Register Calling Convention, Clang changes

[erichkeane]

Summary

The Register Calling Convention (RegCall) was introduced by Intel to optimize parameter transfer on function call.
This calling convention ensures that as many values as possible are passed or returned in registers.

The following review presents the basic additions to the Clang Front-end in order to support RegCall in X86.

See https://reviews.llvm.org/D25022 for CodeGen changes

[ABataev]

Eric, please, prepare a full context review!

lib/AST/ItaniumMangle.cpp

1417–1421 ↗

(On Diff #73309)

Single-line substatements must not be enclosed into braces

1418 ↗

(On Diff #73309)

Line is not clang-formatted

lib/AST/Mangle.cpp

70 ↗

(On Diff #73309)

Seems to me this change is not required

lib/AST/MicrosoftMangle.cpp

440 ↗

(On Diff #73309)

No changes here, restore original line

[erichkeane]

Hi Alexey-
Can you let me know what you mean by "Full Context Review"? I'm unfamiliar with that process. The other fixes I'll look at today.

[erichkeane]

Updated the code, fixed Alexey's concerns. Thanks again for the comments!

[ABataev]

In D25204#560894, @erichkeane wrote:

Hi Alexey-
Can you let me know what you mean by "Full Context Review"? I'm unfamiliar with that process. The other fixes I'll look at today.

Check this page http://llvm.org/docs/Phabricator.html
use

git diff -U999999 other-branch

or

svn diff --diff-cmd=diff -x -U999999

[erichkeane]

Did a full context diff, as requested.

[ABataev]

lib/Sema/SemaDeclAttr.cpp

3833–3837 ↗

(On Diff #73490)

Not formatted

[ABataev]

include/clang-c/Index.h

3029 ↗

(On Diff #73490)

Maybe it is better to use 8, as the previous comment allows it?

/* Value 8 was PnaclCall, but it was never used, so it could safely be re-used. */

In this case you don't need to increase number of bits used for calling conventions

lib/AST/ItaniumMangle.cpp

1236–1237 ↗

(On Diff #73490)

What if function type is not a FunctionProtoType?

lib/AST/MicrosoftMangle.cpp

435–436 ↗

(On Diff #73490)

Again, what if type is not FunctionProtoType?

lib/CodeGen/TargetInfo.cpp

1546–1548 ↗

(On Diff #73490)

Seems to me this code is clang-formatted

3301–3303 ↗

(On Diff #73490)

No braces

3388–3392 ↗

(On Diff #73490)

Not clang-formatted and extra braces

lib/Sema/SemaDecl.cpp

8288 ↗

(On Diff #73490)

Is this formatted?

[rnk]

include/clang/Basic/AttrDocs.td

1263 ↗

(On Diff #73490)

"aims"

lib/CodeGen/TargetInfo.cpp

3306 ↗

(On Diff #73490)

You might want to defend against dynamic C++ records which will have vtable fields.

3323 ↗

(On Diff #73490)

This code doesn't need to worry about windows.h defining max.

3388–3389 ↗

(On Diff #73490)

not llvm style

3742–3743 ↗

(On Diff #73490)

'classify' takes FreeSSERegs by reference and modifies it, so are you sure this is correct? It means if I have this kind of prototype, we won't pass 'd' in registers because we'll consume four registers for the return value:

struct HFA { __m128 f[4]; };
HFA __regcall f(HFA a, HFA b, HFA c, HFA d) {
  ...
}

lib/Sema/SemaDecl.cpp

8288 ↗

(On Diff #73490)

The comment doesn't apply here. Are you sure you don't want some other behavior, like unprototyped functions are actually implicitly void when regcall is used, as in C++?

[erichkeane]

New patch incoming.

lib/AST/ItaniumMangle.cpp

1236–1237 ↗

(On Diff #73490)

Right, good catch. I looked at Mangle.cpp which does something very similar, and assumes that FunctionType is a valid cast here, so I've switched this here too, please let me know if that is a wrong assumption.

lib/CodeGen/TargetInfo.cpp

3306 ↗

(On Diff #73490)

Hmmm... I'm not sure what behavior would be expected in that case. Also, I just looked at a bunch of similar 'for' loops for various other reasons, and I don't really see where that is tested elsewhere (so I don't really see how to, besides I.isVirtual?). In that case, I would suspect that the class is not put into a register?

3742–3743 ↗

(On Diff #73490)

That was my intent, this should allow return values to be in registers as well if I'm reading the spec correctly. The idea is that register use is 'greedy'.

lib/Sema/SemaDecl.cpp

8288 ↗

(On Diff #73490)

I suspect you're right. I am not sure what behavior would be expected, so I think that reverting to the error case is likely the correct "safe" behavior if the spec doesn't say otherwise?

[erichkeane]

Fixes based on Alexey/Ried's feedback

[rnk]

lib/CodeGen/TargetInfo.cpp

3742–3743 ↗

(On Diff #73490)

But, if the return value is returned directly, it doesn't conflict with the free parameter registers. In my example, the return value can use XMM0-3 and the parameters can use XMM0-15. Can you add this test case and validate that it does what you want?

[erichkeane]

lib/CodeGen/TargetInfo.cpp

3742–3743 ↗

(On Diff #73490)

Ah, I see what you mean now. That definitely sounds like the right way to go about it. I'll add a test and validate my assumptions. Thanks for putting up with my misunderstanding.

[rnk]

lib/AST/ItaniumMangle.cpp

1203 ↗

(On Diff #73507)

What mangling should happen for operator overloads and all other kinds of DeclarationName? Please add tests for these cases

1236–1237 ↗

(On Diff #73490)

IMO this would be simpler if we did:

if (IsRegCall)
  mangleRegCallName(II); // implement this elsewhere
else
  mangleSourceName(II);

I don't like that the standard mangleSourceName as you have it has to be concerned with IsRegCall. Without your change, it is a simple function that clearly expresses that the basic unit of Itanium name mangling is decimal length prefixed strings.

lib/CodeGen/TargetInfo.cpp

3306 ↗

(On Diff #73490)

You don't need to check if each base is dynamic, just check CXXRD->isDynamicClass() and make it return indirectly if so.

1954 ↗

(On Diff #73507)

format

3297 ↗

(On Diff #73507)

Please name variables in LLVM style

3321 ↗

(On Diff #73507)

variable naming

3352–3354 ↗

(On Diff #73507)

Since you're touching most of this, can you make this code standardize on LLVM's variable naming convention: http://llvm.org/docs/CodingStandards.html#name-types-functions-variables-and-enumerators-properly

[majnemer]

lib/AST/ItaniumMangle.cpp

1413–1414 ↗

(On Diff #73507)

I'd add an overload of mangleSourceName which takes a Twine. Then the one caller which passes isRegCall would merely concat the identifier with regcall3. The existing implementation using IdentifierInfo would merely pass in II->getName for the Twine.

[majnemer]

lib/AST/ItaniumMangle.cpp

1234 ↗

(On Diff #73507)

auto *

1235 ↗

(On Diff #73507)

The convention is to just do FD instead of (FD != nullptr)

[erichkeane]

Commenting to save my comments (don't seem to survive a refresh). Still working on non-function mangling.

lib/AST/ItaniumMangle.cpp

1203 ↗

(On Diff #73507)

Test cases show that I need to implement these. Going to quit for the evening, but will continue working on this tomorrow.

1413–1414 ↗

(On Diff #73507)

I really like this idea, and had most of an implementation, however it seems that Twine has no real good way to get its 'length', so it required a temporary for the sake of getting the length. Instead I ended up with Ried's suggestion of separate functions for each.

lib/CodeGen/TargetInfo.cpp

3742–3743 ↗

(On Diff #73490)

I added it to the C test, and was convinced that you're right. Fixed, thanks!

[oren_ben_simhon]

include/clang/Basic/AttrDocs.td

1267 ↗

(On Diff #73507)

You might want to use the following link instead because it is most updated: https://software.intel.com/en-us/node/693069

lib/CodeGen/TargetInfo.cpp

3352 ↗

(On Diff #73507)

According to the ABI, there are 12 free int regs for windows and 11 free int regs for non-windows (linux, OSX, etc). Is that taken into account somewhere?

3732 ↗

(On Diff #73507)

Maybe i misinterpret the comment, but AFAIK, RegCall gives us 16 SSE registers for each (return values and passed arguments)

test/CodeGen/regcall.c

26 ↗

(On Diff #73507)

I see that expended structures don't get InReg attribute. IMHO, If you know that the value should be saved in register then you InReg attribute should be added.

[ABataev]

include/clang/AST/Type.h

1381 ↗

(On Diff #73507)

Erich, do you really need this? You don't increase number of required bits anymore, so this code must be restored

2909–2921 ↗

(On Diff #73507)

Also, you don't need these changes anymore

lib/AST/MicrosoftMangle.cpp

434 ↗

(On Diff #73507)

auto*

lib/CodeGen/CodeGenModule.cpp

686 ↗

(On Diff #73507)

const auto *FD

lib/CodeGen/TargetInfo.cpp

1546–1548 ↗

(On Diff #73490)

I mean, did you try to reformat the whole return statement? It looks ugly

3321 ↗

(On Diff #73507)

I believe the whole patch must be clang-tided

[erichkeane]

include/clang/Basic/AttrDocs.td

1267 ↗

(On Diff #73507)

This has changed 2x since I started this project. Is there a way to get a STABLE link? I imagine that much of this documentation is filled with broken links (since MSDN breaks them constantly), but don't really want to add to it.

lib/CodeGen/TargetInfo.cpp

3352 ↗

(On Diff #73507)

Yes. There are separate ABIInfo types for windows.

3732 ↗

(On Diff #73507)

I'd misread that in the spec and Ried corrected my implementation below. Updating the comment.

test/CodeGen/regcall.c

26 ↗

(On Diff #73507)

I am not sure that is the case. That behavior doesn't happen in vectorcall seemingly.

