Automerge: [llvm][DebugInfo] Encode DW_AT_object_pointer on method declarations with DW_FORM_implicit_const (#124790)

We started attaching `DW_AT_object_pointer`s on method declarations in
llvm/llvm-project#122742. However, that caused
the `.debug_info` section size to increase significantly (by around ~10%
on some projects). This was mainly due to the large number of new
`DW_FORM_ref4` values. This patch tries to address that regression by
changing the `DW_FORM_ref4` to a `DW_FORM_implicit_const` for
declarations. The value of `DW_FORM_implicit_const` will be the *index*
of the object parameter in the list of formal parameters of the
subprogram (i.e., if the first `DW_TAG_formal_parameter` is the object
pointer, the `DW_FORM_implicit_const` would be `0`). The DWARFv5 spec
only mentions the use of the `reference` attribute class to for
`DW_AT_object_pointer`. So using a `DW_FORM_impilicit_const` would be an
extension to (and not something mandated/specified by) the standard.
Though it'd make sense to extend the wording in the spec to allow for
this optimization.

That way we don't pay for the 4 byte references on every attribute
occurrence. In a local build of clang this barely affected the
`.debug_info` section size (but did increase `.debug_abbrev` by up to
10%, which doesn't impact the total debug-info size much however).

We guarded this on LLDB tuning (since using `DW_FORM_implicit_const` for
this purpose may surprise consumers) and DWARFv5 (since that's where
`DW_FORM_implicit_const` was first standardized).

Author: Michael137

llvm/lib/CodeGen/AsmPrinter/DwarfUnit.cpp

@@ -895,7 +895,10 @@ void DwarfUnit::constructTypeDIE(DIE &Buffer, const DIDerivedType *DTy) {
   }
 }
 
-void DwarfUnit::constructSubprogramArguments(DIE &Buffer, DITypeRefArray Args) {
+std::optional<unsigned>
+DwarfUnit::constructSubprogramArguments(DIE &Buffer, DITypeRefArray Args) {
+  // Args[0] is the return type.
+  std::optional<unsigned> ObjectPointerIndex;
   for (unsigned i = 1, N = Args.size(); i < N; ++i) {
     const DIType *Ty = Args[i];
     if (!Ty) {
@@ -906,8 +909,16 @@ void DwarfUnit::constructSubprogramArguments(DIE &Buffer, DITypeRefArray Args) {
       addType(Arg, Ty);
       if (Ty->isArtificial())
         addFlag(Arg, dwarf::DW_AT_artificial);
+
+      if (Ty->isObjectPointer()) {
+        assert(!ObjectPointerIndex &&
+               "Can't have more than one object pointer");
+        ObjectPointerIndex = i;
+      }
     }
   }
+
+  return ObjectPointerIndex;
 }
 
 void DwarfUnit::constructTypeDIE(DIE &Buffer, const DISubroutineType *CTy) {
@@ -1458,7 +1469,20 @@ void DwarfUnit::applySubprogramAttributes(const DISubprogram *SP, DIE &SPDie,
 
     // Add arguments. Do not add arguments for subprogram definition. They will
     // be handled while processing variables.
-    constructSubprogramArguments(SPDie, Args);
+    //
+    // Encode the object pointer as an index instead of a DIE reference in order
+    // to minimize the affect on the .debug_info size.
+    if (std::optional<unsigned> ObjectPointerIndex =
+            constructSubprogramArguments(SPDie, Args)) {
+      if (getDwarfDebug().tuneForLLDB() &&
+          getDwarfDebug().getDwarfVersion() >= 5) {
+        // 0th index in Args is the return type, hence adjust by 1. In DWARF
+        // we want the first parameter to be at index 0.
+        assert(*ObjectPointerIndex > 0);
+        addSInt(SPDie, dwarf::DW_AT_object_pointer,
+                dwarf::DW_FORM_implicit_const, *ObjectPointerIndex - 1);
+      }
+    }
   }
 
   addThrownTypes(SPDie, SP->getThrownTypes());

llvm/lib/CodeGen/AsmPrinter/DwarfUnit.h

@@ -273,7 +273,11 @@ class DwarfUnit : public DIEUnit {
   void constructContainingTypeDIEs();
 
   /// Construct function argument DIEs.
-  void constructSubprogramArguments(DIE &Buffer, DITypeRefArray Args);
+  ///
+  /// \returns The index of the object parameter in \c Args if one exists.
+  /// Returns std::nullopt otherwise.
+  std::optional<unsigned> constructSubprogramArguments(DIE &Buffer,
+                                                       DITypeRefArray Args);
 
   /// Create a DIE with the given Tag, add the DIE to its parent, and
   /// call insertDIE if MD is not null.

llvm/test/DebugInfo/X86/DW_AT_object_pointer-non-standard-index.ll

@@ -0,0 +1,79 @@
+; Similar to DW_AT_object_pointer.ll but tests that we correctly
+; encode the object pointer index even if it's not the first argument
+; of the subprogram (which isn't something the major compilers do,
+; but is not mandated by DWARF).
+
+; RUN: llc -mtriple=x86_64-apple-darwin -debugger-tune=lldb -dwarf-version=5 -filetype=obj < %s | \
+; RUN:      llvm-dwarfdump -v -debug-info - | FileCheck %s --check-prefixes=CHECK
+
+; CHECK: DW_TAG_class_type
+; CHECK: [[DECL:0x[0-9a-f]+]]: DW_TAG_subprogram
+; CHECK:                         DW_AT_name {{.*}} "A"
+; CHECK: DW_AT_object_pointer [DW_FORM_implicit_const] (2)
+;
+; CHECK: DW_TAG_subprogram
+; CHECK:   DW_AT_object_pointer [DW_FORM_ref4]     (cu + 0x{{[0-9a-f]*}} => {[[PARAM:0x[0-9a-f]*]]})
+; CHECK:   DW_AT_specification [DW_FORM_ref4] (cu + {{.*}} => {[[DECL]]}
+; CHECK:   DW_TAG_formal_parameter
+; CHECK:   DW_TAG_formal_parameter
+; CHECK-NOT: "this"
+; CHECK: [[PARAM]]: DW_TAG_formal_parameter
+; CHECK: DW_AT_name
+; CHECK-SAME: = "this")
+; CHECK:   DW_TAG_formal_parameter
+
+%class.A = type { i8 }
+
+define linkonce_odr noundef ptr @_ZN1AC1Eii(ptr noundef nonnull returned align 1 dereferenceable(1) %this, i32 noundef %x, i32 noundef %y, i32 noundef %z) !dbg !24 {
+entry:
+  %this.addr = alloca ptr, align 8
+  %x.addr = alloca i32, align 4
+  %y.addr = alloca i32, align 4
+  %z.addr = alloca i32, align 4
+  store ptr %this, ptr %this.addr, align 8
+    #dbg_declare(ptr %this.addr, !26, !DIExpression(), !28)
+  store i32 %x, ptr %x.addr, align 4
+    #dbg_declare(ptr %x.addr, !29, !DIExpression(), !30)
+  store i32 %y, ptr %y.addr, align 4
+    #dbg_declare(ptr %y.addr, !31, !DIExpression(), !32)
+  store i32 %z, ptr %y.addr, align 4
+    #dbg_declare(ptr %z.addr, !36, !DIExpression(), !37)
+  %this1 = load ptr, ptr %this.addr, align 8
+  %0 = load i32, ptr %x.addr, align 4, !dbg !33
+  %1 = load i32, ptr %y.addr, align 4, !dbg !33
+  %2 = load i32, ptr %z.addr, align 4, !dbg !33
+  ret ptr %this1, !dbg !34
+}
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!12, !13}
+
+!0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = distinct !DIGlobalVariable(name: "a", scope: !2, file: !3, line: 3, type: !5, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !3, producer: "clang version 20.0.0git", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, globals: !4, splitDebugInlining: false, nameTableKind: Apple, sysroot: "/")
+!3 = !DIFile(filename: "object_ptr.cpp", directory: "/tmp")
+!4 = !{!0}
+!5 = distinct !DICompositeType(tag: DW_TAG_class_type, name: "A", file: !3, line: 1, size: 8, flags: DIFlagTypePassByValue | DIFlagNonTrivial, elements: !6, identifier: "_ZTS1A")
+!6 = !{!7}
+!7 = !DISubprogram(name: "A", scope: !5, file: !3, line: 2, type: !8, scopeLine: 2, flags: DIFlagPublic | DIFlagPrototyped, spFlags: 0)
+!8 = !DISubroutineType(types: !9)
+!9 = !{null, !11, !11, !10, !35}
+!10 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !5, size: 64, flags: DIFlagArtificial | DIFlagObjectPointer)
+!11 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!12 = !{i32 7, !"Dwarf Version", i32 5}
+!13 = !{i32 2, !"Debug Info Version", i32 3}
+!18 = !{!"clang version 20.0.0git"}
+!24 = distinct !DISubprogram(name: "A", linkageName: "_ZN1AC1Eii", scope: !5, file: !3, line: 2, type: !8, scopeLine: 2, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !2, declaration: !7, retainedNodes: !25)
+!25 = !{}
+!26 = !DILocalVariable(name: "this", arg: 3, scope: !24, type: !27, flags: DIFlagArtificial | DIFlagObjectPointer)
+!27 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !5, size: 64)
+!28 = !DILocation(line: 0, scope: !24)
+!29 = !DILocalVariable(name: "x", arg: 2, scope: !24, file: !3, line: 2, type: !11)
+!30 = !DILocation(line: 2, column: 19, scope: !24)
+!31 = !DILocalVariable(name: "y", arg: 1, scope: !24, file: !3, line: 2, type: !11)
+!32 = !DILocation(line: 2, column: 26, scope: !24)
+!33 = !DILocation(line: 2, column: 29, scope: !24)
+!34 = !DILocation(line: 2, column: 30, scope: !24)
+!35 = !DIBasicType(name: "short", size: 16, encoding: DW_ATE_signed)
+!36 = !DILocalVariable(name: "z", arg: 4, scope: !24, file: !3, line: 2, type: !35)
+!37 = !DILocation(line: 2, column: 35, scope: !24)

