[ARM SEH 3] [ARM] [MC] Add support for writing ARM WinEH unwind info

[mstorsjo]

Summary

This includes .seh_* directives for generating it from assembly.
It is designed fairly similarly to the ARM64 handling.

For .seh_handler directives, such as
".seh_handler __C_specific_handler, @except" (which is supported
on x86_64 and aarch64 so far), the "@except" bit doesn't work in
ARM assembly, as '@' is used as a comment character (on all current
platforms).

Allow using '%' instead of '@' for this purpose. This convention
is used by GAS in similar contexts already,
e.g. [1]:

Note on targets where the @ character is the start of a comment
(eg ARM) then another character is used instead. For example the
ARM port uses the % character.

In practice, this unfortunately means that all such .seh_handler
directives will need ifdefs for ARM.

Contrary to ARM64, on ARM, it's quite common that we can't evaluate
e.g. the function length at this point, due to instructions whose
length is finalized later. (Also, inline jump tables end with
a ".p2align 1".)

If unable to to evaluate the function length immediately, emit
it as an MCExpr instead. If we'd implement splitting the unwind
info for a function (which isn't implemented for ARM64 yet either),
we wouldn't know whether we need to split it though.

TODO: We probably should emit this with an ARM specific fixup, so
that we at least could implement range checking before writing the
final value after evaluating the MCExpr.

We maybe should rename MCWin64EH, as this is used on a 32 bit
architecture too.

[1] https://sourceware.org/binutils/docs/as/Section.html#Section

[efriedma]

It looks like this doesn't include any directive to set the "F" bit in xdata, or the "Condition" field in epilogue scopes. Is the omission intentional? (I assume your LLVM patch won't emit sequences that would require either of those, but that might change, and people writing assembly might want them.)

llvm/include/llvm/Support/Win64EH.h

79

Not sure I like calling this operation UOP_SetFP, given it can set an arbitrary register.

[mstorsjo]

In D125645#3516878, @efriedma wrote:

It looks like this doesn't include any directive to set the "F" bit in xdata,

That's true. If we'd have unwind info splitting implemented, that bit would be set on the follow-up fragments, implicitly. Maybe there's a usecase for manually creating such unwind data though? I think that one can be added without much troubles later, if there's a concrete demand for it though...

(On that note, when rereading the docs about the "F" bit, I'm left wondering how that's supposed to work. If we have a huge function, and describe a prologue-less fragment in the middle of the function - how will unwinding from it work, when there's no prologue opcodes in that fragment to use for unwinding, if unwinding starts from outside of any potential epilogue in it, as there also can be fragments without epilogues?)

or the "Condition" field in epilogue scopes. Is the omission intentional? (I assume your LLVM patch won't emit sequences that would require either of those, but that might change, and people writing assembly might want them.)

Right, that's probably a bit bigger omission. Yeah it's not needed by our generated code, but it would probably be good to be able to express it if we wanted to. Maybe .seh_startepilogue_cond <cond> or something like that? (We'd probably want to have the default .seh_startepilogue as is - I think.)

llvm/include/llvm/Support/Win64EH.h

79

Oh, right, I guess I brought over the design a bit too much from the aarch64 case here, while it's notably different here.

Should we call it UOP_SaveSP and .seh_save_sp then?

[efriedma]

In D125645#3516998, @mstorsjo wrote:

In D125645#3516878, @efriedma wrote:

It looks like this doesn't include any directive to set the "F" bit in xdata,

That's true. If we'd have unwind info splitting implemented, that bit would be set on the follow-up fragments, implicitly. Maybe there's a usecase for manually creating such unwind data though? I think that one can be added without much troubles later, if there's a concrete demand for it though...

We would want this to split a function across multiple section (e.g. for hot/cold splitting).

(On that note, when rereading the docs about the "F" bit, I'm left wondering how that's supposed to work. If we have a huge function, and describe a prologue-less fragment in the middle of the function - how will unwinding from it work, when there's no prologue opcodes in that fragment to use for unwinding, if unwinding starts from outside of any potential epilogue in it, as there also can be fragments without epilogues?)

I'm pretty sure the idea is that you still emit the prologue opcodes from the original prologue. The "F" bit just means that the prologue is not part of the current fragment. (For AArch64, this idea was generalized with the "c_end" opcode.)

or the "Condition" field in epilogue scopes. Is the omission intentional? (I assume your LLVM patch won't emit sequences that would require either of those, but that might change, and people writing assembly might want them.)

Right, that's probably a bit bigger omission. Yeah it's not needed by our generated code, but it would probably be good to be able to express it if we wanted to. Maybe .seh_startepilogue_cond <cond> or something like that? (We'd probably want to have the default .seh_startepilogue as is - I think.)

Sure.

llvm/include/llvm/Support/Win64EH.h

79

That sounds reasonable.

[mstorsjo]

In D125645#3517116, @efriedma wrote:

In D125645#3516998, @mstorsjo wrote:

In D125645#3516878, @efriedma wrote:

It looks like this doesn't include any directive to set the "F" bit in xdata,

That's true. If we'd have unwind info splitting implemented, that bit would be set on the follow-up fragments, implicitly. Maybe there's a usecase for manually creating such unwind data though? I think that one can be added without much troubles later, if there's a concrete demand for it though...

We would want this to split a function across multiple section (e.g. for hot/cold splitting).

Oh, right, true.

(On that note, when rereading the docs about the "F" bit, I'm left wondering how that's supposed to work. If we have a huge function, and describe a prologue-less fragment in the middle of the function - how will unwinding from it work, when there's no prologue opcodes in that fragment to use for unwinding, if unwinding starts from outside of any potential epilogue in it, as there also can be fragments without epilogues?)

I'm pretty sure the idea is that you still emit the prologue opcodes from the original prologue. The "F" bit just means that the prologue is not part of the current fragment. (For AArch64, this idea was generalized with the "c_end" opcode.)

Ah, yes. (I just rechecked this with my empirical tests with RtlVirtualUnwind too.)

Would this look like a reasonable way of handling it?

myfragment:
    .seh_proc myfragment
    .seh_save_regs {r4-r7,lr}
    .seh_stackalloc 1234
    .seh_endprologue_fragment
    // fragment code
    .seh_startepilogue
    // regular epilogue

Thus we'd pile up the phantom prologue opcodes at the start, and end it with an . seh_endprologue_fragment (which also implies that the opcodes aren't expected to exactly match the range between .seh_proc and .seh_endprologue_fragment).

[efriedma]

.seh_endprologue_fragment looks fine.

[mstorsjo]

Renamed the opcode for mov rX, sp to UOP_SaveSP and .seh_save_sp, added .seh_endprologue_fragment and .seh_startepilogue_cond.

[mstorsjo]

llvm/include/llvm/MC/MCWinEH.h

60

I guess it'd be much preferrable here, to change this into a map containing a struct instead, e.g.:

struct Epilog {
    std::vector<Instruction> Instructions;
    unsigned Condition;
};
MapVector<MCSymbol *, Epilog> EpilogMap;

And if doing that, I guess it'd be best to do such a refactoring (without the Condition field) as a separate preceding commit. But posting it in this form for now, for potential comments on the rest of the changes.

llvm/lib/MC/MCWin64EH.cpp

1441

How do I make this into an ARM specific fixup, which could do range checking for FuncLengthExpr and error out if the function is too long?

[efriedma]

llvm/lib/MC/MCWin64EH.cpp

1441

See llvm/lib/Target/ARM/MCTargetDesc/ARMMCExpr.h.

I'd prefer not to put any substantial effort into infrastructure we won't need if we implement splitting, though. Zhaoshi is currently looking into that.

[mstorsjo]

llvm/lib/MC/MCAsmStreamer.cpp

2079

What do you think about this use of % instead of @ here, does it seem reasonable?

The alternative would be to switch to a different comment character, but I'm afraid that would break lots of handwritten assembly out there - while there's very little code that uses these explicit .seh_handler directives, so I guess those will just have to cope with some ifdefs.

llvm/lib/MC/MCWin64EH.cpp

1441

Oh, ok. Sure, I'll stop worrying about that aspect then! (As I've never hit that limit in practice on ARM64 so far, I'm not very worried about it for ARM either. But for implementing splitting, the fact that ARM functions often won't have their length fixed at this point might be problematic...)

[mstorsjo]

Cleaner implementation of conditional epilogues, using the preparatory refactoring from D125879.

[efriedma]

llvm/lib/MC/MCAsmStreamer.cpp

2079

Switching the character seems fine.

llvm/lib/MC/MCWin64EH.cpp

1441

If we need to, we can put off emitting the unwind info until the layout is fixed. See the giant FIXME near the beginning of ARM64EmitUnwindInfo.

[mstorsjo]

llvm/lib/MC/MCWin64EH.cpp

1441

Ah, I see! (I've read it a couple times before but never really confidently understood all the alternatives - but now I think I get the part about a relaxable fragment too).

Anyway, the current code should be fine for practical cases here, and whenever we get support for delaying emitting the unwind info for ARM64, we can apply the same to ARM too.

[mstorsjo]

Any further comments on patches 3-5? AFAIK there are no outstanding issues that have been brought up on them. (I'm still looking into fixing the frame pointer placement for patches 6-7.)

[efriedma]

Did you consider instead of adding .seh_endepilogue_nop, just require writing .seh_nop; .seh_endepilogue? Then we can automatically merge a tailing nop with the "end" opcode.

llvm/lib/Target/ARM/AsmParser/ARMAsmParser.cpp

12106

Do we really need the pc -> lr remapping, as opposed to just forbidding users from specifying pc? Seems a little weird to have multiple ways to write the same unwind code.

[mstorsjo]

In D125645#3528320, @efriedma wrote:

Did you consider instead of adding .seh_endepilogue_nop, just require writing .seh_nop; .seh_endepilogue? Then we can automatically merge a tailing nop with the "end" opcode.

Hmm, I didn't consider that - as they're separate opcodes in the unwind info, I was kinda led into this format. But that sounds like a good idea - I'll try it out and see if it seems better or not!

llvm/lib/Target/ARM/AsmParser/ARMAsmParser.cpp

12106

I guess it's not strictly necessary, but both for consistency with the preceding instruction, you may want to use both; pop {pc} and pop {lr}; bx lr both would use the same .seh_save_regs {lr} opcode, while I think you'd might want to write .seh_save_regs {pc} in the former case.

[mstorsjo]

In D125645#3528343, @mstorsjo wrote:

In D125645#3528320, @efriedma wrote:

Did you consider instead of adding .seh_endepilogue_nop, just require writing .seh_nop; .seh_endepilogue? Then we can automatically merge a tailing nop with the "end" opcode.

Hmm, I didn't consider that - as they're separate opcodes in the unwind info, I was kinda led into this format. But that sounds like a good idea - I'll try it out and see if it seems better or not!

Yep, overall that seems to make things even cleaner and looks like a good direction - thanks!

However, doing that exposes a different underlying issue in the handling of the SEH_* machine instructions used in frame lowering that hadn't been a problem so far - I'll follow up in D125648 with that issue.

[mstorsjo]

Remove the .seh_endepilogue_nop variants and use explicit .seh_nop instructions instead.

[mstorsjo]

Any more comments on this set? I’m generally happy with patches 3-6, and patch 7 seems to work even if it’s not very pretty.

[efriedma]

llvm/lib/CodeGen/AsmPrinter/WinException.cpp

764

I assume this is related somehow, but not sure how, exactly.

llvm/lib/Target/ARM/MCTargetDesc/ARMWinCOFFStreamer.cpp

114

This comment is out-of-date?

[mstorsjo]

llvm/lib/CodeGen/AsmPrinter/WinException.cpp

764

Ah, hmm. I think this might be leftovers from initial attempts at making things work for MSVC mode (for either C++ unwinding or SEH __try). I left that out of scope for now (itanium unwinding for mingw mode works fine without it though), so I guess I should leave it out here too.

llvm/lib/Target/ARM/MCTargetDesc/ARMWinCOFFStreamer.cpp

114

Oh, oops. I guess I'll leave out the specifics of the opcodes here as I don't think it's relevant to elaborate on them here.

[mstorsjo]

llvm/lib/CodeGen/AsmPrinter/WinException.cpp

764

Ah, no, now I remember. We have an existing testcase, test/CodeGen/ARM/Windows/wineh-basic.ll, which tests some C++ exception handling constructs.

Since we enable WinEHEncodingType = WinEH::EncodingType::Itanium; in this commit, we end up trying to do more things than we've done before, and that testcase triggers a failed assert, as we try to call UnwindHelpOffset = getFrameIndexOffset(FuncInfo.UnwindHelpFrameIdx, FuncInfo) when FuncInfo.UnwindHelpFrameIdx is unset (left at its default value of std::numeric_limits<int>::max()).

So these changes could be removed later when MSVC C++ exception handling is properly implemented for ARM, but for now, this avoids breaking the existing testcase.

[mstorsjo]

Removed leftover copypaste bits from AArch64 in a comment.

[mstorsjo]

Rebased after superclass methods have been renamed.

[mstorsjo]

Ping - any more on this? If you want to, I could add comments at the changes in AsmPrinter/WinException.cpp that they can be removed if the MS C++ unwinding is fully implemented for ARM (and/or remove the wineh-basic.ll testcase?).

[efriedma]

I think I'd like some sort of comment on the WinException.cpp, yes.

Otherwise LGTM

llvm/include/llvm/MC/MCStreamer.h

public:
   virtual void emitInst(uint32_t Inst, char Suffix = '\0');
 
   virtual void AnnotateTLSDescriptorSequence(const MCSymbolRefExpr *SRE);
 
   virtual void emitThumbSet(MCSymbol *Symbol, const MCExpr *Value);
 
   void emitConstantPools() override;
 
+  virtual void emitARMWinCFIAllocStack(unsigned Size, bool Wide);
+  virtual void emitARMWinCFISaveRegMask(unsigned Mask, bool Wide);
+  virtual void emitARMWinCFISaveSP(unsigned Reg);
+  virtual void emitARMWinCFISaveFRegs(unsigned First, unsigned Last);
+  virtual void emitARMWinCFISaveLR(unsigned Offset);
+  virtual void emitARMWinCFIPrologEnd(bool Fragment);
+  virtual void emitARMWinCFINop(bool Wide);
+  virtual void emitARMWinCFIEpilogStart(unsigned Condition);
+  virtual void emitARMWinCFIEpilogEnd();
+  virtual void emitARMWinCFICustom(unsigned Opcode);
+
   /// Reset any state between object emissions, i.e. the equivalent of
   /// MCStreamer's reset method.
   virtual void reset();
 
   /// Callback used to implement the ldr= pseudo.
   /// Add a new entry to the constant pool for the current section and return an
   /// MCExpr that can be used to refer to the constant pool location.
   const MCExpr *addConstantPoolEntry(const MCExpr *, SMLoc Loc);

llvm/include/llvm/MC/MCWin64EH.h

 
 class UnwindEmitter : public WinEH::UnwindEmitter {
 public:
   void Emit(MCStreamer &Streamer) const override;
   void EmitUnwindInfo(MCStreamer &Streamer, WinEH::FrameInfo *FI,
                       bool HandlerData) const override;
 };
 
-class ARM64UnwindEmitter : public WinEH::UnwindEmitter {
+class ARMUnwindEmitter : public WinEH::UnwindEmitter {
 public:
   void Emit(MCStreamer &Streamer) const override;
   void EmitUnwindInfo(MCStreamer &Streamer, WinEH::FrameInfo *FI,
                       bool HandlerData) const override;
 };
 
+class ARM64UnwindEmitter : public WinEH::UnwindEmitter {
+public:
+  void Emit(MCStreamer &Streamer) const override;
+  void EmitUnwindInfo(MCStreamer &Streamer, WinEH::FrameInfo *FI,
+                      bool HandlerData) const override;
+};
 }
 } // end namespace llvm
 
 #endif

llvm/include/llvm/MC/MCWinEH.h

 //===- MCWinEH.h - Windows Unwinding Support --------------------*- 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
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_MC_MCWINEH_H
 #define LLVM_MC_MCWINEH_H
 
+#include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/MapVector.h"
 #include <vector>
 
 namespace llvm {
 class MCSection;
 class MCStreamer;
 class MCSymbol;
 
struct FrameInfo {
   const MCSymbol *PrologEnd = nullptr;
   const MCSymbol *Symbol = nullptr;
   MCSection *TextSection = nullptr;
   uint32_t PackedInfo = 0;
 
   bool HandlesUnwind = false;
   bool HandlesExceptions = false;
   bool EmitAttempted = false;
+  bool Fragment = false;
 
   int LastFrameInst = -1;
   const FrameInfo *ChainedParent = nullptr;
   std::vector<Instruction> Instructions;
   struct Epilog {
     std::vector<Instruction> Instructions;
+    unsigned Condition;
   };
   MapVector<MCSymbol *, Epilog> EpilogMap;
 
   FrameInfo() = default;
   FrameInfo(const MCSymbol *Function, const MCSymbol *BeginFuncEHLabel)
       : Begin(BeginFuncEHLabel), Function(Function) {}
   FrameInfo(const MCSymbol *Function, const MCSymbol *BeginFuncEHLabel,
             const FrameInfo *ChainedParent)

llvm/include/llvm/Support/Win64EH.h

 #include "llvm/Support/Endian.h"
 
 namespace llvm {
 namespace Win64EH {
 
 /// UnwindOpcodes - Enumeration whose values specify a single operation in
 /// the prolog of a function.
 enum UnwindOpcodes {
+  // The following set of unwind opcodes is for x86_64.  They are documented at
+  // https://docs.microsoft.com/en-us/cpp/build/exception-handling-x64.
+  // Some generic values from this set are used for other architectures too.
   UOP_PushNonVol = 0,
   UOP_AllocLarge,
   UOP_AllocSmall,
   UOP_SetFPReg,
   UOP_SaveNonVol,
   UOP_SaveNonVolBig,
   UOP_Epilog,
   UOP_SpareCode,
enum UnwindOpcodes {
   UOP_SaveFRegPX,
   UOP_SetFP,
   UOP_AddFP,
   UOP_Nop,
   UOP_End,
   UOP_SaveNext,
   UOP_TrapFrame,
   UOP_Context,
-  UOP_ClearUnwoundToCall
+  UOP_ClearUnwoundToCall,
+  // The following set of unwind opcodes is for ARM.  They are documented at
+  // https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling
+
+  // Stack allocations use UOP_AllocSmall, UOP_AllocLarge from above, plus
+  // the following. AllocSmall, AllocLarge and AllocHuge represent a 16 bit
+  // instruction, while the WideAlloc* opcodes represent a 32 bit instruction.
+  // Small can represent a stack offset of 0x7f*4 (252) bytes, Medium can
+  // represent up to 0x3ff*4 (4092) bytes, Large up to 0xffff*4 (262140) bytes,
+  // and Huge up to 0xffffff*4 (67108860) bytes.
+  UOP_AllocHuge,
+  UOP_WideAllocMedium,
+  UOP_WideAllocLarge,
+  UOP_WideAllocHuge,
+
+  UOP_WideSaveRegMask,
+  UOP_SaveSP,
+  UOP_SaveRegsR4R7LR,
+  UOP_WideSaveRegsR4R11LR,
+  UOP_SaveFRegD8D15,
+  UOP_SaveRegMask,
+  UOP_SaveLR,
+  UOP_SaveFRegD0D15,
+  UOP_SaveFRegD16D31,
+  // Using UOP_Nop from above
+  UOP_WideNop,
+  // Using UOP_End from above
+  UOP_EndNop,
+  UOP_WideEndNop,
+  // A custom unspecified opcode, consisting of one or more bytes. This
+  // allows producing opcodes in the implementation defined/reserved range.
+  UOP_Custom,
 };
 
 /// UnwindCode - This union describes a single operation in a function prolog,
 /// or part thereof.
 union UnwindCode {
   struct {
     uint8_t CodeOffset;
     uint8_t UnwindOpAndOpInfo;

llvm/lib/CodeGen/AsmPrinter/WinException.cpp

if (shouldEmitPersonality) {
     FuncInfoXData =
         Asm->OutContext.getOrCreateSymbol(Twine("$cppxdata$", FuncLinkageName));
     computeIP2StateTable(MF, FuncInfo, IPToStateTable);
   } else {
     FuncInfoXData = Asm->OutContext.getOrCreateLSDASymbol(FuncLinkageName);
   }
 
   int UnwindHelpOffset = 0;
-  if (Asm->MAI->usesWindowsCFI())
+  // TODO: The check for UnwindHelpFrameIdx against max() below (and the
+  // second check further below) can be removed if MS C++ unwinding is
+  // implemented for ARM, when test/CodeGen/ARM/Windows/wineh-basic.ll
+  // passes without the check.
+  if (Asm->MAI->usesWindowsCFI() &&
+      FuncInfo.UnwindHelpFrameIdx != std::numeric_limits<int>::max())
     UnwindHelpOffset =
         getFrameIndexOffset(FuncInfo.UnwindHelpFrameIdx, FuncInfo);
 
   MCSymbol *UnwindMapXData = nullptr;
   MCSymbol *TryBlockMapXData = nullptr;
   MCSymbol *IPToStateXData = nullptr;
   if (!FuncInfo.CxxUnwindMap.empty())
     UnwindMapXData = Asm->OutContext.getOrCreateSymbol(
void WinException::emitCXXFrameHandler3Table(const MachineFunction *MF) {
   OS.emitValue(create32bitRef(TryBlockMapXData), 4);
 
   AddComment("IPMapEntries");
   OS.emitInt32(IPToStateTable.size());
 
   AddComment("IPToStateXData");
   OS.emitValue(create32bitRef(IPToStateXData), 4);
 
-  if (Asm->MAI->usesWindowsCFI()) {
+  if (Asm->MAI->usesWindowsCFI() &&
+      FuncInfo.UnwindHelpFrameIdx != std::numeric_limits<int>::max()) {
     AddComment("UnwindHelp");
     OS.emitInt32(UnwindHelpOffset);
   }
 
   AddComment("ESTypeList");
   OS.emitInt32(0);
 
   AddComment("EHFlags");

llvm/lib/MC/MCAsmStreamer.cpp

 }
 
 void MCAsmStreamer::emitWinEHHandler(const MCSymbol *Sym, bool Unwind,
                                      bool Except, SMLoc Loc) {
   MCStreamer::emitWinEHHandler(Sym, Unwind, Except, Loc);
 
   OS << "\t.seh_handler ";
   Sym->print(OS, MAI);
+  char Marker = '@';
+  const Triple &T = getContext().getTargetTriple();
+  if (T.getArch() == Triple::arm || T.getArch() == Triple::thumb)
+    Marker = '%';
   if (Unwind)
-    OS << ", @unwind";
+    OS << ", " << Marker << "unwind";
   if (Except)
-    OS << ", @except";
+    OS << ", " << Marker << "except";
   EmitEOL();
 }
 
 void MCAsmStreamer::emitWinEHHandlerData(SMLoc Loc) {
   MCStreamer::emitWinEHHandlerData(Loc);
 
   // Switch sections. Don't call SwitchSection directly, because that will
   // cause the section switch to be visible in the emitted assembly.

llvm/lib/MC/MCParser/COFFAsmParser.cpp

 bool COFFAsmParser::ParseSEHDirectiveEndProlog(StringRef, SMLoc Loc) {
   Lex();
   getStreamer().emitWinCFIEndProlog(Loc);
   return false;
 }
 
 bool COFFAsmParser::ParseAtUnwindOrAtExcept(bool &unwind, bool &except) {
   StringRef identifier;
-  if (getLexer().isNot(AsmToken::At))
-    return TokError("a handler attribute must begin with '@'");
+  if (getLexer().isNot(AsmToken::At) && getLexer().isNot(AsmToken::Percent))
+    return TokError("a handler attribute must begin with '@' or '%'");
   SMLoc startLoc = getLexer().getLoc();
   Lex();
   if (getParser().parseIdentifier(identifier))
     return Error(startLoc, "expected @unwind or @except");
   if (identifier == "unwind")
     unwind = true;
   else if (identifier == "except")
     except = true;

llvm/lib/MC/MCWin64EH.cpp

 //===- lib/MC/MCWin64EH.cpp - MCWin64EH implementation --------------------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/MC/MCWin64EH.h"
+#include "llvm/ADT/Optional.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCObjectStreamer.h"
 #include "llvm/MC/MCStreamer.h"
 #include "llvm/MC/MCSymbol.h"
 #include "llvm/Support/Win64EH.h"
 namespace llvm {
void llvm::Win64EH::UnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer,
   // Switch sections (the static function above is meant to be called from
   // here and from Emit().
   MCSection *XData = Streamer.getAssociatedXDataSection(info->TextSection);
   Streamer.SwitchSection(XData);
 
