--- gforth/prims2x.fs	2002/06/02 15:46:16	1.106
+++ gforth/prims2x.fs	2003/10/09 14:15:19	1.144
@@ -1,6 +1,6 @@
 \ converts primitives to, e.g., C code 
 
-\ Copyright (C) 1995,1996,1997,1998,2000 Free Software Foundation, Inc.
+\ Copyright (C) 1995,1996,1997,1998,2000,2003 Free Software Foundation, Inc.
 
 \ This file is part of Gforth.
 
@@ -51,8 +51,15 @@
 \ (stack-in-index-xt and a test for stack==instruction-stream); there
 \ should be only one.
 
+\ for backwards compatibility, jaw
+require compat/strcomp.fs
+
 warnings off
 
+\ redefinitions of kernel words not present in gforth-0.6.1
+: latestxt lastcfa @ ;
+: latest last @ ;
+
 [IFUNDEF] try
 include startup.fs
 [THEN]
@@ -60,10 +67,10 @@ include startup.fs
 : struct% struct ; \ struct is redefined in gray
 
 warnings off
+\ warnings on
 
 include ./gray.fs
-
-32 constant max-effect \ number of things on one side of a stack effect
+128 constant max-effect \ number of things on one side of a stack effect
 4 constant max-stacks  \ the max. number of stacks (including inst-stream).
 255 constant maxchar
 maxchar 1+ constant eof-char
@@ -78,10 +85,25 @@ variable line-start \ pointer to start o
 0 line !
 2variable filename \ filename of original input file
 0 0 filename 2!
+2variable out-filename \ filename of the output file (for sync lines)
+0 0 out-filename 2!
 2variable f-comment
 0 0 f-comment 2!
 variable skipsynclines \ are sync lines ("#line ...") invisible to the parser?
-skipsynclines on 
+skipsynclines on
+variable out-nls \ newlines in output (for output sync lines)
+0 out-nls !
+variable store-optimization \ use store optimization?
+store-optimization off
+
+variable include-skipped-insts
+\ does the threaded code for a combined instruction include the cells
+\ for the component instructions (true) or only the cells for the
+\ inline arguments (false)
+include-skipped-insts off
+
+variable immarg \ values for immediate arguments (to be used in IMM_ARG macros)
+$12340000 immarg !
 
 : th ( addr1 n -- addr2 )
     cells + ;
@@ -121,12 +143,24 @@ skipsynclines on
 	    0
 	recover endtry
 	r> to outfile-id throw
-	abort
+	1 (bye) \ abort
     endif ;
 
 : quote ( -- )
     [char] " emit ;
 
+\ count output lines to generate sync lines for output
+
+: count-nls ( addr u -- )
+    bounds u+do
+	i c@ nl-char = negate out-nls +!
+    loop ;
+
+:noname ( addr u -- )
+    2dup count-nls
+    defers type ;
+is type
+
 variable output          \ xt ( -- ) of output word for simple primitives
 variable output-combined \ xt ( -- ) of output word for combined primitives
 
@@ -135,6 +169,7 @@ struct%
     cell% 2* field stack-pointer \ stackpointer name
     cell%    field stack-type \ name for default type of stack items
     cell%    field stack-in-index-xt \ ( in-size item -- in-index )
+    cell%    field stack-access-transform \ ( nitem -- index )
 end-struct stack%
 
 struct%
@@ -153,8 +188,29 @@ struct%
     cell%    field type-store \ xt of store code generator ( item -- )
 end-struct type%
 
+struct%
+    cell%    field register-number
+    cell%    field register-type \ pointer to type
+    cell% 2* field register-name \ c name
+end-struct register%
+
+struct%
+    cell% 2* field ss-registers  \ addr u; ss-registers[0] is TOS
+                                 \ 0 means: use memory
+    cell%    field ss-offset     \ stack pointer offset: sp[-offset] is TOS
+end-struct ss% \ stack-state
+
+struct%
+    cell%              field state-number
+    cell% max-stacks * field state-sss
+end-struct state%
+
 variable next-stack-number 0 next-stack-number !
 create stacks max-stacks cells allot \ array of stacks
+256 constant max-registers
+create registers max-registers cells allot \ array of registers
+variable nregisters 0 nregisters ! \ number of registers
+variable next-state-number 0 next-state-number ! \ next state number
 
 : stack-in-index ( in-size item -- in-index )
     item-offset @ - 1- ;
@@ -170,9 +226,17 @@ create stacks max-stacks cells allot \ a
     1 next-stack-number +!
     r@ stack-type !
     save-mem r@ stack-pointer 2! 
-    ['] stack-in-index r> stack-in-index-xt ! ;
+    ['] stack-in-index r@ stack-in-index-xt !
+    ['] noop r@ stack-access-transform !
+    rdrop ;
 
 : map-stacks { xt -- }
+    \ perform xt for all stacks
+    next-stack-number @ 0 +do
+	stacks i th @ xt execute
+    loop ;
+
+: map-stacks1 { xt -- }
     \ perform xt for all stacks except inst-stream
     next-stack-number @ 1 +do
 	stacks i th @ xt execute
@@ -226,10 +290,28 @@ end-struct prim%
 0 value combined \ in combined prims the combined prim
 variable in-part \ true if processing a part
  in-part off
+0 value state-in  \ state on entering prim
+0 value state-out \ state on exiting prim
+0 value state-default  \ canonical state at bb boundaries
+
+: prim-context ( ... p xt -- ... )
+    \ execute xt with prim set to p
+    prim >r
+    swap to prim
+    catch
+    r> to prim
+    throw ;
 
 1000 constant max-combined
 create combined-prims max-combined cells allot
 variable num-combined
+variable part-num \ current part number during process-combined
+
+: map-combined { xt -- }
+    \ perform xt for all components of the current combined instruction
+    num-combined @ 0 +do
+	combined-prims i th @ xt execute
+    loop ;
 
 table constant combinations
   \ the keys are the sequences of pointers to primitives
@@ -238,6 +320,19 @@ create current-depth max-stacks cells al
 create max-depth     max-stacks cells allot
 create min-depth     max-stacks cells allot
 
