llvm dev - Apr 2017 - [GlobalISel][AArch64] Toward flipping the switch for O0: Please give it a try!

Hi Kristof,
> On Apr 6, 2017, at 6:53 AM, Kristof Beyls <kristof.beyls at arm.com>
wrote:
> 
> I've been digging a little bit deeper into the biggest performance
regressions I've observed.
> 
> What I've observed so far is:
> * A lot of the biggest regressions are caused by unnecessarily moving
floating point values through general purpose registers. I've raised
http://bugs.llvm.org/show_bug.cgi?id=32550
<http://bugs.llvm.org/show_bug.cgi?id=32550> for this. I think this one
definitely needs fixing before enabling GlobalISel by default at -O0.
I commented in the PR. This is a known problem and we have a solution. Given
this is an optimization in the sense that it does not affect the correctness of
the program, we didn’t push for fixing it now.

For O0 we wanted to focus ourselves on generating correct code. Unless the
regressions you are seeing are preventing debugging/running of the program, I
wouldn’t block the flip of the switch on that.

What do you think? 
> * FastISel seems to transform division-by-constant-power-of-2 into right
shift (see
https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/SelectionDAG/FastISel.cpp#L456-L468
<https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/SelectionDAG/FastISel.cpp#L456-L468>).
GlobalISel does not. It seems to me that at -O0 there may be reasons not perform
this transformation, but maybe there is a good reason why FastISel does this?
I think FastISel tries to generate the best code it can no matter what. For
GISel O0 however, not doing this optimization sounds sensible to me.
Now, I would say that the same remark as the previous bullet point apply: we
shouldn’t do it unless it gets in the way of running/debugging the program.
> * FastISel doesn’t\ seem to handle functions with switch statements, so it
falls back to DAGISel. DAGISel produces code that's a lot better than
GlobalISel for switch statement at -O0. I'm not sure if we need to do
something here before enabling GlobalISel by default. I'm thinking we may
need to add a smarter way to lower switch statements rather than just a cascaded
sequence of conditional branches.
Sounds optimization-ish to me. Same remark.
> 
> I'll try to add the above content to the document Diana created at
https://goo.gl/IS2Bdw <https://goo.gl/IS2Bdw> too.
> 
> Thanks,
> 
> Kristof
> 
> 
> 
>> On 3 Apr 2017, at 17:10, Kristof Beyls <Kristof.Beyls at arm.com
<mailto:Kristof.Beyls at arm.com>> wrote:
>> 
>> I've kicked off a run to compare "-O0 -g" versus
"-O0 -g -mllvm -global-isel -mllvm -global-isel-abort=2".
>> I've selected the test-suite (albeit a version which is a couple of
months old now) and a few short-running proprietary benchmarks to get data back
quickly for an initial feel of where things are.
>> This was running on Cortex-A57 AArch64 Linux.
>> 
>> I saw one assertion failure in GlobalISel, see
http://bugs.llvm.org/show_bug.cgi?id=32471
<http://bugs.llvm.org/show_bug.cgi?id=32471>. This is in a program
compiled at -O2 (my out-dated test-suite still overrides -O0 and instead uses -O
for that program). The root cause of the failure seems to be due to LowLevelType
not supporting vectors of pointers. I think this demonstrates that for
correctness, we should be trying to test more than -O0, or even more than just
LLVM-IR produced by clang, as other front-ends could run into this even at -O0.
>> 
>> Due to this assertion failure and the infrastructure I used, the
numbers below do not include test-suite/MultiSource/Benchmarks results.
>> 
>> On the non-correctness aspects, LNT tells me that:
>> - The programs that report execution time, on geomean are about 17%
slower.
>> - The programs that report scores, on geomean are about 21% slower.
>> - Code size is up on geomean about 11%.
>> I'm afraid I don't have compile time numbers, nor any feel for
debug info quality.
>> 
>> I'll need quite a bit more time to dig into the details to come up
with something actionable, although the fact that LowLevelType doesn't
support vectors of pointers is already actionable.
>> Nevertheless, I thought to share what I see as is, to see if others see
similar results so far.
>> 
>> I thought Diana was going to look into fallback rate on the test-suite
on AArch64 linux?
>> 
>> Thanks,
>> 
>> Kristof
>> 
>>> On 30 Mar 2017, at 10:54, Renato Golin <renato.golin at
linaro.org <mailto:renato.golin at linaro.org>> wrote:
>>> 
>>> On 30 March 2017 at 00:27, Quentin Colombet <qcolombet at
apple.com <mailto:qcolombet at apple.com>> wrote:
>>>> On iOS we are at 100% pass rate in 00 g for the LLVM test
suite, standard
>>>> benchmarks and unit tests. In about 5% of all functions
GlobalIsel falls
>>>> back to SDIsel.
>>>> (Kristof Beyls would have the linux numbers.)
>>>> The self host compiler correctly builds and runs the LLVM test
suite in O0.
>>> 
>>> Having done no tests at all on my side, I think we need to have
>>> similar numbers on Linux to be able to flip across the board.
>>> 
>>> I don't want to flip it only for Darwin and not Linux, as that
will
>>> fragment the effort too much.
>>> 
>>> I'll check with Diana and Kristof to know what's the best
way forward,
>>> but it should be reasonably quick.
>>> 
>>> cheers,
>>> --renato
>> 
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170426/745d5645/attachment.html>

