[Assignment Tracking] Coalesce dbg loc definitions with contiguous fragments

[Orlando]

Summary

MemLocFragmentFill uses an IntervalMap to track which bits of each variable are stack-homed. Intervals with the same value (same stack location base address) are automatically coalesced by the map. This patch changes the analysis to take advantage of that and insert a new dbg loc after each def if any coalescing took place. This results in some additional redundant defs (we insert a def, then another that by definition shadows the previous one if any coalescing took place) but they're all cleaned up thanks to the previous patch in this stack.

This reduces the total number of fragments created by AssignmentTrackingAnalysis which reduces compile time because LiveDebugValues computes SSA for every fragment it encounters. There's a geomean reduction in instructions retired in a CTMark LTO-O3-g build of 0.3% with these two patches.

One small caveat is that this technique can produce partially overlapping fragments (e.g. slice [0, 32) and slice [16, 64)), which we know LiveDebugVariables doesn't really handle correctly. Used in combination with instruction-referencing this isn't a problem, since LiveDebugVariables is effectively side-stepped in instruction-referencing mode.

[jmorse]

Good saving; in the future we might find different ways of representing variable fragments and communicating them to LiveDebugValues, or make LDV smarter, but it's good to get this benefit now. I think the more structured fragment model @scott.linder proposed would deliver the sort of guarantees we'd need.

As discussed offline, I think the only thing missing is making this behaviour conditional on instr-ref variable locations being used, to avoid LiveDebugVariables tripping over it. Rather than seeing it as "assignment tracking needs instruction referencing" and jamming it in the frontnend, we can consider it as being "instruction referencing needs this hack to deal with assignment tracking efficiently". That allows other backends to benefit from the work.

llvm/lib/CodeGen/AssignmentTrackingAnalysis.cpp

605–609

I wonder if there's a pathological case where we effectively store a history of defs that we then have to manually clear up, at reasonable expense.

Not really a criticism, everything in compilers has a pathological case, and this is clearly a net positive for compilation overall.

704–708

In this scenario, would I be right in thinking that the code below shortens "i", inserts "f" in the middle, and then IntervalMap will coalesce them back together if they're the same? (Seems like the right behaviour).

[Orlando]

+ Check if instr-ref is turned on before coalescing fragments (see true table in coalesce-options.ll)
+ Add test coalesce-options.ll

Note that the check uses debuginfoShouldUseDebugInstrRef which misses out the -O0 and ((optnone)) checks that the MachineFunction variation use. Assignment tracking is currently disabled for -O0 (note to self: we should check ((optnone)) in AssignmentTrackingPass too) but either way locations that require coalescing shouldn't be created if there are no optimisations - so no harm should come of this slightly permissive interpretation of instruction-reference enabled-ness.

[Orlando]

llvm/lib/CodeGen/AssignmentTrackingAnalysis.cpp

605–609

I think it would be possible to be smarter about when to insert certain defs or not, but this solution seemed to yield decent results with fairly trivial changes. It could be an area to look at in future rounds of performance improvements.

704–708

Yes exactly that.

[Orlando]

note to self: we should check ((optnone)) in AssignmentTrackingPass too

done D147129

[jmorse]

LGTM

llvm/lib/CodeGen/AssignmentTrackingAnalysis.cpp

 #include "llvm/CodeGen/AssignmentTrackingAnalysis.h"
+#include "LiveDebugValues/LiveDebugValues.h"
 #include "llvm/ADT/DenseMapInfo.h"
 #include "llvm/ADT/IntervalMap.h"
 #include "llvm/ADT/PostOrderIterator.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/UniqueVector.h"
 #include "llvm/Analysis/Interval.h"

 /// Option for debugging the pass, determines if the memory location fragment
 /// filling happens after generating the variable locations.
 static cl::opt<bool> EnableMemLocFragFill("mem-loc-frag-fill", cl::init(true),
                                           cl::Hidden);
 /// Print the results of the analysis. Respects -filter-print-funcs.
 static cl::opt<bool> PrintResults("print-debug-ata", cl::init(false),
                                   cl::Hidden);
 
+/// Coalesce adjacent dbg locs describing memory locations that have contiguous
+/// fragments. This reduces the cost of LiveDebugValues which does SSA
+/// construction for each explicitly stated variable fragment.
+static cl::opt<cl::boolOrDefault>
+    CoalesceAdjacentFragmentsOpt("debug-ata-coalesce-frags", cl::Hidden);
+
 // Implicit conversions are disabled for enum class types, so unfortunately we
 // need to create a DenseMapInfo wrapper around the specified underlying type.
 template <> struct llvm::DenseMapInfo<VariableID> {
   using Wrapped = DenseMapInfo<unsigned>;
   static inline VariableID getEmptyKey() {
     return static_cast<VariableID>(Wrapped::getEmptyKey());
   }
   static inline VariableID getTombstoneKey() {

 using DebugAggregate = std::pair<const DILocalVariable *, const DILocation *>;
 static DebugAggregate getAggregate(const DbgVariableIntrinsic *DII) {
   return DebugAggregate(DII->getVariable(), DII->getDebugLoc().getInlinedAt());
 }
 static DebugAggregate getAggregate(const DebugVariable &Var) {
   return DebugAggregate(Var.getVariable(), Var.getInlinedAt());
 }
 
