f/target.c HOST_BYTES_BIG_ENDIAN

f/target.c HOST_BYTES_BIG_ENDIAN

[Alan Modra]

HOST_BYTES_BIG_ENDIAN and HOST_BITS_BIG_ENDIAN are referenced in
f/target.c when cross-compiling, but are not defined anywhere.
This makes g77 spit the dummy when cross-compiling from a
big-endian host to a big-endian target.

Since arbitrarily assuming host bytes and bits are little endian is
arguably worse than assuming host and target match, may I suggest
the following patch?

gcc/f/ChangeLog
	* target.c (ffetarget_memcpy_): Don't test endian macros when
	NATIVE_CROSS.  Default host endian values to target values.

Index: gcc/f/target.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/f/target.c,v
retrieving revision 1.19
diff -u -p -r1.19 target.c
--- gcc/f/target.c	17 Jun 2002 21:21:01 -0000	1.19
+++ gcc/f/target.c	22 Aug 2002 12:04:18 -0000
@@ -2520,10 +2520,10 @@ ffetarget_verify_character1 (mallocPool 
 void *
 ffetarget_memcpy_ (void *dst, void *src, size_t len)
 {
-#ifdef CROSS_COMPILE
+#if defined (CROSS_COMPILE) && !defined (NATIVE_CROSS)
   int host_words_big_endian =
 #ifndef HOST_WORDS_BIG_ENDIAN
-    0
+    WORDS_BIG_ENDIAN
 #else
     HOST_WORDS_BIG_ENDIAN
 #endif
@@ -2531,7 +2531,7 @@ ffetarget_memcpy_ (void *dst, void *src,
 
   int host_bytes_big_endian =
 #ifndef HOST_BYTES_BIG_ENDIAN
-    0
+    BYTES_BIG_ENDIAN
 #else
     HOST_BYTES_BIG_ENDIAN
 #endif
@@ -2539,7 +2539,7 @@ ffetarget_memcpy_ (void *dst, void *src,
 
   int host_bits_big_endian =
 #ifndef HOST_BITS_BIG_ENDIAN
-    0
+    BITS_BIG_ENDIAN
 #else
     HOST_BITS_BIG_ENDIAN
 #endif

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Hans-Peter Nilsson]

On Thu, 22 Aug 2002, Alan Modra wrote:

> HOST_BYTES_BIG_ENDIAN and HOST_BITS_BIG_ENDIAN are referenced in
> f/target.c when cross-compiling, but are not defined anywhere.

That sounds strange (for a BE host).  f/target.c does include
config.h which includes auto-host.h...  There doesn't seem to be
anything magic about NATIVE_CROSS.  What am I missing?

brgds, H-P

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Alan Modra]

On Thu, Aug 22, 2002 at 09:33:58AM -0400, Hans-Peter Nilsson wrote:
> On Thu, 22 Aug 2002, Alan Modra wrote:
> 
> > HOST_BYTES_BIG_ENDIAN and HOST_BITS_BIG_ENDIAN are referenced in
> > f/target.c when cross-compiling, but are not defined anywhere.
> 
> That sounds strange (for a BE host).  f/target.c does include
> config.h which includes auto-host.h...

HOST_WORDS_BIG_ENDIAN is the only HOST_*_BIG_ENDIAN macro defined.
grep the source.

>  There doesn't seem to be
> anything magic about NATIVE_CROSS.  What am I missing?

Belts and braces.  Strictly speaking, it's not necessary with the
other changes.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Hans-Peter Nilsson]

On Fri, 23 Aug 2002, Alan Modra wrote:

> On Thu, Aug 22, 2002 at 09:33:58AM -0400, Hans-Peter Nilsson wrote:
> > On Thu, 22 Aug 2002, Alan Modra wrote:
> >
> > > HOST_BYTES_BIG_ENDIAN and HOST_BITS_BIG_ENDIAN are referenced in
> > > f/target.c when cross-compiling, but are not defined anywhere.
> >
> > That sounds strange (for a BE host).  f/target.c does include
> > config.h which includes auto-host.h...
>
> HOST_WORDS_BIG_ENDIAN is the only HOST_*_BIG_ENDIAN macro defined.
> grep the source.

Bummer.  I added tests for macros that are not there.  So only a
subset of the host <=> target mismatch situation is detectable:
HOST_WORDS_BIG_ENDIAN <=> BYTES_BIG_ENDIAN.  The semantics for
the host vs. target macros are a little different;
HOST_WORDS_BIG_ENDIAN implies the hostness *BYTES*_BIG_ENDIAN
but also the host *WORDS*_BIG_ENDIAN.  And apparently nothing
else has a need to test these host endian details so it doesn't
seem worthwhile to actually autoconf for it.  Ugh.

Your patch is wrong; don't use the *target* BITS_BIG_ENDIAN (et
al) when the HOST_ one isn't defined.  I suggest to remove that
test; check instead that HOST_WORDS_BIG_ENDIAN matches both
BYTES_BIG_ENDIAN and WORDS_BIG_ENDIAN.  I see there's also a
HOST_FLOAT_WORDS_BIG_ENDIAN FWIW.

(Heh, it seems few people use big-endian hosts for
cross-compiling fortran.)

> >  There doesn't seem to be
> > anything magic about NATIVE_CROSS.  What am I missing?
>
> Belts and braces.  Strictly speaking, it's not necessary with the
> other changes.

