[DebugInfo] Sort the SDDbgValue list before assuming it is in IR order

[rnk]

Summary

This code iterates the 'Orders' vector in parallel with the DbgValue
list, emitting all DBG_VALUEs that occurred between the last IR order
insertion point and the next insertion point. This assumes the
SDDbgValue list is sorted in IR order, which it usually is. However, it
is not sorted when a node with a debug value is replaced with another
one. When this happens, TransferDbgValues is called, and the new value
is added to the end of the list.

The problem can be solved by stably sorting the list by IR order.

[rnk]

• Don't overwrite existing test case

[Ka-Ka]

llvm/test/DebugInfo/X86/dbg-value-transfer-order.ll

27–39 ↗

(On Diff #116418)

I had problem recreating the fault as this testcase passed both with or without your changes in ScheduleDAGSDNodes.cpp, but then I realized that the checks in these lines don't check anything as they are missing a ':' after the CHECK.

[aprantl]

Thanks! It might be good to separate the refactoring into an NFC commit.

llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp

755

Orders.push_back({Order, &*std::prev(IP)});

860

Add a comment why the stable_sort is necessary?

[rnk]

Thanks!

llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp

860

I don't have a testcase to demonstrate any issue, but it will ensure that clang has the same output on all platforms regardless of the (usually deterministic) instability in sort order. Anyways, commented.

llvm/test/DebugInfo/X86/dbg-value-transfer-order.ll

27–39 ↗

(On Diff #116418)

Oops. I added the colons and confirmed that the test fails without the sort. Without the fix, the DBG_VALUE for bit_offset appears at the end of .LBB0_3.

llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp

ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
   if (!Order || !Seen.insert(Order).second) {
     // Process any valid SDDbgValues even if node does not have any order
     // assigned.
     ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
     return;
   }
 
   MachineBasicBlock *BB = Emitter.getBlock();
-  if (Emitter.getInsertPos() == BB->begin() || BB->back().isPHI() ||
+  auto IP = Emitter.getInsertPos();
+  if (IP == BB->begin() || BB->back().isPHI() ||
       // Fast-isel may have inserted some instructions, in which case the
       // BB->back().isPHI() test will not fire when we want it to.
-      std::prev(Emitter.getInsertPos())->isPHI()) {
+      std::prev(IP)->isPHI()) {
     // Did not insert any instruction.
     Orders.push_back(std::make_pair(Order, (MachineInstr*)nullptr));
     return;
   }
 
-  Orders.push_back(std::make_pair(Order, &*std::prev(Emitter.getInsertPos())));
+  Orders.push_back(std::make_pair(Order, &*std::prev(IP)));
   ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
 }
 
 void ScheduleDAGSDNodes::
 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap,
                 MachineBasicBlock::iterator InsertPos) {
   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
        I != E; ++I) {
EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
   }
 
   // Insert all the dbg_values which have not already been inserted in source
   // order sequence.
   if (HasDbg) {
     MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
 
     // Sort the source order instructions and use the order to insert debug
     // values.
-    std::sort(Orders.begin(), Orders.end(), less_first());
+    std::stable_sort(Orders.begin(), Orders.end(), less_first());
+    std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
+                     [](const SDDbgValue *LHS, const SDDbgValue *RHS) {
+                       return LHS->getOrder() < RHS->getOrder();
+                     });
 
     SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
     SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
     // Now emit the rest according to source order.
     unsigned LastOrder = 0;
     for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
       unsigned Order = Orders[i].first;
       MachineInstr *MI = Orders[i].second;
       // Insert all SDDbgValue's whose order(s) are before "Order".
       if (!MI)
         continue;
-      for (; DI != DE &&
-             (*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
+      for (; DI != DE; ++DI) {
+        if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
+          break;
         if ((*DI)->isInvalidated())
           continue;
+
         MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
         if (DbgMI) {
           if (!LastOrder)
             // Insert to start of the BB (after PHIs).
             BB->insert(BBBegin, DbgMI);
           else {
             // Insert at the instruction, which may be in a different
             // block, if the block was split by a custom inserter.
             MachineBasicBlock::iterator Pos = MI;
             MI->getParent()->insert(Pos, DbgMI);
           }
         }
       }
       LastOrder = Order;
     }
     // Add trailing DbgValue's before the terminator. FIXME: May want to add
     // some of them before one or more conditional branches?
     SmallVector<MachineInstr*, 8> DbgMIs;
-    while (DI != DE) {
-      if (!(*DI)->isInvalidated())
-        if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
-          DbgMIs.push_back(DbgMI);
-      ++DI;
+    for (; DI != DE; ++DI) {
+      if ((*DI)->isInvalidated())
+        continue;
+      assert((*DI)->getOrder() >= LastOrder &&
+             "emitting DBG_VALUE out of order");
     }
 
     MachineBasicBlock *InsertBB = Emitter.getBlock();
     MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
     InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
   }
 
   InsertPos = Emitter.getInsertPos();
   return Emitter.getBlock();
 }
 
 /// Return the basic block label.
 std::string ScheduleDAGSDNodes::getDAGName() const {
   return "sunit-dag." + BB->getFullName();
 }

llvm/test/CodeGen/X86/frame-lowering-debug-intrinsic-2.ll

entry:
   tail call void @llvm.dbg.value(metadata i64 %4, i64 0, metadata !23, metadata !33), !dbg !34
   tail call void @llvm.dbg.value(metadata i64 %5, i64 0, metadata !22, metadata !33), !dbg !35
   tail call void (...) @printf()
   tail call void (...) @printf(i64 1, i64 2, i64 3, i64 4, i32 0, i64 0, i64 %4, i64 %5)
   ret void
 }
 
 ; CHECK-LABEL: withDebug
-; CHECK:       #DEBUG_VALUE: test:j <- %RBX
+; CHECK:       callq printf
+; CHECK:       callq printf
 ; CHECK-NEXT:  addq  $24, %rsp
 ; CHECK:       popq  %rbx
 ; CHECK-NEXT:  popq  %r14
 ; CHECK-NEXT:  retq
 
 declare { i64, i1 } @llvm.uadd.with.overflow.i64(i64, i64)
 declare i64 @fn1(i64, i64)
 

llvm/test/DebugInfo/X86/dbg-value-range.ll

-; RUN: llc -mtriple=x86_64-apple-darwin10 < %s | FileCheck %s
+; RUN: llc < %s | FileCheck %s
 
-%struct.a = type { i32 }
+; Reduced manually from C source:
+; unsigned __attribute__((noinline, optnone)) useslot(unsigned v) { return v; }
+; void __attribute__((noinline)) f(unsigned cell_offset, unsigned *page_start_) {
+;   unsigned cell = page_start_[0];
+;   while (cell) {
+;     unsigned bit_offset = cell_offset ? __builtin_ctz(cell) : 32;
+;     unsigned bit_mask = 1U << bit_offset;
+;     unsigned slot = (cell_offset + bit_offset) << /*log2(sizeof(void*))*/ 3;
+;     cell ^= bit_mask + page_start_[slot];
+;   }
+; }
+; int main() {
+;   static unsigned pages[32] = {1, 2, 3, 4, 5, 6, 7};
+;   f(3, &pages[0]);
+; }
+
+; We had a bug where the DBG_VALUE instruction for bit_offset would be emitted
+; at the end of the basic block, long after its actual program point. What's
+; interesting in this example is that the !range metadata produces an AssertSext
+; DAG node that gets replaced during ISel. This leads to an unordered SDDbgValue
+; vector, which has to be sorted by IR order before it is processed in parallel
+; with the Orders insertion point vector.
+
+; CHECK-LABEL: f: # @f
+; CHECK .LBB0_1:                                # %while.body
+; CHECK         movl    $32, %ecx
+; CHECK         testl   {{.*}}
+; CHECK         jne     .LBB0_3
+; CHECK # BB#2:                                 # %if.then
+; CHECK         callq   if_then
+; CHECK         movl    %eax, %ecx
+; CHECK .LBB0_3:                                # %if.end
+;        Check that this DEBUG_VALUE comes before the left shift.
+; CHECK         #DEBUG_VALUE: bit_offset <- %ECX
+; CHECK         .cv_loc 0 1 8 28                # t.c:8:28
+; CHECK         movl    $1, %[[reg:[^ ]*]]
+; CHECK         shll    %cl, %[[reg]]
+
+; ModuleID = 't.c'
+source_filename = "t.c"
+target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-windows-msvc19.0.24215"
 
-define i32 @bar(%struct.a* nocapture %b) nounwind ssp !dbg !0 {
+; Function Attrs: nounwind readnone speculatable
+declare void @llvm.dbg.declare(metadata, metadata, metadata) #1
+
+declare i32 @if_then()
+
+; Function Attrs: noinline nounwind readonly uwtable
+define void @f(i32 %cell_offset, i32* nocapture %page_start_) local_unnamed_addr #0 !dbg !31 {
 entry:
-  tail call void @llvm.dbg.value(metadata %struct.a* %b, metadata !6, metadata !DIExpression()), !dbg !13
-  %tmp1 = getelementptr inbounds %struct.a, %struct.a* %b, i64 0, i32 0, !dbg !14
-  %tmp2 = load i32, i32* %tmp1, align 4, !dbg !14
-  tail call void @llvm.dbg.value(metadata i32 %tmp2, metadata !11, metadata !DIExpression()), !dbg !14
-  %call = tail call i32 (...) @foo(i32 %tmp2) nounwind , !dbg !18
-  %add = add nsw i32 %tmp2, 1, !dbg !19
-  ret i32 %add, !dbg !19
+  %0 = load i32, i32* %page_start_, align 4, !dbg !45
+  %tobool14 = icmp eq i32 %0, 0, !dbg !47
+  br i1 %tobool14, label %while.end, label %while.body, !dbg !47
+
+while.body:                                       ; preds = %while.body.lr.ph, %while.body
+  %cell.015 = phi i32 [ %0, %entry ], [ %xor, %if.end ]
+  %tobool1 = icmp eq i32 %cell_offset, 0
+  br i1 %tobool1, label %if.then, label %if.end
+
+if.then:
+  %1 = call i32 @if_then(), !dbg !48, !range !49
+  br label %if.end
+
+if.end:
+  %cond = phi i32 [ %1, %if.then ], [ 32, %while.body ]
+  tail call void @llvm.dbg.value(metadata i32 %cond, metadata !39, metadata !DIExpression()), !dbg !51
+  %shl = shl i32 1, %cond, !dbg !52
+  %add = add i32 %cond, %cell_offset, !dbg !54
+  %shl2 = shl i32 %add, 3, !dbg !55
+  %idxprom = zext i32 %shl2 to i64, !dbg !57
+  %arrayidx3 = getelementptr inbounds i32, i32* %page_start_, i64 %idxprom, !dbg !57
+  %2 = load i32, i32* %arrayidx3, align 4, !dbg !57
+  %add4 = add i32 %2, %shl, !dbg !58
+  %xor = xor i32 %add4, %cell.015, !dbg !59
+  tail call void @llvm.dbg.value(metadata i32 %xor, metadata !38, metadata !DIExpression()), !dbg !46
+  %tobool = icmp eq i32 %xor, 0, !dbg !47
+  br i1 %tobool, label %while.end, label %while.body, !dbg !47, !llvm.loop !60
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void, !dbg !62
 }
 
-declare i32 @foo(...) 
+; Function Attrs: nounwind readnone speculatable
+declare i32 @llvm.cttz.i32(i32, i1) #1
 
-declare void @llvm.dbg.value(metadata, metadata, metadata) nounwind readnone
+; Function Attrs: nounwind readnone speculatable
+declare void @llvm.dbg.value(metadata, metadata, metadata) #1
 
-!llvm.dbg.cu = !{!2}
-!llvm.module.flags = !{!24}
+attributes #0 = { noinline nounwind readonly uwtable }
+attributes #1 = { nounwind readnone speculatable }
 
