Egcs 1.1 produces incorrect ColdFire code

Egcs 1.1 produces incorrect ColdFire code

[Eric Devolder]

Hey all, I've configured Egcs 1.1b with Binutils 2.9.1 for m68k-coff.
Here is a simple program that illustrates the incorrect code:

typedef struct
{
    char spacing[0xAAAA]; /* value which breaks 16-bit displacement */
    char tryme;
} silly;

void
main (void)
{
    silly *sill = 0;

    sill->tryme = 0xABCD;
}

When compiling with Egcs 1.1b, I obtain the following:

/usr2/styx/comp/m68k/egcs-1.1b/binutils-2.9.1/bin/m68k-coff-gcc -m5200
-S -I../../src/include -I../../src/dbug/v2 -I. -fno-builtin -ansi -Wall
-D MCF5206_EVS src/dbug/t.c -o obj/gnu/board/t.s
src/dbug/t.c:9: warning: return type of `main' is not `int'
src/dbug/t.c: In function `main':
src/dbug/t.c:12: warning: overflow in implicit constant conversion
/usr2/styx/comp/m68k/egcs-1.1b/binutils-2.9.1/bin/m68k-coff-as  -m5200
--register-prefix-optional -ahls=obj/gnu/board/t.lst -o
obj/gnu/board/t.o obj/gnu/board/t.s
obj/gnu/board/t.s: Assembler messages:
obj/gnu/board/t.s:13: Error: displacement too large for this
architecture; needs 68020 or higher -- statement `move.b %d0,43690(%a0)'
ignored

Here is the listing file:

68K GAS  obj/gnu/board/t.s          page 1


   1                    .file   "t.c"
   2                gcc2_compiled.:
   3                __gnu_compiled_c:
   4                .text
   5 ???? 00                .even
   6                .globl main
   7                main:
   8 ???? 4E56 FFFC         link.w %a6,#-4
   9 ???? 4EB9 0000         jsr __main
   9      0000
  10 ???? 42AE FFFC         clr.l -4(%a6)
  11 ???? 206E FFFC         move.l -4(%a6),%a0
  12 ???? 70CD              moveq.l #-51,%d0
  13                    move.b %d0,43690(%a0)        <---- invalid
16-bit displacement
  14                .L1:
  15 ???? 4E5E              unlk %a6
  16                    rts
^L68K GAS  obj/gnu/board/t.s            page 2

The assembler correctly catches the invalid displacement for ColdFire.

For what its worth, this was working in Egcs 1.0.3, as evidenced by the
listing file below:

68K GAS  obj/gnu/board/t.s          page 1


   1                    .file   "t.c"
   2                gcc2_compiled.:
   3                __gnu_compiled_c:
   4                .text
   5                    .even
   6                .globl main
   7                main:
   8 0000 4E56 FFFC         link.w %a6,#-4
   9 0004 4EB9 0000         jsr __main
   9      0000
  10 000a 42AE FFFC         clr.l -4(%a6)
  11 000e 206E FFFC         move.l -4(%a6),%a0
  12 0012 203C 0000         move.l #43690,%d0
  12      AAAA
  13 0018 72CD              moveq.l #-51,%d1
  14 001a 1181 0800         move.b %d1,(%a0,%d0.l)
  15                .L1:
  16 001e 4E5E              unlk %a6
  17 0020 4E75              rts
^L68K GAS  obj/gnu/board/t.s            page 2



Thanks,
Eric

Re: Egcs 1.1 produces incorrect ColdFire code

[Toon Moene]

>  Hey all, I've configured Egcs 1.1b with Binutils 2.9.1
>  for m68k-coff.  Here is a simple program that illustrates
>  the incorrect code:
>
>  [ simple snip ]
>
>  When compiling with Egcs 1.1b, I obtain the following:
>
>  /usr2/styx/comp/m68k/egcs-1.1b/binutils-2.9.1/bin/m68k-coff-gcc
>  -m5200 -S -I../../src/include -I../../src/dbug/v2 -I.
>  -fno-builtin -ansi -Wall -D MCF5206_EVS src/dbug/t.c -o
>  obj/gnu/board/t.s
>  ...
>  obj/gnu/board/t.s: Assembler messages:
>  obj/gnu/board/t.s:13: Error: displacement too large for
>  this architecture; needs 68020 or higher -- statement
>  `move.b %d0,43690(%a0)' ignored

Unfortunately, I don't have egcs-1.1 lying around anymore, but it  
might be instructive to look into .../gcc/config/m68k/m68k.h and  
search for the "TARGET_SWITCHES"; the following is the current entry  
for the m5200 (note the turning off of MASK_68020):

    { "5200", - (MASK_68060|MASK_68040|MASK_68040_ONLY|MASK_68020       \
                |MASK_BITFIELD|MASK_68881)},                            \
    { "5200", (MASK_5200)},                                             \

This bit (TARGET_68020) is used further down in the m68k.h file to  
allow large offsets to address registers, so for the m5200 this bit  
should *not* be set, as above.