[erichkeane]

Updated tests to properly show mangling for C++ types. Required some fixes. Note that the decorating of names doesn't match ICC in non-named functions due to a bug in ICC, and the spec is silent about it, so this patch decorates them in the way that best fits the spec.

[oren_ben_simhon]

include/clang/Basic/AttrDocs.td

1267 ↗

(On Diff #73507)

I am not sure if there is a constant link to the latest and greatest. When i contacted the documentation team they pointed me to the latest published compiler documention version 16

[erichkeane]

Quick point i meant to post earlier Couldn't change ExtInfo size.

include/clang/AST/Type.h

1381 ↗

(On Diff #73507)

After much debugging, I realized that this does NOT store the enum from above, it stores the CallingConv from include/clang/Basic/Specifiers.h, which we actually HAVE put over 16. I didn't see anything that we could lose from that enum to replace it however.

[erichkeane]

I'm going to need to pump-the-brakes on this for a little bit. The name-decoration work I did here highlighted a number of issues with that section of the spec. We're currently considering rev'ing the spec to properly name mangle/decorate when C++ and Microsoft is ready.

In the meantime, the LLVM changes are unaffected.

[erichkeane]

After much debate, the architects have agreed to change the "Decoration" section to the following.

The next patch does these, so I'm ready for continued review. Thanks for your patience!
-Erich

regcall Decoration
Names of functions that use regcall are decorated. This helps avoid improper linking with a function using a different calling convention.
The name of a C function is prefixed with regcalln. For example, the name foo would be decorated as follows: regcall3foo. The n part of the decoration specifies the version of the regcall convention in effect (the current convention revision number is 3). In a GCC compatibility environment, this decoration is also done for a C++ function or template function whose function or template name is an identifier (excluding constructors and operator functions, for example); this happens before C++ name mangling.
In a Microsoft compatibility environment, C++ name mangling uses the (lower case) letter w to encode revision 3 of the regcall calling convention.

[erichkeane]

Corrected Decoration settings to match the soon-to-be-updated spec.

[rnk]

include/clang/Basic/AttrDocs.td

1263 ↗

(On Diff #75903)

If you want __regcall to appear as a link, you want to put backticks and around it and underscore after it, like is done for __fastcall above:

`__regcall`_

1265 ↗

(On Diff #75903)

I don't think you need this trailing single quote.

lib/AST/MicrosoftMangle.cpp

2006 ↗

(On Diff #75903)

Update the EBNF comment

lib/CodeGen/TargetInfo.cpp

3324 ↗

(On Diff #75903)

This will return false for class fields, which I don't think you want. You probably want to do this for record types that aren't unions.

3396 ↗

(On Diff #75903)

You probably wanted isRecordType().

[majnemer]

lib/AST/MicrosoftMangle.cpp

433 ↗

(On Diff #75903)

Please remove this stray newline.

lib/CodeGen/TargetInfo.cpp

3333 ↗

(On Diff #75903)

I think UINT_MAX is more popular than using std::numeric_limits for this purpose in LLVM/Clang.

3337–3338 ↗

(On Diff #75903)

Please avoid else after return: http://llvm.org/docs/CodingStandards.html#don-t-use-else-after-a-return

[erichkeane]

New diff coming that fixes Reid & David's comments

[joerg]

Can we please avoid adding more (pseudo) keywords in the double-underscore name space? Those tend to be used a lot by existing libc implementations and existing attribute cases like strong and weak have created enough trouble.

[erichkeane]

I guess I'm not sure how to respond to that... Calling conventions traditionally use double underscore to prevent from stomping on user keywords. Additionally, this is in a specification that has an existing implementation available, so I'm not sure what could be done.

[majnemer]

The __ namespace is reserved for us and I can't imagine how __regcall would upset any existing code out there.

[joerg]

The __ namespace is shared between all parts of the implementation, not just the compiler. The convention in the past was that compiler keywords will end in a __ as well, but calling conventions and Objective C broke that convention. I still think it is something that should be avoided when possible.

[rnk]

Remember the fight over _Atomic with MSVC's STL? The fallacy of the implementer's namespace is that there is only one implementer.
https://llvm.org/bugs/show_bug.cgi?id=19043

We should prefer adding __attribute__s and __declspecs instead of keywords when possible.

[erichkeane]

In D25204#581469, @rnk wrote:

Remember the fight over _Atomic with MSVC's STL? The fallacy of the implementer's namespace is that there is only one implementer.
https://llvm.org/bugs/show_bug.cgi?id=19043

We should prefer adding __attribute__s and __declspecs instead of keywords when possible.

In general, I can see the benefit of this rule, however in the case of calling conventions, I would think that keeping them all orthogonal is important. Having "most" calling conventions work one way, and a couple a different way seems like a bigger problem.

[rnk]

In D25204#581477, @erichkeane wrote:

In general, I can see the benefit of this rule, however in the case of calling conventions, I would think that keeping them all orthogonal is important. Having "most" calling conventions work one way, and a couple a different way seems like a bigger problem.

I agree. I don't think introducing a __regcall keyword is going to cause much real world breakage.

include/clang/Basic/Attr.td

815 ↗

(On Diff #75919)

Can we not add the single underscore keyword version? That's outside the implementer's namespace.

[erichkeane]

Remove single-underscore version of _regcall kw.

[erichkeane]

@rnk, @majnemer and @ABataev : I believe that I've done everything that has come up in review, and this passes all tests for the convention I can find. Do you guys see anything that is holding this patch up? What is otherwise the 'next step' in getting this into master?

[erichkeane]

It was brought to my attention that regcall isn't a calling convention that should cause MSVC to switch to C++ mangling. Switched that and updated the tests.

[rnk]

Looks good to me.

cfe/trunk/include/clang-c/Index.h

enum CXCallingConv {
   CXCallingConv_Default = 0,
   CXCallingConv_C = 1,
   CXCallingConv_X86StdCall = 2,
   CXCallingConv_X86FastCall = 3,
   CXCallingConv_X86ThisCall = 4,
   CXCallingConv_X86Pascal = 5,
   CXCallingConv_AAPCS = 6,
   CXCallingConv_AAPCS_VFP = 7,
-  /* Value 8 was PnaclCall, but it was never used, so it could safely be re-used. */
+  CXCallingConv_X86RegCall = 8,
   CXCallingConv_IntelOclBicc = 9,
   CXCallingConv_X86_64Win64 = 10,
   CXCallingConv_X86_64SysV = 11,
   CXCallingConv_X86VectorCall = 12,
   CXCallingConv_Swift = 13,
   CXCallingConv_PreserveMost = 14,
   CXCallingConv_PreserveAll = 15,
 

cfe/trunk/include/clang/AST/Type.h

protected:
   class FunctionTypeBitfields {
     friend class FunctionType;
     friend class FunctionProtoType;
 
     unsigned : NumTypeBits;
 
     /// Extra information which affects how the function is called, like
     /// regparm and the calling convention.
-    unsigned ExtInfo : 9;
+    unsigned ExtInfo : 10;
 
     /// Used only by FunctionProtoType, put here to pack with the
     /// other bitfields.
     /// The qualifiers are part of FunctionProtoType because...
     ///
     /// C++ 8.3.5p4: The return type, the parameter type list and the
     /// cv-qualifier-seq, [...], are part of the function type.
     unsigned TypeQuals : 4;
public:
   // * ASTContext::getFooType
   // * ASTContext::mergeFunctionTypes
   // * FunctionNoProtoType::Profile
   // * FunctionProtoType::Profile
   // * TypePrinter::PrintFunctionProto
   // * AST read and write
   // * Codegen
   class ExtInfo {
-    // Feel free to rearrange or add bits, but if you go over 9,
+    // Feel free to rearrange or add bits, but if you go over 10,
     // you'll need to adjust both the Bits field below and
     // Type::FunctionTypeBitfields.
 
     //   |  CC  |noreturn|produces|regparm|
-    //   |0 .. 3|   4    |    5   | 6 .. 8|
+    //   |0 .. 4|   5    |    6   | 7 .. 9|
     //
     // regparm is either 0 (no regparm attribute) or the regparm value+1.
-    enum { CallConvMask = 0xF };
-    enum { NoReturnMask = 0x10 };
-    enum { ProducesResultMask = 0x20 };
+    enum { CallConvMask = 0x1F };
+    enum { NoReturnMask = 0x20 };
+    enum { ProducesResultMask = 0x40 };
     enum { RegParmMask = ~(CallConvMask | NoReturnMask | ProducesResultMask),
-           RegParmOffset = 6 }; // Assumed to be the last field
+           RegParmOffset = 7 }; // Assumed to be the last field
 
     uint16_t Bits;
 
     ExtInfo(unsigned Bits) : Bits(static_cast<uint16_t>(Bits)) {}
 
     friend class FunctionType;
 
    public:
enum Kind {
     LastEnumOperandKind = attr_pcs_vfp,
 
     // No operand.
     attr_noreturn,
     attr_cdecl,
     attr_fastcall,
     attr_stdcall,
     attr_thiscall,
+    attr_regcall,
     attr_pascal,
     attr_swiftcall,
     attr_vectorcall,
     attr_inteloclbicc,
     attr_ms_abi,
     attr_sysv_abi,
     attr_preserve_most,
     attr_preserve_all,

cfe/trunk/include/clang/Basic/Attr.td

 
 def FastCall : InheritableAttr {
   let Spellings = [GCC<"fastcall">, Keyword<"__fastcall">,
                    Keyword<"_fastcall">];
 //  let Subjects = [Function, ObjCMethod];
   let Documentation = [FastCallDocs];
 }
 
+def RegCall : InheritableAttr {
+  let Spellings = [GCC<"regcall">, Keyword<"__regcall">];
+  let Documentation = [RegCallDocs];
+}
+
 def Final : InheritableAttr {
   let Spellings = [Keyword<"final">, Keyword<"sealed">];
   let Accessors = [Accessor<"isSpelledAsSealed", [Keyword<"sealed">]>];
   let SemaHandler = 0;
   let Documentation = [Undocumented];
 }
 
 def MinSize : InheritableAttr {

cfe/trunk/include/clang/Basic/AttrDocs.td

 convention is supported primarily for compatibility with existing code. Users
 seeking register parameters should use the ``regparm`` attribute, which does
 not require callee-cleanup.  See the documentation for `__fastcall`_ on MSDN.
 
 .. _`__fastcall`: http://msdn.microsoft.com/en-us/library/6xa169sk.aspx
   }];
 }
 
+def RegCallDocs : Documentation {
+  let Category = DocCatCallingConvs;
+  let Content = [{
+On x86 targets, this attribute changes the calling convention to
+`__regcall`_ convention. This convention aims to pass as many arguments
+as possible in registers. It also tries to utilize registers for the
+return value whenever it is possible.
+
+.. _`__regcall`: https://software.intel.com/en-us/node/693069
+  }];
+}
+
 def ThisCallDocs : Documentation {
   let Category = DocCatCallingConvs;
   let Content = [{
 On 32-bit x86 targets, this attribute changes the calling convention of a
 function to use ECX for the first parameter (typically the implicit ``this``
 parameter of C++ methods) and clear parameters off of the stack on return. This
 convention does not support variadic calls or unprototyped functions in C, and
 has no effect on x86_64 targets. See the documentation for `__thiscall`_ on

cfe/trunk/include/clang/Basic/Specifiers.h

enum CallingConv {
     CC_C,           // __attribute__((cdecl))
     CC_X86StdCall,  // __attribute__((stdcall))
     CC_X86FastCall, // __attribute__((fastcall))
     CC_X86ThisCall, // __attribute__((thiscall))
     CC_X86VectorCall, // __attribute__((vectorcall))
     CC_X86Pascal,   // __attribute__((pascal))
     CC_X86_64Win64, // __attribute__((ms_abi))
     CC_X86_64SysV,  // __attribute__((sysv_abi))
+    CC_X86RegCall, // __attribute__((regcall))
     CC_AAPCS,       // __attribute__((pcs("aapcs")))
     CC_AAPCS_VFP,   // __attribute__((pcs("aapcs-vfp")))
     CC_IntelOclBicc, // __attribute__((intel_ocl_bicc))
     CC_SpirFunction, // default for OpenCL functions on SPIR target
     CC_OpenCLKernel, // inferred for OpenCL kernels
     CC_Swift,        // __attribute__((swiftcall))
     CC_PreserveMost, // __attribute__((preserve_most))
     CC_PreserveAll,  // __attribute__((preserve_all))
   };
 
   /// \brief Checks whether the given calling convention supports variadic
   /// calls. Unprototyped calls also use the variadic call rules.
   inline bool supportsVariadicCall(CallingConv CC) {
     switch (CC) {
     case CC_X86StdCall:
     case CC_X86FastCall:
     case CC_X86ThisCall:
+    case CC_X86RegCall:
     case CC_X86Pascal:
     case CC_X86VectorCall:
     case CC_SpirFunction:
     case CC_OpenCLKernel:
     case CC_Swift:
       return false;
     default:
       return true;

cfe/trunk/include/clang/Basic/TokenKinds.def

 // disabled via '-fno-declspec'.
 KEYWORD(__declspec                  , 0)
 
 // Microsoft Extension.
 KEYWORD(__cdecl                     , KEYALL)
 KEYWORD(__stdcall                   , KEYALL)
 KEYWORD(__fastcall                  , KEYALL)
 KEYWORD(__thiscall                  , KEYALL)
+KEYWORD(__regcall                   , KEYALL)
 KEYWORD(__vectorcall                , KEYALL)
 KEYWORD(__forceinline               , KEYMS)
 KEYWORD(__unaligned                 , KEYMS)
 KEYWORD(__super                     , KEYMS)
 
 // OpenCL address space qualifiers
 KEYWORD(__global                    , KEYOPENCL)
 KEYWORD(__local                     , KEYOPENCL)

cfe/trunk/lib/AST/Expr.cpp

if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(CurrentDecl)) {
 
     if (IT == FuncSig) {
       switch (FT->getCallConv()) {
       case CC_C: POut << "__cdecl "; break;
       case CC_X86StdCall: POut << "__stdcall "; break;
       case CC_X86FastCall: POut << "__fastcall "; break;
       case CC_X86ThisCall: POut << "__thiscall "; break;
       case CC_X86VectorCall: POut << "__vectorcall "; break;
+      case CC_X86RegCall: POut << "__regcall "; break;
       // Only bother printing the conventions that MSVC knows about.
       default: break;
       }
     }
 
     FD->printQualifiedName(POut, Policy);
 
     POut << "(";

cfe/trunk/lib/AST/ItaniumMangle.cpp

void mangleUnqualifiedName(const NamedDecl *ND, DeclarationName Name,
                              const AbiTagList *AdditionalAbiTags);
   void mangleUnscopedName(const NamedDecl *ND,
                           const AbiTagList *AdditionalAbiTags);
   void mangleUnscopedTemplateName(const TemplateDecl *ND,
                                   const AbiTagList *AdditionalAbiTags);
   void mangleUnscopedTemplateName(TemplateName,
                                   const AbiTagList *AdditionalAbiTags);
   void mangleSourceName(const IdentifierInfo *II);
+  void mangleRegCallName(const IdentifierInfo *II);
   void mangleSourceNameWithAbiTags(
       const NamedDecl *ND, const AbiTagList *AdditionalAbiTags = nullptr);
   void mangleLocalName(const Decl *D,
                        const AbiTagList *AdditionalAbiTags);
   void mangleBlockForPrefix(const BlockDecl *Block);
   void mangleUnqualifiedBlock(const BlockDecl *Block);
   void mangleLambda(const CXXRecordDecl *Lambda);
   void mangleNestedName(const NamedDecl *ND, const DeclContext *DC,
if (II) {
       //   void test() { extern void foo(); }
       //   static void foo();
       // This naming convention is the same as that followed by GCC,
       // though it shouldn't actually matter.
       if (ND && ND->getFormalLinkage() == InternalLinkage &&
           getEffectiveDeclContext(ND)->isFileContext())
         Out << 'L';
 
+      auto *FD = dyn_cast<FunctionDecl>(ND);
+      bool IsRegCall = FD &&
+                       FD->getType()->castAs<FunctionType>()->getCallConv() ==
+                           clang::CC_X86RegCall;
+      if (IsRegCall)
+        mangleRegCallName(II);
+      else
         mangleSourceName(II);
+
       writeAbiTags(ND, AdditionalAbiTags);
       break;
     }
 
     // Otherwise, an anonymous entity.  We must have a declaration.
     assert(ND && "mangling empty name without declaration");
 
     if (const NamespaceDecl *NS = dyn_cast<NamespaceDecl>(ND)) {
case DeclarationName::CXXLiteralOperatorName:
     writeAbiTags(ND, AdditionalAbiTags);
     break;
 
   case DeclarationName::CXXUsingDirective:
     llvm_unreachable("Can't mangle a using directive name!");
   }
 }
 
+void CXXNameMangler::mangleRegCallName(const IdentifierInfo *II) {
+  // <source-name> ::= <positive length number> __regcall3__ <identifier>
+  // <number> ::= [n] <non-negative decimal integer>
+  // <identifier> ::= <unqualified source code identifier>
+  Out << II->getLength() + sizeof("__regcall3__") - 1 << "__regcall3__"
+      << II->getName();
+}
+
 void CXXNameMangler::mangleSourceName(const IdentifierInfo *II) {
   // <source-name> ::= <positive length number> <identifier>
   // <number> ::= [n] <non-negative decimal integer>
   // <identifier> ::= <unqualified source code identifier>
   Out << II->getLength() << II->getName();
 }
 
 void CXXNameMangler::mangleNestedName(const NamedDecl *ND,
StringRef CXXNameMangler::getCallingConvQualifierName(CallingConv CC) {
 
   case CC_X86StdCall:
   case CC_X86FastCall:
   case CC_X86ThisCall:
   case CC_X86VectorCall:
   case CC_X86Pascal:
   case CC_X86_64Win64:
   case CC_X86_64SysV:
+  case CC_X86RegCall:
   case CC_AAPCS:
   case CC_AAPCS_VFP:
   case CC_IntelOclBicc:
   case CC_SpirFunction:
   case CC_OpenCLKernel:
   case CC_PreserveMost:
   case CC_PreserveAll:
     // FIXME: we should be mangling all of the above.

cfe/trunk/lib/AST/Mangle.cpp

else
     Out << "__" << Outer << "_block_invoke_" << discriminator+1; 
 }
 
 void MangleContext::anchor() { }
 
 enum CCMangling {
   CCM_Other,
   CCM_Fast,
+  CCM_RegCall,
   CCM_Vector,
   CCM_Std
 };
 
 static bool isExternC(const NamedDecl *ND) {
   if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(ND))
     return FD->isExternC();
   return cast<VarDecl>(ND)->isExternC();
if (CC == CCM_Other || (MCXX && TI.getCXXABI() == TargetCXXABI::Microsoft)) {
     return;
   }
 
   Out << '\01';
   if (CC == CCM_Std)
     Out << '_';
   else if (CC == CCM_Fast)
     Out << '@';
+  else if (CC == CCM_RegCall)
+    Out << "__regcall3__";
 
   if (!MCXX)
     Out << D->getIdentifier()->getName();
   else if (const ObjCMethodDecl *OMD = dyn_cast<ObjCMethodDecl>(D))
     mangleObjCMethodName(OMD, Out);
   else
     mangleCXXName(D, Out);
 

cfe/trunk/lib/AST/MicrosoftMangle.cpp

void MicrosoftCXXNameMangler::mangleCallingConvention(CallingConv CC) {
   //                      ::= D # __export __pascal
   //                      ::= E # __thiscall
   //                      ::= F # __export __thiscall
   //                      ::= G # __stdcall
   //                      ::= H # __export __stdcall
   //                      ::= I # __fastcall
   //                      ::= J # __export __fastcall
   //                      ::= Q # __vectorcall
+  //                      ::= w # __regcall
   // The 'export' calling conventions are from a bygone era
   // (*cough*Win16*cough*) when functions were declared for export with
   // that keyword. (It didn't actually export them, it just made them so
   // that they could be in a DLL and somebody from another module could call
   // them.)
 
   switch (CC) {
     default:
       llvm_unreachable("Unsupported CC for mangling");
     case CC_X86_64Win64:
     case CC_X86_64SysV:
     case CC_C: Out << 'A'; break;
     case CC_X86Pascal: Out << 'C'; break;
     case CC_X86ThisCall: Out << 'E'; break;
     case CC_X86StdCall: Out << 'G'; break;
     case CC_X86FastCall: Out << 'I'; break;
     case CC_X86VectorCall: Out << 'Q'; break;
+    case CC_X86RegCall: Out << 'w'; break;
   }
 }
 void MicrosoftCXXNameMangler::mangleCallingConvention(const FunctionType *T) {
   mangleCallingConvention(T->getCallConv());
 }
 void MicrosoftCXXNameMangler::mangleThrowSpecification(
                                                 const FunctionProtoType *FT) {
   // <throw-spec> ::= Z # throw(...) (default)

cfe/trunk/lib/AST/Type.cpp

StringRef FunctionType::getNameForCallConv(CallingConv CC) {
   case CC_C: return "cdecl";
   case CC_X86StdCall: return "stdcall";
   case CC_X86FastCall: return "fastcall";
   case CC_X86ThisCall: return "thiscall";
   case CC_X86Pascal: return "pascal";
   case CC_X86VectorCall: return "vectorcall";
   case CC_X86_64Win64: return "ms_abi";
   case CC_X86_64SysV: return "sysv_abi";
+  case CC_X86RegCall : return "regcall";
   case CC_AAPCS: return "aapcs";
   case CC_AAPCS_VFP: return "aapcs-vfp";
   case CC_IntelOclBicc: return "intel_ocl_bicc";
   case CC_SpirFunction: return "spir_function";
   case CC_OpenCLKernel: return "opencl_kernel";
   case CC_Swift: return "swiftcall";
   case CC_PreserveMost: return "preserve_most";
   case CC_PreserveAll: return "preserve_all";
bool AttributedType::isQualifier() const {
   case AttributedType::attr_neon_polyvector_type:
   case AttributedType::attr_pcs:
   case AttributedType::attr_pcs_vfp:
   case AttributedType::attr_noreturn:
   case AttributedType::attr_cdecl:
   case AttributedType::attr_fastcall:
   case AttributedType::attr_stdcall:
   case AttributedType::attr_thiscall:
+  case AttributedType::attr_regcall:
   case AttributedType::attr_pascal:
   case AttributedType::attr_swiftcall:
   case AttributedType::attr_vectorcall:
   case AttributedType::attr_inteloclbicc:
   case AttributedType::attr_preserve_most:
   case AttributedType::attr_preserve_all:
   case AttributedType::attr_ms_abi:
   case AttributedType::attr_sysv_abi:
case attr_objc_kindof:
     return false;
 
   case attr_pcs:
   case attr_pcs_vfp:
   case attr_cdecl:
   case attr_fastcall:
   case attr_stdcall:
   case attr_thiscall:
+  case attr_regcall:
   case attr_swiftcall:
   case attr_vectorcall:
   case attr_pascal:
   case attr_ms_abi:
   case attr_sysv_abi:
   case attr_inteloclbicc:
   case attr_preserve_most:
   case attr_preserve_all:

cfe/trunk/lib/AST/TypePrinter.cpp

case CC_IntelOclBicc:
       OS << " __attribute__((intel_ocl_bicc))";
       break;
     case CC_X86_64Win64:
       OS << " __attribute__((ms_abi))";
       break;
     case CC_X86_64SysV:
       OS << " __attribute__((sysv_abi))";
       break;
+    case CC_X86RegCall:
+      OS << " __attribute__((regcall))";
+      break;
     case CC_SpirFunction:
     case CC_OpenCLKernel:
       // Do nothing. These CCs are not available as attributes.
       break;
     case CC_Swift:
       OS << " __attribute__((swiftcall))";
       break;
     case CC_PreserveMost:
void TypePrinter::printAttributedAfter(const AttributedType *T,
   case AttributedType::attr_fastcall: OS << "fastcall"; break;
   case AttributedType::attr_stdcall: OS << "stdcall"; break;
   case AttributedType::attr_thiscall: OS << "thiscall"; break;
   case AttributedType::attr_swiftcall: OS << "swiftcall"; break;
   case AttributedType::attr_vectorcall: OS << "vectorcall"; break;
   case AttributedType::attr_pascal: OS << "pascal"; break;
   case AttributedType::attr_ms_abi: OS << "ms_abi"; break;
   case AttributedType::attr_sysv_abi: OS << "sysv_abi"; break;
+  case AttributedType::attr_regcall: OS << "regcall"; break;
   case AttributedType::attr_pcs:
   case AttributedType::attr_pcs_vfp: {
     OS << "pcs(";
    QualType t = T->getEquivalentType();
    while (!t->isFunctionType())
      t = t->getPointeeType();
    OS << (t->getAs<FunctionType>()->getCallConv() == CC_AAPCS ?
          "\"aapcs\"" : "\"aapcs-vfp\"");

cfe/trunk/lib/Basic/Targets.cpp

public:
 
   CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
     // Most of the non-ARM calling conventions are i386 conventions.
     switch (CC) {
     case CC_X86ThisCall:
     case CC_X86FastCall:
     case CC_X86StdCall:
     case CC_X86VectorCall:
+    case CC_X86RegCall:
     case CC_C:
     case CC_Swift:
     case CC_X86Pascal:
     case CC_IntelOclBicc:
       return CCCR_OK;
     default:
       return CCCR_Warning;
     }
CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
     switch (CC) {
     case CC_C:
     case CC_Swift:
     case CC_X86VectorCall:
     case CC_IntelOclBicc:
     case CC_X86_64Win64:
     case CC_PreserveMost:
     case CC_PreserveAll:
+    case CC_X86RegCall:
       return CCCR_OK;
     default:
       return CCCR_Warning;
     }
   }
 
   CallingConv getDefaultCallingConv(CallingConvMethodType MT) const override {
     return CC_C;
CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
     case CC_X86ThisCall:
     case CC_X86FastCall:
       return CCCR_Ignore;
     case CC_C:
     case CC_X86VectorCall:
     case CC_IntelOclBicc:
     case CC_X86_64SysV:
     case CC_Swift:
+    case CC_X86RegCall:
       return CCCR_OK;
     default:
       return CCCR_Warning;
     }
   }
 };
 
 // x86-64 Windows Visual Studio target

cfe/trunk/lib/CodeGen/CGCall.cpp

 
 /***/
 
 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
   switch (CC) {
   default: return llvm::CallingConv::C;
   case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
   case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
+  case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
   case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
   case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64;
   case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
   case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
   case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
   case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
   // TODO: Add support for __pascal to LLVM.
   case CC_X86Pascal: return llvm::CallingConv::C;

 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
   // Set the appropriate calling convention for the Function.
   if (D->hasAttr<StdCallAttr>())
     return CC_X86StdCall;
 
   if (D->hasAttr<FastCallAttr>())
     return CC_X86FastCall;
 
+  if (D->hasAttr<RegCallAttr>())
+    return CC_X86RegCall;
+
   if (D->hasAttr<ThisCallAttr>())
     return CC_X86ThisCall;
 
   if (D->hasAttr<VectorCallAttr>())
     return CC_X86VectorCall;
 
   if (D->hasAttr<PascalAttr>())
     return CC_X86Pascal;

cfe/trunk/lib/CodeGen/CGDebugInfo.cpp

static unsigned getDwarfCC(CallingConv CC) {
   case CC_AAPCS:
   case CC_AAPCS_VFP:
   case CC_IntelOclBicc:
   case CC_SpirFunction:
   case CC_OpenCLKernel:
   case CC_Swift:
   case CC_PreserveMost:
   case CC_PreserveAll:
+  case CC_X86RegCall:
     return 0;
   }
   return 0;
 }
 
 llvm::DIType *CGDebugInfo::CreateType(const FunctionType *Ty,
                                       llvm::DIFile *Unit) {
   SmallVector<llvm::Metadata *, 16> EltTys;

cfe/trunk/lib/CodeGen/CodeGenModule.cpp

if (getCXXABI().getMangleContext().shouldMangleDeclName(ND)) {
     else if (const auto *D = dyn_cast<CXXDestructorDecl>(ND))
       getCXXABI().getMangleContext().mangleCXXDtor(D, GD.getDtorType(), Out);
     else
       getCXXABI().getMangleContext().mangleName(ND, Out);
     Str = Out.str();
   } else {
     IdentifierInfo *II = ND->getIdentifier();
     assert(II && "Attempt to mangle unnamed decl.");
+    const auto *FD = dyn_cast<FunctionDecl>(ND);
+
+    if (FD &&
+        FD->getType()->castAs<FunctionType>()->getCallConv() == CC_X86RegCall) {
+      llvm::raw_svector_ostream Out(Buffer);
+      Out << "__regcall3__" << II->getName();
+      Str = Out.str();
+    } else {
       Str = II->getName();
     }
+  }
 
   // Keep the first result in the case of a mangling collision.
   auto Result = Manglings.insert(std::make_pair(Str, GD));
   return FoundStr = Result.first->first();
 }
 
 StringRef CodeGenModule::getBlockMangledName(GlobalDecl GD,
                                              const BlockDecl *BD) {

cfe/trunk/lib/CodeGen/TargetInfo.cpp

 
 ABIArgInfo X86_32ABIInfo::classifyReturnType(QualType RetTy,
                                              CCState &State) const {
   if (RetTy->isVoidType())
     return ABIArgInfo::getIgnore();
 
   const Type *Base = nullptr;
   uint64_t NumElts = 0;
-  if (State.CC == llvm::CallingConv::X86_VectorCall &&
+  if ((State.CC == llvm::CallingConv::X86_VectorCall ||
+       State.CC == llvm::CallingConv::X86_RegCall) &&
       isHomogeneousAggregate(RetTy, Base, NumElts)) {
     // The LLVM struct type for such an aggregate should lower properly.
     return ABIArgInfo::getDirect();
   }
 
   if (const VectorType *VT = RetTy->getAs<VectorType>()) {
     // On Darwin, some vectors are returned in registers.
     if (IsDarwinVectorABI) {
bool X86_32ABIInfo::shouldAggregateUseDirect(QualType Ty, CCState &State,
 
   if (!updateFreeRegs(Ty, State))
     return false;
 
   if (IsMCUABI)
     return true;
 
   if (State.CC == llvm::CallingConv::X86_FastCall ||
-      State.CC == llvm::CallingConv::X86_VectorCall) {
+      State.CC == llvm::CallingConv::X86_VectorCall ||
+      State.CC == llvm::CallingConv::X86_RegCall) {
     if (getContext().getTypeSize(Ty) <= 32 && State.FreeRegs)
       NeedsPadding = true;
 
     return false;
   }
 
   return true;
 }
 
 bool X86_32ABIInfo::shouldPrimitiveUseInReg(QualType Ty, CCState &State) const {
   if (!updateFreeRegs(Ty, State))
     return false;
 
   if (IsMCUABI)
     return false;
 
   if (State.CC == llvm::CallingConv::X86_FastCall ||
-      State.CC == llvm::CallingConv::X86_VectorCall) {
+      State.CC == llvm::CallingConv::X86_VectorCall ||
+      State.CC == llvm::CallingConv::X86_RegCall) {
     if (getContext().getTypeSize(Ty) > 32)
       return false;
 
     return (Ty->isIntegralOrEnumerationType() || Ty->isPointerType() || 
         Ty->isReferenceType());
   }
 
   return true;
if (RAA == CGCXXABI::RAA_Indirect) {
       return ABIArgInfo::getInAlloca(/*FieldIndex=*/0);
     }
   }
 
   // vectorcall adds the concept of a homogenous vector aggregate, similar
   // to other targets.
   const Type *Base = nullptr;
   uint64_t NumElts = 0;
-  if (State.CC == llvm::CallingConv::X86_VectorCall &&
+  if ((State.CC == llvm::CallingConv::X86_VectorCall ||
+       State.CC == llvm::CallingConv::X86_RegCall) &&
       isHomogeneousAggregate(Ty, Base, NumElts)) {
     if (State.FreeSSERegs >= NumElts) {
       State.FreeSSERegs -= NumElts;
       if (Ty->isBuiltinType() || Ty->isVectorType())
         return ABIArgInfo::getDirect();
       return ABIArgInfo::getExpand();
     }
     return getIndirectResult(Ty, /*ByVal=*/false, State);
if (isAggregateTypeForABI(Ty)) {
     // LLVM backend isn't smart enough to remove byval, which inhibits many
     // optimizations.
     // Don't do this for the MCU if there are still free integer registers
     // (see X86_64 ABI for full explanation).
     if (getContext().getTypeSize(Ty) <= 4 * 32 &&
         (!IsMCUABI || State.FreeRegs == 0) && canExpandIndirectArgument(Ty))
       return ABIArgInfo::getExpandWithPadding(
           State.CC == llvm::CallingConv::X86_FastCall ||
-              State.CC == llvm::CallingConv::X86_VectorCall,
+              State.CC == llvm::CallingConv::X86_VectorCall ||
+              State.CC == llvm::CallingConv::X86_RegCall,
           PaddingType);
 
     return getIndirectResult(Ty, true, State);
   }
 
   if (const VectorType *VT = Ty->getAs<VectorType>()) {
     // On Darwin, some vectors are passed in memory, we handle this by passing
     // it as an i8/i16/i32/i64.
if (IsMCUABI)
     State.FreeRegs = 3;
   else if (State.CC == llvm::CallingConv::X86_FastCall)
     State.FreeRegs = 2;
   else if (State.CC == llvm::CallingConv::X86_VectorCall) {
     State.FreeRegs = 2;
     State.FreeSSERegs = 6;
   } else if (FI.getHasRegParm())
     State.FreeRegs = FI.getRegParm();
-  else
+  else if (State.CC == llvm::CallingConv::X86_RegCall) {
+    State.FreeRegs = 5;
+    State.FreeSSERegs = 8;
+  } else
     State.FreeRegs = DefaultNumRegisterParameters;
 
   if (!getCXXABI().classifyReturnType(FI)) {
     FI.getReturnInfo() = classifyReturnType(FI.getReturnType(), State);
   } else if (FI.getReturnInfo().isIndirect()) {
     // The C++ ABI is not aware of register usage, so we have to check if the
     // return value was sret and put it in a register ourselves if appropriate.
     if (State.FreeRegs) {
class X86_64ABIInfo : public SwiftABIInfo {
   /// such that the argument will be passed in memory.
   ///
   /// \param freeIntRegs - The number of free integer registers remaining
   /// available.
   ABIArgInfo getIndirectResult(QualType Ty, unsigned freeIntRegs) const;
 
   ABIArgInfo classifyReturnType(QualType RetTy) const;
 
-  ABIArgInfo classifyArgumentType(QualType Ty,
-                                  unsigned freeIntRegs,
-                                  unsigned &neededInt,
-                                  unsigned &neededSSE,
+  ABIArgInfo classifyArgumentType(QualType Ty, unsigned freeIntRegs,
+                                  unsigned &neededInt, unsigned &neededSSE,
                                   bool isNamedArg) const;
 
+  ABIArgInfo classifyRegCallStructType(QualType Ty, unsigned &NeededInt,
+                                       unsigned &NeededSSE) const;
+
+  ABIArgInfo classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt,
+                                           unsigned &NeededSSE) const;
+
   bool IsIllegalVectorType(QualType Ty) const;
 
   /// The 0.98 ABI revision clarified a lot of ambiguities,
   /// unfortunately in ways that were not always consistent with
   /// certain previous compilers.  In particular, platforms which
   /// required strict binary compatibility with older versions of GCC
   /// may need to exempt themselves.
   bool honorsRevision0_98() const {
ABIArgInfo X86_64ABIInfo::classifyArgumentType(
   // known to pass in the high eightbyte of the result.  We do this by forming a
   // first class struct aggregate with the high and low part: {low, high}
   if (HighPart)
     ResType = GetX86_64ByValArgumentPair(ResType, HighPart, getDataLayout());
 
   return ABIArgInfo::getDirect(ResType);
 }
 
+ABIArgInfo
+X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt,
+                                             unsigned &NeededSSE) const {
+  auto RT = Ty->getAs<RecordType>();
+  assert(RT && "classifyRegCallStructType only valid with struct types");
+
+  if (RT->getDecl()->hasFlexibleArrayMember())
+    return getIndirectReturnResult(Ty);
+
+  // Sum up bases
+  if (auto CXXRD = dyn_cast<CXXRecordDecl>(RT->getDecl())) {
+    if (CXXRD->isDynamicClass()) {
+      NeededInt = NeededSSE = 0;
+      return getIndirectReturnResult(Ty);
+    }
+
+    for (const auto &I : CXXRD->bases())
+      if (classifyRegCallStructTypeImpl(I.getType(), NeededInt, NeededSSE)
+              .isIndirect()) {
+        NeededInt = NeededSSE = 0;
+        return getIndirectReturnResult(Ty);
+      }
+  }
+
+  // Sum up members
+  for (const auto *FD : RT->getDecl()->fields()) {
+    if (FD->getType()->isRecordType() && !FD->getType()->isUnionType()) {
+      if (classifyRegCallStructTypeImpl(FD->getType(), NeededInt, NeededSSE)
+              .isIndirect()) {
+        NeededInt = NeededSSE = 0;
+        return getIndirectReturnResult(Ty);
+      }
+    } else {
+      unsigned LocalNeededInt, LocalNeededSSE;
+      if (classifyArgumentType(FD->getType(), UINT_MAX, LocalNeededInt,
+                               LocalNeededSSE, true)
+              .isIndirect()) {
+        NeededInt = NeededSSE = 0;
+        return getIndirectReturnResult(Ty);
+      }
+      NeededInt += LocalNeededInt;
+      NeededSSE += LocalNeededSSE;
+    }
+  }
+
+  return ABIArgInfo::getDirect();
+}
+
+ABIArgInfo X86_64ABIInfo::classifyRegCallStructType(QualType Ty,
+                                                    unsigned &NeededInt,
+                                                    unsigned &NeededSSE) const {
+
+  NeededInt = 0;
+  NeededSSE = 0;
+
+  return classifyRegCallStructTypeImpl(Ty, NeededInt, NeededSSE);
+}
+
 void X86_64ABIInfo::computeInfo(CGFunctionInfo &FI) const {
 
