[Attr] Introduce [[clang::nonportable_musttail]] as less strict version of [[clang::musttail]]

[xbolva00]

Summary

As noted in https://github.com/llvm/llvm-project/issues/54964 , musttail provide assurances that the tail call can be optimized on all targets.

Some users cares about only one specific target and if they use musttail attribute, they may see an Clang error about signatures mismatch. This mismatch is often a not a real problem in practice and LLVM can still emit a tail call.

So... Introduce [[clang::nonportable_musttail]] which is a less strict version of [[clang::musttail]] (no checking if signatures fully match).

Fixes https://github.com/llvm/llvm-project/issues/54964

[xbolva00]

Mention it in attribute documentation.

[efriedma]

Any thoughts on diagnostics here? If I'm not mistaken, with this patch, if you request an impossible tail call, you get a crash with very little useful information. (Although, see https://discourse.llvm.org/t/rfc-improving-clangs-middle-and-back-end-diagnostics/69261 )

[xbolva00]

In D147714#4249274, @efriedma wrote:

Any thoughts on diagnostics here? If I'm not mistaken, with this patch, if you request an impossible tail call, you get a crash with very little useful information. (Although, see https://discourse.llvm.org/t/rfc-improving-clangs-middle-and-back-end-diagnostics/69261 )

Yes, something like:

fatal error: error in backend: failed to perform tail call elimination on a call site marked musttail

It seems it would be possible to return also "reason" from IsEligibleForTailCallOptimization. On the other hand, IsEligibleForTailCallOptimization is implemented by many backends, so quite a lot of work, so for now maybe we could improve it atleast for X86 / ARM? I could do it for X86.

[efriedma]

An error message pointing at the call in the source code would be good enough, I think. Adding a "reason" might be nice, but not strictly necessary.

[xbolva00]

Added location when emiting diagnostic about failed TCO in the backend.

[xbolva00]

In D147714#4249357, @efriedma wrote:

An error message pointing at the call in the source code would be good enough, I think. Adding a "reason" might be nice, but not strictly necessary.

Good idea. So now for:

short i(void);
int j(void) {
    [[clang::nonportable_musttail]]
    return i();
}

Clang / LLVM on X86 emits:

example.cpp:2:5: error: failed to perform tail call elimination on a call site marked musttail
int j(void) {

^

With debug info:

example.cpp:4:12: error: failed to perform tail call elimination on a call site marked musttail

return i();
       ^

Nice improvement for end users.

So if no concerns, I will do necessary modifications for other backends.

[haberman]

Correct me if I am wrong, but won't this produce LLVM IR that fails to validate?

The checks that exist today for [[clang::musttail]] are necessary to follow LLVM's rules about where musttail may appear: https://llvm.org/docs/LangRef.html#id328

Calls marked musttail must obey the following additional rules:

• The call must immediately precede a ret instruction, or a pointer bitcast followed by a ret instruction.
• The ret instruction must return the (possibly bitcasted) value produced by the call, undef, or void.
• The calling conventions of the caller and callee must match.
• The callee must be varargs iff the caller is varargs. Bitcasting a non-varargs function to the appropriate varargs type is legal so long as the non-varargs prefixes obey the other rules.
• The return type must not undergo automatic conversion to an sret pointer.

In addition, if the calling convention is not swifttailcc or tailcc:

• All ABI-impacting function attributes, such as sret, byval, inreg, returned, and inalloca, must match.
• The caller and callee prototypes must match. Pointer types of parameters or return types may differ in pointee type, but not in address space.

It looks like this change would allow users to produce LLVM IR that violates that final condition:

The caller and callee prototypes must match. Pointer types of parameters or return types may differ in pointee type, but not in address space.

[xbolva00]

Possibly verifier ignores it? Because I dont see any verifier error compiling this testcase:

void NoParams();
void TestParamArityMismatchNonPortable(int x) {
  [[clang::nonportable_musttail]]
  return NoParams();
}

But yeah, this is something we need to solve somehow. Maybe just remove this one sentence or modify it...

[aaron.ballman]

In D147714#4249531, @xbolva00 wrote:

Added location when emiting diagnostic about failed TCO in the backend.

Will the diagnostic behavior change based on optimization level? Will the diagnostic be emitted to the correct location if there's no debug info? Are there frontend tests for this diagnostic?

Higher-level question: are we sure there's enough evidence this will be used in practice to warrant adding it to Clang?

[erichkeane]

Yeah, I think the diagnostics result of this attribute are unacceptable. I understand the goal here, but I suspect the more ergonomic (albeit, perhaps surprising) is a [[should_tail]] attribute sort of thing: a tail-hint where we do 'best effort with no promises', and make no guarantees that we're going to tail it. I'm sure this would require more work in the backends + require an additional IR tag, but even that seems difficult to use properly.

Could we instead encode the platform specific tail-call rules in Sema? We definitely have attributes that differ in behavior between platforms, and this seems like a perfect candidate.

[xbolva00]

Could we instead encode the platform specific tail-call rules in Sema? We definitely have attributes that differ in behavior between platforms, and this seems like a perfect candidate.

But then you break promise of musttail that code with musttail will work on any target supported by LLVM.

Plus, while we can relax frontend checks (based on target info), it is true that current LLVM LangRef says that there must be a match between prototypes.

Higher-level question: are we sure there's enough evidence this will be used in practice to warrant adding it to Clang?

Originally mentioned as request for CTRE project, I have use cases as well, @jacobsa (Github issue) mentioned it:

Personally I would like the semantics "give me an error if this can't be a tail call on this target", not the semantics "give me an error if this can't be a tail call on all targets clang supports". I suspect most people would be looking for the former, since most people care about nearly none of the supported targets by volume.

I agree with that.

[erichkeane]

In D147714#4251208, @xbolva00 wrote:

Could we instead encode the platform specific tail-call rules in Sema? We definitely have attributes that differ in behavior between platforms, and this seems like a perfect candidate.

But then you break promise of musttail that code with musttail will work on any target supported by LLVM.

I should have been clearer: I meant this with a new attribute.

Plus, while we can relax frontend checks (based on target info), it is true that current LLVM LangRef says that there must be a match between prototypes.

Right, this would require its own llvm-level concept as well.

Higher-level question: are we sure there's enough evidence this will be used in practice to warrant adding it to Clang?

Originally mentioned as request for CTRE project, I have use cases as well, @jacobsa (Github issue) mentioned it:

Personally I would like the semantics "give me an error if this can't be a tail call on this target", not the semantics "give me an error if this can't be a tail call on all targets clang supports". I suspect most people would be looking for the former, since most people care about nearly none of the supported targets by volume.

I agree with that.

I see value in that, absolutely, but this isn't the way about it. Said new attribute (whether it be [[should_tail]] for [[nonportable_musttail]]) shouldn't be diagnosing the code generator, it should be diagnosing in SEMA. This requires encoding the platform specific/LLVM-IR specific rules. If LLVM-IR docs require musttail rules (that this patch seems to be breaking!), we perhaps need an alternative at the IR level as well.

[xbolva00]

Will the diagnostic be emitted to the correct location if there's no debug info?

I did some experiments and with no debug info, I got atleast the location of the caller. If we consider than currently there is no location, anything is improvement and this is probably best we can do.

Are there frontend tests for this diagnostic?

Do you mean C ---> ASM tests? I think we dont have such test, are they even allowed in Clang testsuite?

[haberman]

is a [[should_tail]] attribute sort of thing: a tail-hint where we do 'best effort with no promises', and make no guarantees that we're going to tail it.

I'm not sure I see the value in that. The compiler already optimizes tail calls when it can in a best-effort manner. The purpose of [[musttail]] is to support algorithms that would blow the stack if tail calls were not optimized. It's better to get a compiler error than to get a stack overflow at runtime (especially if stack overflow only occurs with certain inputs).

[erichkeane]

In D147714#4251690, @haberman wrote:

is a [[should_tail]] attribute sort of thing: a tail-hint where we do 'best effort with no promises', and make no guarantees that we're going to tail it.

I'm not sure I see the value in that. The compiler already optimizes tail calls when it can in a best-effort manner. The purpose of [[musttail]] is to support algorithms that would blow the stack if tail calls were not optimized. It's better to get a compiler error than to get a stack overflow at runtime (especially if stack overflow only occurs with certain inputs).

I see, yes, that makes sense. Then I go back to my other suggestion: we this needs to encode the per-platform limitations in Sema.

I could definitely see this being an 'alias' for musttail on most platforms, and us reducing the restrictions over time on a per-platform basis though, so long as we knew codegen would be compatible with it.

[xbolva00]

So to make some conclusion - there is a need for less strict version of musttail in LLVM, right? Ideally platform specific, like x86_musttail?

[erichkeane]

In D147714#4251752, @xbolva00 wrote:

So to make some conclusion - there is a need for less strict version of musttail in LLVM, right? Ideally platform specific, like x86_musttail?

I'm not sure what the rules for the LLVM definition of musttail should be, but from the CFE's perspective, something like [[nonportable_musttail]] that always enforced the rules of the CURRENT platform(note: or stricter, and relaxing would be allowed later) would be acceptable.

[haberman]

[[nonportable_musttail]] makes sense to me as a semantic. It indicates that the algorithm requires tail calls, but the author is willing to accept that the algorithm may be non-portable.

"Non-portable" here can mean architecture-specific, but it can also mean "sensitive to compiler flags." For example, I think there are architectures where tail calls can be optimized for statically-linked code but not across a shared library boundary.

Since the circumstances for when tail calls are supported can be complicated and subtle, architecture-specific attributes don't make as much sense to me.

[nickdesaulniers]

In D147714#4249274, @efriedma wrote:

Any thoughts on diagnostics here? If I'm not mistaken, with this patch, if you request an impossible tail call, you get a crash with very little useful information. (Although, see https://discourse.llvm.org/t/rfc-improving-clangs-middle-and-back-end-diagnostics/69261 )

Patching this in, and running through the original test cases as reported in https://github.com/llvm/llvm-project/issues/54964#issue-1207256071:

$ clang++ x.cpp
x.cpp:5:32: error: cannot perform a tail call to function 'base' because its signature is incompatible with the calling function
int f1() { [[clang::musttail]] return base(5); }
                               ^
x.cpp:1:1: note: target function has different number of parameters (expected 0 but has 1)
int base(int);
^
x.cpp:5:14: note: tail call required by 'musttail' attribute here
int f1() { [[clang::musttail]] return base(5); }
             ^
x.cpp:6:57: error: cannot perform a tail call to function 'base' because its signature is incompatible with the calling function
int f2(int x, int y /* unused */) { [[clang::musttail]] return base(5); }
                                                        ^
x.cpp:1:1: note: target function has different number of parameters (expected 2 but has 1)
int base(int);
^
x.cpp:6:39: note: tail call required by 'musttail' attribute here
int f2(int x, int y /* unused */) { [[clang::musttail]] return base(5); }
                                      ^
2 errors generated.

so that's pretty informative to the user. Changing musttail to nonportable_musttail:

$ clang++ x.cpp
cannot guarantee tail call due to mismatched parameter counts
  %call = musttail call noundef i32 @_Z4basei(i32 noundef 5)
in function _Z2f1v
fatal error: error in backend: Broken function found, compilation aborted!
clang++: error: clang frontend command failed with exit code 70 (use -v to see invocation)
clang version 17.0.0 ([e-mail]:llvm/llvm-project.git 36e13ec194d1601499e0069b33b9bd8f69887e1f)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /android0/llvm-project/llvm/build/bin
clang++: note: diagnostic msg: 
********************

PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:
Preprocessed source(s) and associated run script(s) are located at:
clang++: note: diagnostic msg: /tmp/x-299fe3.cpp
clang++: note: diagnostic msg: /tmp/x-299fe3.sh
clang++: note: diagnostic msg: 

********************

So at the very least, this should use the existing clang support for backend diagnostics rather than crashing. Please see how !srcloc works for inline asm and __attribute__((warning(""))). It should be straightforward to attach that ad-hoc metadata to call sites with that statement attribute.

Also, "woof" to the attribute name.

Taking a step back: can't we just fix the backends or modify CheckTypesMatch (or add a new method) to detect partial matches? Is there actually a safety issue on any known target? If not, creating a separate attribute feels like working around a possible bug in a particular backend when that should probably be fixed instead. Reading https://github.com/llvm/llvm-project/issues/54964, I'm curious if folks have pursued @efriedma 's suggestion #2 from https://github.com/llvm/llvm-project/issues/54964#issuecomment-1101612886?

[xbolva00]

I'm curious if folks have pursued @efriedma 's suggestion #2 from https://github.com/llvm/llvm-project/issues/54964#issuecomment-1101612886?

This is something we mentioned here, perform target specific checks in Sema (but still it is not possible to catch everything in Sema!)

But the problem of strict musttail call maker stays here. So either we have to relax musttail or create new one.

[nickdesaulniers]

In D147714#4253247, @xbolva00 wrote:

I'm curious if folks have pursued @efriedma 's suggestion #2 from https://github.com/llvm/llvm-project/issues/54964#issuecomment-1101612886?

This is something we mentioned here, perform target specific checks in Sema (but still it is not possible to catch everything in Sema!)

But the problem of strict musttail call maker stays here. So either we have to relax musttail or create new one.

(After reading more sources to see where musttail fails: https://github.com/llvm/llvm-project/issues/54964#issuecomment-1500708651)

I'm not against a new one, but I'll retain my "Changes Requested" on this patch until the UX is fixed. Restating:

So at the very least, this should use the existing clang support for backend diagnostics rather than crashing. Please see how !srcloc works for inline asm and attribute((warning(""))). It should be straightforward to attach that ad-hoc metadata to call sites with that statement attribute.

Do not crash the compiler like this.

[yamt]

musttail provide assurances that the tail call can be optimized on all targets.

Such a definition doesn't make much sense because for some targets tail call optimization isn't available at all.
Eg. Wasm w/o tail call proposal, xtensa windowed abi.

So I'm not sure what's the point of the new attribute. It seems simpler to fix the existing musttail attribute.

[xbolva00]

In D147714#4446780, @yamt wrote:

musttail provide assurances that the tail call can be optimized on all targets.

Such a definition doesn't make much sense because for some targets tail call optimization isn't available at all.
Eg. Wasm w/o tail call proposal, xtensa windowed abi.

So I'm not sure what's the point of the new attribute. It seems simpler to fix the existing musttail attribute.

I agree, I consider the original design of this feature as a mistake, which made this attribute kinda unreliable and useless.. I dont think people are willing to just change existing impl of musttail attribute...

[haberman]

Such a definition doesn't make much sense because for some targets tail call optimization isn't available at all.
Eg. Wasm w/o tail call proposal, xtensa windowed abi.

That is true. But what is the correct behavior if you try to use [[clang::musttail]] on such a target? In my case, I want to get a compiler error, because if a tail call cannot be performed, the algorithm is incorrect (it will cause stack overflow at runtime).

There are three possible semantics that I can see:

1. Perform a tail call if possible.
2. Perform a tail call, fail to compile if this is not possible ([[clang::nonportable_musttail]])
3. Perform a tail call, fail to compile if this is not possible or if the function signatures between caller and callee differ in ways that make tail calls difficult on some platforms ([[clang::musttail]])

Both (2) and (3) would fail to compile on Wasm without tail calls. If all you want is (1), that doesn't require any attribute (the optimizer will do it automatically).

[xbolva00]

In D147714#4446809, @haberman wrote:

Such a definition doesn't make much sense because for some targets tail call optimization isn't available at all.
Eg. Wasm w/o tail call proposal, xtensa windowed abi.

That is true. But what is the correct behavior if you try to use [[clang::musttail]] on such a target? In my case, I want to get a compiler error, because if a tail call cannot be performed, the algorithm is incorrect (it will cause stack overflow at runtime).

There are three possible semantics that I can see:

1. Perform a tail call if possible.
2. Perform a tail call, fail to compile if this is not possible ([[clang::nonportable_musttail]])
3. Perform a tail call, fail to compile if this is not possible or if the function signatures between caller and callee differ in ways that make tail calls difficult on some platforms ([[clang::musttail]])

Both (2) and (3) would fail to compile on Wasm without tail calls. If all you want is (1), that doesn't require any attribute (the optimizer will do it automatically).

So clearly definition “musttail provide assurances that the tail call can be optimized on all targets.” is broken in reality.

Code with musttail will compile fine on x86, and musttail ensures that should compile to tail calls on other targets - not a case, on eg. wasm you get just an error. Or what is your definition of “all targets”? Seems fragile in practise, there are crashes even on arm..

[xbolva00]

If all you want is (1), that doesn't require any attribute (the optimizer will do it automatically).

Not entirely true, as compiler can do anything. In some cases you really need *musttail* (in llvm IR) to get or preserve tail call.

As current musttail is too restricted, we need nonportable_musttail (aka real musttail one would want)

[haberman]

I think I misunderstood @yamt's comment in https://reviews.llvm.org/D147714#4446780. I take back what I wrote above.

I agree that [[clang::nonportable_musttail]] is a nicer semantic, and the restrictions around the existing [[clang:musttail]] don't seem to buy us very much, since they are not universal enough to give a true assurance.

If someone could land a change to both LLVM and Clang to change the existing attribute, I would have no objection. But I have no idea what is involved in landing a backend (LLVM) change of that magnitude, especially since it would touch all the arch-specific backends.

clang/include/clang/Basic/Attr.td

def NoMerge : DeclOrStmtAttr {
   let Spellings = [Clang<"nomerge">];
   let Documentation = [NoMergeDocs];
   let Subjects = SubjectList<[Function, Stmt], ErrorDiag,
                              "functions and statements">;
   let SimpleHandler = 1;
 }
 
 def MustTail : StmtAttr {
-  let Spellings = [Clang<"musttail">];
+  let Spellings = [Clang<"musttail">, Clang<"nonportable_musttail">];
   let Documentation = [MustTailDocs];
+  let Accessors = [Accessor<"isNonPortableMustTail", [Clang<"nonportable_musttail">]>];
   let Subjects = SubjectList<[ReturnStmt], ErrorDiag, "return statements">;
 }
 
 def FastCall : DeclOrTypeAttr {
   let Spellings = [GCC<"fastcall">, Keyword<"__fastcall">,
                    Keyword<"_fastcall">];
 //  let Subjects = [Function, ObjCMethod];
   let Documentation = [FastCallDocs];

clang/include/clang/Basic/AttrDocs.td

 arguments must be similar according to C++ rules (differing only in cv
 qualifiers or array size), including the implicit "this" argument, if any.
 Any variables in scope, including all arguments to the function and the
 return value must be trivially destructible. The calling convention of the
 caller and callee must match, and they must not be variadic functions or have
 old style K&R C function declarations.
 
 ``clang::musttail`` provides assurances that the tail call can be optimized on
-all targets, not just one.
+all targets, not just one. If you care about one target and you do not need
+these assurances, use ``clang::nonportable_musttail``.
   }];
 }
 
 def AssertCapabilityDocs : Documentation {
   let Category = DocCatFunction;
   let Heading = "assert_capability, assert_shared_capability";
   let Content = [{
 Marks a function that dynamically tests whether a capability is held, and halts

clang/include/clang/Sema/Sema.h

   /// Check to see if the given expression is a valid argument to a variadic
   /// function, issuing a diagnostic if not.
   void checkVariadicArgument(const Expr *E, VariadicCallType CT);
 
   /// Check whether the given statement can have musttail applied to it,
   /// issuing a diagnostic and returning false if not. In the success case,
   /// the statement is rewritten to remove implicit nodes from the return
   /// value.
-  bool checkAndRewriteMustTailAttr(Stmt *St, const Attr &MTA);
+  bool checkAndRewriteMustTailAttr(Stmt *St, const MustTailAttr &MTA);
 
 private:
   /// Check whether the given statement can have musttail applied to it,
   /// issuing a diagnostic and returning false if not.
-  bool checkMustTailAttr(const Stmt *St, const Attr &MTA);
+  bool checkMustTailAttr(const Stmt *St, const MustTailAttr &MTA);
 
 public:
   /// Check to see if a given expression could have '.c_str()' called on it.
   bool hasCStrMethod(const Expr *E);
 
   /// GatherArgumentsForCall - Collector argument expressions for various
   /// form of call prototypes.
   bool GatherArgumentsForCall(SourceLocation CallLoc, FunctionDecl *FDecl,

clang/lib/Sema/SemaStmt.cpp

 
 StmtResult Sema::BuildAttributedStmt(SourceLocation AttrsLoc,
                                      ArrayRef<const Attr *> Attrs,
                                      Stmt *SubStmt) {
   // FIXME: this code should move when a planned refactoring around statement
   // attributes lands.
   for (const auto *A : Attrs) {
     if (A->getKind() == attr::MustTail) {
-      if (!checkAndRewriteMustTailAttr(SubStmt, *A)) {
+      if (!checkAndRewriteMustTailAttr(SubStmt, *cast<MustTailAttr>(A))) {
         return SubStmt;
       }
       setFunctionHasMustTail();
     }
   }
 
   return AttributedStmt::Create(Context, AttrsLoc, Attrs, SubStmt);
 }
 
 StmtResult Sema::ActOnAttributedStmt(const ParsedAttributes &Attrs,
                                      Stmt *SubStmt) {
   SmallVector<const Attr *, 1> SemanticAttrs;
   ProcessStmtAttributes(SubStmt, Attrs, SemanticAttrs);
   if (!SemanticAttrs.empty())
     return BuildAttributedStmt(Attrs.Range.getBegin(), SemanticAttrs, SubStmt);
   // If none of the attributes applied, that's fine, we can recover by
   // returning the substatement directly instead of making an AttributedStmt
   // with no attributes on it.
   return SubStmt;
 }
 
-bool Sema::checkAndRewriteMustTailAttr(Stmt *St, const Attr &MTA) {
+bool Sema::checkAndRewriteMustTailAttr(Stmt *St, const MustTailAttr &MTA) {
   ReturnStmt *R = cast<ReturnStmt>(St);
   Expr *E = R->getRetValue();
 
   if (CurContext->isDependentContext() || (E && E->isInstantiationDependent()))
     // We have to suspend our check until template instantiation time.
     return true;
 
   if (!checkMustTailAttr(St, MTA))
     return false;
 
   // FIXME: Replace Expr::IgnoreImplicitAsWritten() with this function.
   // Currently it does not skip implicit constructors in an initialization
   // context.
   auto IgnoreImplicitAsWritten = [](Expr *E) -> Expr * {
     return IgnoreExprNodes(E, IgnoreImplicitAsWrittenSingleStep,
                            IgnoreElidableImplicitConstructorSingleStep);
   };
 
   // Now that we have verified that 'musttail' is valid here, rewrite the
   // return value to remove all implicit nodes, but retain parentheses.
   R->setRetValue(IgnoreImplicitAsWritten(E));
   return true;
 }
 
-bool Sema::checkMustTailAttr(const Stmt *St, const Attr &MTA) {
+bool Sema::checkMustTailAttr(const Stmt *St, const MustTailAttr &MTA) {
   assert(!CurContext->isDependentContext() &&
          "musttail cannot be checked from a dependent context");
 
   // FIXME: Add Expr::IgnoreParenImplicitAsWritten() with this definition.
   auto IgnoreParenImplicitAsWritten = [](const Expr *E) -> const Expr * {
     return IgnoreExprNodes(const_cast<Expr *>(E), IgnoreParensSingleStep,
                            IgnoreImplicitAsWrittenSingleStep,
                            IgnoreElidableImplicitConstructorSingleStep);
for (size_t I = 0; I < N; I++) {
         PD << static_cast<int>(I) + 1;
         return false;
       }
     }
 
     return true;
   };
 
+  if (MTA.isNonPortableMustTail())
+    return true;
+
   PartialDiagnostic PD = PDiag(diag::note_musttail_mismatch);
   if (!CheckTypesMatch(CallerType, CalleeType, PD)) {
     if (const auto *ND = dyn_cast_or_null<NamedDecl>(CE->getCalleeDecl()))
       Diag(St->getBeginLoc(), diag::err_musttail_mismatch)
           << true << ND->getDeclName();
     else
       Diag(St->getBeginLoc(), diag::err_musttail_mismatch) << false;
     Diag(CalleeLoc, PD);

clang/test/SemaCXX/attr-musttail.cpp

 // RUN: %clang_cc1 -verify -fsyntax-only -fms-extensions -fcxx-exceptions -fopenmp -triple x86_64-linux %s
 
 int ReturnsInt1();
 int Func1() {
   [[clang::musttail]] ReturnsInt1();              // expected-error {{'musttail' attribute only applies to return statements}}
   [[clang::musttail(1, 2)]] return ReturnsInt1(); // expected-error {{'musttail' attribute takes no arguments}}
   [[clang::musttail]] return 5;                   // expected-error {{'musttail' attribute requires that the return value is the result of a function call}}
   [[clang::musttail]] return ReturnsInt1();
+  [[clang::nonportable_musttail]] return ReturnsInt1();
 }
 
 void NoFunctionCall() {
   [[clang::musttail]] return; // expected-error {{'musttail' attribute requires that the return value is the result of a function call}}
 }
 
 [[clang::musttail]] static int int_val = ReturnsInt1(); // expected-error {{'musttail' attribute cannot be applied to a declaration}}
 
 void NoParams(); // expected-note {{target function has different number of parameters (expected 1 but has 0)}}
 void TestParamArityMismatch(int x) {
   [[clang::musttail]] // expected-note {{tail call required by 'musttail' attribute here}}
   return NoParams();  // expected-error {{cannot perform a tail call to function 'NoParams' because its signature is incompatible with the calling function}}
 }
 
+void TestParamArityMismatchNonPortable(int x) {
+  [[clang::nonportable_musttail]]
+  return NoParams();
+}
+
 void LongParam(long x); // expected-note {{target function has type mismatch at 1st parameter (expected 'long' but has 'int')}}
 void TestParamTypeMismatch(int x) {
   [[clang::musttail]]  // expected-note {{tail call required by 'musttail' attribute here}}
   return LongParam(x); // expected-error {{cannot perform a tail call to function 'LongParam' because its signature is incompatible with the calling function}}
 }
 
 long ReturnsLong(); // expected-note {{target function has different return type ('int' expected but has 'long')}}
 int TestReturnTypeMismatch() {
void BadPMF() {
     [[clang::musttail]] return ((*this)->*pmf)(); // expected-error {{left hand operand to ->* must be a pointer to class compatible with the right hand operand, but is 'TestBadPMF'}}
   }
 };
 
 namespace ns {}
 void TestCallNonValue() {
   [[clang::musttail]] return ns; // expected-error {{unexpected namespace name 'ns': expected expression}}
 }
+
+struct Foo {
+  int foo(int) { return 0; }
+};
+
+struct Bar {
+  Foo foo;
+  int bar(int);
+};
+
+int Bar::bar(int x) {
+  [[clang::nonportable_musttail]]
+  return foo.foo(x);
+}

llvm/lib/Target/X86/X86ISelLowering.cpp

if (isTailCall && !IsMustTail) {
     if (!IsGuaranteeTCO && isTailCall)
       IsSibcall = true;
 
     if (isTailCall)
       ++NumTailCalls;
   }
 
   if (IsMustTail && !isTailCall)
-    report_fatal_error("failed to perform tail call elimination on a call "
+    errorUnsupported(DAG, dl, "failed to perform tail call elimination on a call "
                        "site marked musttail");
 
   assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
          "Var args not supported with calling convention fastcc, ghc or hipe");
 
   // Analyze operands of the call, assigning locations to each operand.
   SmallVector<CCValAssign, 16> ArgLocs;
   CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());

llvm/test/CodeGen/X86/tailcc-notail.ll

 ; RUN: not --crash llc -mtriple=x86_64-linux-gnu %s -o - 2>&1 | FileCheck %s
 
-; CHECK: LLVM ERROR: failed to perform tail call elimination on a call site marked musttail
+; CHECK: error: {{.*}} in function caller {{.*}}: failed to perform tail call elimination on a call site marked musttail
 declare tailcc [16 x i64] @callee()
 define tailcc [16 x i64] @caller() {
   %res = musttail call tailcc [16 x i64] @callee()
   ret [16 x i64] %res
 }