Failed cross: Alpha->ARM

Failed cross: Alpha->ARM

[Dave Gilbert]

Hi,
  Host: Alpha Linux RedHat 5.0
  Target: ARM Linux
  Problem: Constant greater than 2^32 generated in assembler file
  Suspected cause: Converting a -1 to 2^32 and then propogating the value.
----------------------------------------------------
Example source file:
typedef struct {
  int a[10];
} foo;

foo* x;
static foo y[128];

int a() {
  do {
  } while (x < y + 128-1);
};

---------------------------------------------------------------
Duff output:

rfp	.req	r9
sl	.req	r10
fp	.req	r11
ip	.req	r12
sp	.req	r13
lr	.req	r14
pc	.req	r15
@ GNU C version egcs-2.91.02 971206 (gcc-2.8.0) (arm-unknown-linuxaout) compiled by GNU C version 2.7.2.3.
@ options passed: 
@ options enabled:  -fpeephole -ffunction-cse -fkeep-static-consts
@ -freg-struct-return -fsjlj-exceptions -fcommon -fverbose-asm -fgnu-linker
@ -fargument-alias

gcc2_compiled.:
___gnu_compiled_c:
.text
	.align	0
	.global	_a
_a:
	@ args = 0, pretend = 0, frame = 0
	@ frame_needed = 1, current_function_anonymous_args = 0
	mov	ip, sp
	stmfd	sp!, {fp, ip, lr, pc}
	sub	fp, ip, #4
	mov	r0, r0	@ nop
L2:
L4:
	ldr	r3, L6
	ldr	r2, [r3, #0]
	ldr	r3, L6+4
	cmp	r2, r3
	bcc	L5
	b	L3
L5:
	b	L4
L7:
	.align	0
L6:
	.word	_x
	.word	_y+4294972376      < Bzzt!!!!!!!!!!
L3:
L1:
	ldmea	fp, {fp, sp, pc}^
	.comm	_x, 4	@ 4
.bss
	.align	0
_y:
	.space	5120


Dave
*  NOTE: I have had mail problems over the last few days - things may  *
*  have bounced                                                        *
---------------------------------------------------- Man can not live  -
 David Alan Gilbert - gro.gilbert @ treblig.org ---- by bread alone. He 
---------------------------------------------------- needs chocolate.  -

Re: Failed cross: Alpha->ARM

[Nick Clifton]

: Date: Mon, 15 Dec 1997 21:38:19 +0000 (GMT)
: From: Dave Gilbert <gilbertd@treblig.org>
: 
: Hi,
:   Host: Alpha Linux RedHat 5.0
:   Target: ARM Linux
:   Problem: Constant greater than 2^32 generated in assembler file
:   Suspected cause: Converting a -1 to 2^32 and then propogating the value.
: ----------------------------------------------------
: Duff output:
:
: @ GNU C version egcs-2.91.02 971206 (gcc-2.8.0) (arm-unknown-linuxaout) compiled by GNU C version 2.7.2.3.

[snip]

: L6:
: 	.word	_x
: 	.word	_y+4294972376      < Bzzt!!!!!!!!!!
: L3:
: L1:

Hmm, I get the right output from the coff toolchain:

   [snip]

  .L6:
	.word	_x
	.word	_y+5080
  .L3:
  .L1:

This is from:

  @ GNU C version cygnus-2.91.00 971101 (gcc2-971021 experimental) (arm-acorn-coff) compiled by GNU C version 2.7-97r1.

so maybe it is a problem with the linux part of the egcs backend.  I
will try creating such a toolchain and see if I can reproduce the
problem.

Cheers
	Nick

Re: Failed cross: Alpha->ARM

[Stephen Williams]

  Host: Alpha Linux RedHat 5.0
  Target: ARM Linux
  Problem: Constant greater than 2^32 generated in assembler file
  Suspected cause: Converting a -1 to 2^32 and then propogating the value.

Try this patch. (In spite of the name, it does apply to the egcs-1.0
sources.) As I understand it, it does indeed mess up detecting if the
constant that is can load in a simple instruction or requires more
complex handling. Here is a ChangeLog and the patch itself.

