Thanks for the reply, but I don’t think I’m ready to submit a patch yet.
Hopefully someday :)

Thanks,
Horsng Junn

On 2014-03-19, at 8:54 PM, Simon Marlow <marlowsd@gmail.com> wrote:

On 19/03/2014 03:06, Horsng Junn wrote:
Hi list!
This is my first mail to the list; please correct me if I made any mistakes.

Welcome :-)

I’ve been studying GHC source code for studying’s sake, and think I’ve found a few problems in this function.

Below comes the function in its current form:
————————————————————
__stg_gc_fun /* explicit stack */
{
    W_ size;
    W_ info;
    W_ type;

    info = %GET_FUN_INFO(UNTAG(R1));

    // cache the size
    type = TO_W_(StgFunInfoExtra_fun_type(info));
    if (type == ARG_GEN) {
size = BITMAP_SIZE(StgFunInfoExtra_bitmap(info));
    } else {
if (type == ARG_GEN_BIG) {
#ifdef TABLES_NEXT_TO_CODE
            // bitmap field holds an offset
            size = StgLargeBitmap_size( StgFunInfoExtra_bitmap(info)
                                        + %GET_ENTRY(UNTAG(R1)) /* ### */ );
#else
    size = StgLargeBitmap_size( StgFunInfoExtra_bitmap(info) );
#endif
} else {
    size = BITMAP_SIZE(W_[stg_arg_bitmaps + WDS(type)]);
}
    }

#ifdef NO_ARG_REGS
    // we don't have to save any registers away
    Sp_adj(-3);
    Sp(2) = R1;
    Sp(1) = size;
    Sp(0) = stg_gc_fun_info;
    jump stg_gc_noregs [];
#else
    W_ type;
    type = TO_W_(StgFunInfoExtra_fun_type(info));
    // cache the size
    if (type == ARG_GEN || type == ARG_GEN_BIG) {
        // regs already saved by the heap check code
        Sp_adj(-3);
        Sp(2) = R1;
        Sp(1) = size;
        Sp(0) = stg_gc_fun_info;
        // DEBUG_ONLY(foreign "C" debugBelch("stg_fun_gc_gen(ARG_GEN)"););
        jump stg_gc_noregs [];
    } else {
        jump W_[stg_stack_save_entries + WDS(type)] [*]; // all regs live
    // jumps to stg_gc_noregs after saving stuff
    }
#endif /* !NO_ARG_REGS */
}

The problems I see:
1. “type” variable is declared and its value calculated twice,

Yes.

2. if (type == ARG_GEN … check is needlessly done twice,
3. “size" value is calculated and then discarded in !NO_ARG_REGS, for types other than ARG_GEN or ARG_GEN_BIG, and

Yes, but this isn't a performance-critical bit of code, and a good backend will be able to optimise it anyway.  Clarity is more important.

4. (this is a minor one) Sp_adj(-3) ~ jmup_stg_gc_noregs code is duplicated. I know they’re in separate #if branch but still...
And my proposed fixes are:

I think your version has more complex control-flow: whether we drop out of the if expression to the code that follows depends on an #ifdef. I'd rather have something like this:

if (MAX_REAL_VANILLA_REG < 2 || type == ARG_GEN || type == ARG_GEN_BIG) {
 Sp_adj(-3);
 ...
} else {
 jump W_[stg_stack_save_entries + WDS(type)] [*];
}

That way the tail-calls are all at the end of the control flow.

If you validate a patch and attach it to a ticket, I'll take a look.

Cheers,
Simon



————————————————————
__stg_gc_fun /* explicit stack */
{
    W_ size;
    W_ info;
    W_ type;

    info = %GET_FUN_INFO(UNTAG(R1));

    // cache the size
    type = TO_W_(StgFunInfoExtra_fun_type(info));
    if (type == ARG_GEN) {
        // regs already saved by the heap check code
size = BITMAP_SIZE(StgFunInfoExtra_bitmap(info));
    } else {
if (type == ARG_GEN_BIG) {
    // regs already saved by the heap check code
#ifdef TABLES_NEXT_TO_CODE
            // bitmap field holds an offset
            size = StgLargeBitmap_size( StgFunInfoExtra_bitmap(info)
                                        + %GET_ENTRY(UNTAG(R1)) /* ### */ );
#else
    size = StgLargeBitmap_size( StgFunInfoExtra_bitmap(info) );
#endif
} else {
#ifdef NO_ARG_REGS
    // we don't have to save any registers away
    size = BITMAP_SIZE(W_[stg_arg_bitmaps + WDS(type)]);
#else
    jump W_[stg_stack_save_entries + WDS(type)] [*]; // all regs live
    // jumps to stg_gc_noregs after saving stuff
#endif
}
    }

    Sp_adj(-3);
    Sp(2) = R1;
    Sp(1) = size;
    Sp(0) = stg_gc_fun_info;
    jump stg_gc_noregs [];
}

What do you guys think?

Thanks,
Horsng Junn

_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://www.haskell.org/mailman/listinfo/ghc-devs