[DebugInfo][DWARF] Add heapallocsite information (llvm#132073)

LLVM currently stores heapallocsite information in CodeView debuginfo,
but not in DWARF debuginfo. Plumb it into DWARF as an LLVM-specific
extension.

heapallocsite debug information is useful when it is combined with
allocator instrumentation that stores caller addresses; I've used a
previous version of this patch for:

 - analyzing memory usage by object type
 - analyzing the distributions of values of class members

Other possible uses might be:

- attributing memory access profiles (for example, on Intel CPUs, from
PEBS records with Linear Data Address) to object types or specific
object members
 - adding type information to crash/ASAN reports

Author: thejh

llvm/include/llvm/BinaryFormat/Dwarf.def

@@ -625,6 +625,9 @@ HANDLE_DW_AT(0x3e0a, LLVM_ptrauth_authentication_mode, 0, LLVM)
 HANDLE_DW_AT(0x3e0b, LLVM_num_extra_inhabitants, 0, LLVM)
 HANDLE_DW_AT(0x3e0c, LLVM_stmt_sequence, 0, LLVM)
 HANDLE_DW_AT(0x3e0d, LLVM_coro_suspend_idx, 0, LLVM)
+// The DWARF v6 working draft defines DW_AT_alloc_type; use this LLVM-private ID
+// until that is released as an official standard.
+HANDLE_DW_AT(0x3e0e, LLVM_alloc_type, 0, LLVM)
 
 // Apple extensions.
 

llvm/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp

@@ -1292,12 +1292,10 @@ DwarfCompileUnit::getDwarf5OrGNULocationAtom(dwarf::LocationAtom Loc) const {
   }
 }
 
-DIE &DwarfCompileUnit::constructCallSiteEntryDIE(DIE &ScopeDIE,
-                                                 const DISubprogram *CalleeSP,
-                                                 bool IsTail,
-                                                 const MCSymbol *PCAddr,
-                                                 const MCSymbol *CallAddr,
-                                                 unsigned CallReg) {
+DIE &DwarfCompileUnit::constructCallSiteEntryDIE(
+    DIE &ScopeDIE, const DISubprogram *CalleeSP, bool IsTail,
+    const MCSymbol *PCAddr, const MCSymbol *CallAddr, unsigned CallReg,
+    DIType *AllocSiteTy) {
   // Insert a call site entry DIE within ScopeDIE.
   DIE &CallSiteDIE = createAndAddDIE(getDwarf5OrGNUTag(dwarf::DW_TAG_call_site),
                                      ScopeDIE, nullptr);
@@ -1306,7 +1304,7 @@ DIE &DwarfCompileUnit::constructCallSiteEntryDIE(DIE &ScopeDIE,
     // Indirect call.
     addAddress(CallSiteDIE, getDwarf5OrGNUAttr(dwarf::DW_AT_call_target),
                MachineLocation(CallReg));
-  } else {
+  } else if (CalleeSP) {
     DIE *CalleeDIE = getOrCreateSubprogramDIE(CalleeSP);
     assert(CalleeDIE && "Could not create DIE for call site entry origin");
     if (AddLinkageNamesToDeclCallOriginsForTuning(DD) &&
@@ -1351,6 +1349,9 @@ DIE &DwarfCompileUnit::constructCallSiteEntryDIE(DIE &ScopeDIE,
                     getDwarf5OrGNUAttr(dwarf::DW_AT_call_return_pc), PCAddr);
   }
 
+  if (AllocSiteTy)
+    addType(CallSiteDIE, AllocSiteTy, dwarf::DW_AT_LLVM_alloc_type);
+
   return CallSiteDIE;
 }
 

llvm/lib/CodeGen/AsmPrinter/DwarfCompileUnit.h

@@ -289,7 +289,8 @@ class DwarfCompileUnit final : public DwarfUnit {
   /// the \p CallReg is set to 0.
   DIE &constructCallSiteEntryDIE(DIE &ScopeDIE, const DISubprogram *CalleeSP,
                                  bool IsTail, const MCSymbol *PCAddr,
-                                 const MCSymbol *CallAddr, unsigned CallReg);
+                                 const MCSymbol *CallAddr, unsigned CallReg,
+                                 DIType *AllocSiteTy);
   /// Construct call site parameter DIEs for the \p CallSiteDIE. The \p Params
   /// were collected by the \ref collectCallSiteParameters.
   /// Note: The order of parameters does not matter, since debuggers recognize

llvm/lib/CodeGen/AsmPrinter/DwarfDebug.cpp

@@ -936,6 +936,8 @@ void DwarfDebug::constructCallSiteEntryDIEs(const DISubprogram &SP,
       if (MI.hasDelaySlot() && !delaySlotSupported(*&MI))
         return;
 
