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

[Victor Leschuk via cfe-commits]

vleschuk added inline comments.


Comment at: lib/CodeGen/CGDebugInfo.cpp:608
@@ -608,2 +607,3 @@
 
   uint64_t Size = CGM.getContext().getTypeSize(Ty);
+  uint64_t Align = 0;

rnk wrote:
> IMO this is what we should be doing everywhere, rather than manually checking 
> AlignedAttr:
>   TypeInfo TI = CGM.getContext().getTypeInfo(Ty);
>   uint64_t Size = TI.Width;
>   uint64_t Align = TI.AlignIsRequired ? TI.Align : 0;
> 
> This saves a hash lookup, and handles some corner cases. AlignIsRequired is 
> already supposed to capture whether the alignment was changed.
Will check if this works in all cases. I think it's worth putting this snippet 
into helper function within anon namespace.


https://reviews.llvm.org/D24426



___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Victor Leschuk via cfe-commits]

vleschuk added inline comments.


Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr())
+AlignInBits = D->getMaxAlignment();
   StringRef DeclName, LinkageName;

probinson wrote:
> 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).
> is max alignment the right thing here? Should it be min alignment?

The DWARF document says: 

> This constant describes the actual alignment used by the compiler (if there 
> are multiple specified by the user, or if the user specified an alignment the 
> compiler couldn't satisfy, then only the strictest alignment will be added 
> through a DW_AT_alignment).

So I think max alignment is right choice here.


https://reviews.llvm.org/D24426



___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Robinson, Paul via cfe-commits]

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:dblai...@gmail.com]
Sent: Monday, September 12, 2016 6:11 PM
To: Robinson, Paul; reviews+d24426+public+6ee6274d38fdf...@reviews.llvm.org; 
vlesc...@accesssoftek.com; echri...@gmail.com; apra...@apple.com; 
mehdi.am...@apple.com
Cc: cfe-commits (cfe-commits@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 
> 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-boun...@lists.llvm.org]
 On Behalf Of David Blaikie via cfe-commits
Sent: Monday, September 12, 2016 5:12 PM
To: 
reviews+d24426+public+6ee6274d38fdf...@reviews.llvm.org;
 vlesc...@accesssoftek.com; 
echri...@gmail.com; 
apra...@apple.com; 
mehdi.am...@apple.com
Cc: cfe-commits@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 
> wrote:
probinson added a subscriber: probinson.


Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr())
+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

___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[David Blaikie via cfe-commits]

On Mon, Sep 12, 2016 at 6:01 PM Robinson, Paul 
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-boun...@lists.llvm.org] *On
> Behalf Of *David Blaikie via cfe-commits
> *Sent:* Monday, September 12, 2016 5:12 PM
> *To:* reviews+d24426+public+6ee6274d38fdf...@reviews.llvm.org;
> vlesc...@accesssoftek.com; echri...@gmail.com; apra...@apple.com;
> mehdi.am...@apple.com
> *Cc:* cfe-commits@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 
> wrote:
>
> probinson added a subscriber: probinson.
>
> 
> Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
> @@ -3635,1 +3690,3 @@
> +  if (D->hasAttr())
> +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
>
>
>
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Robinson, Paul via cfe-commits]

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).

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

From: cfe-commits [mailto:cfe-commits-boun...@lists.llvm.org] On Behalf Of 
David Blaikie via cfe-commits
Sent: Monday, September 12, 2016 5:12 PM
To: reviews+d24426+public+6ee6274d38fdf...@reviews.llvm.org; 
vlesc...@accesssoftek.com; echri...@gmail.com; apra...@apple.com; 
mehdi.am...@apple.com
Cc: cfe-commits@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 
> wrote:
probinson added a subscriber: probinson.


Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr())
+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


___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Reid Kleckner via cfe-commits]

rnk added a subscriber: rnk.


Comment at: lib/CodeGen/CGDebugInfo.cpp:608
@@ -608,2 +607,3 @@
 
   uint64_t Size = CGM.getContext().getTypeSize(Ty);
+  uint64_t Align = 0;

IMO this is what we should be doing everywhere, rather than manually checking 
AlignedAttr:
  TypeInfo TI = CGM.getContext().getTypeInfo(Ty);
  uint64_t Size = TI.Width;
  uint64_t Align = TI.AlignIsRequired ? TI.Align : 0;

This saves a hash lookup, and handles some corner cases. AlignIsRequired is 
already supposed to capture whether the alignment was changed.


https://reviews.llvm.org/D24426



___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Paul Robinson via cfe-commits]

probinson added a subscriber: probinson.


Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr())
+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).


https://reviews.llvm.org/D24426



___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[David Blaikie via cfe-commits]

dblaikie added a comment.

Like the backend patch, should/could this be broken up into separate patches 
for the different places/features that have alignment?

(probably just passing zero for alignment in general could be a simple cleanup 
patch to start, if it's otherwise unused)



Comment at: lib/CodeGen/CGDebugInfo.cpp:609
@@ -609,3 +608,3 @@
   uint64_t Size = CGM.getContext().getTypeSize(Ty);
-  uint64_t Align = CGM.getContext().getTypeAlign(Ty);
+  uint64_t Align = 0;
   return DBuilder.createBasicType("complex", Size, Align, Encoding);

Maybe add a comment here about why we're specifying zero alignment (or possibly 
change the API to not take an alignment at all so we don't have to explain it?)

(similarly for other cases where the alignment is just hardcoded to zero)


Comment at: lib/CodeGen/CGDebugInfo.cpp:3691
@@ -3635,1 +3690,3 @@
+  if (D->hasAttr())
+AlignInBits = D->getMaxAlignment();
   StringRef DeclName, LinkageName;

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?)


https://reviews.llvm.org/D24426



___
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

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

[Victor Leschuk via cfe-commits]

vleschuk retitled this revision from "DebugInfo: use 
llvm::DINode::FlagAlignment to mark forcibly aligned data" to "DebugInfo: Pass 
non-zero alignment to DIBuilder only if aligment was forced".
vleschuk updated the summary for this revision.
vleschuk updated this revision to Diff 70965.
vleschuk added a comment.

Got rid of DINode::FlagAlignment. Check for forced alignment (alignas(), 
_Alignas(), __attribute(aligned (N) when creating DI* objects with 
DIBuilder. If alignment was not forced pass zero, otherwise pass actual 
alignment value. Updated related tests according to new behavior.


https://reviews.llvm.org/D24426

Files:
  lib/CodeGen/CGDebugInfo.cpp
  lib/CodeGen/CGDebugInfo.h
  test/CodeGen/debug-info-packed-struct.c
  test/CodeGen/debug-info-vector.c
  test/CodeGenCXX/debug-info-calling-conventions.cpp
  test/CodeGenCXX/debug-info-enum-class.cpp
  test/CodeGenCXX/debug-info-indirect-field-decl.cpp
  test/CodeGenCXX/debug-info-ms-bitfields.cpp
  test/CodeGenCXX/debug-info-ms-ptr-to-member.cpp
  test/CodeGenCXX/debug-info-rvalue-ref.cpp
  test/CodeGenCXX/debug-info-template-quals.cpp
  test/CodeGenCXX/debug-info-template.cpp
  test/CodeGenCXX/debug-info-union.cpp
  test/CodeGenCXX/debug-info-uuid.cpp
  test/CodeGenCXX/debug-info-vla.cpp
  test/CodeGenCXX/debug-info-zero-length-arrays.cpp
  test/CodeGenCXX/debug-info.cpp
  test/CodeGenCXX/debug-lambda-this.cpp
  test/CodeGenObjC/block-byref-debuginfo.m
  test/CodeGenObjC/debug-info-block-type.m
  test/CodeGenObjC/debug-info-ivars-extension.m
  test/CodeGenObjC/debug-info-ivars-private.m
  test/CodeGenObjC/debug-info-ivars.m
  test/CodeGenObjCXX/debug-info-cyclic.mm

Index: test/CodeGenObjCXX/debug-info-cyclic.mm
===
--- test/CodeGenObjCXX/debug-info-cyclic.mm
+++ test/CodeGenObjCXX/debug-info-cyclic.mm
@@ -3,7 +3,7 @@
 struct B {
 // CHECK: ![[B:[0-9]+]] ={{.*}}!DICompositeType(tag: DW_TAG_structure_type, name: "B"
 // CHECK-SAME: line: [[@LINE-2]],
-// CHECK-SAME: size: 8, align: 8,
+// CHECK-SAME: size: 8,
 // CHECK-NOT:  offset:
 // CHECK-NOT:  DIFlagFwdDecl
 // CHECK-SAME: elements: ![[BMEMBERS:[0-9]+]]
Index: test/CodeGenObjC/debug-info-ivars.m
===
--- test/CodeGenObjC/debug-info-ivars.m
+++ test/CodeGenObjC/debug-info-ivars.m
@@ -21,24 +21,24 @@
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "i"
 // CHECK-SAME:   line: 10
 // CHECK-SAME:   baseType: ![[INT:[0-9]+]]
-// CHECK-SAME:   size: 32, align: 32,
+// CHECK-SAME:   size: 32,
 // CHECK-NOT:offset:
 // CHECK-SAME:   flags: DIFlagProtected
 // CHECK: ![[INT]] = !DIBasicType(name: "int"
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "flag_1"
 // CHECK-SAME:   line: 11
 // CHECK-SAME:   baseType: ![[UNSIGNED:[0-9]+]]
-// CHECK-SAME:   size: 9, align: 32,
+// CHECK-SAME:   size: 9,
 // CHECK-NOT:offset:
 // CHECK-SAME:   flags: DIFlagProtected
 // CHECK: ![[UNSIGNED]] = !DIBasicType(name: "unsigned int"
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "flag_2"
 // CHECK-SAME:   line: 12
 // CHECK-SAME:   baseType: ![[UNSIGNED]]
-// CHECK-SAME:   size: 9, align: 32, offset: 1,
+// CHECK-SAME:   size: 9, offset: 1,
 // CHECK-SAME:   flags: DIFlagProtected
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "flag_3"
 // CHECK-SAME:   line: 14
 // CHECK-SAME:   baseType: ![[UNSIGNED]]
-// CHECK-SAME:   size: 9, align: 32, offset: 3,
+// CHECK-SAME:   size: 9, offset: 3,
 // CHECK-SAME:   flags: DIFlagProtected
Index: test/CodeGenObjC/debug-info-ivars-private.m
===
--- test/CodeGenObjC/debug-info-ivars-private.m
+++ test/CodeGenObjC/debug-info-ivars-private.m
@@ -35,13 +35,13 @@
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "foo"
 // CHECK-SAME:   line: 14
 // CHECK-SAME:   baseType: ![[INT:[0-9]+]]
-// CHECK-SAME:   size: 32, align: 32,
+// CHECK-SAME:   size: 32,
 // CHECK-NOT:offset:
 // CHECK-SAME:   flags: DIFlagProtected
 // CHECK: ![[INT]] = !DIBasicType(name: "int"
 // CHECK: !DIDerivedType(tag: DW_TAG_member, name: "bar"
 // CHECK-SAME:   line: 27
 // CHECK-SAME:   baseType: ![[INT:[0-9]+]]
-// CHECK-SAME:   size: 32, align: 32,
+// CHECK-SAME:   size: 32,
 // CHECK-NOT:offset:
 // CHECK-SAME:   flags: DIFlagPrivate
Index: test/CodeGenObjC/debug-info-ivars-extension.m
===
--- test/CodeGenObjC/debug-info-ivars-extension.m
+++