[tree-ssa] recent C++ Altivec breakage?

[tree-ssa] recent C++ Altivec breakage?

[Timothy J. Wood]

   Sometime recently the Altivec splat instructions seem to have gone 
south:

<http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14425>
<http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14426>

   This was working a week or so ago and I was hoping to retest the 
example that caused me to log 
<http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14042> a while back (now 
that it is supposedly fixed).  Unfortunately, I can't build even fairly 
simple Altivec programs with tree-ssa right now.

-tim

Re: [tree-ssa] recent C++ Altivec breakage?

[Andrew Pinski]

On Mar 3, 2004, at 21:54, Timothy J. Wood wrote:

>
>
>   Sometime recently the Altivec splat instructions seem to have gone 
> south:
>
> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14425>
> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14426>
>
>   This was working a week or so ago and I was hoping to retest the 
> example that caused me to log 
> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14042> a while back (now 
> that it is supposedly fixed).  Unfortunately, I can't build even 
> fairly simple Altivec programs with tree-ssa right now.

Both of these were broken by:
2004-02-21  Ziemowit Laski  <zlaski@apple.com>

         * config/darwin.h (TARGET_OPTION_TRANSLATE_TABLE): Refer to
         SUBTARGET_OPTION_TRANSLATE_TABLE for architecture-specific 
options.
         * config/i386/darwin.h (SUBTARGET_OPTION_TRANSLATE_TABLE): 
Define it.
         * config/rs6000/altivec.h: #error out if '-maltivec' not 
specified.
         (vector, pixel, bool): #define to __vector, __pixel and __bool.
         (__un_args_eq, __bin_args_eq, __tern_args_eq): Move to 
C-specific
         portion of header.
         (__altivec_link_error_invalid_argument): Remove prototype; will 
use
         __builtin_altivec_compiletime_error("vec_*") instead.
         (vec_*): Fix/complete set of available operation overloads 
given the
         existence of distinct 'vector bool ...' and 'vector pixel' 
types; tighten
         cv-correctness of pointer arguments; in C, always check for 
correct
         argument types before macro expansion.
         * config/rs6000/darwin.h (SUBTARGET_OPTION_TRANSLATE_TABLE): 
New macro
         defining Darwin/PowerPC-specific '-f[no-]altivec' and
         '-W[no-]altivec-long-deprecated' switches.
         * config/rs6000/rs6000-c (rs6000_cpu_cpp_builtins): Pre-define
         '__vector', '__pixel' and '__bool' macros using
         '__attribute__((altivec(...)))' types.
         * config/rs6000/rs6000.c (bool_char_type_node, 
bool_short_type_node,
         bool_int_type_node, pixel_type_node, bool_V16QI_type_node,
         bool_V8HI_type_node, bool_V4SI_type_node, pixel_V8HI_type_node):
         New type nodes.
         (rs6000_warn_altivec_long, rs6000_warn_altivec_long_switch): 
New, for
         handling '-W[no-]altivec-long-deprecated'.
         (rs6000_override_options): Handle 
'-W[no-]altivec-long-deprecated'.
         (rs6000_expand_binop_builtin, rs6000_expand_ternop_builtin,
         altivec_expand_dst_builtin): Remove casts from integer literals.
         (altivec_expand_builtin): Likewise; handle expansion of new
         '__builtin_altivec_compiletime_error' function.
         (rs6000_init_builtins): Initialize 'vector bool ...' and 
'vector pixel'
         types, and make them distinct from other vector types; register
         '__builtin_altivec_compiletime_error' function.
         (print_operand): For 'P', print a full target register name 
instead of
         merely its number.
         (rs6000_attribute_table): Add "altivec" attribute.
         (rs6000_handle_altivec_attribute): New function.
         * config/rs6000/rs6000.h (TARGET_OPTIONS): Describe
         '-m[no-]-warn-altivec-long' (which 
'-W[no-]altivec-long-deprecated'
         maps to).
         (rs6000_warn_altivec_long, rs6000_warn_altivec_long_switch): 
Forward
         declare.
         (ALTIVEC_BUILTIN_COMPILETIME_ERROR): New built-in enumeration.


