[llvm-dev] Seeking clarification and way forward on limited scope variables.

Tomar, Sourabh Singh via llvm-dev llvm-dev at lists.llvm.org
Sun Jun 21 04:25:37 PDT 2020


Gentle Ping!

From: Tomar, Sourabh Singh
Sent: Tuesday, June 9, 2020 5:22 PM
To: David Blaikie <dblaikie at gmail.com>; Vedant Kumar <vedant_kumar at apple.com>; aprantl at apple.com
Cc: llvm-dev at lists.llvm.org; George, Jini Susan <JiniSusan.George at amd.com>
Subject: RE: [llvm-dev] Seeking clarification and way forward on limited scope variables.

Hi Everyone,

We’ve reached to a working prototype (With some caveats mentioned at the end) for this problem, I need to have your opinion on this before we move further WRT.


  *   @David Blaikie<mailto:dblaikie at gmail.com>I have no idea what the LLVM DWARF representation for this would look like - short of making even more fine-grained scopes in the DILexicalScope hierarchy, which sounds really expensive from a memory perspective. That's really where I worry that the cost to this feature might outweigh the benefit (& might be why no one's done this in the past) - but data should tell us that. As much as in-tree development is preferred, this might be the sort of thing worth prototyping out of tree first to see if it can be made viable before adding the complexity to the LLVM project proper - but I'm open to ideas/suggestions.


Problem breakdown:
End representation at (Object or ASM level):

  *   Need a label(Symbol) to extract **exact location(Or address at which the variable defined(or allocated))**, for which an Machine Instruction is needed(in this case DBG_VALUE leveraged).
  *   Take the difference of this label(address) and **low_pc** of the **Lexical Block**, synthesize the “DW_AT_start_scope” using this difference(Offset) attach this attribute to the Scoped variable.


Implementation LLVM-to-DWARF:

  *   Implementation strategy is a bit conservative in a sense, I didn’t introduce any new metadata or new instruction(IR or MachineInstr).



  *   For the unoptimized case, this variable will be stack allocated(with Frame Index per alloca instruction) and dbg.declare is used.
  *   Lower this dbg.declare(associated with the variable of interest) to generate a DBG_VALUE instruction for this dbg.declare intrinsic(in both FastISel and SelectionDAGISel Instruction Selection) Distinguishing with other local variables(if present) is accomplished based on the premises of enclosing scope(i.e Lexical Block).
  *   Request a label(Symbol) before this instruction(DBG_VALUE). This label will be use later for associating the variable with it and calculating offset to be inserted as DW_AT_start_scope.
  *   Extend DbgVariable to contain 2 symbols
     *   VarSym -> Symbol associated with this variable
     *   ScopeBeginSym -> Symbol associated with the enclosing Lexical Scope LowPC
  *   When constructing actual DIE and DW_AT_start_scope use these symbols difference to emit as offset within the variable DIE. Again distinguishing with other variables is accomplished using above mentioned simple check.


Resultant DWARF after this for the test case:
0x0000006d:     DW_TAG_lexical_block
                  DW_AT_low_pc  (0x00000000002016d1)
                  DW_AT_high_pc (0x0000000000201729)

0x0000007a:       DW_TAG_variable
                    DW_AT_location      (0x00000000:
                       [0x00000000002016e8, 0x0000000000201731): DW_OP_breg6 RBP-24)
                    DW_AT_name  ("Local")
                    DW_AT_decl_file     ("test.c")
                    DW_AT_decl_line     (7)
                    DW_AT_type  (0x0000008b "int")
                    DW_AT_start_scope   (0x00000017)

Summarizing results:

  *   Debugging behavior as expected(of course with a small change in GDB(as mentioned in very first mail)).
  *   No regressions found on trunk!
  *   Cost to benefit: Compile time costs and debug info size increment not yet evaluated.


Concerns still need to be addressed:

  *   DBG_VALUE instructions in most cases end up creating location lists(due to their special handling), unfortunately in this case too apart from “DW_AT_start_scope” Location lists are also

getting attached to the variable. Is this Okay ?


Could you guys please let us know your thoughts/comments on the overall approach taken here.

Note:
This sort of bug(improvement based on DW_AT_start_scope) is also reported(By someone else, not me) in GCC here:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93844

Thank You!
Sourabh.