llvm/test/DebugInfo/X86/DW_AT_object_pointer.ll

@@ -1,14 +1,30 @@
-; RUN: llc -mtriple=x86_64-apple-darwin %s -o %t -filetype=obj
-; RUN: llvm-dwarfdump -v -debug-info %t | FileCheck %s
+; RUN: llc -mtriple=x86_64-apple-darwin -debugger-tune=gdb -dwarf-version=5 -filetype=obj < %s | \
+; RUN:      llvm-dwarfdump -v -debug-info - | FileCheck %s --check-prefixes=CHECK,CHECK-GDB
+
+; RUN: llc -mtriple=x86_64-apple-darwin -debugger-tune=lldb -dwarf-version=4 -filetype=obj < %s | \
+; RUN:      llvm-dwarfdump -v -debug-info - | FileCheck %s --check-prefixes=CHECK,CHECK-LLDB-DWARF4
+
+; RUN: llc -mtriple=x86_64-apple-darwin -debugger-tune=lldb -dwarf-version=5 -filetype=obj < %s | \
+; RUN:      llvm-dwarfdump -v -debug-info - | FileCheck %s --check-prefixes=CHECK,CHECK-LLDB-DWARF5
 
 ; CHECK: DW_TAG_formal_parameter [
 ; CHECK-NOT: ""
 ; CHECK: DW_TAG
 ; CHECK: DW_TAG_class_type
-; CHECK: DW_AT_object_pointer [DW_FORM_ref4]     (cu + 0x{{[0-9a-f]*}} => {[[PARAM:0x[0-9a-f]*]]})
+; CHECK: [[DECL:0x[0-9a-f]+]]: DW_TAG_subprogram
+; CHECK:                         DW_AT_name {{.*}} "A"
+; CHECK-LLDB-DWARF5:             DW_AT_object_pointer [DW_FORM_implicit_const] (0)
+; CHECK-GDB-NOT:                 DW_AT_object_pointer
+; CHECK-LLDB-DWARF4-NOT:         DW_AT_object_pointer
+; CHECK: DW_TAG_formal_parameter
+;
+; CHECK: DW_TAG_subprogram
+; CHECK:   DW_AT_object_pointer [DW_FORM_ref4]     (cu + 0x{{[0-9a-f]*}} => {[[PARAM:0x[0-9a-f]*]]})
+; CHECK:   DW_AT_specification [DW_FORM_ref4] (cu + {{.*}} => {[[DECL]]}
 ; CHECK: [[PARAM]]:     DW_TAG_formal_parameter
 ; CHECK-NOT: DW_TAG
-; CHECK: DW_AT_name [DW_FORM_strp]     ( .debug_str[0x{{[0-9a-f]*}}] = "this")
+; CHECK: DW_AT_name
+; CHECK-SAME        = "this")
 
 %class.A = type { i32 }
 

llvm/test/tools/llvm-dwarfdump/X86/statistics.ll

@@ -1,4 +1,4 @@
-; RUN: llc -O0 %s -o - -filetype=obj \
+; RUN: llc -O0 %s -o - -filetype=obj -debugger-tune=gdb -accel-tables=Apple \
 ; RUN:   | llvm-dwarfdump -statistics - | FileCheck %s
 ; CHECK: "version": 9,
 
@@ -55,7 +55,7 @@
 ; CHECK:      "#bytes within functions": [[FUNCSIZE:[0-9]+]]
 ; CHECK:      "#bytes within inlined functions": [[INLINESIZE:[0-9]+]]
 ; CHECK:      "#bytes in __debug_loc": 35,
-; CHECK-NEXT: "#bytes in __debug_abbrev": 384,
+; CHECK-NEXT: "#bytes in __debug_abbrev": 375,
 ; CHECK-NEXT: "#bytes in __debug_info": 459,
 ; CHECK-NEXT: "#bytes in __debug_str": 231,
 ; CHECK-NEXT: "#bytes in __apple_names": 348,