[llvm] r266548 - IR: Use ODR to unique DICompositeType members

David Blaikie via llvm-commits llvm-commits at lists.llvm.org
Sun Apr 17 15:34:55 PDT 2016


On Sun, Apr 17, 2016 at 3:09 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

>
> > On 2016-Apr-17, at 15:08, David Blaikie <dblaikie at gmail.com> wrote:
> >
> >
> >
> > On Sat, Apr 16, 2016 at 7:30 PM, Duncan P. N. Exon Smith via
> llvm-commits <llvm-commits at lists.llvm.org> wrote:
> > Author: dexonsmith
> > Date: Sat Apr 16 21:30:20 2016
> > New Revision: 266548
> >
> > URL: http://llvm.org/viewvc/llvm-project?rev=266548&view=rev
> > Log:
> > IR: Use ODR to unique DICompositeType members
> >
> > Merge members that are describing the same member of the same ODR type,
> > even if other bits differ.  If the file or line differ, we don't care;
> > if anything else differs, it's an ODR violation (and we still don't
> > really care).
> >
> > For DISubprogram declarations, this looks at the LinkageName and Scope.
>

What we put linkage names on is a bit haphazard... I think I mentioned this
in a recent review for Paul. I included some patches to Clang that might
fix those - I'd suggest you check them out or I can go & find them. (and/or
test this deduplication with some function-local types (in inline
functions) and a type in an anonymous namespace)

Why look at the scope as well? What does that add/change?


> > For DW_TAG_member instances of DIDerivedType, this looks at the Name and
> > Scope.  In both cases, we know that the Scope follows ODR rules if it
> > has a non-empty identifier.
> >
> > Not quite following this paragraph ^ why does this need to look at
> subprograms and members, rather than just the 'identifier' of the
> DICompositeType itself?
>
> It's the subprograms and members that are being uniqued.
>

*reads the commit message again* Right.

*ponders*

If it's any help, I think it should be fine to never touch/merge the
"members" list in a DICompositeType (perhaps you're already doing this) -
if a member is in the member list, just take the equivalent (linkage name,
I guess).

But I guess you've got to deduplicate the things that aren't in the member
list too anyway (& the order of things, and their simple names aren't
enough to identify them, etc) so there's not much point doing something
special for the things in the member list, perhaps.


