ICP: gcm-avx: Support architectures lacking the MOVBE instruction #10029
Conversation
AttilaFueloep
changed the title
|
I opened this mainly as a reference for comparing against #10025. It has not been tested yet, which I'll do once time permits. |
Codecov Report
@@ Coverage Diff @@
## master #10029 +/- ##
==========================================
- Coverage 79% 61% -18%
==========================================
Files 386 365 -21
Lines 122448 118117 -4331
==========================================
- Hits 96998 71748 -25250
- Misses 25450 46369 +20919
Continue to review full report at Codecov.
|
|
Looks good. Obviously in #10025 I was going to lengths to avoid duplicating the assembly but that's hardly worth the effort since we don't expect to have to update the code often / ever. |
|
Yes, and since we have the mostly unmodified original Now I wish I'd now why the tests are failing. I suspect accessing the global variable |
|
Looks better now, the ZTS |
|
With unmodified master@392556f0ef I can reproduce the ztest crashes, so let's wait until it stabilizes again. @adamdmoss Could you test if this changes work for you? I can't access my testing VMs right now, so have no way to test on affected CPUs. |
|
@AttilaFueloep master@392556f0ef should be stable. I haven't observed any increase in ztest failures in other PRs. Definitely nothing like the number of failures observed here. I know you're busy but you may want to double check you really can reproduce the issue with a clean build of master. |
|
@behlendorf Yes, you're right, master@392556f0ef is fine. There must have been an error on my part. Sorry for the noise. |
|
@adamdmoss Please don't try this out yet! Although ZTS passes, there seems to be a sincere bug triggered by ztest. |
|
Once I had the chance to run this through |
|
The Ubuntu 16.04 failure seems unrelated but I don't know what to make out of the Fedora failure. Everything else seems fine. I manually verified that the correct code paths are taken on a Ivy Bridge and Skylake virtual CPU. Rebasing on master. |
AttilaFueloep
changed the title
AttilaFueloep
force-pushed
the
icp-gcm-movbe-optional
branch
from
9a7fb36 to
445cde1
Compare
|
Feel free to ignore the Fedora failure. It appears to be caused by some updated packages, we haven't yet run down exactly what they changed. |
Follow-up for openzfs#9749 There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Signed-off-by: Attila Fülöp <attila@fueloep.org>
AttilaFueloep
force-pushed
the
icp-gcm-movbe-optional
branch
from
445cde1 to
ecc9dee
Compare
|
Rebased on master in the hope that the test failures go away. |
|
Looks better now. @behlendorf @adamdmoss I think this is ready for review and testing. |
behlendorf
added
Status: Code Review Needed
Codecov Report
@@ Coverage Diff @@
## master #10029 +/- ##
========================================
+ Coverage 79% 80% +<1%
========================================
Files 385 385
Lines 122385 122390 +5
========================================
+ Hits 97235 97480 +245
+ Misses 25150 24910 -240
Continue to review full report at Codecov.
|
|
@adamdmoss would you mind reviewing this and approving the PR if everything looks good to you. |
|
sorry, behind on github reqs - will take a look! |
|
Code looks good and tests well in casual-but-nontrivial testing on my non-movbe machine. |
behlendorf
added
Status: Accepted
|
@adamdmoss thanks for the quick feedback. This has held up well in my testing as well, I'll get it merged. |
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup openzfs#9749 Closes openzfs#10029
Motivation and Context
Follow-up for #9749
There are a couple of x86_64 architectures which support all needed
features to make the accelerated GCM implementation work but the
MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge
and AMD Bulldozer, Piledriver, and Steamroller.
Description
By using MOVBE only if available and replacing it with a MOV
followed by a BSWAP if not, those architectures now benefit from
the new GCM routines and performance is considerably better
compared to the original implementation.
Signed-off-by: Attila Fülöp attila@fueloep.org
Types of changes
Checklist:
Signed-off-by.