[DebugInfo][InstrRef][X86] Instrument expanded DYN_ALLOCA_ instructions with instruction numbers

[jmorse]

Summary

If we have a DYN_ALLOCA_* instruction, it will eventually be expanded to a stack probe and subtract-from-SP. Add debug-info instrumentation to X86FrameLowering::emitStackProbe so that it can redirect debug-info for the DYN_ALLOCA to the lowered stack probe.

In practice, this means putting an instruction number label either the call instruction to _chkstk for win32, or more commonly on the subtract from SP instruction. The two tests added cover both of these cases.

NB:

• I don't think there are any paths with the Large code-model and 64-bit Windows that will use the NumCallOps number, because it's only win32 that uses _chkstck to modify SP,
• There are other code paths from emitStackProbe that implement probes, such as emitStackProbeInline, but I believe they all deal with fixed-size allocations in the prologue, and so aren't of interest to lowered DYN_ALLOCA instructions

(Adding Craig as this touches X86 things)

[Orlando]

Makes sense to me from a debug info perspective, and tests LGTM. I've added a couple of nits inline. It might be worth someone else having a quick look as I'm not familiar with the .cpps touched here.

llvm/lib/Target/X86/X86DynAllocaExpander.cpp

217

super nit: I think the style guide prefers braces around a single statement if if there is also a nested comment?

llvm/lib/Target/X86/X86FrameLowering.cpp

1050

nits: Do you need NumCallOps? Is it not the same as ModInst->getNumOperands() here?

And since ModInst == CI on this code path (right?) IMO it would be clearer to refer to CI directly here.

[jmorse]

Avoid counting number of operands when we can just query the instruction

[jmorse]

llvm/lib/Target/X86/X86FrameLowering.cpp

1050

Updated to use getNumOperands. Also, apparently I couldn't count in the previous version, because now we need to subtract two (one for it being penultimate op, one for it being zero-based).

IMHO it's better to use ModInst -- it avoids having to think about the earlier control flow. YMMV?

llvm/lib/Target/X86/X86DynAllocaExpander.cpp

void X86DynAllocaExpander::lower(MachineInstr *MI, Lowering L) {
   }
 
   // These two variables differ on x32, which is a 64-bit target with a
   // 32-bit alloca.
   bool Is64Bit = STI->is64Bit();
   bool Is64BitAlloca = MI->getOpcode() == X86::DYN_ALLOCA_64;
   assert(SlotSize == 4 || SlotSize == 8);
 
+  Optional<MachineFunction::DebugInstrOperandPair> InstrNum = None;
+  if (unsigned Num = MI->peekDebugInstrNum())
+    // Operand 2 of DYN_ALLOCAs contains the stack def.
+    InstrNum = {Num, 2};
+
   switch (L) {
   case TouchAndSub: {
     assert(Amount >= SlotSize);
 
     // Use a push to touch the top of the stack.
     unsigned RegA = Is64Bit ? X86::RAX : X86::EAX;
     BuildMI(*MBB, I, DL, TII->get(Is64Bit ? X86::PUSH64r : X86::PUSH32r))
         .addReg(RegA, RegState::Undef);
case Probe:
     if (!NoStackArgProbe) {
       // The probe lowering expects the amount in RAX/EAX.
       unsigned RegA = Is64BitAlloca ? X86::RAX : X86::EAX;
       BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::COPY), RegA)
           .addReg(MI->getOperand(0).getReg());
 
       // Do the probe.
       STI->getFrameLowering()->emitStackProbe(*MBB->getParent(), *MBB, MI, DL,
-                                              /*InProlog=*/false);
+                                              /*InProlog=*/false, InstrNum);
     } else {
       // Sub
       BuildMI(*MBB, I, DL,
               TII->get(Is64BitAlloca ? X86::SUB64rr : X86::SUB32rr), StackPtr)
           .addReg(StackPtr)
           .addReg(MI->getOperand(0).getReg());
     }
     break;

llvm/lib/Target/X86/X86FrameLowering.h

 //===-- X86TargetFrameLowering.h - Define frame lowering for X86 -*- C++ -*-==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This class implements X86-specific bits of TargetFrameLowering class.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_X86_X86FRAMELOWERING_H
 #define LLVM_LIB_TARGET_X86_X86FRAMELOWERING_H
 
+#include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/TargetFrameLowering.h"
 #include "llvm/Support/TypeSize.h"
 
 namespace llvm {
 
 class MachineInstrBuilder;
 class MCCFIInstruction;
 class X86InstrInfo;
public:
   /// instruction operands should be used to manipulate StackPtr and FramePtr.
   bool Uses64BitFramePtr;
 
   unsigned StackPtr;
 
   /// Emit target stack probe code. This is required for all
   /// large stack allocations on Windows. The caller is required to materialize
   /// the number of bytes to probe in RAX/EAX.
-  void emitStackProbe(MachineFunction &MF, MachineBasicBlock &MBB,
-                      MachineBasicBlock::iterator MBBI, const DebugLoc &DL,
-                      bool InProlog) const;
+  /// \p InstrNum optionally contains a debug-info instruction number for the
+  ///    new stack pointer.
+  void emitStackProbe(
+      MachineFunction &MF, MachineBasicBlock &MBB,
+      MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog,
+      Optional<MachineFunction::DebugInstrOperandPair> InstrNum = None) const;
 
   bool stackProbeFunctionModifiesSP() const override;
 
   /// Replace a StackProbe inline-stub with the actual probe code inline.
   void inlineStackProbe(MachineFunction &MF,
                         MachineBasicBlock &PrologMBB) const override;
 
   void

 private:
   bool isWin64Prologue(const MachineFunction &MF) const;
 
   bool needsDwarfCFI(const MachineFunction &MF) const;
 
   uint64_t calculateMaxStackAlign(const MachineFunction &MF) const;
 
   /// Emit target stack probe as a call to a helper function
-  void emitStackProbeCall(MachineFunction &MF, MachineBasicBlock &MBB,
-                          MachineBasicBlock::iterator MBBI, const DebugLoc &DL,
-                          bool InProlog) const;
+  void emitStackProbeCall(
+      MachineFunction &MF, MachineBasicBlock &MBB,
+      MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog,
+      Optional<MachineFunction::DebugInstrOperandPair> InstrNum) const;
 
   /// Emit target stack probe as an inline sequence.
   void emitStackProbeInline(MachineFunction &MF, MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator MBBI,
                             const DebugLoc &DL, bool InProlog) const;
   void emitStackProbeInlineWindowsCoreCLR64(MachineFunction &MF,
                                             MachineBasicBlock &MBB,
                                             MachineBasicBlock::iterator MBBI,

llvm/lib/Target/X86/X86FrameLowering.cpp

if (IsPrologue) {
                MCCFIInstruction::createOffset(nullptr, DwarfReg, Offset));
     } else {
       BuildCFI(MBB, MBBI, DL,
                MCCFIInstruction::createRestore(nullptr, DwarfReg));
     }
   }
 }
 
-void X86FrameLowering::emitStackProbe(MachineFunction &MF,
-                                      MachineBasicBlock &MBB,
-                                      MachineBasicBlock::iterator MBBI,
-                                      const DebugLoc &DL, bool InProlog) const {
+void X86FrameLowering::emitStackProbe(
+    MachineFunction &MF, MachineBasicBlock &MBB,
+    MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog,
+    Optional<MachineFunction::DebugInstrOperandPair> InstrNum) const {
   const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
   if (STI.isTargetWindowsCoreCLR()) {
     if (InProlog) {
       BuildMI(MBB, MBBI, DL, TII.get(X86::STACKALLOC_W_PROBING))
           .addImm(0 /* no explicit stack size */);
     } else {
       emitStackProbeInline(MF, MBB, MBBI, DL, false);
     }
   } else {
-    emitStackProbeCall(MF, MBB, MBBI, DL, InProlog);
+    emitStackProbeCall(MF, MBB, MBBI, DL, InProlog, InstrNum);
   }
 }
 
 bool X86FrameLowering::stackProbeFunctionModifiesSP() const {
   return STI.isOSWindows() && !STI.isTargetWin64();
 }
 
 void X86FrameLowering::inlineStackProbe(MachineFunction &MF,
if (InProlog) {
     }
     for (MachineBasicBlock::iterator CMBBI = ContinueMBB->begin();
          CMBBI != ContinueMBBI; ++CMBBI) {
       CMBBI->setFlag(MachineInstr::FrameSetup);
     }
   }
 }
 
-void X86FrameLowering::emitStackProbeCall(MachineFunction &MF,
-                                          MachineBasicBlock &MBB,
-                                          MachineBasicBlock::iterator MBBI,
-                                          const DebugLoc &DL,
-                                          bool InProlog) const {
+void X86FrameLowering::emitStackProbeCall(
+    MachineFunction &MF, MachineBasicBlock &MBB,
+    MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog,
+    Optional<MachineFunction::DebugInstrOperandPair> InstrNum) const {
   bool IsLargeCodeModel = MF.getTarget().getCodeModel() == CodeModel::Large;
 
   // FIXME: Add indirect thunk support and remove this.
   if (Is64Bit && IsLargeCodeModel && STI.useIndirectThunkCalls())
     report_fatal_error("Emitting stack probe calls on 64-bit with the large "
                        "code model and indirect thunks not yet implemented.");
 
   unsigned CallOp;
   if (Is64Bit)
     CallOp = IsLargeCodeModel ? X86::CALL64r : X86::CALL64pcrel32;
   else
     CallOp = X86::CALLpcrel32;
 
   StringRef Symbol = STI.getTargetLowering()->getStackProbeSymbolName(MF);
 
   MachineInstrBuilder CI;
   MachineBasicBlock::iterator ExpansionMBBI = std::prev(MBBI);
 
   // All current stack probes take AX and SP as input, clobber flags, and
   // preserve all registers. x86_64 probes leave RSP unmodified.
+  unsigned NumCallOps = 0;
   if (Is64Bit && MF.getTarget().getCodeModel() == CodeModel::Large) {
     // For the large code model, we have to call through a register. Use R11,
     // as it is scratch in all supported calling conventions.
     BuildMI(MBB, MBBI, DL, TII.get(X86::MOV64ri), X86::R11)
         .addExternalSymbol(MF.createExternalSymbolName(Symbol));
     CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp)).addReg(X86::R11);
+    NumCallOps = 2;
   } else {
     CI = BuildMI(MBB, MBBI, DL, TII.get(CallOp))
         .addExternalSymbol(MF.createExternalSymbolName(Symbol));
+    NumCallOps = 1;
   }
 
   unsigned AX = Uses64BitFramePtr ? X86::RAX : X86::EAX;
   unsigned SP = Uses64BitFramePtr ? X86::RSP : X86::ESP;
   CI.addReg(AX, RegState::Implicit)
       .addReg(SP, RegState::Implicit)
       .addReg(AX, RegState::Define | RegState::Implicit)
       .addReg(SP, RegState::Define | RegState::Implicit)
       .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
+  NumCallOps += 6;
 
+  MachineInstr *ModInst = CI;
   if (STI.isTargetWin64() || !STI.isOSWindows()) {
     // MSVC x32's _chkstk and cygwin/mingw's _alloca adjust %esp themselves.
     // MSVC x64's __chkstk and cygwin/mingw's ___chkstk_ms do not adjust %rsp
     // themselves. They also does not clobber %rax so we can reuse it when
     // adjusting %rsp.
     // All other platforms do not specify a particular ABI for the stack probe
     // function, so we arbitrarily define it to not adjust %esp/%rsp itself.
+    ModInst =
         BuildMI(MBB, MBBI, DL, TII.get(getSUBrrOpcode(Uses64BitFramePtr)), SP)
             .addReg(SP)
             .addReg(AX);
   }
 
+  // DebugInfo variable locations -- if there's an instruction number for the
+  // allocation (i.e., DYN_ALLOC_*), substitute it for the instruction that
+  // modifies SP.
+  if (InstrNum) {
+    if (STI.isTargetWin64() || !STI.isOSWindows()) {
+      // Label destination operand of the subtract.
+      MF.makeDebugValueSubstitution(*InstrNum,
+                                    {ModInst->getDebugInstrNum(), 0});
+    } else {
+      // Label the call. The operand number is the penultimate operand.
+      MF.makeDebugValueSubstitution(
+          *InstrNum, {ModInst->getDebugInstrNum(), NumCallOps - 1});
+    }
+  }
+
   if (InProlog) {
     // Apply the frame setup flag to all inserted instrs.
     for (++ExpansionMBBI; ExpansionMBBI != MBBI; ++ExpansionMBBI)
       ExpansionMBBI->setFlag(MachineInstr::FrameSetup);
   }
 }
 