>
> >
> >
> > Added:
> >     llvm/trunk/test/Assembler/dicompositetype-members.ll
> > Modified:
> >     llvm/trunk/docs/LangRef.rst
> >     llvm/trunk/lib/IR/LLVMContextImpl.h
> >     llvm/trunk/test/Linker/type-unique-odr-a.ll
> >
> > Modified: llvm/trunk/docs/LangRef.rst
> > URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/docs/LangRef.rst?rev=266548&r1=266547&r2=266548&view=diff
> >
> ==============================================================================
> > --- llvm/trunk/docs/LangRef.rst (original)
> > +++ llvm/trunk/docs/LangRef.rst Sat Apr 16 21:30:20 2016
> > @@ -3976,7 +3976,10 @@ The following ``tag:`` values are valid:
> >
> >  ``DW_TAG_member`` is used to define a member of a :ref:`composite type
> >  <DICompositeType>`. The type of the member is the ``baseType:``. The
> > -``offset:`` is the member's bit offset.
> > +``offset:`` is the member's bit offset.  If the composite type has a
> non-empty
> > +``identifier:``, then it respects ODR rules.  In that case, the
> ``scope:``
> > +reference will be a :ref:`metadata string <metadata-string>`, and the
> member
> > +will be uniqued solely based on its ``name:`` and ``scope:``.
> >
> >  ``DW_TAG_inheritance`` and ``DW_TAG_friend`` are used in the
> ``elements:``
> >  field of :ref:`composite types <DICompositeType>` to describe parents
> and
> > @@ -4125,6 +4128,12 @@ metadata. The ``variables:`` field point
> >  that must be retained, even if their IR counterparts are optimized out
> of
> >  the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
> >
> > +When ``isDefinition: false``, subprograms describe a declaration in the
> type
> > +tree as opposed to a definition of a funciton.  If the scope is a
> > +:ref:`metadata string <metadata-string>` then the composite type
> follows ODR
> > +rules, and the subprogram declaration is uniqued based only on its
> > +``linkageName:`` and ``scope:``.
> > +
> >  .. code-block:: llvm
> >
> >      define void @_Z3foov() !dbg !0 {
> > @@ -4133,7 +4142,7 @@ the IR. The ``type:`` field must point a
> >
> >      !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov",
> scope: !1,
> >                                  file: !2, line: 7, type: !3, isLocal:
> true,
> > -                                isDefinition: false, scopeLine: 8,
> > +                                isDefinition: true, scopeLine: 8,
> >                                  containingType: !4,
> >                                  virtuality: DW_VIRTUALITY_pure_virtual,
> >                                  virtualIndex: 10, flags:
> DIFlagPrototyped,
> >
> > Modified: llvm/trunk/lib/IR/LLVMContextImpl.h
> > URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/IR/LLVMContextImpl.h?rev=266548&r1=266547&r2=266548&view=diff
> >
> ==============================================================================
> > --- llvm/trunk/lib/IR/LLVMContextImpl.h (original)
> > +++ llvm/trunk/lib/IR/LLVMContextImpl.h Sat Apr 16 21:30:20 2016
> > @@ -32,6 +32,7 @@
> >  #include "llvm/IR/LLVMContext.h"
> >  #include "llvm/IR/Metadata.h"
> >  #include "llvm/IR/ValueHandle.h"
> > +#include "llvm/Support/Dwarf.h"
> >  #include <vector>
> >
> >  namespace llvm {
> > @@ -376,6 +377,12 @@ template <> struct MDNodeKeyImpl<DIDeriv
> >             ExtraData == RHS->getRawExtraData();
> >    }
> >    unsigned getHashValue() const {
> > +    // If this is a member inside an ODR type, only hash the type and
> the name.
> > +    // Otherwise the hash will be stronger than
> > +    // MDNodeSubsetEqualImpl::isODRMember().
> > +    if (Tag == dwarf::DW_TAG_member && Name && Scope &&
> isa<MDString>(Scope))
> > +      return hash_combine(Name, Scope);
> > +
> >      // Intentionally computes the hash on a subset of the operands for
> >      // performance reason. The subset has to be significant enough to
> avoid
> >      // collision "most of the time". There is no correctness issue in
> case of
> > @@ -384,6 +391,30 @@ template <> struct MDNodeKeyImpl<DIDeriv
> >    }
> >  };
> >
> > +template <> struct MDNodeSubsetEqualImpl<DIDerivedType> {
> > +  typedef MDNodeKeyImpl<DIDerivedType> KeyTy;
> > +  static bool isSubsetEqual(const KeyTy &LHS, const DIDerivedType *RHS)
> {
> > +    return isODRMember(LHS.Tag, LHS.Scope, LHS.Name, RHS);
> > +  }
> > +  static bool isSubsetEqual(const DIDerivedType *LHS, const
> DIDerivedType *RHS) {
> > +    return isODRMember(LHS->getTag(), LHS->getRawScope(),
> LHS->getRawName(),
> > +                       RHS);
> > +  }
> > +
> > +  /// Subprograms compare equal if they declare the same function in an
> ODR
> > +  /// type.
> > +  static bool isODRMember(unsigned Tag, const Metadata *Scope,
> > +                          const MDString *Name, const DIDerivedType
> *RHS) {
> > +    // Check whether the LHS is eligible.
> > +    if (Tag != dwarf::DW_TAG_member || !Name || !Scope ||
> !isa<MDString>(Scope))
> > +      return false;
> > +
> > +    // Compare to the RHS.
> > +    return Tag == RHS->getTag() && Name == RHS->getRawName() &&
> > +           Scope == RHS->getRawScope();
> > +  }
> > +};
> > +
> >  template <> struct MDNodeKeyImpl<DICompositeType> {
> >    unsigned Tag;
> >    MDString *Name;
> > @@ -537,6 +568,12 @@ template <> struct MDNodeKeyImpl<DISubpr
> >             Variables == RHS->getRawVariables();
> >    }
> >    unsigned getHashValue() const {
> > +    // If this is a declaration inside an ODR type, only hash the type
> and the
> > +    // name.  Otherwise the hash will be stronger than
> > +    // MDNodeSubsetEqualImpl::isDeclarationOfODRMember().
> > +    if (!IsDefinition && LinkageName && Scope && isa<MDString>(Scope))
> > +      return hash_combine(LinkageName, Scope);
> > +
> >      // Intentionally computes the hash on a subset of the operands for
> >      // performance reason. The subset has to be significant enough to
> avoid
> >      // collision "most of the time". There is no correctness issue in
> case of
> > @@ -545,6 +582,33 @@ template <> struct MDNodeKeyImpl<DISubpr
> >    }
> >  };
> >
> > +template <> struct MDNodeSubsetEqualImpl<DISubprogram> {
> > +  typedef MDNodeKeyImpl<DISubprogram> KeyTy;
> > +  static bool isSubsetEqual(const KeyTy &LHS, const DISubprogram *RHS) {
> > +    return isDeclarationOfODRMember(LHS.IsDefinition, LHS.Scope,
> > +                                    LHS.LinkageName, RHS);
> > +  }
> > +  static bool isSubsetEqual(const DISubprogram *LHS, const DISubprogram
> *RHS) {
> > +    return isDeclarationOfODRMember(LHS->isDefinition(),
> LHS->getRawScope(),
> > +                                    LHS->getRawLinkageName(), RHS);
> > +  }
> > +
> > +  /// Subprograms compare equal if they declare the same function in an
> ODR
> > +  /// type.
> > +  static bool isDeclarationOfODRMember(bool IsDefinition, const
> Metadata *Scope,
> > +                                       const MDString *LinkageName,
> > +                                       const DISubprogram *RHS) {
> > +    // Check whether the LHS is eligible.
> > +    if (IsDefinition || !Scope || !LinkageName || !Scope ||
> > +        !isa<MDString>(Scope))
> > +      return false;
> > +
> > +    // Compare to the RHS.
> > +    return IsDefinition == RHS->isDefinition() && Scope ==
> RHS->getRawScope() &&
> > +           LinkageName == RHS->getRawLinkageName();
> > +  }
> > +};
> > +
> >  template <> struct MDNodeKeyImpl<DILexicalBlock> {
> >    Metadata *Scope;
> >    Metadata *File;
> >
> > Added: llvm/trunk/test/Assembler/dicompositetype-members.ll
> > URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Assembler/dicompositetype-members.ll?rev=266548&view=auto
> >
> ==============================================================================
> > --- llvm/trunk/test/Assembler/dicompositetype-members.ll (added)
> > +++ llvm/trunk/test/Assembler/dicompositetype-members.ll Sat Apr 16
> 21:30:20 2016
> > @@ -0,0 +1,54 @@
> > +; RUN: llvm-as < %s | llvm-dis | llvm-as | llvm-dis | FileCheck %s
> > +; RUN: verify-uselistorder %s
> > +
> > +; Anchor the order of the nodes.
> > +!named = !{!0, !1, !2, !3, !4, !5, !6, !7, !8, !9, !10, !11, !12, !13,
> !14, !15, !16, !17}
> > +
> > +; Some basic building blocks.
> > +; CHECK:      !0 = !DIBasicType
> > +; CHECK-NEXT: !1 = !DIFile
> > +; CHECK-NEXT: !2 = !DIFile
> > +!0 = !DIBasicType(tag: DW_TAG_base_type, name: "name", size: 1, align:
> 2, encoding: DW_ATE_unsigned_char)
> > +!1 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
> > +!2 = !DIFile(filename: "path/to/other", directory: "/path/to/dir")
> > +
> > +; Define an identified type with fields and functions.
> > +; CHECK-NEXT: !3 = !DICompositeType(tag: DW_TAG_structure_type, name:
> "has-uuid",
> > +; CHECK-NEXT: !4 = !DIDerivedType(tag: DW_TAG_member, name: "field1",
