--- gforth/prims2x.fs	2002/08/20 07:59:01	1.110
+++ gforth/prims2x.fs	2002/10/12 19:06:37	1.120
@@ -78,10 +78,22 @@ 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
 
 : th ( addr1 n -- addr2 )
     cells + ;
@@ -121,12 +133,24 @@ 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
 
@@ -173,6 +197,12 @@ create stacks max-stacks cells allot \ a
     ['] stack-in-index r> stack-in-index-xt ! ;
 
 : 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
@@ -227,9 +257,24 @@ end-struct prim%
 variable in-part \ true if processing a part
  in-part off
 
+: 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 +283,19 @@ 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 -- )
@@ -281,11 +339,14 @@ Variable function-number 0 function-numb
 \ forward declaration for inst-stream (breaks cycle in definitions)
 defer inst-stream-f ( -- stack )
 
+: 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
@@ -295,10 +356,42 @@ 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 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 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
+    endif ;
+
+: stack-write ( n stack -- )
+    \ print a stack access at index n of stack
+    in-part @ if
+	part-stack-write
     else
 	normal-stack-access
     endif ;
@@ -316,7 +409,7 @@ 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 +420,8 @@ 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 ;
@@ -360,22 +453,20 @@ 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 ;
+    >r
+    store-optimization @ in-part @ 0= and r@ same-as-in? and 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-double ( item -- )
 \ !! store optimization is not performed, because it is not yet needed
@@ -384,8 +475,8 @@ 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 ;
 
@@ -524,12 +615,11 @@ stack inst-stream IP Cell
 : compute-offset-out ( addr1 addr2 -- )
     ['] stack-out compute-offset ;
 
-: clear-stack { -- }
+: clear-stack ( stack -- )
     dup stack-in off stack-out off ;
 
 : compute-offsets ( -- )
     ['] clear-stack map-stacks
-    inst-stream clear-stack
     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 ;
@@ -548,7 +638,7 @@ stack inst-stream IP Cell
     endif ;
 
 : flush-tos ( -- )
-    ['] flush-a-tos map-stacks ;
+    ['] flush-a-tos map-stacks1 ;
 
 : fill-a-tos { stack -- }
     stack stack-out @ 0= stack stack-in @ 0<> and
@@ -559,7 +649,7 @@ stack inst-stream IP Cell
 
 : fill-tos ( -- )
     \ !! inst-stream for prefetching?
-    ['] fill-a-tos map-stacks ;
+    ['] fill-a-tos map-stacks1 ;
 
 : fetch ( addr -- )
     dup item-type @ type-fetch @ execute ;
@@ -569,18 +659,16 @@ stack inst-stream IP Cell
 
 : stack-pointer-update { stack -- }
     \ stack grow downwards
-    stack stack-in @ stack stack-out @ -
+    stack stack-diff
     ?dup-if \ this check is not necessary, gcc would do this for us
-	stack stack-pointer 2@ type ."  += " 0 .r ." ;" cr
-    endif ;
-
-: inst-pointer-update ( -- )
-    inst-stream stack-in @ ?dup-if
-	." INC_IP(" 0 .r ." );" cr
+	stack inst-stream = if
+	    ." INC_IP(" 0 .r ." );" cr
+	else
+	    stack stack-pointer 2@ type ."  += " 0 .r ." ;" cr
+	endif
     endif ;
 
 : stack-pointer-updates ( -- )
-    inst-pointer-update
     ['] stack-pointer-update map-stacks ;
 
 : store ( item -- )
@@ -625,28 +713,44 @@ stack inst-stream IP Cell
     endif
     2drop ;
 
-: output-c-tail1 ( -- )
-    \ the final part of the generated C code except LABEL2 and NEXT_P2
+: output-label2 ( -- )
+    ." LABEL2(" prim prim-c-name 2@ type ." )" 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 ;
+    fill-tos 
+    xt execute
+    ." NEXT_P2;" cr ;
+
+: output-c-tail1-no-stores { xt -- }
+    \ the final part of the generated C code for combinations
+    output-super-end
+    ." NEXT_P1;" cr
+    fill-tos 
+    xt execute
+    ." NEXT_P2;" cr ;
 
