Re: [GHC] #8275: Loopification breaks profiling

16 Sep 2013

      #8275: Loopification breaks profiling
----------------------------------------+----------------------------------
        Reporter:  jstolarek            |            Owner:  jstolarek
            Type:  bug                  |           Status:  new
        Priority:  highest              |        Milestone:
       Component:  Profiling            |          Version:  7.7
      Resolution:                       |         Keywords:
Operating System:  Unknown/Multiple     |     Architecture:
 Type of failure:  Building GHC failed  |  Unknown/Multiple
       Test Case:                       |       Difficulty:  Unknown
        Blocking:  8298                 |       Blocked By:
                                        |  Related Tickets:
----------------------------------------+----------------------------------

Comment (by jstolarek):

 Loopification triggers for fully-saturated (but not over-saturated!) tail
 calls. So in this code:

 {{{
 f 0 = 4
 f 1 = 5
 f n = case f (n - 2) of
         4 -> 4
         5 -> f (n - 1)
 }}}

 If will trigger for call to `f (n - 1)` in second branch of case, but will
 not trigger for `f (n - 2)` in case scrutinee. I checked and this code
 actually causes segfault when compiled with `-prof -fprof-auto -rtsopts`
 (assuming you add `main = print (f 5)` in the file), but it only happens
 when `f :: Integer -> Integer` and not when `f :: Int -> Int`, so I
 suspect that the bug might actually be hidden somewhere in the libraries
 and I might be looking at wrong code.

 The idea behind the loopification is that it should put parameters in the
 local variables (instead of global registers) and make a jump (instead of
 call). `f` function begins like this in Cmm:

 {{{
        cYp:
            _sUP::P64 = R2;
            _sUO::P64 = R1;
            if (%MO_UU_Conv_W32_W64(I32[era]) <= 0) goto cW1; else goto
 cVZ;
        cVZ:
            I64[R1 + 15] = I64[R1 + 15] & 1152921503533105152 |
 %MO_UU_Conv_W32_W64(I32[era]) | 1152921504606846976;
            goto cW1;
        cW1:
            if (Sp - 104 < SpLim) goto cYq; else goto cYr;
        cYr:
            Hp = Hp + 40;
            if (Hp > HpLim) goto cYt; else goto cYs
 }}}

 Without loopification tail call will be a normal call:

 {{{
        cYM:
            I64[CCCS + 72] = I64[CCCS + 72] + %MO_UU_Conv_W64_W64(6 - 2);
            I64[Hp - 40] = sat_sVa_info;
            I64[Hp - 32] = CCCS;
            I64[Hp - 24] = (%MO_UU_Conv_W32_W64(I32[era]) << 30) | 0;
            P64[Hp - 8] = _sUN::P64;
            P64[Hp] = _sUP::P64;
            _cXT::P64 = Hp - 40;
            R2 = _cXT::P64;
            R1 = _sUO::P64;
            Sp = Sp + 40;
            call f1_sUQ_info(R2, R1) args: 8, res: 0, upd: 8;
 }}}

 With loopification we get:

 {{{
        cYM:
            I64[CCCS + 72] = I64[CCCS + 72] + %MO_UU_Conv_W64_W64(6 - 2);
            I64[Hp - 40] = sat_sV8_info;
            I64[Hp - 32] = CCCS;
            I64[Hp - 24] = (%MO_UU_Conv_W32_W64(I32[era]) << 30) | 0;
            P64[Hp - 8] = _sUL::P64;
            P64[Hp] = _sUP::P64;
            _cXT::P64 = Hp - 40;
            _sUP::P64 = _cXT::P64;
            goto cW2;
        cW2:
            if (Sp - 104 < SpLim) goto uZq; else goto uZp;
        uZq:
            Sp = Sp + 40;
            goto cYq;
        cYq:
            R2 = _sUP::P64;
            R1 = _sUO::P64;
            call (stg_gc_fun)(R2, R1) args: 8, res: 0, upd: 8;
        uZp:
            Sp = Sp + 40;
            goto cYr;
 }}}

 What might be surprising is that value of `_sUO` is not set before making
 tail call, but that *seems* to be OK - it is only shuffled between the
 stack and local variable. Note also that loopified call doesn't jump
 directly to second label `cVZ`, but instead it jumps to `cYr`. In
 principle this is OK (we want to skip stack check but not heap check), but
 TBH I can't tell whether in this case that is correct - I don't know what
 the magical numbers in `cVZ` do.

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8275#comment:5
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler