[llvm] r266548 - IR: Use ODR to unique DICompositeType members
David Blaikie via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 18 12:47:28 PDT 2016
On Mon, Apr 18, 2016 at 11:29 AM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:
> >
> >> On 2016-Apr-18, at 09:10, David Blaikie <dblaikie at gmail.com> wrote:
> >>
> >>
> >>
> >> On Sun, Apr 17, 2016 at 4:27 PM, Duncan P. N. Exon Smith <
> dexonsmith at apple.com> wrote:
> >>
> >> > On 2016-Apr-17, at 15:34, David Blaikie <dblaikie at gmail.com> wrote:
> >> >
> >> >
> >> >
> >> > 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?
> >>
> >> You need to check the scope on LHS to see if it follows to ODR type
> >> uniquing (has an 'identifier:' field). When both LHS and RHS follow
> >> ODR, then I think you're right: if the 'linkageName:' matches, that
> >> implies that the 'scope' matches. But I think it's cheaper to just
> >> compare the values of LHS and RHS than to inspect the scope of RHS.
> >
> > I'm lost here - not sure which left/right hand sides you're referring to.
>
> From the commit. LHS has already been splatted out, but this is the
> code I was talking about:
>
> >> > > + 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();
> >> > > + }
> >> > > +};
>
> ^ Note that isa<MDString> implies the type follows ODR.
>
Still confused here - you said "cheaper to just compare the values of LHS
and RHS than to inspect the scope of RHS" - I'm confused why we aren't
just inspecting the linkage name? That's not about inspecting the scope...
but I'm getting lost here.
>
> >> If the linkage names are poor quality, then it's probably best to use
> >> the scopes anyway.
> >
> > I'd like to get more consistency into the linkage names, really. But
> it's possible we'd end up with competing goals between reusing linkage
> names for uniquing versus other things that we want to use them for in
> DWARF, etc.
> >
> >> I think it might be necessary to check the scopes when there's no
> >> DITypeMap in the LLVMContext, since two subprograms might look the
> >> same from their LinkageName, but be in a different DICompositeType.
> >
> > Not sure I follow there.
>
> I'm not sure I follow any more either :/.
>
> >> > > 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.
> >>
> >> I think merging these should happen implicitly now when the
> >> DITypeMap is enabled. Actually the motivation for this was that
> >> llvm/trunk/test/Linker/type-unique-odr-a.ll started failing when
> >> wrote my DITypeMap patch (r266549). This was necessary to make
> >> it all work.
> >>
> >> >
> >> > >
> >> > >
> >> > > 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/20160418/7f906d68/attachment.html>
More information about the llvm-commits
mailing list