+create sp-update-in max-stacks cells allot
+\ where max-depth occured the first time
+create max-depths max-stacks max-combined 1+ * cells allot
+\ maximum depth at start of each part: array[parts] of array[stack]
+create max-back-depths max-stacks max-combined 1+ * cells allot
+\ maximun depth from end of the combination to the start of the each part
+
+: s-c-max-depth ( nstack ncomponent -- addr )
+    max-stacks * + cells max-depths + ;
+
+: s-c-max-back-depth ( nstack ncomponent -- addr )
+    max-stacks * + cells max-back-depths + ;
+
 wordlist constant primitives
 
 : create-prim ( prim -- )
@@ -258,6 +353,13 @@ variable name-line
 2variable name-filename
 2variable last-name-filename
 Variable function-number 0 function-number !
+Variable function-old 0 function-old !
+: function-diff ( n -- )
+    ." GROUPADD(" function-number @ function-old @ - 0 .r ." )" cr
+    function-number @ function-old ! ;
+: forth-fdiff ( -- )
+    function-number @ function-old @ - 0 .r ."  groupadd" cr
+    function-number @ function-old ! ;
 
 \ a few more set ops
 
@@ -267,25 +369,56 @@ Variable function-number 0 function-numb
 : complement ( set1 -- set2 )
  empty ['] bit-equivalent binary-set-operation ;
 
+\ forward declaration for inst-stream (breaks cycle in definitions)
+defer inst-stream-f ( -- stack )
+
 \ stack access stuff
 
-: normal-stack-access ( n stack -- )
-    stack-pointer 2@ type
-    dup
-    if
-	." [" 0 .r ." ]"
+: normal-stack-access0 { n stack -- }
+    \ n has the ss-offset already applied (see ...-access1)
+    n stack stack-access-transform @ execute ." [" 0 .r ." ]" ;
+
+: state-ss { stack state -- ss }
+    state state-sss stack stack-number @ th @ ;
+
+: stack-reg { n stack state -- reg }
+    \ n is the index (TOS=0); reg is 0 if the access is to memory
+    stack state state-ss ss-registers 2@ n u> if ( addr ) \ in ss-registers?
+	n th @
     else
-	drop ." TOS"
+	drop 0
     endif ;
 
-\ forward declaration for inst-stream (breaks cycle in definitions)
-defer inst-stream-f ( -- stack )
+: .reg ( reg -- )
+    register-name 2@ type ;
+
+: stack-offset ( stack state -- n )
+    \ offset for stack in state
+    state-ss ss-offset @ ;
+
+: normal-stack-access1 { n stack state -- }
+    n stack state stack-reg ?dup-if
+	.reg exit
+    endif
+    stack stack-pointer 2@ type
+    n stack state stack-offset - stack normal-stack-access0 ;
+
+: normal-stack-access ( n stack state -- )
+    over inst-stream-f = if
+	." IMM_ARG(" normal-stack-access1 ." ," immarg ? ." )"
+	1 immarg +!
+    else
+	normal-stack-access1
+    endif ;
+
+: stack-depth { stack -- n }
+    current-depth stack stack-number @ th @ ;
 
 : part-stack-access { n stack -- }
     \ print _<stack><x>, x=inst-stream? n : maxdepth-currentdepth-n-1
     ." _" stack stack-pointer 2@ type
     stack stack-number @ { stack# }
-    current-depth stack# th @ n + { access-depth }
+    stack stack-depth n + { access-depth }
     stack inst-stream-f = if
 	access-depth
     else
@@ -295,12 +428,44 @@ defer inst-stream-f ( -- stack )
     endif
     0 .r ;
 
-: stack-access ( n stack -- )
+: part-stack-read { n stack -- }
+    stack stack-depth n + ( ndepth )
+    stack stack-number @ part-num @ s-c-max-depth @
+\    max-depth stack stack-number @ th @ ( ndepth nmaxdepth )
+    over <= if ( ndepth ) \ load from memory
+	stack state-in normal-stack-access
+    else
+	drop n stack part-stack-access
+    endif ;
+
+: stack-diff ( stack -- n )
+    \ in-out
+    dup stack-in @ swap stack-out @ - ;
+
+: part-stack-write { n stack -- }
+    stack stack-depth n +
+    stack stack-number @ part-num @ s-c-max-back-depth @
+    over <= if ( ndepth )
+	stack combined ['] stack-diff prim-context -
+	stack state-out normal-stack-access
+    else
+	drop n stack part-stack-access
+    endif ;
+
+: stack-read ( n stack -- )
     \ print a stack access at index n of stack
     in-part @ if
-	part-stack-access
+	part-stack-read
     else
-	normal-stack-access
+	state-in normal-stack-access
+    endif ;
+
+: stack-write ( n stack -- )
+    \ print a stack access at index n of stack
+    in-part @ if
+	part-stack-write
+    else
+	state-out normal-stack-access
     endif ;
 
 : item-in-index { item -- n }
@@ -316,7 +481,7 @@ defer inst-stream-f ( -- stack )
     >r
     ." vm_" r@ item-stack-type-name type
     ." 2" r@ item-type @ print-type-prefix ." ("
-    r@ item-in-index r@ item-stack @ stack-access ." ,"
+    r@ item-in-index r@ item-stack @ stack-read ." ,"
     r@ item-name 2@ type
     ." );" cr
     rdrop ; 
@@ -327,8 +492,8 @@ defer inst-stream-f ( -- stack )
     ." vm_two"
     r@ item-stack-type-name type ." 2"
     r@ item-type @ print-type-prefix ." ("
-    r@ item-in-index r@ item-stack @ 2dup ." (Cell)" stack-access
-    ." , "                      -1 under+ ." (Cell)" stack-access
+    r@ item-in-index r@ item-stack @ 2dup ." (Cell)" stack-read
+    ." , "                      -1 under+ ." (Cell)" stack-read
     ." , " r@ item-name 2@ type
     ." )" cr
     rdrop ;
@@ -351,7 +516,7 @@ defer inst-stream-f ( -- stack )
  rdrop ;
 
 : item-out-index ( item -- n )
-    \ n is the index of item (in the in-effect)
+    \ n is the index of item (in the out-effect)
     >r r@ item-stack @ stack-out @ r> item-offset @ - 1- ;
 