+static bool shouldCoalesceFragments(Function &F) {
+  // Enabling fragment coalescing reduces compiler run time when instruction
+  // referencing is enabled. However, it may cause LiveDebugVariables to create
+  // incorrect locations. Since instruction-referencing mode effectively
+  // bypasses LiveDebugVariables we only enable coalescing if the cl::opt flag
+  // has not been explicitly set and instruction-referencing is turned on.
+  switch (CoalesceAdjacentFragmentsOpt) {
+  case cl::boolOrDefault::BOU_UNSET:
+    return debuginfoShouldUseDebugInstrRef(
+        Triple(F.getParent()->getTargetTriple()));
+  case cl::boolOrDefault::BOU_TRUE:
+    return true;
+  case cl::boolOrDefault::BOU_FALSE:
+    return false;
+  }
+  llvm_unreachable("Unknown boolOrDefault value");
+}
+
 namespace {
 /// In dwarf emission, the following sequence
 ///    1. dbg.value ... Fragment(0, 64)
 ///    2. dbg.value ... Fragment(0, 32)
 /// effectively sets Fragment(32, 32) to undef (each def sets all bits not in
 /// the intersection of the fragments to having "no location"). This makes
 /// sense for implicit location values because splitting the computed values
 /// could be troublesome, and is probably quite uncommon.  When we convert
 /// dbg.assigns to dbg.value+deref this kind of thing is common, and describing
 /// a location (memory) rather than a value means we don't need to worry about
 /// splitting any values, so we try to recover the rest of the fragment
 /// location here.
 /// This class performs a(nother) dataflow analysis over the function, adding
 /// variable locations so that any bits of a variable with a memory location
 /// have that location explicitly reinstated at each subsequent variable
 /// location definition that that doesn't overwrite those bits. i.e. after a
 /// variable location def, insert new defs for the memory location with
 /// fragments for the difference of "all bits currently in memory" and "the
 /// fragment of the second def".
 class MemLocFragmentFill {
   Function &Fn;
   FunctionVarLocsBuilder *FnVarLocs;
   const DenseSet<DebugAggregate> *VarsWithStackSlot;
+  bool CoalesceAdjacentFragments;
 
   // 0 = no memory location.
   using BaseAddress = unsigned;
   using OffsetInBitsTy = unsigned;
   using FragTraits = IntervalMapHalfOpenInfo<OffsetInBitsTy>;
   using FragsInMemMap = IntervalMap<
       OffsetInBitsTy, BaseAddress,
       IntervalMapImpl::NodeSizer<OffsetInBitsTy, BaseAddress>::LeafSize,
void insertMemLoc(BasicBlock &BB, Instruction &Before, unsigned Var,
     assert(Base && "Expected a non-zero ID for Base address");
     Loc.Base = Base;
     Loc.DL = DL;
     BBInsertBeforeMap[&BB][&Before].push_back(Loc);
     LLVM_DEBUG(dbgs() << "Add mem def for " << Aggregates[Var].first->getName()
                       << " bits [" << StartBit << ", " << EndBit << ")\n");
   }
 
+  /// Inserts a new dbg def if the interval found when looking up \p StartBit
+  /// in \p FragMap starts before \p StartBit or ends after \p EndBit (which
+  /// indicates - assuming StartBit->EndBit has just been inserted - that the
+  /// slice has been coalesced in the map).
+  void coalesceFragments(BasicBlock &BB, Instruction &Before, unsigned Var,
+                         unsigned StartBit, unsigned EndBit, unsigned Base,
+                         DebugLoc DL, const FragsInMemMap &FragMap) {
+    if (!CoalesceAdjacentFragments)
+      return;
+    // We've inserted the location into the map. The map will have coalesced
+    // adjacent intervals (variable fragments) that describe the same memory
+    // location. Use this knowledge to insert a debug location that describes
+    // that coalesced fragment. This may eclipse other locs we've just
+    // inserted. This is okay as redundant locs will be cleaned up later.
+    auto CoalescedFrag = FragMap.find(StartBit);
+    // Bail if no coalescing has taken place.
+    if (CoalescedFrag.start() == StartBit && CoalescedFrag.stop() == EndBit)
+      return;
+
+    LLVM_DEBUG(dbgs() << "- Insert loc for bits " << CoalescedFrag.start()
+                      << " to " << CoalescedFrag.stop() << "\n");
+    insertMemLoc(BB, Before, Var, CoalescedFrag.start(), CoalescedFrag.stop(),
+                 Base, DL);
+  }
+
   void addDef(const VarLocInfo &VarLoc, Instruction &Before, BasicBlock &BB,
               VarFragMap &LiveSet) {
     DebugVariable DbgVar = FnVarLocs->getVariable(VarLoc.VariableID);
     if (skipVariable(DbgVar.getVariable()))
       return;
     // Don't bother doing anything for this variables if we know it's fully
     // promoted. We're only interested in variables that (sometimes) live on
     // the stack here.
void addDef(const VarLocInfo &VarLoc, Instruction &Before, BasicBlock &BB,
     // The variable has an entry in the map.
 
     FragsInMemMap &FragMap = FragIt->second;
     // First check the easy case: the new fragment `f` doesn't overlap with any
     // intervals.
     if (!FragMap.overlaps(StartBit, EndBit)) {
       LLVM_DEBUG(dbgs() << "- No overlaps\n");
       FragMap.insert(StartBit, EndBit, Base);
+      coalesceFragments(BB, Before, Var, StartBit, EndBit, Base, VarLoc.DL,
+                        FragMap);
       return;
     }
     // There is at least one overlap.
 
     // Does StartBit intersect an existing fragment?
     auto FirstOverlap = FragMap.find(StartBit);
     assert(FirstOverlap != FragMap.end());
     bool IntersectStart = FirstOverlap.start() < StartBit;
 
     // Does EndBit intersect an existing fragment?
     auto LastOverlap = FragMap.find(EndBit);
     bool IntersectEnd = LastOverlap.valid() && LastOverlap.start() < EndBit;
 
     // Check if both ends of `f` intersect the same interval `i`.
     if (IntersectStart && IntersectEnd && FirstOverlap == LastOverlap) {
       LLVM_DEBUG(dbgs() << "- Intersect single interval @ both ends\n");
       // Shorten `i` so that there's space to insert `f`.
       //      [ f ]
       // [  -   i   -  ]
       // +
       // [ i ][ f ][ i ]
 
       // Save values for use after inserting a new interval.
       auto EndBitOfOverlap = FirstOverlap.stop();
       unsigned OverlapValue = FirstOverlap.value();
 
