--- gforth/prims2x.fs 2003/08/27 12:13:49 1.143 +++ gforth/prims2x.fs 2003/11/03 20:13:35 1.147 @@ -188,8 +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- ; @@ -246,6 +267,7 @@ struct% cell% 2* field prim-name cell% 2* field prim-wordset cell% 2* field prim-c-name + cell% 2* field prim-c-name-orig \ for reprocessed prims, the original name cell% 2* field prim-doc cell% 2* field prim-c-code cell% 2* field prim-forth-code @@ -269,6 +291,9 @@ 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 @@ -278,6 +303,9 @@ variable in-part \ true if processing a r> to prim throw ; +: prim-c-name-2! ( c-addr u -- ) + 2dup prim prim-c-name 2! prim prim-c-name-orig 2! ; + 1000 constant max-combined create combined-prims max-combined cells allot variable num-combined @@ -351,18 +379,36 @@ defer inst-stream-f ( -- stack ) \ stack access stuff : normal-stack-access0 { n stack -- } + \ n has the ss-offset already applied (see ...-access1) n stack stack-access-transform @ execute ." [" 0 .r ." ]" ; - -: normal-stack-access1 { n stack -- } - stack stack-pointer 2@ type - n if - n stack normal-stack-access0 + +: 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 - ." TOS" + drop 0 endif ; -: normal-stack-access ( n stack -- ) - dup inst-stream-f = if +: .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 @@ -391,7 +437,7 @@ defer inst-stream-f ( -- stack ) 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 + stack state-in normal-stack-access else drop n stack part-stack-access endif ; @@ -405,7 +451,7 @@ defer inst-stream-f ( -- stack ) stack stack-number @ part-num @ s-c-max-back-depth @ over <= if ( ndepth ) stack combined ['] stack-diff prim-context - - stack normal-stack-access + stack state-out normal-stack-access else drop n stack part-stack-access endif ; @@ -415,7 +461,7 @@ defer inst-stream-f ( -- stack ) in-part @ if part-stack-read else - normal-stack-access + state-in normal-stack-access endif ; : stack-write ( n stack -- ) @@ -423,7 +469,7 @@ defer inst-stream-f ( -- stack ) in-part @ if part-stack-write else - normal-stack-access + state-out normal-stack-access endif ; : item-in-index { item -- n } @@ -474,7 +520,7 @@ 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 -- ) @@ -486,17 +532,14 @@ defer inst-stream-f ( -- stack ) r@ item-out-index r@ item-stack @ stack-write ." );" rdrop ; -: store-single ( item -- ) - >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-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 \ in reg/mem + item item-out-index stack state-out stack-reg = and \ out reg/mem + 0= if + item really-store-single cr + endif ; : store-double ( item -- ) \ !! store optimization is not performed, because it is not yet needed @@ -614,21 +657,68 @@ 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 latestxt >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 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 @@ -658,26 +748,42 @@ 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 stack normal-stack-access0 ." = " type ." TOS);" cr - endif ; +: stack-state-items ( stack state -- n ) + state-ss ss-registers 2@ nip ; -: flush-tos ( -- ) - ['] flush-a-tos map-stacks1 ; +: 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 stack normal-stack-access0 ." );" 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-stacks1 ; + ['] fill-stack map-stacks1 ; : fetch ( addr -- ) dup item-type @ type-fetch @ execute ; @@ -685,6 +791,39 @@ stack inst-stream IP Cell : fetches ( -- ) prim prim-effect-in prim prim-effect-in-end @ ['] fetch map-items ; +: reg-reg-move ( reg-from reg-to -- ) + 2dup = if + 2drop + else + .reg ." = " .reg ." ;" cr + endif ; + +: 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 @@ -692,8 +831,11 @@ stack inst-stream IP Cell 0 r> execute - ; : stack-pointer-update { stack -- } + \ and moves \ stacks grow downwards - stack stack-diff + 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 @@ -701,7 +843,8 @@ stack inst-stream IP Cell stack stack-pointer 2@ type ." += " stack stack-update-transform 0 .r ." ;" cr endif - endif ; + endif + stack stack-moves ; : stack-pointer-updates ( -- ) ['] stack-pointer-update map-stacks ; @@ -748,6 +891,17 @@ stack inst-stream IP Cell endif 2drop ; + +defer output-nextp0 +:noname ( -- ) + ." NEXT_P0;" cr ; +is output-nextp0 + +defer output-nextp1 +:noname ( -- ) + ." NEXT_P1;" cr ; +is output-nextp1 + : output-nextp2 ( -- ) ." NEXT_P2;" cr ; @@ -762,16 +916,16 @@ variable tail-nextp2 \ xt to execute for \ the final part of the generated C code, with xt printing LABEL2 or not. output-super-end print-debug-results - ." NEXT_P1;" cr + output-nextp1 stores - fill-tos + 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-tos + output-nextp1 + fill-state xt execute ; : output-c-tail ( -- ) @@ -813,8 +967,8 @@ variable tail-nextp2 \ xt to execute for ." {" cr ." DEF_CA" cr print-declarations - ." NEXT_P0;" cr - flush-tos + output-nextp0 + spill-state fetches print-debug-args stack-pointer-updates @@ -936,7 +1090,7 @@ variable tail-nextp2 \ xt to execute for \ 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 @@ -945,7 +1099,7 @@ variable tail-nextp2 \ xt to execute for \ prim prim-c-code 2@ type \ ." }" cr \ stores -\ fill-tos +\ fill-state \ ." return (sp);" cr \ ." }" cr \ cr ; @@ -1172,6 +1326,82 @@ variable tail-nextp2 \ xt to execute for 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 -- } @@ -1239,8 +1469,8 @@ variable tail-nextp2 \ xt to execute for ." {" cr ." DEF_CA" cr print-declarations-combined - ." NEXT_P0;" cr - flush-tos + output-nextp0 + spill-state \ fetches \ now in parts \ print-debug-args \ stack-pointer-updates now in parts @@ -1279,11 +1509,14 @@ variable tail-nextp2 \ xt to execute for \ 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 */ +\ char loads; /* number of stack loads */ +\ char stores; /* number of stack stores */ +\ char updates; /* number of stack pointer updates */ +\ char branch; /* is it a branch (SET_IP) */ +\ char state_in; /* state on entry */ +\ char state_out; /* state on exit */ +\ short offset; /* offset into super2 table */ +\ char length; /* number of components */ \ }; \ How do you know which primitive or combined instruction this @@ -1313,7 +1546,7 @@ variable tail-nextp2 \ xt to execute for loop ; : output-num-part ( p -- ) - ." N_" prim-c-name 2@ type ." ," ; + ." N_" prim-c-name-orig 2@ type ." ," ; \ prim-num @ 4 .r ." ," ; : output-name-comment ( -- ) @@ -1324,7 +1557,9 @@ variable offset-super2 0 offset-super2 : output-costs-prefix ( -- ) ." {" prim compute-costs rot 2 .r ." ," swap 2 .r ." ," 2 .r ." , " - prim prim-branch? negate . ." ," ; + prim prim-branch? negate . ." ," + state-in state-number @ 2 .r ." ," + state-out state-number @ 2 .r ." ," ; : output-costs-gforth-simple ( -- ) output-costs-prefix @@ -1349,13 +1584,15 @@ variable offset-super2 0 offset-super2 output-name-comment cr ; -: output-super2 ( -- ) - \ table of superinstructions without requirement for existing prefixes - combined if - ['] output-num-part map-combined - else +: output-super2-simple ( -- ) + prim prim-c-name 2@ prim prim-c-name-orig 2@ d= if prim output-num-part - endif + output-name-comment + cr + endif ; + +: output-super2-combined ( -- ) + ['] output-num-part map-combined output-name-comment cr ; @@ -1516,7 +1753,7 @@ Variable c-flag (( {{ prim create-prim }} ` ( white ** {{ start }} stack-effect {{ end prim prim-stack-string 2! }} ` ) white ** (( {{ start }} forth-ident {{ end prim prim-wordset 2! }} white ** - (( {{ start }} c-ident {{ end prim prim-c-name 2! }} )) ?? + (( {{ start }} c-ident {{ end 2dup 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 }} @@ -1538,8 +1775,8 @@ Variable c-flag line @ name-line ! filename 2@ name-filename 2! function-number @ prim prim-num ! 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 ** )) ?? + 2dup prim prim-name 2! prim-c-name-2! }} white ** + (( ` / white ** {{ start }} c-ident {{ end prim-c-name-2! }} white ** )) ?? (( simple-primitive || combined-primitive )) {{ 1 function-number +! }} )) <- primitive ( -- )