) ; group value
|
|
or 0b01110000 ; second upcode
|
|
ld (instrUpcode+1), a
|
|
ld a, (curArg1+1) ; IXY displacement
|
|
ld (instrUpcode+2), a
|
|
ld c, 3
|
|
ret
|
|
.error:
|
|
xor c
|
|
ret
|
|
|
|
; Compute the upcode for argspec row at (DE) and arguments in curArg{1,2} and
|
|
; writes the resulting upcode in instrUpcode. A is the number if bytes written
|
|
; to instrUpcode (can be zero if something went wrong).
|
|
getUpcode:
|
|
push ix
|
|
push de
|
|
push hl
|
|
push bc
|
|
; First, let's go in IX mode. It's easier to deal with offsets here.
|
|
ld ixh, d
|
|
ld ixl, e
|
|
|
|
; Are we a "special instruction"?
|
|
bit 5, (ix+3)
|
|
jr z, .normalInstr ; not set: normal instruction
|
|
; We are a special instruction. Fetch handler (little endian, remember).
|
|
ld l, (ix+4)
|
|
ld h, (ix+5)
|
|
call callHL
|
|
; We have our result written in instrUpcode and C is set.
|
|
jp .end
|
|
|
|
.normalInstr:
|
|
; we begin by writing our "base upcode", which can be one or two bytes
|
|
ld a, (ix+4) ; first upcode
|
|
ld (instrUpcode), a
|
|
ld de, instrUpcode ; from this point, DE points to "where we are"
|
|
; in terms of upcode writing.
|
|
inc de ; make DE point to where we should write next.
|
|
ld a, (ix+5) ; second upcode
|
|
cp 0 ; do we have a second upcode?
|
|
jr z, .onlyOneUpcode
|
|
; we have two upcodes
|
|
ld (de), a
|
|
inc de
|
|
.onlyOneUpcode:
|
|
; now, let's see if we're dealing with a group here
|
|
ld a, (ix+1) ; first argspec
|
|
call isGroupId
|
|
jr z, .firstArgIsGroup
|
|
; First arg not a group. Maybe second is?
|
|
ld a, (ix+2) ; 2nd argspec
|
|
call isGroupId
|
|
jr nz, .writeExtraBytes ; not a group? nothing to do. go to
|
|
; next step: write extra bytes
|
|
; Second arg is group
|
|
ld hl, curArg2
|
|
jr .isGroup
|
|
.firstArgIsGroup:
|
|
ld hl, curArg1
|
|
.isGroup:
|
|
; A is a group, good, now let's get its value. HL is pointing to
|
|
; the argument. Our group value is at (HL+1).
|
|
inc hl
|
|
ld a, (hl)
|
|
; Now, we have our arg "group value" in A. Were going to need to
|
|
; displace it left by the number of steps specified in the table.
|
|
push af
|
|
ld a, (ix+3) ; displacement bit
|
|
and a, 0xf ; we only use the lower nibble.
|
|
ld b, a
|
|
pop af
|
|
call rlaX
|
|
|
|
; At this point, we have a properly displaced value in A. We'll want
|
|
; to OR it with the opcode.
|
|
; However, we first have to verify whether this ORing takes place on
|
|
; the second upcode or the first.
|
|
bit 6, (ix+3)
|
|
jr z, .firstUpcode ; not set: first upcode
|
|
or (ix+5) ; second upcode
|
|
ld (instrUpcode+1), a
|
|
jr .writeExtraBytes
|
|
.firstUpcode:
|
|
or (ix+4) ; first upcode
|
|
ld (instrUpcode), a
|
|
jr .writeExtraBytes
|
|
.writeExtraBytes:
|
|
; Good, we are probably finished here for many primary opcodes. However,
|
|
; some primary opcodes take 8 or 16 bit constants as an argument and
|
|
; if that's the case here, we need to write it too.
|
|
; We still have our instruction row in IX and we have DE pointing to
|
|
; where we should write next (which could be the second or the third
|
|
; byte of instrUpcode).
|
|
ld a, (ix+1) ; first argspec
|
|
ld hl, curArg1
|
|
call checkNOrM
|
|
jr z, .withWord
|
|
call checknmxy
|
|
jr z, .withByte
|
|
ld a, (ix+2) ; second argspec
|
|
ld hl, curArg2
|
|
call checkNOrM
|
|
jr z, .withWord
|
|
call checknmxy
|
|
jr z, .withByte
|
|
; nope, no number, alright, we're finished here
|
|
ld c, 1
|
|
jr .computeBytesWritten
|
|
.withByte:
|
|
; verify that the MSB in argument is zero
|
|
inc hl
|
|
inc hl ; MSB is 2nd byte
|
|
ld a, (hl)
|
|
dec hl ; HL now points to LSB
|
|
cp 0
|
|
jr nz, .numberTruncated
|
|
; HL points to our number
|
|
; one last thing to check. Is the 7th bit on the displacement value set?
|
|
; if yes, we have to decrease our value by 2. Uses for djnz and jr.
|
|
bit 7, (ix+3)
|
|
jr z, .skipDecrease
|
|
; Yup, it's set.
|
|
dec (hl)
|
|
dec (hl)
|
|
.skipDecrease:
|
|
ldi
|
|
ld c, 2
|
|
jr .computeBytesWritten
|
|
|
|
.withWord:
|
|
inc hl ; HL now points to LSB
|
|
; Clear to proceed. HL already points to our number
|
|
ldi ; LSB written, we point to MSB now
|
|
ldi ; MSB written
|
|
ld c, 3
|
|
jr .computeBytesWritten
|
|
.computeBytesWritten:
|
|
; At this point, everything that we needed to write in instrUpcode is
|
|
; written an C is 1 if we have no extra byte, 2 if we have an extra
|
|
; byte and 3 if we have an extra word. What we need to do here is check
|
|
; if ix+5 is non-zero and increase C if it is.
|
|
ld a, (ix+5)
|
|
cp 0
|
|
jr z, .end ; no second upcode? nothing to do.
|
|
; We have 2 base upcodes
|
|
inc c
|
|
jr .end
|
|
.numberTruncated:
|
|
; problem: not zero, so value is truncated. error
|
|
xor c
|
|
.end:
|
|
ld a, c
|
|
pop bc
|
|
pop hl
|
|
pop de
|
|
pop ix
|
|
ret
|
|
|
|
; Parse next argument in string (HL) and place it in (DE)
|
|
; Sets Z on success, reset on error.
|
|
processArg:
|
|
call toWord
|
|
call readWord
|
|
; Read word is in scratchpad, HL is properly advanced. Now, let's push
|
|
; that HL value and replace it with (scratchpad) so that we can parse
|
|
; that arg.
|
|
push hl
|
|
ld hl, scratchpad
|
|
|
|
call parseArg
|
|
cp 0xff
|
|
jr z, .error
|
|
ld (de), a
|
|
; When A is a number, IX is set with the value of that number. Because
|
|
; We don't use the space allocated to store those numbers in any other
|
|
; occasion, we store IX there unconditonally, LSB first.
|
|
inc de
|
|
ld a, ixl
|
|
ld (de), a
|
|
inc de
|
|
ld a, ixh
|
|
ld (de), a
|
|
cp a ; ensure Z is set
|
|
jr .end
|
|
.error:
|
|
call JUMP_UNSETZ
|
|
.end:
|
|
pop hl
|
|
ret
|
|
|
|
; Parse instruction specified in A (I_* const) with args in (HL) and write
|
|
; resulting opcode(s) in (instrUpcode). Returns the number of bytes written in A.
|
|
parseInstruction:
|
|
push bc
|
|
push hl
|
|
push de
|
|
; A is reused in matchPrimaryRow but that register is way too changing.
|
|
; Let's keep a copy in a more cosy register.
|
|
ld c, a
|
|
ld de, curArg1
|
|
call processArg
|
|
jr nz, .error
|
|
ld de, curArg2
|
|
call processArg
|
|
jr nz, .error
|
|
; Parsing done, no error, let's move forward to instr row matching!
|
|
ld de, instrTBl
|
|
ld b, INSTR_TBL_CNT
|
|
.loop:
|
|
ld a, c ; recall A param
|
|
call matchPrimaryRow
|
|
jr z, .match
|
|
ld a, INSTR_TBL_ROWSIZE
|
|
call JUMP_ADDDE
|
|
djnz .loop
|
|
; no match
|
|
xor a
|
|
jr .end
|
|
.match:
|
|
; We have our matching instruction row. We're getting pretty near our
|
|
; goal here!