       // Shorten the overlapping interval.
       FirstOverlap.setStop(StartBit);
       insertMemLoc(BB, Before, Var, FirstOverlap.start(), StartBit,
if (IntersectStart && IntersectEnd && FirstOverlap == LastOverlap) {
         LLVM_DEBUG(dbgs() << "- Erase " << toString(It));
         It.erase(); // This increments It after removing the interval.
       }
       // We've dealt with all the overlaps now!
       assert(!FragMap.overlaps(StartBit, EndBit));
       LLVM_DEBUG(dbgs() << "- Insert DEF into now-empty space\n");
       FragMap.insert(StartBit, EndBit, Base);
     }
+
+    coalesceFragments(BB, Before, Var, StartBit, EndBit, Base, VarLoc.DL,
+                      FragMap);
   }
 
   bool skipVariable(const DILocalVariable *V) { return !V->getSizeInBits(); }
 
   void process(BasicBlock &BB, VarFragMap &LiveSet) {
     BBInsertBeforeMap[&BB].clear();
     for (auto &I : BB) {
       if (const auto *Locs = FnVarLocs->getWedge(&I)) {
         for (const VarLocInfo &Loc : *Locs) {
           addDef(Loc, I, *I.getParent(), LiveSet);
         }
       }
     }
   }
 
 public:
   MemLocFragmentFill(Function &Fn,
-                     const DenseSet<DebugAggregate> *VarsWithStackSlot)
-      : Fn(Fn), VarsWithStackSlot(VarsWithStackSlot) {}
+                     const DenseSet<DebugAggregate> *VarsWithStackSlot,
+                     bool CoalesceAdjacentFragments)
+      : Fn(Fn), VarsWithStackSlot(VarsWithStackSlot),
+        CoalesceAdjacentFragments(CoalesceAdjacentFragments) {}
 
   /// Add variable locations to \p FnVarLocs so that any bits of a variable
   /// with a memory location have that location explicitly reinstated at each
   /// subsequent variable location definition that that doesn't overwrite those
   /// bits. i.e. after a variable location def, insert new defs for the memory
   /// location with fragments for the difference of "all bits currently in
   /// memory" and "the fragment of the second def". e.g.
   ///
for (auto Pair : BBInsertBeforeMap) {
       for (auto Pair : Map) {
         Instruction *InsertBefore = Pair.first;
         assert(InsertBefore && "should never be null");
         auto FragMemLocs = Pair.second;
         auto &Ctx = Fn.getContext();
 
         for (auto FragMemLoc : FragMemLocs) {
           DIExpression *Expr = DIExpression::get(Ctx, std::nullopt);
+          if (FragMemLoc.SizeInBits !=
+              *Aggregates[FragMemLoc.Var].first->getSizeInBits())
             Expr = *DIExpression::createFragmentExpression(
                 Expr, FragMemLoc.OffsetInBits, FragMemLoc.SizeInBits);
           Expr = DIExpression::prepend(Expr, DIExpression::DerefAfter,
                                        FragMemLoc.OffsetInBits / 8);
           DebugVariable Var(Aggregates[FragMemLoc.Var].first, Expr,
                             FragMemLoc.DL.getInlinedAt());
           FnVarLocs->addVarLoc(InsertBefore, Var, Expr, FragMemLoc.DL,
                                Bases[FragMemLoc.Base]);
         }
       }
static void analyzeFunction(Function &Fn, const DataLayout &Layout,
   // Use a scope block to clean up AssignmentTrackingLowering before running
   // MemLocFragmentFill to reduce peak memory consumption.
   {
     AssignmentTrackingLowering Pass(Fn, Layout, &VarsWithStackSlot);
     Changed = Pass.run(FnVarLocs);
   }
 
   if (Changed) {
-    MemLocFragmentFill Pass(Fn, &VarsWithStackSlot);
+    MemLocFragmentFill Pass(Fn, &VarsWithStackSlot,
+                            shouldCoalesceFragments(Fn));
     Pass.run(FnVarLocs);
 
     // Remove redundant entries. As well as reducing memory consumption and
     // avoiding waiting cycles later by burning some now, this has another
     // important job. That is to work around some SelectionDAG quirks. See
     // removeRedundantDbgLocsUsingForwardScan comments for more info on that.
     for (auto &BB : Fn)
       removeRedundantDbgLocs(&BB, *FnVarLocs);

