[Issue 4443] Optimizer produces wrong code for || or && with struct arrays

[d-bugmail at puremagic.com]

http://d.puremagic.com/issues/show_bug.cgi?id=4443


Don <clugdbug at yahoo.com.au> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch


--- Comment #3 from Don <clugdbug at yahoo.com.au> 2010-08-03 21:56:32 PDT ---
This is a really subtle code generation bug. The buggy bit of code is below.
In this section of code, the expression size is one register. 
In the bit of code at L13, it checks to see if the addition factor is already
in a register.
In this case, it is, BUT it's actually in a double register: the {ptr, length}
pair is a ulong, and here it's been cast to uint to get the length.
So isregvar() returns BOTH registers in regm.
Then, at the end of this section of code, the register to use is selected with
a call to findreg(). But it just returns "the first register" which is correct
only if there is only one register.

In this patch, I just use the LSW returned from isregvar. It would also be
possible to change isregvar() to only return the LSW in the case where a cast
to smaller type has occured, but I don't know what other code expects from
isregvar. 

Not sure if the tysize() check is necessary. Maybe it could just be retregs = 1
<< reg1;

PATCH: cod2.c, cdorth(), line 362. (Does 1<<reg1 work if reg1 is 0?).


            L13:
                regm_t regm;
                if (e11->Eoper == OPvar && isregvar(e11,&regm,&reg1))
                {
+                    if (tysize[tybasic(e11->Ety)]<= REGSIZE)
+                       retregs = 1 << reg1; // Only want the LSW
+                    else
                    retregs = regm;
                    c1 = NULL;
                    freenode(e11);
                }
                else
                    c1 = codelem(e11,&retregs,FALSE);
            }
            rretregs = ALLREGS & ~retregs;
            c2 = scodelem(ebase,&rretregs,retregs,TRUE);
            {
                regm_t sregs = *pretregs & ~rretregs;
                if (!sregs)
                    sregs = ALLREGS & ~rretregs;
                c3 = allocreg(&sregs,&reg,ty);
            }
+ // BUG: We should ensure there is only register in retregs
            reg1 = findreg(retregs);


Also noticed that in cod4.c, cdmsw() line 2838, there's an incorrect comment.
    retregs &= mMSW;                    // want LSW only
This should of course be // want MSW only

-- 
Configure issuemail: http://d.puremagic.com/issues/userprefs.cgi?tab=email
------- You are receiving this mail because: -------