   ::EmitUnwindInfo(Streamer, info);
 }
 
-static int64_t GetAbsDifference(MCStreamer &Streamer, const MCSymbol *LHS,
+static const MCExpr *GetSubDivExpr(MCStreamer &Streamer, const MCSymbol *LHS,
+                                   const MCSymbol *RHS, int Div) {
+  MCContext &Context = Streamer.getContext();
+  const MCExpr *Expr =
+      MCBinaryExpr::createSub(MCSymbolRefExpr::create(LHS, Context),
+                              MCSymbolRefExpr::create(RHS, Context), Context);
+  if (Div != 1)
+    Expr = MCBinaryExpr::createDiv(Expr, MCConstantExpr::create(Div, Context),
+                                   Context);
+  return Expr;
+}
+
+static Optional<int64_t> GetOptionalAbsDifference(MCStreamer &Streamer,
+                                                  const MCSymbol *LHS,
                                                   const MCSymbol *RHS) {
   MCContext &Context = Streamer.getContext();
   const MCExpr *Diff =
       MCBinaryExpr::createSub(MCSymbolRefExpr::create(LHS, Context),
                               MCSymbolRefExpr::create(RHS, Context), Context);
   MCObjectStreamer *OS = (MCObjectStreamer *)(&Streamer);
   // It should normally be possible to calculate the length of a function
   // at this point, but it might not be possible in the presence of certain
   // unusual constructs, like an inline asm with an alignment directive.
   int64_t value;
   if (!Diff->evaluateAsAbsolute(value, OS->getAssembler()))
-    report_fatal_error("Failed to evaluate function length in SEH unwind info");
+    return None;
   return value;
 }
 
+static int64_t GetAbsDifference(MCStreamer &Streamer, const MCSymbol *LHS,
+                                const MCSymbol *RHS) {
+  Optional<int64_t> MaybeDiff = GetOptionalAbsDifference(Streamer, LHS, RHS);
+  if (!MaybeDiff)
+    report_fatal_error("Failed to evaluate function length in SEH unwind info");
+  return *MaybeDiff;
+}
+
 static uint32_t ARM64CountOfUnwindCodes(ArrayRef<WinEH::Instruction> Insns) {
   uint32_t Count = 0;
   for (const auto &I : Insns) {
     switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) {
     default:
       llvm_unreachable("Unsupported ARM64 unwind code");
     case Win64EH::UOP_AllocSmall:
       Count += 1;
static void ARM64EmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info,
 
   if (info->HandlesExceptions)
     streamer.emitValue(
         MCSymbolRefExpr::create(info->ExceptionHandler,
                                 MCSymbolRefExpr::VK_COFF_IMGREL32, context),
         4);
 }
 
-static void ARM64EmitRuntimeFunction(MCStreamer &streamer,
-                                     const WinEH::FrameInfo *info) {
+static uint32_t ARMCountOfUnwindCodes(ArrayRef<WinEH::Instruction> Insns) {
+  uint32_t Count = 0;
+  for (const auto &I : Insns) {
+    switch (static_cast<Win64EH::UnwindOpcodes>(I.Operation)) {
+    default:
+      llvm_unreachable("Unsupported ARM unwind code");
+    case Win64EH::UOP_AllocSmall:
+      Count += 1;
+      break;
+    case Win64EH::UOP_AllocLarge:
+      Count += 3;
+      break;
+    case Win64EH::UOP_AllocHuge:
+      Count += 4;
+      break;
+    case Win64EH::UOP_WideAllocMedium:
+      Count += 2;
+      break;
+    case Win64EH::UOP_WideAllocLarge:
+      Count += 3;
+      break;
+    case Win64EH::UOP_WideAllocHuge:
+      Count += 4;
+      break;
+    case Win64EH::UOP_WideSaveRegMask:
+      Count += 2;
+      break;
+    case Win64EH::UOP_SaveSP:
+      Count += 1;
+      break;
+    case Win64EH::UOP_SaveRegsR4R7LR:
+      Count += 1;
+      break;
+    case Win64EH::UOP_WideSaveRegsR4R11LR:
+      Count += 1;
+      break;
+    case Win64EH::UOP_SaveFRegD8D15:
+      Count += 1;
+      break;
+    case Win64EH::UOP_SaveRegMask:
+      Count += 2;
+      break;
+    case Win64EH::UOP_SaveLR:
+      Count += 2;
+      break;
+    case Win64EH::UOP_SaveFRegD0D15:
+      Count += 2;
+      break;
+    case Win64EH::UOP_SaveFRegD16D31:
+      Count += 2;
+      break;
+    case Win64EH::UOP_Nop:
+    case Win64EH::UOP_WideNop:
+    case Win64EH::UOP_End:
+    case Win64EH::UOP_EndNop:
+    case Win64EH::UOP_WideEndNop:
+      Count += 1;
+      break;
+    case Win64EH::UOP_Custom: {
+      int J;
+      for (J = 3; J > 0; J--)
+        if (I.Offset & (0xffu << (8 * J)))
+          break;
+      Count += J + 1;
+      break;
+    }
+    }
+  }
+  return Count;
+}
+
+// Unwind opcode encodings and restrictions are documented at
+// https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling
+static void ARMEmitUnwindCode(MCStreamer &streamer,
+                              const WinEH::Instruction &inst) {
+  uint32_t w, lr;
+  int i;
+  switch (static_cast<Win64EH::UnwindOpcodes>(inst.Operation)) {
+  default:
+    llvm_unreachable("Unsupported ARM unwind code");
+  case Win64EH::UOP_AllocSmall:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0x7f);
+    streamer.emitInt8(inst.Offset / 4);
+    break;
+  case Win64EH::UOP_WideSaveRegMask:
+    assert((inst.Register & ~0x5fff) == 0);
+    lr = (inst.Register >> 14) & 1;
+    w = 0x8000 | (inst.Register & 0x1fff) | (lr << 13);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_SaveSP:
+    assert(inst.Register <= 0x0f);
+    streamer.emitInt8(0xc0 | inst.Register);
+    break;
+  case Win64EH::UOP_SaveRegsR4R7LR:
+    assert(inst.Register >= 4 && inst.Register <= 7);
+    assert(inst.Offset <= 1);
+    streamer.emitInt8(0xd0 | (inst.Register - 4) | (inst.Offset << 2));
+    break;
+  case Win64EH::UOP_WideSaveRegsR4R11LR:
+    assert(inst.Register >= 8 && inst.Register <= 11);
+    assert(inst.Offset <= 1);
+    streamer.emitInt8(0xd8 | (inst.Register - 8) | (inst.Offset << 2));
+    break;
+  case Win64EH::UOP_SaveFRegD8D15:
+    assert(inst.Register >= 8 && inst.Register <= 15);
+    streamer.emitInt8(0xe0 | (inst.Register - 8));
+    break;
+  case Win64EH::UOP_WideAllocMedium:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0x3ff);
+    w = 0xe800 | (inst.Offset / 4);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_SaveRegMask:
+    assert((inst.Register & ~0x40ff) == 0);
+    lr = (inst.Register >> 14) & 1;
+    w = 0xec00 | (inst.Register & 0x0ff) | (lr << 8);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_SaveLR:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0x0f);
+    streamer.emitInt8(0xef);
+    streamer.emitInt8(inst.Offset / 4);
+    break;
+  case Win64EH::UOP_SaveFRegD0D15:
+    assert(inst.Register <= 15);
+    assert(inst.Offset <= 15);
+    assert(inst.Register <= inst.Offset);
+    streamer.emitInt8(0xf5);
+    streamer.emitInt8((inst.Register << 4) | inst.Offset);
+    break;
+  case Win64EH::UOP_SaveFRegD16D31:
+    assert(inst.Register >= 16 && inst.Register <= 31);
+    assert(inst.Offset >= 16 && inst.Offset <= 31);
+    assert(inst.Register <= inst.Offset);
+    streamer.emitInt8(0xf6);
+    streamer.emitInt8(((inst.Register - 16) << 4) | (inst.Offset - 16));
+    break;
+  case Win64EH::UOP_AllocLarge:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0xffff);
+    w = inst.Offset / 4;
+    streamer.emitInt8(0xf7);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_AllocHuge:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0xffffff);
+    w = inst.Offset / 4;
+    streamer.emitInt8(0xf8);
+    streamer.emitInt8((w >> 16) & 0xff);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_WideAllocLarge:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0xffff);
+    w = inst.Offset / 4;
+    streamer.emitInt8(0xf9);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_WideAllocHuge:
+    assert((inst.Offset & 3) == 0);
+    assert(inst.Offset / 4 <= 0xffffff);
+    w = inst.Offset / 4;
+    streamer.emitInt8(0xfa);
+    streamer.emitInt8((w >> 16) & 0xff);
+    streamer.emitInt8((w >> 8) & 0xff);
+    streamer.emitInt8((w >> 0) & 0xff);
+    break;
+  case Win64EH::UOP_Nop:
+    streamer.emitInt8(0xfb);
+    break;
+  case Win64EH::UOP_WideNop:
+    streamer.emitInt8(0xfc);
+    break;
+  case Win64EH::UOP_EndNop:
+    streamer.emitInt8(0xfd);
+    break;
+  case Win64EH::UOP_WideEndNop:
+    streamer.emitInt8(0xfe);
+    break;
+  case Win64EH::UOP_End:
+    streamer.emitInt8(0xff);
+    break;
+  case Win64EH::UOP_Custom:
+    for (i = 3; i > 0; i--)
+      if (inst.Offset & (0xffu << (8 * i)))
+        break;
+    for (; i >= 0; i--)
+      streamer.emitInt8((inst.Offset >> (8 * i)) & 0xff);
+    break;
+  }
+}
+
+// Populate the .xdata section.  The format of .xdata on ARM is documented at
+// https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling
+static void ARMEmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info,
+                              bool TryPacked = true) {
+  // If this UNWIND_INFO already has a symbol, it's already been emitted.
+  if (info->Symbol)
+    return;
+  // If there's no unwind info here (not even a terminating UOP_End), the
+  // unwind info is considered bogus and skipped. If this was done in
+  // response to an explicit .seh_handlerdata, the associated trailing
+  // handler data is left orphaned in the xdata section.
+  if (info->empty()) {
+    info->EmitAttempted = true;
+    return;
+  }
+  if (info->EmitAttempted) {
+    // If we tried to emit unwind info before (due to an explicit
+    // .seh_handlerdata directive), but skipped it (because there was no
+    // valid information to emit at the time), and it later got valid unwind
+    // opcodes, we can't emit it here, because the trailing handler data
+    // was already emitted elsewhere in the xdata section.
+    streamer.getContext().reportError(
+        SMLoc(), "Earlier .seh_handlerdata for " + info->Function->getName() +
+                     " skipped due to no unwind info at the time "
+                     "(.seh_handlerdata too early?), but the function later "
+                     "did get unwind info that can't be emitted");
+    return;
+  }
+
+  MCContext &context = streamer.getContext();
+  MCSymbol *Label = context.createTempSymbol();
+
+  streamer.emitValueToAlignment(4);
+  streamer.emitLabel(Label);
+  info->Symbol = Label;
+
+  Optional<int64_t> RawFuncLength;
+  const MCExpr *FuncLengthExpr = nullptr;
+  if (!info->FuncletOrFuncEnd) {
+    report_fatal_error("FuncletOrFuncEnd not set");
+  } else {
+    // As the size of many thumb2 instructions isn't known until later,
+    // we can't always rely on being able to calculate the absolute
+    // length of the function here. If we can't calculate it, defer it
+    // to a relocation.
+    //
+    // In such a case, we won't know if the function is too long so that
+    // the unwind info would need to be split (but this isn't implemented
+    // anyway).
+    RawFuncLength =
+        GetOptionalAbsDifference(streamer, info->FuncletOrFuncEnd, info->Begin);
+    if (!RawFuncLength)
+      FuncLengthExpr =
+          GetSubDivExpr(streamer, info->FuncletOrFuncEnd, info->Begin, 2);
+  }
+  uint32_t FuncLength = 0;
+  if (RawFuncLength)
+    FuncLength = (uint32_t)*RawFuncLength / 2;
+  if (FuncLength > 0x3FFFF)
+    report_fatal_error("SEH unwind data splitting not yet implemented");
+  uint32_t PrologCodeBytes = ARMCountOfUnwindCodes(info->Instructions);
+  uint32_t TotalCodeBytes = PrologCodeBytes;
+
+  // Process epilogs.
+  MapVector<MCSymbol *, uint32_t> EpilogInfo;
+  // Epilogs processed so far.
+  std::vector<MCSymbol *> AddedEpilogs;
+
+  for (auto &I : info->EpilogMap) {
+    MCSymbol *EpilogStart = I.first;
+    auto &EpilogInstrs = I.second.Instructions;
+    uint32_t CodeBytes = ARMCountOfUnwindCodes(EpilogInstrs);
+
+    MCSymbol *MatchingEpilog =
+        FindMatchingEpilog(EpilogInstrs, AddedEpilogs, info);
+    if (MatchingEpilog) {
+      assert(EpilogInfo.find(MatchingEpilog) != EpilogInfo.end() &&
+             "Duplicate epilog not found");
+      EpilogInfo[EpilogStart] = EpilogInfo.lookup(MatchingEpilog);
+      // Clear the unwind codes in the EpilogMap, so that they don't get output
+      // in the logic below.
+      EpilogInstrs.clear();
+    } else {
+      EpilogInfo[EpilogStart] = TotalCodeBytes;
+      TotalCodeBytes += CodeBytes;
+      AddedEpilogs.push_back(EpilogStart);
+    }
+  }
+
+  // Code Words, Epilog count, F, E, X, Vers, Function Length
+  uint32_t row1 = 0x0;
+  uint32_t CodeWords = TotalCodeBytes / 4;
+  uint32_t CodeWordsMod = TotalCodeBytes % 4;
+  if (CodeWordsMod)
+    CodeWords++;
+  uint32_t EpilogCount = info->EpilogMap.size();
+  bool ExtensionWord = EpilogCount > 31 || CodeWords > 15;
+  if (!ExtensionWord) {
+    row1 |= (EpilogCount & 0x1F) << 23;
+    row1 |= (CodeWords & 0x0F) << 28;
+  }
+  if (info->HandlesExceptions) // X
+    row1 |= 1 << 20;
+  if (info->Fragment) // F
+    row1 |= 1 << 22;
+  row1 |= FuncLength & 0x3FFFF;
+  if (RawFuncLength)
+    streamer.emitInt32(row1);
+  else
+    streamer.emitValue(
+        MCBinaryExpr::createOr(FuncLengthExpr,
+                               MCConstantExpr::create(row1, context), context),
+        4);
+
+  // Extended Code Words, Extended Epilog Count
+  if (ExtensionWord) {
+    // FIXME: We should be able to split unwind info into multiple sections.
+    if (CodeWords > 0xFF || EpilogCount > 0xFFFF)
+      report_fatal_error("SEH unwind data splitting not yet implemented");
+    uint32_t row2 = 0x0;
+    row2 |= (CodeWords & 0xFF) << 16;
+    row2 |= (EpilogCount & 0xFFFF);
+    streamer.emitInt32(row2);
+  }
+
+  // Epilog Start Index, Epilog Start Offset
+  for (auto &I : EpilogInfo) {
+    MCSymbol *EpilogStart = I.first;
+    uint32_t EpilogIndex = I.second;
+
+    Optional<int64_t> MaybeEpilogOffset =
+        GetOptionalAbsDifference(streamer, EpilogStart, info->Begin);
+    const MCExpr *OffsetExpr = nullptr;
+    uint32_t EpilogOffset = 0;
+    if (MaybeEpilogOffset)
+      EpilogOffset = *MaybeEpilogOffset / 2;
+    else
+      OffsetExpr = GetSubDivExpr(streamer, EpilogStart, info->Begin, 2);
+
+    assert(info->EpilogMap.find(EpilogStart) != info->EpilogMap.end());
+    unsigned Condition = info->EpilogMap[EpilogStart].Condition;
+    assert(Condition <= 0xf);
+
+    uint32_t row3 = EpilogOffset;
+    row3 |= Condition << 20;
+    row3 |= (EpilogIndex & 0x3FF) << 24;
+    if (MaybeEpilogOffset)
+      streamer.emitInt32(row3);
+    else
+      streamer.emitValue(
+          MCBinaryExpr::createOr(
+              OffsetExpr, MCConstantExpr::create(row3, context), context),
+          4);
+  }
+
+  // Emit prolog unwind instructions (in reverse order).
+  uint8_t numInst = info->Instructions.size();
+  for (uint8_t c = 0; c < numInst; ++c) {
+    WinEH::Instruction inst = info->Instructions.back();
+    info->Instructions.pop_back();
+    ARMEmitUnwindCode(streamer, inst);
+  }
+
+  // Emit epilog unwind instructions
+  for (auto &I : info->EpilogMap) {
+    auto &EpilogInstrs = I.second.Instructions;
+    for (uint32_t i = 0; i < EpilogInstrs.size(); i++) {
+      WinEH::Instruction inst = EpilogInstrs[i];
+      ARMEmitUnwindCode(streamer, inst);
+    }
+  }
+
+  int32_t BytesMod = CodeWords * 4 - TotalCodeBytes;
+  assert(BytesMod >= 0);
+  for (int i = 0; i < BytesMod; i++)
+    streamer.emitInt8(0xFB);
+
+  if (info->HandlesExceptions)
+    streamer.emitValue(
+        MCSymbolRefExpr::create(info->ExceptionHandler,
+                                MCSymbolRefExpr::VK_COFF_IMGREL32, context),
+        4);
+}
+
+static void ARMEmitRuntimeFunction(MCStreamer &streamer,
   MCContext &context = streamer.getContext();
 
   streamer.emitValueToAlignment(4);
   EmitSymbolRefWithOfs(streamer, info->Begin, info->Begin);
   if (info->PackedInfo)
     streamer.emitInt32(info->PackedInfo);
   else
     streamer.emitValue(
for (const auto &CFI : Streamer.getWinFrameInfos()) {
     WinEH::FrameInfo *Info = CFI.get();
     // ARM64EmitUnwindInfo above clears the info struct, so we can't check
     // empty here. But if a Symbol is set, we should create the corresponding
     // pdata entry.
     if (!Info->Symbol)
       continue;
     MCSection *PData = Streamer.getAssociatedPDataSection(CFI->TextSection);
     Streamer.SwitchSection(PData);
-    ARM64EmitRuntimeFunction(Streamer, Info);
+    ARMEmitRuntimeFunction(Streamer, Info);
   }
 }
 
 void llvm::Win64EH::ARM64UnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer,
                                                        WinEH::FrameInfo *info,
                                                        bool HandlerData) const {
   // Called if there's an .seh_handlerdata directive before the end of the
   // function. This forces writing the xdata record already here - and
   // in this case, the function isn't actually ended already, but the xdata
   // record needs to know the function length. In these cases, if the funclet
   // end hasn't been marked yet, the xdata function length won't cover the
   // whole function, only up to this point.
   if (!info->FuncletOrFuncEnd) {
     Streamer.SwitchSection(info->TextSection);
     info->FuncletOrFuncEnd = Streamer.emitCFILabel();
   }
   // Switch sections (the static function above is meant to be called from
   // here and from Emit().
   MCSection *XData = Streamer.getAssociatedXDataSection(info->TextSection);
   Streamer.SwitchSection(XData);
   ARM64EmitUnwindInfo(Streamer, info, /* TryPacked = */ !HandlerData);
 }
+
+void llvm::Win64EH::ARMUnwindEmitter::Emit(MCStreamer &Streamer) const {
+  // Emit the unwind info structs first.
+  for (const auto &CFI : Streamer.getWinFrameInfos()) {
+    WinEH::FrameInfo *Info = CFI.get();
+    if (Info->empty())
+      continue;
+    MCSection *XData = Streamer.getAssociatedXDataSection(CFI->TextSection);
+    Streamer.SwitchSection(XData);
+    ARMEmitUnwindInfo(Streamer, Info);
+  }
+
+  // Now emit RUNTIME_FUNCTION entries.
+  for (const auto &CFI : Streamer.getWinFrameInfos()) {
+    WinEH::FrameInfo *Info = CFI.get();
+    // ARMEmitUnwindInfo above clears the info struct, so we can't check
+    // empty here. But if a Symbol is set, we should create the corresponding
+    // pdata entry.
+    if (!Info->Symbol)
+      continue;
+    MCSection *PData = Streamer.getAssociatedPDataSection(CFI->TextSection);
+    Streamer.SwitchSection(PData);
+    ARMEmitRuntimeFunction(Streamer, Info);
+  }
+}
+
+void llvm::Win64EH::ARMUnwindEmitter::EmitUnwindInfo(MCStreamer &Streamer,
+                                                     WinEH::FrameInfo *info,
+                                                     bool HandlerData) const {
+  // Called if there's an .seh_handlerdata directive before the end of the
+  // function. This forces writing the xdata record already here - and
+  // in this case, the function isn't actually ended already, but the xdata
+  // record needs to know the function length. In these cases, if the funclet
+  // end hasn't been marked yet, the xdata function length won't cover the
+  // whole function, only up to this point.
+  if (!info->FuncletOrFuncEnd) {
+    Streamer.SwitchSection(info->TextSection);
+    info->FuncletOrFuncEnd = Streamer.emitCFILabel();
+  }
+  // Switch sections (the static function above is meant to be called from
+  // here and from Emit().
+  MCSection *XData = Streamer.getAssociatedXDataSection(info->TextSection);
+  Streamer.SwitchSection(XData);
+  ARMEmitUnwindInfo(Streamer, info, /* TryPacked = */ !HandlerData);
+}

llvm/lib/Target/ARM/AsmParser/ARMAsmParser.cpp

class ARMAsmParser : public MCTargetAsmParser {
 
   int tryParseRegister();
   bool tryParseRegisterWithWriteBack(OperandVector &);
   int tryParseShiftRegister(OperandVector &);
   bool parseRegisterList(OperandVector &, bool EnforceOrder = true,
                          bool AllowRAAC = false);
   bool parseMemory(OperandVector &);
   bool parseOperand(OperandVector &, StringRef Mnemonic);
+  bool parseImmExpr(int64_t &Out);
   bool parsePrefix(ARMMCExpr::VariantKind &RefKind);
   bool parseMemRegOffsetShift(ARM_AM::ShiftOpc &ShiftType,
                               unsigned &ShiftAmount);
   bool parseLiteralValues(unsigned Size, SMLoc L);
   bool parseDirectiveThumb(SMLoc L);
   bool parseDirectiveARM(SMLoc L);
   bool parseDirectiveThumbFunc(SMLoc L);
   bool parseDirectiveCode(SMLoc L);
class ARMAsmParser : public MCTargetAsmParser {
   bool parseDirectiveUnwindRaw(SMLoc L);
   bool parseDirectiveTLSDescSeq(SMLoc L);
   bool parseDirectiveMovSP(SMLoc L);
   bool parseDirectiveObjectArch(SMLoc L);
   bool parseDirectiveArchExtension(SMLoc L);
   bool parseDirectiveAlign(SMLoc L);
   bool parseDirectiveThumbSet(SMLoc L);
 
+  bool parseDirectiveSEHAllocStack(SMLoc L, bool Wide);
+  bool parseDirectiveSEHSaveRegs(SMLoc L, bool Wide);
+  bool parseDirectiveSEHSaveSP(SMLoc L);
+  bool parseDirectiveSEHSaveFRegs(SMLoc L);
+  bool parseDirectiveSEHSaveLR(SMLoc L);
+  bool parseDirectiveSEHPrologEnd(SMLoc L, bool Fragment);
+  bool parseDirectiveSEHNop(SMLoc L, bool Wide);
+  bool parseDirectiveSEHEpilogStart(SMLoc L, bool Condition);
+  bool parseDirectiveSEHEpilogEnd(SMLoc L);
+  bool parseDirectiveSEHCustom(SMLoc L);
+
   bool isMnemonicVPTPredicable(StringRef Mnemonic, StringRef ExtraToken);
   StringRef splitMnemonic(StringRef Mnemonic, StringRef ExtraToken,
                           unsigned &PredicationCode,
                           unsigned &VPTPredicationCode, bool &CarrySetting,
                           unsigned &ProcessorIMod, StringRef &ITMask);
   void getMnemonicAcceptInfo(StringRef Mnemonic, StringRef ExtraToken,
                              StringRef FullInst, bool &CanAcceptCarrySet,
                              bool &CanAcceptPredicationCode,
case AsmToken::Equal: {
     // execute-only: we assume that assembly programmers know what they are
     // doing and allow literal pool creation here
     Operands.push_back(ARMOperand::CreateConstantPoolImm(SubExprVal, S, E));
     return false;
   }
   }
 }
 
+bool ARMAsmParser::parseImmExpr(int64_t &Out) {
+  const MCExpr *Expr = nullptr;
+  SMLoc L = getParser().getTok().getLoc();
+  if (check(getParser().parseExpression(Expr), L, "expected expression"))
+    return true;
+  const MCConstantExpr *Value = dyn_cast_or_null<MCConstantExpr>(Expr);
+  if (check(!Value, L, "expected constant expression"))
+    return true;
+  Out = Value->getValue();
+  return false;
+}
+
 // parsePrefix - Parse ARM 16-bit relocations expression prefix, i.e.
 //  :lower16: and :upper16:.
 bool ARMAsmParser::parsePrefix(ARMMCExpr::VariantKind &RefKind) {
   MCAsmParser &Parser = getParser();
   RefKind = ARMMCExpr::VK_ARM_None;
 
   // consume an optional '#' (GNU compatibility)
   if (getLexer().is(AsmToken::Hash))
else if (!IsMachO && !IsCOFF) {
     else if (IDVal == ".fnstart")
       parseDirectiveFnStart(DirectiveID.getLoc());
     else if (IDVal == ".object_arch")
       parseDirectiveObjectArch(DirectiveID.getLoc());
     else if (IDVal == ".tlsdescseq")
       parseDirectiveTLSDescSeq(DirectiveID.getLoc());
     else
       return true;
+  } else if (IsCOFF) {
+    if (IDVal == ".seh_stackalloc")
+      parseDirectiveSEHAllocStack(DirectiveID.getLoc(), /*Wide=*/false);
+    else if (IDVal == ".seh_stackalloc_w")
+      parseDirectiveSEHAllocStack(DirectiveID.getLoc(), /*Wide=*/true);
+    else if (IDVal == ".seh_save_regs")
+      parseDirectiveSEHSaveRegs(DirectiveID.getLoc(), /*Wide=*/false);
+    else if (IDVal == ".seh_save_regs_w")
+      parseDirectiveSEHSaveRegs(DirectiveID.getLoc(), /*Wide=*/true);
+    else if (IDVal == ".seh_save_sp")
+      parseDirectiveSEHSaveSP(DirectiveID.getLoc());
+    else if (IDVal == ".seh_save_fregs")
+      parseDirectiveSEHSaveFRegs(DirectiveID.getLoc());
+    else if (IDVal == ".seh_save_lr")
+      parseDirectiveSEHSaveLR(DirectiveID.getLoc());
+    else if (IDVal == ".seh_endprologue")
+      parseDirectiveSEHPrologEnd(DirectiveID.getLoc(), /*Fragment=*/false);
+    else if (IDVal == ".seh_endprologue_fragment")
+      parseDirectiveSEHPrologEnd(DirectiveID.getLoc(), /*Fragment=*/true);
+    else if (IDVal == ".seh_nop")
+      parseDirectiveSEHNop(DirectiveID.getLoc(), /*Wide=*/false);
+    else if (IDVal == ".seh_nop_w")
+      parseDirectiveSEHNop(DirectiveID.getLoc(), /*Wide=*/true);
+    else if (IDVal == ".seh_startepilogue")
+      parseDirectiveSEHEpilogStart(DirectiveID.getLoc(), /*Condition=*/false);
+    else if (IDVal == ".seh_startepilogue_cond")
+      parseDirectiveSEHEpilogStart(DirectiveID.getLoc(), /*Condition=*/true);
+    else if (IDVal == ".seh_endepilogue")
+      parseDirectiveSEHEpilogEnd(DirectiveID.getLoc());
+    else if (IDVal == ".seh_custom")
+      parseDirectiveSEHCustom(DirectiveID.getLoc());
+    else
+      return true;
   } else
     return true;
   return false;
 }
 
 /// parseLiteralValues
 ///  ::= .hword expression [, expression]*
 ///  ::= .short expression [, expression]*
bool ARMAsmParser::parseDirectiveThumbSet(SMLoc L) {
   if (MCParserUtils::parseAssignmentExpression(Name, /* allow_redef */ true,
                                                Parser, Sym, Value))
     return true;
 
   getTargetStreamer().emitThumbSet(Sym, Value);
   return false;
 }
 
+/// parseDirectiveSEHAllocStack
+/// ::= .seh_stackalloc
+/// ::= .seh_stackalloc_w
+bool ARMAsmParser::parseDirectiveSEHAllocStack(SMLoc L, bool Wide) {
+  int64_t Size;
+  if (parseImmExpr(Size))
+    return true;
+  getTargetStreamer().emitARMWinCFIAllocStack(Size, Wide);
+  return false;
+}
+
+/// parseDirectiveSEHSaveRegs
+/// ::= .seh_save_regs
+/// ::= .seh_save_regs_w
+bool ARMAsmParser::parseDirectiveSEHSaveRegs(SMLoc L, bool Wide) {
+  SmallVector<std::unique_ptr<MCParsedAsmOperand>, 1> Operands;
+
+  if (parseRegisterList(Operands) ||
+      parseToken(AsmToken::EndOfStatement, "unexpected token in directive"))
+    return true;
+  ARMOperand &Op = (ARMOperand &)*Operands[0];
+  if (!Op.isRegList())
+    return Error(L, ".seh_save_regs{_w} expects GPR registers");
+  const SmallVectorImpl<unsigned> &RegList = Op.getRegList();
+  uint32_t Mask = 0;
+  for (size_t i = 0; i < RegList.size(); ++i) {
+    unsigned Reg = MRI->getEncodingValue(RegList[i]);
+    if (Reg == 15) // pc -> lr
+      Reg = 14;
+    if (Reg == 13)
+      return Error(L, ".seh_save_regs{_w} can't include SP");
+    assert(Reg < 16U && "Register out of range");
+    unsigned Bit = (1u << Reg);
+    Mask |= Bit;
+  }
+  if (!Wide && (Mask & 0x1f00) != 0)
+    return Error(L,
+                 ".seh_save_regs cannot save R8-R12, needs .seh_save_regs_w");
+  getTargetStreamer().emitARMWinCFISaveRegMask(Mask, Wide);
+  return false;
+}
+
+/// parseDirectiveSEHSaveSP
+/// ::= .seh_save_sp
+bool ARMAsmParser::parseDirectiveSEHSaveSP(SMLoc L) {
+  int Reg = tryParseRegister();
+  if (Reg == -1 || !MRI->getRegClass(ARM::GPRRegClassID).contains(Reg))
+    return Error(L, "expected GPR");
+  unsigned Index = MRI->getEncodingValue(Reg);
+  if (Index > 14 || Index == 13)
+    return Error(L, "invalid register for .seh_save_sp");
+  getTargetStreamer().emitARMWinCFISaveSP(Index);
+  return false;
+}
+
+/// parseDirectiveSEHSaveFRegs
+/// ::= .seh_save_fregs
+bool ARMAsmParser::parseDirectiveSEHSaveFRegs(SMLoc L) {
+  SmallVector<std::unique_ptr<MCParsedAsmOperand>, 1> Operands;
+
+  if (parseRegisterList(Operands) ||
+      parseToken(AsmToken::EndOfStatement, "unexpected token in directive"))
+    return true;
+  ARMOperand &Op = (ARMOperand &)*Operands[0];
+  if (!Op.isDPRRegList())
+    return Error(L, ".seh_save_fregs expects DPR registers");
+  const SmallVectorImpl<unsigned> &RegList = Op.getRegList();
+  uint32_t Mask = 0;
+  for (size_t i = 0; i < RegList.size(); ++i) {
+    unsigned Reg = MRI->getEncodingValue(RegList[i]);
+    assert(Reg < 32U && "Register out of range");
+    unsigned Bit = (1u << Reg);
+    Mask |= Bit;
+  }
+
+  if (Mask == 0)
+    return Error(L, ".seh_save_fregs missing registers");
+
+  unsigned First = 0;
+  while ((Mask & 1) == 0) {
+    First++;
+    Mask >>= 1;
+  }
+  if (((Mask + 1) & Mask) != 0)
+    return Error(L,
+                 ".seh_save_fregs must take a contiguous range of registers");
+  unsigned Last = First;
+  while ((Mask & 2) != 0) {
+    Last++;
+    Mask >>= 1;
+  }
+  if (First < 16 && Last >= 16)
+    return Error(L, ".seh_save_fregs must be all d0-d15 or d16-d31");
+  getTargetStreamer().emitARMWinCFISaveFRegs(First, Last);
+  return false;
+}
+
+/// parseDirectiveSEHSaveLR
+/// ::= .seh_save_lr
+bool ARMAsmParser::parseDirectiveSEHSaveLR(SMLoc L) {
+  int64_t Offset;
+  if (parseImmExpr(Offset))
+    return true;
+  getTargetStreamer().emitARMWinCFISaveLR(Offset);
+  return false;
+}
+
+/// parseDirectiveSEHPrologEnd
+/// ::= .seh_endprologue
+/// ::= .seh_endprologue_fragment
+bool ARMAsmParser::parseDirectiveSEHPrologEnd(SMLoc L, bool Fragment) {
+  getTargetStreamer().emitARMWinCFIPrologEnd(Fragment);
+  return false;
+}
+
+/// parseDirectiveSEHNop
+/// ::= .seh_nop
+/// ::= .seh_nop_w
+bool ARMAsmParser::parseDirectiveSEHNop(SMLoc L, bool Wide) {
+  getTargetStreamer().emitARMWinCFINop(Wide);
+  return false;
+}
+
+/// parseDirectiveSEHEpilogStart
+/// ::= .seh_startepilogue
+/// ::= .seh_startepilogue_cond
+bool ARMAsmParser::parseDirectiveSEHEpilogStart(SMLoc L, bool Condition) {
+  unsigned CC = ARMCC::AL;
+  if (Condition) {
+    MCAsmParser &Parser = getParser();
+    SMLoc S = Parser.getTok().getLoc();
+    const AsmToken &Tok = Parser.getTok();
+    if (!Tok.is(AsmToken::Identifier))
+      return Error(S, ".seh_startepilogue_cond missing condition");
+    CC = ARMCondCodeFromString(Tok.getString());
+    if (CC == ~0U)
+      return Error(S, "invalid condition");
+    Parser.Lex(); // Eat the token.
+  }
+
+  getTargetStreamer().emitARMWinCFIEpilogStart(CC);
+  return false;
+}
+
+/// parseDirectiveSEHEpilogEnd
+/// ::= .seh_endepilogue
+bool ARMAsmParser::parseDirectiveSEHEpilogEnd(SMLoc L) {
+  getTargetStreamer().emitARMWinCFIEpilogEnd();
+  return false;
+}
+
+/// parseDirectiveSEHCustom
+/// ::= .seh_custom
+bool ARMAsmParser::parseDirectiveSEHCustom(SMLoc L) {
+  unsigned Opcode = 0;
+  do {
+    int64_t Byte;
+    if (parseImmExpr(Byte))
+      return true;
+    if (Byte > 0xff || Byte < 0)
+      return Error(L, "Invalid byte value in .seh_custom");
+    if (Opcode > 0x00ffffff)
+      return Error(L, "Too many bytes in .seh_custom");
+    // Store the bytes as one big endian number in Opcode. In a multi byte
+    // opcode sequence, the first byte can't be zero.
+    Opcode = (Opcode << 8) | Byte;
+  } while (parseOptionalToken(AsmToken::Comma));
+  getTargetStreamer().emitARMWinCFICustom(Opcode);
+  return false;
+}
+
 /// Force static initialization.
 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeARMAsmParser() {
   RegisterMCAsmParser<ARMAsmParser> X(getTheARMLETarget());
   RegisterMCAsmParser<ARMAsmParser> Y(getTheARMBETarget());
   RegisterMCAsmParser<ARMAsmParser> A(getTheThumbLETarget());
   RegisterMCAsmParser<ARMAsmParser> B(getTheThumbBETarget());
 }
 

llvm/lib/Target/ARM/MCTargetDesc/ARMELFStreamer.cpp

class ARMTargetAsmStreamer : public ARMTargetStreamer {
   void emitObjectArch(ARM::ArchKind Arch) override;
   void emitFPU(unsigned FPU) override;
   void emitInst(uint32_t Inst, char Suffix = '\0') override;
   void finishAttributeSection() override;
 
   void AnnotateTLSDescriptorSequence(const MCSymbolRefExpr *SRE) override;
   void emitThumbSet(MCSymbol *Symbol, const MCExpr *Value) override;
 
+  void emitARMWinCFIAllocStack(unsigned Size, bool Wide) override;
+  void emitARMWinCFISaveRegMask(unsigned Mask, bool Wide) override;
+  void emitARMWinCFISaveSP(unsigned Reg) override;
+  void emitARMWinCFISaveFRegs(unsigned First, unsigned Last) override;
+  void emitARMWinCFISaveLR(unsigned Offset) override;
+  void emitARMWinCFIPrologEnd(bool Fragment) override;
+  void emitARMWinCFINop(bool Wide) override;
+  void emitARMWinCFIEpilogStart(unsigned Condition) override;
+  void emitARMWinCFIEpilogEnd() override;
+  void emitARMWinCFICustom(unsigned Opcode) override;
+
 public:
   ARMTargetAsmStreamer(MCStreamer &S, formatted_raw_ostream &OS,
                        MCInstPrinter &InstPrinter, bool VerboseAsm);
 };
 
 ARMTargetAsmStreamer::ARMTargetAsmStreamer(MCStreamer &S,
                                            formatted_raw_ostream &OS,
                                            MCInstPrinter &InstPrinter,

 void ARMTargetAsmStreamer::emitUnwindRaw(int64_t Offset,
                                       const SmallVectorImpl<uint8_t> &Opcodes) {
   OS << "\t.unwind_raw " << Offset;
   for (uint8_t Opcode : Opcodes)
     OS << ", 0x" << Twine::utohexstr(Opcode);
   OS << '\n';
 }
 
+void ARMTargetAsmStreamer::emitARMWinCFIAllocStack(unsigned Size, bool Wide) {
+  if (Wide)
+    OS << "\t.seh_stackalloc_w\t" << Size << "\n";
+  else
+    OS << "\t.seh_stackalloc\t" << Size << "\n";
+}
+
+static void printRegs(formatted_raw_ostream &OS, ListSeparator &LS, int First,
+                      int Last) {
+  if (First != Last)
+    OS << LS << "r" << First << "-r" << Last;
+  else
+    OS << LS << "r" << First;
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFISaveRegMask(unsigned Mask, bool Wide) {
+  if (Wide)
+    OS << "\t.seh_save_regs_w\t";
+  else
+    OS << "\t.seh_save_regs\t";
+  ListSeparator LS;
+  int First = -1;
+  OS << "{";
+  for (int I = 0; I <= 12; I++) {
+    if (Mask & (1 << I)) {
+      if (First < 0)
+        First = I;
+    } else {
+      if (First >= 0) {
+        printRegs(OS, LS, First, I - 1);
+        First = -1;
+      }
+    }
+  }
+  if (First >= 0)
+    printRegs(OS, LS, First, 12);
+  if (Mask & (1 << 14))
+    OS << LS << "lr";
+  OS << "}\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFISaveSP(unsigned Reg) {
+  OS << "\t.seh_save_sp\tr" << Reg << "\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFISaveFRegs(unsigned First,
+                                                  unsigned Last) {
+  if (First != Last)
+    OS << "\t.seh_save_fregs\t{d" << First << "-d" << Last << "}\n";
+  else
+    OS << "\t.seh_save_fregs\t{d" << First << "}\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFISaveLR(unsigned Offset) {
+  OS << "\t.seh_save_lr\t" << Offset << "\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFIPrologEnd(bool Fragment) {
+  if (Fragment)
+    OS << "\t.seh_endprologue_fragment\n";
+  else
+    OS << "\t.seh_endprologue\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFINop(bool Wide) {
+  if (Wide)
+    OS << "\t.seh_nop_w\n";
+  else
+    OS << "\t.seh_nop\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFIEpilogStart(unsigned Condition) {
+  if (Condition == ARMCC::AL)
+    OS << "\t.seh_startepilogue\n";
+  else
+    OS << "\t.seh_startepilogue_cond\t"
+       << ARMCondCodeToString(static_cast<ARMCC::CondCodes>(Condition)) << "\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFIEpilogEnd() {
+  OS << "\t.seh_endepilogue\n";
+}
+
+void ARMTargetAsmStreamer::emitARMWinCFICustom(unsigned Opcode) {
+  int I;
+  for (I = 3; I > 0; I--)
+    if (Opcode & (0xffu << (8 * I)))
+      break;
+  ListSeparator LS;
+  OS << "\t.seh_custom\t";
+  for (; I >= 0; I--)
+    OS << LS << ((Opcode >> (8 * I)) & 0xff);
+  OS << "\n";
+}
+
 class ARMTargetELFStreamer : public ARMTargetStreamer {
 private:
   StringRef CurrentVendor;
   unsigned FPU = ARM::FK_INVALID;
   ARM::ArchKind Arch = ARM::ArchKind::INVALID;
   ARM::ArchKind EmittedArch = ARM::ArchKind::INVALID;
 
   MCSection *AttributeSection = nullptr;
MCTargetStreamer *createARMTargetAsmStreamer(MCStreamer &S,
                                              bool isVerboseAsm) {
   return new ARMTargetAsmStreamer(S, OS, *InstPrint, isVerboseAsm);
 }
 
 MCTargetStreamer *createARMNullTargetStreamer(MCStreamer &S) {
   return new ARMTargetStreamer(S);
 }
 
-MCTargetStreamer *createARMObjectTargetStreamer(MCStreamer &S,
-                                                const MCSubtargetInfo &STI) {
-  const Triple &TT = STI.getTargetTriple();
-  if (TT.isOSBinFormatELF())
-    return new ARMTargetELFStreamer(S);
-  return new ARMTargetStreamer(S);
+MCTargetStreamer *createARMObjectTargetELFStreamer(MCStreamer &S) {
 }
 
 MCELFStreamer *createARMELFStreamer(MCContext &Context,
                                     std::unique_ptr<MCAsmBackend> TAB,
                                     std::unique_ptr<MCObjectWriter> OW,
                                     std::unique_ptr<MCCodeEmitter> Emitter,
                                     bool RelaxAll, bool IsThumb,
                                     bool IsAndroid) {

llvm/lib/Target/ARM/MCTargetDesc/ARMMCAsmInfo.cpp

 }
 
 void ARMCOFFMCAsmInfoMicrosoft::anchor() { }
 
 ARMCOFFMCAsmInfoMicrosoft::ARMCOFFMCAsmInfoMicrosoft() {
   AlignmentIsInBytes = false;
   SupportsDebugInformation = true;
   ExceptionsType = ExceptionHandling::WinEH;
+  WinEHEncodingType = WinEH::EncodingType::Itanium;
   PrivateGlobalPrefix = "$M";
   PrivateLabelPrefix = "$M";
   CommentString = "@";
 
   // Conditional Thumb 4-byte instructions can have an implicit IT.
   MaxInstLength = 6;
 }
 

llvm/lib/Target/ARM/MCTargetDesc/ARMMCTargetDesc.h

 
 MCTargetStreamer *createARMNullTargetStreamer(MCStreamer &S);
 MCTargetStreamer *createARMTargetAsmStreamer(MCStreamer &S,
                                              formatted_raw_ostream &OS,
                                              MCInstPrinter *InstPrint,
                                              bool isVerboseAsm);
 MCTargetStreamer *createARMObjectTargetStreamer(MCStreamer &S,
                                                 const MCSubtargetInfo &STI);
+MCTargetStreamer *createARMObjectTargetELFStreamer(MCStreamer &S);
+MCTargetStreamer *createARMObjectTargetWinCOFFStreamer(MCStreamer &S);
 
 MCCodeEmitter *createARMLEMCCodeEmitter(const MCInstrInfo &MCII,
                                         MCContext &Ctx);
 
 MCCodeEmitter *createARMBEMCCodeEmitter(const MCInstrInfo &MCII,
                                         MCContext &Ctx);
 
 MCAsmBackend *createARMLEAsmBackend(const Target &T, const MCSubtargetInfo &STI,

llvm/lib/Target/ARM/MCTargetDesc/ARMTargetStreamer.cpp

 void ARMTargetStreamer::emitArchExtension(uint64_t ArchExt) {}
 void ARMTargetStreamer::emitObjectArch(ARM::ArchKind Arch) {}
 void ARMTargetStreamer::emitFPU(unsigned FPU) {}
 void ARMTargetStreamer::finishAttributeSection() {}
 void
 ARMTargetStreamer::AnnotateTLSDescriptorSequence(const MCSymbolRefExpr *SRE) {}
 void ARMTargetStreamer::emitThumbSet(MCSymbol *Symbol, const MCExpr *Value) {}
 
+void ARMTargetStreamer::emitARMWinCFIAllocStack(unsigned Size, bool Wide) {}
+void ARMTargetStreamer::emitARMWinCFISaveRegMask(unsigned Mask, bool Wide) {}
+void ARMTargetStreamer::emitARMWinCFISaveSP(unsigned Reg) {}
+void ARMTargetStreamer::emitARMWinCFISaveFRegs(unsigned First, unsigned Last) {}
+void ARMTargetStreamer::emitARMWinCFISaveLR(unsigned Offset) {}
+void ARMTargetStreamer::emitARMWinCFINop(bool Wide) {}
+void ARMTargetStreamer::emitARMWinCFIPrologEnd(bool Fragment) {}
+void ARMTargetStreamer::emitARMWinCFIEpilogStart(unsigned Condition) {}
+void ARMTargetStreamer::emitARMWinCFIEpilogEnd() {}
+void ARMTargetStreamer::emitARMWinCFICustom(unsigned Opcode) {}
+
 static ARMBuildAttrs::CPUArch getArchForCPU(const MCSubtargetInfo &STI) {
   if (STI.getCPU() == "xscale")
     return ARMBuildAttrs::v5TEJ;
 
   if (STI.hasFeature(ARM::HasV9_0aOps))
     return ARMBuildAttrs::v9_A;
   else if (STI.hasFeature(ARM::HasV8Ops)) {
     if (STI.hasFeature(ARM::FeatureRClass))
else if (STI.hasFeature(ARM::FeatureVirtualization))
     emitAttribute(ARMBuildAttrs::Virtualization_use,
                   ARMBuildAttrs::AllowVirtualization);
 
   if (STI.hasFeature(ARM::FeaturePACBTI)) {
     emitAttribute(ARMBuildAttrs::PAC_extension, ARMBuildAttrs::AllowPAC);
     emitAttribute(ARMBuildAttrs::BTI_extension, ARMBuildAttrs::AllowBTI);
   }
 }
+
+MCTargetStreamer *
+llvm::createARMObjectTargetStreamer(MCStreamer &S, const MCSubtargetInfo &STI) {
+  const Triple &TT = STI.getTargetTriple();
+  if (TT.isOSBinFormatELF())
+    return createARMObjectTargetELFStreamer(S);
+  if (TT.isOSBinFormatCOFF())
+    return createARMObjectTargetWinCOFFStreamer(S);
+  return new ARMTargetStreamer(S);
+}

llvm/lib/Target/ARM/MCTargetDesc/ARMWinCOFFStreamer.cpp

 //===-- ARMWinCOFFStreamer.cpp - ARM Target WinCOFF Streamer ----*- 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
 //
 //===----------------------------------------------------------------------===//
 
 #include "ARMMCTargetDesc.h"
+#include "Utils/ARMBaseInfo.h"
 #include "llvm/MC/MCAsmBackend.h"
 #include "llvm/MC/MCAssembler.h"
 #include "llvm/MC/MCCodeEmitter.h"
 #include "llvm/MC/MCObjectWriter.h"
+#include "llvm/MC/MCWin64EH.h"
 #include "llvm/MC/MCWinCOFFStreamer.h"
 
 using namespace llvm;
 
 namespace {
 class ARMWinCOFFStreamer : public MCWinCOFFStreamer {
+  Win64EH::ARMUnwindEmitter EHStreamer;
+
 public:
   ARMWinCOFFStreamer(MCContext &C, std::unique_ptr<MCAsmBackend> AB,
                      std::unique_ptr<MCCodeEmitter> CE,
                      std::unique_ptr<MCObjectWriter> OW)
       : MCWinCOFFStreamer(C, std::move(AB), std::move(CE), std::move(OW)) {}
 
+  void emitWinEHHandlerData(SMLoc Loc) override;
+  void emitWindowsUnwindTables() override;
+  void emitWindowsUnwindTables(WinEH::FrameInfo *Frame) override;
+
   void emitThumbFunc(MCSymbol *Symbol) override;
   void finishImpl() override;
 };
 
+void ARMWinCOFFStreamer::emitWinEHHandlerData(SMLoc Loc) {
+  MCStreamer::emitWinEHHandlerData(Loc);
+
+  // We have to emit the unwind info now, because this directive
+  // actually switches to the .xdata section!
+  EHStreamer.EmitUnwindInfo(*this, getCurrentWinFrameInfo(),
+                            /* HandlerData = */ true);
+}
+
+void ARMWinCOFFStreamer::emitWindowsUnwindTables(WinEH::FrameInfo *Frame) {
+  EHStreamer.EmitUnwindInfo(*this, Frame, /* HandlerData = */ false);
+}
+
+void ARMWinCOFFStreamer::emitWindowsUnwindTables() {
+  if (!getNumWinFrameInfos())
+    return;
+  EHStreamer.Emit(*this);
+}
+
 void ARMWinCOFFStreamer::emitThumbFunc(MCSymbol *Symbol) {
   getAssembler().setIsThumbFunc(Symbol);
 }
 
 void ARMWinCOFFStreamer::finishImpl() {
   emitFrames(nullptr);
+  emitWindowsUnwindTables();
 
   MCWinCOFFStreamer::finishImpl();
 }
 }
 
 MCStreamer *llvm::createARMWinCOFFStreamer(
     MCContext &Context, std::unique_ptr<MCAsmBackend> &&MAB,
     std::unique_ptr<MCObjectWriter> &&OW,
     std::unique_ptr<MCCodeEmitter> &&Emitter, bool RelaxAll,
     bool IncrementalLinkerCompatible) {
   auto *S = new ARMWinCOFFStreamer(Context, std::move(MAB), std::move(Emitter),
                                    std::move(OW));
   S->getAssembler().setIncrementalLinkerCompatible(IncrementalLinkerCompatible);
   return S;
 }
 
+namespace {
+class ARMTargetWinCOFFStreamer : public llvm::ARMTargetStreamer {
+private:
+  // True if we are processing SEH directives in an epilogue.
+  bool InEpilogCFI = false;
+
+  // Symbol of the current epilog for which we are processing SEH directives.
+  MCSymbol *CurrentEpilog = nullptr;
+
+public:
+  ARMTargetWinCOFFStreamer(llvm::MCStreamer &S) : ARMTargetStreamer(S) {}
+
+  // The unwind codes on ARM Windows are documented at
+  // https://docs.microsoft.com/en-us/cpp/build/arm-exception-handling
+  void emitARMWinCFIAllocStack(unsigned Size, bool Wide) override;
+  void emitARMWinCFISaveRegMask(unsigned Mask, bool Wide) override;
+  void emitARMWinCFISaveSP(unsigned Reg) override;
+  void emitARMWinCFISaveFRegs(unsigned First, unsigned Last) override;
+  void emitARMWinCFISaveLR(unsigned Offset) override;
+  void emitARMWinCFIPrologEnd(bool Fragment) override;
+  void emitARMWinCFINop(bool Wide) override;
+  void emitARMWinCFIEpilogStart(unsigned Condition) override;
+  void emitARMWinCFIEpilogEnd() override;
+  void emitARMWinCFICustom(unsigned Opcode) override;
+
+private:
+  void emitARMWinUnwindCode(unsigned UnwindCode, int Reg, int Offset);
+};
+
+// Helper function to common out unwind code setup for those codes that can
+// belong to both prolog and epilog.
+void ARMTargetWinCOFFStreamer::emitARMWinUnwindCode(unsigned UnwindCode,
+                                                    int Reg, int Offset) {
+  auto &S = getStreamer();
+  WinEH::FrameInfo *CurFrame = S.EnsureValidWinFrameInfo(SMLoc());
+  if (!CurFrame)
+    return;
+  MCSymbol *Label = S.emitCFILabel();
+  auto Inst = WinEH::Instruction(UnwindCode, Label, Reg, Offset);
+  if (InEpilogCFI)
+    CurFrame->EpilogMap[CurrentEpilog].Instructions.push_back(Inst);
+  else
+    CurFrame->Instructions.push_back(Inst);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFIAllocStack(unsigned Size,
+                                                       bool Wide) {
+  unsigned Op = Win64EH::UOP_AllocSmall;
+  if (!Wide) {
+    if (Size / 4 > 0xffff)
+      Op = Win64EH::UOP_AllocHuge;
+    else if (Size / 4 > 0x7f)
+      Op = Win64EH::UOP_AllocLarge;
+  } else {
+    Op = Win64EH::UOP_WideAllocMedium;
+    if (Size / 4 > 0xffff)
+      Op = Win64EH::UOP_WideAllocHuge;
+    else if (Size / 4 > 0x3ff)
+      Op = Win64EH::UOP_WideAllocLarge;
+  }
+  emitARMWinUnwindCode(Op, -1, Size);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFISaveRegMask(unsigned Mask,
+                                                        bool Wide) {
+  assert(Mask != 0);
+  int Lr = (Mask & 0x4000) ? 1 : 0;
+  Mask &= ~0x4000;
+  if (Wide)
+    assert((Mask & ~0x1fff) == 0);
+  else
+    assert((Mask & ~0x00ff) == 0);
+  if (Mask && ((Mask + (1 << 4)) & Mask) == 0) {
+    if (Wide && (Mask & 0x1000) == 0 && (Mask & 0xff) == 0xf0) {
+      // One continuous range from r4 to r8-r11
+      for (int I = 11; I >= 8; I--) {
+        if (Mask & (1 << I)) {
+          emitARMWinUnwindCode(Win64EH::UOP_WideSaveRegsR4R11LR, I, Lr);
+          return;
+        }
+      }
+      // If it actually was from r4 to r4-r7, continue below.
+    } else if (!Wide) {
+      // One continuous range from r4 to r4-r7
+      for (int I = 7; I >= 4; I--) {
+        if (Mask & (1 << I)) {
+          emitARMWinUnwindCode(Win64EH::UOP_SaveRegsR4R7LR, I, Lr);
+          return;
+        }
+      }
+      llvm_unreachable("logic error");
+    }
+  }
+  Mask |= Lr << 14;
+  if (Wide)
+    emitARMWinUnwindCode(Win64EH::UOP_WideSaveRegMask, Mask, 0);
+  else
+    emitARMWinUnwindCode(Win64EH::UOP_SaveRegMask, Mask, 0);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFISaveSP(unsigned Reg) {
+  emitARMWinUnwindCode(Win64EH::UOP_SaveSP, Reg, 0);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFISaveFRegs(unsigned First,
+                                                      unsigned Last) {
+  assert(First <= Last);
+  assert(First >= 16 || Last < 16);
+  assert(First <= 31 && Last <= 31);
+  if (First == 8)
+    emitARMWinUnwindCode(Win64EH::UOP_SaveFRegD8D15, Last, 0);
+  else if (First <= 15)
+    emitARMWinUnwindCode(Win64EH::UOP_SaveFRegD0D15, First, Last);
+  else
+    emitARMWinUnwindCode(Win64EH::UOP_SaveFRegD16D31, First, Last);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFISaveLR(unsigned Offset) {
+  emitARMWinUnwindCode(Win64EH::UOP_SaveLR, 0, Offset);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFINop(bool Wide) {
+  if (Wide)
+    emitARMWinUnwindCode(Win64EH::UOP_WideNop, -1, 0);
+  else
+    emitARMWinUnwindCode(Win64EH::UOP_Nop, -1, 0);
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFIPrologEnd(bool Fragment) {
+  auto &S = getStreamer();
+  WinEH::FrameInfo *CurFrame = S.EnsureValidWinFrameInfo(SMLoc());
+  if (!CurFrame)
+    return;
+
+  MCSymbol *Label = S.emitCFILabel();
+  CurFrame->PrologEnd = Label;
+  WinEH::Instruction Inst =
+      WinEH::Instruction(Win64EH::UOP_End, /*Label=*/nullptr, -1, 0);
+  auto it = CurFrame->Instructions.begin();
+  CurFrame->Instructions.insert(it, Inst);
+  CurFrame->Fragment = Fragment;
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFIEpilogStart(unsigned Condition) {
+  auto &S = getStreamer();
+  WinEH::FrameInfo *CurFrame = S.EnsureValidWinFrameInfo(SMLoc());
+  if (!CurFrame)
+    return;
+
+  InEpilogCFI = true;
+  CurrentEpilog = S.emitCFILabel();
+  CurFrame->EpilogMap[CurrentEpilog].Condition = Condition;
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFIEpilogEnd() {
+  auto &S = getStreamer();
+  WinEH::FrameInfo *CurFrame = S.EnsureValidWinFrameInfo(SMLoc());
+  if (!CurFrame)
+    return;
+
+  std::vector<WinEH::Instruction> &Epilog =
+      CurFrame->EpilogMap[CurrentEpilog].Instructions;
+
+  unsigned UnwindCode = Win64EH::UOP_End;
+  if (!Epilog.empty()) {
+    WinEH::Instruction EndInstr = Epilog.back();
+    if (EndInstr.Operation == Win64EH::UOP_Nop) {
+      UnwindCode = Win64EH::UOP_EndNop;
+      Epilog.pop_back();
+    } else if (EndInstr.Operation == Win64EH::UOP_WideNop) {
+      UnwindCode = Win64EH::UOP_WideEndNop;
+      Epilog.pop_back();
+    }
+  }
+
+  InEpilogCFI = false;
+  WinEH::Instruction Inst = WinEH::Instruction(UnwindCode, nullptr, -1, 0);
+  CurFrame->EpilogMap[CurrentEpilog].Instructions.push_back(Inst);
+  CurrentEpilog = nullptr;
+}
+
+void ARMTargetWinCOFFStreamer::emitARMWinCFICustom(unsigned Opcode) {
+  emitARMWinUnwindCode(Win64EH::UOP_Custom, 0, Opcode);
+}
+
+} // end anonymous namespace
+
+MCTargetStreamer *llvm::createARMObjectTargetWinCOFFStreamer(MCStreamer &S) {
+  return new ARMTargetWinCOFFStreamer(S);
+}

llvm/test/MC/ARM/seh.s

+// This test checks that the SEH directives emit the correct unwind data.
+
+// RUN: llvm-mc -triple thumbv7-pc-win32 -filetype=obj %s | llvm-readobj -S -r -u - | FileCheck %s
+
+// Check that the output assembler directives also can be parsed, and
+// that they produce equivalent output:
+
+// RUN: llvm-mc -triple thumbv7-pc-win32 -filetype=asm %s | llvm-mc -triple thumbv7-pc-win32 -filetype=obj - | llvm-readobj -S -r -u - | FileCheck %s
+
+// CHECK:      Sections [
+// CHECK:        Section {
+// CHECK:          Name: .text
+// CHECK:          RelocationCount: 1
+// CHECK:          Characteristics [
+// CHECK-NEXT:       ALIGN_4BYTES
+// CHECK-NEXT:       CNT_CODE
+// CHECK-NEXT:       MEM_16BIT
+// CHECK-NEXT:       MEM_EXECUTE
+// CHECK-NEXT:       MEM_PURGEABLE
+// CHECK-NEXT:       MEM_READ
+// CHECK-NEXT:     ]
+// CHECK-NEXT:   }
+// CHECK:        Section {
+// CHECK:          Name: .xdata
+// CHECK:          RawDataSize: 120
+// CHECK:          RelocationCount: 1
+// CHECK:          Characteristics [
+// CHECK-NEXT:       ALIGN_4BYTES
+// CHECK-NEXT:       CNT_INITIALIZED_DATA
+// CHECK-NEXT:       MEM_READ
+// CHECK-NEXT:     ]
+// CHECK-NEXT:   }
+// CHECK:        Section {
+// CHECK:          Name: .pdata
+// CHECK:          RelocationCount: 10
+// CHECK:          Characteristics [
+// CHECK-NEXT:       ALIGN_4BYTES
+// CHECK-NEXT:       CNT_INITIALIZED_DATA
+// CHECK-NEXT:       MEM_READ
+// CHECK-NEXT:     ]
+// CHECK-NEXT:   }
+// CHECK-NEXT: ]
+
+// CHECK-NEXT: Relocations [
+// CHECK-NEXT:   Section (1) .text {
+// CHECK-NEXT:     0x5C IMAGE_REL_ARM_BRANCH24T tailcall
+// CHECK-NEXT:   }
+// CHECK-NEXT:   Section (4) .xdata {
+// CHECK-NEXT:     0x38 IMAGE_REL_ARM_ADDR32NB __C_specific_handler
+// CHECK-NEXT:   }
+// CHECK-NEXT:   Section (5) .pdata {
+// CHECK-NEXT:     0x0 IMAGE_REL_ARM_ADDR32NB .text
+// CHECK-NEXT:     0x4 IMAGE_REL_ARM_ADDR32NB .xdata
+// CHECK-NEXT:     0x8 IMAGE_REL_ARM_ADDR32NB .text
+// CHECK-NEXT:     0xC IMAGE_REL_ARM_ADDR32NB .xdata
+// CHECK-NEXT:     0x10 IMAGE_REL_ARM_ADDR32NB .text
+// CHECK-NEXT:     0x14 IMAGE_REL_ARM_ADDR32NB .xdata
+// CHECK-NEXT:     0x18 IMAGE_REL_ARM_ADDR32NB .text
+// CHECK-NEXT:     0x1C IMAGE_REL_ARM_ADDR32NB .xdata
+// CHECK-NEXT:     0x20 IMAGE_REL_ARM_ADDR32NB .text
+// CHECK-NEXT:     0x24 IMAGE_REL_ARM_ADDR32NB .xdata
+// CHECK-NEXT:   }
+// CHECK-NEXT: ]
+
+// CHECK-NEXT: UnwindInformation [
+// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: func
+// CHECK-NEXT:     ExceptionRecord: .xdata
+// CHECK-NEXT:     ExceptionData {
+// CHECK-NEXT:       FunctionLength: 86
+// CHECK:            Fragment: No
+// CHECK:            Prologue [
+// CHECK-NEXT:         0xed 0xf8           ; push {r3-r7, lr}
+// CHECK-NEXT:         0xf6 0x27           ; vpush {d18-d23}
+// CHECK-NEXT:         0xf5 0x7e           ; vpush {d7-d14}
+// CHECK-NEXT:         0xfb                ; nop
+// CHECK-NEXT:         0xce                ; mov r14, sp
+// CHECK-NEXT:         0xe3                ; vpush {d8-d11}
+// CHECK-NEXT:         0xe6                ; vpush {d8-d14}
+// CHECK-NEXT:         0xed 0xf8           ; push {r3-r7, lr}
+// CHECK-NEXT:         0xbd 0x50           ; push.w {r4, r6, r8, r10-r12, lr}
+// CHECK-NEXT:         0xd7                ; push {r4-r7, lr}
+// CHECK-NEXT:         0xdd                ; push.w {r4-r9, lr}
+// CHECK-NEXT:         0xfa 0x01 0x00 0x00 ; sub.w sp, sp, #(65536 * 4)
+// CHECK-NEXT:         0xfc                ; nop.w
+// CHECK-NEXT:         0xfc                ; nop.w
+// CHECK-NEXT:         0xf9 0x04 0x00      ; sub.w sp, sp, #(1024 * 4)
+// CHECK-NEXT:         0xe8 0x80           ; sub.w sp, #(128 * 4)
+// CHECK-NEXT:         0xe8 0x80           ; sub.w sp, #(128 * 4)
+// CHECK-NEXT:         0x06                ; sub sp, #(6 * 4)
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       EpilogueScopes [
+// CHECK-NEXT:         EpilogueScope {
+// CHECK-NEXT:           StartOffset: 31
+// CHECK-NEXT:           Condition: 14
+// CHECK-NEXT:           EpilogueStartIndex: 31
+// CHECK-NEXT:           Opcodes [
+// CHECK-NEXT:             0xfc                ; nop.w
+// CHECK-NEXT:             0xf7 0x00 0x80      ; add sp, sp, #(128 * 4)
+// CHECK-NEXT:             0xfc                ; nop.w
+// CHECK-NEXT:             0xfc                ; nop.w
+// CHECK-NEXT:             0xf8 0x01 0x00 0x00 ; add sp, sp, #(65536 * 4)
+// CHECK-NEXT:             0x06                ; add sp, #(6 * 4)
+// CHECK-NEXT:             0xef 0x04           ; ldr.w lr, [sp], #16
+// CHECK-NEXT:             0xfd                ; bx <reg>
+// CHECK-NEXT:           ]
+// CHECK-NEXT:         }
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       ExceptionHandler [
+// CHECK-NEXT:         Routine: __C_specific_handler
+// CHECK-NEXT:         Parameter: 0x0
+// CHECK-NEXT:       ]
+// CHECK-NEXT:     }
+// CHECK-NEXT:   }
+// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: func2
+// CHECK:            Prologue [
+// CHECK-NEXT:         0xd3                ; push {r4-r7}
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       EpilogueScopes [
+// CHECK-NEXT:         EpilogueScope {
+// CHECK:                Opcodes [
+// CHECK-NEXT:             0xd2                ; pop {r4-r6}
+// CHECK-NEXT:             0xfe                ; b.w <target>
+// CHECK-NEXT:           ]
+// CHECK-NEXT:         }
+// CHECK-NEXT:       ]
+// CHECK-NEXT:     }
+// CHECK-NEXT:   }
+// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: func3
+// CHECK:            FunctionLength: 8
+// CHECK:            Prologue [
+// CHECK-NEXT:         0xd5                ; push {r4-r5, lr}
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       EpilogueScopes [
+// CHECK-NEXT:         EpilogueScope {
+// CHECK-NEXT:           StartOffset: 3
+// CHECK-NEXT:           Condition: 14
+// CHECK-NEXT:           EpilogueStartIndex: 2
+// CHECK-NEXT:           Opcodes [
+// CHECK-NEXT:             0xd6                ; pop {r4-r6, pc}
+// CHECK-NEXT:           ]
+// CHECK-NEXT:         }
+// CHECK-NEXT:       ]
+// CHECK-NEXT:     }
+// CHECK-NEXT:   }
+// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: fragment
+// CHECK:            FunctionLength: 6
+// CHECK:            Fragment: Yes
+// CHECK:            Prologue [
+// CHECK-NEXT:         0xcb                ; mov r11, sp
+// CHECK-NEXT:         0x10                ; sub sp, #(16 * 4)
+// CHECK-NEXT:         0xd5                ; push {r4-r5, lr}
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       EpilogueScopes [
+// CHECK-NEXT:         EpilogueScope {
+// CHECK-NEXT:           StartOffset: 1
+// CHECK-NEXT:           Condition: 14
+// CHECK-NEXT:           EpilogueStartIndex: 4
+// CHECK-NEXT:           Opcodes [
+// CHECK-NEXT:             0x10                ; add sp, #(16 * 4)
+// CHECK-NEXT:             0xd5                ; pop {r4-r5, pc}
+// CHECK-NEXT:           ]
+// CHECK-NEXT:         }
+// CHECK-NEXT:       ]
+// CHECK-NEXT:     }
+// CHECK-NEXT:   }
+// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: condepilog
+// CHECK:            FunctionLength: 8
+// CHECK:            Prologue [
+// CHECK-NEXT:         0xd5                ; push {r4-r5, lr}
+// CHECK-NEXT:       ]
+// CHECK-NEXT:       EpilogueScopes [
+// CHECK-NEXT:         EpilogueScope {
+// CHECK-NEXT:           StartOffset: 3
+// CHECK-NEXT:           Condition: 10
+// CHECK-NEXT:           EpilogueStartIndex: 2
+// CHECK-NEXT:           Opcodes [
+// CHECK-NEXT:             0xd5                ; pop {r4-r5, pc}
+// CHECK-NEXT:           ]
+// CHECK-NEXT:         }
+// CHECK-NEXT:       ]
+// CHECK-NEXT:     }
+// CHECK-NEXT:   }
+// CHECK-NEXT: ]
+
+        .text
+        .syntax unified
+        .globl func
+        .def func
+        .scl 2
+        .type 32
+        .endef
+        .seh_proc func
+func:
+        sub sp, sp, #24
+        .seh_stackalloc 24
+        sub sp, sp, #512
+        .seh_stackalloc_w 512
+        sub sp, sp, #512
+        .seh_stackalloc_w 512
+        sub sp, sp, #4096
+        .seh_stackalloc_w 4096
+        movw r7, #0
+        .seh_nop_w
+        movt r7, #0x4 // 0x40000
+        .seh_nop_w
+        sub sp, sp, r7
+        .seh_stackalloc_w 0x40000
+        push {r4-r8,lr}
+        .seh_save_regs_w {r4-r9,lr}
+        push {r4-r7,lr}
+        .seh_save_regs {r4-r7,lr}
+        push {r4,r6,r8,r10,r11,r12,lr}
+        .seh_save_regs_w {r4,r6,r8,r10,r11,r12,lr}
+        push {r3-r7,lr}
+        .seh_save_regs {r3-r7,lr}
+        vpush {d8-d14}
+        .seh_save_fregs {d8-d14}
+        vpush {q4-q5}
+        .seh_save_fregs {q4-q5}
+        mov lr, sp
+        .seh_save_sp lr
+        nop
+        .seh_nop
+        vpush {d7-d14}
+        .seh_save_fregs {d7-d14}
+        vpush {d18-d23}
+        .seh_save_fregs {d18-d23}
+        push {r3-r7,lr}
+        .seh_custom 0xed, 0xf8
+        .seh_endprologue
+        nop
+        .seh_startepilogue
+        mov r7, #512
+        .seh_nop_w
+        add sp, sp, r7
+        .seh_stackalloc 512
+        movw r7, #0
+        .seh_nop_w
+        movt r7, #0x4 // 0x40000
+        .seh_nop_w
+        add sp, sp, r7
+        .seh_stackalloc 0x40000
+        add sp, sp, #24
+        .seh_stackalloc 24
+        ldr lr, [sp], #16
+        .seh_save_lr 16
+        bx lr
+        .seh_nop
+        .seh_endepilogue
+        .seh_handler __C_specific_handler, %except
+        .seh_handlerdata
+        .long 0
+        .text
+        .seh_endproc
+
+        .seh_proc func2
+func2:
+        push {r4-r7}
+        .seh_save_regs {r4-r7}
+        .seh_endprologue
+        nop
+        .seh_startepilogue
+        pop {r4-r6}
+        .seh_save_regs {r4-r6}
+        b.w tailcall
+        .seh_nop_w
+        .seh_endepilogue
+        .seh_endproc
+
+        .seh_proc func3
+func3:
+        push {r4-r5,lr}
+        .seh_save_regs {r4-r5,lr}
+        .seh_endprologue
+        nop
+        // The p2align causes the length of the function to be unknown.
+        .p2align 1
+        nop
+        .seh_startepilogue
+        pop {r4-r6,pc}
+        .seh_save_regs {r4-r6,pc}
+        .seh_endepilogue
+        .seh_endproc
+
+        .seh_proc fragment
+fragment:
+        // Prologue opcodes without matching instructions
+        .seh_save_regs {r4-r5,lr}
+        .seh_stackalloc 64
+        .seh_save_sp r11
+        .seh_endprologue_fragment
+        nop
+        .seh_startepilogue
+        add sp, sp, #64
+        .seh_stackalloc 64
+        pop {r4-r5,pc}
+        .seh_save_regs {r4-r5,pc}
+        .seh_endepilogue
+        .seh_endproc
+
+        .seh_proc condepilog
+condepilog:
+        push {r4-r5,lr}
+        .seh_save_regs {r4-r5,lr}
+        .seh_endprologue
+        nop
+        it ge
+        .seh_startepilogue_cond ge
+        popge {r4-r5,pc}
+        .seh_save_regs {r4-r5,pc}
+        .seh_endepilogue
+        .seh_endproc
+
+        // Function with no .seh directives; no pdata/xdata entries are
+        // generated.
+        .globl smallFunc
+        .def smallFunc
+        .scl 2
+        .type 32
+        .endef
+        .seh_proc smallFunc
+smallFunc:
+        bx lr
+        .seh_endproc
+
+        // Function with no .seh directives, but with .seh_handlerdata.
+        // No xdata/pdata entries are generated, but the custom handler data
+        // (the .long after .seh_handlerdata) is left orphaned in the xdata
+        // section.
+        .globl handlerFunc
+        .def handlerFunc
+        .scl 2
+        .type 32
+        .endef
+        .seh_proc handlerFunc
+handlerFunc:
+        bx lr
+        .seh_handler __C_specific_handler, %except
+        .seh_handlerdata
+        .long 0
+        .text
+        .seh_endproc