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

Fri Jan 28 12:07:01 PST 2022

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

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of David Blaikie via llvm-dev
Sent: Friday, January 28, 2022 2:37 PM
To: Reid Kleckner <rnk at google.com>
Cc: llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] about __attribute__((artificial)) and inline pass

On Fri, Jan 28, 2022 at 11:26 AM Reid Kleckner <rnk at google.com<mailto:rnk at 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.

On Fri, Jan 28, 2022 at 11:03 AM David Blaikie <dblaikie at gmail.com<mailto:dblaikie at gmail.com>> wrote:
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.

On Fri, Jan 28, 2022 at 9:44 AM Fangqing Du via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
Dear all,

In clang/llvm 7.0, `__attribute__((artificial))` can be identified by clang and attach `DIFlagArtificial` into function metadata in this change.<https://urldefense.com/v3/__https:/reviews.llvm.org/D43259*change-jbvcnfxttyAz__;Iw!!JmoZiZGBv3RvKRSx!pEsMF8kDNPtKteuvhreMp6mZjOVtB-7XVCgfOi_ELf630AbjxHC2VDLB3BoX9IMFWg$>

And according to the specification<https://urldefense.com/v3/__https:/clang.llvm.org/docs/AttributeReference.html*artificial__;Iw!!JmoZiZGBv3RvKRSx!pEsMF8kDNPtKteuvhreMp6mZjOVtB-7XVCgfOi_ELf630AbjxHC2VDLB3BrUKhXpvw$> 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<https://urldefense.com/v3/__https:/godbolt.org/z/oW8Td5d4M__;!!JmoZiZGBv3RvKRSx!pEsMF8kDNPtKteuvhreMp6mZjOVtB-7XVCgfOi_ELf630AbjxHC2VDLB3BqqabukgA$>

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.
_______________________________________________
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://urldefense.com/v3/__https:/lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev__;!!JmoZiZGBv3RvKRSx!pEsMF8kDNPtKteuvhreMp6mZjOVtB-7XVCgfOi_ELf630AbjxHC2VDLB3Bp86eErug$>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20220128/6a33d02c/attachment.html>