Hi Quentin,

On 27 Apr 2017, at 00:48, Quentin Colombet <qcolombet at
apple.com<mailto:qcolombet at apple.com>> wrote:

Hi Kristof,

On Apr 6, 2017, at 6:53 AM, Kristof Beyls <kristof.beyls at
arm.com<mailto:kristof.beyls at arm.com>> wrote:

I've been digging a little bit deeper into the biggest performance
regressions I've observed.

What I've observed so far is:
* A lot of the biggest regressions are caused by unnecessarily moving floating
point values through general purpose registers. I've raised
http://bugs.llvm.org/show_bug.cgi?id=32550 for this. I think this one definitely
needs fixing before enabling GlobalISel by default at -O0.

I commented in the PR. This is a known problem and we have a solution. Given
this is an optimization in the sense that it does not affect the correctness of
the program, we didn’t push for fixing it now.

For O0 we wanted to focus ourselves on generating correct code. Unless the
regressions you are seeing are preventing debugging/running of the program, I
wouldn’t block the flip of the switch on that.

What do you think?

For O0, I think most users care about fast compile time and an excellent debug
experience.
Next to that, I am told that some users also use -O0 to have a straightforward,
simple, mapping between source code and the generated assembly.

Out of the issues I've seen so far, I think this is the single
"optimization" issue that I feel gives a negative first impression of
GlobalISel.
If users look at the generated assembly for floating point code it looks more
like active de-optimization rather than "no optimization".
My guess is also that this may be the biggest reason for the about 20%
performance slow-down and 11% code size increase I measured earlier.
OTOH, I clearly agree this is an optimization issue, not a correctness issue.
Combining the above, I think it would be best if this issue got fixed before we
turn on GlobalISel by default at -O0, or we may end up hearing quite a few
(non-critical) complaints about this from users.
Basically I think this is a tradeoff between giving a better first impression of
GlobalISel vs getting more people to use and test it earlier.

Thanks for the write-up on the PR, that is very useful.
Do you have any rough idea how much effort would be involved in getting this
fixed? I got the impression Daniel is making good progress on the tablegen
generation, which is key to getting this issue fixed?
I think that no matter whether we decide to switch the default before or after
fixing this issue, this is one of the most urgent issues to fix as far as I can
see.

If there is some way I can help contribute to fixing PR32550, I would like to
help; but with the dependency on tablegen generation, I'm not sure what the
best way is to help make that PR get fixed faster?


* FastISel seems to transform division-by-constant-power-of-2 into right shift
(see
https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/SelectionDAG/FastISel.cpp#L456-L468).
GlobalISel does not. It seems to me that at -O0 there may be reasons not perform
this transformation, but maybe there is a good reason why FastISel does this?

I think FastISel tries to generate the best code it can no matter what. For
GISel O0 however, not doing this optimization sounds sensible to me.
Now, I would say that the same remark as the previous bullet point apply: we
shouldn’t do it unless it gets in the way of running/debugging the program.

I agree that these optimizations should not be done at -O0. I think not doing
them is actually an improvement: you give the user what they asked, i.e.
"no optimization", and an as-straigthforward-as-possible mapping from
source to assembly.

* FastISel doesn’t\ seem to handle functions with switch statements, so it falls
back to DAGISel. DAGISel produces code that's a lot better than GlobalISel
for switch statement at -O0. I'm not sure if we need to do something here
before enabling GlobalISel by default. I'm thinking we may need to add a
smarter way to lower switch statements rather than just a cascaded sequence of
conditional branches.

Sounds optimization-ish to me. Same remark.

Agreed, optimization-ish. In comparison to the above point on FastISel
"peepholes", I think that lowering switch statements to something else
than a cascaded sequence of conditional branches doesn't make the generated
code harder to map to the source. So, in comparison to the above point on
FastISel "peepholes", I'm not actively against being smarter about
lowering switch statements at -O0.
But I agree, this shouldn't hold up turning on GlobalISel by default at -O0.


Other than the above remarks, before turning on GlobalISel by default, we'd
better test/verify that debug info quality remains good.
I haven't looked into that at all, but am hoping to start looking into that
soon, with the help of the DIVA tool (https://github.com/SNSystems/DIVA)
presented at last EuroLLVM
(http://llvm.org/devmtg/2017-03//assets/slides/diva_debug_information_visual_analyzer.pdf).
I don't recall anyone so far making any statements about the quality of the
debug info they've measured or experienced with GlobalISel enabled?

Other than the all of the above, I just wanted to mention that Oliver recently
started running csmith on AArch64 GlobalISel and found one issue so far that
already got fixed (https://bugs.llvm.org//show_bug.cgi?id=32733).
If he finds other correctness issues, I'm sure he'll keep on reporting
them.

I'll try to add the above content to the document Diana created at
https://goo.gl/IS2Bdw too.

Thanks,

Kristof



On 3 Apr 2017, at 17:10, Kristof Beyls <Kristof.Beyls at
arm.com<mailto:Kristof.Beyls at arm.com>> wrote:

I've kicked off a run to compare "-O0 -g" versus "-O0 -g
-mllvm -global-isel -mllvm -global-isel-abort=2".
I've selected the test-suite (albeit a version which is a couple of months
old now) and a few short-running proprietary benchmarks to get data back quickly
for an initial feel of where things are.
This was running on Cortex-A57 AArch64 Linux.

I saw one assertion failure in GlobalISel, see
http://bugs.llvm.org/show_bug.cgi?id=32471. This is in a program compiled at -O2
(my out-dated test-suite still overrides -O0 and instead uses -O for that
program). The root cause of the failure seems to be due to LowLevelType not
supporting vectors of pointers. I think this demonstrates that for correctness,
we should be trying to test more than -O0, or even more than just LLVM-IR
produced by clang, as other front-ends could run into this even at -O0.

Due to this assertion failure and the infrastructure I used, the numbers below
do not include test-suite/MultiSource/Benchmarks results.

On the non-correctness aspects, LNT tells me that:
- The programs that report execution time, on geomean are about 17% slower.
- The programs that report scores, on geomean are about 21% slower.
- Code size is up on geomean about 11%.
I'm afraid I don't have compile time numbers, nor any feel for debug
info quality.

I'll need quite a bit more time to dig into the details to come up with
something actionable, although the fact that LowLevelType doesn't support
vectors of pointers is already actionable.
Nevertheless, I thought to share what I see as is, to see if others see similar
results so far.

I thought Diana was going to look into fallback rate on the test-suite on
AArch64 linux?

Thanks,

Kristof

On 30 Mar 2017, at 10:54, Renato Golin <renato.golin at
linaro.org<mailto:renato.golin at linaro.org>> wrote:

On 30 March 2017 at 00:27, Quentin Colombet <qcolombet at
apple.com<mailto:qcolombet at apple.com>> wrote:
On iOS we are at 100% pass rate in 00 g for the LLVM test suite, standard
benchmarks and unit tests. In about 5% of all functions GlobalIsel falls
back to SDIsel.
(Kristof Beyls would have the linux numbers.)
The self host compiler correctly builds and runs the LLVM test suite in O0.

Having done no tests at all on my side, I think we need to have
similar numbers on Linux to be able to flip across the board.

I don't want to flip it only for Darwin and not Linux, as that will
fragment the effort too much.

I'll check with Diana and Kristof to know what's the best way forward,
but it should be reasonably quick.

cheers,
--renato




-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170427/b4dbdaa2/attachment-0001.html>