-  if (!getCXXABI().classifyReturnType(FI))
-    FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
+  bool IsRegCall = FI.getCallingConvention() == llvm::CallingConv::X86_RegCall;
 
   // Keep track of the number of assigned registers.
-  unsigned freeIntRegs = 6, freeSSERegs = 8;
+  unsigned FreeIntRegs = IsRegCall ? 11 : 6;
+  unsigned FreeSSERegs = IsRegCall ? 16 : 8;
+  unsigned NeededInt, NeededSSE;
+
+  if (IsRegCall && FI.getReturnType()->getTypePtr()->isRecordType() &&
+      !FI.getReturnType()->getTypePtr()->isUnionType()) {
+    FI.getReturnInfo() =
+        classifyRegCallStructType(FI.getReturnType(), NeededInt, NeededSSE);
+    if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) {
+      FreeIntRegs -= NeededInt;
+      FreeSSERegs -= NeededSSE;
+    } else {
+      FI.getReturnInfo() = getIndirectReturnResult(FI.getReturnType());
+    }
+  } else if (!getCXXABI().classifyReturnType(FI))
+    FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
 
   // If the return value is indirect, then the hidden argument is consuming one
   // integer register.
   if (FI.getReturnInfo().isIndirect())
-    --freeIntRegs;
+    --FreeIntRegs;
 
   // The chain argument effectively gives us another free register.
   if (FI.isChainCall())
-    ++freeIntRegs;
+    ++FreeIntRegs;
 
   unsigned NumRequiredArgs = FI.getNumRequiredArgs();
   // AMD64-ABI 3.2.3p3: Once arguments are classified, the registers
   // get assigned (in left-to-right order) for passing as follows...
   unsigned ArgNo = 0;
   for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end();
        it != ie; ++it, ++ArgNo) {
     bool IsNamedArg = ArgNo < NumRequiredArgs;
 
-    unsigned neededInt, neededSSE;
-    it->info = classifyArgumentType(it->type, freeIntRegs, neededInt,
-                                    neededSSE, IsNamedArg);
+    if (IsRegCall && it->type->isStructureOrClassType())
+      it->info = classifyRegCallStructType(it->type, NeededInt, NeededSSE);
+    else
+      it->info = classifyArgumentType(it->type, FreeIntRegs, NeededInt,
+                                      NeededSSE, IsNamedArg);
 
     // AMD64-ABI 3.2.3p3: If there are no registers available for any
     // eightbyte of an argument, the whole argument is passed on the
     // stack. If registers have already been assigned for some
     // eightbytes of such an argument, the assignments get reverted.
-    if (freeIntRegs >= neededInt && freeSSERegs >= neededSSE) {
-      freeIntRegs -= neededInt;
-      freeSSERegs -= neededSSE;
+    if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) {
+      FreeIntRegs -= NeededInt;
+      FreeSSERegs -= NeededSSE;
     } else {
-      it->info = getIndirectResult(it->type, freeIntRegs);
+      it->info = getIndirectResult(it->type, FreeIntRegs);
     }
   }
 }
 
 static Address EmitX86_64VAArgFromMemory(CodeGenFunction &CGF,
                                          Address VAListAddr, QualType Ty) {
   Address overflow_arg_area_p = CGF.Builder.CreateStructGEP(
       VAListAddr, 2, CharUnits::fromQuantity(8), "overflow_arg_area_p");
ABIArgInfo WinX86_64ABIInfo::classify(QualType Ty, unsigned &FreeSSERegs,
   }
 
   return ABIArgInfo::getDirect();
 }
 
 void WinX86_64ABIInfo::computeInfo(CGFunctionInfo &FI) const {
   bool IsVectorCall =
       FI.getCallingConvention() == llvm::CallingConv::X86_VectorCall;
+  bool IsRegCall = FI.getCallingConvention() == llvm::CallingConv::X86_RegCall;
 
+  unsigned FreeSSERegs = 0;
+  if (IsVectorCall) {
     // We can use up to 4 SSE return registers with vectorcall.
-  unsigned FreeSSERegs = IsVectorCall ? 4 : 0;
+    FreeSSERegs = 4;
+  } else if (IsRegCall) {
+    // RegCall gives us 16 SSE registers.
+    FreeSSERegs = 16;
+  }
+
   if (!getCXXABI().classifyReturnType(FI))
     FI.getReturnInfo() = classify(FI.getReturnType(), FreeSSERegs, true);
 
+  if (IsVectorCall) {
     // We can use up to 6 SSE register parameters with vectorcall.
-  FreeSSERegs = IsVectorCall ? 6 : 0;
+    FreeSSERegs = 6;
+  } else if (IsRegCall) {
+    FreeSSERegs = 16;
+  }
+
   for (auto &I : FI.arguments())
     I.info = classify(I.type, FreeSSERegs, false);
 }
 
 Address WinX86_64ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                                     QualType Ty) const {
 
   bool IsIndirect = false;

cfe/trunk/lib/Parse/ParseDecl.cpp

 
 void Parser::ParseMicrosoftTypeAttributes(ParsedAttributes &attrs) {
   // Treat these like attributes
   while (true) {
     switch (Tok.getKind()) {
     case tok::kw___fastcall:
     case tok::kw___stdcall:
     case tok::kw___thiscall:
+    case tok::kw___regcall:
     case tok::kw___cdecl:
     case tok::kw___vectorcall:
     case tok::kw___ptr64:
     case tok::kw___w64:
     case tok::kw___ptr32:
     case tok::kw___sptr:
     case tok::kw___uptr: {
       IdentifierInfo *AttrName = Tok.getIdentifierInfo();
while (1) {
     case tok::kw___uptr:
     case tok::kw___ptr64:
     case tok::kw___ptr32:
     case tok::kw___w64:
     case tok::kw___cdecl:
     case tok::kw___stdcall:
     case tok::kw___fastcall:
     case tok::kw___thiscall:
+    case tok::kw___regcall:
     case tok::kw___vectorcall:
       ParseMicrosoftTypeAttributes(DS.getAttributes());
       continue;
 
     // Borland single token adornments.
     case tok::kw___pascal:
       ParseBorlandTypeAttributes(DS.getAttributes());
       continue;
case tok::annot_typename:
     // GNU ObjC bizarre protocol extension: <proto1,proto2> with implicit 'id'.
   case tok::less:
     return getLangOpts().ObjC1;
 
   case tok::kw___cdecl:
   case tok::kw___stdcall:
   case tok::kw___fastcall:
   case tok::kw___thiscall:
+  case tok::kw___regcall:
   case tok::kw___vectorcall:
   case tok::kw___w64:
   case tok::kw___ptr64:
   case tok::kw___ptr32:
   case tok::kw___pascal:
   case tok::kw___unaligned:
 
   case tok::kw__Nonnull:
case tok::annot_typename:
     return !DisambiguatingWithExpression ||
            !isStartOfObjCClassMessageMissingOpenBracket();
 
   case tok::kw___declspec:
   case tok::kw___cdecl:
   case tok::kw___stdcall:
   case tok::kw___fastcall:
   case tok::kw___thiscall:
+  case tok::kw___regcall:
   case tok::kw___vectorcall:
   case tok::kw___w64:
   case tok::kw___sptr:
   case tok::kw___uptr:
   case tok::kw___ptr64:
   case tok::kw___ptr32:
   case tok::kw___forceinline:
   case tok::kw___pascal:
while (1) {
     case tok::kw___sptr:
     case tok::kw___w64:
     case tok::kw___ptr64:
     case tok::kw___ptr32:
     case tok::kw___cdecl:
     case tok::kw___stdcall:
     case tok::kw___fastcall:
     case tok::kw___thiscall:
+    case tok::kw___regcall:
     case tok::kw___vectorcall:
       if (AttrReqs & AR_DeclspecAttributesParsed) {
         ParseMicrosoftTypeAttributes(DS.getAttributes());
         continue;
       }
       goto DoneWithTypeQuals;
     case tok::kw___pascal:
       if (AttrReqs & AR_VendorAttributesParsed) {

cfe/trunk/lib/Parse/ParseTentative.cpp

if (mayBeAbstract &&
       if (TPR != TPResult::Ambiguous)
         return TPR;
     } else {
       // '(' declarator ')'
       // '(' attributes declarator ')'
       // '(' abstract-declarator ')'
       if (Tok.isOneOf(tok::kw___attribute, tok::kw___declspec, tok::kw___cdecl,
                       tok::kw___stdcall, tok::kw___fastcall, tok::kw___thiscall,
-                      tok::kw___vectorcall))
+                      tok::kw___regcall, tok::kw___vectorcall))
         return TPResult::True; // attributes indicate declaration
       TPResult TPR = TryParseDeclarator(mayBeAbstract, mayHaveIdentifier);
       if (TPR != TPResult::Ambiguous)
         return TPR;
       if (Tok.isNot(tok::r_paren))
         return TPResult::False;
       ConsumeParen();
     }
#include "clang/Basic/TokenKinds.def"
   case tok::kw_thread_local:
   case tok::kw__Thread_local:
   case tok::kw_typeof:
   case tok::kw___underlying_type:
   case tok::kw___cdecl:
   case tok::kw___stdcall:
   case tok::kw___fastcall:
   case tok::kw___thiscall:
+  case tok::kw___regcall:
   case tok::kw___vectorcall:
   case tok::kw___unaligned:
   case tok::kw___vector:
   case tok::kw___pixel:
   case tok::kw___bool:
   case tok::kw__Atomic:
 #define GENERIC_IMAGE_TYPE(ImgType, Id) case tok::kw_##ImgType##_t:
 #include "clang/Basic/OpenCLImageTypes.def"
case tok::kw___auto_type:
     return TPResult::True;
 
