<div dir="ltr"><div dir="ltr"><div>Hi Alexey,</div><div><br></div><div>I should've looked at this earlier. I went through the thread again and I've</div><div>made some comments, mostly from the dsymutil point of view.</div><div><br></div><div>> Current DWARFEmitter/DWARFStreamer has an implementation for DWARF</div><div>> generation, which does not support DWARF5(only debug_names table). At the</div><div>> same time, there already exists code in CodeGen/AsmPrinter/DwarfDebug.h,</div><div>> which implements most of DWARF5. It seems that DWARFEmitter/DWARFStreamer</div><div>> should be rewritten using DwarfDebug/DwarfFile. Though I am not sure</div><div>> whether it would be easy to re-use DwarfDebug/DwarfFile. It would probably</div><div>> be necessary to separate some intermediate level of DwarfDebug/DwarfFile.</div><div><br></div><div>These classes serve very different purposes. Last time I looked at them there</div><div>was very little overlap in functionality. In the compiler we're mostly</div><div>concerned with generating the DWARF, while in dsymutil we try to copy</div><div>everything we don't need to parse, and fix up what we have to. I don't want</div><div>to say it's not possible, but I think supporting DWARF5 in those classes is</div><div>going to be a lot less work than trying to reuse the CodeGen variants.</div><div><br></div><div>> Measurements show that it is spent ~10 sec in</div><div>> llvm::StringMapImpl::LookupBucketFor(). The problem is that the same</div><div>> strings, again and again, are added to the string pool. Two attributes</div><div>> having the same string value would be analyzed (hash calculated) and</div><div>> searched inside the string pool. Even if these strings are already in</div><div>> string table(DW_FORM_strp, DW_FORM_strx). The process could be optimized</div><div>> for string tables. So that if some string from the string table were</div><div>> accessed previously then, it would keep a reference into the string pool.</div><div>> This would eliminate a lot of string pool searches.</div><div><br></div><div>I'm not sure I fully understand the optimization, but I'd love to speed this</div><div>up, if only for dsymutil's sake. I'd love to talk about this in a separate</div><div>thread or offline.</div><div><br></div><div>> Currently, all object files are analyzed sequentially and cloned</div><div>> sequentially. Cloning is started in parallel with analyzing. That scheme</div><div>> could be changed: analyzing and cloning could be done in parallel for each</div><div>> object file. That requires refactoring of DWARFLinker and making string</div><div>> pools and DeclContextTree thread-safe.</div><div><br></div><div>I'm less familiar with the way that LLD uses the DWARFOptimizer but this is</div><div>not possible for dsymutil as it is trying to deduplicate DIEs from different</div><div>compile units.</div><div><br></div><div>> I think improving dsymutil is a valuable thing. Though there are several</div><div>> directions which might be considered to make it more robust:</div><div>></div><div>> 1. support of latest DWARF - DWARF5/DWARF64...</div><div><br></div><div>Strong +1 on DWARF5. I haven't had the bandwidth yet to really look at this.</div><div>Right now we can't find (at least some) rellocations so we bail out. I'd need</div><div>to fix that to assess the current state of things and figure out how much</div><div>work would be needed.</div><div><br></div><div>I don't think anything in LLVM supports generating DWARF64 though.</div><div><br></div><div>> 2. implement multi-threaded execution.</div><div><br></div><div>See my earlier comment. At least for the dsymutil case, the current approach</div><div>is the best we can do, but I'd love to be proven wrong. :-)</div><div><br></div><div>> 3. support of split DWARF.</div><div>> 4. implement dsymutil for non-darwin platform.</div><div><br></div><div>These two seem to go together. Given the work you did to split off the DWARF</div><div>optimization part I think we're closer to this than ever. Thanks again for</div><div>doing that.</div><div><br></div><div>> We considered three options:</div><div>></div><div>> 1. add new functionality into dsymutil. So that dsymutil behaves</div><div>> differently on a non-darwin platform and supports another set of</div><div>> command-line options.</div><div>></div><div>> 2. add new functionality into llvm-objcopy. llvm-objcopy already supports</div><div>> various binary objects formats(MachO,ELF,COFF,wasm). It also has several</div><div>> options to work with debug-info.</div><div>></div><div>> 3. create new utility llvm-dwarfutil which would implement the above</div><div>> functionality and reuse DWARFLinker(extracted from dsymutil) library and</div><div>> new library ObjectCopy(extracted from llvm-objcopy).</div><div>></div><div>> So far our preference is number three. The reason for this is that separate</div><div>> utility specifically working with debug info looks as good separation of</div><div>> concepts. Adding another behavior to dsymutil looks not very good.</div><div><br></div><div>In its current state dsymutil itself is a pretty small tool on top of the</div><div>DWARFOptimizer/Linker. I'm curious what the benefits of another tool are</div><div>compared to a different frontend (like objcopy) for MachO and ELF. It seems</div><div>like that would allow for separation of concerns, while still being able to</div><div>share common code without having to push it all the way up into LLVM.</div><div><br></div><div>> Extending the already rich interface of llvm-objcopy looks also not very</div><div>> good. Having in mind that actual implementation would be shared by</div><div>> libraries, the separate utility, working specifically with debug info,</div><div>> looks like the right choice. That is our current idea.</div><div><br></div><div>> My personal thought would be that extending dsymutil should be ok as the</div><div>> functionality goes well with everything else dsymutil does (other than not</div><div>> support ELF which the dsymutil maintainers are on board with last I</div><div>> checked). That said, I definitely think a write-up will be helpful. No</div><div>> matter what I support extracting all of the behavior into libraries and</div><div>> using that somewhere :)</div><div><br></div><div>Ha, so basically what I was trying to say above.</div><div><br></div><div>I look forward to seeing the proposal!</div><div><br></div><div>Cheers,</div><div>Jonas</div><div><br></div></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Tue, Aug 4, 2020 at 11:33 PM Eric Christopher <<a href="mailto:echristo@gmail.com">echristo@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr">Hi Alexey,<div><br></div><div><br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Aug 3, 2020 at 8:32 AM Alexey Lapshin <<a href="mailto:avl.lapshin@gmail.com" target="_blank">avl.lapshin@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex">
<div>
<p>Hi Eric, please <br>
</p>
<div>On 31.07.2020 22:02, Eric Christopher
wrote:<br>
</div>
<blockquote type="cite">
<div dir="ltr">
<div>Hi Alexey,</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Fri, Jul 31, 2020 at 4:02
AM Alexey Lapshin via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><br>
On 28.07.2020 19:28, David Blaikie wrote:<br>
> On Tue, Jul 28, 2020 at 8:55 AM Alexey Lapshin <<a href="mailto:avl.lapshin@gmail.com" target="_blank">avl.lapshin@gmail.com</a>>
wrote:<br>
>><br>
>> On 28.07.2020 10:29, David Blaikie via llvm-dev
wrote:<br>
>>> On Fri, Jun 26, 2020 at 9:28 AM Alexey Lapshin<br>
>>> <<a href="mailto:alapshin@accesssoftek.com" target="_blank">alapshin@accesssoftek.com</a>>
wrote:<br>
>>>>>>>>>>>> This idea
goes in another direction than fragmenting dwarf<br>
>>>>>>>>>>>> using elf
sections&tricks. It seems to me that the cost of
fragmenting is too high.<br>
>>>>>>>>>>> I tend to agree
- but I'm sort of leaning towards trying to use object<br>
>>>>>>>>>>> features as
much as possible, then implementing just enough custom<br>
>>>>>>>>>>> handling in the
linker to recoup overhead, etc. (eg: add some kind of<br>
>>>>>>>>>>> small
header/brief description that makes it easy for the linker
to<br>
>>>>>>>>>>> slice-and-dice
- but hopefully a domain-specific such header can be a<br>
>>>>>>>>>>> bit more
compact than the fully general ELF form)<br>
>>>>>>>>>> I think this indeed
should be implemented and evaluated.<br>
>>>>>>>>>> So that various
approaches could be compared.<br>
>>>>>>>>>><br>
>>>>>>>>>>>> It is not
only the sizes of structures describing fragments but also
the complexity<br>
>>>>>>>>>>>> of tools
that should be taught to work with fragmented DWARF.<br>
>>>>>>>>>>>> (f.e.
llvm-dwarfdump applied to object file should be able to read
fragmented DWARF,<br>
>>>>>>>>>>>> but applied
to linked executable it should work with non-fragmented
DWARF).<br>
>>>>>>>>>>>> That idea
is for the tool which works the same way as dsymutil ODR.<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> I will
shortly describe the idea of making DWARF be easier
processed by dsymutil/DWARFLinker:<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> The idea is
to have only one "type table" per object file(special
section .debug_types_table).<br>
>>>>>>>>>>>> This "type
table" would contain all types.<br>
>>>>>>>>>>>> There could
be a special type of reference - type_offset - that offset
points into the type table.<br>
>>>>>>>>>>>> Basic types
could always be placed into the start of "type table" thus,
offsets to basic types<br>
>>>>>>>>>>>> most often
would be 1 byte. There also would be a special kind of
reference - reference inside the type.<br>
>>>>>>>>>>>> Type units
sig8 system - would not be used to reference types.<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> Types
deduplication is assumed to be done, not by linker mechanism
for COMDAT,<br>
>>>>>>>>>>>> but by a
tool like dsymutil. This tool would create resulting
.debug_types_table by putting there<br>
>>>>>>>>>>>> types from
source .debug_types_table-s. Only one copy of the type would
be placed into the<br>
>>>>>>>>>>>> resulting
table. All references pointing to the deleted copy would be
corrected to point<br>
>>>>>>>>>>>> to the
single copy inside "type table". (that is how dsymutil works
currently)<br>
>>>>>>>>>>> ^ that's the
step that's probably a bit expensive for a general-use<br>
>>>>>>>>>>> tool - it
implies parsing all the DWARF to find those references and<br>
>>>>>>>>>>> rewrite them, I
think. For a high-performance solution that could be<br>
>>>>>>>>>>> run by the
linker I think it'd be necessary to have a solution that<br>
>>>>>>>>>>> doesn't involve
parsing all the DIEs.<br>
>>>>>>>>>> According to the
current dsymutil processing,<br>
>>>>>>>>>> exactly this
process is not the most time-consuming.<br>
>>>>>>>>>> That could be done
relatively fast.<br>
>>>>>>>>> Fair enough - though
I'd still imagine any solution that involves<br>
>>>>>>>>> parsing all the DIEs
still wouldn't be fast enough (maybe an order of<br>
>>>>>>>>> magnitude faster than
the current solution even - but that's stuill,<br>
>>>>>>>>> what, 6 or 7x slower
than linking without the feature?) for most users<br>
>>>>>>>>> to consider it a good
trade-off.<br>
>>>>>>>> It seems to me that even
the current 6x-7x slowdown could be useful.<br>
>>>>>>>> Users who already use
dsymutil or llvm-dwp(assuming DWARFLinker<br>
>>>>>>>> would be taught to work
with a split dwarf) tools spend this time and,<br>
>>>>>>>> in some scenarios, waste
disk space by inter-mediate files.<br>
>>>>>>> FWIW, dwp (llvm-dwp hasn't
really been optimized compared to binutils<br>
>>>>>>> dwp) is designed to be very
quick - by not needing to do a lot of<br>
>>>>>>> parsing/fixups. Which, yes,
means larger output files than would be<br>
>>>>>>> possible with more parsing/etc.
It also doesn't take any input from<br>
>>>>>>> the linker (so it can run in
parallel with the linker) - so it can't<br>
>>>>>>> remove dead subprograms. Given
Google's the major (perhaps only<br>
>>>>>>> significant?) user of Split
DWARF - I can say that the needs don't<br>
>>>>>>> necessarily overlap well with
something that would take significantly<br>
>>>>>>> longer to run or use
significantly more memory. Faster/cheaper/with<br>
>>>>>>> somewhat bigger output files is
probably the right tradeoff for<br>
>>>>>>> Google's use case, at least.<br>
>>>>>>><br>
>>>>>>> I imagine Apple's use for
dsymutil is somewhat similar - it's not used<br>
>>>>>>> in the iterative development
cycle, only in final releases - well,<br>
>>>>>>> maybe their situation is more
"neutral" - not a major pain point in<br>
>>>>>>> any case I'd guess.<br>
>>>>>>><br>
>>>>>>><br>
>>>>>> I see. FWIW, Comparison
splitdwarf+dwp and DWARFLinker from lld:<br>
>>>>>><br>
>>>>>> 1. split-dwarf+llvm-dwp = linking
time for clang 6 sec,<br>
>>>>>> generating time for .dwp 53
sec, clang=997M clang.dwp=1.1G.<br>
>>>>> FWIW, llvm-dwp is not very well
optimized (which is to say: it is not<br>
>>>>> optimized), binutils dwp might be a
better comparison (& even that<br>
>>>>> doesn't have the parallelism & some
potential further memory savings<br>
>>>>> that lld has that we could take
advantage of in a dwp-like tool)<br>
>>>>><br>
>>>>> What build mode was the clang binary
built in? Optimized or unoptimized?<br>
>>>> right, that is unoptimized build with
-ffunction-sections.<br>
>>>><br>
>>>>>> 2. DWARFLinker from lld = linking
time for clang 72 sec, clang=760M.<br>
>>> And this is without Split DWARF? Without linker
DWARF compression? -<br>
>>> that seems quite a bit surprising, that the
deduplication of DWARF<br>
>>> could fit into less space than the
wasted/reclaimed space in ranges (&<br>
>>> line)?<br>
>> that was without split dwarf, without linker
compression.<br>
>><br>
>>> Could you double check these numbers &
provide a clearer summary?<br>
>> sure, I would re-check it.<br>
>><br>
>>> Here's my attempt at numbers (all with
function-sections+gc-sections)...<br>
>>><br>
>>> Split DWARF tests didn't seem meaningful -
gc-debuginfo + split DWARF<br>
>>> seemed to drop all the debug info (except
gdb_index) so wasn't<br>
>>> working/comparison wasn't meaningful for Apples
to Apples, but<br>
>>> included it for comparing gc'd non-split to
non-gc'd split (disabled<br>
>>> gnu-pubnames/gdb-index (-gsplit-dwarf
-gno-gnu-pubnames) (which turns<br>
>>> on by default with Split DWARF because gdb
needs it - but a bit of an<br>
>>> unfair comparison without turning on
gnu-pubnames/gdb-index in other<br>
>>> build modes too, since it... /shouldn't/ be
necessary) which might've<br>
>>> been a factor in the data you were looking at)<br>
>> that might be the case. i.e. clang=997M for split
dwarf(from my previous<br>
>> measurement) might include gnu-pubnames.<br>
>><br>
>> would recheck it and if that is the case then it is
a unfair comparison.<br>
>><br>
>><br>
>> My point was that "DWARFLinker from lld" takes less
space than singleton<br>
>> split dwarf file+.dwp file.<br>
>><br>
>> for -O0 uncompressed:<br>
>><br>
>> - .dwp took 1.1G(if I built it correctly),
singleton clang(from your<br>
>> measurements) 566 MB<br>
>><br>
>> overall 1.6G.<br>
> Oh, yeah, even if there are some measurement issues,
linked executable<br>
> + .dwp is going to be larger than a linked executable
using non-split<br>
> DWARF (in v5), since v5 uses all the same
representations as non-split<br>
> DWARF, and split DWARF adds the indirection overhead of
a split file,<br>
> etc.<br>
><br>
> Even without DWARF linking, it's true that split DWARF
has overhead<br>
> (dwp+executable will be larger than executable
non-split).<br>
><br>
> But maybe we've ended up down a bit of a tangent in any
case.<br>
><br>
> Trying to bring this back to "should this be committed
to lld" seems<br>
> valuable, and I'm not sure what the right criteria are
for that.<br>
I think it would be useful to do "removing obsolete debug
info"<br>
in the linker. First thing is that it would be the fastest
way(no need<br>
to copy data/create temp files/built address map...) Second
thing<br>
is that it would be a good separation of concepts. All debug
info<br>
processing, currently done in the linker(gdb_index, upcoming<br>
debug_names), could be moved into separate library
processing<br>
debug info. When gdb_index/debug_names should be built
without<br>
"removing of obsolete debug info" it would have the same<br>
performance results as it currently has.<br>
<br>
We decided to give the idea of "removing of obsolete debug
info"<br>
another try and are going to implement it as a separate
utility<br>
working with built binary. Making it to be multi-thread
would<br>
probably show better performance results and then it could<br>
probably be considered as acceptable to use from the linker.<br>
<br>
</blockquote>
<div><br>
</div>
<div>I'm quite interested in this direction. One thought I had
was to incorporate such a library into dsymutil but with
support for ELF. If you get a proposal written up I'd love
to take a look and comment.</div>
<div><br>
</div>
</div>
</div>
</blockquote>
<p><br>
</p>
yes, I would share the proposal in a separate thread within a week
or two.<br>
<br></div></blockquote><div><br></div><div>Excellent, thanks :)</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div>
Shortly: we decided to move in slightly other direction than adding
this functionality <br>
into dsymutil. Though if there is a preference to implement it as
part of dsymutil <br>
we are OK to do this way.<br>
<br></div></blockquote><div><br></div><div>I have a vague preference since a lot of functionality already exists there on one platform and extending that seems straight forward, however...</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div>
In its first version, this new utility supposed to receive built
binary with debug info <br>
as input(with the new marking for references to removed code
sections -1/-2 <br>
-<a href="https://reviews.llvm.org/D84825" target="_blank">https://reviews.llvm.org/D84825</a>) and create a new binary with
removed obsolete <br>
debug info according to the above marking. In the next versions, it
could be extended <br>
with other debug info optimizations tasks. F.e. generation new index
tables, debug info <br>
optimizing... etc...<br>
<br>
We considered three options:<br>
<br>
1. add new functionality into dsymutil. So that dsymutil behaves
differently <br>
on a non-darwin platform and supports another set of
command-line options.<br>
<br>
2. add new functionality into llvm-objcopy. llvm-objcopy already
supports various <br>
binary objects formats(MachO,ELF,COFF,wasm). It also has
several options <br>
to work with debug-info.<br>
<br>
3. create new utility llvm-dwarfutil which would implement the above
functionality <br>
and reuse DWARFLinker(extracted from dsymutil) library and new
library <br>
ObjectCopy(extracted from llvm-objcopy).<br>
<br>
So far our preference is number three. The reason for this is that
separate <br>
utility specifically working with debug info looks as good
separation of concepts. <br>
Adding another behavior to dsymutil looks not very good. Extending
the already <br>
rich interface of llvm-objcopy looks also not very good. Having in
mind that actual <br>
implementation would be shared by libraries, the separate utility,
working specifically <br>
with debug info, looks like the right choice. That is our current
idea. <br>
<p>I would publish the proposal shortly to discuss it.<br>
</p>
<br></div></blockquote><div><br></div><div>These are solid arguments - in particular, I agree with not extending llvm-objcopy :)</div><div><br></div><div><a class="gmail_plusreply" id="gmail-m_144640436407649066plusReplyChip-0" href="mailto:jonas@devlieghere.com" target="_blank">+Jonas Devlieghere</a> and <a class="gmail_plusreply" id="gmail-m_144640436407649066plusReplyChip-1" href="mailto:aprantl@apple.com" target="_blank">+Adrian Prantl</a> for dsymutil comments.</div><div><br></div><div>My personal thought would be that extending dsymutil should be ok as the functionality goes well with everything else dsymutil does (other than not support ELF which the dsymutil maintainers are on board with last I checked). That said, I definitely think a write-up will be helpful. No matter what I support extracting all of the behavior into libraries and using that somewhere :)</div><div><br></div><div>Thanks!</div><div><br></div><div>-eric</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div>
Thank you, Alexey.<br>
<blockquote type="cite">
<div dir="ltr">
<div class="gmail_quote">
<div>Thanks!</div>
<div><br>
</div>
<div>-eric</div>
<div> </div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex">
Alexey.<br>
<br>
><br>
> Ray's the best person to weigh in on that. My 2c is
that I think it<br>
> probably is worthwhile, even just as an experiment,
assuming it's not<br>
> too intrusive to lld.<br>
><br>
>> - The "DWARFLinker from lld" 820 MB(from your
measurements).<br>
>><br>
>><br>
>> So "DWARFLinker from lld" looks two times better.<br>
>><br>
>><br>
>> Anyway, thank you for pointing me to possible
mistake. I would recheck<br>
>> it and update results.<br>
>><br>
>><br>
>> Alexey.<br>
>><br>
>><br>
>>> * -O0: (baseline, just using strip -g: 356 MB)<br>
>>> * compressed: 25% smaller with gc-debuginfo
(481 MB / 641 MB) (407<br>
>>> MB split/non-gc)<br>
>>> * uncompressed: 30% smaller (820 MB / 1.2
GB) (566 MB split/non-gc)<br>
>>> * -O3: (baseline: 116 MB)<br>
>>> * compressed: 16% smaller (361 MB / 462 MB)
(283 MB split/non-gc)<br>
>>> * uncompressed: 22% smaller (1022 MB / 1.2
GB) (156 MB split/non-gc)<br>
>>><br>
>>><br>
>>><br>
>>><br>
>>> On Fri, Jun 26, 2020 at 9:28 AM Alexey Lapshin<br>
>>> <<a href="mailto:alapshin@accesssoftek.com" target="_blank">alapshin@accesssoftek.com</a>>
wrote:<br>
>>>>>>>>>>>> This idea
goes in another direction than fragmenting dwarf<br>
>>>>>>>>>>>> using elf
sections&tricks. It seems to me that the cost of
fragmenting is too high.<br>
>>>>>>>>>>> I tend to agree
- but I'm sort of leaning towards trying to use object<br>
>>>>>>>>>>> features as
much as possible, then implementing just enough custom<br>
>>>>>>>>>>> handling in the
linker to recoup overhead, etc. (eg: add some kind of<br>
>>>>>>>>>>> small
header/brief description that makes it easy for the linker
to<br>
>>>>>>>>>>> slice-and-dice
- but hopefully a domain-specific such header can be a<br>
>>>>>>>>>>> bit more
compact than the fully general ELF form)<br>
>>>>>>>>>> I think this indeed
should be implemented and evaluated.<br>
>>>>>>>>>> So that various
approaches could be compared.<br>
>>>>>>>>>><br>
>>>>>>>>>>>> It is not
only the sizes of structures describing fragments but also
the complexity<br>
>>>>>>>>>>>> of tools
that should be taught to work with fragmented DWARF.<br>
>>>>>>>>>>>> (f.e.