-!0 = distinct !DISubprogram(name: "bar", line: 5, isLocal: false, isDefinition: true, virtualIndex: 6, flags: DIFlagPrototyped, isOptimized: true, unit: !2, file: !22, scope: !1, type: !3, variables: !21)
-!1 = !DIFile(filename: "bar.c", directory: "/private/tmp")
-!2 = distinct !DICompileUnit(language: DW_LANG_C99, producer: "clang version 2.9 (trunk 122997)", isOptimized: true, emissionKind: FullDebug, file: !22, enums: !23, retainedTypes: !23, imports:  null)
-!3 = !DISubroutineType(types: !4)
-!4 = !{!5}
-!5 = !DIBasicType(tag: DW_TAG_base_type, name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
-!6 = !DILocalVariable(name: "b", line: 5, arg: 1, scope: !0, file: !1, type: !7)
-!7 = !DIDerivedType(tag: DW_TAG_pointer_type, size: 64, align: 64, scope: !2, baseType: !8)
-!8 = !DICompositeType(tag: DW_TAG_structure_type, name: "a", line: 1, size: 32, align: 32, file: !22, scope: !2, elements: !9)
-!9 = !{!10}
-!10 = !DIDerivedType(tag: DW_TAG_member, name: "c", line: 2, size: 32, align: 32, file: !22, scope: !1, baseType: !5)
-!11 = !DILocalVariable(name: "x", line: 6, scope: !12, file: !1, type: !5)
-!12 = distinct !DILexicalBlock(line: 5, column: 22, file: !22, scope: !0)
-!13 = !DILocation(line: 5, column: 19, scope: !0)
-!14 = !DILocation(line: 6, column: 14, scope: !12)
-!18 = !DILocation(line: 7, column: 2, scope: !12)
-!19 = !DILocation(line: 8, column: 2, scope: !12)
-!21 = !{!6, !11}
-!22 = !DIFile(filename: "bar.c", directory: "/private/tmp")
-!23 = !{}
-
-; Check that variable bar:b value range is appropriately truncated in debug info.
-; The variable is in %rdi which is clobbered by 'movl %ebx, %edi'
-; Here Ltmp7 is the end of the location range.
-
-;CHECK: .loc  1 7 2
-;CHECK: movl
-;CHECK-NEXT: [[CLOBBER:Ltmp[0-9]*]]
-
-;CHECK:Ldebug_loc0:
-;CHECK-NEXT: Lset{{.*}} =
-;CHECK-NEXT:  .quad
-;CHECK-NEXT: [[CLOBBER_OFF:Lset.*]] = [[CLOBBER]]-{{.*}}
-;CHECK-NEXT:  .quad [[CLOBBER_OFF]]
-;CHECK-NEXT:  .short 1 ## Loc expr size
-;CHECK-NEXT:  .byte 85 ## DW_OP_reg
-;CHECK-NEXT:  .quad 0
-;CHECK-NEXT:  .quad 0
-!24 = !{i32 1, !"Debug Info Version", i32 3}
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!14, !15, !16, !17}
+!llvm.ident = !{!18}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 6.0.0 ", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, globals: !3)
+!1 = !DIFile(filename: "t.c", directory: "C:\5Csrc\5Cllvm-project\5Cbuild", checksumkind: CSK_MD5, checksum: "f80d0003faf76554dfdec6c95da285cc")
+!2 = !{}
+!3 = !{!4}
+!4 = !DIGlobalVariableExpression(var: !5, expr: !DIExpression())
+!5 = distinct !DIGlobalVariable(name: "pages", scope: !6, file: !1, line: 18, type: !10, isLocal: true, isDefinition: true)
+!6 = distinct !DISubprogram(name: "main", scope: !1, file: !1, line: 17, type: !7, isLocal: false, isDefinition: true, scopeLine: 17, isOptimized: true, unit: !0, variables: !2)
+!7 = !DISubroutineType(types: !8)
+!8 = !{!9}
+!9 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!10 = !DICompositeType(tag: DW_TAG_array_type, baseType: !11, size: 1024, elements: !12)
+!11 = !DIBasicType(name: "unsigned int", size: 32, encoding: DW_ATE_unsigned)
+!12 = !{!13}
+!13 = !DISubrange(count: 32)
+!14 = !{i32 2, !"CodeView", i32 1}
+!15 = !{i32 2, !"Debug Info Version", i32 3}
+!16 = !{i32 1, !"wchar_size", i32 2}
+!17 = !{i32 7, !"PIC Level", i32 2}
+!18 = !{!"clang version 6.0.0 "}
+!31 = distinct !DISubprogram(name: "f", scope: !1, file: !1, line: 4, type: !32, isLocal: false, isDefinition: true, scopeLine: 4, flags: DIFlagPrototyped, isOptimized: true, unit: !0, variables: !35)
+!32 = !DISubroutineType(types: !33)
+!33 = !{null, !11, !34}
+!34 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !11, size: 64)
+!35 = !{!36, !37, !38, !39, !41, !42}
+!36 = !DILocalVariable(name: "page_start_", arg: 2, scope: !31, file: !1, line: 4, type: !34)
+!37 = !DILocalVariable(name: "cell_offset", arg: 1, scope: !31, file: !1, line: 4, type: !11)
+!38 = !DILocalVariable(name: "cell", scope: !31, file: !1, line: 5, type: !11)
+!39 = !DILocalVariable(name: "bit_offset", scope: !40, file: !1, line: 7, type: !11)
+!40 = distinct !DILexicalBlock(scope: !31, file: !1, line: 6, column: 16)
+!41 = !DILocalVariable(name: "bit_mask", scope: !40, file: !1, line: 8, type: !11)
+!42 = !DILocalVariable(name: "slot", scope: !40, file: !1, line: 9, type: !11)
+!43 = !DILocation(line: 4, column: 66, scope: !31)
+!44 = !DILocation(line: 4, column: 43, scope: !31)
+!45 = !DILocation(line: 5, column: 19, scope: !31)
+!46 = !DILocation(line: 5, column: 12, scope: !31)
+!47 = !DILocation(line: 6, column: 3, scope: !31)
+!48 = !DILocation(line: 7, column: 41, scope: !40)
+!49 = !{i32 0, i32 33}
+!50 = !DILocation(line: 7, column: 27, scope: !40)
+!51 = !DILocation(line: 7, column: 14, scope: !40)
+!52 = !DILocation(line: 8, column: 28, scope: !40)
+!53 = !DILocation(line: 8, column: 14, scope: !40)
+!54 = !DILocation(line: 9, column: 34, scope: !40)
+!55 = !DILocation(line: 9, column: 48, scope: !40)
+!56 = !DILocation(line: 9, column: 14, scope: !40)
+!57 = !DILocation(line: 13, column: 24, scope: !40)
+!58 = !DILocation(line: 13, column: 22, scope: !40)
+!59 = !DILocation(line: 13, column: 10, scope: !40)
+!60 = distinct !{!60, !47, !61}
+!61 = !DILocation(line: 14, column: 3, scope: !31)
+!62 = !DILocation(line: 15, column: 1, scope: !31)