Re: Problem with unsigned long long on ultrasparc

Re: Problem with unsigned long long on ultrasparc

[Brad Lucier]

> From davem@dm.cobaltmicro.com  Thu Aug 20 18:55:42 1998
> Date: Thu, 20 Aug 1998 16:53:46 -0700
> 
>    From: Brad Lucier <lucier@math.purdue.edu>
>    Date: Thu, 20 Aug 1998 18:51:17 -0500 (EST)
> 
>    It is important to me to use 32 double-precision floating-point registers
>    for some of my codes; does -mtume=ultrasparc provide this?
> 
> It should.
>

Well, actually, it doesn't---it only uses the first 16 double precision
registers.

How much 64bit support can egcs provide in Solaris 2.5.1 and 2.6?

Brad Lucier   lucier@math.purdue.edu

Re: Problem with unsigned long long on ultrasparc

[Jeffrey A Law]

  In message <199808202122.QAA22453@polya.math.purdue.edu>you write:
  > Here's a simple program:
  > 
  > #include <stdlib.h>
  > #include <stdio.h>
  > 
  > int main() {
  >   unsigned long long a[10];
  > 	
  > #define BASE_DIV_2 (((unsigned long long) 1) << 63)
  > 
  >  a[0] = -1;
  >  printf("%ull\n", a[0]);
  >  printf("%ull\n", BASE_DIV_2);
  >  if (a[0] < BASE_DIV_2)
  >    printf("Problem\n");
  >  else
  >    printf("OK\n");
  > }
  > 
  > With this version of gcc:
  > 
  > peano-30% gcc -v
  > Reading specs from /opt/egcs-19980816/lib/gcc-lib/sparc-sun-solaris2.5.1/eg
  > cs-2.91.54/specs
  > gcc version egcs-2.91.54 19980816 (gcc2 ss-980609 experimental)
  > 
  > I get:
  > 
  > peano-31% gcc -O1 -o testlong testlong.c
  > peano-32% testlong
  > 4294967295ll
  > 2147483648ll
  > OK
  > 
  > but
  > 
  > peano-27% gcc -O1 -mcpu=ultrasparc -o testlong testlong.c
  > peano-28% testlong
  > 4294967295ll
  > 2147483648ll
  > Problem
  > 
  > This is on Solaris 2.5.1.
I don't know if anyone ever responded to this bug report.

This problem has been corrected in the mainline egcs sources, but is not
expected to be fixed for the upcoming egcs-1.1.2 minor release.

jeff

Re: Problem with unsigned long long on ultrasparc

[Todd Vierling]

On Thu, 20 Aug 1998, Brad Lucier wrote:

: > Two things, -1 is an int not a long long (although does GCC promote it
: > to a long long in this case?),
: 
: Yes, it does seem to promote it; I printed out the two values
: as unsigned long longs in order to check this.

IIRC, your test case used printf("%ull", ...) which is NOT printing "long
long".  It's printing "unsigned int" followed by two "l" characters.

If you want to print "unsigned long long", use "%llu".  For "long long",
"%lld".  "%qu" and "%qd" may also work, but I do not have Slowlaris handy to
test.  :)

-- 
-- Todd Vierling (Personal tv@pobox.com; Bus. todd_vierling@xn.xerox.com)

Re: Problem with unsigned long long on ultrasparc

[Brad Lucier]

OK, I have to admit, free software is great.  Bugs are fixed before
they're even reported!  As you suggested, the the current mainline branch
doesn't exhibit the bug I mentioned earlier.

But the 1.1 branch still has the bug.  Can this be fixed?  The mainline
is not stable enough for use.  For example, there was another problem
installing the mainline branch:

  cd ../../../libio; \
  for FILE in *.h; do if [ $FILE != _G_config.h ]; then \
    rm -f /opt/egcs-mainline/include/g++/$FILE ; \
    /bin/sh /home/c/lucier/programs/egcs/install-sh -c -m 644 $FILE /opt/egcs-mainline/include/g++/$FILE ; \
    chmod a-x /opt/egcs-mainline/include/g++/$FILE ; \
  else true; fi; done ; \
else true; \
fi
install:  _G_config.h does not exist
make[1]: *** [install] Error 1
make[1]: Leaving directory `/home/c/lucier/programs/egcs/objdir/sparc-sun-solaris2.5.1/libio'
make: *** [install-target-libio] Error 2

Then, when I tried to compile a longer file, I got:

peano-51% gcc -fPIC -mcpu=ultrasparc -O1 -c -D___DYNAMIC -D___SINGLE_HOST -o routines.c routines.c
../../gcc/emit-rtl.c:509: Internal compiler error in function gen_reg_rtx
peano-52% gcc -v
Reading specs from /opt/egcs-mainline/lib/gcc-lib/sparc-sun-solaris2.5.1/egcs-2.92.02/specs
gcc version egcs-2.92.02 19980820 (gcc2 ss-980609 experimental)

Brad Lucier   lucier@math.purdue.edu

Re: Problem with unsigned long long on ultrasparc

[David S. Miller]

   From: Brad Lucier <lucier@math.purdue.edu>
   Date: Fri, 21 Aug 1998 03:31:52 -0500 (EST)

   Then, when I tried to compile a longer file, I got:

   peano-51% gcc -fPIC -mcpu=ultrasparc -O1 -c -D___DYNAMIC -D___SINGLE_HOST -o routines.c routines.c
   ../../gcc/emit-rtl.c:509: Internal compiler error in function gen_reg_rtx
   peano-52% gcc -v
   Reading specs from /opt/egcs-mainline/lib/gcc-lib/sparc-sun-solaris2.5.1/egcs-2.92.02/specs
   gcc version egcs-2.92.02 19980820 (gcc2 ss-980609 experimental)

Send me the pre-processed source and I will fix the bug.

For the mainline issue try "-mno-v8plus"

Later,
David S. Miller
davem@dm.cobaltmicro.com

Re: Problem with unsigned long long on ultrasparc

[Brad Lucier]

I used cvs to download the latest sources on the 1.1 branch and had the same
problem.  I also downloaded the mainline branch, but part of it wouldn't link.
The error message is listed below.

I tried simpler programs but could not trigger the error.

Brad Lucier   lucier@math.purdue.edu

...
stage1/xgcc -Bstage1/  -DIN_GCC  -DSVR4  -W -Wall -O2 -g -O2   -o ../cc1plus \
      call.o decl.o errfn.o expr.o pt.o sig.o typeck2.o class.o decl2.o error.o lex.o parse.o ptree.o rtti.o spew.o typeck.o cvt.o except.o friend.o init.o method.o search.o semantics.o tree.o xref.o repo.o  `cat ../stamp-objlist` ../c-common.o ../c-pragma.o `if [ xobstack.o != x ]; then echo ../obstack.o; else true; fi` `case "stage1/xgcc -Bstage1/" in "cc") if [ xalloca.o != x ]; then echo ../alloca.o; else true; fi ;; esac` `if [ x != x ]; then echo ../; else true; fi` 
Undefined                       first referenced
 symbol                             in file
walk_tree                           tree.o
ld: fatal: Symbol referencing errors. No output written to ../cc1plus
collect2: ld returned 1 exit status
make[3]: *** [../cc1plus] Error 1
make[3]: Leaving directory `/home/c/lucier/programs/egcs/objdir/gcc/cp'
make[2]: *** [cc1plus] Error 2
make[2]: Leaving directory `/home/c/lucier/programs/egcs/objdir/gcc'
make[1]: *** [bootstrap] Error 2
make[1]: Leaving directory `/home/c/lucier/programs/egcs/objdir/gcc'
make: *** [bootstrap] Error 2
bessel-130% 

Re: Problem with unsigned long long on ultrasparc

[David S. Miller]

   From: Brad Lucier <lucier@math.purdue.edu>
   Date: Fri, 21 Aug 1998 01:51:28 -0500 (EST)

   Undefined                       first referenced
    symbol                             in file
   walk_tree                           tree.o
   ld: fatal: Symbol referencing errors. No output written to ../cc1plus

Jason Merril just commited a fix to this.

Later,
David S. Miller
davem@dm.cobaltmicro.com

Re: Problem with unsigned long long on ultrasparc

[David S. Miller]

   Date: Thu, 20 Aug 1998 16:22:03 -0500 (EST)
   From: Brad Lucier <lucier@math.purdue.edu>

   With this version of gcc:

   peano-30% gcc -v
   Reading specs from /opt/egcs-19980816/lib/gcc-lib/sparc-sun-solaris2.5.1/egcs-2.91.54/specs
   gcc version egcs-2.91.54 19980816 (gcc2 ss-980609 experimental)

...

   peano-27% gcc -O1 -mcpu=ultrasparc -o testlong testlong.c
   peano-28% testlong
   4294967295ll
   2147483648ll
   Problem

   This is on Solaris 2.5.1.