llvm-dwarfdump applied to object file should be able to read
fragmented DWARF,<br>
>>>>>>>>>>>> but applied
to linked executable it should work with non-fragmented
DWARF).<br>
>>>>>>>>>>>> That idea
is for the tool which works the same way as dsymutil ODR.<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> I will
shortly describe the idea of making DWARF be easier
processed by dsymutil/DWARFLinker:<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> The idea is
to have only one "type table" per object file(special
section .debug_types_table).<br>
>>>>>>>>>>>> This "type
table" would contain all types.<br>
>>>>>>>>>>>> There could
be a special type of reference - type_offset - that offset
points into the type table.<br>
>>>>>>>>>>>> Basic types
could always be placed into the start of "type table" thus,
offsets to basic types<br>
>>>>>>>>>>>> most often
would be 1 byte. There also would be a special kind of
reference - reference inside the type.<br>
>>>>>>>>>>>> Type units
sig8 system - would not be used to reference types.<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> Types
deduplication is assumed to be done, not by linker mechanism
for COMDAT,<br>
>>>>>>>>>>>> but by a
tool like dsymutil. This tool would create resulting
.debug_types_table by putting there<br>
>>>>>>>>>>>> types from
source .debug_types_table-s. Only one copy of the type would
be placed into the<br>
>>>>>>>>>>>> resulting
table. All references pointing to the deleted copy would be
corrected to point<br>
>>>>>>>>>>>> to the
single copy inside "type table". (that is how dsymutil works
currently)<br>
>>>>>>>>>>> ^ that's the
step that's probably a bit expensive for a general-use<br>
>>>>>>>>>>> tool - it
implies parsing all the DWARF to find those references and<br>
>>>>>>>>>>> rewrite them, I
think. For a high-performance solution that could be<br>
>>>>>>>>>>> run by the
linker I think it'd be necessary to have a solution that<br>
>>>>>>>>>>> doesn't involve
parsing all the DIEs.<br>
>>>>>>>>>> According to the
current dsymutil processing,<br>
>>>>>>>>>> exactly this
process is not the most time-consuming.<br>
>>>>>>>>>> That could be done
relatively fast.<br>
>>>>>>>>> Fair enough - though
I'd still imagine any solution that involves<br>
>>>>>>>>> parsing all the DIEs
still wouldn't be fast enough (maybe an order of<br>
>>>>>>>>> magnitude faster than
the current solution even - but that's stuill,<br>
>>>>>>>>> what, 6 or 7x slower
than linking without the feature?) for most users<br>
>>>>>>>>> to consider it a good
trade-off.<br>
>>>>>>>> It seems to me that even
the current 6x-7x slowdown could be useful.<br>
>>>>>>>> Users who already use
dsymutil or llvm-dwp(assuming DWARFLinker<br>
>>>>>>>> would be taught to work
with a split dwarf) tools spend this time and,<br>
>>>>>>>> in some scenarios, waste
disk space by inter-mediate files.<br>
>>>>>>> FWIW, dwp (llvm-dwp hasn't
really been optimized compared to binutils<br>
>>>>>>> dwp) is designed to be very
quick - by not needing to do a lot of<br>
>>>>>>> parsing/fixups. Which, yes,
means larger output files than would be<br>
>>>>>>> possible with more parsing/etc.
It also doesn't take any input from<br>
>>>>>>> the linker (so it can run in
parallel with the linker) - so it can't<br>
>>>>>>> remove dead subprograms. Given
Google's the major (perhaps only<br>
>>>>>>> significant?) user of Split
DWARF - I can say that the needs don't<br>
>>>>>>> necessarily overlap well with
something that would take significantly<br>
>>>>>>> longer to run or use
significantly more memory. Faster/cheaper/with<br>
>>>>>>> somewhat bigger output files is
probably the right tradeoff for<br>
>>>>>>> Google's use case, at least.<br>
>>>>>>><br>
>>>>>>> I imagine Apple's use for
dsymutil is somewhat similar - it's not used<br>
>>>>>>> in the iterative development
cycle, only in final releases - well,<br>
>>>>>>> maybe their situation is more
"neutral" - not a major pain point in<br>
>>>>>>> any case I'd guess.<br>
>>>>>>><br>
>>>>>>><br>
>>>>>> I see. FWIW, Comparison
splitdwarf+dwp and DWARFLinker from lld:<br>
>>>>>><br>
>>>>>> 1. split-dwarf+llvm-dwp = linking
time for clang 6 sec,<br>
>>>>>> generating time for .dwp 53
sec, clang=997M clang.dwp=1.1G.<br>
>>>>> FWIW, llvm-dwp is not very well
optimized (which is to say: it is not<br>
>>>>> optimized), binutils dwp might be a
better comparison (& even that<br>
>>>>> doesn't have the parallelism & some
potential further memory savings<br>
>>>>> that lld has that we could take
advantage of in a dwp-like tool)<br>
>>>>><br>
>>>>> What build mode was the clang binary
built in? Optimized or unoptimized?<br>
>>>> right, that is unoptimized build with
-ffunction-sections.<br>
>>>><br>
>>>>>> 2. DWARFLinker from lld = linking
time for clang 72 sec, clang=760M.<br>
>>>>> It does seem a tad strange that the
clang binary would be smaller<br>
>>>>> non-split with DWARF linking than it
was split. Though I could imagine<br>
>>>>> this might be possible in an optimized
build (wehre debug_ranges<br>
>>>>> become quite relatively expensive in
the .o file contribution with<br>
>>>>> Split DWARF)<br>
>>>>> Could you compare the section sizes
between these two clang binaries, perhaps?<br>
>>>> .debug_ranges is three times bigger and
.debug_line is twice bigger.<br>
>>>><br>
>>>>>>>> Thus if they would use this
LLD feature in its current state<br>
>>>>>>>> - they would still receive
benefits.<br>
>>>>>>>><br>
>>>>>>>> Speaking of performance
results - LLD is a multi-thread linker;<br>
>>>>>>>> it handles sections in
parallel. DWARFLinker generates DWARF using<br>
>>>>>>>> AsmPrinter which is a
stream - so it could make resulting DWARF only<br>
>>>>>>>> continuously. It is not
surprising that the parallel solution works faster.<br>
>>>>>>>> Making DWARFLinker truly
multi-threaded would probably allow us<br>
>>>>>>>> to make slowdown to be at
2x-4x range.<br>
>>>>>>> *nod* that's still a really
expensive link - but I understand that's a<br>
>>>>>>> suitable tradeoff for your
users<br>
>>>>>>><br>
>>>>>> Btw, 2x or 7x is for pure linking
time. Overall compilation slowdown<br>
>>>>>> is not so significant. Building
LLVM codebase has only 20% slowdown.<br>
>>>>> Understood - that's still quite
significant to most users, I'd imagine.<br>
>>>> I see.<br>
>>>><br>
>>>>>>>>>> Anyway, I think the
dsymutil approach is still valuable, and it<br>
>>>>>>>>>> would be useful to
optimize it.<br>
>>>>>>>>>> Do you think it
would be useful to make dsymutil/DWARFLinker truly
multi-thread?<br>
>>>>>>>>>> (To make
dsymutil/DWARFLinker able to process each object file in a
separate thread)<br>
>>>>>>>>> Perhaps - that I'd
probably leave up to the folks who are more<br>
>>>>>>>>> invested in dsymutil
(Adrian Prantl et al). Maybe one day we'll get it<br>
>>>>>>>>> integrated into
llvm-dwp and then I'll be interested in getting as<br>
>>>>>>>>> much performance out of
it as lld - so multithreading and things would<br>
>>>>>>>>> be on the books.<br>
>>>>>>>> I think improving dsymutil
is a valuable thing.<br>
>>>>>>>> Though there are several
directions which might be considered<br>
>>>>>>>> to make it more robust:<br>
>>>>>>>><br>
>>>>>>>> 1. support of latest DWARF
- DWARF5/DWARF64...<br>
>>>>>>> I expect/though some of the
Apple folks had already worked on DWARF5 support?<br>
>>>>>>> DWARF64 - that's been around
for a while, and just hasn't been needed<br>
>>>>>>> by LLVM users thus far, it
seems (until recently - where some<br>
>>>>>>> developers have started working
on that)<br>
>>>>>> There already implemented
debug_names table, but debug_rnglists,<br>
>>>>>> debug_loclists, type units - are
not implemented yet.<br>
>>>>> Superficially, type units wouldn't be
on the list of features (like<br>
>>>>> DWARF64 - it's optional) I'd try to
support in dsymutil - since their<br>
>>>>> size overhead is more justified for a
DWARF-agnostic linker that's<br>
>>>>> using comdat groups. With a DWARF-aware
linker I'd be specifically<br>
>>>>> hoping to avoid using type units to
help<br>
>>>>>> The thing which<br>
>>>>>> should probably be changed is that
dsymutil should not have its version<br>
>>>>>> of code generating DWARF tables. It
should call already existed<br>
>>>>>> DWARF5/DWARF64 implementations.
