a more efficient tail call by using the new TUCK instruction

This commit is contained in:
kr.angelov
2014-10-30 13:09:50 +00:00
parent 0519493ca9
commit 9b0f354c7c
5 changed files with 139 additions and 54 deletions

View File

@@ -7,6 +7,7 @@ import PGF(CId,utf8CId)
import PGF.Internal(CodeLabel,Instr(..),IVal(..),TailInfo(..),Literal(..)) import PGF.Internal(CodeLabel,Instr(..),IVal(..),TailInfo(..),Literal(..))
import qualified Data.Map as Map import qualified Data.Map as Map
import Data.List(nub,mapAccumL) import Data.List(nub,mapAccumL)
import Data.Maybe(fromMaybe)
generateByteCode :: SourceGrammar -> Int -> [L Equation] -> [[Instr]] generateByteCode :: SourceGrammar -> Int -> [L Equation] -> [[Instr]]
generateByteCode gr arity eqs = generateByteCode gr arity eqs =
@@ -88,10 +89,9 @@ compileBody gr arity st vs e bs =
let eval fun args let eval fun args
| arity == 0 = let (st1,is) = pushArgs (st+2) (reverse args) | arity == 0 = let (st1,is) = pushArgs (st+2) (reverse args)
fun' = shiftIVal st1 fun fun' = shiftIVal st1 fun
in [PUSH_FRAME]++is++[EVAL fun' (UpdateCall st st1)] in [PUSH_FRAME]++is++[EVAL fun' UpdateCall]
| otherwise = let (st1,is) = pushArgs st (reverse args) | otherwise = let (st1,fun',is) = tuckArgs arity st fun args
fun' = shiftIVal st1 fun in is++[EVAL fun' (TailCall (st1-length args-1))]
in is++[EVAL fun' (TailCall arity st st1)]
(heap,bs1,is) = compileFun gr eval st vs e 0 bs [] (heap,bs1,is) = compileFun gr eval st vs e 0 bs []
in (bs1,if heap > 0 then (ALLOC heap : is) else is) in (bs1,if heap > 0 then (ALLOC heap : is) else is)
@@ -119,7 +119,13 @@ compileFun gr eval st vs (Q (m,id)) h0 bs args =
in (h1,bs,PUT_CONSTR (i2i id):is1++eval (HEAP h0) []) in (h1,bs,PUT_CONSTR (i2i id):is1++eval (HEAP h0) [])
else let h1 = h0 + 1 + n_args else let h1 = h0 + 1 + n_args
is2 = [SET (FREE_VAR i) | i <- [0..n_args-1]] ++ [SET (ARG_VAR (i+1)) | i <- [0..diff-1]] is2 = [SET (FREE_VAR i) | i <- [0..n_args-1]] ++ [SET (ARG_VAR (i+1)) | i <- [0..diff-1]]
b = CHECK_ARGS diff : ALLOC (c_arity+2) : PUT_CONSTR (i2i id) : is2 ++ [EVAL (HEAP h0) (TailCall diff (diff+1) (diff+1))] b = CHECK_ARGS diff :
ALLOC (c_arity+2) :
PUT_CONSTR (i2i id) :
is2 ++
TUCK (ARG_VAR 0) diff :
EVAL (HEAP h0) (TailCall diff) :
[]
in (h1,b:bs,PUT_CLOSURE (length bs):is1++eval (HEAP h0) []) in (h1,b:bs,PUT_CLOSURE (length bs):is1++eval (HEAP h0) [])
compileFun gr eval st vs (QC qid) h0 bs args = compileFun gr eval st vs (QC qid) h0 bs args =
compileFun gr eval st vs (Q qid) h0 bs args compileFun gr eval st vs (Q qid) h0 bs args
@@ -159,7 +165,13 @@ compileArg gr st vs (Q(m,id)) h0 bs =
in if c_arity == 0 in if c_arity == 0
then (h0,bs,GLOBAL (i2i id),[]) then (h0,bs,GLOBAL (i2i id),[])