Hi Kristof,
> On Apr 27, 2017, at 9:47 AM, Kristof Beyls <kristof.beyls at arm.com>
wrote:
> 
> Hi Quentin,
> 
>> On 27 Apr 2017, at 00:48, Quentin Colombet <qcolombet at apple.com
<mailto:qcolombet at apple.com>> wrote:
>> 
>> Hi Kristof,
>> 
>>> On Apr 6, 2017, at 6:53 AM, Kristof Beyls <kristof.beyls at
arm.com <mailto:kristof.beyls at arm.com>> wrote:
>>> 
>>> I've been digging a little bit deeper into the biggest
performance regressions I've observed.
>>> 
>>> What I've observed so far is:
>>> * A lot of the biggest regressions are caused by unnecessarily
moving floating point values through general purpose registers. I've raised
http://bugs.llvm.org/show_bug.cgi?id=32550
<http://bugs.llvm.org/show_bug.cgi?id=32550> for this. I think this one
definitely needs fixing before enabling GlobalISel by default at -O0.
>> 
>> I commented in the PR. This is a known problem and we have a solution.
Given this is an optimization in the sense that it does not affect the
correctness of the program, we didn’t push for fixing it now.
>> 
>> For O0 we wanted to focus ourselves on generating correct code. Unless
the regressions you are seeing are preventing debugging/running of the program,
I wouldn’t block the flip of the switch on that.
>> 
>> What do you think? 
> 
> For O0, I think most users care about fast compile time and an excellent
debug experience.
> Next to that, I am told that some users also use -O0 to have a
straightforward, simple, mapping between source code and the generated assembly.
> 
> Out of the issues I've seen so far, I think this is the single
"optimization" issue that I feel gives a negative first impression of
GlobalISel.
> If users look at the generated assembly for floating point code it looks
more like active de-optimization rather than "no optimization".
> My guess is also that this may be the biggest reason for the about 20%
performance slow-down and 11% code size increase I measured earlier.
> OTOH, I clearly agree this is an optimization issue, not a correctness
issue.
> Combining the above, I think it would be best if this issue got fixed
before we turn on GlobalISel by default at -O0, or we may end up hearing quite a
few (non-critical) complaints about this from users.
> Basically I think this is a tradeoff between giving a better first
impression of GlobalISel vs getting more people to use and test it earlier.
> 
> Thanks for the write-up on the PR, that is very useful.
> Do you have any rough idea how much effort would be involved in getting
this fixed?
I’d say 2-3 weeks. Could actually be shorter if we don’t do all the refactoring
I have in mind to make the table for the alternative mappings smaller, but I
don’t think it is worth taking any shortcut here.
> I got the impression Daniel is making good progress on the tablegen
generation, which is key to getting this issue fixed?
To be accurate, Daniel’s work on table gen is for the select phase, not
regbankselect. In other words, right now, all the table for mappings for
regbankselect are hand written and Daniel’s work is not changing that. The
(weak) rationale is that it is not on the critical path :).
> I think that no matter whether we decide to switch the default before or
after fixing this issue, this is one of the most urgent issues to fix as far as
I can see.
Agree. This is one of the top items of my todo list.
> 
> If there is some way I can help contribute to fixing PR32550, I would like
to help; but with the dependency on tablegen generation, I'm not sure what
the best way is to help make that PR get fixed faster?
Thanks for offering to help. I agree that with the dependency on the tabelgen
generation in the way, this is probably not a good use of your time. Depending
when I can spare some time on this, I’ll either do it or explain the refactoring
in the google doc.

In the meantime, I’d suggest to focus on validating the debug info on your side.
> 
> 
>>> * FastISel seems to transform division-by-constant-power-of-2 into
right shift (see
https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/SelectionDAG/FastISel.cpp#L456-L468
<https://github.com/llvm-mirror/llvm/blob/master/lib/CodeGen/SelectionDAG/FastISel.cpp#L456-L468>).
GlobalISel does not. It seems to me that at -O0 there may be reasons not perform
this transformation, but maybe there is a good reason why FastISel does this?
>> 
>> I think FastISel tries to generate the best code it can no matter what.
For GISel O0 however, not doing this optimization sounds sensible to me.
>> Now, I would say that the same remark as the previous bullet point
apply: we shouldn’t do it unless it gets in the way of running/debugging the
program.
> 
> I agree that these optimizations should not be done at -O0. I think not
doing them is actually an improvement: you give the user what they asked, i.e.
"no optimization", and an as-straigthforward-as-possible mapping from
source to assembly.
> 
>>> * FastISel doesn’t\ seem to handle functions with switch
statements, so it falls back to DAGISel. DAGISel produces code that's a lot
better than GlobalISel for switch statement at -O0. I'm not sure if we need
to do something here before enabling GlobalISel by default. I'm thinking we
may need to add a smarter way to lower switch statements rather than just a
cascaded sequence of conditional branches.
>> 
>> Sounds optimization-ish to me. Same remark.
> 
> Agreed, optimization-ish. In comparison to the above point on FastISel
"peepholes", I think that lowering switch statements to something else
than a cascaded sequence of conditional branches doesn't make the generated
code harder to map to the source. So, in comparison to the above point on
FastISel "peepholes", I'm not actively against being smarter about
lowering switch statements at -O0.
> But I agree, this shouldn't hold up turning on GlobalISel by default at
-O0.
> 
> 
> Other than the above remarks, before turning on GlobalISel by default,
we'd better test/verify that debug info quality remains good.
> I haven't looked into that at all, but am hoping to start looking into
that soon, with the help of the DIVA tool (https://github.com/SNSystems/DIVA
<https://github.com/SNSystems/DIVA>) presented at last EuroLLVM
(http://llvm.org/devmtg/2017-03//assets/slides/diva_debug_information_visual_analyzer.pdf
<http://llvm.org/devmtg/2017-03//assets/slides/diva_debug_information_visual_analyzer.pdf>).
I don't recall anyone so far making any statements about the quality of the
debug info they've measured or experienced with GlobalISel enabled?
We ran the lldb test suite with GISel. IIRC, Tim has the details, GISel was on
part with SDISel.
> 
> Other than the all of the above, I just wanted to mention that Oliver
recently started running csmith on AArch64 GlobalISel and found one issue so far
that already got fixed (https://bugs.llvm.org//show_bug.cgi?id=32733
<https://bugs.llvm.org//show_bug.cgi?id=32733>).
> If he finds other correctness issues, I'm sure he'll keep on
reporting them.
Great!

Thanks,
-Quentin
> 
>>> I'll try to add the above content to the document Diana created
at https://goo.gl/IS2Bdw <https://goo.gl/IS2Bdw> too.
>>> 
>>> Thanks,
>>> 
>>> Kristof
>>> 
>>> 
>>> 
>>>> On 3 Apr 2017, at 17:10, Kristof Beyls <Kristof.Beyls at
arm.com <mailto:Kristof.Beyls at arm.com>> wrote:
>>>> 
>>>> I've kicked off a run to compare "-O0 -g" versus
"-O0 -g -mllvm -global-isel -mllvm -global-isel-abort=2".
>>>> I've selected the test-suite (albeit a version which is a
couple of months old now) and a few short-running proprietary benchmarks to get
data back quickly for an initial feel of where things are.
>>>> This was running on Cortex-A57 AArch64 Linux.
>>>> 
>>>> I saw one assertion failure in GlobalISel, see
http://bugs.llvm.org/show_bug.cgi?id=32471
<http://bugs.llvm.org/show_bug.cgi?id=32471>. This is in a program
compiled at -O2 (my out-dated test-suite still overrides -O0 and instead uses -O
for that program). The root cause of the failure seems to be due to LowLevelType
not supporting vectors of pointers. I think this demonstrates that for
correctness, we should be trying to test more than -O0, or even more than just
LLVM-IR produced by clang, as other front-ends could run into this even at -O0.
>>>> 
>>>> Due to this assertion failure and the infrastructure I used,
the numbers below do not include test-suite/MultiSource/Benchmarks results.
>>>> 
>>>> On the non-correctness aspects, LNT tells me that:
>>>> - The programs that report execution time, on geomean are about
17% slower.
>>>> - The programs that report scores, on geomean are about 21%
slower.
>>>> - Code size is up on geomean about 11%.
>>>> I'm afraid I don't have compile time numbers, nor any
feel for debug info quality.
>>>> 
>>>> I'll need quite a bit more time to dig into the details to
come up with something actionable, although the fact that LowLevelType
doesn't support vectors of pointers is already actionable.
>>>> Nevertheless, I thought to share what I see as is, to see if
others see similar results so far.
>>>> 
>>>> I thought Diana was going to look into fallback rate on the
test-suite on AArch64 linux?
>>>> 
>>>> Thanks,
>>>> 
>>>> Kristof
>>>> 
>>>>> On 30 Mar 2017, at 10:54, Renato Golin <renato.golin at
linaro.org <mailto:renato.golin at linaro.org>> wrote:
>>>>> 
>>>>> On 30 March 2017 at 00:27, Quentin Colombet <qcolombet
at apple.com <mailto:qcolombet at apple.com>> wrote:
>>>>>> On iOS we are at 100% pass rate in 00 g for the LLVM
test suite, standard
>>>>>> benchmarks and unit tests. In about 5% of all functions
GlobalIsel falls
>>>>>> back to SDIsel.
>>>>>> (Kristof Beyls would have the linux numbers.)
>>>>>> The self host compiler correctly builds and runs the
LLVM test suite in O0.
>>>>> 
>>>>> Having done no tests at all on my side, I think we need to
have
>>>>> similar numbers on Linux to be able to flip across the
board.
>>>>> 
>>>>> I don't want to flip it only for Darwin and not Linux,
as that will
>>>>> fragment the effort too much.
>>>>> 
>>>>> I'll check with Diana and Kristof to know what's
the best way forward,
>>>>> but it should be reasonably quick.
>>>>> 
>>>>> cheers,
>>>>> --renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170427/ef718887/attachment.html>