factorial: let's get ahead of jhc! :)
Hello glasgow-haskell-users, i propose to do for start rather small change that will allow to make things go several times faster and in particular outperform jhc for the small "leaf" loops (i.e. loops that use only GHC primitives and don't call other functions). that include factorial, "a[]+=b[i]" loop that was discussed here several weeks ago, some shootout examples, my own arrays serialization code and so on, so on the idea is that the following STG code f :: Int# -> Double# -> ... -> Int# (i.e. function use only unboxed values/results) f x y z = code1 in case expr1 of A -> codeA B -> codeB C -> codeC in f x_c y_c z_c D -> codeD in f x_d y_d z_d _ -> code0 in f x0 y0 z0 can be compiled to the following C/C-- code: f() { x = sp[0]; y = sp[4]; z = sp[8]; sp+=12; code1; while (expr1!=A) { if (expr1==B) then return codeB; else if (expr1==C) then {x=x_c; y=y_c; z=z_c;} else if (expr1==D) then {x=x_d; y=y_d; z=z_d;} else {x=x0; y=y0; z=z0;} code1; } return codeA; } this compilation don't require any changes in GHC memory model. all we need: 1) add for/if/while to the C-- statement types list (data CmmStmt) 2) implement recognizer for such simple STG functions (leaf unboxed procedures recognizer) 3) implement the translation itself as i said, it's should be no more than one day of work. i even think that it's one day for me and one hour for Simon Marlow :) Simon, how about this? i can even make the patches over current 6.6 sources and you will apply them at morning and we will see whether it work :) i yet never tried to rebuild the whole ghc :) -- Best regards, Bulat mailto:Bulat.Ziganshin@gmail.com
Bulat Ziganshin wrote:
i propose to do for start rather small change that will allow to make things go several times faster and in particular outperform jhc for the small "leaf" loops (i.e. loops that use only GHC primitives and don't call other functions). that include factorial, "a[]+=b[i]" loop that was discussed here several weeks ago, some shootout examples, my own arrays serialization code and so on, so on
the idea is that the following STG code
f :: Int# -> Double# -> ... -> Int#
(i.e. function use only unboxed values/results)
f x y z = code1 in case expr1 of A -> codeA B -> codeB C -> codeC in f x_c y_c z_c D -> codeD in f x_d y_d z_d _ -> code0 in f x0 y0 z0
can be compiled to the following C/C-- code:
f() { x = sp[0]; y = sp[4]; z = sp[8]; sp+=12;
code1; while (expr1!=A) { if (expr1==B) then return codeB; else if (expr1==C) then {x=x_c; y=y_c; z=z_c;} else if (expr1==D) then {x=x_d; y=y_d; z=z_d;} else {x=x0; y=y0; z=z0;} code1; } return codeA; }
this compilation don't require any changes in GHC memory model. all we need:
1) add for/if/while to the C-- statement types list (data CmmStmt)
Please don't extend the C-- data type - it's very small and simple because that makes it easy to manipulate and reason about. I don't see why you need to, either: you can already express for, if and while in C-- using conditional branches. I don't think gcc cares whether you write your code using labels and goto or while/for/if, it generates the same code either way.