|
|
call getUpcode
|
|
jr .end
|
|
.error:
|
|
xor a
|
|
.end:
|
|
pop de
|
|
pop hl
|
|
pop bc
|
|
ret
|
|
|
|
|
|
; In instruction metadata below, argument types arge indicated with a single
|
|
; char mnemonic that is called "argspec". This is the table of correspondance.
|
|
; Single letters are represented by themselves, so we don't need as much
|
|
; metadata.
|
|
; Special meaning:
|
|
; 0 : no arg
|
|
; 1-10 : group id (see Groups section)
|
|
; 0xff: error
|
|
|
|
; Format: 1 byte argspec + 4 chars string
|
|
argspecTbl:
|
|
.db 'A', "A", 0, 0, 0
|
|
.db 'B', "B", 0, 0, 0
|
|
.db 'C', "C", 0, 0, 0
|
|
.db 'k', "(C)", 0
|
|
.db 'D', "D", 0, 0, 0
|
|
.db 'E', "E", 0, 0, 0
|
|
.db 'H', "H", 0, 0, 0
|
|
.db 'L', "L", 0, 0, 0
|
|
.db 'I', "I", 0, 0, 0
|
|
.db 'R', "R", 0, 0, 0
|
|
.db 'h', "HL", 0, 0
|
|
.db 'l', "(HL)"
|
|
.db 'd', "DE", 0, 0
|
|
.db 'e', "(DE)"
|
|
.db 'b', "BC", 0, 0
|
|
.db 'c', "(BC)"
|
|
.db 'a', "AF", 0, 0
|
|
.db 'f', "AF'", 0
|
|
.db 'X', "IX", 0, 0
|
|
.db 'Y', "IY", 0, 0
|
|
.db 'x', "(IX)" ; always come with displacement
|
|
.db 'y', "(IY)" ; with JP
|
|
.db 's', "SP", 0, 0
|
|
.db 'p', "(SP)"
|
|
; we also need argspecs for the condition flags
|
|
.db 'Z', "Z", 0, 0, 0
|
|
.db 'z', "NZ", 0, 0
|
|
; C is in conflict with the C register. The situation is ambiguous, but
|
|
; doesn't cause actual problems.
|
|
.db '=', "NC", 0, 0
|
|
.db '+', "P", 0, 0, 0
|
|
.db '-', "M", 0, 0, 0
|
|
.db '1', "PO", 0, 0
|
|
.db '2', "PE", 0, 0
|
|
|
|
; argspecs not in the list:
|
|
; n -> N
|
|
; N -> NN
|
|
; m -> (N) (running out of mnemonics. 'm' for 'memory pointer')
|
|
; M -> (NN)
|
|
|
|
; Groups
|
|
; Groups are specified by strings of argspecs. To facilitate jumping to them,
|
|
; we have a fixed-sized table. Because most of them are 2 or 4 bytes long, we
|
|
; have a table that is 4 in size to minimize consumed space. We treat the two
|
|
; groups that take 8 bytes in a special way.