Please check the latest in CVS as I believe a change I made on the
17th cures this problem, I cannot make the compiler output the code
you are seeing.

Later,
David S. Miller
davem@dm.cobaltmicro.com

Re: Problem with unsigned long long on ultrasparc

[Brad Lucier]

> Two things, -1 is an int not a long long (although does GCC promote it
> to a long long in this case?),

Yes, it does seem to promote it; I printed out the two values
as unsigned long longs in order to check this.

> secondarily setting cpu type to
> ultrasparc is an error on Solaris-2.5, perhaps you meant
> "-mtune=ultrasparc".  "-mcpu=ultrasparc" turns on the V9 instructions
> in the compiler which won't work correctly in all cases on Solaris-2.5
> since it is not providing a 64-bit userland environment for the cpu.

It is important to me to use 32 double-precision floating-point registers
for some of my codes; does -mtume=ultrasparc provide this?

> 
> "-mtune=ultrasparc" on the other hand turns on scheduling and cost
> parameters independent of the instruction ISA being used.
> 
> Later,
> David S. Miller
> davem@dm.cobaltmicro.com
> 

The bottom line for me is that egcs-1.0.3a provides this functionality
and the current egcs-1.1 does not.  I consider this a new problem, i.e.,
a bug, not a feature, or at least a change.  I'd like somebody to
look at the assembly code that's generated to see if it's correct.
If it is, I'll crawl back into my hole and wait until Sun comes out
with a version of Solaris that's truly 64-bit..

For informational purposes, I'm including the assembly language output
for this routine from egcs-1.0.3a and the 19980816 snapshot of egcs.
I forgot to mention in my first message that I built egcs with
--haifa-enabled in both cases.

Brad Lucier   lucier@math.purdue.edu

	.file	"testlong.c"
! GNU C version egcs-2.90.29 980515 (egcs-1.0.3 release) (sparc-sun-solaris2.5.1) compiled by GNU C version egcs-2.90.29 980515 (egcs-1.0.3 release).
! options passed:  -mcpu=ultrasparc -O1
! options enabled:  -fdefer-pop -fomit-frame-pointer -fthread-jumps
! -fpeephole -ffunction-cse -finline -fkeep-static-consts
! -fpcc-struct-return -fdelayed-branch -fsched-interblock -fsched-spec
! -fcommon -fverbose-asm -fgnu-linker -falias-check -fargument-alias
! -mepilogue -mapp-regs -mcpu=ultrasparc

gcc2_compiled.:
.section	".rodata"
	.align 8
.LLC0:
	.asciz	"%ull\n"
	.align 8
.LLC1:
	.asciz	"Problem\n"
	.align 8
.LLC2:
	.asciz	"OK\n"
.section	".text"
	.align 32
	.global main
	.type	 main,#function
	.proc	04
main:
	!#PROLOGUE# 0
	save %sp,-192,%sp
	!#PROLOGUE# 1
	mov -1,%o0
	mov -1,%o1
	std %o0,[%fp-96]
	sethi %hi(.LLC0),%l0
	mov -1,%o1
	mov -1,%o2
	call printf,0
	or %l0,%lo(.LLC0),%o0
	or %l0,%lo(.LLC0),%o0
	sethi %hi(0),%o2
	sethi %hi(-2147483648),%o1
	call printf,0
	or %o2,%lo(0),%o2
	ld [%fp-96],%o1
	sethi %hi(2147483647),%o0
	or %o0,%lo(2147483647),%o0
	cmp %o1,%o0
	bgu %icc,.LL2
	sethi %hi(.LLC1),%o0
	b .LL5
	or %o0,%lo(.LLC1),%o0
.LL2:
	sethi %hi(.LLC2),%o0
	or %o0,%lo(.LLC2),%o0
.LL5:
	call printf,0
	nop
	ret
	restore
.LLfe1:
	.size	 main,.LLfe1-main
	.ident	"GCC: (GNU) egcs-2.90.29 980515 (egcs-1.0.3 release)"

with the 19980816 egcs snapshot:

	.file	"testlong.c"
gcc2_compiled.:
.section	".rodata"
	.align 8
.LLC0:
	.asciz	"%ull\n"
	.align 8
.LLC1:
	.asciz	"Problem\n"
	.align 8
.LLC2:
	.asciz	"OK\n"
.section	".text"
	.align 32
	.global main
	.type	 main,#function
	.proc	04
main:
	!#PROLOGUE# 0
	save %sp,-192,%sp
	!#PROLOGUE# 1
	mov -1,%o0
	mov -1,%o1
	st %o0,[%fp-96]
	st %o1,[%fp-92]
	sethi %hi(.LLC0),%l0
	mov -1,%o1
	mov -1,%o2
	call printf,0
	or %l0,%lo(.LLC0),%o0
	sethi %hi(-2147483648),%o1
	mov 0,%o2
	call printf,0
	or %l0,%lo(.LLC0),%o0
	ld [%fp-96],%o2
	ld [%fp-92],%o3
	sethi %hi(2147483647),%o0
	or %o0,%lo(2147483647),%o0
	mov -1,%o1
	sllx %o2,32,%o4
	sllx %o0,32,%o5
	or %o3,%o4,%o4
	or %o1,%o5,%o5
	cmp %o4,%o5
	bgu %xcc,.LL2
	sethi %hi(.LLC1),%o0
	b .LL4
	or %o0,%lo(.LLC1),%o0
.LL2:
	sethi %hi(.LLC2),%o0
	or %o0,%lo(.LLC2),%o0
.LL4:
	call printf,0
	nop
	return %i7+8
	nop
.LLfe1:
	.size	 main,.LLfe1-main
	.ident	"GCC: (GNU) egcs-2.91.54 19980816 (gcc2 ss-980609 experimental)"

Re: Problem with unsigned long long on ultrasparc

[David S. Miller]

   From: Brad Lucier <lucier@math.purdue.edu>
   Date: Thu, 20 Aug 1998 18:51:17 -0500 (EST)

   It is important to me to use 32 double-precision floating-point registers
   for some of my codes; does -mtume=ultrasparc provide this?

It should.

   The bottom line for me is that egcs-1.0.3a provides this functionality
   and the current egcs-1.1 does not.

Ok it does look like a bug, thanks for reporting it.  I'll work on a
fix.

Later,
David S. Miller
davem@dm.cobaltmicro.com

Re: Problem with unsigned long long on ultrasparc

[David S. Miller]

   Date: Thu, 20 Aug 1998 16:22:03 -0500 (EST)
   From: Brad Lucier <lucier@math.purdue.edu>

   but

   peano-27% gcc -O1 -mcpu=ultrasparc -o testlong testlong.c
   peano-28% testlong
   4294967295ll
   2147483648ll
   Problem

Two things, -1 is an int not a long long (although does GCC promote it
to a long long in this case?), secondarily setting cpu type to
ultrasparc is an error on Solaris-2.5, perhaps you meant
"-mtune=ultrasparc".  "-mcpu=ultrasparc" turns on the V9 instructions
in the compiler which won't work correctly in all cases on Solaris-2.5
since it is not providing a 64-bit userland environment for the cpu.

"-mtune=ultrasparc" on the other hand turns on scheduling and cost
parameters independent of the instruction ISA being used.

Later,
David S. Miller
davem@dm.cobaltmicro.com

Problem with unsigned long long on ultrasparc

[Brad Lucier]

Here's a simple program:

#include <stdlib.h>
#include <stdio.h>

int main() {
  unsigned long long a[10];
	
#define BASE_DIV_2 (((unsigned long long) 1) << 63)

 a[0] = -1;
 printf("%ull\n", a[0]);
 printf("%ull\n", BASE_DIV_2);
 if (a[0] < BASE_DIV_2)
   printf("Problem\n");
 else
   printf("OK\n");
}

With this version of gcc:

peano-30% gcc -v
Reading specs from /opt/egcs-19980816/lib/gcc-lib/sparc-sun-solaris2.5.1/egcs-2.91.54/specs
gcc version egcs-2.91.54 19980816 (gcc2 ss-980609 experimental)

I get:

peano-31% gcc -O1 -o testlong testlong.c
peano-32% testlong
4294967295ll
2147483648ll
OK

but

peano-27% gcc -O1 -mcpu=ultrasparc -o testlong testlong.c
peano-28% testlong
4294967295ll
2147483648ll
Problem

This is on Solaris 2.5.1.

Brad Lucier    lucier@math.purdue.edu