llvm dev - Nov 2013 - [LLVMdev] How to reduce the footprint of MDNodes? (About the comment you made at BOF LTO)

Hi Manman (and llvmdev),

I filed these two bugs to track the ideas that I was cooking:

http://llvm.org/bugs/show_bug.cgi?id=17891
http://llvm.org/bugs/show_bug.cgi?id=17892

TL;DR: I'm saying we should go from:

	!14 = metadata !{i32 786445, metadata !1, metadata !10, metadata
!"y", i32 3, i64 32, i64 32, i64 32, i32 0, metadata !13}
to:
	!14 = metadata !"v12,14,y,3,0,32,32,32"(metadata !1, metadata !13)

I think that this is along the lines of what you're thinking of, but
doesn't require doing specific optimizations for constant values (since most
of those would be encoded into strings instead of showing up as operands).

This is a pretty big project, but I think the payoff would be a huge constant
factor improvement in memory use.  We should probably also use a similar
approach for profile info.

-Chris


On Nov 11, 2013, at 4:09 PM, Manman Ren <mren at apple.com> wrote:
> 
> Hi Chris,
> 
> I missed the BOF LTO session and Bob mentioned to me that you had a comment
about how to reduce the footprint of MDNodes.

On Mon, Nov 11, 2013 at 11:29 PM, Chris Lattner <clattner at apple.com>
wrote:
> Hi Manman (and llvmdev),
>
> I filed these two bugs to track the ideas that I was cooking:
>
> http://llvm.org/bugs/show_bug.cgi?id=17891
> http://llvm.org/bugs/show_bug.cgi?id=17892
>
> TL;DR: I'm saying we should go from:
>
>         !14 = metadata !{i32 786445, metadata !1, metadata !10, metadata
!"y", i32 3, i64 32, i64 32, i64 32, i32 0, metadata !13}
> to:
>         !14 = metadata !"v12,14,y,3,0,32,32,32"(metadata !1,
metadata !13)
>
> I think that this is along the lines of what you're thinking of, but
doesn't require doing specific optimizations for constant values (since most
of those would be encoded into strings instead of showing up as operands).
>
> This is a pretty big project, but I think the payoff would be a huge
constant factor improvement in memory use.  We should probably also use a
similar approach for profile info.
>
We can take discussion of specific format into the bugs, but a more
compact encoding is definitely a good thing. Thanks for filing the
bugs to keep track.

-eric
> -Chris
>
>
> On Nov 11, 2013, at 4:09 PM, Manman Ren <mren at apple.com> wrote:
>
>>
>> Hi Chris,
>>
>> I missed the BOF LTO session and Bob mentioned to me that you had a
comment about how to reduce the footprint of MDNodes.
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On Mon, Nov 11, 2013 at 11:29 PM, Chris Lattner <clattner at apple.com>
wrote:
> Hi Manman (and llvmdev),
>
> I filed these two bugs to track the ideas that I was cooking:
>
> http://llvm.org/bugs/show_bug.cgi?id=17891
> http://llvm.org/bugs/show_bug.cgi?id=17892
>
> TL;DR: I'm saying we should go from:
>
>         !14 = metadata !{i32 786445, metadata !1, metadata !10, metadata
> !"y", i32 3, i64 32, i64 32, i64 32, i32 0, metadata !13}
> to:
>         !14 = metadata !"v12,14,y,3,0,32,32,32"(metadata !1,
metadata !13)
>
So, I like where you're going here, but a few tweaks.

First, there are two things going on here: removing an indirection through
a referenced metadata node and flattening N values into a string inclusion.
Removing the indirection seems obvious strict goodness, my comments are
about the second part.

I'm moderately opposed to just encoding these in a string format. I think
we can do something substantially better both for space, time, and
readability. Fundamentally, there is no reason for the original metadata
node you describe to not *encode* its operands into a dense bit-packed blob
of memory. We can still expose APIs that manipulate them as separate
entities, and have the AsmPrinter and AsmParser map back and forth with
nice human-readable forms. But even a simple varint encoding will be both
smaller and faster than ascii.

Just to be clear, I still want the nice format (much like your proposed
format, but maybe with the numbers outside of the "s) in the textual IR, I
just think we should use a more direct and efficient in-memory encoding
(and in-bitcode encoding if that isn't already suitably dense).
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131112/9d6086ce/attachment.html>

On Nov 12, 2013, at 1:28 PM, Chandler Carruth <chandlerc at google.com>
wrote:
> On Mon, Nov 11, 2013 at 11:29 PM, Chris Lattner <clattner at
apple.com> wrote:
> Hi Manman (and llvmdev),
> 
> I filed these two bugs to track the ideas that I was cooking:
> 
> http://llvm.org/bugs/show_bug.cgi?id=17891
> http://llvm.org/bugs/show_bug.cgi?id=17892
> 
> TL;DR: I'm saying we should go from:
> 
>         !14 = metadata !{i32 786445, metadata !1, metadata !10, metadata
!"y", i32 3, i64 32, i64 32, i64 32, i32 0, metadata !13}
> to:
>         !14 = metadata !"v12,14,y,3,0,32,32,32"(metadata !1,
metadata !13)
> 
> So, I like where you're going here, but a few tweaks.
> 
> First, there are two things going on here: removing an indirection through
a referenced metadata node and flattening N values into a string inclusion.
Removing the indirection seems obvious strict goodness, my comments are about
the second part.
David pointed out that the PR17892 is a microoptimization that may not even be
worthwhile.  I think that the flattening into a string is the important part.
> I'm moderately opposed to just encoding these in a string format. I
think we can do something substantially better both for space, time, and
readability. Fundamentally, there is no reason for the original metadata node
you describe to not *encode* its operands into a dense bit-packed blob of
memory. We can still expose APIs that manipulate them as separate entities, and
have the AsmPrinter and AsmParser map back and forth with nice human-readable
forms. But even a simple varint encoding will be both smaller and faster than
ascii.
I guess you could make it work, but would that actually be simpler than what is
proposed?  If it is denser, how much denser would it have to be to justify the
complexity?
> Just to be clear, I still want the nice format (much like your proposed
format, but maybe with the numbers outside of the "s) in the textual IR, I
just think we should use a more direct and efficient in-memory encoding (and
in-bitcode encoding if that isn't already suitably dense).
Where would the encoding schema be specified?

Note that there are simple things that can be done to make MDNodes more
efficient in common cases.  The CallbackVH is only necessary when pointing to
Value*’s that are not MDNode/MDString, and Constants-other-than-GlobalValue.  If
we make MDNode detect when it has “all-immortal” operands (like most debug info
nodes) then we could just store Value*’s directly.  This would be a completely
invisible implementation improvement, but would not provide the same level of
improvement as the “flatten into strings” approach.  The two are quite
complementary.

-Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131112/b94228a1/attachment.html>