 static unsigned calculateSetFPREG(uint64_t SPAdjust) {

llvm/test/DebugInfo/X86/instr-ref-dyn-alloca-win32.ll

+; RUN: llc %s -o - -stop-after=machine-sink \
+; RUN:     -experimental-debug-variable-locations=true \
+; RUN:    | FileCheck %s --check-prefixes=DYN_LABEL
+; RUN: llc %s -o - -stop-before=phi-node-elimination \
+; RUN:     -experimental-debug-variable-locations=true \
+; RUN:    | FileCheck %s --check-prefixes=DYN_LOWERED
+
+;; This is a copy of DebugInfo/COFF/types-array-advanced.ll, checking that
+;; DYN_ALLOCAs correctly have instruction references updated. This is compiled
+;; from LLVM-IR because the "hasDynAlloca" flag isn't serialised to MIR, for
+;; complicated reasons, that means we can't store as MIR.
+
+;; The alloca instruction should be labelled, and we should refer to operand 2,
+;; which happens to be a def of $esp
+; DYN_LABEL: DYN_ALLOCA_32 {{.*}} debug-instr-number 1,
+; DYN_LABEL: DBG_INSTR_REF 1, 2
+
+;; Once lowered, on win32 _chkstk alters the stack pointer. We should label the
+;; call and it's SP operand, plus check for a value substitution.
+; DYN_LOWERED:      debugValueSubstitutions:
+; DYN_LOWERED-NEXT: - { srcinst: 1, srcop: 2, dstinst: 2, dstop: 6, subreg: 0 }
+;; Test that it's operand six, and labelled.
+; DYN_LOWERED:      CALLpcrel32 &_chkstk, implicit $esp, implicit $ssp,
+; DYN_LOWERED-SAME: implicit killed $eax, implicit $esp, implicit-def dead $eax,
+; DYN_LOWERED-SAME: implicit-def $esp, implicit-def dead $eflags,
+; DYN_LOWERED-SAME:  debug-instr-number 2,
+
+source_filename = "test/DebugInfo/COFF/types-array-advanced.ll"
+target datalayout = "e-m:x-p:32:32-i64:64-f80:32-n8:16:32-a:0:32-S32"
+target triple = "i686-pc-windows-msvc18.0.31101"
+
+%struct.incomplete_struct = type { i32 }
+
+@"\01?multi_dim_arr@@3PAY146DA" = global [2 x [5 x [7 x i8]]] zeroinitializer, align 1, !dbg !0
+@"\01?p_incomplete_struct_arr@@3PAY02Uincomplete_struct@@A" = global [3 x i8]* null, align 4, !dbg !6
+@"\01?incomplete_struct_arr@@3PAUincomplete_struct@@A" = global [3 x %struct.incomplete_struct] zeroinitializer, align 4, !dbg !16
+@"\01?typedef_arr@@3SDHD" = constant [4 x i32] zeroinitializer, align 4, !dbg !18
+
+; Function Attrs: nounwind
+define void @"\01?foo@@YAXH@Z"(i32 %x) #0 !dbg !35 {
+entry:
+  %x.addr = alloca i32, align 4
+  %saved_stack = alloca i8*, align 4
+  store i32 %x, i32* %x.addr, align 4
+  call void @llvm.dbg.declare(metadata i32* %x.addr, metadata !38, metadata !39), !dbg !40
+  %0 = load i32, i32* %x.addr, align 4, !dbg !41
+  %1 = call i8* @llvm.stacksave(), !dbg !42
+  store i8* %1, i8** %saved_stack, align 4, !dbg !42
+  %vla = alloca i32, i32 %0, align 4, !dbg !42
+  call void @llvm.dbg.declare(metadata i32* %vla, metadata !43, metadata !47), !dbg !48
+  %arrayidx = getelementptr inbounds i32, i32* %vla, i32 0, !dbg !49
+  store i32 0, i32* %arrayidx, align 4, !dbg !50
+  %2 = load i8*, i8** %saved_stack, align 4, !dbg !51
+  call void @llvm.stackrestore(i8* %2), !dbg !51
+  ret void, !dbg !51
+}
+
+; Function Attrs: nounwind readnone
+declare void @llvm.dbg.declare(metadata, metadata, metadata)
+
+; Function Attrs: nounwind
+declare i8* @llvm.stacksave()
+
+; Function Attrs: nounwind
+declare void @llvm.stackrestore(i8*)
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!32, !33}
+!llvm.ident = !{!34}
+
+!0 = distinct !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = !DIGlobalVariable(name: "multi_dim_arr", linkageName: "\01?multi_dim_arr@@3PAY146DA", scope: !2, file: !3, line: 1, type: !26, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !3, producer: "clang version 3.9.0 (trunk 273874)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !4, globals: !5)
+!3 = !DIFile(filename: "t.cpp", directory: "/")
+!4 = !{}
+!5 = !{!0, !6, !16, !18}
+!6 = distinct !DIGlobalVariableExpression(var: !7, expr: !DIExpression())
+!7 = !DIGlobalVariable(name: "p_incomplete_struct_arr", linkageName: "\01?p_incomplete_struct_arr@@3PAY02Uincomplete_struct@@A", scope: !2, file: !3, line: 3, type: !8, isLocal: false, isDefinition: true)
+!8 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !9, size: 32, align: 32)
+!9 = !DICompositeType(tag: DW_TAG_array_type, baseType: !10, elements: !14)
+!10 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "incomplete_struct", file: !3, line: 4, size: 32, align: 32, elements: !11, identifier: ".?AUincomplete_struct@@")
+!11 = !{!12}
+!12 = !DIDerivedType(tag: DW_TAG_member, name: "s1", scope: !10, file: !3, line: 5, baseType: !13, size: 32, align: 32)
+!13 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
+!14 = !{!15}
+!15 = !DISubrange(count: 3)
+!16 = distinct !DIGlobalVariableExpression(var: !17, expr: !DIExpression())
+!17 = !DIGlobalVariable(name: "incomplete_struct_arr", linkageName: "\01?incomplete_struct_arr@@3PAUincomplete_struct@@A", scope: !2, file: !3, line: 6, type: !9, isLocal: false, isDefinition: true)
+!18 = distinct !DIGlobalVariableExpression(var: !19, expr: !DIExpression())
+!19 = !DIGlobalVariable(name: "typedef_arr", linkageName: "\01?typedef_arr@@3SDHD", scope: !2, file: !3, line: 14, type: !20, isLocal: false, isDefinition: true)
+!20 = !DICompositeType(tag: DW_TAG_array_type, baseType: !21, size: 128, align: 32, elements: !24)
+!21 = !DIDerivedType(tag: DW_TAG_typedef, name: "T_INT", file: !3, line: 13, baseType: !22)
+!22 = !DIDerivedType(tag: DW_TAG_const_type, baseType: !23)
+!23 = !DIDerivedType(tag: DW_TAG_volatile_type, baseType: !13)
+!24 = !{!25}
+!25 = !DISubrange(count: 4)
+!26 = !DICompositeType(tag: DW_TAG_array_type, baseType: !27, size: 560, align: 8, elements: !28)
+!27 = !DIBasicType(name: "char", size: 8, align: 8, encoding: DW_ATE_signed_char)
+!28 = !{!29, !30, !31}
+!29 = !DISubrange(count: 2)
+!30 = !DISubrange(count: 5)
+!31 = !DISubrange(count: 7)
+!32 = !{i32 2, !"CodeView", i32 1}
+!33 = !{i32 2, !"Debug Info Version", i32 3}
+!34 = !{!"clang version 3.9.0 (trunk 273874)"}
+!35 = distinct !DISubprogram(name: "foo", linkageName: "\01?foo@@YAXH@Z", scope: !3, file: !3, line: 8, type: !36, isLocal: false, isDefinition: true, scopeLine: 8, flags: DIFlagPrototyped, isOptimized: false, unit: !2, retainedNodes: !4)
+!36 = !DISubroutineType(types: !37)
+!37 = !{null, !13}
+!38 = !DILocalVariable(name: "x", arg: 1, scope: !35, file: !3, line: 8, type: !13)
+!39 = !DIExpression()
+!40 = !DILocation(line: 8, column: 14, scope: !35)
+!41 = !DILocation(line: 9, column: 21, scope: !35)
+!42 = !DILocation(line: 9, column: 4, scope: !35)
+!43 = !DILocalVariable(name: "dyn_size_arr", scope: !35, file: !3, line: 9, type: !44)
+!44 = !DICompositeType(tag: DW_TAG_array_type, baseType: !13, align: 32, elements: !45)
+!45 = !{!46}
+!46 = !DISubrange(count: -1)
+!47 = !DIExpression(DW_OP_deref)
+!48 = !DILocation(line: 9, column: 8, scope: !35)
+!49 = !DILocation(line: 10, column: 4, scope: !35)
+!50 = !DILocation(line: 10, column: 20, scope: !35)
+!51 = !DILocation(line: 11, column: 1, scope: !35)
+

