High-level Cmm code and stack allocation

Herbert Valerio Riedel hvriedel at gmail.com
Sun Jan 5 11:46:04 UTC 2014


On 2014-01-05 at 01:15:53 +0100, Carter Schonwald wrote:
> hey Herbert,
> I generally start with looking at the primops.cmm file for examples
> https://github.com/ghc/ghc/blob/master/rts/PrimOps.cmm#L572-L588

    stg_decodeFloatzuIntzh ( F_ arg )
    {
        W_ p, mp_tmp1, W_ mp_tmp_w;
    
        STK_CHK_GEN_N (WDS(2));
    
        mp_tmp1  = Sp - WDS(1);
        mp_tmp_w = Sp - WDS(2);
    
        ccall __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg);
    
        return (W_[mp_tmp1], W_[mp_tmp_w]);
    }
    
that function in particular is compiled to

    [stg_decodeFloatzuIntzh() //  [F1]
             { info_tbl: []
               stack_info: arg_space: 8 updfr_space: Just 8
             }
         {offset
           cc: _c0::F32 = F1;
               goto c4;
           c4: if ((old + 0) - 2 * 8 < SpLim) goto c6; else goto c7;
           c6: I64[(young<c5> + 8)] = c5;
               call stg_gc_noregs() returns to c5, args: 8, res: 8, upd: 8;
           c5: goto c4;
           c7: _c2::I64 = Sp - 1 * 8;
               _c3::I64 = Sp - 2 * 8;
               _c8::I64 = __decodeFloat_Int;
               _c9::I64 = _c2::I64;
               _ca::I64 = _c3::I64;
               _cb::F32 = _c0::F32;
               call "ccall" arg hints:  [PtrHint, PtrHint,]  result hints:  []
                    (_c8::I64)(_c9::I64, _ca::I64, _cb::F32);
               R2 = I64[_c3::I64];
               R1 = I64[_c2::I64];
               call (P64[(old + 8)])(R2, R1) args: 8, res: 0, upd: 8;
         }
     }]
    
But I see no effort to adjust Sp (i.e. `Sp = Sp - 16`) right before the
call to __decodeFloat_Int; how is it ensured that __decodeFloat_Int
doesn't use the locations Sp-8 and Sp-16 for as its local stack?


> otoh, the comments in cmmparse.y indicate that's not quite "kosher"? or
> maybe the comments are a lie?
> https://github.com/ghc/ghc/blob/master/compiler/cmm/CmmParse.y#L24-L28


More information about the ghc-devs mailing list