     // Microsoft
   case tok::kw___declspec:
   case tok::kw___cdecl:
   case tok::kw___stdcall:
   case tok::kw___fastcall:
   case tok::kw___thiscall:
+  case tok::kw___regcall:
   case tok::kw___vectorcall:
   case tok::kw___w64:
   case tok::kw___sptr:
   case tok::kw___uptr:
   case tok::kw___ptr64:
   case tok::kw___ptr32:
   case tok::kw___forceinline:
   case tok::kw___unaligned:

cfe/trunk/lib/Sema/SemaDeclAttr.cpp

D->addAttr(::new (S.Context)
                MSABIAttr(Attr.getRange(), S.Context,
                          Attr.getAttributeSpellingListIndex()));
     return;
   case AttributeList::AT_SysVABI:
     D->addAttr(::new (S.Context)
                SysVABIAttr(Attr.getRange(), S.Context,
                            Attr.getAttributeSpellingListIndex()));
     return;
+  case AttributeList::AT_RegCall:
+    D->addAttr(::new (S.Context) RegCallAttr(
+        Attr.getRange(), S.Context, Attr.getAttributeSpellingListIndex()));
+    return;
   case AttributeList::AT_Pcs: {
     PcsAttr::PCSType PCS;
     switch (CC) {
     case CC_AAPCS:
       PCS = PcsAttr::AAPCS;
       break;
     case CC_AAPCS_VFP:
       PCS = PcsAttr::AAPCS_VFP;
bool Sema::CheckCallingConvAttr(const AttributeList &attr, CallingConv &CC,
   switch (attr.getKind()) {
   case AttributeList::AT_CDecl: CC = CC_C; break;
   case AttributeList::AT_FastCall: CC = CC_X86FastCall; break;
   case AttributeList::AT_StdCall: CC = CC_X86StdCall; break;
   case AttributeList::AT_ThisCall: CC = CC_X86ThisCall; break;
   case AttributeList::AT_Pascal: CC = CC_X86Pascal; break;
   case AttributeList::AT_SwiftCall: CC = CC_Swift; break;
   case AttributeList::AT_VectorCall: CC = CC_X86VectorCall; break;
+  case AttributeList::AT_RegCall: CC = CC_X86RegCall; break;
   case AttributeList::AT_MSABI:
     CC = Context.getTargetInfo().getTriple().isOSWindows() ? CC_C :
                                                              CC_X86_64Win64;
     break;
   case AttributeList::AT_SysVABI:
     CC = Context.getTargetInfo().getTriple().isOSWindows() ? CC_X86_64SysV :
                                                              CC_C;
     break;
static void ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D,
   case AttributeList::AT_NoInstrumentFunction: // Interacts with -pg.
     handleSimpleAttribute<NoInstrumentFunctionAttr>(S, D, Attr);
     break;
   case AttributeList::AT_StdCall:
   case AttributeList::AT_CDecl:
   case AttributeList::AT_FastCall:
   case AttributeList::AT_ThisCall:
   case AttributeList::AT_Pascal:
+  case AttributeList::AT_RegCall:
   case AttributeList::AT_SwiftCall:
   case AttributeList::AT_VectorCall:
   case AttributeList::AT_MSABI:
   case AttributeList::AT_SysVABI:
   case AttributeList::AT_Pcs:
   case AttributeList::AT_IntelOclBicc:
   case AttributeList::AT_PreserveMost:
   case AttributeList::AT_PreserveAll:

cfe/trunk/lib/Sema/SemaType.cpp

#define OBJC_POINTER_TYPE_ATTRS_CASELIST \
     case AttributeList::AT_ObjCOwnership
 
 // Calling convention attributes.
 #define CALLING_CONV_ATTRS_CASELIST \
     case AttributeList::AT_CDecl: \
     case AttributeList::AT_FastCall: \
     case AttributeList::AT_StdCall: \
     case AttributeList::AT_ThisCall: \
+    case AttributeList::AT_RegCall: \
     case AttributeList::AT_Pascal: \
     case AttributeList::AT_SwiftCall: \
     case AttributeList::AT_VectorCall: \
     case AttributeList::AT_MSABI: \
     case AttributeList::AT_SysVABI: \
     case AttributeList::AT_Pcs: \
     case AttributeList::AT_IntelOclBicc: \
     case AttributeList::AT_PreserveMost: \
static AttributeList::Kind getAttrListKind(AttributedType::Kind kind) {
   case AttributedType::attr_cdecl:
     return AttributeList::AT_CDecl;
   case AttributedType::attr_fastcall:
     return AttributeList::AT_FastCall;
   case AttributedType::attr_stdcall:
     return AttributeList::AT_StdCall;
   case AttributedType::attr_thiscall:
     return AttributeList::AT_ThisCall;
+  case AttributedType::attr_regcall:
+    return AttributeList::AT_RegCall;
   case AttributedType::attr_pascal:
     return AttributeList::AT_Pascal;
   case AttributedType::attr_swiftcall:
     return AttributeList::AT_SwiftCall;
   case AttributedType::attr_vectorcall:
     return AttributeList::AT_VectorCall;
   case AttributedType::attr_pcs:
   case AttributedType::attr_pcs_vfp:
static AttributedType::Kind getCCTypeAttrKind(AttributeList &Attr) {
   case AttributeList::AT_CDecl:
     return AttributedType::attr_cdecl;
   case AttributeList::AT_FastCall:
     return AttributedType::attr_fastcall;
   case AttributeList::AT_StdCall:
     return AttributedType::attr_stdcall;
   case AttributeList::AT_ThisCall:
     return AttributedType::attr_thiscall;
+  case AttributeList::AT_RegCall:
+    return AttributedType::attr_regcall;
   case AttributeList::AT_Pascal:
     return AttributedType::attr_pascal;
   case AttributeList::AT_SwiftCall:
     return AttributedType::attr_swiftcall;
   case AttributeList::AT_VectorCall:
     return AttributedType::attr_vectorcall;
   case AttributeList::AT_Pcs: {
     // The attribute may have had a fixit applied where we treated an