else let is2 = [SET (ARG_VAR (i+1)) | i <- [0..c_arity-1]] else let is2 = [SET (ARG_VAR (i+1)) | i <- [0..c_arity-1]]
b = CHECK_ARGS c_arity : ALLOC (c_arity+2) : PUT_CONSTR (i2i id) : is2 ++ [EVAL (HEAP h0) (TailCall c_arity (c_arity+1) (c_arity+1))] b = CHECK_ARGS c_arity :
ALLOC (c_arity+2) :
PUT_CONSTR (i2i id) :
is2 ++
TUCK (ARG_VAR 0) c_arity :
EVAL (HEAP h0) (TailCall c_arity) :
[]
h1 = h0 + 2 h1 = h0 + 2
in (h1,b:bs,HEAP h0,[PUT_CLOSURE (length bs),SET_PAD]) in (h1,b:bs,HEAP h0,[PUT_CLOSURE (length bs),SET_PAD])
compileArg gr st vs (QC qid) h0 bs = compileArg gr st vs (QC qid) h0 bs =
@@ -206,7 +218,13 @@ compileArg gr st vs e h0 bs =
in (h2,bs1,HEAP h1,is1 ++ (PUT_CONSTR (i2i id) : is2)) in (h2,bs1,HEAP h1,is1 ++ (PUT_CONSTR (i2i id) : is2))
else let h2 = h1 + 1 + n_args else let h2 = h1 + 1 + n_args
is2 = [SET (FREE_VAR i) | i <- [0..n_args-1]] ++ [SET (ARG_VAR (i+1)) | i <- [0..diff-1]] is2 = [SET (FREE_VAR i) | i <- [0..n_args-1]] ++ [SET (ARG_VAR (i+1)) | i <- [0..diff-1]]
b = CHECK_ARGS diff : ALLOC (c_arity+2) : PUT_CONSTR (i2i id) : is2 ++ [EVAL (HEAP h0) (TailCall diff (diff+1) (diff+1))] b = CHECK_ARGS diff :
ALLOC (c_arity+2) :
PUT_CONSTR (i2i id) :
is2 ++
TUCK (ARG_VAR 0) diff :
EVAL (HEAP h0) (TailCall diff) :
[]
in (h2,b:bs1,HEAP h1,is1 ++ (PUT_CLOSURE (length bs):is2)) in (h2,b:bs1,HEAP h1,is1 ++ (PUT_CLOSURE (length bs):is2))
Nothing -> compileLambda gr st vs [] e h0 bs Nothing -> compileLambda gr st vs [] e h0 bs
@@ -241,6 +259,28 @@ pushArgs st [] = (st,[])
pushArgs st (arg:args) = let (st1,is) = pushArgs (st+1) args pushArgs st (arg:args) = let (st1,is) = pushArgs (st+1) args
in (st1, PUSH (shiftIVal st arg) : is) in (st1, PUSH (shiftIVal st arg) : is)
tuckArgs arity st fun args = (st2,shiftIVal st2 fun',is1++is2)
where
(st2,fun',is2) = tucks st1 0 fun tas
(st1,is1) = pushArgs st pas
(tas,pas) = splitAt st args'
args' = reverse (ARG_VAR arity : args)
tucks st i fun [] = (st,fun,[])
tucks st i fun (arg:args)
| arg == ARG_VAR i = tucks st (i+1) fun args
| otherwise = case save st (ARG_VAR i) (fun:args) of
Just (fun:args) -> let (st1,fun',is) = tucks (st+1) (i+1) fun args
in (st1, fun', PUSH (ARG_VAR (st-i-1)) :
TUCK (shiftIVal (st+1) arg) (st-i) : is)
Nothing -> let (st1,fun',is) = tucks st (i+1) fun args
in (st1, fun', TUCK (shiftIVal st arg) (st-i-1) : is)
save st arg0 [] = Nothing
save st arg0 (arg:args)
| arg0 == arg = Just (ARG_VAR st1 : fromMaybe args (save st arg0 args))
| otherwise = fmap (arg :) (save st arg0 args)
setArgs st [] = [] setArgs st [] = []
setArgs st (arg:args) = SET (shiftIVal st arg) : setArgs st args setArgs st (arg:args) = SET (shiftIVal st arg) : setArgs st args

View File

@@ -136,11 +136,12 @@ typedef enum {
PGF_INSTR_SET_PAD = 9, PGF_INSTR_SET_PAD = 9,
PGF_INSTR_PUSH_FRAME = 10, PGF_INSTR_PUSH_FRAME = 10,
PGF_INSTR_PUSH = 11, PGF_INSTR_PUSH = 11,
PGF_INSTR_EVAL = 12, PGF_INSTR_TUCK = 12,
PGF_INSTR_DROP = 15, PGF_INSTR_EVAL = 13,
PGF_INSTR_JUMP = 16, PGF_INSTR_DROP = 16,
PGF_INSTR_FAIL = 17, PGF_INSTR_JUMP = 17,
PGF_INSTR_ADD = 18, PGF_INSTR_FAIL = 18,
PGF_INSTR_ADD = 19,
} PgfInstruction; } PgfInstruction;
typedef GuSeq PgfConcrs; typedef GuSeq PgfConcrs;

View File

@@ -1000,6 +1000,61 @@ pgf_jit_function(PgfReader* rdr, PgfAbstr* abstr,
} }
break; break;
} }
case PGF_INSTR_TUCK: {
switch (mod) {
case 0: {
size_t offset = pgf_read_int(rdr);
size_t sindex = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG
gu_printf(out, err, "TUCK hp(%d) %d\n", offset, sindex);
#endif
if (offset == 0) {
jit_stxi_p(sizeof(PgfClosure*)*sindex, JIT_SP, JIT_VHEAP);
} else {
jit_addi_p(JIT_R0, JIT_VHEAP, offset*sizeof(void*));
jit_stxi_p(sizeof(PgfClosure*)*sindex, JIT_SP, JIT_R0);
}
break;
}
case 1: {
size_t index = pgf_read_int(rdr);
size_t sindex = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG
gu_printf(out, err, "TUCK stk(%d) %d\n", index, sindex);
#endif
jit_ldxi_p(JIT_R0, JIT_SP, index*sizeof(PgfClosure*));
jit_stxi_p(sizeof(PgfClosure*)*sindex, JIT_SP, JIT_R0);
break;
}
case 2: {
size_t index = pgf_read_int(rdr);
size_t sindex = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG
gu_printf(out, err, "TUCK env(%d) %d\n", index, sindex);
#endif
jit_ldxi_p(JIT_R0, JIT_VCLOS, sizeof(PgfClosure)+index*sizeof(PgfClosure*));
jit_stxi_p(sizeof(PgfClosure*)*sindex, JIT_SP, JIT_R0);
break;
}
case 3: {
PgfCId id = pgf_read_cid(rdr, rdr->tmp_pool);
size_t sindex = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG
gu_printf(out, err, "TUCK %s %d\n", id, sindex);
#endif
PgfCallPatch patch;
patch.cid = id;
patch.ref = jit_movi_p(JIT_R0, jit_forward());
gu_buf_push(rdr->jit_state->call_patches, PgfCallPatch, patch);
jit_stxi_p(sizeof(PgfClosure*)*sindex, JIT_SP, JIT_R0);
break;
}
default:
gu_impossible();
}
break;
}
case PGF_INSTR_EVAL+0: case PGF_INSTR_EVAL+0:
case PGF_INSTR_EVAL+2: case PGF_INSTR_EVAL+2:
case PGF_INSTR_EVAL+1: { case PGF_INSTR_EVAL+1: {
@@ -1057,36 +1112,18 @@ pgf_jit_function(PgfReader* rdr, PgfAbstr* abstr,
} }
case 1: { case 1: {
size_t a = pgf_read_int(rdr); size_t a = pgf_read_int(rdr);
size_t b = pgf_read_int(rdr);
size_t c = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG #ifdef PGF_JIT_DEBUG
gu_printf(out, err, " tail(%d,%d,%d)\n", a, b, c); gu_printf(out, err, " tail(%d)\n", a);
#endif #endif
jit_addi_p(JIT_SP, JIT_SP, a*sizeof(PgfClosure*));
jit_ldxi_p(JIT_R2, JIT_SP, sizeof(PgfClosure*)*(c-a-1));
for (size_t i = 0; i < c-b; i++) {
jit_ldxi_p(JIT_R1, JIT_SP, sizeof(PgfClosure*)*((c-b-1)-i));
jit_stxi_p(sizeof(PgfClosure*)*((c-1)-i), JIT_SP, JIT_R1);
}
jit_addi_p(JIT_SP, JIT_SP, b*sizeof(PgfClosure*));
jit_pushr_p(JIT_R2);
jit_jmpr(JIT_R0); jit_jmpr(JIT_R0);
break; break;
} }
case 2: { case 2: {
size_t b = pgf_read_int(rdr);
size_t c = pgf_read_int(rdr);
#ifdef PGF_JIT_DEBUG #ifdef PGF_JIT_DEBUG
gu_printf(out, err, " update(%d,%d)\n", b, c); gu_printf(out, err, " update\n");
#endif #endif
if (b > 1) {
for (size_t i = 0; i < c-b; i++) {
jit_ldxi_p(JIT_R1, JIT_SP, sizeof(PgfClosure*)*(c-b-1-i));
jit_stxi_p(sizeof(PgfClosure*)*(c-2-i), JIT_SP, JIT_R1);
}
jit_addi_p(JIT_SP, JIT_SP, (b-1)*sizeof(PgfClosure*));
}
jit_movi_p(JIT_R1, abstr->eval_gates->update_closure); jit_movi_p(JIT_R1, abstr->eval_gates->update_closure);
jit_pushr_p(JIT_R1); jit_pushr_p(JIT_R1);
jit_jmpr(JIT_R0); jit_jmpr(JIT_R0);