+      DIType *AllocSiteTy = dyn_cast_or_null<DIType>(MI.getHeapAllocMarker());
+
       // If this is a direct call, find the callee's subprogram.
       // In the case of an indirect call find the register that holds
       // the callee.
@@ -950,23 +952,23 @@ void DwarfDebug::constructCallSiteEntryDIEs(const DISubprogram &SP,
         PhysRegCalleeOperand =
             PhysRegCalleeOperand && MCOI.OperandType == MCOI::OPERAND_REGISTER;
       }
-      if (!CalleeOp.isGlobal() && !PhysRegCalleeOperand)
-        continue;
 
       unsigned CallReg = 0;
       const DISubprogram *CalleeSP = nullptr;
       const Function *CalleeDecl = nullptr;
       if (PhysRegCalleeOperand) {
-        CallReg = CalleeOp.getReg();
-        if (!CallReg)
-          continue;
-      } else {
+        CallReg = CalleeOp.getReg(); // might be zero
+      } else if (CalleeOp.isGlobal()) {
         CalleeDecl = dyn_cast<Function>(CalleeOp.getGlobal());
-        if (!CalleeDecl || !CalleeDecl->getSubprogram())
-          continue;
-        CalleeSP = CalleeDecl->getSubprogram();
+        if (CalleeDecl)
+          CalleeSP = CalleeDecl->getSubprogram(); // might be nullptr
       }
 
+      // Omit DIE if we can't tell where the call goes *and* we don't want to
+      // add metadata to it.
+      if (CalleeSP == nullptr && CallReg == 0 && AllocSiteTy == nullptr)
+        continue;
+
       // TODO: Omit call site entries for runtime calls (objc_msgSend, etc).
 
       bool IsTail = TII->isTailCall(MI);
@@ -1000,7 +1002,7 @@ void DwarfDebug::constructCallSiteEntryDIEs(const DISubprogram &SP,
                         << (IsTail ? " [IsTail]" : "") << "\n");
 
       DIE &CallSiteDIE = CU.constructCallSiteEntryDIE(
-          ScopeDIE, CalleeSP, IsTail, PCAddr, CallAddr, CallReg);
+          ScopeDIE, CalleeSP, IsTail, PCAddr, CallAddr, CallReg, AllocSiteTy);
 
       // Optionally emit call-site-param debug info.
       if (emitDebugEntryValues()) {

llvm/lib/DWARFLinker/Classic/DWARFLinker.cpp

@@ -109,6 +109,7 @@ static bool isODRAttribute(uint16_t Attr) {
   case dwarf::DW_AT_specification:
   case dwarf::DW_AT_abstract_origin:
   case dwarf::DW_AT_import:
+  case dwarf::DW_AT_LLVM_alloc_type:
     return true;
   }
   llvm_unreachable("Improper attribute.");

llvm/test/DebugInfo/X86/DW_AT_alloc_type.ll

@@ -0,0 +1,34 @@
+; RUN: llc -O3 -o %t -filetype=obj %s
+; RUN: llvm-dwarfdump %t | FileCheck %s
+
+; based on clang++ output for `int *alloc_int() { return new int; }`
+
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define dso_local ptr @alloc_int() !dbg !3 {
+; CHECK: DW_TAG_subprogram
+entry:
+  %call = call ptr @alloc(i64 noundef 4), !heapallocsite !7
+; CHECK: DW_TAG_call_site
+; CHECK: DW_AT_LLVM_alloc_type ([[ALLOCSITE:.*]])
+  ret ptr %call
+}
+
+; CHECK: {{.*}}[[ALLOCSITE]]: DW_TAG_base_type
+; CHECK: DW_AT_name ("int")
+
+declare dso_local ptr @alloc(i64 noundef)
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2,!8}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, emissionKind: FullDebug)
+!1 = !DIFile(filename: "a.cpp", directory: "/")
+!2 = !{i32 2, !"Debug Info Version", i32 3}
+!3 = distinct !DISubprogram(name: "alloc_int", scope: !1, file: !1, line: 1, type: !4, scopeLine: 1, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition, unit: !0)
+!4 = !DISubroutineType(types: !5)
+!5 = !{!6}
+!6 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !7, size: 64)
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{i32 2, !"Dwarf Version", i32 5}