|
|
;
|
|
; The table below is in order, starting with group 0x01
|
|
argGrpTbl:
|
|
.db "bdha" ; 0x01
|
|
.db "ZzC=" ; 0x02
|
|
.db "bdhs" ; 0x03
|
|
.db "bdXs" ; 0x04
|
|
.db "bdYs" ; 0x05
|
|
|
|
argGrpCC:
|
|
.db "zZ=C12+-" ; 0xa
|
|
argGrpABCDEHL:
|
|
.db "BCDEHL_A" ; 0xb
|
|
|
|
; Each row is 4 bytes wide, fill with zeroes
|
|
instrNames:
|
|
.db "ADC", 0
|
|
.db "ADD", 0
|
|
.db "AND", 0
|
|
.db "BIT", 0
|
|
.db "CALL"
|
|
.db "CCF", 0
|
|
.db "CP",0,0
|
|
.db "CPD", 0
|
|
.db "CPDR"
|
|
.db "CPI", 0
|
|
.db "CPIR"
|
|
.db "CPL", 0
|
|
.db "DAA", 0
|
|
.db "DEC", 0
|
|
.db "DI",0,0
|
|
.db "DJNZ"
|
|
.db "EI",0,0
|
|
.db "EX",0,0
|
|
.db "EXX", 0
|
|
.db "HALT"
|
|
.db "IM",0,0
|
|
.db "IN",0,0
|
|
.db "INC", 0
|
|
.db "IND", 0
|
|
.db "INDR"
|
|
.db "INI", 0
|
|
.db "INIR"
|
|
.db "JP",0,0
|
|
.db "JR",0,0
|
|
.db "LD",0,0
|
|
.db "LDD", 0
|
|
.db "LDDR"
|
|
.db "LDI", 0
|
|
.db "LDIR"
|
|
.db "NEG", 0
|
|
.db "NOP", 0
|
|
.db "OR",0,0
|
|
.db "OTDR"
|
|
.db "OTIR"
|
|
.db "OUT", 0
|
|
.db "POP", 0
|
|
.db "PUSH"
|
|
.db "RET", 0
|
|
.db "RLA", 0
|
|
.db "RLCA"
|
|
.db "RRA", 0
|
|
.db "RRCA"
|
|
.db "SBC", 0
|
|
.db "SCF", 0
|
|
.db "SUB", 0
|
|
.db "XOR", 0
|
|
|
|
; This is a list of all supported instructions. Each row represent a combination
|
|
; of instr/argspecs (which means more than one row per instr). Format:
|
|
;
|
|
; 1 byte for the instruction ID
|
|
; 1 byte for arg constant
|
|
; 1 byte for 2nd arg constant
|
|
; 1 byte displacement for group arguments + flags
|
|
; 2 bytes for upcode (2nd byte is zero if instr is one byte)
|
|
;
|
|
; An "arg constant" is a char corresponding to either a row in argspecTbl or
|
|
; a group index in argGrpTbl (values < 0x10 are considered group indexes).
|
|
;
|
|
; The displacement bit is split in 2 nibbles: lower nibble is the displacement
|
|
; value, upper nibble is for flags:
|
|
;
|
|
; Bit 7: indicates that the numerical argument is of the 'e' type and has to be
|
|
; decreased by 2 (djnz, jr).
|
|
; Bit 6: it indicates that the group argument's value is to be placed on the
|
|
; second upcode rather than the first.
|
|
; Bit 5: Indicates that this row is handled very specially: the next two bytes
|
|
; aren't upcode bytes, but a routine address to call to handle this case with
|