View File

@@ -158,22 +158,26 @@ instance Binary Instr where
put (PUSH (ARG_VAR n)) = putWord8 45 >> put n put (PUSH (ARG_VAR n)) = putWord8 45 >> put n
put (PUSH (FREE_VAR n)) = putWord8 46 >> put n put (PUSH (FREE_VAR n)) = putWord8 46 >> put n
put (PUSH (GLOBAL id)) = putWord8 47 >> put id put (PUSH (GLOBAL id)) = putWord8 47 >> put id
put (EVAL (HEAP n) (RecCall )) = putWord8 48 >> put n put (TUCK (HEAP n) i) = putWord8 48 >> put (n,i)
put (EVAL (ARG_VAR n) (RecCall )) = putWord8 49 >> put n put (TUCK (ARG_VAR n) i) = putWord8 49 >> put (n,i)
put (EVAL (FREE_VAR n) (RecCall )) = putWord8 50 >> put n put (TUCK (FREE_VAR n) i) = putWord8 50 >> put (n,i)
put (EVAL (GLOBAL id) (RecCall )) = putWord8 51 >> put id put (TUCK (GLOBAL id) i) = putWord8 51 >> put (id,i)
put (EVAL (HEAP n) (TailCall a b c)) = putWord8 52 >> put n >> put (a,b,c) put (EVAL (HEAP n) RecCall) = putWord8 52 >> put n
put (EVAL (ARG_VAR n) (TailCall a b c)) = putWord8 53 >> put n >> put (a,b,c) put (EVAL (ARG_VAR n) RecCall) = putWord8 53 >> put n
put (EVAL (FREE_VAR n) (TailCall a b c)) = putWord8 54 >> put n >> put (a,b,c) put (EVAL (FREE_VAR n) RecCall) = putWord8 54 >> put n
put (EVAL (GLOBAL id) (TailCall a b c)) = putWord8 55 >> put id >> put (a,b,c) put (EVAL (GLOBAL id) RecCall) = putWord8 55 >> put id
put (EVAL (HEAP n) (UpdateCall b c)) = putWord8 56 >> put n >> put (b,c) put (EVAL (HEAP n) (TailCall a)) = putWord8 56 >> put n >> put a
put (EVAL (ARG_VAR n) (UpdateCall b c)) = putWord8 57 >> put n >> put (b,c) put (EVAL (ARG_VAR n) (TailCall a)) = putWord8 57 >> put n >> put a
put (EVAL (FREE_VAR n) (UpdateCall b c)) = putWord8 58 >> put n >> put (b,c) put (EVAL (FREE_VAR n) (TailCall a)) = putWord8 58 >> put n >> put a
put (EVAL (GLOBAL id) (UpdateCall b c)) = putWord8 59 >> put id >> put (b,c) put (EVAL (GLOBAL id) (TailCall a)) = putWord8 59 >> put id >> put a
put (DROP n ) = putWord8 60 >> put n put (EVAL (HEAP n) UpdateCall) = putWord8 60 >> put n
put (JUMP l ) = putWord8 64 >> put l put (EVAL (ARG_VAR n) UpdateCall) = putWord8 61 >> put n
put (FAIL ) = putWord8 68 put (EVAL (FREE_VAR n) UpdateCall) = putWord8 62 >> put n
put (ADD ) = putWord8 72 put (EVAL (GLOBAL id) UpdateCall) = putWord8 63 >> put id
put (DROP n ) = putWord8 64 >> put n
put (JUMP l ) = putWord8 68 >> put l
put (FAIL ) = putWord8 72
put (ADD ) = putWord8 76
instance Binary Type where instance Binary Type where
put (DTyp hypos cat exps) = put (hypos,cat,exps) put (DTyp hypos cat exps) = put (hypos,cat,exps)