> scope: !"has-uuid", file: !1
> > +; CHECK-NEXT: !5 = !DIDerivedType(tag: DW_TAG_member, name: "field2",
> scope: !"has-uuid", file: !1
> > +; CHECK-NEXT: !6 = !DISubprogram(name: "foo", linkageName: "foo1",
> scope: !"has-uuid", file: !1
> > +; CHECK-NEXT: !7 = !DISubprogram(name: "foo", linkageName: "foo2",
> scope: !"has-uuid", file: !1
> > +!3 = !DICompositeType(tag: DW_TAG_structure_type, name: "has-uuid",
> file: !1, line: 2, size: 64, align: 32, identifier: "uuid")
> > +!4 = !DIDerivedType(tag: DW_TAG_member, name: "field1", scope:
> !"has-uuid", file: !1, line: 4, baseType: !0, size: 32, align: 32, offset:
> 32)
> > +!5 = !DIDerivedType(tag: DW_TAG_member, name: "field2", scope:
> !"has-uuid", file: !1, line: 4, baseType: !0, size: 32, align: 32, offset:
> 32)
> > +!6 = !DISubprogram(name: "foo", linkageName: "foo1", scope:
> !"has-uuid", file: !1, isDefinition: false)
> > +!7 = !DISubprogram(name: "foo", linkageName: "foo2", scope:
> !"has-uuid", file: !1, isDefinition: false)
> > +
> > +; Define an un-identified type with fields and functions.
> > +; CHECK-NEXT: !8 = !DICompositeType(tag: DW_TAG_structure_type, name:
> "no-uuid", file: !1
> > +; CHECK-NEXT: !9 = !DIDerivedType(tag: DW_TAG_member, name: "field1",
> scope: !8, file: !1
> > +; CHECK-NEXT: !10 = !DIDerivedType(tag: DW_TAG_member, name: "field2",
> scope: !8, file: !1
> > +; CHECK-NEXT: !11 = !DISubprogram(name: "foo", linkageName: "foo1",
> scope: !8, file: !1
> > +; CHECK-NEXT: !12 = !DISubprogram(name: "foo", linkageName: "foo2",
> scope: !8, file: !1
> > +!8 = !DICompositeType(tag: DW_TAG_structure_type, name: "no-uuid",
> file: !1, line: 2, size: 64, align: 32)
> > +!9 = !DIDerivedType(tag: DW_TAG_member, name: "field1", scope: !8,
> file: !1, line: 4, baseType: !0, size: 32, align: 32, offset: 32)
> > +!10 = !DIDerivedType(tag: DW_TAG_member, name: "field2", scope: !8,
> file: !1, line: 4, baseType: !0, size: 32, align: 32, offset: 32)
> > +!11 = !DISubprogram(name: "foo", linkageName: "foo1", scope: !8, file:
> !1, isDefinition: false)
> > +!12 = !DISubprogram(name: "foo", linkageName: "foo2", scope: !8, file:
> !1, isDefinition: false)
> > +
> > +; Add duplicate fields and members of "no-uuid" in a different file.
> These
> > +; should stick around, since "no-uuid" does not have an "identifier:"
> field.
> > +; CHECK-NEXT: !13 = !DIDerivedType(tag: DW_TAG_member, name: "field1",
> scope: !8, file: !2,
> > +; CHECK-NEXT: !14 = !DISubprogram(name: "foo", linkageName: "foo1",
> scope: !8, file: !2,
> > +!13 = !DIDerivedType(tag: DW_TAG_member, name: "field1", scope: !8,
> file: !2, line: 4, baseType: !0, size: 32, align: 32, offset: 32)
> > +!14 = !DISubprogram(name: "foo", linkageName: "foo1", scope: !8, file:
> !2, isDefinition: false)
> > +
> > +; Add duplicate fields and members of "has-uuid" in a different file.
> These
> > +; should be merged.
> > +!15 = !DIDerivedType(tag: DW_TAG_member, name: "field1", scope:
> !"has-uuid", file: !2, line: 4, baseType: !0, size: 32, align: 32, offset:
> 32)
> > +!16 = !DISubprogram(name: "foo", linkageName: "foo1", scope:
> !"has-uuid", file: !2, isDefinition: false)
> > +
> > +; CHECK-NEXT: !15 = !{!4, !6}
> > +; CHECK-NOT: !DIDerivedType
> > +; CHECK-NOT: !DISubprogram
> > +!17 = !{!15, !16}
> >
> > Modified: llvm/trunk/test/Linker/type-unique-odr-a.ll
> > URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Linker/type-unique-odr-a.ll?rev=266548&r1=266547&r2=266548&view=diff
> >
> ==============================================================================
> > --- llvm/trunk/test/Linker/type-unique-odr-a.ll (original)
> > +++ llvm/trunk/test/Linker/type-unique-odr-a.ll Sat Apr 16 21:30:20 2016
> > @@ -4,6 +4,10 @@
> >  ; RUN:   | %llc_dwarf -dwarf-linkage-names=Enable -filetype=obj -O0 \
> >  ; RUN:   | llvm-dwarfdump -debug-dump=info - \
> >  ; RUN:   | FileCheck %s
> > +; RUN: llvm-link %p/type-unique-odr-b.ll %s -S -o - \
> > +; RUN:   | %llc_dwarf -dwarf-linkage-names=Enable -filetype=obj -O0 \
> > +; RUN:   | llvm-dwarfdump -debug-dump=info - \
> > +; RUN:   | FileCheck %s
> >  ;
> >  ; Test ODR-based type uniquing for C++ class members.
> >  ; rdar://problem/15851313.
> >
> >
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160417/92f6dd78/attachment.html>


More information about the llvm-commits mailing list