|
; custom code.
|
|
|
|
instrTBl:
|
|
.db I_ADC, 'A', 'l', 0, 0x8e , 0 ; ADC A, (HL)
|
|
.db I_ADC, 'A', 0xb, 0, 0b10001000 , 0 ; ADC A, r
|
|
.db I_ADC, 'A', 'n', 0, 0xce , 0 ; ADC A, n
|
|
.db I_ADC, 'h', 0x3, 0x44, 0xed, 0b01001010 ; ADC HL, ss
|
|
.db I_ADD, 'A', 'l', 0, 0x86 , 0 ; ADD A, (HL)
|
|
.db I_ADD, 'A', 0xb, 0, 0b10000000 , 0 ; ADD A, r
|
|
.db I_ADD, 'A', 'n', 0, 0xc6 , 0 ; ADD A, n
|
|
.db I_ADD, 'h', 0x3, 4, 0b00001001 , 0 ; ADD HL, ss
|
|
.db I_ADD, 'X', 0x4, 0x44, 0xdd, 0b00001001 ; ADD IX, pp
|
|
.db I_ADD, 'Y', 0x5, 0x44, 0xfd, 0b00001001 ; ADD IY, rr
|
|
.db I_ADD, 'A', 'x', 0, 0xdd, 0x86 ; ADD A, (IX+d)
|
|
.db I_ADD, 'A', 'y', 0, 0xfd, 0x86 ; ADD A, (IY+d)
|
|
.db I_AND, 'l', 0, 0, 0xa6 , 0 ; AND (HL)
|
|
.db I_AND, 0xb, 0, 0, 0b10100000 , 0 ; AND r
|
|
.db I_AND, 'n', 0, 0, 0xe6 , 0 ; AND n
|
|
.db I_AND, 'x', 0, 0, 0xdd, 0xa6 ; AND (IX+d)
|
|
.db I_AND, 'y', 0, 0, 0xfd, 0xa6 ; AND (IY+d)
|
|
.db I_BIT, 'n', 'l', 0x20 \ .dw handleBITHL ; BIT b, (HL)
|
|
.db I_BIT, 'n', 'x', 0x20 \ .dw handleBITIX ; BIT b, (IX+d)
|
|
.db I_BIT, 'n', 'y', 0x20 \ .dw handleBITIY ; BIT b, (IY+d)
|
|
.db I_BIT, 'n', 0xb, 0x20 \ .dw handleBITR ; BIT b, r
|
|
.db I_CALL,0xa, 'N', 3, 0b11000100 , 0 ; CALL cc, NN
|
|
.db I_CALL,'N', 0, 0, 0xcd , 0 ; CALL NN
|
|
.db I_CCF, 0, 0, 0, 0x3f , 0 ; CCF
|
|
.db I_CP, 'l', 0, 0, 0xbe , 0 ; CP (HL)
|
|
.db I_CP, 0xb, 0, 0, 0b10111000 , 0 ; CP r
|
|
.db I_CP, 'n', 0, 0, 0xfe , 0 ; CP n
|
|
.db I_CP, 'x', 0, 0, 0xdd, 0xbe ; CP (IX+d)
|
|
.db I_CP, 'y', 0, 0, 0xfd, 0xbe ; CP (IY+d)
|
|
.db I_CPD, 0, 0, 0, 0xed, 0xa9 ; CPD
|
|
.db I_CPDR,0, 0, 0, 0xed, 0xb9 ; CPDR
|
|
.db I_CPI, 0, 0, 0, 0xed, 0xa1 ; CPI
|
|
.db I_CPIR,0, 0, 0, 0xed, 0xb1 ; CPIR
|
|
.db I_CPL, 0, 0, 0, 0x2f , 0 ; CPL
|
|
.db I_DAA, 0, 0, 0, 0x27 , 0 ; DAA
|
|
.db I_DEC, 'l', 0, 0, 0x35 , 0 ; DEC (HL)
|
|
.db I_DEC, 'X', 0, 0, 0xdd, 0x2b ; DEC IX
|
|
.db I_DEC, 'x', 0, 0, 0xdd, 0x35 ; DEC (IX+d)
|
|
.db I_DEC, 'Y', 0, 0, 0xfd, 0x2b ; DEC IY
|
|
.db I_DEC, 'y', 0, 0, 0xfd, 0x35 ; DEC (IY+d)
|
|
.db I_DEC, 0xb, 0, 3, 0b00000101 , 0 ; DEC r
|
|
.db I_DEC, 0x3, 0, 4, 0b00001011 , 0 ; DEC ss
|
|
.db I_DI, 0, 0, 0, 0xf3 , 0 ; DI
|
|
.db I_DJNZ,'n', 0, 0x80, 0x10 , 0 ; DJNZ e
|
|
.db I_EI, 0, 0, 0, 0xfb , 0 ; EI
|
|
.db I_EX, 'p', 'h', 0, 0xe3 , 0 ; EX (SP), HL