From: David Blaikie <dblaikie at gmail.com<mailto:dblaikie at gmail.com>>
Sent: Thursday, April 16, 2020 1:46 AM
To: Vedant Kumar <vedant_kumar at apple.com<mailto:vedant_kumar at apple.com>>
Cc: Sourabh Singh Tomar <sourav0311 at gmail.com<mailto:sourav0311 at gmail.com>>; Achra, Nitika <Nitika.Achra at amd.com<mailto:Nitika.Achra at amd.com>>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>; Kumar N, Bhuvanendra <Bhuvanendra.KumarN at amd.com<mailto:Bhuvanendra.KumarN at amd.com>>; George, Jini Susan <JiniSusan.George at amd.com<mailto:JiniSusan.George at amd.com>>; Sharma, Alok Kumar <AlokKumar.Sharma at amd.com<mailto:AlokKumar.Sharma at amd.com>>; Tomar, Sourabh Singh <SourabhSingh.Tomar at amd.com<mailto:SourabhSingh.Tomar at amd.com>>; Adrian Prantl <aprantl at apple.com<mailto:aprantl at apple.com>>
Subject: Re: [llvm-dev] Seeking clarification and way forward on limited scope variables.

[CAUTION: External Email]


On Wed, Apr 15, 2020 at 11:52 AM Vedant Kumar <vedant_kumar at apple.com<mailto:vedant_kumar at apple.com>> wrote:
Hi Sourabh,

Thanks for raising this issue. To answer your question, (afaik) there isn’t anyone working on DW_AT_start_scope support in tree. We’re looking for a solution to this problem for Swift debugging, where it's important not to make a debug location for a variable available until its (guaranteed) initialization is complete.

If at all possible, I’d /much/ rather we use the existing location list machinery to solve this problem. Fundamentally, we’re looking for a way to express when a location for a variable becomes available, and location lists give us that already.

To test this out, I took the IR for your test case, replaced all calls to “dbg.declare(…)” with “dbg.value(…, DW_OP_deref)”, and compiled with `-g -O0 -mllvm -fast-isel=false` (apparently FastISel doesn’t know what to do with frame-index dbg.values, but this is simple to fix). This seemed to work pretty well, and the DWARF looked legit:

```
0x0000005f:     DW_TAG_variable
                  DW_AT_location        (DW_OP_fbreg -20)
                  DW_AT_name    ("Local")

0x0000006d:     DW_TAG_lexical_block
                  DW_AT_low_pc  (0x0000000100000f4f)
                  DW_AT_high_pc (0x0000000100000f84)

0x0000007a:       DW_TAG_variable
                    DW_AT_location      (0x00000000
                       [0x0000000100000f63,  0x0000000100000f8c): DW_OP_breg6 RBP-24)
                    DW_AT_name  ("Local”)
```
I did find one lldb bug where it didn’t know to look up a variable in the parent scope when stopped at your second printf() call (line 6). But that's a general bug: we’d have to fix it even if we used DW_AT_start_scope.

The upshot of sticking to location lists is that fixing any bugs we find improves optimized debugging workflows. And Swift -Onone debugging workflows as well, since Swift runs certain mandatory optimizations which can make the DWARF at -Onone fairly complex.

I think these are different problems, and that location lists can't be used to express when the scope of a variable starts - Sourabh's example at the end shows why: that the debugger wouldn't search outside the scope, it would instead (correctly) report the variable's value as unavailable.

"I did find one lldb bug where it didn’t know to look up a variable in the parent scope when stopped at your second printf() call (line 6)."

I don't think that's not a bug, though/that change would be incorrect in general (& no way to differentiate the correct and incorrect places without scope_start), for example:


void f1();
int i;

void f2() {

  int i = 3;

  f1();

  int j;

  i = j;

  f1();

}

If you search into outer scopes whenever a location list doesn't cover an address - then at the second call to 'f1' the debugger would interpret 'i' as referring to global 'i' when it should be referring to local 'i' just with no known value. Printing global 'i' could be quite confusing/misleading.

best,
vedant

On Apr 15, 2020, at 9:53 AM, David Blaikie via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

As always, concerned about the size growth in object files this might produce - though looks like the DWARF spec avoids the worst of this in unoptimized code by using an offset relative to the start of the enclosing scope, so it doesn't require a relocation in that case.

I have no idea what the LLVM DWARF representation for this would look like - short of making even more fine-grained scopes in the DILexicalScope hierarchy, which sounds really expensive from a memory perspective. That's really where I worry that the cost to this feature might outweigh the benefit (& might be why no one's done this in the past) - but data should tell us that. As much as in-tree development is preferred, this might be the sort of thing worth prototyping out of tree first to see if it can be made viable before adding the complexity to the LLVM project proper - but I'm open to ideas/suggestions.

On Wed, Apr 15, 2020 at 3:15 AM Sourabh Singh Tomar <sourav0311 at gmail.com<mailto:sourav0311 at gmail.com>> wrote:
Hello Everyone,

I need to have your thoughts on this.

Consider the following test case --
-------------------------------------------
 1    int main(int Argc, char **Argv) {
  2         int Local = 6;
  3         printf("%d\n",Local);
  4
  5         {
  6         printf("%d\n",Local);
  7         int Local = 7;
  8         printf("%d\n",Local);
  9         }
 10
 11         return 0;
 12  }
--------------------------------------------
When compiled in debug mode with compilers including (trunk gcc and trunk clang) and debugging with GDB at Line No.6, the following behavior is observed
Breakpoint 1, main (Argc=1, Argv=0x7fffffffe458) at MainScope.c:6
6               printf("%d\n",Local);
(gdb) print Local
$1 = 2102704   -- some Garbage value,
(gdb) info addr Local
Symbol "Local" is a variable at frame base reg $rbp offset 0+-24.   -- This is location of *Local* declared inside scope, but as you may notice that the variable being referred here is from the outer scope.

This problem persists with both GDB and LLDB. Since we have entered the Lexical Scope and when we try to print value of *Local*,  it will look into the *current scope* and fetch the value if the variable exists in scope(in case variable doesn't exist, GDB searches for it in the outer scope).

This is regardless of whether the variable has actually came into scope(or actually defined) at Line No. 7. Since DWARF already defined the location(on stack) which will be valid for the lifetime of the variable, contrary to when the variable is actually defined(or allocated) which is in this case Line No. 7.
---------------------------------------------
  0x0000006d:     DW_TAG_lexical_block
                  DW_AT_low_pc  (0x00000000002016d1)
                  DW_AT_high_pc (0x000000000020170b)
0x0000007a:       DW_TAG_variable
                    DW_AT_location      (DW_OP_fbreg -24)
                    DW_AT_name  ("Local")
                    DW_AT_decl_file     ("MainScope.c")
                    DW_AT_decl_line     (7)
                    DW_AT_type  (0x0000008a "int")
----------------------------------------------

The DWARF specification provides the DW_AT_start_scope attribute to deal with this issue (Sec 3.9 Declarations with Reduced Scope DWARFv5). This attribute aims at limiting the scope of variables within the lexical scope in which it is defined to from where it has been declared/ defined.

In order to fix this issue, we want to modify llvm so that DW_AT_start_scope is emitted for the variable in the inner block (in the above example). This limits the scope of the inner block variable to start from the point of its declaration.

For POC, we inserted DW_AT_start_scope in this inner *Local* variable, resultant dwarf after this.
-----------------------------
0x0000006d:     DW_TAG_lexical_block
                  DW_AT_low_pc  (0x00000000002016d1)
                  DW_AT_high_pc (0x000000000020170b)
0x0000007a:       DW_TAG_variable
                     DW_AT_start_scope   (0x17) -- restricted within a subset(starting from the point of definition(specified as an offset)) of entire ranges covered by Lex Block.
                    DW_AT_location      (DW_OP_fbreg -24)
                    DW_AT_name  ("Local")
                    DW_AT_decl_file     ("MainScope.c")
                    DW_AT_decl_line     (7)
                    DW_AT_type  (0x00000092 "int")
----------------------------


We also modified ‘gdb’ to interpret DW_AT_start_scope so that the scope of the variable is limited from the PC where the value of DW_AT_start_scope is. If the debugger is stopped at a point within the same lexical block but at a PC before DW_AT_start_scope, then gdb follows the normal search mechanism of searching in consecutive super blocks till it gets a match or it reaches the global block. After the modification,  GDB is able to correctly show the value *6* in our example.


After incorporating changes --
  Breakpoint 1, main (Argc=1, Argv=0x7fffffffe458) at MainScope.c:6
6               printf("%d\n",Local);
(gdb) print Local
$1 = 6 --- Value retrieved from outer scope
(gdb) info addr Local
Symbol "Local" is a variable at frame base reg $rbp offset 0+-20.

Could you guys please let us know your thoughts or suggestions on this? Was/ Is there is an existing effort already going on to deal with this problem?

Even though location lists can be used to deal with this scenario, in my experience, location lists are emitted at higher optimization levels, and with the usage of location lists in this example, gdb prints out <optimized out> (as expected) if it is stopped at a PC in the same lexical block but before the point of declaration of the local variable.

Thank You,
Sourabh Singh Tomar.
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev<https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fllvm-dev&data=02%7C01%7Csourabhsingh.tomar%40amd.com%7Ce380b64c3f054f9c7ead08d7e179c87b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225785481382842&sdata=TxcGddPstXlJIBDdarp6KCKd4OqE2YJkJF2JBrSnANc%3D&reserved=0>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200621/a3a0022b/attachment-0001.html>


More information about the llvm-dev mailing list