llvm/test/DebugInfo/assignment-tracking/X86/coalesce-cfg.ll

+; RUN: llc %s -o - -stop-after=finalize-isel \
+; RUN: | FileCheck %s --implicit-check-not=DBG_
+
+;; Test coalescing of contiguous fragments in adjacent location definitions.
+;; Further details and check directives inline.
+
+target triple = "x86_64-unknown-linux-gnu"
+
+@cond = dso_local global i8 0, align 1
+
+;; The final store and linked dbg.assign indicate the whole variable is located
+;; on the stack. Coalesce the two fragment defs that are generated (0-32,
+;; 32-64) at the final dbg.assign into one (0-64, which covers the whole
+;; variable meaning we don't need a fragment expression). And check the two
+;; DBG_VALUEs in if.then are not coalesced, since they specify different
+;; locations. This is the same as the first test in coalesce-simple.ll except
+;; the dbg intrinsics are split up over a simple diamond CFG to check the info
+;; is propagated betweeen blocks correctly.
+
+; CHECK-LABEL: bb.0.entry:
+; CHECK-NEXT: successors:
+; CHECK-NEXT: {{^ *$}}
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref)
+; CHECK-NEXT: TEST8mi
+; CHECK-NEXT: JCC_1 %bb.2
+; CHECK-NEXT: JMP_1 %bb.1
+
+; CHECK-LABEL: bb.1.if.then:
+; CHECK-NEXT: successors:
+; CHECK-NEXT: {{^ *$}}
+; CHECK-NEXT: MOV8mi $rip, 1, $noreg, @cond, $noreg, 0 :: (store (s8) into @cond)
+; CHECK-NEXT: DBG_VALUE 1, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
+; CHECK-NEXT: JMP_1 %bb.3
+
+; CHECK-LABEL: bb.2.if.else:
+; CHECK-NEXT: successors:
+; CHECK-NEXT: {{^ *$}}
+; CHECK-NEXT:   MOV8mi $rip, 1, $noreg, @cond, $noreg, 1 :: (store (s8) into @cond)
+
+; CHECK-LABEL: bb.3.if.end:
+; CHECK-NEXT:   MOV32mi %stack.0.a, 1, $noreg, 0, $noreg, 5 :: (store (s32) into %ir.a, align 8)
+; CHECK-NEXT:   DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref)
+; CHECK-NEXT:   RET 0
+
+define dso_local void @_Z3funv() local_unnamed_addr !dbg !16 {
+entry:
+  %a = alloca i64, !DIAssignID !37
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !20, metadata !DIExpression(), metadata !37, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  %0 = load i8, ptr @cond, align 1
+  %tobool = trunc i8 %0 to i1
+  br i1 %tobool, label %if.then, label %if.else
+
+if.then:
+  store i1 false, ptr @cond
+  call void @llvm.dbg.value(metadata i32 1, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32)), !dbg !25
+  br label %if.end
+
+if.else:
+  store i1 true, ptr @cond
+  br label %if.end
+
+if.end:
+  store i32 5, ptr %a, !DIAssignID !38
+  call void @llvm.dbg.assign(metadata i32 5, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !38, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  ret void
+}
+
+declare void @llvm.dbg.value(metadata, metadata, metadata)
+declare void @llvm.dbg.assign(metadata, metadata, metadata, metadata, metadata, metadata)
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!8, !9, !10, !11, !12, !13, !14}
+!llvm.ident = !{!15}
+
+!0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = distinct !DIGlobalVariable(name: "G", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !3, producer: "clang version 17.0.0", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, globals: !4, splitDebugInlining: false, nameTableKind: None)
+!3 = !DIFile(filename: "test.cpp", directory: "/")
+!4 = !{!0, !5}
+!5 = !DIGlobalVariableExpression(var: !6, expr: !DIExpression())
+!6 = distinct !DIGlobalVariable(name: "F", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{i32 7, !"Dwarf Version", i32 5}
+!9 = !{i32 2, !"Debug Info Version", i32 3}
+!10 = !{i32 1, !"wchar_size", i32 4}
+!11 = !{i32 8, !"PIC Level", i32 2}
+!12 = !{i32 7, !"PIE Level", i32 2}
+!13 = !{i32 7, !"uwtable", i32 2}
+!14 = !{i32 7, !"debug-info-assignment-tracking", i1 true}
+!15 = !{!"clang version 17.0.0"}
+!16 = distinct !DISubprogram(name: "fun", linkageName: "_Z3funv", scope: !3, file: !3, line: 3, type: !17, scopeLine: 3, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !19)
+!17 = !DISubroutineType(types: !18)
+!18 = !{null}
+!19 = !{!20}
+!20 = !DILocalVariable(name: "X", scope: !16, file: !3, line: 4, type: !21)
+!21 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Pair", file: !3, line: 2, size: 64, flags: DIFlagTypePassByValue, elements: !22, identifier: "_ZTS4Pair")
+!22 = !{}
+!25 = !DILocation(line: 0, scope: !16)
+!26 = !DILocation(line: 7, column: 7, scope: !27)
+!27 = distinct !DILexicalBlock(scope: !16, file: !3, line: 7, column: 7)
+!28 = distinct !DIAssignID()
+!37 = distinct !DIAssignID()
+!38 = distinct !DIAssignID()

llvm/test/DebugInfo/assignment-tracking/X86/coalesce-options.ll

+;; Test coalescing of contiguous fragments in adjacent location definitions.
+;; This test contains the first function from coalesce-simple.ll. Just use it
+;; to check whether coalescing happens or not with different flag settings.
+;;
+;; +=================+==============================+======================+
+;; | Coalescing flag | Instruction-Referencing flag | Coalescing behaviour |
+;; +=================+==============================+======================+
+;; | default         | enabled                      | enabled              |
+;; | default         | disabled                     | disabled             |
+;; | enabled         | *                            | enabled              |
+;; | disabled        | *                            | disabled             |
+;; +-----------------+------------------------------+----------------------+
+
+;; Coalescing default + instructino-referencing enabled = enable.
+; RUN: llc %s -o - -stop-after=finalize-isel -experimental-debug-variable-locations=true \
+; RUN: | FileCheck %s --check-prefixes=CHECK,ENABLE
+
+;; Coalescing default + instructino-referencing disabled = disable.
+; RUN: llc %s -o - -stop-after=finalize-isel -experimental-debug-variable-locations=false \
+; RUN: | FileCheck %s --check-prefixes=CHECK,DISABLE
+
+;; Coalescing enabled + instructino-referencing disabled = enable.
+; RUN: llc %s -o - -stop-after=finalize-isel -experimental-debug-variable-locations=false \
+; RUN:     -debug-ata-coalesce-frags=true \
+; RUN: | FileCheck %s --check-prefixes=CHECK,ENABLE
+
+;; Coalescing disabled + instructino-referencing enabled = disable.
+; RUN: llc %s -o - -stop-after=finalize-isel -experimental-debug-variable-locations=true \
+; RUN:     -debug-ata-coalesce-frags=false \
+; RUN: | FileCheck %s --check-prefixes=CHECK,DISABLE
+
+; CHECK: MOV32mi %stack.0.a, 1, $noreg, 0, $noreg, 5
+; ENABLE-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref)
+; DISABLE-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32)
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define dso_local void @_Z3funv() local_unnamed_addr !dbg !16 {
+entry:
+  %a = alloca i64, !DIAssignID !37
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !20, metadata !DIExpression(), metadata !37, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  call void @llvm.dbg.value(metadata i32 1, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32)), !dbg !25
+  store i32 5, ptr %a, !DIAssignID !38
+  call void @llvm.dbg.assign(metadata i32 5, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !38, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  ret void
+}
+
+declare void @llvm.dbg.value(metadata, metadata, metadata)
+declare void @llvm.dbg.assign(metadata, metadata, metadata, metadata, metadata, metadata)
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!8, !9, !10, !11, !12, !13, !14}
+!llvm.ident = !{!15}
+
+!0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = distinct !DIGlobalVariable(name: "G", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !3, producer: "clang version 17.0.0", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, globals: !4, splitDebugInlining: false, nameTableKind: None)
+!3 = !DIFile(filename: "test.cpp", directory: "/")
+!4 = !{!0, !5}
+!5 = !DIGlobalVariableExpression(var: !6, expr: !DIExpression())
+!6 = distinct !DIGlobalVariable(name: "F", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{i32 7, !"Dwarf Version", i32 5}
+!9 = !{i32 2, !"Debug Info Version", i32 3}
+!10 = !{i32 1, !"wchar_size", i32 4}
+!11 = !{i32 8, !"PIC Level", i32 2}
+!12 = !{i32 7, !"PIE Level", i32 2}
+!13 = !{i32 7, !"uwtable", i32 2}
+!14 = !{i32 7, !"debug-info-assignment-tracking", i1 true}
+!15 = !{!"clang version 17.0.0"}
+!16 = distinct !DISubprogram(name: "fun", linkageName: "_Z3funv", scope: !3, file: !3, line: 3, type: !17, scopeLine: 3, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !19)
+!17 = !DISubroutineType(types: !18)
+!18 = !{null}
+!19 = !{!20}
+!20 = !DILocalVariable(name: "X", scope: !16, file: !3, line: 4, type: !21)
+!21 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Pair", file: !3, line: 2, size: 64, flags: DIFlagTypePassByValue, elements: !22, identifier: "_ZTS4Pair")
+!22 = !{}
+!25 = !DILocation(line: 0, scope: !16)
+!26 = !DILocation(line: 7, column: 7, scope: !27)
+!27 = distinct !DILexicalBlock(scope: !16, file: !3, line: 7, column: 7)
+!37 = distinct !DIAssignID()
+!38 = distinct !DIAssignID()

llvm/test/DebugInfo/assignment-tracking/X86/coalesce-simple.ll

+; RUN: llc %s -o - -stop-after=finalize-isel \
+; RUN: | FileCheck %s --implicit-check-not=DBG_
+
+;; Test coalescing of contiguous fragments in adjacent location definitions.
+;; Further details and check directives inline.
+
+target triple = "x86_64-unknown-linux-gnu"
+
+;; The final store and linked dbg.assign indicate the whole variable is located
+;; on the stack. Coalesce the two fragment defs that are generated (0-32,
+;; 32-64) at the final dbg.assign into one (0-64, which covers the whole
+;; variable meaning we don't need a fragment expression). And check the
+;; first two DBG_VALUEs are not coalesced, since they specify different
+;; locations.
+; CHECK: _Z3fun
+; CHECK-LABEL: bb.0.entry:
+; CHECK-NEXT: DBG_VALUE 1, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
+; CHECK-NEXT: MOV32mi %stack.0.a, 1, $noreg, 0, $noreg, 5
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref)
+; CHECK-NEXT: RET
+define dso_local void @_Z3funv() local_unnamed_addr !dbg !16 {
+entry:
+  %a = alloca i64, !DIAssignID !37
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !20, metadata !DIExpression(), metadata !37, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  call void @llvm.dbg.value(metadata i32 1, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32)), !dbg !25
+  store i32 5, ptr %a, !DIAssignID !38
+  call void @llvm.dbg.assign(metadata i32 5, metadata !20, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !38, metadata ptr %a, metadata !DIExpression()), !dbg !25
+  ret void
+}
+
+;; Similar to the test above except that the variable has been split over two
+;; allocas, so coalescing should not take place (different memory location for
+;; the fragments).
+; CHECK: Z3funv2
+; CHECK-LABEL: bb.0.entry:
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT: DBG_VALUE %stack.1.b, $noreg, ![[#]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
+; CHECK-NEXT: DBG_VALUE 1, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT: MOV32mi %stack.0.a, 1, $noreg, 0, $noreg, 5
+;; Both fragments 0-32 and 32-64 are in memory now, but located in different stack slots.
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT: RET
+define dso_local void @_Z3funv2() local_unnamed_addr !dbg !39 {
+entry:
+  %a = alloca i32, !DIAssignID !42
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !41, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !42, metadata ptr %a, metadata !DIExpression()), !dbg !44
+  %b = alloca i32, !DIAssignID !45
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !41, metadata !DIExpression(DW_OP_LLVM_fragment, 32, 32), metadata !45, metadata ptr %b, metadata !DIExpression()), !dbg !44
+  call void @llvm.dbg.value(metadata i32 1, metadata !41, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32)), !dbg !44
+  store i32 5, ptr %a, !DIAssignID !43
+  call void @llvm.dbg.assign(metadata i32 5, metadata !41, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !43, metadata ptr %a, metadata !DIExpression()), !dbg !44
+  ret void
+}
+
+;; Similar to the first test above except the part that it's the slice 16-32
+;; that is "partially promoted". The dbg defs after the alloca cannot be
+;; coalesced (slices 0-16 and 32-64 are in memory but 16-32 isn't). The entire
+;; variable is on the stack after the store.
+; CHECK: _Z2funv3
+; CHECK-LABEL: bb.0.entry:
+; CHECK-NEXT: DBG_VALUE 2, $noreg, ![[#]], !DIExpression(DW_OP_LLVM_fragment, 16, 16)
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 16)
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
+; CHECK-NEXT: MOV32mi %stack.0.a, 1, $noreg, 0, $noreg, 5
+; CHECK-NEXT: DBG_VALUE %stack.0.a, $noreg, ![[#]], !DIExpression(DW_OP_deref)
+; CHECK-NEXT: RET
+define dso_local void @_Z2funv3() local_unnamed_addr !dbg !46 {
+entry:
+  %a = alloca i64, !DIAssignID !49
+  call void @llvm.dbg.assign(metadata i64 poison, metadata !48, metadata !DIExpression(), metadata !49, metadata ptr %a, metadata !DIExpression()), !dbg !51
+  call void @llvm.dbg.value(metadata i32 2, metadata !48, metadata !DIExpression(DW_OP_LLVM_fragment, 16, 16)), !dbg !51
+  store i32 5, ptr %a, !DIAssignID !50
+  call void @llvm.dbg.assign(metadata i32 5, metadata !48, metadata !DIExpression(DW_OP_LLVM_fragment, 16, 16), metadata !50, metadata ptr %a, metadata !DIExpression(DW_OP_plus_uconst, 2)), !dbg !51
+  ret void
+}
+
+declare void @llvm.dbg.value(metadata, metadata, metadata)
+declare void @llvm.dbg.assign(metadata, metadata, metadata, metadata, metadata, metadata)
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!8, !9, !10, !11, !12, !13, !14}
+!llvm.ident = !{!15}
+
+!0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = distinct !DIGlobalVariable(name: "G", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !3, producer: "clang version 17.0.0", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, globals: !4, splitDebugInlining: false, nameTableKind: None)
+!3 = !DIFile(filename: "test.cpp", directory: "/")
+!4 = !{!0, !5}
+!5 = !DIGlobalVariableExpression(var: !6, expr: !DIExpression())
+!6 = distinct !DIGlobalVariable(name: "F", scope: !2, file: !3, line: 1, type: !7, isLocal: false, isDefinition: true)
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{i32 7, !"Dwarf Version", i32 5}
+!9 = !{i32 2, !"Debug Info Version", i32 3}
+!10 = !{i32 1, !"wchar_size", i32 4}
+!11 = !{i32 8, !"PIC Level", i32 2}
+!12 = !{i32 7, !"PIE Level", i32 2}
+!13 = !{i32 7, !"uwtable", i32 2}
+!14 = !{i32 7, !"debug-info-assignment-tracking", i1 true}
+!15 = !{!"clang version 17.0.0"}
+!16 = distinct !DISubprogram(name: "fun", linkageName: "_Z3funv", scope: !3, file: !3, line: 3, type: !17, scopeLine: 3, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !19)
+!17 = !DISubroutineType(types: !18)
+!18 = !{null}
+!19 = !{!20}
+!20 = !DILocalVariable(name: "X", scope: !16, file: !3, line: 4, type: !21)
+!21 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Pair", file: !3, line: 2, size: 64, flags: DIFlagTypePassByValue, elements: !22, identifier: "_ZTS4Pair")
+!22 = !{}
+!25 = !DILocation(line: 0, scope: !16)
+!26 = !DILocation(line: 7, column: 7, scope: !27)
+!27 = distinct !DILexicalBlock(scope: !16, file: !3, line: 7, column: 7)
+!37 = distinct !DIAssignID()
+!38 = distinct !DIAssignID()
+!39 = distinct !DISubprogram(name: "fun2", linkageName: "_Z3funv2", scope: !3, file: !3, line: 3, type: !17, scopeLine: 3, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !40)
+!40 = !{!41}
+!41 = !DILocalVariable(name: "X", scope: !39, file: !3, line: 10, type: !21)
+!42 = distinct !DIAssignID()
+!43 = distinct !DIAssignID()
+!44 = !DILocation(line: 0, scope: !39)
+!45 = distinct !DIAssignID()
+!46 = distinct !DISubprogram(name: "fun3", linkageName: "_Z3funv3", scope: !3, file: !3, line: 3, type: !17, scopeLine: 3, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !2, retainedNodes: !47)
+!47 = !{!48}
+!48 = !DILocalVariable(name: "X", scope: !46, file: !3, line: 10, type: !21)
+!49 = distinct !DIAssignID()
+!50 = distinct !DIAssignID()
+!51 = !DILocation(line: 0, scope: !46)

llvm/test/DebugInfo/assignment-tracking/X86/lower-to-value.ll

 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=false \
+; RUN:    -debug-ata-coalesce-frags=true \
 ; RUN: | FileCheck %s --check-prefixes=CHECK,DBGVALUE --implicit-check-not=DBG_VALUE
 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=true \
 ; RUN: | FileCheck %s --check-prefixes=CHECK,INSTRREF --implicit-check-not=DBG_VALUE \
 ; RUN:    --implicit-check-not=DBG_INSTR_REF
 
 ;; Check that dbg.assigns for an aggregate variable which lives on the stack
 ;; for some of its lifetime are lowered into an appropriate set of DBG_VALUEs.

 ; CHECK-NEXT:     debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
 
 ;; Initially the whole variable is on the stack.
 ; CHECK: bb.0.entry:
 ; CHECK-NEXT: DBG_VALUE %stack.0.X, $noreg, ![[VAR]], !DIExpression(DW_OP_deref), debug-location
 
 ;; Then there is a store to the upper 64 bits.
 ; CHECK: MOV64mi32 %stack.0.X, 1, $noreg, 8, $noreg, 0, debug-location
-;; This DBG_VALUE is added by the mem-loc-frag-fill pass because bits [0, 64)
-;; are still live in memory.
-; CHECK-NEXT: DBG_VALUE %stack.0.X, $noreg, ![[VAR]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 64)
-; CHECK-NEXT: DBG_VALUE %stack.0.X, $noreg, ![[VAR]], !DIExpression(DW_OP_plus_uconst, 8, DW_OP_deref, DW_OP_LLVM_fragment, 64, 64), debug-location
+;; No change in variable location: the stack home is still valid.
 
 ;; The final assignment (X.B += 2) doesn't get stored back to the alloca. This
 ;; means that that the stack location isn't valid for the entire lifetime of X.
 ; DBGVALUE: %2:gr64 = nsw ADD64ri8 %1, 2, implicit-def dead $eflags, debug-location
 ; DBGVALUE-NEXT: DBG_VALUE %2, $noreg, ![[VAR]], !DIExpression(DW_OP_LLVM_fragment, 64, 64), debug-location
 ; INSTRREF: %2:gr64 = nsw ADD64ri8 %1, 2, implicit-def dead $eflags, debug-instr-number 1
 ; INSTRREF-NEXT: DBG_INSTR_REF ![[VAR]], !DIExpression(DW_OP_LLVM_arg, 0, DW_OP_LLVM_fragment, 64, 64), dbg-instr-ref(1, 0), debug-location
 

llvm/test/DebugInfo/assignment-tracking/X86/mem-loc-frag-fill-cfg.ll

 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=false \
+; RUN:    -debug-ata-coalesce-frags=true \
 ; RUN: | FileCheck %s --implicit-check-not=DBG_
 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=true \
 ; RUN: | FileCheck %s --implicit-check-not=DBG_
 
 ;; Check that the mem-loc-frag-fill pseudo-pass works on a simple CFG. When
 ;; LLVM sees a dbg.value with an overlapping fragment it essentially considers
 ;; the previous location as valid for all bits in that fragment. The pass
-;; inserts dbg.value fragments to preserve memory locations for bits in memory
-;; when overlapping fragments are encountered.
+;; tracks which bits are in memory and inserts dbg.values preserve memory
+;; locations for bits in memory when overlapping fragments are encountered.
 
 ;; nums lives in mem, except prior to the second call to step() where there has
 ;; been some DSE. At this point, the memory loc for nums.c is invalid.  But the
 ;; rest of num's bits, [0, 64), are in memory, so check there's a dbg.value for
 ;; them.
 
 ;; $ cat test.cpp
 ;; struct Nums { int a, b, c; };

 if.end:                                           ; preds = %if.else, %if.then
 ; CHECK: bb.3.if.end:
 ; CHECK-NEXT: DBG_VALUE 2, $noreg, ![[nums]], !DIExpression(DW_OP_LLVM_fragment, 64, 32), debug-location
 ; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 64)
   %c = getelementptr inbounds %struct.Nums, ptr %nums, i64 0, i32 2, !dbg !55
   call void @llvm.dbg.assign(metadata i32 2, metadata !44, metadata !DIExpression(DW_OP_LLVM_fragment, 64, 32), metadata !56, metadata ptr %c, metadata !DIExpression()), !dbg !46
   %call1 = tail call noundef zeroext i1 @_Z4stepv(), !dbg !57
   store i32 1, ptr %c, align 4, !dbg !58, !DIAssignID !61
+;; Store to bits 64 to 96 - the whole variable is in memory again.
 ; CHECK:      MOV32mi %stack.0.nums, 1, $noreg, 8, $noreg, 1
-; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_plus_uconst, 8, DW_OP_deref, DW_OP_LLVM_fragment, 64, 32)
+; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_deref)
   call void @llvm.dbg.assign(metadata i32 1, metadata !44, metadata !DIExpression(DW_OP_LLVM_fragment, 64, 32), metadata !61, metadata ptr %c, metadata !DIExpression()), !dbg !46
   call void @_Z4esc1P4Nums(ptr noundef nonnull %nums), !dbg !62
   ret i32 0, !dbg !64
 }
 
 declare void @llvm.lifetime.start.p0i8(i64 immarg, ptr nocapture) #4
 declare !dbg !65 dso_local noundef zeroext i1 @_Z4stepv() local_unnamed_addr #5
 declare void @llvm.lifetime.end.p0i8(i64 immarg, ptr nocapture) #4

llvm/test/DebugInfo/assignment-tracking/X86/mem-loc-frag-fill.ll

 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=false \
+; RUN:    -debug-ata-coalesce-frags=true \
 ; RUN: | FileCheck %s --implicit-check-not=DBG_
 ; RUN: llc %s -stop-before finalize-isel -o - \
 ; RUN:    -experimental-debug-variable-locations=true \
 ; RUN: | FileCheck %s  --implicit-check-not=DBG_
 
 ;; Check that the mem-loc-frag-fill analysis works on a simple case; ensure
 ;; that location definitions are added to preserve memory locations of
 ;; fragments of variables at subsequent location definitions for other
; CHECK: MOV64mr %stack.0.nums, 1, $noreg, 0, $noreg, killed %0
   call void @llvm.dbg.assign(metadata i1 undef, metadata !12, metadata !DIExpression(DW_OP_LLVM_fragment, 64, 32), metadata !22, metadata ptr undef, metadata !DIExpression()), !dbg !19
 ; CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[nums]], !DIExpression(DW_OP_LLVM_fragment, 64, 32)
 ; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 64)
   tail call void @_Z4stepv(), !dbg !23
   %c = getelementptr inbounds %struct.Nums, ptr %nums, i64 0, i32 2, !dbg !24
   store i32 2, ptr %c, align 8, !dbg !25, !DIAssignID !31
 ; CHECK: MOV32mi %stack.0.nums, 1, $noreg, 8, $noreg, 2
   call void @llvm.dbg.assign(metadata i32 2, metadata !12, metadata !DIExpression(DW_OP_LLVM_fragment, 64, 32), metadata !31, metadata ptr %c, metadata !DIExpression()), !dbg !19
-; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_plus_uconst, 8, DW_OP_deref, DW_OP_LLVM_fragment, 64, 32)
+; CHECK-NEXT: DBG_VALUE %stack.0.nums, $noreg, ![[nums]], !DIExpression(DW_OP_deref)
   tail call void @_Z4stepv(), !dbg !32
 ;; Next dbg.assign added by hand to test that the bits [64, 32) have been
 ;; correctly tracked as in memory - we know this has worked if
 ;; mem-loc-frag-fill reinstates the memory location after this kill-location
 ;; for bits [80, 96).
   call void @llvm.dbg.assign(metadata i32 poison, metadata !12, metadata !DIExpression(DW_OP_LLVM_fragment, 80, 16), metadata !44, metadata ptr poison, metadata !DIExpression()), !dbg !19
 ; CHECK: CALL64pcrel32 @_Z4stepv
 ; CHECK-NEXT: ADJCALLSTACKUP64

llvm/test/DebugInfo/assignment-tracking/X86/nested-loop-frags.ll

 ; CHECK-LABEL: bb.0.entry:
 ; CHECK-NEXT: successors
 ; CHECK-NEXT: {{^ *$}}
 ; CHECK-NEXT:    DBG_VALUE %stack.0.a.addr, $noreg, ![[a]], !DIExpression(DW_OP_deref)
 ; CHECK-NEXT:    DBG_VALUE %stack.1.b.addr, $noreg, ![[b]], !DIExpression(DW_OP_deref)
 ; CHECK-NEXT:    DBG_VALUE %stack.4.e.addr, $noreg, ![[e]], !DIExpression(DW_OP_deref)
 ; CHECK-NEXT:    MOV64mi32 %stack.0.a.addr, 1, $noreg, 0, $noreg, 1
 ; CHECK-NEXT:    MOV64mi32 %stack.1.b.addr, 1, $noreg, 0, $noreg, 2
-; CHECK-NEXT:    DBG_VALUE %stack.4.e.addr, $noreg, ![[e]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
-; CHECK-NEXT:    DBG_VALUE %stack.4.e.addr, $noreg, ![[e]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32)
 ; CHECK-NEXT:    MOV32mi %stack.0.a.addr, 1, $noreg, 0, $noreg, 3
 ; CHECK-NEXT:    DBG_VALUE $noreg, $noreg, ![[a]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
 ; CHECK-NEXT:    DBG_VALUE %stack.0.a.addr, $noreg, ![[a]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
 ; CHECK-NEXT:    MOV32mi %stack.1.b.addr, 1, $noreg, 0, $noreg, 4
 ; CHECK-NEXT:    DBG_VALUE $noreg, $noreg, ![[b]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
 ; CHECK-NEXT:    DBG_VALUE %stack.1.b.addr, $noreg, ![[b]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
 ; CHECK-NEXT:    MOV64mi32 %stack.2.c.addr, 1, $noreg, 0, $noreg, 5
 ; CHECK-NEXT:    MOV64mi32 %stack.3.d.addr, 1, $noreg, 0, $noreg, 6

