[Lldb-commits] [PATCH] D94064: lldb: Add support for printing variables with DW_AT_ranges on DW_TAG_subprograms

[David Blaikie via Phabricator via lldb-commits]

dblaikie added a comment.

In D94064#2485222 <https://reviews.llvm.org/D94064#2485222>, @labath wrote:

> In D94064#2483578 <https://reviews.llvm.org/D94064#2483578>, @dblaikie wrote:
>
>> In D94064#2481925 <https://reviews.llvm.org/D94064#2481925>, @labath wrote:
>>
>>> I don't think that simply setting func_lo_pc to zero will be sufficient to make this work. I would expect this to break in more complicated scenarios (like, even if we just have two of these functions). I think the only reason it works in this case is because for this particular function, it's base address (relative to the CU base) *is* zero.
>>
>> I certainly don't have high confidence in the change, to be sure - but I think it's a bit more robust than that.
>
> Yes, it seems it is, though just by a bit. :)
>
> The thing which is missing is the location list on the variable DIE -- without it, it does not matter which address gets put here (so long as it does not trigger the assertion), as the value is totally unused. The necessity of the location list is kind of obvious from my description of the problem, though even I did not realize when I was writing it that this requires a more elaborate test case.
>
> Adjusting your test case to produce a location list for `i` I got this (probably not minimal) snippet (add -O1 to CFLAGS):
>
>   #define optnone __attribute__((optnone))
>   #define noinline __attribute__((noinline))
>   
>   void optnone noinline stop() {}
>   void optnone noinline consume(int i) {}
>   void noinline f1(int x) {
>     int i = x;
>     stop();
>     consume(i);
>     i = 5;
>     consume(i);
>   }
>   int main() {
>     int j = 12;
>     f1(j);
>     stop();
>   }
>
> Here lldb fails to print the value of `i`, but if I rearrange the code so that the `f1` functions comes first, the value is printed correctly:
>
>   #define optnone __attribute__((optnone))
>   #define noinline __attribute__((noinline))
>   
>   void optnone noinline stop();
>   void optnone noinline consume(int i);
>   void noinline f1(int x) {
>     int i = x;
>     stop();
>     consume(i);
>     i = 5;
>     consume(i);
>   }
>   void optnone noinline stop() {}
>   void optnone noinline consume(int i) {}
>   int main() {
>     int j = 12;
>     f1(j);
>     stop();
>   }

Ah, thanks - think I figured out a representative test & understand better.
I ended up with this:

  __attribute__((nodebug)) volatile int i;
  int main() {
    int var = 3;
    i = 1;
    var = 5;
    i = 2;
  }

By using volatile writes, I was able to get the 'var' variable live range to start at the start of the function (so that `image lookup -v -s main` would render the "var" variable since it's now live at the very start of the function (rather than only after the push instruction)). And use nodebug to reduce the DWARF since the 'i' isn't interesting.

Indeed without the fix you suggested to use "lowest address" rather than zero, this test above would/was failing (running the image lookup command would not show the "var" variable).

Also, I figured out how to run the API tests, and that showed a bunch more failures when using ranges-everywhere (actually a more aggressive version of ranges-everywhere - using it no matter the DWARF version and even when it wouldn't reduce the address pool size (eg: using ranges on a subprogram even when low_pc is an address already in the pool (such as for the start of a CU range))) - and with the change to use the lowest address of the ranges, all those failures now go away.

So *fingers crossed* this is ready.

>>> The purpose of func_lo_pc is pretty weird, but it's essentially used for runtime relocation of location lists within the function -- we take the static "base" of the function, and the runtime "base", and the difference between the two produces the load bias. Applying the same bias to all variable location lists "relocates" the variables. (The reason this is so convoluted is reading location lists straight from (unrelocated) .o files on mac. I seriously considered changing this when working on debug_rnglists, but it turned out it wasn't really necessary, so I let it go.)
>>
>> Yep, I figured a bunch of this was for DWARF in unrelocated MachO files - and that they wouldn't be able to/want to use Split DWARF or this feature (which is particularly relevant to Split DWARF). Does any of this logic apply outside that unrelocated MachO scenario?
>
> Yes, this code is used everywhere, though it's role is less important (and one that could be implemented in an easier manner, were it not for MachO) -- it adjusts the variable address for ASLR, i.e., the variable/function being loaded at a different address than what the static debug info says.

Ah, good to know!

>>> LLDB's representation of a function (lldb_private::Function) assumes that all functions will be contiguous (Function::GetAddressRange returns a single range). If we make it so that this range matches the first block of the function (maybe that's what happens now),
>>
>> Actually seems to adjust for discontiguous ranges and not just pick the start of the first (as it appears in the DWARF - which isn't guaranteed to be ordered in any way) range, instead finding the smallest and largest addresses: https://github.com/llvm-mirror/lldb/blob/master/source/Plugins/SymbolFile/DWARF/DWARFASTParserClang.cpp#L2357
>
> Ah.. interesting. In that case, I believe things should work if we use the same algorithm to compute `func_lo_pc`. We may still run into weird issues for functions which are truly discontinuous, but it should be ok for single-range range lists, at least.

Fair.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D94064/new/

https://reviews.llvm.org/D94064