Then dsymutil would always<br>
>>>>>> use last DWARF generators.<br>
>>>>> Possibly - I don't know what the
architectural tradeoffs for that look<br>
>>>>> like - I'd imagine DWARFLinker has
sufficiently different<br>
>>>>> needs/tradeoffs than LLVM's DWARF
generation code (rewriting existing<br>
>>>>> DIEs compared to building new ones from
scratch, etc) that it might be<br>
>>>>> hard for them to share a lot of their
implementation.<br>
>>>> It is not easy, and would require some
additions, but it would benefit<br>
>>>> in that all format implementation is in one
place. Thus changing that place<br>
>>>> would reflect in other places. There are at
least three implementations for<br>
>>>> .debug_ranges, .debug_aranges currently...<br>
>>>><br>
>>>><br>
>>>>>>>> 2. implement multi-threaded
execution.<br>
>>>>>>>> 3. support of split DWARF.<br>
>>>>>>> Maybe, though I'm still not
sure it'd be the right tradeoff -<br>
>>>>>>> especially if it involved
having to wait to run the .dwo merger (call<br>
>>>>>>> it DWARF-aware dwp, or dsymutil
with dwp support) until after the<br>
>>>>>>> linker ran.<br>
>>>>>>><br>
>>>>>>>> 4. implement dsymutil for
non-darwin platform.<br>
>>>>>>> That's probably, essentially
(3), more-or-less. Split DWARF is<br>
>>>>>>> somewhat of a formalization of
Apple's/MachO DWARF distribution model<br>
>>>>>>> (leave DWARF it in files that
aren't linked/use them from a debugger,<br>
>>>>>>> but also be able to merge them
into some final file (dsym or dwp) for<br>
>>>>>>> archival purposes)<br>
>>>>>>><br>
>>>>>>>> All of this is a massive
piece of work.<br>
>>>>>>>> Our original investment was
to solve two problems:<br>
>>>>>>>><br>
>>>>>>>> 1. Overlapped address
ranges, which is currently close to being solved. Thank you
for helping with that!<br>
>>>>>>> Yeah, again, sorry that's taken
quite so long/somewhat circuitous route.<br>
>>>>>>><br>
>>>>>>>> 2. Size of debug info. That
still becomes an issue, but we are unsure whether we are
ready to<br>
>>>>>>>> invest in solving all
the above 1-4 problems and how much community interested in
it.<br>
>>>>>>> Fair, for sure - I don't think
you'd need to sign up to solve all of<br>
>>>>>>> them (don't think they
necessarily need solving). Potentially moving<br>
>>>>>>> the logic out into a separate
tool as Fangrui's considering - a<br>
>>>>>>> post-link DWARF optimizer,
rather than in-linker DWARF optimization.<br>
>>>>>>><br>
>>>>>>> I really don't want to give you
the runaround like this - but multiple<br>
>>>>>>> times slower links is something
that seems pretty problematic for most<br>
>>>>>>> users, to the point of weighing
the maintainability of lld against the<br>
>>>>>>> convenience of having this
functionality in-linker rather than in a<br>
>>>>>>> post-link optimizer.<br>
>>>>>>><br>
>>>>>>> (I know you've spoken a bit
before about your users needs - but if<br>
>>>>>>> it's possible, could you
explain (again :/) why they have such a<br>
>>>>>>> strong need for smaller DWARF?
While DWARF size is an ongoing concern<br>
>>>>>>> for many users (Google
certainly - hence the invention of Split DWARF,<br>
>>>>>>> use of type units and
compressed DWARF, etc) - usually it's in rather<br>
>>>>>>> large programs, but it sounds
like you're dealing with relatively<br>
>>>>>>> small ones (otherwise the
increase in link time, I'd imagine, would be<br>
>>>>>>> prohibitive for your users?)?<br>
>>>>>> We have many large programs and
keep Dayly/Nightly debug builds,<br>
>>>>>> which takes a lot of disk space.
Compilation time for these programs is big.<br>
>>>>>> The scenario is "compile once".(not
compile-debug-compile-debug).<br>
>>>>>> So we think that solution(like
dsymutil/DWARFLinker) would not slowdown<br>
>>>>>> the compilation time of overall
build significantly(see above numbers for<br>
>>>>>> llvm codebase) and would allow us
to reduce disk space required to keep<br>
>>>>>> all of these builds.<br>
>>>>> Ah, OK - for archival purposes. So the
interactive developers wouldn't<br>
>>>>> necessarily be using this feature.
Makes sense - similar to dsymutil<br>
>>>>> and dwp, mostly used for archival
purposes & you can debug straight<br>
>>>> >from .o/.dwos for interactive/iterative
development.<br>
>>>><br>
>>>>> In that case, it seems more likely that
a separate tool might suffice.<br>
>>>> agreed: if to continue the work on this
then it makes sense to<br>
>>>> do it as separate tool. Make it fast
enough. And if there would be interest<br>
>>>> in it - then it would probably be possible
to return to idea calling it from linker.<br>
>>>><br>
>>>>> Also, out of curiosity - have you tried
just compressing the output<br>
>>>>> (-gz (I think that does the right thing
for the linker level<br>
>>>>> compression too, otherwise
-Wl,-compress-debug-sections might do it))<br>
>>>>> or are you already doing that in
addition?<br>
>>>> sure. we use -Wl,-compress-debug-sections.<br>
>>>><br>
>>>> Thank you, Alexey.<br>
>>>><br>
>>>>>>> You mentioned that the
usability cost of<br>
>>>>>>> Split DWARF for your users was
too high (or high enough to justify<br>
>>>>>>> this alternative work of
DWARF-aware linking)? That all seems a bit<br>
>>>>>>> surprising to me - though I
understand the deployment issues of Split<br>
>>>>>>> DWARF do present some
challenges to users in more heterogenous<br>
>>>>>>> environments than Google's...