Mon Oct 20 08:49:16 1997  Stephen Williams  <steve@icarus.com>

       * gcc/config/arm/arm.c: Fixed problem with 64bit hosts generating
       bogus assembler for long long constants.


diff --exclude=*~ --exclude=*.orig --exclude=testsuite -Npcr 
egcs-971023/gcc/config/arm/arm.c ../egcs-971023/gcc/config/arm/arm.c
*** egcs-971023/gcc/config/arm/arm.c    Mon Aug 11 08:57:24 1997
--- ../egcs-971023/gcc/config/arm/arm.c Mon Oct 27 20:38:41 1997
*************** const_ok_for_arm (i)
*** 407,412 ****
--- 407,415 ----
  {
    unsigned HOST_WIDE_INT mask = ~0xFF;

+   if (i & ~(unsigned HOST_WIDE_INT) 0xffffffff)
+     return FALSE;
+
    /* Fast return for 0 and powers of 2 */
    if ((i & (i - 1)) == 0)
      return TRUE;

-- 
Steve Williams
steve@icarus.com
steve@picturel.com

"The woods are lovely, dark and deep.  But I have promises to keep,
And lines to code before I sleep, And lines to code before I sleep."

Re: Failed cross: Alpha->ARM

[Richard Earnshaw]

gilbertd@treblig.org said:
> Hi,
>   Host: Alpha Linux RedHat 5.0
>   Target: ARM Linux
>   Problem: Constant greater than 2^32 generated in assembler file
>   Suspected cause: Converting a -1 to 2^32 and then propogating the
>   value.
> ----------------------------------------------------
> Example source file: 
> typedef struct {
>   int a[10];
> } foo;
>
> foo* x;
> static foo y[128];
>
> int a() {
>   do {
>   } while (x < y + 128-1);
> }; 


Stephen Williams <steve@icarus.icarus.com> wrote:
>   Host: Alpha Linux RedHat 5.0
>   Target: ARM Linux
>   Problem: Constant greater than 2^32 generated in assembler file
>   Suspected cause: Converting a -1 to 2^32 and then propogating the value.
> 
> Try this patch. (In spite of the name, it does apply to the egcs-1.0
> sources.) As I understand it, it does indeed mess up detecting if the
> constant that is can load in a simple instruction or requires more
> complex handling. Here is a ChangeLog and the patch itself.
> 
> Mon Oct 20 08:49:16 1997  Stephen Williams  <steve@icarus.com>
> 
>        * gcc/config/arm/arm.c: Fixed problem with 64bit hosts generating
>        bogus assembler for long long constants.
> 
> 

First this patch doesn't fix the problem.  Second it is wrong, since it 
doesn't take into account assumptions elsewhere about use of HOST_WIDE_INT 
when that is more than 32-bits.  (The latest egcs snapshot contains the 
correct fixes for this).

The patch below fixes the problem reported.

Tue Dec 16 17:13:43 1997  Richard Earnshaw (rearnsha@arm.com)

	* explow.c (plus_constant_wide, case PLUS): Truncate any sum of
	two integers to the width of the mode.

Re: Failed cross: Alpha->ARM

[Jeffrey A Law]

  In message < 199712160500.VAA04153@icarus.icarus.com >you write:
  >   Host: Alpha Linux RedHat 5.0
  >   Target: ARM Linux
  >   Problem: Constant greater than 2^32 generated in assembler file
  >   Suspected cause: Converting a -1 to 2^32 and then propogating the value.
  > 
  > Try this patch. (In spite of the name, it does apply to the egcs-1.0
  > sources.) As I understand it, it does indeed mess up detecting if the
  > constant that is can load in a simple instruction or requires more
  > complex handling. Here is a ChangeLog and the patch itself.
  > 
  > Mon Oct 20 08:49:16 1997  Stephen Williams  <steve@icarus.com>
  > 
  >        * gcc/config/arm/arm.c: Fixed problem with 64bit hosts generating
  >        bogus assembler for long long constants.
Hmmm, egcs already has code in this function to try and handle this kind
of problem.  Why don't you pick up the latest snapshot and see why the
code to handle this problem isn't working?  That would be a big help to
us.

jeff