 : output-c-tail ( -- )
-    \ the final part of the generated C code, without LABEL2
-    output-c-tail1
-    ." NEXT_P2;" ;
+    ['] noop output-c-tail1 ;
 
 : output-c-tail2 ( -- )
-    \ the final part of the generated C code, including LABEL2
-    output-c-tail1
-    ." LABEL2(" prim prim-c-name 2@ type ." )" cr
-    ." NEXT_P2;" cr ;
+    ['] output-label2 output-c-tail1 ;
+
+: output-c-tail-no-stores ( -- )
+    ['] noop 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 "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" INST_TAIL;" search
     while ( c-addr1 u1 c-addr3 u3 )
@@ -655,30 +759,29 @@ stack inst-stream IP Cell
 	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 ." )" ;
     
 : 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-tail2
- ." }" 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
+    flush-tos
+    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 -- }
@@ -715,14 +818,15 @@ 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
@@ -808,6 +912,9 @@ stack inst-stream IP Cell
 : output-alias ( -- ) 
     ( primitive-number @ . ." alias " ) ." Primitive " prim prim-name 2@ type cr ;
 
+: output-prim-num ( -- )
+    prim prim-num @ 8 + 4 .r space prim prim-name 2@ type cr ;
+
 : output-forth ( -- )  
     prim prim-forth-code @ 0=
     IF    	\ output-alias
@@ -919,6 +1026,7 @@ 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  !
@@ -930,24 +1038,33 @@ 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
@@ -980,6 +1097,27 @@ 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
@@ -987,6 +1125,7 @@ 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 ;
 
 \ C output
@@ -1012,10 +1151,20 @@ 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 ( -- )
+    max-stacks 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
@@ -1025,20 +1174,18 @@ 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 ;
 
@@ -1050,11 +1197,11 @@ stack inst-stream IP Cell
     print-declarations-combined
     ." NEXT_P0;" cr
     flush-tos
-    fetches
+    \ fetches \ now in parts
     \ print-debug-args
-    stack-pointer-updates
+    \ stack-pointer-updates now in parts
     output-parts
-    output-c-tail2
+    output-c-tail2-no-stores
     ." }" cr
     cr ;
 
@@ -1064,13 +1211,16 @@ 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 ." ,"
@@ -1078,15 +1228,53 @@ 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
-    cr ;
 
+\ 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 length;      /* number of components */
+\    int *components; /* array of vm_prim indexes 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.
+
+: 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, nor that IP updates are combined with
+    \ other stuff
+    0 max-stacks 0 +do
+	p prim-stacks-in i th @ +
+    loop
+    0 max-stacks 0 +do
+	p prim-stacks-out i th @ +
+    loop
+    0 max-stacks 0 +do
+	p prim-stacks-in i th @ p prim-stacks-out i th @ <> -
+    loop ;
+
+: output-num-part ( p -- )
+    prim-num @ 4 .r ." ," ;
+
+: output-costs ( -- )
+    ." {" prim compute-costs
+    rot 2 .r ." ," swap 2 .r ." ," 2 .r ." ,"
+    combined if
+	num-combined @ 2 .r
+	." , ((int []){" ['] output-num-part map-combined ." })}, /* "
+    else
+	."  1, ((int []){" prim prim-num @ 4 .r ." })}, /* "
+    endif
+    prim prim-name 2@ type ."  */"
+    cr ;
 
 \ the parser
 
@@ -1116,7 +1304,7 @@ print-token !
     \ 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
+    s" #line " r@ over compare if
 	rdrop 1 line +! EXIT
     endif
     0. r> 6 chars + 20 >number drop >r drop line ! r> ( c-addr )