[Bug c++/98207] New: ARM64 IEEE math results change with -O2

[Bug c++/98207] New: ARM64 IEEE math results change with -O2

[pramsey at cleverelephant dot ca]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98207

            Bug ID: 98207
           Summary: ARM64 IEEE math results change with -O2
           Product: gcc
           Version: 7.3.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pramsey at cleverelephant dot ca
  Target Milestone: ---

Created attachment 49707
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49707&action=edit
Test program temp files built with O0

On the ARM64 platform (in particular testing on an AWS Graviton).

Linux ip-172-31-55-124.us-west-2.compute.internal
4.14.203-156.332.amzn2.aarch64 #1 SMP Fri Oct 30 19:19:46 UTC 2020 aarch64
aarch64 aarch64 GNU/Linux

A "small" program that carries out only normal arithmetic operations on double
returns a different results if g++ is called with -O0 (or with no O at all) and
-O2. 

Attached are dd-o0.ii and dd-o2.ii, the temp file outputs for compiling with
-O0 and -O2. 

On x84_64, g++ compiled code returns the same answer for both O0 and O2. 

On ARM64 and x84_64, clang++ returns the same answer for both O0 and O2. 

This test is extracted from a small routine that provides "double double"
precision math using only arithmetic operations on IEEE compliant doubles.
https://github.com/libgeos/geos/blob/master/src/math/DD.cpp

[Bug c++/98207] ARM64 IEEE math results change with -O2

[pramsey at cleverelephant dot ca]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98207

--- Comment #1 from Paul Ramsey <pramsey at cleverelephant dot ca> ---
Created attachment 49708
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49708&action=edit
Test program temp files built with O2

[Bug c++/98207] ARM64 IEEE math results change with -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98207

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The default is to use fused multiple add.
If you don't want to use FMA, then use -ffp-contract=off (the default =fast).


x84_64 by default does not have a fused-multiple-add instruction without extra
options (to enable the ISA for fused multiple add instruction).

[Bug c++/98207] ARM64 IEEE math results change with -O2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98207

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #2)
> The default is to use fused multiple add.
> If you don't want to use FMA, then use -ffp-contract=off (the default =fast).
> 
> 
> x84_64 by default does not have a fused-multiple-add instruction without
> extra options (to enable the ISA for fused multiple add instruction).

Note this is still IEEE compliant FP and C standard compliant too.

[Bug c++/98207] ARM64 IEEE math results change with -O2

[pramsey at cleverelephant dot ca]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98207

--- Comment #4 from Paul Ramsey <pramsey at cleverelephant dot ca> ---
Thanks for the pointer!