2) implement recognizer for such simple STG functions (leaf unboxed procedures recognizer)
3) implement the translation itself
By all means try this. What you want is to compile recursive functions like this: f() { x = arg1; y = arg2; L: ... body of f, with args mapped to x & y, and recursive calls jumping to L passing args in x & y. } It's quite hard to do this as a C-- to C-- optimisation, at least without implementing a lot of other optimisations that we don't already have. I was hoping that gcc would do it for us, if we compile code like this: f() { L: ... body of f ... goto L; } but sadly it doesn't do the whole job. (see cmmLoopifyForC in cmm/CmmOpt.hs, which I added today). So you might try the hacky way of doing this transformation at a higher level, when generating the C-- in the first place. Putting the args in temporaries is the easy bit; generating different code for the recursive call will be more tricky. Cheers, Simon
Hello Simon, Friday, February 24, 2006, 7:18:30 PM, you wrote:
1) add for/if/while to the C-- statement types list (data CmmStmt)
SM> Please don't extend the C-- data type - it's very small and simple SM> because that makes it easy to manipulate and reason about. SM> I don't see why you need to, either: you can already express for, if and SM> while in C-- using conditional branches. I don't think gcc cares SM> whether you write your code using labels and goto or while/for/if, it SM> generates the same code either way. i tested gcc (3.4.2, included in ghc 6.4.1) and my tests show that gcc unroll loops only when a loop variable modified directly. This loop unrolls: loop: ... i=i-1; if(i>1) goto loop; and this loop don't unrolls: loop: ... x=i-1; i=x; if(i>1) goto loop; but using of while/goto don't matters. i attached fac.c that shows this - here only fac() and fac3() are unrolled
2) implement recognizer for such simple STG functions (leaf unboxed procedures recognizer)
3) implement the translation itself
SM> By all means try this. What you want is to compile recursive functions SM> like this: SM> f() { SM> x = arg1; SM> y = arg2; SM> L: SM> ... body of f, with args mapped to x & y, SM> and recursive calls jumping to L passing args in x & y. SM> } SM> It's quite hard to do this as a C-- to C-- optimisation, at least SM> without implementing a lot of other optimisations that we don't already SM> have. I was hoping that gcc would do it for us, if we compile code like SM> this: SM> f() { SM> L: SM> ... body of f ... SM> goto L; SM> } SM> but sadly it doesn't do the whole job. (see cmmLoopifyForC in SM> cmm/CmmOpt.hs, which I added today). SM> So you might try the hacky way of doing this transformation at a higher SM> level, when generating the C-- in the first place. Putting the args in SM> temporaries is the easy bit; generating different code for the recursive SM> call will be more tricky. i just downloaded fresh sources. cmmLoopifyForC may be not what we need. the whole idea of transformation is to move f() parameters to local variables, and then update contents of these local variables and perform `goto` on each recursive call. why cmmLoopifyForC is not good for this? only because before processing by this function the code should look like this: f() { x = arg1; y = arg2; L: ... body of f, with args mapped to x & y, x = newarg1; y = newarg2 jump f(); } what is just not proper code, especially if x&y is local f() variables. so to use this function we should start from generating improper code and anyway we should change STG->C-- code generation i think that the discussed optimization can be done in another way: 1) in function mkFunEntryCode add check that f() is a "leaf" "unboxed" function 2) if it's true then replace call to bindArgsToStack with generation of "x=sp[0]; y=sp[4]; L:" assignments and bind corresponding args to x&y (instead of binding them to sp[0]/sp[4]) 3) in cgTailCall function check that it's a self-recursive call and in this case generate sequence "x=newarg1; y=newarg2; goto L" instead of "jump f()". i think that this check will require adding info about f() to the FCode monad. so, the whole story should look like this: CgClosure.lhs: - ; bindArgsToStack stk_args + ; if isLeafUnboxedFunction cl_info body + then do local_args <- makeNewLocals (length stk_args) -- generate x&y locals + generateLoads local_args stk_args -- generate "x=sp[0]; y=sp[4]" assignments + lbl <- generateNewLabel -- generate "L:" + bindStkArgstoLocals local_args stk_args -- bind arguments to x&y local variables + modifyFCodeState (Just cl_info) lbl -- put in FCode monad state info about + -- now-processed leaf function and label + -- marking start of its self-recursive cycle + else do bindArgsToStack stk_args + modifyFCodeState Nothing undefined CgTailCall.lhs: - = do { fun_info <- getCgIdInfo fun + = do { fun_info <- getCgIdInfo fun + ; (optimized_fun, lbl) <- getFromFCodeState -- optimized_fun/=Nothing only if we + -- process "leaf" "unboxed" function - else -- Normal case, fun is boxed + else if (Just fun == optimized_fun) then + do generateAssignments -- generate "x=newarg1; y=newarg2" + generateGoto lbl -- generate "goto L" + else -- Normal case, fun is boxed -- Best regards, Bulat mailto:Bulat.Ziganshin@gmail.com
participants (2)
-
Bulat Ziganshin -
Simon Marlow