Thanks,
Andrew Pinski

Re: [tree-ssa] recent C++ Altivec breakage?

[Timothy J. Wood]

On Wednesday, March 3, 2004, at 11:08  PM, Andrew Pinski wrote:

>
> On Mar 3, 2004, at 21:54, Timothy J. Wood wrote:
>
>>
>>
>>   Sometime recently the Altivec splat instructions seem to have gone 
>> south:
>>
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14425>
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14426>
>>
>>   This was working a week or so ago and I was hoping to retest the 
>> example that caused me to log 
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14042> a while back (now 
>> that it is supposedly fixed).  Unfortunately, I can't build even 
>> fairly simple Altivec programs with tree-ssa right now.
>
> Both of these were broken by:
> 2004-02-21  Ziemowit Laski  <zlaski@apple.com>

   So I guess this answers a question I had -- apparently Apple doesn't 
have an internal (more comprehensive) Altivec test suite.  Does 
Motorola or IBM have a test suite that could be used or would a 
volunteer need to build one?

   It seems reasonable to expect a minimal test suite to at least:

- Call every PIM-specified function at least once, checking the output
	- Pseudo things like vec_splat
	- 'Real' instructions like vec_vspltisw
- Test error reporting for inputs (non-const arguments)

-tim

Re: [tree-ssa] recent C++ Altivec breakage?

[Ziemowit Laski]

Tim, I'll take a look at these pronto.  Sorry for the inconvenience.

--Zem

On 3 Mar, 2004, at 23.08, Andrew Pinski wrote:

>
> On Mar 3, 2004, at 21:54, Timothy J. Wood wrote:
>
>>
>>
>>   Sometime recently the Altivec splat instructions seem to have gone 
>> south:
>>
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14425>
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14426>
>>
>>   This was working a week or so ago and I was hoping to retest the 
>> example that caused me to log 
>> <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14042> a while back (now 
>> that it is supposedly fixed).  Unfortunately, I can't build even 
>> fairly simple Altivec programs with tree-ssa right now.
>
> Both of these were broken by:
> 2004-02-21  Ziemowit Laski  <zlaski@apple.com>
>
>         * config/darwin.h (TARGET_OPTION_TRANSLATE_TABLE): Refer to
>         SUBTARGET_OPTION_TRANSLATE_TABLE for architecture-specific 
> options.
>         * config/i386/darwin.h (SUBTARGET_OPTION_TRANSLATE_TABLE): 
> Define it.
>         * config/rs6000/altivec.h: #error out if '-maltivec' not 
> specified.
>         (vector, pixel, bool): #define to __vector, __pixel and __bool.
>         (__un_args_eq, __bin_args_eq, __tern_args_eq): Move to 
> C-specific
>         portion of header.
>         (__altivec_link_error_invalid_argument): Remove prototype; 
> will use
>         __builtin_altivec_compiletime_error("vec_*") instead.
>         (vec_*): Fix/complete set of available operation overloads 
> given the
>         existence of distinct 'vector bool ...' and 'vector pixel' 
> types; tighten
>         cv-correctness of pointer arguments; in C, always check for 
> correct
>         argument types before macro expansion.
>         * config/rs6000/darwin.h (SUBTARGET_OPTION_TRANSLATE_TABLE): 
> New macro
>         defining Darwin/PowerPC-specific '-f[no-]altivec' and
>         '-W[no-]altivec-long-deprecated' switches.
>         * config/rs6000/rs6000-c (rs6000_cpu_cpp_builtins): Pre-define
>         '__vector', '__pixel' and '__bool' macros using
>         '__attribute__((altivec(...)))' types.
>         * config/rs6000/rs6000.c (bool_char_type_node, 
> bool_short_type_node,
>         bool_int_type_node, pixel_type_node, bool_V16QI_type_node,
>         bool_V8HI_type_node, bool_V4SI_type_node, 
> pixel_V8HI_type_node):
>         New type nodes.
>         (rs6000_warn_altivec_long, rs6000_warn_altivec_long_switch): 
> New, for
>         handling '-W[no-]altivec-long-deprecated'.
>         (rs6000_override_options): Handle 
> '-W[no-]altivec-long-deprecated'.
>         (rs6000_expand_binop_builtin, rs6000_expand_ternop_builtin,
>         altivec_expand_dst_builtin): Remove casts from integer 
> literals.
>         (altivec_expand_builtin): Likewise; handle expansion of new
>         '__builtin_altivec_compiletime_error' function.
>         (rs6000_init_builtins): Initialize 'vector bool ...' and 
> 'vector pixel'
>         types, and make them distinct from other vector types; register
>         '__builtin_altivec_compiletime_error' function.
>         (print_operand): For 'P', print a full target register name 
> instead of
>         merely its number.
>         (rs6000_attribute_table): Add "altivec" attribute.
>         (rs6000_handle_altivec_attribute): New function.
>         * config/rs6000/rs6000.h (TARGET_OPTIONS): Describe
>         '-m[no-]-warn-altivec-long' (which 
> '-W[no-]altivec-long-deprecated'
>         maps to).
>         (rs6000_warn_altivec_long, rs6000_warn_altivec_long_switch): 
> Forward
>         declare.
>         (ALTIVEC_BUILTIN_COMPILETIME_ERROR): New built-in enumeration.
>
>
> Thanks,
> Andrew Pinski
>
>
--------------------------------------------------------------
Ziemowit Laski                 1 Infinite Loop, MS 301-2K
Mac OS X Compiler Group        Cupertino, CA USA  95014-2083
Apple Computer, Inc.           +1.408.974.6229  Fax .5477