llvm/test/DebugInfo/X86/instr-ref-dyn-alloca.ll

+; RUN: llc %s -o - -stop-before=phi-node-elimination \
+; RUN:     -experimental-debug-variable-locations=true \
+; RUN:    | FileCheck %s --check-prefixes=DYN_LOWERED
+
+;; Copy of instr-ref-dyn-alloca-win32.ll, targetting 64 bit Windows. Here,
+;; _chkstk doesn't return with a modified stack pointer, instead we have to
+;; edit it with our own subtract operation. Check that it's labelled and the
+;; substitution is correct. This also covers the code paths for non-Windows on
+;; x86.
+
+; DYN_LOWERED:      debugValueSubstitutions:
+; DYN_LOWERED-NEXT: - { srcinst: 1, srcop: 2, dstinst: 2, dstop: 0, subreg: 0 }
+; DYN_LOWERED:      SUB64rr $rsp, killed $rax, {{.*}} debug-instr-number 2,
+
+source_filename = "test/DebugInfo/COFF/types-array-advanced.ll"
+target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-windows-msvc19.0.24210"
+
+%struct.incomplete_struct = type { i32 }
+
+@"\01?multi_dim_arr@@3PAY146DA" = global [2 x [5 x [7 x i8]]] zeroinitializer, align 1, !dbg !0
+@"\01?p_incomplete_struct_arr@@3PAY02Uincomplete_struct@@A" = global [3 x i8]* null, align 4, !dbg !6
+@"\01?incomplete_struct_arr@@3PAUincomplete_struct@@A" = global [3 x %struct.incomplete_struct] zeroinitializer, align 4, !dbg !16
+@"\01?typedef_arr@@3SDHD" = constant [4 x i32] zeroinitializer, align 4, !dbg !18
+
+; Function Attrs: nounwind
+define void @"\01?foo@@YAXH@Z"(i32 %x) #0 !dbg !35 {
+entry:
+  %x.addr = alloca i32, align 4
+  %saved_stack = alloca i8*, align 4
+  store i32 %x, i32* %x.addr, align 4
+  call void @llvm.dbg.declare(metadata i32* %x.addr, metadata !38, metadata !39), !dbg !40
+  %0 = load i32, i32* %x.addr, align 4, !dbg !41
+  %1 = call i8* @llvm.stacksave(), !dbg !42
+  store i8* %1, i8** %saved_stack, align 4, !dbg !42
+  %vla = alloca i32, i32 %0, align 4, !dbg !42
+  call void @llvm.dbg.declare(metadata i32* %vla, metadata !43, metadata !47), !dbg !48
+  %arrayidx = getelementptr inbounds i32, i32* %vla, i32 0, !dbg !49
+  store i32 0, i32* %arrayidx, align 4, !dbg !50
+  %2 = load i8*, i8** %saved_stack, align 4, !dbg !51
+  call void @llvm.stackrestore(i8* %2), !dbg !51
+  ret void, !dbg !51
+}
+
+; Function Attrs: nounwind readnone
+declare void @llvm.dbg.declare(metadata, metadata, metadata)
+
+; Function Attrs: nounwind
+declare i8* @llvm.stacksave()
+
+; Function Attrs: nounwind
+declare void @llvm.stackrestore(i8*)
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!32, !33}
+!llvm.ident = !{!34}
+
+!0 = distinct !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
+!1 = !DIGlobalVariable(name: "multi_dim_arr", linkageName: "\01?multi_dim_arr@@3PAY146DA", scope: !2, file: !3, line: 1, type: !26, isLocal: false, isDefinition: true)
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !3, producer: "clang version 3.9.0 (trunk 273874)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !4, globals: !5)
+!3 = !DIFile(filename: "t.cpp", directory: "/")
+!4 = !{}
+!5 = !{!0, !6, !16, !18}
+!6 = distinct !DIGlobalVariableExpression(var: !7, expr: !DIExpression())
+!7 = !DIGlobalVariable(name: "p_incomplete_struct_arr", linkageName: "\01?p_incomplete_struct_arr@@3PAY02Uincomplete_struct@@A", scope: !2, file: !3, line: 3, type: !8, isLocal: false, isDefinition: true)
+!8 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !9, size: 32, align: 32)
+!9 = !DICompositeType(tag: DW_TAG_array_type, baseType: !10, elements: !14)
+!10 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "incomplete_struct", file: !3, line: 4, size: 32, align: 32, elements: !11, identifier: ".?AUincomplete_struct@@")
+!11 = !{!12}
+!12 = !DIDerivedType(tag: DW_TAG_member, name: "s1", scope: !10, file: !3, line: 5, baseType: !13, size: 32, align: 32)
+!13 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
+!14 = !{!15}
+!15 = !DISubrange(count: 3)
+!16 = distinct !DIGlobalVariableExpression(var: !17, expr: !DIExpression())
+!17 = !DIGlobalVariable(name: "incomplete_struct_arr", linkageName: "\01?incomplete_struct_arr@@3PAUincomplete_struct@@A", scope: !2, file: !3, line: 6, type: !9, isLocal: false, isDefinition: true)
+!18 = distinct !DIGlobalVariableExpression(var: !19, expr: !DIExpression())
+!19 = !DIGlobalVariable(name: "typedef_arr", linkageName: "\01?typedef_arr@@3SDHD", scope: !2, file: !3, line: 14, type: !20, isLocal: false, isDefinition: true)
+!20 = !DICompositeType(tag: DW_TAG_array_type, baseType: !21, size: 128, align: 32, elements: !24)
+!21 = !DIDerivedType(tag: DW_TAG_typedef, name: "T_INT", file: !3, line: 13, baseType: !22)
+!22 = !DIDerivedType(tag: DW_TAG_const_type, baseType: !23)
+!23 = !DIDerivedType(tag: DW_TAG_volatile_type, baseType: !13)
+!24 = !{!25}
+!25 = !DISubrange(count: 4)
+!26 = !DICompositeType(tag: DW_TAG_array_type, baseType: !27, size: 560, align: 8, elements: !28)
+!27 = !DIBasicType(name: "char", size: 8, align: 8, encoding: DW_ATE_signed_char)
+!28 = !{!29, !30, !31}
+!29 = !DISubrange(count: 2)
+!30 = !DISubrange(count: 5)
+!31 = !DISubrange(count: 7)
+!32 = !{i32 2, !"CodeView", i32 1}
+!33 = !{i32 2, !"Debug Info Version", i32 3}
+!34 = !{!"clang version 3.9.0 (trunk 273874)"}
+!35 = distinct !DISubprogram(name: "foo", linkageName: "\01?foo@@YAXH@Z", scope: !3, file: !3, line: 8, type: !36, isLocal: false, isDefinition: true, scopeLine: 8, flags: DIFlagPrototyped, isOptimized: false, unit: !2, retainedNodes: !4)
+!36 = !DISubroutineType(types: !37)
+!37 = !{null, !13}
+!38 = !DILocalVariable(name: "x", arg: 1, scope: !35, file: !3, line: 8, type: !13)
+!39 = !DIExpression()
+!40 = !DILocation(line: 8, column: 14, scope: !35)
+!41 = !DILocation(line: 9, column: 21, scope: !35)
+!42 = !DILocation(line: 9, column: 4, scope: !35)
+!43 = !DILocalVariable(name: "dyn_size_arr", scope: !35, file: !3, line: 9, type: !44)
+!44 = !DICompositeType(tag: DW_TAG_array_type, baseType: !13, align: 32, elements: !45)
+!45 = !{!46}
+!46 = !DISubrange(count: -1)
+!47 = !DIExpression(DW_OP_deref)
+!48 = !DILocation(line: 9, column: 8, scope: !35)
+!49 = !DILocation(line: 10, column: 4, scope: !35)
+!50 = !DILocation(line: 10, column: 20, scope: !35)
+!51 = !DILocation(line: 11, column: 1, scope: !35)
+