llvm dev - Nov 2018 - DWARF: Ranges base address specifier entries & Gold's gdb-index 32 bit bug

On Tue, Aug 8, 2017 at 10:50 AM Robinson, Paul <paul.robinson at sony.com>
wrote:
> Can gdb handle these?  i.e. is it just gold that has the problem?
>
Yep, it's just gold when it's building the gdb-index (an accelerator
table
for GDB)

>   Conditioning on debugger tuning when it's not the debugger that has
the
> problem… icky.
>
It does. Though to a lesser degree even gnu-pubnames isn't a perfect
signal. One, I think, could use that even without a linker-generated
gdb-index - presumably GDB's still faster by reading gnu-pubnames than by
reading the raw DIEs to do lookup.

>   If that's the case then asking people to add –ggnu-pubnames seems
okay.
> Assuming that gets handed down through IR/metadata then the LTO problem
> goes away too.
>
gnu-pubnames has such problems (it's done as a backend option, passed down
from the clang driver via -mllvm -generate-gnu-dwarf-pub-sections, but
they're already there... - mostly the discussion around flags was "If
we're
adding a new one, what's the right tradeoff here for a workaround flag
compared to the platonic ideals of CU-based flags as much as possible") but
it's already that way...

- Dave

> --paulr
>
>
>
> *From:* David Blaikie [mailto:dblaikie at gmail.com]
> *Sent:* Tuesday, August 08, 2017 10:59 AM
> *To:* Robinson, Paul; llvm-dev; Adrian Prantl; Eric Christopher; Nico
> Weber
> *Subject:* Re: DWARF: Ranges base address specifier entries &
Gold's
> gdb-index 32 bit bug
>
>
>
> Adrian: any thoughts? Has LLDB been fixed to support this yet?
>
> On Tue, Aug 8, 2017 at 6:33 AM Robinson, Paul <paul.robinson at
sony.com>
> wrote:
>
> My inclination would be to use "disable if 32-bit and
–ggnu-pubnames" as
> the default,
>
>
> Unfortunately Nico points out that Chrome doesn't currently use
> -ggnu-pubnames :/ So to continue to work "out of the box"
we'd have to
> broaden the surface to "disable if 32 bit and targeting gdb".
>
> -ggnu-pubnames makes gdb-index creation much more efficient by allowing
> gdb-index generation not to have to parse most of the DWARF DIEs, and just
> create the index straight from a table. (this is necessary for correctness
> in -gsplit-dwarf mode (where the DIEs are not available at link time as
> they're in the .dwo not the .o), and an efficiency gain otherwise)
>
> If "disable if 32 bit and targeting gdb" is overly broad (I feel
it kind
> of is, but I'm not wedded to it) we could require those working around
the
> gold bug to add -ggnu-pubnames to add a bit more of a specific hint that
> they're intending to build gdb-index from these objects.
>
>
> and have a flag.
>
> This would be determined in DwarfDebug, right?
>
>
> Yes
>
>
> If the check is in CodeGen then I'd think it would be insensitive to
LTO.
>
>
> There's some desire that LTO works as-if it was non-LTO. So if you pass
> flags (eg: debugger tuning flags, in this case) to each separate compile,
> you get the same behavior if you LTO as when you didn't - so we
preserve
> things in the CU, and then do the specified thing per-CU. (eg: emission
> kind (gmlt/limited/full), split-debug-inlining, debug-info-for-profiling,
> etc)
>
> In this case if, say, you're on an LLDB platform but prefer to use GDB
(&
> happen to use Gold and gdb-index on 32 bit and LTO) - so you pass the
> debugger tuning=gdb flag to your compiles, it wouldn't disable the
feature,
> because the flag would be consulted at CodeGen time and not at compile time
> when you passed the tuning flag. (oh, actually the tuning is passed down
> via TM Options anyway, not on the CU - so I guess you wouldn't be
getitng
> the tuning at all today because that flag already doesn't follow the
per-CU
> model... well I guess that makes it easier)
>
>
> Setting it via IR/metadata is adding a bunch of complexity just to work
> around a gold bug.  Not worth it IMO.
>
>
> Fair enough. Any thoughts on whether it should be a frontend-visible
> option or just a cc1 or even backend option? I'll probably just leave
it as
> a backend option at this point, could even implement the defaulting down in
> LLVM without any changes to Clang, I suppose...
>
>
> --paulr
>
>
>
> *From:* David Blaikie [mailto:dblaikie at gmail.com]
> *Sent:* Monday, August 07, 2017 6:58 PM
> *To:* llvm-dev; Robinson, Paul; Adrian Prantl; Eric Christopher; Nico
> Weber
> *Subject:* DWARF: Ranges base address specifier entries & Gold's
> gdb-index 32 bit bug
>
>
>
> Context:
>
> In r309526 (with a followup fix in r309529) I implemented the use of
> DWARF's debug_ranges base address specifier entries to reduce the
number of
> object file relocations needed for debug_ranges*.
>
> * in a particular binary internally, an optimized build had a 70%
> reduction in debug_ranges.reloc, a 16% decrease in total object size (with
> compressed debug info and fission)
>
> Nico noted that this broke the Chromium build (
> https://bugs.llvm.org/show_bug.cgi?id=34007 ). Turns out GNU Binutils
> Gold has a bug (I've reported it as:
> https://sourceware.org/bugzilla/show_bug.cgi?id=21894 - wouldn't expect
> much action on it, though) where it fails to parse a base address specifier
> entry when trying to build gdb-index (-Wl,-gdb-index) for a 32 bit build.
>
> To address this in the short term, I added a flag that disables this
> feature by default for now.
>
> So, two questions:
>
> 1) What does everyone reckon the best solution to this is?
>   * Keep the feature disabled by default - with a flag to enable it for
> those who want it/can use it
>   * Enable the feature everywhere - with a flag to disable it
>   * Conditionally disable (or conditionally enable) the feature, with a
> flag to enable/disable it
>     * best condition we probably have is "disabled if 32 bit and
> -ggnu-pubnames" (gnu-pubnames are necessary for efficient building of
> gdb-index, and isn't the default)
> 2) How to implement the flag:
>   * backend option as it is currently (-mllvm
> -use-dwarf-ranges-base-address-specifier (maybe drop "-specifier"
to make
> it a bit more terse))
>   * clang option that maps to
>     * backend option (as currently)
>     * MCOption (programmatic/C++ API rather than command line of a backend
> option)
>     * LLVM IR attribute on the CU (this feature could be supported on a
> per-CU basis easiyl enough, thus survive LTO, etc, exactly as the user
> requested (but really if the user requested this on at least one CU, they
> probably might as well have it on all their CUs))
>     * LLVM IR module metadata
>
>
> Any ideas/thoughts/other aspects?
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170808/fdbcc06e/attachment.html>

Since it seems like there's unlikely to be a good heuristic that
wouldn't
potentially need to be overridden by some reasonable use cases, we'll need
a flag one way or the thore - so let's add the flag & if someone wants
to
push on the defaults (I'll probably just leave it as default off without
any heuristic for turning it on by default for any cases) they can.

So I've sent out a couple of reviews for the LLVM and Clang parts of adding
a flag to opt-in to this functionality:

Clang: https://reviews.llvm.org/D54243
LLVM: https://reviews.llvm.org/D54242

Main thing is just to bikeshed the naming of the metadata and flag.

Thanks everyone!
- Dave

On Tue, Aug 8, 2017 at 10:57 AM David Blaikie <dblaikie at gmail.com>
wrote:
> On Tue, Aug 8, 2017 at 10:50 AM Robinson, Paul <paul.robinson at
sony.com>
> wrote:
>
>> Can gdb handle these?  i.e. is it just gold that has the problem?
>>
>
> Yep, it's just gold when it's building the gdb-index (an
accelerator table
> for GDB)
>
>
>>   Conditioning on debugger tuning when it's not the debugger that
has the
>> problem… icky.
>>
>
> It does. Though to a lesser degree even gnu-pubnames isn't a perfect
> signal. One, I think, could use that even without a linker-generated
> gdb-index - presumably GDB's still faster by reading gnu-pubnames than
by
> reading the raw DIEs to do lookup.
>
>
>>   If that's the case then asking people to add –ggnu-pubnames seems
>> okay.  Assuming that gets handed down through IR/metadata then the LTO
>> problem goes away too.
>>
>
> gnu-pubnames has such problems (it's done as a backend option, passed
down
> from the clang driver via -mllvm -generate-gnu-dwarf-pub-sections, but
> they're already there... - mostly the discussion around flags was
"If we're
> adding a new one, what's the right tradeoff here for a workaround flag
> compared to the platonic ideals of CU-based flags as much as
possible") but
> it's already that way...
>
> - Dave
>
>
>> --paulr
>>
>>
>>
>> *From:* David Blaikie [mailto:dblaikie at gmail.com]
>> *Sent:* Tuesday, August 08, 2017 10:59 AM
>> *To:* Robinson, Paul; llvm-dev; Adrian Prantl; Eric Christopher; Nico
>> Weber
>> *Subject:* Re: DWARF: Ranges base address specifier entries &
Gold's
>> gdb-index 32 bit bug
>>
>>
>>
>> Adrian: any thoughts? Has LLDB been fixed to support this yet?
>>
>> On Tue, Aug 8, 2017 at 6:33 AM Robinson, Paul <paul.robinson at
sony.com>
>> wrote:
>>
>> My inclination would be to use "disable if 32-bit and
–ggnu-pubnames" as
>> the default,
>>
>>
>> Unfortunately Nico points out that Chrome doesn't currently use
>> -ggnu-pubnames :/ So to continue to work "out of the box"
we'd have to
>> broaden the surface to "disable if 32 bit and targeting gdb".
>>
>> -ggnu-pubnames makes gdb-index creation much more efficient by allowing
>> gdb-index generation not to have to parse most of the DWARF DIEs, and
just
>> create the index straight from a table. (this is necessary for
correctness
>> in -gsplit-dwarf mode (where the DIEs are not available at link time as
>> they're in the .dwo not the .o), and an efficiency gain otherwise)
>>
>> If "disable if 32 bit and targeting gdb" is overly broad (I
feel it kind
>> of is, but I'm not wedded to it) we could require those working
around the
>> gold bug to add -ggnu-pubnames to add a bit more of a specific hint
that
>> they're intending to build gdb-index from these objects.
>>
>>
>> and have a flag.
>>
>> This would be determined in DwarfDebug, right?
>>
>>
>> Yes
>>
>>
>> If the check is in CodeGen then I'd think it would be insensitive
to LTO.
>>
>>
>> There's some desire that LTO works as-if it was non-LTO. So if you
pass
>> flags (eg: debugger tuning flags, in this case) to each separate
compile,
>> you get the same behavior if you LTO as when you didn't - so we
preserve
>> things in the CU, and then do the specified thing per-CU. (eg: emission
>> kind (gmlt/limited/full), split-debug-inlining,
debug-info-for-profiling,
>> etc)
>>
>> In this case if, say, you're on an LLDB platform but prefer to use
GDB (&
>> happen to use Gold and gdb-index on 32 bit and LTO) - so you pass the
>> debugger tuning=gdb flag to your compiles, it wouldn't disable the
feature,
>> because the flag would be consulted at CodeGen time and not at compile
time
>> when you passed the tuning flag. (oh, actually the tuning is passed
down
>> via TM Options anyway, not on the CU - so I guess you wouldn't be
getitng
>> the tuning at all today because that flag already doesn't follow
the per-CU
>> model... well I guess that makes it easier)
>>
>>
>> Setting it via IR/metadata is adding a bunch of complexity just to work
>> around a gold bug.  Not worth it IMO.
>>
>>
>> Fair enough. Any thoughts on whether it should be a frontend-visible
>> option or just a cc1 or even backend option? I'll probably just
leave it as
>> a backend option at this point, could even implement the defaulting
down in
>> LLVM without any changes to Clang, I suppose...
>>
>>
>> --paulr
>>
>>
>>
>> *From:* David Blaikie [mailto:dblaikie at gmail.com]
>> *Sent:* Monday, August 07, 2017 6:58 PM
>> *To:* llvm-dev; Robinson, Paul; Adrian Prantl; Eric Christopher; Nico
>> Weber
>> *Subject:* DWARF: Ranges base address specifier entries &
Gold's
>> gdb-index 32 bit bug
>>
>>
>>
>> Context:
>>
>> In r309526 (with a followup fix in r309529) I implemented the use of
>> DWARF's debug_ranges base address specifier entries to reduce the
number of
>> object file relocations needed for debug_ranges*.
>>
>> * in a particular binary internally, an optimized build had a 70%
>> reduction in debug_ranges.reloc, a 16% decrease in total object size
(with
>> compressed debug info and fission)
>>
>> Nico noted that this broke the Chromium build (
>> https://bugs.llvm.org/show_bug.cgi?id=34007 ). Turns out GNU Binutils
>> Gold has a bug (I've reported it as:
>> https://sourceware.org/bugzilla/show_bug.cgi?id=21894 - wouldn't
expect
>> much action on it, though) where it fails to parse a base address
specifier
>> entry when trying to build gdb-index (-Wl,-gdb-index) for a 32 bit
build.
>>
>> To address this in the short term, I added a flag that disables this
>> feature by default for now.
>>
>> So, two questions:
>>
>> 1) What does everyone reckon the best solution to this is?
>>   * Keep the feature disabled by default - with a flag to enable it for
>> those who want it/can use it
>>   * Enable the feature everywhere - with a flag to disable it
>>   * Conditionally disable (or conditionally enable) the feature, with a
>> flag to enable/disable it
>>     * best condition we probably have is "disabled if 32 bit and
>> -ggnu-pubnames" (gnu-pubnames are necessary for efficient building
of
>> gdb-index, and isn't the default)
>> 2) How to implement the flag:
>>   * backend option as it is currently (-mllvm
>> -use-dwarf-ranges-base-address-specifier (maybe drop
"-specifier" to make
>> it a bit more terse))
>>   * clang option that maps to
>>     * backend option (as currently)
>>     * MCOption (programmatic/C++ API rather than command line of a
>> backend option)
>>     * LLVM IR attribute on the CU (this feature could be supported on a
>> per-CU basis easiyl enough, thus survive LTO, etc, exactly as the user
>> requested (but really if the user requested this on at least one CU,
they
>> probably might as well have it on all their CUs))
>>     * LLVM IR module metadata
>>
>>
>> Any ideas/thoughts/other aspects?
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20181107/357a8703/attachment.html>

Flag's in, off by default, -fdebug-ranges-base-address - if anyone feels
like taking on/advocating for making this on by default under some subset
of conditions (non-32 bit builds (to avoid the gold 32 bit gdb-index bug),
or some other slice), I'm happy to help with that, but adding a flag was
the thing I was trying to avoid - so since it needs/has a flag, that's
enough for me & I'll switch Google over to passing that flag once we
pick
up the changes.

- Dave

On Wed, Nov 7, 2018 at 5:13 PM David Blaikie <dblaikie at gmail.com>
wrote:
> Since it seems like there's unlikely to be a good heuristic that
wouldn't
> potentially need to be overridden by some reasonable use cases, we'll
need
> a flag one way or the thore - so let's add the flag & if someone
wants to
> push on the defaults (I'll probably just leave it as default off
without
> any heuristic for turning it on by default for any cases) they can.
>
> So I've sent out a couple of reviews for the LLVM and Clang parts of
> adding a flag to opt-in to this functionality:
>
> Clang: https://reviews.llvm.org/D54243
> LLVM: https://reviews.llvm.org/D54242
>
> Main thing is just to bikeshed the naming of the metadata and flag.
>
> Thanks everyone!
> - Dave
>
>
> On Tue, Aug 8, 2017 at 10:57 AM David Blaikie <dblaikie at gmail.com>
wrote:
>
>> On Tue, Aug 8, 2017 at 10:50 AM Robinson, Paul <paul.robinson at
sony.com>
>> wrote:
>>
>>> Can gdb handle these?  i.e. is it just gold that has the problem?
>>>
>>
>> Yep, it's just gold when it's building the gdb-index (an
accelerator
>> table for GDB)
>>
>>
>>>   Conditioning on debugger tuning when it's not the debugger
that has
>>> the problem… icky.
>>>
>>
>> It does. Though to a lesser degree even gnu-pubnames isn't a
perfect
>> signal. One, I think, could use that even without a linker-generated
>> gdb-index - presumably GDB's still faster by reading gnu-pubnames
than by
>> reading the raw DIEs to do lookup.
>>
>>
>>>   If that's the case then asking people to add –ggnu-pubnames
seems
>>> okay.  Assuming that gets handed down through IR/metadata then the
LTO
>>> problem goes away too.
>>>
>>
>> gnu-pubnames has such problems (it's done as a backend option,
passed
>> down from the clang driver via -mllvm -generate-gnu-dwarf-pub-sections,
but
>> they're already there... - mostly the discussion around flags was
"If we're
>> adding a new one, what's the right tradeoff here for a workaround
flag
>> compared to the platonic ideals of CU-based flags as much as
possible") but
>> it's already that way...
>>
>> - Dave
>>
>>
>>> --paulr
>>>
>>>
>>>
>>> *From:* David Blaikie [mailto:dblaikie at gmail.com]
>>> *Sent:* Tuesday, August 08, 2017 10:59 AM
>>> *To:* Robinson, Paul; llvm-dev; Adrian Prantl; Eric Christopher;
Nico
>>> Weber
>>> *Subject:* Re: DWARF: Ranges base address specifier entries &
Gold's
>>> gdb-index 32 bit bug
>>>
>>>
>>>
>>> Adrian: any thoughts? Has LLDB been fixed to support this yet?
>>>
>>> On Tue, Aug 8, 2017 at 6:33 AM Robinson, Paul <paul.robinson at
sony.com>
>>> wrote:
>>>
>>> My inclination would be to use "disable if 32-bit and
–ggnu-pubnames" as
>>> the default,
>>>
>>>
>>> Unfortunately Nico points out that Chrome doesn't currently use
>>> -ggnu-pubnames :/ So to continue to work "out of the box"
we'd have to
>>> broaden the surface to "disable if 32 bit and targeting
gdb".
>>>
>>> -ggnu-pubnames makes gdb-index creation much more efficient by
allowing
>>> gdb-index generation not to have to parse most of the DWARF DIEs,
and just
>>> create the index straight from a table. (this is necessary for
correctness
>>> in -gsplit-dwarf mode (where the DIEs are not available at link
time as
>>> they're in the .dwo not the .o), and an efficiency gain
otherwise)
>>>
>>> If "disable if 32 bit and targeting gdb" is overly broad
(I feel it kind
>>> of is, but I'm not wedded to it) we could require those working
around the
>>> gold bug to add -ggnu-pubnames to add a bit more of a specific hint
that
>>> they're intending to build gdb-index from these objects.
>>>
>>>
>>> and have a flag.
>>>
>>> This would be determined in DwarfDebug, right?
>>>
>>>
>>> Yes
>>>
>>>
>>> If the check is in CodeGen then I'd think it would be
insensitive to LTO.
>>>
>>>
>>> There's some desire that LTO works as-if it was non-LTO. So if
you pass
>>> flags (eg: debugger tuning flags, in this case) to each separate
compile,
>>> you get the same behavior if you LTO as when you didn't - so we
preserve
>>> things in the CU, and then do the specified thing per-CU. (eg:
emission
>>> kind (gmlt/limited/full), split-debug-inlining,
debug-info-for-profiling,
>>> etc)
>>>
>>> In this case if, say, you're on an LLDB platform but prefer to
use GDB
>>> (& happen to use Gold and gdb-index on 32 bit and LTO) - so you
pass the
>>> debugger tuning=gdb flag to your compiles, it wouldn't disable
the feature,
>>> because the flag would be consulted at CodeGen time and not at
compile time
>>> when you passed the tuning flag. (oh, actually the tuning is passed
down
>>> via TM Options anyway, not on the CU - so I guess you wouldn't
be getitng
>>> the tuning at all today because that flag already doesn't
follow the per-CU
>>> model... well I guess that makes it easier)
>>>
>>>
>>> Setting it via IR/metadata is adding a bunch of complexity just to
work
>>> around a gold bug.  Not worth it IMO.
>>>
>>>
>>> Fair enough. Any thoughts on whether it should be a
frontend-visible
>>> option or just a cc1 or even backend option? I'll probably just
leave it as
>>> a backend option at this point, could even implement the defaulting
down in
>>> LLVM without any changes to Clang, I suppose...
>>>
>>>
>>> --paulr
>>>
>>>
>>>
>>> *From:* David Blaikie [mailto:dblaikie at gmail.com]
>>> *Sent:* Monday, August 07, 2017 6:58 PM
>>> *To:* llvm-dev; Robinson, Paul; Adrian Prantl; Eric Christopher;
Nico
>>> Weber
>>> *Subject:* DWARF: Ranges base address specifier entries &
Gold's
>>> gdb-index 32 bit bug
>>>
>>>
>>>
>>> Context:
>>>
>>> In r309526 (with a followup fix in r309529) I implemented the use
of
>>> DWARF's debug_ranges base address specifier entries to reduce
the number of
>>> object file relocations needed for debug_ranges*.
>>>
>>> * in a particular binary internally, an optimized build had a 70%
>>> reduction in debug_ranges.reloc, a 16% decrease in total object
size (with
>>> compressed debug info and fission)
>>>
>>> Nico noted that this broke the Chromium build (
>>> https://bugs.llvm.org/show_bug.cgi?id=34007 ). Turns out GNU
Binutils
>>> Gold has a bug (I've reported it as:
>>> https://sourceware.org/bugzilla/show_bug.cgi?id=21894 -
wouldn't expect
>>> much action on it, though) where it fails to parse a base address
specifier
>>> entry when trying to build gdb-index (-Wl,-gdb-index) for a 32 bit
build.
>>>
>>> To address this in the short term, I added a flag that disables
this
>>> feature by default for now.
>>>
>>> So, two questions:
>>>
>>> 1) What does everyone reckon the best solution to this is?
>>>   * Keep the feature disabled by default - with a flag to enable it
for
>>> those who want it/can use it
>>>   * Enable the feature everywhere - with a flag to disable it
>>>   * Conditionally disable (or conditionally enable) the feature,
with a
>>> flag to enable/disable it
>>>     * best condition we probably have is "disabled if 32 bit
and
>>> -ggnu-pubnames" (gnu-pubnames are necessary for efficient
building of
>>> gdb-index, and isn't the default)
>>> 2) How to implement the flag:
>>>   * backend option as it is currently (-mllvm
>>> -use-dwarf-ranges-base-address-specifier (maybe drop
"-specifier" to make
>>> it a bit more terse))
>>>   * clang option that maps to
>>>     * backend option (as currently)
>>>     * MCOption (programmatic/C++ API rather than command line of a
>>> backend option)
>>>     * LLVM IR attribute on the CU (this feature could be supported
on a
>>> per-CU basis easiyl enough, thus survive LTO, etc, exactly as the
user
>>> requested (but really if the user requested this on at least one
CU, they
>>> probably might as well have it on all their CUs))
>>>     * LLVM IR module metadata
>>>
>>>
>>> Any ideas/thoughts/other aspects?
>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20181113/26a15480/attachment.html>