Re: [tree-ssa] recent C++ Altivec breakage?

[Jim Wilson]

Timothy J. Wood wrote:
> have an internal (more comprehensive) Altivec test suite.  Does Motorola 
> or IBM have a test suite that could be used or would a volunteer need to 
> build one?

Motorola has nice testsuites for both Altivec and SPE.  When I was at 
Red Hat, I asked if they could be contributed or made publicly 
available.  I left not long afterwards, so I don't know what happened.

By the way, Zack of CodeSourcery tried to contribute a version of the 
Motorola Altivec testsuite, but I pointed out that it was copyright 
Motorola, and we couldn't distribute it without their permission, so it 
did not go in.  See
     http://gcc.gnu.org/ml/gcc-patches/2002-11/msg01391.html
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com

Re: [tree-ssa] recent C++ Altivec breakage?

[Zack Weinberg]

Jim Wilson <wilson@specifixinc.com> writes:

> By the way, Zack of CodeSourcery tried to contribute a version of the
> Motorola Altivec testsuite, but I pointed out that it was copyright
> Motorola, and we couldn't distribute it without their permission, so
> it did not go in.  See
>      http://gcc.gnu.org/ml/gcc-patches/2002-11/msg01391.html

I did find out who to ask at Motorola, and sent them mail periodically
for about three months before giving up.  Never heard a word.

zw

Re: [tree-ssa] recent C++ Altivec breakage?

[Aldy Hernandez]

>>>>> "Jim" == Jim Wilson <wilson@specifixinc.com> writes:

 > Timothy J. Wood wrote:
 >> have an internal (more comprehensive) Altivec test suite.  Does
 >> Motorola or IBM have a test suite that could be used or would a
 >> volunteer need to build one?

 > Motorola has nice testsuites for both Altivec and SPE.  When I was at
 > Red Hat, I asked if they could be contributed or made publicly
 > available.  I left not long afterwards, so I don't know what happened.

 > By the way, Zack of CodeSourcery tried to contribute a version of the
 > Motorola Altivec testsuite, but I pointed out that it was copyright
 > Motorola, and we couldn't distribute it without their permission, so
 > it did not go in.  See

It was my understanding there were both GPL'ed, but needed to be
brought into the dejagnu framework.

Aldy