On Fri, Oct 13, 2017 at 10:25 PM, Richard Earnshaw <richard.earns...@foss.arm.com> wrote: > On 13/10/17 18:28, vladimir.mezent...@oracle.com wrote: >> On 10/12/2017 03:40 AM, Richard Earnshaw wrote: >>> On 12/10/17 06:21, vladimir.mezent...@oracle.com wrote: >>>> From: Vladimir Mezentsev <vladimir.mezent...@oracle.com> >>>> >>>> FMA (floating-point multiply-add) instructions are supported on aarch64. >>>> These instructions can produce different result if two operations executed >>>> separately. >>>> -ffp-contract=off doesn't allow the FMA instructions. >>>> >>>> Tested on aarch64-linux-gnu. >>>> No regression. Two failed tests now passed. >>>> >>>> ChangeLog: >>>> 2017-10-11 Vladimir Mezentsev <vladimir.mezent...@oracle.com> >>>> >>>> PR libgcc/59714 >>>> * libgcc/config/aarch64/t-aarch64 (HOST_LIBGCC2_CFLAGS): Add >>>> -ffp-contract=off >>>> --- >>>> libgcc/config/aarch64/t-aarch64 | 1 + >>>> 1 file changed, 1 insertion(+) >>>> >>>> diff --git a/libgcc/config/aarch64/t-aarch64 >>>> b/libgcc/config/aarch64/t-aarch64 >>>> index 3af933c..e33bef0 100644 >>>> --- a/libgcc/config/aarch64/t-aarch64 >>>> +++ b/libgcc/config/aarch64/t-aarch64 >>>> @@ -18,4 +18,5 @@ >>>> # along with GCC; see the file COPYING3. If not see >>>> # <http://www.gnu.org/licenses/>. >>>> >>>> +HOST_LIBGCC2_CFLAGS += -ffp-contract=off >>>> LIB2ADD += $(srcdir)/config/aarch64/sync-cache.c >>>> >>> Why would we want to do this on AArch64 only? If it's right for us, >>> then surely it would be right for everyone and the option should be >>> applied at the top level. >> >> It is a machine dependent option. >> We don't need this option, for example, on sparc or intel machines. >> > > No, it's a target-independent option (it's in the -f... option space). > It might be a no-op on some machines, but either it should be on > globally, or it should be off globally. If it's a no-op then it won't > matter whether it's on or off.
Indeed and we do have this on AArch32 with vfpv4 so that's 2 architectures already affected. It maybe that the default builds on x86 don't tickle fma but we may see the same behavior elsewhere depending on (fp) architecture defaults. I haven't tried this testcase yet on AArch32 but I suspect we're likely to hit the same issue there too ... regards Ramana > > R. > >> -Vladimir >>> >>> Hint: I'm not convinced on the evidence here that it is right across the >>> whole of libgcc. Before moving forward on this particular PR I think we >>> need to understand what behaviour we do want out of the compiler. >>> >>> R. >> >
