Improving the quality of debug locations / DbgValueHistoryCalculator

Hi all,

An early implementation of extending debug ranges and providing multiple location support is done here: http://reviews.llvm.org/D11933

Design document: https://docs.google.com/document/d/1noDVWTvTWBdYdweICPBwvwyt8QvX4KHl7j3XKNSg1nE/edit?usp=sharing

Hello,

Regarding the problem of debug range for optimized code.
Currently a DEBUG_VALUE will be inserted after the vregX
DEBUG_VALUE are only valid until the end of the current MachineBasicBlock. That’s the main problem.
Why not simply iterate over all uses of vregX and insert an DEBUG_VALUE in all the MachineBasicBlocks where vregX is used. (pre regalloc)

I prototyped a small pass to do that and at first it seems to improve .debug_loc range validity and enhance the debugging experience for optimized code.

Hello,

Regarding the problem of debug range for optimized code.
Currently a DEBUG_VALUE will be inserted after the vregX
DEBUG_VALUE are only valid until the end of the current MachineBasicBlock. That’s the main problem.
Why not simply iterate over all uses of vregX and insert an DEBUG_VALUE in all the MachineBasicBlocks where vregX is used. (pre regalloc)

I prototyped a small pass to do that and at first it seems to improve .debug_loc range validity and enhance the debugging experience for optimized code.

This thread is almost a year old now :slight_smile:
Have you looked at the LiveDebugValues pass that resulted from this conversation since?

– adrian

Yes I looked at LiveDebugValues and it is an improvement. But in some cases it doesn’t give me a debug value for variables that are obviously in a register.

Good debug experience for optimized code is critical for my (out-of-tree) target.

So I was looking at improving LiveDebugValues and I realized that unilaterally extending DEBUG_VALUE in ssa form was simpler and it gave me good results. I know I am lacking specifics here but I just wonder if its a good approach.

Hello,

Regarding the problem of debug range for optimized code.
Currently a DEBUG_VALUE will be inserted after the <def>vregX
DEBUG_VALUE are only valid until the end of the current MachineBasicBlock. That's the main problem.
Why not simply iterate over all uses of vregX and insert an DEBUG_VALUE in all the MachineBasicBlocks where vregX is used. (pre regalloc)

I prototyped a small pass to do that and at first it seems to improve .debug_loc range validity and enhance the debugging experience for optimized code.

The problem that I see with this approach is that DEBUG_VALUEs are only valid until the next DEBUG_VALUE that describes the same variable. How does your pass handle:

  BB0:
    DEBUG_VALUE “x”, vreg0

  > >
  > BB1:
  > DEBUG_VALUE “x”, vreg1
  > >
  \_____BB2:
    > >
    >____/
  BB2:
    vreg0 // still not clobbered here.

Does it insert a DEBUG_VALUE “x”, vreg0 into BB2?

-- adrian

Good point.

Currently yes a DEBUG_VALUE “x”, vreg0 will be added in BB2. Now I realize this might be wrong in some (corner?) cases where vreg0 no longer refer to “x”

My fix would be to propagate the DEBUG_VALUE only if “x” is associated with only a single virtual register.
BTW, my goal is to generally improve DEBUG_VALUE for optimized code, not make it 100% correct.

Good point.

Currently yes a DEBUG_VALUE “x”, vreg0 will be added in BB2. Now I realize this might be wrong in some (corner?) cases where vreg0 no longer refer to “x”

My fix would be to propagate the DEBUG_VALUE only if “x” is associated with only a single virtual register.
BTW, my goal is to generally improve DEBUG_VALUE for optimized code, not make it 100% correct.

I hold the (perhaps somewhat extreme) position that having debug info the may be correct is worse than having no debug info at all, because it means that — in the end — you then cannot trust anything reported by the debugger. (There are some debatable corner cases, for example, in contrast to the debugger, the user may know over which path the control flow arrived at the current break point, but these are far in between.)

That said, I’m very open to improvements to the quality and correctness, so please don’t feel discouraged :slight_smile:

– adrian

In retrospect I totally agree with you. I am looking at LiveDebugValue again to see if I can improve certain specific cases.

The most obvious place where it is lacking at the moment is that it only supports DBG_VALUEs in registers. Adding support for constant values, memory locations, and fp constants would be a big win!

thanks,
Adrian

FYI, Wolfgang has been working on a patch for making LiveDebugValues handle spills and reloads. Hopefully it will be presentable upstream in the not-too-distant future.

–paulr

Here is a specific case that make the debugging experiences degraded on my target:

This is a loop simplified CFG:

BB#0:
%R5 = OR_rr %R0, %R49 // this is %R5 only def.
DBG_VALUE %R5, %noreg, !“argc”, <!18>; line no:4
Successors according to CFG: BB#1

BB#1:
Live Ins: %R5
Predecessors according to CFG: BB#0 BB#3

Successors according to CFG: BB#2 BB#4

BB#2:
Live Ins: %R5
Predecessors according to CFG: BB#1

Successors according to CFG: BB#3

BB#3:
Live Ins: %R5
Predecessors according to CFG: BB#2

Successors according to CFG: BB#1

BB#4:
Predecessors according to CFG: BB#1

Its obvious to me that the DEBUG_VALUE %R5, %noreg, !“argc” should be propagated to BB#1, BB#2 and BB#3.

LiveDebugValue will currently not handle this case. The propagation will not be done for BB#1 because one of its predecessor BB3# doesn’t have a DEBUG_VALUE %R5. But R5 is still guaranteed to correspond to argc.

I am investigating how to extend LiveDebugValue for a case like this.

Here is a specific case that make the debugging experiences degraded on my target:

This is a loop simplified CFG:

BB#0:
%R5 = OR_rr %R0, %R49 // this is %R5 only def.
DBG_VALUE %R5, %noreg, !“argc”, <!18>; line no:4
Successors according to CFG: BB#1

BB#1:
Live Ins: %R5
Predecessors according to CFG: BB#0 BB#3

Successors according to CFG: BB#2 BB#4

BB#2:
Live Ins: %R5
Predecessors according to CFG: BB#1

Successors according to CFG: BB#3

BB#3:
Live Ins: %R5
Predecessors according to CFG: BB#2

Successors according to CFG: BB#1

BB#4:
Predecessors according to CFG: BB#1

Its obvious to me that the DEBUG_VALUE %R5, %noreg, !“argc” should be propagated to BB#1, BB#2 and BB#3.

LiveDebugValue will currently not handle this case. The propagation will not be done for BB#1 because one of its predecessor BB3# doesn’t have a DEBUG_VALUE %R5. But R5 is still guaranteed to correspond to argc.

I am investigating how to extend LiveDebugValue for a case like this.

That sounds like a bug in LiveDebugValues, or there must be something missing from your example. LiveDebugValues should propagate the DBG_VALUE in this case. Are you sure that R5 isn’t clobbered or regmask’ed by any instruction, and that there is no conflicting DBG_VALUE?

– adrian

No, R5 is not clobbered or having a conflicting DBG_VALUE

That sounds more like a missing feature in LiveDebugValue: There is no DBG_VALUE for R5 in BB#3.
LiveDebugValue will only join when all predecessors have the same DBG_VALUE.

No, R5 is not clobbered or having a conflicting DBG_VALUE

That sounds more like a missing feature in LiveDebugValue: There is no
DBG_VALUE for R5 in BB#3.
LiveDebugValue will only join when all predecessors have the same
DBG_VALUE.

True. And currently, LiveDebugValues cannot propagate <R5, argc> to BB1 as
join fails with one of the predecessor - BB3.

Here is my thought about handling similar cases:
I would regard BB1 to be special because join fails to propagate the debug
value which is dependent on itself through a cycle. Such blocks need to be
identified and handled slightly differently. I think one way to identify
them is by using dominator info - look for blocks that dominate its
predecessors and propagate debug values from other predecessors
unconditionally (also note down the propagated values). This is to be done
early during OutLocs initialization. Further, transfer-join would happen
normally.

If join() fails for these special blocks, the propagated debug values need
to be reverted (followed by another round of convergence for its
successors) and surely this will need extra book-keeping.