Re: [PATCH 1/3] Rework 128-bit complex multiply and divide, PR target/107299

["Kewen.Lin" <linkw@linux.ibm.com>]

on 2022/12/9 06:04, Michael Meissner wrote:
> On Wed, Dec 07, 2022 at 03:55:41PM +0800, Kewen.Lin wrote:
>> Hi Mike,
>>
>> on 2022/12/7 14:44, Michael Meissner wrote:
>>> On Tue, Dec 06, 2022 at 05:36:54PM +0800, Kewen.Lin wrote:
>>>> Hi Mike,
>>>>
>>>> Thanks for fixing this!
>>>>
>>>> Could you help to elaborate why we need to disable it during libgcc building?
>>>
>>> When you are building libgcc, you are building the __mulkc3, __divkc3
>>> functions.  The mapping in the compiler interferes with those functions,
>>> because at the moment, libgcc uses an alternate IEEE 128-bit type.
>>>
>>
>> But I'm still confused.  For __mulkc3 (__divkc3 is similar),
>>
>> 1) with -mabi=ieeelongdouble (TARGET_IEEEQUAD true, define __LONG_DOUBLE_IEEE128__),
>>    the used types are:
>>
>>    typedef float TFtype __attribute__ ((mode (TF)));
>>    typedef __complex float TCtype __attribute__ ((mode (TC)));
>>
>> 2) with -mabi=ibmlongdouble (TARGET_IEEEQUAD false, not __LONG_DOUBLE_IEEE128__ defined),
>>    the used types are:
>>
>>    typedef float TFtype __attribute__ ((mode (KF)));
>>    typedef __complex float TCtype __attribute__ ((mode (KC)));
>>
>> The proposed mapping in the current patch is:
>>
>> +
>> +      if (id == complex_multiply_builtin_code (KCmode))
>> +	newname = "__mulkc3";
>> +
>> +      else if (id == complex_multiply_builtin_code (ICmode))
>> +	newname = "__multc3";
>> +
>> +      else if (id == complex_multiply_builtin_code (TCmode))
>> +	newname = (TARGET_IEEEQUAD) ? "__mulkc3" : "__multc3";
>>
>> for 1), TCmode && TARGET_IEEEQUAD => "__mulkc3"
>> for 2), KCmode => "__mulkc3"
>>
>> Both should be still with name "__mulkc3", do I miss anything?
>>
>> BR,
>> Kewen
> 
> The reason is due to the different internal types, the value range propigation
> pass throws an error when we are trying to build libgcc.  This is due to the
> underlying problem of different IEEE 128-bit types within the compiler.
> 

But this is the reason why we need patch #2 and #3, not the reason why we need
the special handling for building_libgcc in patch #1, right?

Without or with patch #1, the below ICE in libgcc exists, the ICE should have
nothing to do with the special handling for building_libgcc in patch #1.  I
think patch #2 which makes _Float128 and __float128 use the same internal
type fixes that ICE.

I still don't get the point why we need the special handling for building_libgcc,
I also tested on top of patch #1 and #2 w/ and w/o the special handling for
building_libgcc, both bootstrapped and regress-tested.

Could you have a double check?

> The 128-bit IEEE support in libgcc was written before _Float128 was added to
> GCC.  One consequence is that you can't get to the complex variant of
> __float128.  So libgcc needs to use the attribute mode to get to that type.
> 
> But with the support for IEEE 128-bit long double changing things, it makes the
> libgcc code use the wrong code.
> 
> /home/meissner/fsf-src/work102/libgcc/config/rs6000/_mulkc3.c: In function ‘__mulkc3_sw’:
> /home/meissner/fsf-src/work102/libgcc/config/rs6000/_mulkc3.c:97:1: internal compiler error: in fold_stmt, at gimple-range-fold.cc:522
>    97 | }
>       | ^
> 0x122784f3 fold_using_range::fold_stmt(vrange&, gimple*, fur_source&, tree_node*)
>         /home/meissner/fsf-src/work102/gcc/gimple-range-fold.cc:522
> 0x1226477f gimple_ranger::fold_range_internal(vrange&, gimple*, tree_node*)
>         /home/meissner/fsf-src/work102/gcc/gimple-range.cc:257
> 0x12264b1f gimple_ranger::range_of_stmt(vrange&, gimple*, tree_node*)
>         /home/meissner/fsf-src/work102/gcc/gimple-range.cc:318
> 0x113bdd8b range_query::value_of_stmt(gimple*, tree_node*)
>         /home/meissner/fsf-src/work102/gcc/value-query.cc:134
> 0x1134838f rvrp_folder::value_of_stmt(gimple*, tree_node*)
>         /home/meissner/fsf-src/work102/gcc/tree-vrp.cc:1023
> 0x111344cf substitute_and_fold_dom_walker::before_dom_children(basic_block_def*)
>         /home/meissner/fsf-src/work102/gcc/tree-ssa-propagate.cc:819
> 0x121ecbd3 dom_walker::walk(basic_block_def*)
>         /home/meissner/fsf-src/work102/gcc/domwalk.cc:311
> 0x11134ee7 substitute_and_fold_engine::substitute_and_fold(basic_block_def*)
>         /home/meissner/fsf-src/work102/gcc/tree-ssa-propagate.cc:998
> 0x11346bb7 execute_ranger_vrp(function*, bool, bool)
>         /home/meissner/fsf-src/work102/gcc/tree-vrp.cc:1084
> 0x11347063 execute
>         /home/meissner/fsf-src/work102/gcc/tree-vrp.cc:1165
> Please submit a full bug report, with preprocessed source (by using -freport-bug).
> Please include the complete backtrace with any bug report.
> See <https://gcc.gnu.org/bugs/> for instructions.
> make[1]: *** [/home/meissner/fsf-src/work102/libgcc/shared-object.mk:14: _mulkc3.o] Error 1
> make[1]: Leaving directory '/home/meissner/fsf-build-ppc64le/work102/powerpc64le-unknown-linux-gnu/libgcc'
> make: *** [Makefile:20623: all-target-libgcc] Error 2
> 
>>> I have a patch for making libgcc use the 'right' type that I haven't submitted
>>> yet.  This is because the more general fix that these 3 patches do impacts other
>>> functions (due to __float128 and _Float128 being different in the current
>>> compiler when -mabi=ieeelongdouble).
>>>
> 
> The patch is to use _Float128 and _Complex _Float128 in libgcc.h instead of
> trying to use attribute((mode(TF))) and attribute((mode(TC))) in libgcc.
> 

Since your patch #2 (and #3) fixes ICE and some exposed problems, and _Float128
is to use the same internal type as __float128, types with attribute((mode(TF)))
and attribute((mode(TC))) should be correct, I assume that this patch is just
to make the types explicit be with _Float128 (for better readability and
maintainance), but not for any correctness issues.

> Now, this patch fixes the specific problem of not being able to build libgcc
> (along with patch #1 of the series).  But other things show the differences
> from time time because we are using different internal types and the middle end
> doesn't know that these types are really the same bits.
> 
> It is better long term (IMHO) if we have the two types (__float128 and
> _Float128) use the same internal type (which is what is done in patches #2 and
> #3).  This fixes the other issues that show up, such as creating signaling NaNs
> for one internal type, and converting it to the other internal type, loses that
> the NaN is signalling.
> 

I see, nice!

BR,
Kewen