HTH,
Toon.

Re: Egcs 1.1 produces incorrect ColdFire code

[Jeffrey A Law]

  In message < 9809151847.AA23990@moene.indiv.nluug.nl >you write:
  > >  obj/gnu/board/t.s: Assembler messages:
  > >  obj/gnu/board/t.s:13: Error: displacement too large for
  > >  this architecture; needs 68020 or higher -- statement
  > >  `move.b %d0,43690(%a0)' ignored
  > 
  > Unfortunately, I don't have egcs-1.1 lying around anymore, but it  
  > might be instructive to look into .../gcc/config/m68k/m68k.h and  
  > search for the "TARGET_SWITCHES"; the following is the current entry  
  > for the m5200 (note the turning off of MASK_68020):
  > 
  >     { "5200", - (MASK_68060|MASK_68040|MASK_68040_ONLY|MASK_68020       \
  >                 |MASK_BITFIELD|MASK_68881)},                            \
  >     { "5200", (MASK_5200)},                                             \
  > 
  > This bit (TARGET_68020) is used further down in the m68k.h file to  
  > allow large offsets to address registers, so for the m5200 this bit  
  > should *not* be set, as above.
I do not think changing TARGET_SWITCHES is the right fix since that
will have other effects than just changing the allowable offsets.
It is probably better to fix GO_IF_LEGITIMATE_ADDRESS and friends to
test TARGET_5200 has needed.

jeff

Re: Egcs 1.1 produces incorrect ColdFire code

[Toon Moene]

  In message < 9809151847.AA23990@moene.indiv.nluug.nl > I wrote:

  > >  obj/gnu/board/t.s: Assembler messages:
  > >  obj/gnu/board/t.s:13: Error: displacement too large for
  > >  this architecture; needs 68020 or higher -- statement
  > >  `move.b %d0,43690(%a0)' ignored
  >
  > Unfortunately, I don't have egcs-1.1 lying around anymore, but it  
  > might be instructive to look into .../gcc/config/m68k/m68k.h and  
  > search for the "TARGET_SWITCHES"; the following is the current  
entry
  > for the m5200 (note the turning off of MASK_68020):
  >
  >     { "5200", -  
(MASK_68060|MASK_68040|MASK_68040_ONLY|MASK_68020       \
  >                 |MASK_BITFIELD|MASK_68881)},                     
        \
  >     { "5200", (MASK_5200)},                                      
        \
  >
  > This bit (TARGET_68020) is used further down in the m68k.h file to  
  > allow large offsets to address registers, so for the m5200 this bit  
  > should *not* be set, as above.

and Jeff replied:

>  I do not think changing TARGET_SWITCHES is the right fix
>  since that will have other effects than just changing
>  the allowable offsets.  It is probably better to fix
>  GO_IF_LEGITIMATE_ADDRESS and friends to test TARGET_5200
>  has needed.

According to my reading of the documentation, the above definition  
turns *off* flags TARGET_68060, TARGET_68040, TARGET_68040_ONLY,  
TARGET_68020, TARGET_BITFIELD and TARGET_68881 when specifying  
-m5200 on the command line.

QUOTE

TARGET_SWITCHES

	...

	Each subgrouping contains a string constant, that defines the
	option name, and a number, which contains the bits to set in
	`target_flags'.  A negative number says to clear these bits
	instead; the negative of the number is which bits to clear.
	The actual option name is made by appending `-m' to the
	specified name [ meant here is `prepending', I think (TM) ]

	...

UNQUOTE

So gcc -m5200 should clear TARGET_68020 (amongst others) and set  
TARGET_5200.  If this interpretation is correct, the backend  
shouldn't generate large offsets, because TARGET_68020 isn't set.

Confusingly yours,
Toon.

Re: Egcs 1.1 produces incorrect ColdFire code

[Jeffrey A Law]

  In message <35FE7D3D.E38EC066@email.sps.mot.com>you write:
  > Hey all, I've configured Egcs 1.1b with Binutils 2.9.1 for m68k-coff.
  > Here is a simple program that illustrates the incorrect code:
  > 
  > typedef struct
  > {
  >     char spacing[0xAAAA]; /* value which breaks 16-bit displacement */
  >     char tryme;
  > } silly;
  > 
  > void
  > main (void)
  > {
  >     silly *sill = 0;
  > 
  >     sill->tryme = 0xABCD;
  > }
  > 
  > When compiling with Egcs 1.1b, I obtain the following:
This has been fixed for the upcoming egcs-1.1.1 release:

main:
        link.w %a6,#-4
        jsr __main
        clr.l -4(%a6)
        move.l -4(%a6),%a0
        move.l #43690,%d0
        moveq.l #-51,%d1
        move.b %d1,(%a0,%d0.l)
.L1:
        unlk %a6
        rts