View File

@@ -27,6 +27,7 @@ data Instr
| SET_PAD | SET_PAD
| PUSH_FRAME | PUSH_FRAME
| PUSH IVal | PUSH IVal
| TUCK IVal {-# UNPACK #-} !Int
| EVAL IVal TailInfo | EVAL IVal TailInfo
| DROP {-# UNPACK #-} !Int | DROP {-# UNPACK #-} !Int
| JUMP {-# UNPACK #-} !CodeLabel | JUMP {-# UNPACK #-} !CodeLabel
@@ -38,11 +39,12 @@ data IVal
| ARG_VAR {-# UNPACK #-} !Int | ARG_VAR {-# UNPACK #-} !Int
| FREE_VAR {-# UNPACK #-} !Int | FREE_VAR {-# UNPACK #-} !Int
| GLOBAL CId | GLOBAL CId
deriving Eq
data TailInfo data TailInfo
= RecCall = RecCall
| TailCall {-# UNPACK #-} !Int {-# UNPACK #-} !Int {-# UNPACK #-} !Int | TailCall {-# UNPACK #-} !Int
| UpdateCall {-# UNPACK #-} !Int {-# UNPACK #-} !Int | UpdateCall
ppLit (LStr s) = text (show s) ppLit (LStr s) = text (show s)
ppLit (LInt n) = int n ppLit (LInt n) = int n
@@ -65,6 +67,7 @@ ppInstr (SET_PAD ) = text "SET_PAD"
ppInstr (PUSH_FRAME ) = text "PUSH_FRAME" ppInstr (PUSH_FRAME ) = text "PUSH_FRAME"
ppInstr (PUSH v) = text "PUSH " <+> ppIVal v ppInstr (PUSH v) = text "PUSH " <+> ppIVal v
ppInstr (EVAL v ti) = text "EVAL " <+> ppIVal v <+> ppTailInfo ti ppInstr (EVAL v ti) = text "EVAL " <+> ppIVal v <+> ppTailInfo ti
ppInstr (TUCK v n ) = text "TUCK " <+> ppIVal v <+> int n
ppInstr (DROP n ) = text "DROP " <+> int n ppInstr (DROP n ) = text "DROP " <+> int n
ppInstr (JUMP l ) = text "JUMP " <+> ppLabel l ppInstr (JUMP l ) = text "JUMP " <+> ppLabel l
ppInstr (FAIL ) = text "FAIL" ppInstr (FAIL ) = text "FAIL"
@@ -75,8 +78,8 @@ ppIVal (ARG_VAR n) = text "stk" <> parens (int n)
ppIVal (FREE_VAR n) = text "env" <> parens (int n) ppIVal (FREE_VAR n) = text "env" <> parens (int n)
ppIVal (GLOBAL id) = ppCId id ppIVal (GLOBAL id) = ppCId id
ppTailInfo RecCall = empty ppTailInfo RecCall = empty
ppTailInfo (TailCall a b c) = text "tail" <> parens (int a <> comma <> int b <> comma <> int c) ppTailInfo (TailCall n) = text "tail" <> parens (int n)
ppTailInfo (UpdateCall b c) = text "update" <> parens (int b <> comma <> int c) ppTailInfo UpdateCall = text "update"
ppLabel l = text (let s = show l in replicate (3-length s) '0' ++ s) ppLabel l = text (let s = show l in replicate (3-length s) '0' ++ s)