llvm dev - Mar 2019 - [RFC] Making space for a flush-to-zero flag in FastMathFlags

Hi,

I need to add a flush-denormals-to-zero (FTZ) flag to FastMathFlags,
but  we've already used up the 7 bits available in
Value::SubclassOptionalData (the "backing storage" for
FPMathOperator::getFastMathFlags()).  These are the possibilities I
can think of:

1. Increase the size of FPMathOperator.  This gives us some additional
bits for FTZ and other fastmath flags we'd want to add in the future.
Obvious downside is that it increases LLVM's memory footprint.

2. Steal some low bits from pointers already present in Value and
expose them as part of SubclassOptionalData.  We can at least steal 3
bits from the first two words in Value which are both pointers.  The
LSB of the first pointer needs to be 0, otherwise we could steal 4
bits.

3. Allow only specific combinations in FastMathFlags.  In practice, I
don't think folks are equally interested in all the 2^N combinations
present in FastMathFlags, so we could compromise and allow only the
most "typical" 2^7 combinations (e.g. we could nonan and noinf into a
single bit, under the assumption that users want to enable-disable
them as a unit).  I'm unsure if establishing the most typical 2^7
combinations will be straightforward though.

4. Function level attributes.  Instead of wasting precious
instruction-level space, we could move all FP math attributes on the
containing function.  I'm not sure if this will work for all frontends
and it also raises annoying tradeoffs around inlining and other
inter-procedural passes.


My gut feeling is to go with (2).  It should be semantically
invisible, have no impact on memory usage, and the ugly bit
manipulation can be abstracted away.  What do you think?  Any other
possibilities I missed?


Why I need an FTZ flag:  some ARM Neon vector instructions have FTZ
semantics, which means we can't vectorize instructions when compiling
for Neon unless we know the user is okay with FTZ.  Today we pretend
that the "fast" variant of FastMathFlags implies FTZ
(https://reviews.llvm.org/rL266363), which is not ideal.  Moreover
(this is the immediate reason), for XLA CPU I'm trying to generate FP
instructions without nonan and noinf, which breaks vectorization on
ARM Neon for this reason.  An explicit bit for FTZ will let me
generate FP operations tagged with FTZ and all fast math flags except
nonan and noinf, and still have them vectorize on Neon.

-- Sanjoy

Can we move HasValueHandle out of the byte used for SubClassOptionalData
and move it to the flags at the bottom of value by
shrinking NumUserOperands to 27?

~Craig


On Sat, Mar 16, 2019 at 12:51 PM Sanjoy Das via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi,
>
> I need to add a flush-denormals-to-zero (FTZ) flag to FastMathFlags,
> but  we've already used up the 7 bits available in
> Value::SubclassOptionalData (the "backing storage" for
> FPMathOperator::getFastMathFlags()).  These are the possibilities I
> can think of:
>
> 1. Increase the size of FPMathOperator.  This gives us some additional
> bits for FTZ and other fastmath flags we'd want to add in the future.
> Obvious downside is that it increases LLVM's memory footprint.
>
> 2. Steal some low bits from pointers already present in Value and
> expose them as part of SubclassOptionalData.  We can at least steal 3
> bits from the first two words in Value which are both pointers.  The
> LSB of the first pointer needs to be 0, otherwise we could steal 4
> bits.
>
> 3. Allow only specific combinations in FastMathFlags.  In practice, I
> don't think folks are equally interested in all the 2^N combinations
> present in FastMathFlags, so we could compromise and allow only the
> most "typical" 2^7 combinations (e.g. we could nonan and noinf
into a
> single bit, under the assumption that users want to enable-disable
> them as a unit).  I'm unsure if establishing the most typical 2^7
> combinations will be straightforward though.
>
> 4. Function level attributes.  Instead of wasting precious
> instruction-level space, we could move all FP math attributes on the
> containing function.  I'm not sure if this will work for all frontends
> and it also raises annoying tradeoffs around inlining and other
> inter-procedural passes.
>
>
> My gut feeling is to go with (2).  It should be semantically
> invisible, have no impact on memory usage, and the ugly bit
> manipulation can be abstracted away.  What do you think?  Any other
> possibilities I missed?
>
>
> Why I need an FTZ flag:  some ARM Neon vector instructions have FTZ
> semantics, which means we can't vectorize instructions when compiling
> for Neon unless we know the user is okay with FTZ.  Today we pretend
> that the "fast" variant of FastMathFlags implies FTZ
> (https://reviews.llvm.org/rL266363), which is not ideal.  Moreover
> (this is the immediate reason), for XLA CPU I'm trying to generate FP
> instructions without nonan and noinf, which breaks vectorization on
> ARM Neon for this reason.  An explicit bit for FTZ will let me
> generate FP operations tagged with FTZ and all fast math flags except
> nonan and noinf, and still have them vectorize on Neon.
>
> -- Sanjoy
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190317/cc7f20af/attachment.html>

We knew the day when we needed another FMF bit was coming back in:
https://reviews.llvm.org/D39304
...it was just a question of 'when'. :)

I'm guessing that an FTZ bit won't be the last new bit needed if we
consider permutations between strict FP and fast-math. Even without that,
denormals-as-zero (DAZ) might also be useful?
So rather than continuing to carve these out bit-by-bit, it's worth
considering a more general solution: instruction-level metadata.

IIUC, the main argument for making FMF part of the instruction was that
per-instruction metadata gets expensive if we're applying it to a
significant chunk of the instructions.
But let's think about that - even the most FP-heavy code tops out around
10% FP math ops out of the total instruction count. Typical FP benchmark
code is only 2-5% FP ops. The rest is the same load/store/control-flow/ALU
stuff found in integer code.

I'm not exactly sure yet what it would take to do the experiment, but it
seems worth exploring moving the existing FMF to metadata.

One point in favor of this approach is that we already have an
"MD_fpmath"
enum. It's currently only used to convey reduced precision requirements to
the AMDGPU backend. We could extend that to include arbitrary FMF settings.

A couple of related points for FMF-as-metadata:
1. It might encourage fixing a hack added for reciprocals: we use a
function-level attribute for those (grep for "reciprocal-estimates").
IIRC,
that was just a quicker fix than using MD_fpmath. The existing squished
boolean FMF can't convey the more general settings that we need for
reciprocal optimizations.
2. These don't require new bits, but FMF isn't applied correctly today
as-is:
https://reviews.llvm.org/D48085
https://bugs.llvm.org/show_bug.cgi?id=38086
https://bugs.llvm.org/show_bug.cgi?id=39535
https://reviews.llvm.org/D51701
...so we need to make FMF changes regardless of FTZ.

On Sun, Mar 17, 2019 at 2:47 PM Craig Topper via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Can we move HasValueHandle out of the byte used for SubClassOptionalData
> and move it to the flags at the bottom of value by
> shrinking NumUserOperands to 27?
>
> ~Craig
>
>
> On Sat, Mar 16, 2019 at 12:51 PM Sanjoy Das via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Hi,
>>
>> I need to add a flush-denormals-to-zero (FTZ) flag to FastMathFlags,
>> but  we've already used up the 7 bits available in
>> Value::SubclassOptionalData (the "backing storage" for
>> FPMathOperator::getFastMathFlags()).  These are the possibilities I
>> can think of:
>>
>> 1. Increase the size of FPMathOperator.  This gives us some additional
>> bits for FTZ and other fastmath flags we'd want to add in the
future.
>> Obvious downside is that it increases LLVM's memory footprint.
>>
>> 2. Steal some low bits from pointers already present in Value and
>> expose them as part of SubclassOptionalData.  We can at least steal 3
>> bits from the first two words in Value which are both pointers.  The
>> LSB of the first pointer needs to be 0, otherwise we could steal 4
>> bits.
>>
>> 3. Allow only specific combinations in FastMathFlags.  In practice, I
>> don't think folks are equally interested in all the 2^N
combinations
>> present in FastMathFlags, so we could compromise and allow only the
>> most "typical" 2^7 combinations (e.g. we could nonan and
noinf into a
>> single bit, under the assumption that users want to enable-disable
>> them as a unit).  I'm unsure if establishing the most typical 2^7
>> combinations will be straightforward though.
>>
>> 4. Function level attributes.  Instead of wasting precious
>> instruction-level space, we could move all FP math attributes on the
>> containing function.  I'm not sure if this will work for all
frontends
>> and it also raises annoying tradeoffs around inlining and other
>> inter-procedural passes.
>>
>>
>> My gut feeling is to go with (2).  It should be semantically
>> invisible, have no impact on memory usage, and the ugly bit
>> manipulation can be abstracted away.  What do you think?  Any other
>> possibilities I missed?
>>
>>
>> Why I need an FTZ flag:  some ARM Neon vector instructions have FTZ
>> semantics, which means we can't vectorize instructions when
compiling
>> for Neon unless we know the user is okay with FTZ.  Today we pretend
>> that the "fast" variant of FastMathFlags implies FTZ
>> (https://reviews.llvm.org/rL266363), which is not ideal.  Moreover
>> (this is the immediate reason), for XLA CPU I'm trying to generate
FP
>> instructions without nonan and noinf, which breaks vectorization on
>> ARM Neon for this reason.  An explicit bit for FTZ will let me
>> generate FP operations tagged with FTZ and all fast math flags except
>> nonan and noinf, and still have them vectorize on Neon.
>>
>> -- Sanjoy
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190318/4952e6df/attachment.html>

On Sun, Mar 17, 2019 at 1:47 PM Craig Topper <craig.topper at gmail.com>
wrote:
> Can we move HasValueHandle out of the byte used for SubClassOptionalData
and move it to the flags at the bottom of value by shrinking NumUserOperands to
27?
I like this approach because it is less work for me. :)

But I agree with Sanjay below that this only kicks the can slightly
further down the road (solutions (2) and (3) also have the same
problem).  Let's see if we can agree on a more future proof solution.

-- Sanjoy
>
> ~Craig
>
>
> On Sat, Mar 16, 2019 at 12:51 PM Sanjoy Das via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
>>
>> Hi,
>>
>> I need to add a flush-denormals-to-zero (FTZ) flag to FastMathFlags,
>> but  we've already used up the 7 bits available in
>> Value::SubclassOptionalData (the "backing storage" for
>> FPMathOperator::getFastMathFlags()).  These are the possibilities I
>> can think of:
>>
>> 1. Increase the size of FPMathOperator.  This gives us some additional
>> bits for FTZ and other fastmath flags we'd want to add in the
future.
>> Obvious downside is that it increases LLVM's memory footprint.
>>
>> 2. Steal some low bits from pointers already present in Value and
>> expose them as part of SubclassOptionalData.  We can at least steal 3
>> bits from the first two words in Value which are both pointers.  The
>> LSB of the first pointer needs to be 0, otherwise we could steal 4
>> bits.
>>
>> 3. Allow only specific combinations in FastMathFlags.  In practice, I
>> don't think folks are equally interested in all the 2^N
combinations
>> present in FastMathFlags, so we could compromise and allow only the
>> most "typical" 2^7 combinations (e.g. we could nonan and
noinf into a
>> single bit, under the assumption that users want to enable-disable
>> them as a unit).  I'm unsure if establishing the most typical 2^7
>> combinations will be straightforward though.
>>
>> 4. Function level attributes.  Instead of wasting precious
>> instruction-level space, we could move all FP math attributes on the
>> containing function.  I'm not sure if this will work for all
frontends
>> and it also raises annoying tradeoffs around inlining and other
>> inter-procedural passes.
>>
>>
>> My gut feeling is to go with (2).  It should be semantically
>> invisible, have no impact on memory usage, and the ugly bit
>> manipulation can be abstracted away.  What do you think?  Any other
>> possibilities I missed?
>>
>>
>> Why I need an FTZ flag:  some ARM Neon vector instructions have FTZ
>> semantics, which means we can't vectorize instructions when
compiling
>> for Neon unless we know the user is okay with FTZ.  Today we pretend
>> that the "fast" variant of FastMathFlags implies FTZ
>> (https://reviews.llvm.org/rL266363), which is not ideal.  Moreover
>> (this is the immediate reason), for XLA CPU I'm trying to generate
FP
>> instructions without nonan and noinf, which breaks vectorization on
>> ARM Neon for this reason.  An explicit bit for FTZ will let me
>> generate FP operations tagged with FTZ and all fast math flags except
>> nonan and noinf, and still have them vectorize on Neon.
>>
>> -- Sanjoy
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev