Files
LithosAnanake/v4/capsule/codegen.v4
T
rajamesandClaude Opus 5.5 806f876ee0 fix(v4.0.0): shorter chains of calls in the compiler capsule
Measured, compiling a DO loop left one return entry spare under
INTERPRET and a defining word called from inside another word
overflowed the nine-entry return stack (D-2); the prompt loop will take
one more.

C@ and C! shift without a loop.  The code generator shifts each opcode
into the word being built, so nothing is placed by a counted shift, and
the address masks come from a table.  The dictionary search, the
in-liner, the end of a loop and the making of a data word are each one
word reached by a jump.

Every kind of line now leaves at least three return entries spare while
it compiles; a word run from the interpreter has six; and CREATE ...
DOES> works from the prompt, from a word, and from a word that calls
that.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-04 11:17:58 -04:00

120 lines
4.9 KiB
Plaintext

\ codegen.v4 -- the code generator: opcodes, literals and branches packed into
\ instruction words at HERE.
\
\ This is what the compiling words ( : ; IF LOOP LITERAL ... ) are built on.
\ It lays code down exactly as DECOMPOSITION.md 1.2 and 2 require and as the
\ test harness's own packers do (v4/include/v4/asm.h, text.h):
\
\ - opcodes fill slots 0 .. 5 of a word left to right; a full word is
\ appended to the dictionary and a new one started; unused slots are nop
\ - a literal is @p in a slot and its value in the cell after the word;
\ several in one word follow it in order
\ - ; and ex close the word, since nothing after them runs
\ - a branch closes the word: the bits to its right are its address
\ - a branch goes only in a slot whose address field reaches all of the
\ node's memory; otherwise the word is closed first and the branch starts
\ the next one. So no branch ever needs a reach check.
\
\ An instruction word is 32 bits at every cell width (D-9): slot k is the five
\ bits whose lowest is bit 27 - 5k, and a branch there has 27 - 5k address bits.
\
\ Rests on core.v4 and dict.v4. Constants the loader supplies:
\ (CG) word address of fourteen cells of scratch for this file
\ CG-BSLOTS how many slots, counted from slot 0, may hold a branch
\ Like WORD, these run under whatever the user has on the stacks, so what they
\ work on is in (CG), they keep at most three cells of their own on the data
\ stack, and they call each other as little as they can (D-2):
\ (CG)+0 the word being built (CG)+1 its next free slot
\ (CG)+2 how many literals it owes, and (CG)+3 .. (CG)+8 the literals
\ (CG)+9 the opcode (OP,) was given (CG)+10 the ref (BRANCH0) made
\ (CG)+11 the target a branch is to reach (CG)+12 which literal (FLUSH) is at
\ (CG)+13 what (FLUSH) fills unused slots with: nop, or 0 after a branch
\
\ The word being built is kept right-justified: each opcode is shifted in
\ below the ones before it, and (FLUSH) shifts in what fills the rest and the
\ two spare bits. So nothing is ever shifted by a counted loop, which would
\ want a return-stack entry for its count.
\ ( k -- mask ) the address bits of a branch in slot k: 2^(27-5k) - 1
: (MASK)
if K0 -1 + if K1 -1 + if K2
drop 4095 ;
K2: drop 131071 ;
K1: drop 4194303 ;
K0: drop 134217727 ;
\ ( -- ) forget the word being built
: (CG-RESET) 0 (CG) a! ! 0 (CG)+1 a! ! 0 (CG)+2 a! ! 28 (CG)+13 a! ! ;
\ ( x -- ) five bits into the next free slot, which must be free
: (PUT)
(CG) a! @ 2* 2* 2* 2* 2* + !
(CG)+1 a! @ 1 + ! ;
\ ( -- ) append the word being built, its unused slots filled, then the
\ literals it owes; nothing if no slot is in use
: (FLUSH)
(CG)+1 a! @ if EMPTY drop
P: (CG)+1 a! @ -6 + if FULL drop (CG)+13 a! @ (PUT) jump P
FULL: drop
(CG) a! @ 2* 2* , \ the two spare bits
0 (CG)+12 a! !
L: (CG)+12 a! @ (CG)+2 a! @ xor if DONE
drop (CG)+12 a! @ dup 1 + ! (CG)+3 + a! @ ,
jump L
DONE: drop
jump (CG-RESET)
EMPTY: drop ;
\ ( op -- ) any opcode that is neither a branch nor @p
: (OP,)
dup (CG)+9 a! ! (PUT)
(CG)+9 a! @ if ENDS -1 + if ENDS \ ; and ex end the word
drop (CG)+1 a! @ -6 + if ENDS drop ; \ and so does its last slot
ENDS: drop jump (FLUSH)
\ ( x -- ) @p and its value
: (LIT,)
(CG)+2 a! @ dup 1 + ! \ x n
(CG)+3 + a! !
8 (PUT)
(CG)+1 a! @ -6 + if FULL drop ;
FULL: drop jump (FLUSH)
\ ( -- addr ) where the next word will go: a branch target
: (LABEL) (FLUSH) HERE ;
\ ( op -- ) a branch (jump 2, call 3, next 5, if 6, -if 7) whose address is
\ not filled in yet. Its ref -- the branch's word address times 8, plus its
\ slot -- is left in (CG)+10. The rest of its word is filled with 0, which
\ is its address field. It ends by jumping to (FLUSH), so that the chain of
\ calls under it is one shorter.
: (BRANCH0)
(CG)+1 a! @ CG-BSLOTS inv + 1 + -if MOVE drop jump PLACED \ slot - CG-BSLOTS
MOVE: drop (FLUSH)
PLACED:
\ the word will be appended at HERE; its literals, if any, come after it
HERE 2* 2* 2* (CG)+1 a! @ + (CG)+10 a! !
(PUT)
0 (CG)+13 a! ! jump (FLUSH)
\ ( op -- ref ) the same, leaving the ref, for (RESOLVE) to fill in later
: (BRANCH>) (BRANCH0) (CG)+10 a! @ ;
\ ( ref -- ) put the address in (CG)+11 into the branch at ref
: (RESOLVE1)
dup 7 and (MASK) push \ ref R: mask
2/ 2/ 2/ a! \ A: the branch's word
(CG)+11 b! @b pop dup push and \ the address bits
@ pop inv and + ! ;
\ ( target ref -- ) put the address into the branch at ref
: (RESOLVE) SWAP (CG)+11 a! ! jump (RESOLVE1)
\ ( target op -- ) a branch to a known address
: (BRANCH,) SWAP (CG)+11 a! ! (BRANCH0) (CG)+10 a! @ jump (RESOLVE1)
: (JUMP,) ( addr -- ) 2 jump (BRANCH,)
: (CALL,) ( addr -- ) 3 jump (BRANCH,)