 : really-store-single ( item -- )
@@ -360,22 +525,17 @@ defer inst-stream-f ( -- stack )
     r@ item-type @ print-type-prefix ." 2"
     r@ item-stack-type-name type ." ("
     r@ item-name 2@ type ." ,"
-    r@ item-out-index r@ item-stack @ stack-access ." );"
+    r@ item-out-index r@ item-stack @ stack-write ." );"
     rdrop ;
 
-: store-single ( item -- )
- >r
- r@ same-as-in?
- if
-   r@ item-in-index 0= r@ item-out-index 0= xor
-   if
-       ." IF_" r@ item-stack @ stack-pointer 2@ type
-       ." TOS(" r@ really-store-single ." );" cr
-   endif
- else
-   r@ really-store-single cr
- endif
- rdrop ;
+: store-single { item -- }
+    item item-stack @ { stack }
+    store-optimization @ in-part @ 0= and item same-as-in? and
+    item item-in-index  stack state-in  stack-reg 0= and \  in in memory?
+    item item-out-index stack state-out stack-reg 0= and \ out in memory?
+    0= if
+	item really-store-single cr
+    endif ;
 
 : store-double ( item -- )
 \ !! store optimization is not performed, because it is not yet needed
@@ -384,8 +544,8 @@ defer inst-stream-f ( -- stack )
  r@ item-type @ print-type-prefix ." 2two"
  r@ item-stack-type-name type ." ("
  r@ item-name 2@ type ." , "
- r@ item-out-index r@ item-stack @ 2dup stack-access
- ." , "                       -1 under+ stack-access
+ r@ item-out-index r@ item-stack @ 2dup stack-write
+ ." , "                       -1 under+ stack-write
  ." )" cr
  rdrop ;
 
@@ -493,21 +653,68 @@ does> ( item -- )
 wordlist constant type-names \ this is here just to meet the requirement
                     \ that a type be a word; it is never used for lookup
 
+: define-type ( addr u -- xt )
+    \ define single type with name addr u, without stack
+    get-current type-names set-current >r
+    2dup nextname stack-type-name
+    r> set-current
+    latestxt ;
+
 : stack ( "name" "stack-pointer" "type" -- )
     \ define stack
     name { d: stack-name }
     name { d: stack-pointer }
     name { d: stack-type }
-    get-current type-names set-current
-    stack-type 2dup nextname stack-type-name
-    set-current
-    stack-pointer lastxt >body stack-name nextname make-stack ;
+    stack-type define-type
+    stack-pointer rot >body stack-name nextname make-stack ;
 
 stack inst-stream IP Cell
 ' inst-in-index inst-stream stack-in-index-xt !
 ' inst-stream <is> inst-stream-f
 \ !! initialize stack-in and stack-out
 
+\ registers
+
+: make-register ( type addr u -- )
+    \ define register with type TYPE and name ADDR U.
+    nregisters @ max-registers < s" too many registers" ?print-error
+    2dup nextname create register% %allot >r
+    r@ register-name 2!
+    r@ register-type !
+    nregisters @ r@ register-number !
+    1 nregisters +!
+    rdrop ;
+
+: register ( "name" "type" -- )
+    \ define register
+    name { d: reg-name }
+    name { d: reg-type }
+    reg-type define-type >body
+    reg-name make-register ;
+
+\ stack-states
+
+: stack-state ( a-addr u uoffset "name" -- )
+    create ss% %allot >r
+    r@ ss-offset !
+    r@ ss-registers 2!
+    rdrop ;
+
+0 0 0 stack-state default-ss
+
+\ state
+
+: state ( "name" -- )
+    \ create a state initialized with default-sss
+    create state% %allot { s }
+    next-state-number @ s state-number ! 1 next-state-number +!
+    max-stacks 0 ?do
+	default-ss s state-sss i th !
+    loop ;
+
+: set-ss ( ss stack state -- )
+    state-sss swap stack-number @ th ! ;
+
 \ offset computation
 \ the leftmost (i.e. deepest) item has offset 0
 \ the rightmost item has the highest offset
@@ -524,12 +731,9 @@ stack inst-stream IP Cell
 : compute-offset-out ( addr1 addr2 -- )
     ['] stack-out compute-offset ;
 
-: clear-stack { -- }
-    dup stack-in off stack-out off ;
-
 : compute-offsets ( -- )
-    ['] clear-stack map-stacks
-    inst-stream clear-stack
+    prim prim-stacks-in  max-stacks cells erase
+    prim prim-stacks-out max-stacks cells erase
     prim prim-effect-in  prim prim-effect-in-end  @ ['] compute-offset-in  map-items
     prim prim-effect-out prim prim-effect-out-end @ ['] compute-offset-out map-items
     inst-stream stack-out @ 0= s" # can only be on the input side" ?print-error ;
@@ -540,26 +744,42 @@ stack inst-stream IP Cell
     declarations compute-offsets
     output @ execute ;
 
-: flush-a-tos { stack -- }
-    stack stack-out @ 0<> stack stack-in @ 0= and
-    if
-	." IF_" stack stack-pointer 2@ 2dup type ." TOS("
-	2dup type ." [0] = " type ." TOS);" cr
-    endif ;
+: stack-state-items ( stack state -- n )
+    state-ss ss-registers 2@ nip ;
 
-: flush-tos ( -- )
-    ['] flush-a-tos map-stacks ;
+: unused-stack-items { stack -- n-in n-out }
+    \ n-in  are the stack items in state-in  not used    by prim
+    \ n-out are the stack items in state-out not written by prim
+    stack state-in  stack-state-items stack stack-in  @ - 0 max
+    stack state-out stack-state-items stack stack-out @ - 0 max ;
+
+: spill-stack { stack -- }
+    \ spill regs of state-in that are not used by prim and are not in state-out
+    stack state-in stack-offset { offset }
+    stack state-in stack-state-items ( items )
+    dup stack unused-stack-items - - +do
+	\ loop through the bottom items
+	stack stack-pointer 2@ type
+	i offset - stack normal-stack-access0 ."  = "
+	i stack state-in normal-stack-access1 ." ;" cr
+    loop ;
 
-: fill-a-tos { stack -- }
-    stack stack-out @ 0= stack stack-in @ 0<> and
-    if
-	." IF_" stack stack-pointer 2@ 2dup type ." TOS("
-	2dup type ." TOS = " type ." [0]);" cr
-    endif ;
+: spill-state ( -- )
+    ['] spill-stack map-stacks1 ;
 
-: fill-tos ( -- )
+: fill-stack { stack -- }
+    stack state-out stack-offset { offset }
+    stack state-out stack-state-items ( items )
+    dup stack unused-stack-items - + +do
+	\ loop through the bottom items
+	i stack state-out normal-stack-access1 ."  = "
+	stack stack-pointer 2@ type
+	i offset - stack normal-stack-access0 ." ;" cr
+    loop ;
+
+: fill-state ( -- )
     \ !! inst-stream for prefetching?
-    ['] fill-a-tos map-stacks ;
+    ['] fill-stack map-stacks1 ;
 
 : fetch ( addr -- )
     dup item-type @ type-fetch @ execute ;
@@ -567,20 +787,62 @@ stack inst-stream IP Cell
 : fetches ( -- )
     prim prim-effect-in prim prim-effect-in-end @ ['] fetch map-items ;
 
-: stack-pointer-update { stack -- }
-    \ stack grow downwards
-    stack stack-in @ stack stack-out @ -
-    ?dup-if \ this check is not necessary, gcc would do this for us
-	stack stack-pointer 2@ type ."  += " 0 .r ." ;" cr
+: reg-reg-move ( reg-from reg-to -- )
+    2dup = if
+	2drop
+    else
+	.reg ."  = " .reg ." ;" cr
     endif ;
 
-: inst-pointer-update ( -- )
-    inst-stream stack-in @ ?dup-if
-	." INC_IP(" 0 .r ." );" cr
+: stack-bottom-reg { n stack state -- reg }
+    stack state stack-state-items n - 1- stack state stack-reg ;
+
+: stack-moves { stack -- }
+    \ generate moves between registers in state-in/state-out that are
+    \ not spilled or consumed/produced by prim.
+    \ !! this works only for a simple stack cache, not e.g., for
+    \ rotating stack caches, or registers shared between stacks (the
+    \ latter would also require a change in interface)
+    \ !! maybe place this after NEXT_P1?
+    stack unused-stack-items 2dup < if ( n-in n-out )
+	\ move registers from 0..n_in-1 to n_out-n_in..n_out-1
+	over - { diff } ( n-in )
+	-1 swap 1- -do
+	    i stack state-in stack-bottom-reg ( reg-from )
+	    i diff + stack state-out stack-bottom-reg reg-reg-move
+	1 -loop
+    else
+	\ move registers from n_in-n_out..n_in-1 to 0..n_out-1
+	swap over - { diff } ( n-out )
+	0 +do
+	    i diff + stack state-in stack-bottom-reg ( reg-from )
+	    i stack state-out stack-bottom-reg reg-reg-move
+	loop
     endif ;
 
+: stack-update-transform ( n1 stack -- n2 )
+    \ n2 is the number by which the stack pointer should be
+    \ incremented to pop n1 items
+    stack-access-transform @ dup >r execute
+    0 r> execute - ;
+
+: stack-pointer-update { stack -- }
+    \ and moves
+    \ stacks grow downwards
+    stack stack-diff ( in-out )
+    stack state-in  stack-offset -
+    stack state-out stack-offset + ( [in-in_offset]-[out-out_offset] )
+    ?dup-if \ this check is not necessary, gcc would do this for us
+	stack inst-stream = if
+	    ." INC_IP(" 0 .r ." );" cr
+	else
+	    stack stack-pointer 2@ type ."  += "
+	    stack stack-update-transform 0 .r ." ;" cr
+	endif
+    endif
+    stack stack-moves ;
+
 : stack-pointer-updates ( -- )
-    inst-pointer-update
     ['] stack-pointer-update map-stacks ;
 
 : store ( item -- )
@@ -625,60 +887,89 @@ stack inst-stream IP Cell
     endif
     2drop ;
 
-: output-c-tail ( -- )
-    \ the final part of the generated C code
+: output-nextp2 ( -- )
+    ." NEXT_P2;" cr ;
+
+variable tail-nextp2 \ xt to execute for printing NEXT_P2 in INST_TAIL
+' output-nextp2 tail-nextp2 !
+
+: output-label2 ( -- )
+    ." LABEL2(" prim prim-c-name 2@ type ." )" cr
+    ." NEXT_P2;" cr ;
+
+: output-c-tail1 { xt -- }
+    \ the final part of the generated C code, with xt printing LABEL2 or not.
     output-super-end
     print-debug-results
     ." NEXT_P1;" cr
     stores
-    fill-tos
-    ." NEXT_P2;" ;
+    fill-state 
+    xt execute ;
+
+: output-c-tail1-no-stores { xt -- }
+    \ the final part of the generated C code for combinations
+    output-super-end
+    ." NEXT_P1;" cr
+    fill-state 
+    xt execute ;
+
+: output-c-tail ( -- )
+    tail-nextp2 @ output-c-tail1 ;
+
+: output-c-tail2 ( -- )
+    ['] output-label2 output-c-tail1 ;
+
+: output-c-tail-no-stores ( -- )
+    tail-nextp2 @ output-c-tail1-no-stores ;
+
+: output-c-tail2-no-stores ( -- )
+    ['] output-label2 output-c-tail1-no-stores ;
 
 : type-c-code ( c-addr u xt -- )
-    \ like TYPE, but replaces "TAIL;" with tail code produced by xt
+    \ like TYPE, but replaces "INST_TAIL;" with tail code produced by xt
     { xt }
+    ." {" cr
+    ." #line " c-line @ . quote c-filename 2@ type quote cr
     begin ( c-addr1 u1 )
-	2dup s" TAIL;" search
+	2dup s" INST_TAIL;" search
     while ( c-addr1 u1 c-addr3 u3 )
 	2dup 2>r drop nip over - type
 	xt execute
-	2r> 5 /string
+	2r> 10 /string
 	\ !! resync #line missing
     repeat
-    2drop type ;
+    2drop type
+    ." #line " out-nls @ 2 + . quote out-filename 2@ type quote cr
+    ." }" cr ;
 
 : print-entry ( -- )
-    ." LABEL(" prim prim-c-name 2@ type ." ):" ;
+    ." LABEL(" prim prim-c-name 2@ type ." )" ;
     