still, I'd have thought there was some<br>
>>>>>>> hope there)<br>
>>>>>> Our tools does not support split
dwarf yet. Though we plan to implement it.<br>
>>>>>> When we would have support of split
dwarf then it would be<br>
>>>>>> convenient to have easy way to
share built debug binaries. llvm-dwp is the<br>
>>>>>> answer to this. DWARFLinker could
probably be another answer.<br>
>>>>> Ah, fair enough - thanks for the
context!<br>
>>>>>>>>> One way to do that
would be to have a CU-local type indirection table.<br>
>>>>>>>>> DIEs reference local
type numbers (like local address/string numbers -<br>
>>>>>>>>> addrx/strx/rnglistx)
and that table contains either sig8 (no linker<br>
>>>>>>>>> fixups required) or the
local type offsets you describe - the linker<br>
>>>>>>>>> would then only need to
read this type number indirection table and<br>
>>>>>>>>> rewrite them to the
final type numbers.<br>
>>>>>>>> Yes, that could be
additionally done if this process would be time-consuming.<br>
>>>>>>>><br>
>>>>>>>> David, thank you for all
your comments and explanations. They are extremely helpful.<br>
>>>>>>> Sure thing - really appreciate
your patience with all this - it's... a<br>
>>>>>>> lot of moving parts.<br>
>>>>>>> - Dave<br>
>>>>>>> Thank you, Alexey.<br>
>>>>>>><br>
>>>>>>>> sig8 hash-id would be used
to compare types and to deduplicate them.<br>
>>>>>>>> It would speed up the
current dsymutil context analysis.<br>
>>>>>>>> Types having the same
hash-id could be deduplicated.<br>
>>>>>>>> This would allow
deduplicating a more number of types than current dsymutil.<br>
>>>>>>>> Incomplete type definitions
having a similar set of members are not deduplicated by
dsymutil currently.<br>
>>>>>>>> In this case they would
have the same hash-id.<br>
>>>>>>>><br>
>>>>>>>> This "type table" would
take less space than current "type units" and current ODR
solution.<br>
>>>>>>>><br>
>>>>>>>> Above is just an idea on
how to help DWARF-aware linker(based on idea removing
obsolete debug info)<br>
>>>>>>>> to work faster(if that is
interesting).<br>
>>>>>>>><br>
>>>>>>>> Alexey.<br>
>>>>>>>><br>
>>>>>>>>> From: llvm-dev <<a href="mailto:llvm-dev-bounces@lists.llvm.org" target="_blank">llvm-dev-bounces@lists.llvm.org</a>>
On Behalf Of James Henderson via llvm-dev<br>
>>>>>>>>> Sent: Wednesday, June
3, 2020 3:48 AM<br>
>>>>>>>>> To: David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>><br>
>>>>>>>>> Cc: <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>>>>>>>> Subject: Re: [llvm-dev]
[Debuginfo][DWARF][LLD] Remove obsolete debug info in lld.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> It makes me sad that
the linker (via a library or otherwise) has to be
"DWARF-aware" to be able to effectively handle
--gc-sections, COMDATs, --icf etc for debug info, without
leaving large blocks of data kicking around.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> The patching to -1 (or
equivalent) is probably a good lightweight solution (though
I'd love it if it could be done based on section type in the
future rather than section name, but that's probably outside
the realm of DWARF), as it requires only minimal
understanding in the linker, but anything beyond that seems
to be complicated logic that is mostly due to the structure
of DWARF. Patching to -1 does feel a bit like a sticking
plaster/band aid to patch over the issue rather than
properly solving it too - there will still be debug data
(potentially significant amounts in COMDAT-heavy objects)
that the linker has to write and the debugger has to somehow
know how to skip (even if it knows that -1 is special-case
due to the standard being updated, it needs to get as far as
the -1), which is all wasted effort.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> We've already seen from
Alexey's prototyping, and from our own experiences with the
Sony proprietary linker (which tried to rewrite .debug_line
only) that deconstructing the DWARF so that it can be more
optimally reassembled at link time is slow going, and will
probably inevitably be however much effort is put into
optimising it. For a start, given the current standards,
it's impossible to know how to deconstruct it without having
to parse vast amounts of DWARF, which is typically going to
mean a lot more parsing work than the linker would normally
have to deal with. Additionally, much of this parsing work
is wasted effort, since it seems unlikely in many links that
large amounts of the DWARF will be redundant. Having an
option to opt-in doesn't help much there, since it just
means the logic exists without most people using it, due to
it not being good enough, or potentially they don't even
know it exists.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> I don't have
particularly concrete suggestions as to how to solve the
structural problems with DWARF at this point. The only thing
that seems obvious to me is a more "blessed" approach to
fragmentation of sections, similar to what I tried with my
prototype mentioned earlier in the thread, although we'd
need to figure out the previously stated performance issues.
Other ideas might tie into this, like somehow sharing the
various table headers a bit like CIEs in .eh_frame that
could be merged by the linker - each object could have
separate table header sections, which are referenced by the
individual .debug_* blocks, which in turn are one per
function/data piece and easily discardable/merged by the
linker.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> Just some thoughts.<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> James<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>> On Tue, 2 Jun 2020 at
19:24, David Blaikie via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>>
wrote:<br>
>>>>>>>>><br>
>>>>>>>>> On Tue, May 19, 2020 at
7:17 AM Alexey Lapshin<br>
>>>>>>>>> <<a href="mailto:alapshin@accesssoftek.com" target="_blank">alapshin@accesssoftek.com</a>>
wrote:<br>
>>>>>>>>>> Hi David, please
find my comments inside:<br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>>>>> Broad
question: Do you have any specific motivation/users/etc in
implementing this (if you can speak about it)?<br>
>>>>>>>>>>>>> - it
might help motivate the work, understand what tradeoffs
might be suitable for you/your users, etc.<br>
>>>>>>>>>>>> There are
two general requirements:<br>
>>>>>>>>>>>> 1) Remove
(or clean) invalid debug info.<br>
>>>>>>>>>>> Perhaps a
simpler direct solution for your immediate needs might be a
much narrower,<br>
>>>>>>>>>>> and more
efficient linker-DWARF-awareness feature:<br>
>>>>>>>>>>><br>
>>>>>>>>>>> With DWARFv5,
rnglists present an opportunity for a DWARF linker to
rewrite the ranges<br>
>>>>>>>>>>> without parsing
the rest of the DWARF. /technically/ this isn't guaranteed -
rnglist entries<br>
>>>>>>>>>>> can be
referenced either directly, or by index. If all rnglists are
referenced by index, then<br>
>>>>>>>>>>> a linker could
parse only the debug_rnglists section and rewrite ranges to
remove any<br>
>>>>>>>>>>> address ranges
that refer to optimized-out code.<br>
>>>>>>>>>>><br>
>>>>>>>>>>> This would only
be correct for rnglists that had no direct references to
them (that only were<br>
>>>>>>>>>>> referenced via
the indexes) - but we could either implement it with that
assumption, or could<br>
>>>>>>>>>>> add an LLVM
extension attribute on the CU that would say "I promise I
only referenced rnglists<br>
>>>>>>>>>>> via rnglistx
forms/indexes). If this DWARF-aware linking would have to
read the CU DIE (not<br>
>>>>>>>>>>> all the other
DIEs) it /could/ also then rewrite high/low_pc if the CU
wasn't using ranges...<br>
>>>>>>>>>>> but that
wouldn't come up in the function-removal case, because then
you'd have ranges anyway,<br>
>>>>>>>>>>> so no need for
that.<br>
>>>>>>>>>>><br>
>>>>>>>>>>> Such a
DWARF-aware rnglist linking could also simplify rnglists, in
cases where functions<br>
>>>>>>>>>>> ended up being
laid out next to each other, the linker could coalesce their
ranges together.<br>
>>>>>>>>>>><br>
>>>>>>>>>>> I imagine this
could be implemented with very little overhead to linking,
especially compared<br>
>>>>>>>>>>> to the overhead
of full DWARF-aware linking.<br>
>>>>>>>>>>><br>
>>>>>>>>>>> Though none of
this fixes Split DWARF, where the linker doesn't get a
chance to see the<br>
>>>>>>>>>>> addresses being
used - but if you only want/need the CU-level ranges to be
correct, this<br>
>>>>>>>>>>> might be a
viable fix, and quite efficient.<br>
>>>>>>>>>> Yes, we think about
that alternative. This would resolve our problem of invalid
debug info<br>
>>>>>>>>>> and would work much
faster. Thus, if we would not have good results for D74169
then we<br>
>>>>>>>>>> will implement it.
