[llvm-dev] about __attribute__((artificial)) and inline pass

[Fangqing Du via llvm-dev]

Dear all,

In clang/llvm 7.0, `__attribute__((artificial))` can be identified by clang and attach `DIFlagArtificial` into function metadata in this change.

And according to the specification of `artificial`, if I understand correctly, when the function (which is marked with `artificial` attribute) is inlined, the inlined function instructions should be associated with the callsite line number, instead of the line of inlined callee.

But from my observation, with and without this attribute, the inline pass behavior is not affected, and inlined instructions debug location still point to the callee instead of callsite.

Here is my experiment: godbolt experiment link

Is my understanding correct?

Thanks,

Fangqing

Xilinx Inc.

== ------------ example -------------===

  1 int bar(int);

  2 

  3 inline int __attribute__((artificial)) __attribute__((always_inline)) foo(int x)

  4 {

  5     return bar(x + 1); 

  6 }

  7 

  8 void baz(void)

  9 {

 10     auto x = foo(1);

 11     x = foo(2);

 12 }

When function 'foo' is inlined, the debug location of call instruction 'bar' still points to line #5, instead of line #10 and #11.

[David Blaikie via llvm-dev]

Hmm, seems I missed this feature (or at least don't remember) when it went in.

Looks like the GCC spec might be more accurate to the current implementation than what Clang's documentation says:

"artificial

This attribute is useful for small inline wrappers which if possible should appear during debugging as a unit, depending on the debug info format it will either mean marking the function as artificial or using the caller location for all instructions within the inlined body."

& the "marking the function as artificial" is the thing that this patch implemented in Clang. This did not have the same effect as nodebug, and doesn't "use the caller location for all instructions within the inlined body".

It's possible we could implement the "artificial" attribute as a an alias for "nodebug" which would also be conforming with GCC's spec, but would be less expressive than the current implementation (though would save debug info size in a way the current Clang implementation does not).

It looks like Clang's implementation is consistent with GCC's implementation for DWARF - "marking the function as artificial", not "using the caller location for instructions within the inlined body", but Clang's documentation is incorrect, since it documents the latter and not the former.

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Reid Kleckner via llvm-dev]

I thought `artificial` and `nodebug` were supposed to be synonyms, just different names invented by different communities. The linked patch doesn't implement the LLVM side of things, so I'm not sure when that got added.

[David Blaikie via llvm-dev]

On Fri, Jan 28, 2022 at 11:26 AM Reid Kleckner <r...@google.com> wrote:

I thought `artificial` and `nodebug` were supposed to be synonyms, just different names invented by different communities.

Doesn't look like it, based on the GCC documentation and behavior - GCC puts the DW_AT_artificial attribute on the entity, but still fully describes it (DWARF remains the same except for that attribute). I think the second option in the GCC docs is for debug info formats that don't have an equivalent of DW_AT_artificial, in which case it downgrades to the equivalent of "nodebug".
 

The linked patch doesn't implement the LLVM side of things, so I'm not sure when that got added.

The patch added the artificial flag on the metadata which was already implemented/supported in LLVM (used for things like implicit constructors, etc, which don't have user-written code in them) and matches the GCC behavior - no extra work was needed on the LLVM side.
 

[Robinson, Paul via llvm-dev]

In order to get the described effect, I think we’d need to put DW_AT_artificial on the subprogram, and then not attach source locations to the generated code.  That would have the effect of not allowing break/step inside, but the subprogram name is still available. This is not the same as `nodebug`, where you lose the subprogram name too.

Shouldn’t be at all hard to get this to work.

 

Fangqing, would you be able to file an issue for this?

Thanks,

--paulr

[David Blaikie via llvm-dev]

On Fri, Jan 28, 2022 at 12:07 PM Robinson, Paul <paul.r...@sony.com> wrote:

In order to get the described effect, I think we’d need to put DW_AT_artificial on the subprogram, and then not attach source locations to the generated code.  That would have the effect of not allowing break/step inside, but the subprogram name is still available. This is not the same as `nodebug`, where you lose the subprogram name too.

Shouldn’t be at all hard to get this to work.

That doesn't seem to be the described or actual effect (testing GCC's debug info, even for __attribute__((artificial)) functions, the locations inside the function (inlined or non-inlined) are attributed to their original source location, not to the call site) - I believe the intent is that only effect of the attribute is to add the DW_AT_artificial attribute to the function (& that seems to be the implemented effect) - but if GCC is emitting some other debug info format (not sure what other formats it supports) that doesn't have an equivalent to DW_AT_artificial, it does the next best thing - which is to attribute the inlined source locations to the call site. (unclear what it does for non-inlined code in that case - though interestingly the attribute is only valid in GCC on functions that are marked "inline"... which seems like a weird limitation to me).

I believe Clang is implementing the same behavior as GCC, as far as DWARF is concerned at least. It appears to be a documentation issue to me.

- Dave
 

[Fangqing Du via llvm-dev]

Thank you all for replying!

As Paul suggested, today, if clang doesn't attach source locations for 'artificial' marked functions, then inline pass will automatically attach call site location on the inlined instructions from callee.

Anyway, I filed an issue for this.

Fangqing