 : output-c ( -- ) 
- print-entry ."  /* " prim prim-name 2@ type ."  ( " prim prim-stack-string 2@ type ." ) */" cr
- ." /* " prim prim-doc 2@ type ."  */" cr
- ." NAME(" quote prim prim-name 2@ type quote ." )" cr \ debugging
- ." {" cr
- ." DEF_CA" cr
- print-declarations
- ." NEXT_P0;" cr
- flush-tos
- fetches
- print-debug-args
- stack-pointer-updates
- ." {" cr
- ." #line " c-line @ . quote c-filename 2@ type quote cr
- prim prim-c-code 2@ ['] output-c-tail type-c-code
- ." }" cr
- output-c-tail
- ." }" cr
- cr
+    print-entry ."  /* " prim prim-name 2@ type ."  ( " prim prim-stack-string 2@ type ." ) */" cr
+    ." /* " prim prim-doc 2@ type ."  */" cr
+    ." NAME(" quote prim prim-name 2@ type quote ." )" cr \ debugging
+    ." {" cr
+    ." DEF_CA" cr
+    print-declarations
+    ." NEXT_P0;" cr
+    spill-state
+    fetches
+    print-debug-args
+    stack-pointer-updates
+    prim prim-c-code 2@ ['] output-c-tail type-c-code
+    output-c-tail2
+    ." }" cr
+    cr
 ;
 
 : disasm-arg { item -- }
     item item-stack @ inst-stream = if
-	."   fputc(' ', vm_out); "
-	\ !! change this to first convert args to the right type and
-	\ then print them
-	." /* printarg_" item item-type @ print-type-prefix
-	." ((" item item-type @ type-c-name 2@ type ." )"
-	." ip[" item item-offset @ 1+ 0 .r ." ]); */" cr
+	." {" cr
+	item print-declaration
+	item fetch
+	item print-debug-arg
+	." }" cr
     endif ;
 
 : disasm-args ( -- )
@@ -706,14 +997,15 @@ stack inst-stream IP Cell
     endif
     ." }" cr ;
 
+: output-profile-part ( p )
+    ."   add_inst(b, " quote
+    prim-name 2@ type
+    quote ." );" cr ;
+    
 : output-profile-combined ( -- )
     \ generate code for postprocessing the VM block profile stuff
     ." if (VM_IS_INST(*ip, " function-number @ 0 .r ." )) {" cr
-    num-combined @ 0 +do
-	."   add_inst(b, " quote
-	combined-prims i th @ prim-name 2@ type
-	quote ." );" cr
-    loop
+    ['] output-profile-part map-combined
     ."   ip += " inst-stream stack-in @ 1+ 0 .r ." ;" cr
     combined-prims num-combined @ 1- th @ prim-c-code 2@  s" SET_IP"    search nip nip
     combined-prims num-combined @ 1- th @ prim-c-code 2@  s" SUPER_END" search nip nip or if
@@ -723,10 +1015,14 @@ stack inst-stream IP Cell
     endif
     ." }" cr ;
 
+: prim-branch? { prim -- f }
+    \ true if prim is a branch or super-end
+    prim prim-c-code 2@  s" SET_IP" search nip nip 0<> ;
+
 : output-superend ( -- )
     \ output flag specifying whether the current word ends a dynamic superinst
-    prim prim-c-code 2@  s" SET_IP"    search nip nip
-    prim prim-c-code 2@  s" SUPER_END" search nip nip or 0<>
+    prim prim-branch?
+    prim prim-c-code 2@  s" SUPER_END" search nip nip 0<> or
     prim prim-c-code 2@  s" SUPER_CONTINUE" search nip nip 0= and
     negate 0 .r ." , /* " prim prim-name 2@ type ."  */" cr ;
 
@@ -779,7 +1075,7 @@ stack inst-stream IP Cell
 \      data-stack   stack-used? IF ." Cell *sp=SP;" cr THEN
 \      fp-stack     stack-used? IF ." Cell *fp=*FP;" cr THEN
 \      return-stack stack-used? IF ." Cell *rp=*RP;" cr THEN
-\      flush-tos
+\      spill-state
 \      fetches
 \      stack-pointer-updates
 \      fp-stack   stack-used? IF ." *FP=fp;" cr THEN
@@ -788,7 +1084,7 @@ stack inst-stream IP Cell
 \      prim prim-c-code 2@ type
 \      ." }" cr
 \      stores
-\      fill-tos
+\      fill-state
 \      ." return (sp);" cr
 \      ." }" cr
 \      cr ;
@@ -799,6 +1095,9 @@ stack inst-stream IP Cell
 : output-alias ( -- ) 
     ( primitive-number @ . ." alias " ) ." Primitive " prim prim-name 2@ type cr ;
 
+: output-c-prim-num ( -- )
+    ." N_" prim prim-c-name 2@ type ." ," cr ;
+
 : output-forth ( -- )  
     prim prim-forth-code @ 0=
     IF    	\ output-alias
@@ -900,9 +1199,7 @@ stack inst-stream IP Cell
 \  #line 516 "./prim"
 \  n = n1+n2;
 \  }
-\  NEXT_P1;
 \  _x_sp0 = (Cell)n;
-\  NEXT_P2;
 \  }
 \  NEXT_P1;
 \  spTOS = (Cell)_x_sp0;
@@ -912,6 +1209,7 @@ stack inst-stream IP Cell
     prim to combined
     0 num-combined !
     current-depth max-stacks cells erase
+    include-skipped-insts @ current-depth 0 th !
     max-depth     max-stacks cells erase
     min-depth     max-stacks cells erase
     prim prim-effect-in  prim prim-effect-in-end  !
@@ -923,24 +1221,33 @@ stack inst-stream IP Cell
 : min! ( n addr -- )
     tuck @ min swap ! ;
 
+: inst-stream-adjustment ( nstack -- n )
+    \ number of stack items to add for each part
+    0= include-skipped-insts @ and negate ;
+
 : add-depths { p -- }
     \ combine stack effect of p with *-depths
     max-stacks 0 ?do
 	current-depth i th @
-	p prim-stacks-in  i th @ +
+	p prim-stacks-in  i th @ + i inst-stream-adjustment +
 	dup max-depth i th max!
 	p prim-stacks-out i th @ -
 	dup min-depth i th min!
 	current-depth i th !
     loop ;
 