Do you think it could be useful to have this solution in
upstream?<br>
>>>>>>>>> A pure rnglist
rewriting - I think it'd be OK to have in upstream -<br>
>>>>>>>>> again, cost/benefit/etc
would have to be weighed. I'm not sure it<br>
>>>>>>>>> would save enough space
to be particularly valuable beyond the<br>
>>>>>>>>> correctness issue - and
it doesn't completely solve the correctness<br>
>>>>>>>>> issue for zero-address
usage or low-address usage (because you could<br>
>>>>>>>>> still have overlapping
subprograms inside a CU - so if you were<br>
>>>>>>>>> symbolizing you could
use the correct rnglist to filter, but then go<br>
>>>>>>>>> look inside the CU only
to find two subprograms that had that address<br>
>>>>>>>>> & not know which
one was the correct one an which one was the<br>
>>>>>>>>> discarded one).<br>
>>>>>>>>><br>
>>>>>>>>> rnglist rewriting might
be easy enough to prototype - but depends what<br>
>>>>>>>>> you want to spend your
time on, I know this whole issue has been a<br>
>>>>>>>>> huge investment of your
time already - but maybe this recent<br>
>>>>>>>>> revitalization of the
conversation around having an explicit value in<br>
>>>>>>>>> the linker might be
sufficient to address everyone's needs... *fingers<br>
>>>>>>>>> crossed*)<br>
>>>>>>>>><br>
>>>>>>>>><br>
>>>>>>>>>>>> 2) Optimize
the DWARF size.<br>
>>>>>>>>>>> Do your users
care much about this? I imagine if they had significant
DWARF size issues,<br>
>>>>>>>>>>> they'd have
significant link time issues and the kind of cost to link
time this feature has would<br>
>>>>>>>>>>> be prohibitive
- but perhaps they're sharing linked binaries much more
often than they're<br>
>>>>>>>>>>> actually
performing linking.<br>
>>>>>>>>>> Yes, they do. They
also have significant link-time issues.<br>
>>>>>>>>>> So current
performance results of D74169 are not very acceptable.<br>
>>>>>>>>>> We hope to improve
it.<br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>>>> The
specifics which our users have:<br>
>>>>>>>>>>>> -
embedded platform which uses 0 as start of .text section.<br>
>>>>>>>>>>>> - custom
toolset which does not support all features yet(f.e. split
dwarf).<br>
>>>>>>>>>>>> -
tolerant of the link-time increase.<br>
>>>>>>>>>>>> - need a
useful way to share debug builds.<br>
>>>>>>>>>>> Sharing two
files (executable and dwp) is significantly less useful than
sharing one file?<br>
>>>>>>>>>> Probably not
significantly, but yes, it looks less useful comparing to
D74169.<br>
>>>>>>>>>> Having only two
files (executable and .dwp) looks significantly better than
having executable and multiple .dwo files.<br>
>>>>>>>>>> Having only one
file(executable) with minimal size looks better than the two
files with a bigger size.<br>
>>>>>>>>>><br>
>>>>>>>>>> clang compiled with
-gsplitdwarf takes 0.9G for executable and 0.9G for .dwp.<br>
>>>>>>>>>> clang compiled with
-gc-debuginfo takes only 0.76G for single executable.<br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>>>> For the
first point: we have a problem "Overlapping address ranges
starting from 0"(D59553).<br>
>>>>>>>>>>>> We use
custom solution, but the general solution like D74169 would
be better here.<br>
>>>>>>>>>>> If CU ranges
are the only ones that need fixing, then I think the above
solution might be as<br>
>>>>>>>>>>> good/better -
if more than CU ranges need fixing, then I think we might
want to start talking about<br>
>>>>>>>>>>> how to fix
DWARF itself (split and non-split) to signal certain
addresses point to dead code with a<br>
>>>>>>>>>>> specific
blessed value that linkers would need to implement - because
with Split DWARF there's<br>
>>>>>>>>>>> no way to solve
the non-CU addresses at the linker.<br>
>>>>>>>>>> I think the
worthful solution for that signal value would be LowPC >
HighPC.<br>
>>>>>>>>>> That does not
require additional bits in DWARF.<br>
>>>>>>>>>> It would be natural
to skip such address ranges since they explicitly marked as
invalid.<br>
>>>>>>>>>> It could be
implemented in a linker very easily. Probably, it would make
sense to describe that<br>
>>>>>>>>>> usage in DWARF
standard.<br>
>>>>>>>>>><br>
>>>>>>>>>> As to the addresses
which are not seen by the linker(since they are in .dwo
files) - yes,<br>
>>>>>>>>>> they need to have
another solution. Could you show an example of such a case,
please?<br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>><br>
>>>>>>>>>>>>> 2.
Support of type units.<br>
>>>>>>>>>>>>>>
That could be implemented further.<br>
>>>>>>>>>>>>>
Enabling type units increases object size to make it easier
to deduplicate at link time by a DWARF-unaware<br>
>>>>>>>>>>>>> linker.
With a DWARF aware linker it'd be generally desirable not to
have to add that object size overhead to<br>
>>>>>>>>>>>>> get the
linking improvements.<br>
>>>>>>>>>>>> But,
DWARFLinker should adequately work with type units since
they are already implemented.<br>
>>>>>>>>>>> Maybe - it'd be
nice & all, but I don't think it's an outright necessity
- if someone knows they're using<br>
>>>>>>>>>>> a DWARF-aware
linker, they'd probably not use type units in their object
files. It's possible someone<br>
>>>>>>>>>>> doesn't know
for sure & maybe they have pre-canned debug object files
from someone else, etc.<br>
>>>>>>>>>> I see.<br>
>>>>>>>>>><br>
>>>>>>>>>>>> Another
thing is that the idea behind type units has the potential
to help Dwarf-aware linker to work faster.<br>
>>>>>>>>>>>> Currently,
DWARFLinker analyzes context to understand whether types are
the same or not.<br>
>>>>>>>>>>> When you say
"analyzes context" what do you mean? Usually I'd take that
to mean<br>
>>>>>>>>>>> "looks at
things outside the type itself - like what namespace it's
in, etc" - which, yes,<br>
>>>>>>>>>>> it should do
that, but it doesn't seem very expensive to do. But I guess
you actually<br>
>>>>>>>>>>> mean something
about doing structural equivalence in some way, looking at
things inside the type?<br>
>>>>>>>>>> I think it could be
useful for both cases. Currently, dsymutil does only first
thing<br>
>>>>>>>>>> (look at type name,
namespace name, etc..) and does not do the second thing<br>
>>>>>>>>>> (doing structural
equivalence). Analyzing type names is currently quite
expensive<br>
>>>>>>>>>> (the only search in
string pool takes ~10 sec from 70 sec of overall time).<br>
>>>>>>>>>> That is expensive
because of many things should be done to work with strings:<br>
>>>>>>>>>> parse DWARF, search
and resolve relocations, compute a hash for strings,<br>
>>>>>>>>>> put data into a
string pool, create a fully qualified name(like
namespace::function::name).<br>
>>>>>>>>>> It looks like it
could be optimized and finally require less time, but it
still would be a noticeable<br>
>>>>>>>>>> part of the overall
time.<br>
>>>>>>>>>><br>
>>>>>>>>>> If dsymutil starts
to check for the structural equivalence, then the process
would be even more slowly.<br>
>>>>>>>>>> So, If instead of
comparing types structure, there would be checked single
hash-id - then this process<br>
>>>>>>>>>> would also be
faster.<br>
>>>>>>>>>><br>
>>>>>>>>>> Thus I think using
hash-id to compare types would allow to make current
implementation faster and would<br>
>>>>>>>>>> allow handling
incomplete types by DWARFLinker without massive performance
degradation also.<br>
>>>>>>>>>><br>
>>>>>>>>>>>> But the
context is known when types are generated. So, no need to
spent the time analyzing it.<br>
>>>>>>>>>>>> If types
could be compared without analyzing context, then
Dwarf-aware linker would work faster.<br>
>>>>>>>>>>>> That is
just an idea(not for immediate implementation): If types
would be stored in some "type table"<br>
>>>>>>>>>>>> (instead of
COMDAT section group) and could be accessed through
hash-id(like type units<br>
>>>>>>>>>>>> - then it
would be the solution requiring fewer bits to store but
allowing to compare types<br>
>>>>>>>>>>>> by
hash-id(not analysing context).<br>
>>>>>>>>>>>> In this
case, size increasing would be small. And processing time
could be done faster.<br>
>>>>>>>>>>>><br>
>>>>>>>>>>>> this is
just an idea and could be discussed separately from the
problem of integrating of D74169.<br>
>>>>>>>>>>>>>> 6.