|
|
.db I_EX, 'p', 'X', 0, 0xdd, 0xe3 ; EX (SP), IX
|
|
.db I_EX, 'p', 'Y', 0, 0xfd, 0xe3 ; EX (SP), IY
|
|
.db I_EX, 'a', 'f', 0, 0x08 , 0 ; EX AF, AF'
|
|
.db I_EX, 'd', 'h', 0, 0xeb , 0 ; EX DE, HL
|
|
.db I_EXX, 0, 0, 0, 0xd9 , 0 ; EXX
|
|
.db I_HALT,0, 0, 0, 0x76 , 0 ; HALT
|
|
.db I_IM, 'n', 0, 0x20 \ .dw handleIM ; IM {0,1,2}
|
|
.db I_IN, 'A', 'm', 0, 0xdb , 0 ; IN A, (n)
|
|
.db I_IN, 0xb, 'k', 0x43, 0xed, 0b01000000 ; IN r, (C)
|
|
.db I_INC, 'l', 0, 0, 0x34 , 0 ; INC (HL)
|
|
.db I_INC, 'X', 0, 0, 0xdd , 0x23 ; INC IX
|
|
.db I_INC, 'x', 0, 0, 0xdd , 0x34 ; INC (IX+d)
|
|
.db I_INC, 'Y', 0, 0, 0xfd , 0x23 ; INC IY
|
|
.db I_INC, 'y', 0, 0, 0xfd , 0x34 ; INC (IY+d)
|
|
.db I_INC, 0xb, 0, 3, 0b00000100 , 0 ; INC r
|
|
.db I_INC, 0x3, 0, 4, 0b00000011 , 0 ; INC ss
|
|
.db I_IND, 0, 0, 0, 0xed, 0xaa ; IND
|
|
.db I_INDR,0, 0, 0, 0xed, 0xba ; INDR
|
|
.db I_INI, 0, 0, 0, 0xed, 0xa2 ; INI
|
|
.db I_INIR,0, 0, 0, 0xed, 0xb2 ; INIR
|
|
.db I_JP, 'l', 0, 0, 0xe9 , 0 ; JP (HL)
|
|
.db I_JP, 0xa, 'N', 3, 0b11000010 , 0 ; JP cc, NN
|
|
.db I_JP, 'N', 0, 0, 0xc3 , 0 ; JP NN
|
|
.db I_JP, 'x', 0, 0x20 \ .dw handleJPIX ; JP (IX)
|
|
.db I_JP, 'y', 0, 0x20 \ .dw handleJPIY ; JP (IY)
|
|
.db I_JR, 'n', 0, 0x80, 0x18 , 0 ; JR e
|
|
.db I_JR, 'C', 'n', 0x80, 0x38 , 0 ; JR C, e
|
|
.db I_JR, '=', 'n', 0x80, 0x30 , 0 ; JR NC, e
|
|
.db I_JR, 'Z', 'n', 0x80, 0x28 , 0 ; JR Z, e
|
|
.db I_JR, 'z', 'n', 0x80, 0x20 , 0 ; JR NZ, e
|
|
.db I_LD, 'c', 'A', 0, 0x02 , 0 ; LD (BC), A
|
|
.db I_LD, 'e', 'A', 0, 0x12 , 0 ; LD (DE), A
|
|
.db I_LD, 'A', 'c', 0, 0x0a , 0 ; LD A, (BC)
|
|
.db I_LD, 'A', 'e', 0, 0x1a , 0 ; LD A, (DE)
|
|
.db I_LD, 's', 'h', 0, 0xf9 , 0 ; LD SP, HL
|
|
.db I_LD, 'A', 'I', 0, 0xed, 0x57 ; LD A, I
|
|
.db I_LD, 'I', 'A', 0, 0xed, 0x47 ; LD I, A
|
|
.db I_LD, 'A', 'R', 0, 0xed, 0x5f ; LD A, R
|
|
.db I_LD, 'R', 'A', 0, 0xed, 0x4f ; LD R, A
|
|
.db I_LD, 'l', 0xb, 0, 0b01110000 , 0 ; LD (HL), r
|
|
.db I_LD, 0xb, 'l', 3, 0b01000110 , 0 ; LD r, (HL)
|
|
.db I_LD, 'l', 'n', 0, 0x36 , 0 ; LD (HL), n
|
|
.db I_LD, 0xb, 'n', 3, 0b00000110 , 0 ; LD r, (HL)
|
|
.db I_LD, 0x3, 'N', 4, 0b00000001 , 0 ; LD dd, n
|
|
.db I_LD, 'M', 'A', 0, 0x32 , 0 ; LD (NN), A
|
|
.db I_LD, 'A', 'M', 0, 0x3a , 0 ; LD A, (NN)
|
|
.db I_LD, 'M', 'h', 0, 0x22 , 0 ; LD (NN), HL
