[Embench] LLVM 11.0.0 update

paolo paolo.savini at embecosm.com
Thu Oct 22 18:33:48 CEST 2020


Hi all,

as the new release of LLVM has been out for some days now I generated
new benchmark results.

Here are the geomean relative (to the current baseline, see slide 7)
results per optimization flag for RI5CY:

Options               Geomean   +ve   -ve
-------               -------   ---    --
-O0                      0.43  0.23  0.16
-O1                      0.87  0.33  0.24
-O2                      0.97  0.37  0.28
-O3                      0.97  0.35  0.27
-O3 -finline-limit=50    0.97  0.35  0.27
-O3 -funroll-loops       0.97  0.35  0.27
-Os                      0.88  0.34  0.25
-Os -msave-restore       0.87  0.33  0.24
-Oz                      0.87  0.33  0.25
-Oz -msave-restore       0.86  0.33  0.25

Size results
============

Options               Geomean   +ve   -ve
-------               -------   ---    --
-O0                      2.65  0.82  0.63
-O1                      1.24  0.24  0.21
-O2                      1.37  0.42  0.33
-O3                      1.56  0.76  0.52
-O3 -finline-limit=50    1.56  0.76  0.52
-O3 -funroll-loops       1.56  0.76  0.52
-Os                      1.23  0.28  0.23
-Os -msave-restore       1.14  0.28  0.23
-Oz                      1.18  0.24  0.21
-Oz -msave-restore       1.09  0.23  0.20

And here the same for ARM Cortex-M4:

Speed results
=============

Options               Geomean   +ve   -ve
-------               -------   ---    --
-O0                      0.42  0.25  0.16
-O1                      0.94  0.14  0.13
-O2                      1.04  0.13  0.12
-O3                      1.05  0.17  0.16
-O3 -finline-limit=50    1.05  0.17  0.16
-O3 -funroll-loops       1.05  0.17  0.16
-Os                      0.95  0.13  0.12
-Oz                      0.92  0.17  0.15

Size results
============

Options               Geomean   +ve   -ve
-------               -------   ---    --
-O0                      1.85  0.33  0.29
-O1                      1.18  0.14  0.13
-O2                      2.04  1.22  0.77
-O3                      2.18  1.59  0.92
-O3 -finline-limit=50    2.18  1.59  0.92
-O3 -funroll-loops       2.18  1.59  0.92
-Os                      1.16  0.17  0.16
-Oz                      1.02  0.12  0.11

You'll notice that the ARM has no results for the flag combination "-Oz
-msave-restore". That's only because ARM LLVM doesn't support the
"-msave-restore" flag. This is now being specified in the slides as well.

Also note that the top of tree speed result for RI5CY in silde 23 of the
Embench presentation showed during last meeting had a quite surprisingly
worse geomean than LLVM 10.0.1.

That was probably due to a lack of some optimization flags. Nevertheless
the results from LLVM 11 (that follows the top of tree example of the
slides) confirm that there is actually a slight decrease of speed
optimization between LLVM 10.0.1 (a geomean of 0.98) and LLVM 11.0.0 (a
goemean of 0.97) as we mentioned during the call.

The slides are being currently reviewed and updated to integrate the
LLVM 11 results.

For any question/observation don't hesitate to reply.


Speak to you all soon.

Best wishes,

Paolo Savini

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201022/9e6c63c3/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201022/9e6c63c3/attachment.sig>


More information about the Embench mailing list