Is there a way to tell the compiler to defer updating RAM loop variables?

[rlw]

I am using Code Warrior version 5.9.0, build 2836 for S12XE.

In my listing, below, I see the gnerated code needlessly reloading and resaving registers in the do loop that starts at source line 589.

I tried the -Or and -Ol options. -Or did not make a difference, but -Onr made things worse, so I assume -Or is a default.

-Ol3 helped, but the effect was equivalent to -Ol1. Only one of the 3 loop variables was "optimized".

Neither the presence nor absence of the "Register" qualifier made a difference. Nor did the location of the definitions of d and s.

Other observations: In source line 588, the compiler "remembers" that B was loaded with the needed value in line 581, but in line 590, it seems to have "forgotten".

Also, I noticed that the peephole optimizer removed the extranious "CPD #0" from source line 591, but not the extranious "LDD 0,SP"

(Just some observations.)

  578:  void DiagReadMemoryByAddress( uint32 address, uint16 len)  579:  {Function: DiagReadMemoryByAddressSource  : apdi.cOptions : -CPUHCS12XE -F2 -Lasm=%n.lst -LmCfg=ilmu -Mb -Ol3 -Onbf -OnPMNC -WmsgNu=abcet  0000 6ca9         [2]     STD   5,-SP  580:     /* Block reading code area */  581:     if ((uint8)(address >> 16) > 0x40)  0002 e689         [3]     LDAB  9,SP  0004 c140         [1]     CMPB  #64  0006 2307         [3/1]   BLS   *+9 ;abs = 000f  582:     {  583:        DiagNRCInvalidKey(); // Closest error code to security violation  0008 c635         [1]     LDAB  #53  000a 7b0000       [3]     STAB  diagErrorCode  584:     }  000d 201f         [3]     BRA   *+33 ;abs = 002e  585:     else  586:     {  587:        register uint8 * d = DiagBuffer + LengthServiceID + LengthAddressParameter;  000f fe0000       [3]     LDX   DiagBuffer  0012 1a05         [2]     LEAX  5,X  588:        register uint8 * __far s = (uint8 * __far)address;  0014 ed8a         [3]     LDY   10,SP  0016 6d83         [2]     STY   3,SP  0018 6b82         [2]     STAB  2,SP  589:        do {  590:           *d++ = *s++;  001a ed83         [3]     LDY   3,SP  001c e682         [3]     LDAB  2,SP  001e 5b10         [2]     STAB  /*GPAGE*/16  0020 18a670       [4]     GLDAA 1,Y+  0023 6a30         [2]     STAA  1,X+  0025 6d83         [2]     STY   3,SP  591:        } while (--len);  0027 186380       [4]     DECW  0,SP  002a ec80         [3]     LDD   0,SP  002c 26ec         [3/1]   BNE   *-18 ;abs = 001a  592:     }  593:     DiagDataLength = len;  002e ecb4         [3]     LDD   5,SP+  0030 7c0000       [3]     STD   DiagDataLength  594:  }  0033 0a           [7]     RTC  

[kef]

I don't know how to force loop variables to registers, it seems you know all required switches. But using

#pragma PAGE_UPDATE OFF // see Release_Notes_HC12.chm in Help folder

you could remove

  001c e682         [3]     LDAB  2,SP
  001e 5b10         [2]     STAB  /*GPAGE*/16

from the loop. Of course GPAGE has to be setup before the loop.

[rlw]

Kef, thanks. Handy to know how to get at least the GPAGE update out of  the loop.

That did give me the thought that the one pointer being a far pointer might have inhibited fully optimizing the loop variables, though I'd still expect the other pointer to have been optimized. Not sure if I can get anywhere with this thought, but I will reread the sections about the various page registers. tomorrow.

[kef]

Maybe make service request on this here . If there's no answer how to optimize loop variables in you case, then maybe Freescale will accept your request to improve compiler regarding subject.