|
|
.db I_LD, 'h', 'M', 0, 0x2a , 0 ; LD HL, (NN)
|
|
.db I_LD, 'M', 'X', 0, 0xdd, 0x22 ; LD (NN), IX
|
|
.db I_LD, 'X', 'M', 0, 0xdd, 0x2a ; LD IX, (NN)
|
|
.db I_LD, 'M', 'Y', 0, 0xfd, 0x22 ; LD (NN), IY
|
|
.db I_LD, 'Y', 'M', 0, 0xfd, 0x2a ; LD IY, (NN)
|
|
.db I_LD, 'M', 0x3, 0x44, 0xed, 0b01000011 ; LD (NN), dd
|
|
.db I_LD, 0x3, 'M', 0x44, 0xed, 0b01001011 ; LD dd, (NN)
|
|
.db I_LD, 'x', 'n', 0x20 \ .dw handleLDIXn ; LD (IX+d), n
|
|
.db I_LD, 'y', 'n', 0x20 \ .dw handleLDIYn ; LD (IY+d), n
|
|
.db I_LD, 'x', 0xb, 0x20 \ .dw handleLDIXr ; LD (IX+d), r
|
|
.db I_LD, 'y', 0xb, 0x20 \ .dw handleLDIYr ; LD (IY+d), r
|
|
.db I_LDD, 0, 0, 0, 0xed, 0xa8 ; LDD
|
|
.db I_LDDR,0, 0, 0, 0xed, 0xb8 ; LDDR
|
|
.db I_LDI, 0, 0, 0, 0xed, 0xa0 ; LDI
|
|
.db I_LDIR,0, 0, 0, 0xed, 0xb0 ; LDIR
|
|
.db I_NEG, 0, 0, 0, 0xed, 0x44 ; NEG
|
|
.db I_NOP, 0, 0, 0, 0x00 , 0 ; NOP
|
|
.db I_OR, 'l', 0, 0, 0xb6 , 0 ; OR (HL)
|
|
.db I_OR, 0xb, 0, 0, 0b10110000 , 0 ; OR r
|
|
.db I_OR, 'n', 0, 0, 0xf6 , 0 ; OR n
|
|
.db I_OR, 'x', 0, 0, 0xdd, 0xb6 ; OR (IX+d)
|
|
.db I_OR, 'y', 0, 0, 0xfd, 0xb6 ; OR (IY+d)
|
|
.db I_OTDR,0, 0, 0, 0xed, 0xbb ; OTDR
|
|
.db I_OTIR,0, 0, 0, 0xed, 0xb3 ; OTIR
|
|
.db I_OUT, 'm', 'A', 0, 0xd3 , 0 ; OUT (n), A
|
|
.db I_OUT, 'k', 0xb, 0x43, 0xed, 0b01000001 ; OUT (C), r
|
|
.db I_POP, 0x1, 0, 4, 0b11000001 , 0 ; POP qq
|
|
.db I_PUSH,0x1, 0, 4, 0b11000101 , 0 ; PUSH qq
|
|
.db I_RET, 0, 0, 0, 0xc9 , 0 ; RET
|
|
.db I_RET, 0xa, 0, 3, 0b11000000 , 0 ; RET cc
|
|
.db I_RLA, 0, 0, 0, 0x17 , 0 ; RLA
|
|
.db I_RLCA,0, 0, 0, 0x07 , 0 ; RLCA
|
|
.db I_RRA, 0, 0, 0, 0x1f , 0 ; RRA
|
|
.db I_RRCA,0, 0, 0, 0x0f , 0 ; RRCA
|
|
.db I_SBC, 'A', 'l', 0, 0x9e , 0 ; SBC A, (HL)
|
|
.db I_SBC, 'A', 0xb, 0, 0b10011000 , 0 ; SBC A, r
|
|
.db I_SCF, 0, 0, 0, 0x37 , 0 ; SCF
|
|
.db I_SUB, 'A', 'l', 0, 0x96 , 0 ; SUB A, (HL)
|
|
.db I_SUB, 'A', 0xb, 0, 0b10010000 , 0 ; SUB A, r
|
|
.db I_SUB, 'n', 0, 0, 0xd6 , 0 ; SUB n
|
|
.db I_XOR, 'l', 0, 0, 0xae , 0 ; XOR (HL)
|
|
.db I_XOR, 0xb, 0, 0, 0b10101000 , 0 ; XOR r
|
|
|
|
|
|
; *** Variables ***
|
|
; Args are 3 bytes: argspec, then values of numerical constants (when that's
|
|
; appropriate)
|
|
curArg1:
|
|
.db 0, 0, 0
|
|
curArg2:
|
|
.db 0, 0, 0
|
|
|
|
instrUpcode:
|
|
.db 0, 0, 0, 0
|
|
|