-flto=thin<br>
>>>>>>>>>>>>>>
That problem was described in this review <a href="https://reviews.llvm.org/D54747#1503720" rel="noreferrer" target="_blank">https://reviews.llvm.org/D54747#1503720</a>.
It also exists in<br>
>>>>>>>>>>>>>>
current DWARFLinker/dsymutil implementation. I think that
problem should be discussed more: it could<br>
>>>>>>>>>>>>>>
probably be fixed by avoiding generation of such incomplete
declaration during thinlto,<br>
>>>>>>>>>>>>>>
That would be costly to produce extra/redundant debug info
in ThinLTO - actually ThinLTO could be doing<br>
>>>>>>>>>>>>>>
more to reduce that redundancy early on (actually removing
definitions from some llvm Modules if the type<br>
>>>>>>>>>>>>>>
definition is known to exist in another Module, etc)<br>
>>>>>>>>>>>>> I don't
know if it's a problem since that patch was reverted.<br>
>>>>>>>>>>>> Yes. That
patch was reverted, but this patch(D74169) has the same
problem.<br>
>>>>>>>>>>>> if D74169
would be applied and --gc-debuginfo used then structure type<br>
>>>>>>>>>>>> definition
would be removed.<br>
>>>>>>>>>>>> DWARFLinker
could handle that case - "removing definitions from some
llvm Modules if the type<br>
>>>>>>>>>>>> definition
is known to exist in another Module".<br>
>>>>>>>>>>>> i.e.
DWARFLinker could replace the declaration with the
definition.<br>
>>>>>>>>>>>> But that
problem could be more easily resolved when debug info is
generated(probably without<br>
>>>>>>>>>>>> significant
increase of debug info size):<br>
>>>>>>>>>>>> Here we
have:<br>
>>>>>>>>>>>>
DW_TAG_compile_unit(0x0000000b) - compile unit containing
concrete instance for function "f".<br>
>>>>>>>>>>>>
DW_TAG_compile_unit(0x00000073) - compile unit containing
abstract instance root for function "f".<br>
>>>>>>>>>>>>
DW_TAG_compile_unit(0x000000c1) - compile unit containing
function "f" definition.<br>
>>>>>>>>>>>> Code for
function "f" was deleted. gc-debuginfo deletes compile unit
DW_TAG_compile_unit(0x000000c1)<br>
>>>>>>>>>>>> containing
"f" definition (since there is no corresponding code). But
it has structure "Foo" definition<br>
>>>>>>>>>>>>
DW_TAG_structure_type(0x0000011e) referenced from
DW_TAG_compile_unit(0x00000073)<br>
>>>>>>>>>>>> by
declaration DW_TAG_structure_type(0x000000ae). That
declaration is exactly the case when definition<br>
>>>>>>>>>>>> was removed
by thinlto and replaced with declaration.<br>
>>>>>>>>>>>> Would it
cost too much if type definition would not be replaced with
declaration for "abstract instance root"?<br>
>>>>>>>>>>>> The number
of concrete instances is bigger than number of abstract
instance roots.<br>
>>>>>>>>>>>> Probably,
it would not be too costly to leave definition in abstract
instance root?<br>
>>>>>>>>>><br>
>>>>>>>>>>>>
Alternatively, Would it cost too much if type definition
would not be replaced with declaration when<br>
>>>>>>>>>>>> declaration
references type from not used function? (lto could
understand that concrete function is not used).<br>
>>>>>>>>>>> I don't follow
this example - could you provide a small concrete test case
I could reproduce?<br>
>>>>>>>>>> I would provide a
test case if necessary. But it looks like this issue is
finally clear, and you already commented on that.<br>
>>>>>>>>>><br>
>>>>>>>>>>> Oh, I guess
this is happening perhaps because ThinLTO can't know for
sure that a standalone<br>
>>>>>>>>>>> definition of
'f' won't be needed - so it produces one in case one of the
inlining opportunities<br>
>>>>>>>>>>> doesn't end up
inlining. Then it turns out all calls got inlined, so the
external definition wasn't needed.<br>
>>>>>>>>>>> Oh, you're
suggesting that these 3 CUs got emitted into one object file
during LTO, but that DWARFLinker<br>
>>>>>>>>>>> drops a CU
without any code in it - even though... So far as I know, in
LTO, LLVM directly references<br>
>>>>>>>>>>> types across
units if the CUs are all emitted in the same object file.
(and if they weren't in the same<br>
>>>>>>>>>>> object file -
then the abstract_origin couldn't be pointing cross-CU).<br>
>>>>>>>>>>> I guess some
basic things to say:<br>
>>>>>>>>>>> With ThinLTO,
the concrete/standalone function definition is emitted in
case some call sites don't end up<br>
>>>>>>>>>>> being inlined.
So we know it'll be emitted (but might not be needed by the
actual linker)<br>
>>>>>>>>>>> ANy number of
inline calls might exist - but we shouldn't put the type
information into those, because<br>
>>>>>>>>>>> they aren't
guaranteed to emit it (if the inline function gets optimized
away, there would be nothing to<br>
>>>>>>>>>>> enforce the
type being emitted) - and even if we forced the type
information to be emitted into one<br>
>>>>>>>>>>> object file
that has an inline copy of the function - there's no
guarantee that object file will get linked in either.<br>
>>>>>>>>>>> So, no, I don't
think there's much we can do to keep the size of object
files down, while guaranteeing<br>
>>>>>>>>>>> the type
information will be emitted with the usual linker semantics.<br>
>>>>>>>>>> Then
dsymutil/DWARFLinker could be changed to handle that(though
it would probably be not very efficient).<br>
>>>>>>>>>> If thinlto would
understand that function is not used finally(and then must
not contain referenced type definition),<br>
>>>>>>>>>> then this situation
could be handled more effectively.<br>
>>>>>>>>>><br>
>>>>>>>>>> Thank you, Alexey.<br>
>>>>>>>>>><br>
>>>>>>>>>>>><br>
>>>>>>>>>>>><br>
>>>>>>>>>>>>
_______________________________________________<br>
>>>>>>>>>>>> LLVM
Developers mailing list<br>
>>>>>>>>>>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>>>>>>>>>>> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
>>>>>>>>>
_______________________________________________<br>
>>>>>>>>> LLVM Developers mailing
list<br>
>>>>>>>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>>>>>>>> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
>>> _______________________________________________<br>
>>> LLVM Developers mailing list<br>
>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
</blockquote>
</div>
</div>
</blockquote>
</div>
</blockquote></div></div>
</blockquote></div>