+: copy-maxdepths ( n -- )
+    max-depth max-depths rot max-stacks * th max-stacks cells move ;
+
 : add-prim ( addr u -- )
     \ add primitive given by "addr u" to combined-prims
     primitives search-wordlist s" unknown primitive" ?print-error
     execute { p }
     p combined-prims num-combined @ th !
+    num-combined @ copy-maxdepths
     1 num-combined +!
-    p add-depths ;
+    p add-depths
+    num-combined @ copy-maxdepths ;
 
 : compute-effects { q -- }
     \ compute the stack effects of q from the depths
@@ -973,6 +1280,27 @@ stack inst-stream IP Cell
 	i q prim-stacks-out i th @ q prim-effect-out-end make-effect-items
     loop ;
 
+: compute-stack-max-back-depths ( stack -- )
+    stack-number @ { stack# }
+    current-depth stack# th @ dup
+    dup stack# num-combined @ s-c-max-back-depth !
+    -1 num-combined @ 1- -do ( max-depth current-depth )
+	combined-prims i th @ { p }
+	p prim-stacks-out stack# th @ +
+	dup >r max r>
+	over stack# i s-c-max-back-depth !
+	p prim-stacks-in stack# th @ -
+	stack# inst-stream-adjustment -
+    1 -loop
+    assert( dup stack# inst-stream-adjustment negate = )
+    assert( over max-depth stack# th @ = )
+    2drop ;
+
+: compute-max-back-depths ( -- )
+    \ compute max-back-depths.
+    \ assumes that current-depths is correct for the end of the combination
+    ['] compute-stack-max-back-depths map-stacks ;
+
 : process-combined ( -- )
     combined combined-prims num-combined @ cells
     combinations ['] constant insert-wordlist
@@ -980,14 +1308,91 @@ stack inst-stream IP Cell
     @ prim-c-code 2@ prim prim-c-code 2! \ used by output-super-end
     prim compute-effects
     prim init-effects
+    compute-max-back-depths
     output-combined perform ;
 
+\ reprocessing (typically to generate versions for another cache states)
+\ !! use prim-context
+
+variable reprocessed-num 0 reprocessed-num !
+
+: new-name ( -- c-addr u )
+    reprocessed-num @ 0
+    1 reprocessed-num +!
+    <# #s 'p hold '_ hold #> save-mem ;
+
+: reprocess-simple ( prim -- )
+    to prim
+    new-name prim prim-c-name 2!
+    output @ execute ;
+
+: lookup-prim ( c-addr u -- prim )
+    primitives search-wordlist 0= -13 and throw execute ;
+
+: state-prim1 { in-state out-state prim -- }
+    in-state out-state state-default dup d= ?EXIT
+    in-state  to state-in
+    out-state to state-out
+    prim reprocess-simple ;
+
+: state-prim ( in-state out-state "name" -- )
+    parse-word lookup-prim state-prim1 ;
+
+\ reprocessing with default states
+
+\ This is a simple scheme and should be generalized
+\ assumes we only cache one stack and use simple states for that
+
+0 value cache-stack  \ stack that we cache
+2variable cache-states \ states of the cache, starting with the empty state
+
+: compute-default-state-out ( n-in -- n-out )
+    \ for the current prim
+    cache-stack stack-in @ - 0 max
+    cache-stack stack-out @ + cache-states 2@ nip 1- min ;
+
+: gen-prim-states ( prim -- )
+    to prim
+    cache-states 2@ swap { states } ( nstates )
+    cache-stack stack-in @ +do
+	states i th @
+	states i compute-default-state-out th @
+	prim state-prim1
+    loop ;
+
+: prim-states ( "name" -- )
+    parse-word lookup-prim gen-prim-states ;
+
+: gen-branch-states ( prim -- )
+    \ generate versions that produce state-default; useful for branches
+    to prim
+    cache-states 2@ swap { states } ( nstates )
+    cache-stack stack-in @ +do
+	states i th @ state-default prim state-prim1
+    loop ;
+
+: branch-states ( out-state "name" -- )
+    parse-word lookup-prim gen-branch-states ;
+
+\ producing state transitions
+
+: gen-transitions ( "name" -- )
+    parse-word lookup-prim { prim }
+    cache-states 2@ { states nstates }
+    nstates 0 +do
+	nstates 0 +do
+	    i j <> if
+		states i th @ states j th @ prim state-prim1
+	    endif
+	loop
+    loop ;
+
 \ C output
 
 : print-item { n stack -- }
     \ print nth stack item name
     stack stack-type @ type-c-name 2@ type space
-    ." _" stack stack-pointer 2@ type n 0 .r ;
+    ." MAYBE_UNUSED _" stack stack-pointer 2@ type n 0 .r ;
 
 : print-declarations-combined ( -- )
     max-stacks 0 ?do
@@ -1005,10 +1410,20 @@ stack inst-stream IP Cell
 
 : output-combined-tail ( -- )
     part-output-c-tail
-    prim >r combined to prim
     in-part @ >r in-part off
-    output-c-tail
-    r> in-part ! r> to prim ;
+    combined ['] output-c-tail-no-stores prim-context
+    r> in-part ! ;
+
+: part-stack-pointer-updates ( -- )
+    next-stack-number @ 0 +do
+	i part-num @ 1+ s-c-max-depth @ dup
+	i num-combined @ s-c-max-depth @ =    \ final depth
+	swap i part-num @ s-c-max-depth @ <> \ just reached now
+	part-num @ 0= \ first part
+	or and if
+	    stacks i th @ stack-pointer-update
+	endif
+    loop ;
 
 : output-part ( p -- )
     to prim
@@ -1018,20 +1433,18 @@ stack inst-stream IP Cell
     print-declarations
     part-fetches
     print-debug-args
+    combined ['] part-stack-pointer-updates prim-context
+    1 part-num +!
     prim add-depths \ !! right place?
-    ." {" cr
-    ." #line " c-line @ . quote c-filename 2@ type quote cr
     prim prim-c-code 2@ ['] output-combined-tail type-c-code
-    ." }" cr
     part-output-c-tail
     ." }" cr ;
 
 : output-parts ( -- )
     prim >r in-part on
     current-depth max-stacks cells erase
-    num-combined @ 0 +do
-	combined-prims i th @ output-part
-    loop
+    0 part-num !
+    ['] output-part map-combined
     in-part off
     r> to prim ;
 
@@ -1042,12 +1455,12 @@ stack inst-stream IP Cell
     ." DEF_CA" cr
     print-declarations-combined
     ." NEXT_P0;" cr
-    flush-tos
-    fetches
+    spill-state
+    \ fetches \ now in parts
     \ print-debug-args
-    stack-pointer-updates
+    \ stack-pointer-updates now in parts
     output-parts
-    output-c-tail
+    output-c-tail2-no-stores
     ." }" cr
     cr ;
 
@@ -1057,13 +1470,16 @@ stack inst-stream IP Cell
 
 \ peephole optimization rules
 
+\ data for a simple peephole optimizer that always tries to combine
+\ the currently compiled instruction with the last one.
+
 \ in order for this to work as intended, shorter combinations for each
 \ length must be present, and the longer combinations must follow
 \ shorter ones (this restriction may go away in the future).
   
 : output-peephole ( -- )
     combined-prims num-combined @ 1- cells combinations search-wordlist
-    s" the prefix for this combination must be defined earlier" ?print-error
+    s" the prefix for this superinstruction must be defined earlier" ?print-error
     ." {"
     execute prim-num @ 5 .r ." ,"
     combined-prims num-combined @ 1- th @ prim-num @ 5 .r ." ,"
@@ -1071,15 +1487,94 @@ stack inst-stream IP Cell
     combined prim-c-name 2@ type ."  */"
     cr ;
 
-: output-forth-peephole ( -- )
-    combined-prims num-combined @ 1- cells combinations search-wordlist
-    s" the prefix for this combination must be defined earlier" ?print-error
-    execute prim-num @ 5 .r
-    combined-prims num-combined @ 1- th @ prim-num @ 5 .r
-    combined prim-num @ 5 .r ."  prim, \ "
-    combined prim-c-name 2@ type
+
+\ cost and superinstruction data for a sophisticated combiner (e.g.,
+\ shortest path)
+
+\ This is intended as initializer for a structure like this
+
+\  struct cost {
+\    int loads;       /* number of stack loads */
+\    int stores;      /* number of stack stores */
+\    int updates;     /* number of stack pointer updates */
+\    int offset;      /* offset into super2 table */
+\    int length;      /* number of components */
+\  };
+
+\ How do you know which primitive or combined instruction this
+\ structure refers to?  By the order of cost structures, as in most
+\ other cases.
+
+: super2-length ( -- n )
+    combined if
+	num-combined @
+    else
+	1
+    endif ;
+
+: compute-costs { p -- nloads nstores nupdates }
+    \ compute the number of loads, stores, and stack pointer updates
+    \ of a primitive or combined instruction; does not take TOS
+    \ caching into account
+    0 max-stacks 0 +do
+	p prim-stacks-in i th @ +
+    loop
+    super2-length 1- - \ don't count instruction fetches of subsumed insts
+    0 max-stacks 0 +do
+	p prim-stacks-out i th @ +
+    loop
+    0 max-stacks 1 +do \ don't count ip updates, therefore "1 +do"
+	p prim-stacks-in i th @ p prim-stacks-out i th @ <> -
+    loop ;
+
+: output-num-part ( p -- )
+    ." N_" prim-c-name 2@ type ." ," ;
+    \ prim-num @ 4 .r ." ," ;
+
+: output-name-comment ( -- )
+    ."  /* " prim prim-name 2@ type ."  */" ;
+
+variable offset-super2  0 offset-super2 ! \ offset into the super2 table
+
+: output-costs-prefix ( -- )
+    ." {" prim compute-costs
+    rot 2 .r ." ," swap 2 .r ." ," 2 .r ." , "
+    prim prim-branch? negate . ." ,"
+    state-in  state-number @ 2 .r ." ,"
+    state-out state-number @ 2 .r ." ," ;
+
+: output-costs-gforth-simple ( -- )
+    output-costs-prefix
+    prim output-num-part
+    1 2 .r ." },"
+    output-name-comment
+    cr ;
+
+: output-costs-gforth-combined ( -- )
+    output-costs-prefix
+    ." N_START_SUPER+" offset-super2 @ 5 .r ." ,"
+    super2-length dup 2 .r ." }," offset-super2 +!
+    output-name-comment
+    cr ;
+
+: output-costs ( -- )
+    \ description of superinstructions and simple instructions
+    ." {" prim compute-costs
+    rot 2 .r ." ," swap 2 .r ." ," 2 .r ." ,"
+    offset-super2 @ 5 .r ." ,"
+    super2-length dup 2 .r ." }," offset-super2 +!
+    output-name-comment
     cr ;
 
+: output-super2 ( -- )
+    \ table of superinstructions without requirement for existing prefixes
+    combined if
+	['] output-num-part map-combined 
+    else
+	prim output-num-part
+    endif
+    output-name-comment
+    cr ;   
 
 \ the parser
 
@@ -1105,25 +1600,25 @@ print-token !
  getinput member? ;
 ' testchar? test-vector !
 
-: checksyncline ( -- )
+: checksynclines ( -- )
     \ when input points to a newline, check if the next line is a
     \ sync line.  If it is, perform the appropriate actions.
-    rawinput @ >r
-    s" #line " r@ over compare 0<> if
-	rdrop 1 line +! EXIT
-    endif
-    0. r> 6 chars + 20 >number drop >r drop line ! r> ( c-addr )
-    dup c@ bl = if
-	char+ dup c@ [char] " <> 0= s" sync line syntax" ?print-error
-	char+ dup 100 [char] " scan drop swap 2dup - save-mem filename 2!
-	char+
-    endif
-    dup c@ nl-char <> 0= s" sync line syntax" ?print-error
-    skipsynclines @ if
-	dup char+ rawinput !
-	rawinput @ c@ cookedinput @ c!
-    endif
-    drop ;
+    rawinput @ begin >r
+	s" #line " r@ over compare if
+	    rdrop 1 line +! EXIT
+	endif
+	0. r> 6 chars + 20 >number drop >r drop line ! r> ( c-addr )
+	dup c@ bl = if
+	    char+ dup c@ [char] " <> 0= s" sync line syntax" ?print-error
+	    char+ dup 100 [char] " scan drop swap 2dup - save-mem filename 2!
+	    char+
+	endif
+	dup c@ nl-char <> 0= s" sync line syntax" ?print-error
+	skipsynclines @ if
+	    char+ dup rawinput !
+	    rawinput @ c@ cookedinput @ c!
+	endif
+    again ;
 
