llvm dev - Jun 2018 - [RFC] Removing debug locations from ConstantSDNodes

> On Jun 19, 2018, at 6:36 PM, Reid Kleckner <rnk at google.com> wrote:
> 
> On Tue, Jun 19, 2018 at 5:46 PM Vedant Kumar <vsk at apple.com
<mailto:vsk at apple.com>> wrote:
> Someone (Reid?) mentioned that we could try sinking constants to their
point of first use as an alternative, and (IIUC) create new nodes with distinct
DebugLocs for each use of a constant. I don't recall the details of that
alternative clearly. Based on my (likely incorrect) understanding of it,
dropping locations from constants outright might be simpler.
> 
> Our use case was in fastisel, so things are different. I don't think my
solution will help you.
> 
> In our case, users were complaining about code like this:
> volatile int do_something;
> void f() {
>   ++do_something;
>   foo(1, 2, 3);
>   ++do_something;
> }
> 
> We'd generate locations like:
>   .loc line 1
>   incl do_something(%rip)
>   movl $1, %ecx
>   movl $2, %edx
>   movl $3, %r8d
>   .loc line 2 # line 2 starts here, instead of 3 instructions earlier
>   callq foo
>   .loc line 3
>   incl do_something(%rip)
> 
> Our users really wanted the line table entry for line 2 to include the
constant materialization code for some VS debugger feature.
Got it. (For others following along, there's more discussion about this
debugger feature in the commit description for r327581.)

> I think if you remove locations from ConstantSDNodes, you might want to add
a late pass that propagates source locations backwards onto location-less
instructions. This would also avoid some special cases when a basic block starts
with an instruction that lacks location information. See
CodeViewDebug::beginInstruction and DwarfDebug::beginInstruction for what we do
today.
Interesting, I hadn't considered doing this.

What happens in the special case where a block starts with a location-less
instruction? Ah, I see that CodeViewDebug::beginInstruction does a forward scan
to find the first instruction with location. This can fail though: there might
not be such an instruction, in which case... I assume we either apply a line-0
location, or nothing at all, similar to Dwarf::beginInstruction?

It's a bit unclear to me what the benefits of a late pass that
backwards-propagated locations would be. I suppose we'd have fewer forward
scans in CodeViewDebug::beginInstruction, so is the benefit a compile-time win
(which would offset the cost of a late pass)? Would the final debug info quality
would be better in some way I'm just missing?

thanks!
vedant
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180620/b1d5f899/attachment.html>

DwarfDebug::beginInstruction sets the location to line 0 because otherwise the
location is implicitly the same location as the last instruction of the
physically preceding block.  That location is often completely unrelated to
what's at the top of the new block.  Line 0 isn't great, but at least
it's not a complete lie.
--paulr

From: vsk at apple.com [mailto:vsk at apple.com]
Sent: Wednesday, June 20, 2018 1:48 PM
To: Reid Kleckner
Cc: Robinson, Paul; llvm-dev; Justin Bogner; David Li; David Blaikie
Subject: Re: [RFC] Removing debug locations from ConstantSDNodes

On Jun 19, 2018, at 6:36 PM, Reid Kleckner <rnk at google.com<mailto:rnk
at google.com>> wrote:

On Tue, Jun 19, 2018 at 5:46 PM Vedant Kumar <vsk at apple.com<mailto:vsk
at apple.com>> wrote:
Someone (Reid?) mentioned that we could try sinking constants to their point of
first use as an alternative, and (IIUC) create new nodes with distinct DebugLocs
for each use of a constant. I don't recall the details of that alternative
clearly. Based on my (likely incorrect) understanding of it, dropping locations
from constants outright might be simpler.

Our use case was in fastisel, so things are different. I don't think my
solution will help you.

In our case, users were complaining about code like this:
volatile int do_something;
void f() {
  ++do_something;
  foo(1, 2, 3);
  ++do_something;
}

We'd generate locations like:
  .loc line 1
  incl do_something(%rip)
  movl $1, %ecx
  movl $2, %edx
  movl $3, %r8d
  .loc line 2 # line 2 starts here, instead of 3 instructions earlier
  callq foo
  .loc line 3
  incl do_something(%rip)


Our users really wanted the line table entry for line 2 to include the constant
materialization code for some VS debugger feature.

Got it. (For others following along, there's more discussion about this
debugger feature in the commit description for r327581.)



I think if you remove locations from ConstantSDNodes, you might want to add a
late pass that propagates source locations backwards onto location-less
instructions. This would also avoid some special cases when a basic block starts
with an instruction that lacks location information. See
CodeViewDebug::beginInstruction and DwarfDebug::beginInstruction for what we do
today.

Interesting, I hadn't considered doing this.

What happens in the special case where a block starts with a location-less
instruction? Ah, I see that CodeViewDebug::beginInstruction does a forward scan
to find the first instruction with location. This can fail though: there might
not be such an instruction, in which case... I assume we either apply a line-0
location, or nothing at all, similar to Dwarf::beginInstruction?

It's a bit unclear to me what the benefits of a late pass that
backwards-propagated locations would be. I suppose we'd have fewer forward
scans in CodeViewDebug::beginInstruction, so is the benefit a compile-time win
(which would offset the cost of a late pass)? Would the final debug info quality
would be better in some way I'm just missing?

thanks!
vedant
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180620/f5ab5684/attachment.html>

FWIW: Debug information on constants feels odd to me. They are just values not
something that is executed so conceptually I would not expect them to
"happen" at a specific time/place in the program. That said most
numbers are copied into registers or stored into memory and that is of course an
interesting action. So in the original example I would hope to see debug info on
whatever instructions are used to fill the array with values.

That said I'm not familiar with the inner workings of dwarf or other
debugger formats, so it may very well be reasonable to backfill the information
in a late pass to avoid having assembler instructions without debug info as some
people proposed.

- Matthias
> On Jun 20, 2018, at 11:09 AM, via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> DwarfDebug::beginInstruction sets the location to line 0 because otherwise
the location is implicitly the same location as the last instruction of the
physically preceding block.  That location is often completely unrelated to
what's at the top of the new block.  Line 0 isn't great, but at least
it's not a complete lie.
> --paulr
>   <>
> From: vsk at apple.com [mailto:vsk at apple.com] 
> Sent: Wednesday, June 20, 2018 1:48 PM
> To: Reid Kleckner
> Cc: Robinson, Paul; llvm-dev; Justin Bogner; David Li; David Blaikie
> Subject: Re: [RFC] Removing debug locations from ConstantSDNodes
>  
> On Jun 19, 2018, at 6:36 PM, Reid Kleckner <rnk at google.com
<mailto:rnk at google.com>> wrote:
>  
> On Tue, Jun 19, 2018 at 5:46 PM Vedant Kumar <vsk at apple.com
<mailto:vsk at apple.com>> wrote:
> Someone (Reid?) mentioned that we could try sinking constants to their
point of first use as an alternative, and (IIUC) create new nodes with distinct
DebugLocs for each use of a constant. I don't recall the details of that
alternative clearly. Based on my (likely incorrect) understanding of it,
dropping locations from constants outright might be simpler.
>  
> Our use case was in fastisel, so things are different. I don't think my
solution will help you.
>  
> In our case, users were complaining about code like this:
> volatile int do_something;
> void f() {
>   ++do_something;
>   foo(1, 2, 3);
>   ++do_something;
> }
>  
> We'd generate locations like:
>   .loc line 1
>   incl do_something(%rip)
>   movl $1, %ecx
>   movl $2, %edx
>   movl $3, %r8d
>   .loc line 2 # line 2 starts here, instead of 3 instructions earlier
>   callq foo
>   .loc line 3
>   incl do_something(%rip)
> 
> 
> Our users really wanted the line table entry for line 2 to include the
constant materialization code for some VS debugger feature.
>  
> Got it. (For others following along, there's more discussion about this
debugger feature in the commit description for r327581.)
> 
> 
> 
> I think if you remove locations from ConstantSDNodes, you might want to add
a late pass that propagates source locations backwards onto location-less
instructions. This would also avoid some special cases when a basic block starts
with an instruction that lacks location information. See
CodeViewDebug::beginInstruction and DwarfDebug::beginInstruction for what we do
today.
>  
> Interesting, I hadn't considered doing this.
>  
> What happens in the special case where a block starts with a location-less
instruction? Ah, I see that CodeViewDebug::beginInstruction does a forward scan
to find the first instruction with location. This can fail though: there might
not be such an instruction, in which case... I assume we either apply a line-0
location, or nothing at all, similar to Dwarf::beginInstruction?
>  
> It's a bit unclear to me what the benefits of a late pass that
backwards-propagated locations would be. I suppose we'd have fewer forward
scans in CodeViewDebug::beginInstruction, so is the benefit a compile-time win
(which would offset the cost of a late pass)? Would the final debug info quality
would be better in some way I'm just missing?
>  
> thanks!
> vedant
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://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/20180620/4f7e7be5/attachment.html>