Re: [PATCH] PR libgcc/59714 complex division is surprising on aarch64

[vladimir.mezentsev@oracle.com]

On 10/19/2017 06:07 AM, Wilco Dijkstra wrote:
> Vladimir wrote:
>
> +# Disable floating-point expression contraction
> +LIBGCC2_FFP_CONTRAST_CFLAGS = -ffp-contract=off
> +
>
> It looks like this disables fp-contract in all of libgcc...
> What is the the number of FMAs in libgcc before/after?
In original (without my fix) libgcc_s.so:
% objdump -d libgcc_s.so| egrep '(fmadd|fmsub|fnmadd|fnmsub)'|wc -l
  38

In my new libgcc_s.so ( see my fix below)
% objdump -d libgcc_s.so| egrep '(fmadd|fmsub|fnmadd|fnmsub)'|wc -l
8
> If it disables anything other than the ones in complex division, it
> would have an unintended performance impact. It's best to do
> this just for complex division.

On 10/19/2017 10:17 AM, Wilco Dijkstra wrote:
> Vladimir Mezentsev <vladimir.mezentsev@oracle.com>
>> On 10/19/2017 06:37 AM, Richard Earnshaw (lists) wrote:
>>> On 19/10/17 14:07, Wilco Dijkstra wrote:
>>>> Vladimir wrote:
>>>>
>>>> +# Disable floating-point expression contraction
>>>> +LIBGCC2_FFP_CONTRAST_CFLAGS = -ffp-contract=off
>>>> +
>>>>
>>>> It looks like this disables fp-contract in all of libgcc...
>>>> What is the the number of FMAs in libgcc before/after?
>>    How can I find this number ?
>> dis <all functions in libgcc> | grep <all FMAs> | wc
> Eg. objdump -d ~/install/gcc/lib64/libgcc_s.so | grep -c fmadd
> Also grep fmsub, fnmadd, fnmsub.
>
>>> It's probably better to do this with an attribute
>>>
>>>         __attribute__((optimize("fp-contract=off")))
>>>
>>> on the affected functions.
>>    I like your suggestion.
> I don't think this will work in general, IIRC only a few targets correctly
> implement the optimize pragma. Changing the makefile to build only
> __divdc3/__divtc3/__divsc3/__divhc3 without FMA looks like a better
> approach.

Only 3 (_divdc3 / _divhc3 / _divsc3) have FMAs.
The other FMAs are in:
_fixunsdfdi.o _fixunsdfdi_s.o _fixunssfdi.o _fixunssfdi_s.o _muldc3.o
_muldc3_s.o _mulhc3.o _mulhc3_s.o _mulsc3.o _mulsc3_s.o

I suggest the following fix:

% git diff
diff --git a/libgcc/Makefile.in b/libgcc/Makefile.in
index a1a392d..f313556 100644
--- a/libgcc/Makefile.in
+++ b/libgcc/Makefile.in
@@ -455,6 +455,14 @@ ifeq ($(LIB2_SIDITI_CONV_FUNCS),)
   lib2funcs += $(subst XX,di,$(dwfloatfuncs))
 endif
 
+# Disable FMA (floating-point multiply-add) instructions for complex
division.
+# These instructions can produce different result if two operations
executed separately.
+LIBGCC2_FFP_CONTRAST_CFLAGS = -ffp-contract=off
+LIB2_DIV3_FUNCS = _divdc3 _divhc3 _divsc3
+lib2-div3-o = $(patsubst %,%$(objext),$(LIB2_DIV3_FUNCS))
+lib2-div3-s-o = $(patsubst %,%_s$(objext),$(LIB2_DIV3_FUNCS))
+$(lib2-div3-o) $(lib2-div3-s-o): LIBGCC2_CFLAGS +=
$(LIBGCC2_FFP_CONTRAST_CFLAGS)
+



 If my fix will be approved I send ChangeLog and a patch file tomorrow.

-Vladimir

> However we also need to decide whether this is the best possible fix.
> It will affect accuracy of other inputs as well...
>
> Wilco