 : ?nextchar ( f -- )
     s" syntax error, wrong char" ?print-error
@@ -1132,10 +1627,11 @@ print-token !
 	1 chars rawinput +!
 	1 chars cookedinput +!
 	nl-char = if
-	    checksyncline
+	    checksynclines
 	    rawinput @ line-start !
 	endif
-	rawinput @ c@ cookedinput @ c!
+	rawinput @ c@
+	cookedinput @ c!
     endif ;
 
 : charclass ( set "name" -- )
@@ -1158,6 +1654,8 @@ bl singleton tab-char over add-member
 nl-char singleton eof-char over add-member complement	charclass nonl
 nl-char singleton eof-char over add-member
     char : over add-member complement                   charclass nocolonnl
+nl-char singleton eof-char over add-member
+    char } over add-member complement                   charclass nobracenl
 bl 1+ maxchar .. char \ singleton complement intersection
                                                         charclass nowhitebq
 bl 1+ maxchar ..                                        charclass nowhite
@@ -1169,7 +1667,7 @@ nl-char singleton eof-char over add-memb
 (( letter (( letter || digit )) **
 )) <- c-ident ( -- )
 
-(( ` # ?? (( letter || digit || ` : )) **
+(( ` # ?? (( letter || digit || ` : )) ++
 )) <- stack-ident ( -- )
 
 (( nowhitebq nowhite ** ))
@@ -1191,27 +1689,33 @@ Variable c-flag
 )) <- c-comment ( -- )
 
 (( ` - nonl ** {{ 
-	forth-flag @ IF ." [ELSE]" cr THEN
-	c-flag @ IF ." #else" cr THEN }}
+	forth-flag @ IF forth-fdiff ." [ELSE]" cr THEN
+	c-flag @ IF
+	    function-diff
+	    ." #else /* " function-number @ 0 .r ."  */" cr THEN }}
 )) <- else-comment
 
 (( ` + {{ start }} nonl ** {{ end
 	dup
 	IF	c-flag @
-		IF    ." #ifdef HAS_" bounds ?DO  I c@ toupper emit  LOOP cr
+	    IF
+		function-diff
+		." #ifdef HAS_" bounds ?DO  I c@ toupper emit  LOOP cr
 		THEN
 		forth-flag @
-		IF  ." has? " type ."  [IF]"  cr THEN
+		IF  forth-fdiff  ." has? " type ."  [IF]"  cr THEN
 	ELSE	2drop
-	    c-flag @      IF  ." #endif"  cr THEN
-	    forth-flag @  IF  ." [THEN]"  cr THEN
+	    c-flag @      IF
+		function-diff  ." #endif" cr THEN
+	    forth-flag @  IF  forth-fdiff  ." [THEN]"  cr THEN
 	THEN }}
 )) <- if-comment
 
 (( (( ` g || ` G )) {{ start }} nonl **
    {{ end
-      forth-flag @ IF  ." group " type cr  THEN
-      c-flag @     IF  ." GROUP(" type ." )" cr  THEN }}
+      forth-flag @ IF  forth-fdiff  ." group " type cr  THEN
+      c-flag @     IF  function-diff
+	  ." GROUP(" type ." , " function-number @ 0 .r ." )" cr  THEN }}
 )) <- group-comment
 
 (( (( eval-comment || forth-comment || c-comment || else-comment || if-comment || group-comment )) ?? nonl ** )) <- comment-body
@@ -1232,7 +1736,10 @@ Variable c-flag
       (( {{ start }}  c-ident {{ end prim prim-c-name 2! }} )) ??
    )) ??  nleof
    (( ` " ` "  {{ start }} (( noquote ++ ` " )) ++ {{ end 1- prim prim-doc 2! }} ` " white ** nleof )) ??
-   {{ skipsynclines off line @ c-line ! filename 2@ c-filename 2! start }} (( nocolonnl nonl **  nleof white ** )) ** {{ end prim prim-c-code 2! skipsynclines on }}
+   {{ skipsynclines off line @ c-line ! filename 2@ c-filename 2! start }}
+   (( (( ` { nonl ** nleof (( (( nobracenl {{ line @ drop }} nonl ** )) ?? nleof )) ** ` } white ** nleof white ** ))
+   || (( nocolonnl nonl **  nleof white ** )) ** ))
+   {{ end prim prim-c-code 2! skipsynclines on }}
    (( ` :  white ** nleof
       {{ start }} (( nonl ++  nleof white ** )) ++ {{ end prim prim-forth-code 2! }}
    )) ?? {{ process-simple }}
@@ -1247,9 +1754,11 @@ Variable c-flag
 (( {{ make-prim to prim 0 to combined
       line @ name-line ! filename 2@ name-filename 2!
       function-number @ prim prim-num !
-      start }} forth-ident {{ end 2dup prim prim-name 2! prim prim-c-name 2! }}  white ++
+      start }} [ifdef] vmgen c-ident [else] forth-ident [then] {{ end
+      2dup prim prim-name 2! prim prim-c-name 2! }}  white **
    (( ` / white ** {{ start }} c-ident {{ end prim prim-c-name 2! }} white ** )) ??
-   (( simple-primitive || combined-primitive )) {{ 1 function-number +! }}
+   (( simple-primitive || combined-primitive ))
+   {{ 1 function-number +! }}
 )) <- primitive ( -- )
 
 (( (( comment || primitive || nl white ** )) ** eof ))
@@ -1274,13 +1783,24 @@ warnings @ [IF]
     \ process the string at addr u
     over dup rawinput ! dup line-start ! cookedinput !
     + endrawinput !
-    checksyncline
+    checksynclines
     primitives2something ;    
 
+: unixify ( c-addr u1 -- c-addr u2 )
+    \ delete crs from the string
+    bounds tuck tuck ?do ( c-addr1 )
+	i c@ dup #cr <> if
+	    over c! char+
+	else
+	    drop
+	endif
+    loop
+    over - ;
+
 : process-file ( addr u xt-simple x-combined -- )
     output-combined ! output !
     save-mem 2dup filename 2!
-    slurp-file
+    slurp-file unixify
     warnings @ if
 	." ------------ CUT HERE -------------" cr  endif
     primfilter ;