 do.body1:                                         ; preds = %do.cond, %do.body
   %0 = phi i32 [ %.pre10, %do.body ], [ %1, %do.cond ], !dbg !27
   ;call void @llvm.dbg.assign(metadata i32 11, metadata !79, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !81, metadata ptr %f.addr, metadata !DIExpression()), !dbg !27 ; VAR:f
   %tobool.not = icmp eq i32 %0, 0, !dbg !27
   br i1 %tobool.not, label %if.else, label %if.then, !dbg !35
 ; CHECK-LABEL: bb.2.do.body1:
 ; CHECK-NEXT: successors
 ; CHECK-NEXT: {{^ *$}}
-; CxHECK-NEXT:    DBG_VALUE 11, $noreg, ![[f]], !DIExpression(DW_OP_LLVM_fragment, 0, 32)
 ; CHECK:         JMP_1
 ; CHECK-NEXT: {{^ *$}}
 
 if.then:                                          ; preds = %do.body1
   %.pre = load i32, ptr @g_a, align 4, !dbg !27
   store i32 %.pre, ptr %b.addr, !DIAssignID !66 ; VAR:b
   call void @llvm.dbg.assign(metadata i32 %.pre, metadata !22, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !66, metadata ptr %b.addr, metadata !DIExpression()), !dbg !27 ; VAR:b
   store i32 6, ptr %d.addr, !DIAssignID !74 ; VAR:d
   call void @llvm.dbg.assign(metadata i32 6, metadata !72, metadata !DIExpression(DW_OP_LLVM_fragment, 32, 32), metadata !74, metadata ptr %d.addr, metadata !DIExpression()), !dbg !27 ; VAR:d
   call void @llvm.dbg.assign(metadata i32 8, metadata !75, metadata !DIExpression(DW_OP_LLVM_fragment, 32, 32), metadata !82, metadata ptr %e.addr, metadata !DIExpression()), !dbg !27 ; VAR:e
   store i32 8, ptr %e.addr, !DIAssignID !82 ; VAR:e
   br label %do.cond, !dbg !39
 ; CHECK-LABEL: bb.3.if.then:
 ; CHECK-NEXT: successors
 ; CHECK-NEXT: {{^ *$}}
 ; CHECK-NEXT:    %5:gr32 = MOV32rm $rip, 1, $noreg, @g_a, $noreg
 ; CHECK-NEXT:    MOV32mr %stack.1.b.addr, 1, $noreg, 0, $noreg, killed %5
-; CHECK-NEXT:    DBG_VALUE %stack.1.b.addr, $noreg, ![[b]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32)
+; CHECK-NEXT:    DBG_VALUE %stack.1.b.addr, $noreg, ![[b]], !DIExpression(DW_OP_deref)
 ; CHECK-NEXT:    MOV32mi %stack.3.d.addr, 1, $noreg, 0, $noreg, 6
 ; CHECK-NEXT:    DBG_VALUE 8, $noreg, ![[e]], !DIExpression(DW_OP_LLVM_fragment, 32, 32)
 ; CHECK-NEXT:    MOV32mi %stack.4.e.addr, 1, $noreg, 0, $noreg, 8
 ; CHECK-NEXT:    DBG_VALUE %stack.4.e.addr, $noreg, ![[e]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
 ; CHECK-NEXT:    JMP_1 %bb.5
-; CHECxK-NEXT: {{^ *$}}
 
 if.else:                                          ; preds = %do.body1
   store i32 6, ptr %d.addr, !DIAssignID !74 ; VAR:d
   call void @llvm.dbg.assign(metadata i32 6, metadata !72, metadata !DIExpression(DW_OP_LLVM_fragment, 0, 32), metadata !74, metadata ptr %d.addr, metadata !DIExpression()), !dbg !27 ; VAR:d
   %e.high32 = getelementptr i32, ptr %e.addr, i32 1
   store i32 15, ptr %e.high32 ; VAR:e
   ;store i32 10, ptr %f.addr ; VAR:f
   br label %do.cond

llvm/test/DebugInfo/assignment-tracking/X86/use-known-value-at-early-mem-def-2.ll

 ;; No DBG_VALUE required because the stack location is still valid.
 
 ; CHECK: MOV32mi %stack.0.c, 1, $noreg, 0, $noreg, 1
 ; CHECK-NEXT: DBG_VALUE $noreg, $noreg, ![[var]], !DIExpression(DW_OP_LLVM_fragment, 0, 32), debug-location
 ;; This DBG_VALUE is added by the frag-agg pass because bits [32, 64) are still
 ;; live in memory.
 ; CHECK-NEXT: DBG_VALUE %stack.0.c, $noreg, ![[var]], !DIExpression(DW_OP_plus_uconst, 4, DW_OP_deref, DW_OP_LLVM_fragment, 32, 32)
 
+;; After the call to @d there's a dbg.assign linked to the store to bits 0-32
+;; that comes before it. Meaning the stack location for bits 0-32 are valid
+;; from here (bits 32-64 for the variable are already located in memory).
 ; CHECK: CALL64pcrel32 @d
 ; CHECK-NEXT: ADJCALLSTACKUP64
-; CHECK-NEXT: DBG_VALUE %stack.0.c, $noreg, ![[var]], !DIExpression(DW_OP_deref, DW_OP_LLVM_fragment, 0, 32), debug-location
+; CHECK-NEXT: DBG_VALUE %stack.0.c, $noreg, ![[var]], !DIExpression(DW_OP_deref), debug-location
 
 target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 ; Function Attrs: nounwind uwtable
 define dso_local void @b() local_unnamed_addr #0 !dbg !7 {
 entry:
   %c = alloca i64, align 1, !DIAssignID !13