Well it's confusing.  I think there should be no NATIVE_CROSS
test.  (Actually it's the wrong direction of the belts and
braces; belts and braces would have been to remove the #ifdef
CROSS_COMPILE and *always* test for endian mismatch. ;-)

brgds, H-P

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Alan Modra]

On Thu, Aug 22, 2002 at 10:32:07PM -0400, Hans-Peter Nilsson wrote:
> Your patch is wrong; don't use the *target* BITS_BIG_ENDIAN (et
> al) when the HOST_ one isn't defined.  I suggest to remove that
> test; check instead that HOST_WORDS_BIG_ENDIAN matches both
> BYTES_BIG_ENDIAN and WORDS_BIG_ENDIAN.

Please, can you fix your code then?

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Hans-Peter Nilsson]

(gcc-patches@ added for archival purposes.)

On Fri, 23 Aug 2002, Alan Modra wrote:

> On Thu, Aug 22, 2002 at 10:32:07PM -0400, Hans-Peter Nilsson wrote:
> > Your patch is wrong; don't use the *target* BITS_BIG_ENDIAN (et
> > al) when the HOST_ one isn't defined.  I suggest to remove that
> > test; check instead that HOST_WORDS_BIG_ENDIAN matches both
> > BYTES_BIG_ENDIAN and WORDS_BIG_ENDIAN.
>
> Please, can you fix your code then?

Fair enough; here.  I've just sanity tested that this compiles
on a i686-pc-linux-gnu, with "make" on top of a previously
compiled tree.  Could you please test if it works for you?

	* target.c (ffetarget_memcpy_): Don't test nonexistent
	HOST_BYTES_BIG_ENDIAN, HOST_BITS_BIG_ENDIAN.  Check
	HOST_WORDS_BIG_ENDIAN against both WORDS_BIG_ENDIAN and
	BYTES_BIG_ENDIAN.  Check HOST_FLOAT_WORDS_BIG_ENDIAN against
	FLOAT_WORDS_BIG_ENDIAN.

Index: target.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/f/target.c,v
retrieving revision 1.19
diff -p -c -r1.19 target.c
*** target.c	17 Jun 2002 21:21:01 -0000	1.19
--- target.c	23 Aug 2002 20:30:19 -0000
*************** void *
*** 2521,2526 ****
--- 2521,2530 ----
  ffetarget_memcpy_ (void *dst, void *src, size_t len)
  {
  #ifdef CROSS_COMPILE
+
+   /* HOST_WORDS_BIG_ENDIAN corresponds to both WORDS_BIG_ENDIAN and
+      BYTES_BIG_ENDIAN (i.e. a difference in the two latter is not
+      representable on the host).  */
    int host_words_big_endian =
  #ifndef HOST_WORDS_BIG_ENDIAN
      0
*************** ffetarget_memcpy_ (void *dst, void *src,
*** 2529,2547 ****
  #endif
      ;

!   int host_bytes_big_endian =
! #ifndef HOST_BYTES_BIG_ENDIAN
!     0
! #else
!     HOST_BYTES_BIG_ENDIAN
! #endif
!     ;
!
!   int host_bits_big_endian =
! #ifndef HOST_BITS_BIG_ENDIAN
      0
  #else
!     HOST_BITS_BIG_ENDIAN
  #endif
      ;

--- 2533,2543 ----
  #endif
      ;

!   int host_float_words_big_endian =
! #ifndef HOST_FLOAT_WORDS_BIG_ENDIAN
      0
  #else
!     HOST_FLOAT_WORDS_BIG_ENDIAN
  #endif
      ;

*************** ffetarget_memcpy_ (void *dst, void *src,
*** 2555,2562 ****
       for instance in g77.f-torture/execute/980628-[4-6].f and alpha2.f.
       Still, we compile *some* code.  FIXME: Rewrite handling of numbers.  */
    if (!WORDS_BIG_ENDIAN != !host_words_big_endian
!       || !BYTES_BIG_ENDIAN != !host_bytes_big_endian
!       || !BITS_BIG_ENDIAN != !host_bits_big_endian)
      sorry ("data initializer on host with different endianness");

  #endif /* CROSS_COMPILE */
--- 2551,2558 ----
       for instance in g77.f-torture/execute/980628-[4-6].f and alpha2.f.
       Still, we compile *some* code.  FIXME: Rewrite handling of numbers.  */
    if (!WORDS_BIG_ENDIAN != !host_words_big_endian
!       || !BYTES_BIG_ENDIAN != !host_words_big_endian
!       || !FLOAT_WORDS_BIG_ENDIAN != !host_float_words_big_endian)
      sorry ("data initializer on host with different endianness");

  #endif /* CROSS_COMPILE */

brgds, H-P

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Alan Modra]

On Fri, Aug 23, 2002 at 04:42:46PM -0400, Hans-Peter Nilsson wrote:
> Fair enough; here.  I've just sanity tested that this compiles
> on a i686-pc-linux-gnu, with "make" on top of a previously
> compiled tree.  Could you please test if it works for you?

It does.  Thanks.

> 	* target.c (ffetarget_memcpy_): Don't test nonexistent
> 	HOST_BYTES_BIG_ENDIAN, HOST_BITS_BIG_ENDIAN.  Check
> 	HOST_WORDS_BIG_ENDIAN against both WORDS_BIG_ENDIAN and
> 	BYTES_BIG_ENDIAN.  Check HOST_FLOAT_WORDS_BIG_ENDIAN against
> 	FLOAT_WORDS_BIG_ENDIAN.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: f/target.c HOST_BYTES_BIG_ENDIAN

[Toon Moene]

Hans-Peter Nilsson wrote: