[PATCH] D24426: DebugInfo: Pass non-zero alignment to DIBuilder only if aligment was forced

Tue Sep 13 11:00:29 PDT 2016

I hadn't thought Clang wanted to be *quite* so knowledgeable about targets, and similarly not so tightly tied to byte-addressable targets.  But if both of those things are actually okay, then it's fine to set the alignment value here to what would be passed through to DWARF.
--paulr

From: David Blaikie [mailto:dblaikie at gmail.com]
Sent: Monday, September 12, 2016 6:11 PM
To: Robinson, Paul; reviews+D24426+public+6ee6274d38fdf0d6 at reviews.llvm.org; vleschuk at accesssoftek.com; echristo at gmail.com; aprantl at apple.com; mehdi.amini at apple.com
Cc: cfe-commits (cfe-commits at lists.llvm.org)
Subject: Re: [PATCH] D24426: DebugInfo: Pass non-zero alignment to DIBuilder only if aligment was forced


On Mon, Sep 12, 2016 at 6:01 PM Robinson, Paul <paul.robinson at sony.com<mailto:paul.robinson at sony.com>> wrote:
The text in the committee draft is different (e.g., the exhortation about non-default alignment is gone), with an example to the effect that a value of 8 means the entity's address is a multiple of 8 (not 2^8).  So, alignment is conceived in terms of address bits, whatever those represent (not always bytes).
Not sure I quite follow. OK, so in an octet addressable world (which LLVM is - there have been some attempts to support non-octet addressing, but I don't think any have been near to successful) then DW_AT_alignment is byte alignment (1 means there are no zero bits in the address, 2 means there's 1 trailing zero bit in the address, etc).
If Clang is being infested with more target knowledge, okay, but that means tolerating the weirder targets in these cases.  Dividing by CHAR_BITS makes an assumption that isn't necessarily correct.
Clang has the knowledge already - it knows the alignment of the types its allocating, etc. So I'm not sure what infestation you're referring to.

I've sort of lost track of what we're discussing here.

Essentially what I'm suggesting is that Clang should put whatever number is going to go in the DWARF, into the metadata. I don't believe the LLVM backends have greater knowledge than the frontend does in this domain - have I missed something there, are there examples where that could/would be true?

- David

--paulr
P.S. The committee is hoping to get a draft out for public comment Real Soon Now.
Looking forward to it :)

From: cfe-commits [mailto:cfe-commits-bounces at lists.llvm.org<mailto:cfe-commits-bounces at lists.llvm.org>] On Behalf Of David Blaikie via cfe-commits
Sent: Monday, September 12, 2016 5:12 PM
To: reviews+D24426+public+6ee6274d38fdf0d6 at reviews.llvm.org<mailto:reviews%2BD24426%2Bpublic%2B6ee6274d38fdf0d6 at reviews.llvm.org>; vleschuk at accesssoftek.com<mailto:vleschuk at accesssoftek.com>; echristo at gmail.com<mailto:echristo at gmail.com>; aprantl at apple.com<mailto:aprantl at apple.com>; mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>
Cc: cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>
Subject: Re: [PATCH] D24426: DebugInfo: Pass non-zero alignment to DIBuilder only if aligment was forced


On Mon, Sep 12, 2016 at 5:00 PM Paul Robinson <paul.robinson at sony.com<mailto:paul.robinson at sony.com>> wrote:
probinson added a subscriber: probinson.

================
Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr<AlignedAttr>())
+    AlignInBits = D->getMaxAlignment();
   StringRef DeclName, LinkageName;
----------------
dblaikie wrote:
> is max alignment the right thing here? Should it be min alignment?
> (is alignment in bits the desired thing across all of this too? It looked like in the backend patch there was some division by CHAR_BITS, etc?)
I should think bits is the right choice here; seems more the province of the backend to convert it into the appropriate addressable units (commonly but not universally chars).

The alternative thinking is that we've a generally sense we want to make more of this type information opaque to LLVM - so I'm somewhat inclined to make the frontend do the work of choosing what to emit and the backend just being as simple as possible.

Hmm, seems like the DWARF spec details I can find: http://www.dwarfstd.org/ShowIssue.php?issue=140528.1 don't really specify what the value of DW_AT_alignment is, it's sort of assumed, by the looks of it? I'm assuming it's bytes, the same as the byte_size attribute.





https://reviews.llvm.org/D24426

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20160913/71f42fa4/attachment-0001.html>