cfe/trunk/test/CodeGen/regcall.c

+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -triple=i386-pc-win32       | FileCheck %s --check-prefix=Win32
+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -triple=x86_64-pc-win32     | FileCheck %s --check-prefix=Win64
+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -triple=i386-pc-linux-gnu   | FileCheck %s --check-prefix=Lin32
+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -triple=x86_64-pc-linux-gnu | FileCheck %s --check-prefix=Lin64
+
+#include <xmmintrin.h>
+
+void __regcall v1(int a, int b) {}
+// Win32: define x86_regcallcc void @__regcall3__v1(i32 inreg %a, i32 inreg %b)
+// Win64: define x86_regcallcc void @__regcall3__v1(i32 %a, i32 %b)
+// Lin32: define x86_regcallcc void @__regcall3__v1(i32 inreg %a, i32 inreg %b)
+// Lin64: define x86_regcallcc void @__regcall3__v1(i32 %a, i32 %b)
+
+void __attribute__((regcall)) v1b(int a, int b) {}
+// Win32: define x86_regcallcc void @__regcall3__v1b(i32 inreg %a, i32 inreg %b)
+// Win64: define x86_regcallcc void @__regcall3__v1b(i32 %a, i32 %b)
+// Lin32: define x86_regcallcc void @__regcall3__v1b(i32 inreg %a, i32 inreg %b)
+// Lin64: define x86_regcallcc void @__regcall3__v1b(i32 %a, i32 %b)
+
+void __regcall v2(char a, char b) {}
+// Win32: define x86_regcallcc void @__regcall3__v2(i8 inreg signext %a, i8 inreg signext %b)
+// Win64: define x86_regcallcc void @__regcall3__v2(i8 %a, i8 %b)
+// Lin32: define x86_regcallcc void @__regcall3__v2(i8 inreg signext %a, i8 inreg signext %b)
+// Lin64: define x86_regcallcc void @__regcall3__v2(i8 signext %a, i8 signext %b)
+
+struct Small { int x; };
+void __regcall v3(int a, struct Small b, int c) {}
+// Win32: define x86_regcallcc void @__regcall3__v3(i32 inreg %a, i32 %b.0, i32 inreg %c)
+// Win64: define x86_regcallcc void @__regcall3__v3(i32 %a, i32 %b.coerce, i32 %c)
+// Lin32: define x86_regcallcc void @__regcall3__v3(i32 inreg %a, i32 inreg, i32 %b.0, i32 inreg %c)
+// Lin64: define x86_regcallcc void @__regcall3__v3(i32 %a, i32 %b.coerce, i32 %c)
+
+struct Large { int a[5]; };
+void __regcall v4(int a, struct Large b, int c) {}
+// Win32: define x86_regcallcc void @__regcall3__v4(i32 inreg %a, %struct.Large* byval align 4 %b, i32 inreg %c)
+// Win64: define x86_regcallcc void @__regcall3__v4(i32 %a, %struct.Large* %b, i32 %c)
+// Lin32: define x86_regcallcc void @__regcall3__v4(i32 inreg %a, %struct.Large* byval align 4 %b, i32 %c)
+// Lin64: define x86_regcallcc void @__regcall3__v4(i32 %a, [5 x i32] %b.coerce, i32 %c)
+
+struct HFA2 { double x, y; };
+struct HFA4 { double w, x, y, z; };
+struct HFA5 { double v, w, x, y, z; };
+
+void __regcall hfa1(int a, struct HFA4 b, int c) {}
+// Win32: define x86_regcallcc void @__regcall3__hfa1(i32 inreg %a, double %b.0, double %b.1, double %b.2, double %b.3, i32 inreg %c)
+// Win64: define x86_regcallcc void @__regcall3__hfa1(i32 %a, double %b.0, double %b.1, double %b.2, double %b.3, i32 %c)
+// Lin32: define x86_regcallcc void @__regcall3__hfa1(i32 inreg %a, double %b.0, double %b.1, double %b.2, double %b.3, i32 inreg %c)
+// Lin64: define x86_regcallcc void @__regcall3__hfa1(i32 %a, double %b.coerce0, double %b.coerce1, double %b.coerce2, double %b.coerce3, i32 %c)
+
+// HFAs that would require more than six total SSE registers are passed
+// indirectly. Additional vector arguments can consume the rest of the SSE
+// registers.
+void __regcall hfa2(struct HFA4 a, struct HFA4 b, double c) {}
+// Win32: define x86_regcallcc void @__regcall3__hfa2(double %a.0, double %a.1, double %a.2, double %a.3, double %b.0, double %b.1, double %b.2, double %b.3, double* inreg)
+// Win64: define x86_regcallcc void @__regcall3__hfa2(double %a.0, double %a.1, double %a.2, double %a.3, double %b.0, double %b.1, double %b.2, double %b.3, double %c)
+// Lin32: define x86_regcallcc void @__regcall3__hfa2(double %a.0, double %a.1, double %a.2, double %a.3, double %b.0, double %b.1, double %b.2, double %b.3, double* inreg)
+// Lin64: define x86_regcallcc void @__regcall3__hfa2(double %a.coerce0, double %a.coerce1, double %a.coerce2, double %a.coerce3, double %b.coerce0, double %b.coerce1, double %b.coerce2, double %b.coerce3, double %c)
+
+// Ensure that we pass builtin types directly while counting them against the
+// SSE register usage.
+void __regcall hfa3(double a, double b, double c, double d, double e, struct HFA2 f) {}
+// Win32: define x86_regcallcc void @__regcall3__hfa3(double %a, double %b, double %c, double %d, double %e, double %f.0, double %f.1)
+// Win64: define x86_regcallcc void @__regcall3__hfa3(double %a, double %b, double %c, double %d, double %e, double %f.0, double %f.1)
+// Lin32: define x86_regcallcc void @__regcall3__hfa3(double %a, double %b, double %c, double %d, double %e, double %f.0, double %f.1)
+// Lin64: define x86_regcallcc void @__regcall3__hfa3(double %a, double %b, double %c, double %d, double %e, double %f.coerce0, double %f.coerce1)
+
+// Aggregates with more than four elements are not HFAs and are passed byval.
+// Because they are not classified as homogeneous, they don't get special
+// handling to ensure alignment.
+void __regcall hfa4(struct HFA5 a) {}
+// Win32: define x86_regcallcc void @__regcall3__hfa4(%struct.HFA5* byval align 4)
+// Win64: define x86_regcallcc void @__regcall3__hfa4(%struct.HFA5* %a)
+// Lin32: define x86_regcallcc void @__regcall3__hfa4(%struct.HFA5* byval align 4 %a)
+// Lin64: define x86_regcallcc void @__regcall3__hfa4(double %a.coerce0, double %a.coerce1, double %a.coerce2, double %a.coerce3, double %a.coerce4)
+
+// Return HFAs of 4 or fewer elements in registers.
+static struct HFA2 g_hfa2;
+struct HFA2 __regcall hfa5(void) { return g_hfa2; }
+// Win32: define x86_regcallcc %struct.HFA2 @__regcall3__hfa5()
+// Win64: define x86_regcallcc %struct.HFA2 @__regcall3__hfa5()
+// Lin32: define x86_regcallcc %struct.HFA2 @__regcall3__hfa5()
+// Lin64: define x86_regcallcc %struct.HFA2 @__regcall3__hfa5()
+
+typedef float __attribute__((vector_size(16))) v4f32;
+struct HVA2 { v4f32 x, y; };
+struct HVA4 { v4f32 w, x, y, z; };
+
+void __regcall hva1(int a, struct HVA4 b, int c) {}
+// Win32: define x86_regcallcc void @__regcall3__hva1(i32 inreg %a, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, i32 inreg %c)
+// Win64: define x86_regcallcc void @__regcall3__hva1(i32 %a, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, i32 %c)
+// Lin32: define x86_regcallcc void @__regcall3__hva1(i32 inreg %a, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, i32 inreg %c)
+// Lin64: define x86_regcallcc void @__regcall3__hva1(i32 %a, <4 x float> %b.coerce0, <4 x float> %b.coerce1, <4 x float> %b.coerce2, <4 x float> %b.coerce3, i32 %c)
+
+void __regcall hva2(struct HVA4 a, struct HVA4 b, v4f32 c) {}
+// Win32: define x86_regcallcc void @__regcall3__hva2(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, <4 x float>* inreg)
+// Win64: define x86_regcallcc void @__regcall3__hva2(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, <4 x float> %c)
+// Lin32: define x86_regcallcc void @__regcall3__hva2(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, <4 x float>* inreg)
+// Lin64: define x86_regcallcc void @__regcall3__hva2(<4 x float> %a.coerce0, <4 x float> %a.coerce1, <4 x float> %a.coerce2, <4 x float> %a.coerce3, <4 x float> %b.coerce0, <4 x float> %b.coerce1, <4 x float> %b.coerce2, <4 x float> %b.coerce3, <4 x float> %c)
+
+void __regcall hva3(v4f32 a, v4f32 b, v4f32 c, v4f32 d, v4f32 e, struct HVA2 f) {}
+// Win32: define x86_regcallcc void @__regcall3__hva3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e, <4 x float> %f.0, <4 x float> %f.1)
+// Win64: define x86_regcallcc void @__regcall3__hva3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e, <4 x float> %f.0, <4 x float> %f.1)
+// Lin32: define x86_regcallcc void @__regcall3__hva3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e, <4 x float> %f.0, <4 x float> %f.1)
+// Lin64: define x86_regcallcc void @__regcall3__hva3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e, <4 x float> %f.coerce0, <4 x float> %f.coerce1)
+
+typedef float __attribute__((ext_vector_type(3))) v3f32;
+struct OddSizeHVA { v3f32 x, y; };
+
+void __regcall odd_size_hva(struct OddSizeHVA a) {}
+// Win32: define x86_regcallcc void @__regcall3__odd_size_hva(<3 x float> %a.0, <3 x float> %a.1)
+// Win64: define x86_regcallcc void @__regcall3__odd_size_hva(<3 x float> %a.0, <3 x float> %a.1)
+// Lin32: define x86_regcallcc void @__regcall3__odd_size_hva(<3 x float> %a.0, <3 x float> %a.1)
+// Lin64: define x86_regcallcc void @__regcall3__odd_size_hva(<3 x float> %a.coerce0, <3 x float> %a.coerce1)
+
+struct HFA6 { __m128 f[4]; };
+struct HFA6 __regcall ret_reg_reused(struct HFA6 a, struct HFA6 b, struct HFA6 c, struct HFA6 d){ struct HFA6 h; return h;}
+// Win32: define x86_regcallcc %struct.HFA6 @__regcall3__ret_reg_reused(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, %struct.HFA6* inreg %c, %struct.HFA6* inreg %d)
+// Win64: define x86_regcallcc %struct.HFA6 @__regcall3__ret_reg_reused(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, <4 x float> %c.0, <4 x float> %c.1, <4 x float> %c.2, <4 x float> %c.3, <4 x float> %d.0, <4 x float> %d.1, <4 x float> %d.2, <4 x float> %d.3)
+// Lin32: define x86_regcallcc %struct.HFA6 @__regcall3__ret_reg_reused(<4 x float> %a.0, <4 x float> %a.1, <4 x float> %a.2, <4 x float> %a.3, <4 x float> %b.0, <4 x float> %b.1, <4 x float> %b.2, <4 x float> %b.3, %struct.HFA6* inreg %c, %struct.HFA6* inreg %d)
+// Lin64: define x86_regcallcc %struct.HFA6 @__regcall3__ret_reg_reused([4 x <4 x float>] %a.coerce, [4 x <4 x float>] %b.coerce, [4 x <4 x float>] %c.coerce, [4 x <4 x float>] %d.coerce)

cfe/trunk/test/CodeGenCXX/regcall.cpp

+// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm -std=c++11     %s -o - | FileCheck -check-prefix=CHECK-LIN -check-prefix=CHECK-LIN64 %s
+// RUN: %clang_cc1 -triple i386-linux-gnu -emit-llvm -std=c++11     %s -o -   | FileCheck -check-prefix=CHECK-LIN -check-prefix=CHECK-LIN32 %s
+// RUN: %clang_cc1 -triple x86_64-windows-msvc -emit-llvm -std=c++11  %s -o - -DWIN_TEST | FileCheck -check-prefix=CHECK-WIN64 %s
+// RUN: %clang_cc1 -triple i386-windows-msvc -emit-llvm -std=c++11  %s -o - -DWIN_TEST   | FileCheck -check-prefix=CHECK-WIN32 %s
+
+int __regcall foo(int i);
+
+int main()
+{
+  int p = 0, _data;
+  auto lambda = [&](int parameter) -> int {
+    _data = foo(parameter);
+    return _data;
+  };
+  return lambda(p);
+}
+// CHECK-LIN: call x86_regcallcc {{.+}} @_Z15__regcall3__foo
+// CHECK-WIN64: call x86_regcallcc {{.+}} @"\01?foo@@YwHH@Z"
+// CHECK-WIN32: call x86_regcallcc {{.+}} @"\01?foo@@YwHH@Z"
+
+int __regcall foo (int i){
+  return i;
+}
+// CHECK-LIN: define x86_regcallcc {{.+}}@_Z15__regcall3__foo
+// CHECK-WIN64: define x86_regcallcc {{.+}}@"\01?foo@@YwHH@Z"
+// CHECK-WIN32: define x86_regcallcc {{.+}}@"\01?foo@@YwHH@Z"
+
+// used to give a body to test_class functions
+static int x = 0;
+class test_class {
+  int a;
+public:
+#ifndef WIN_TEST
+  __regcall 
+#endif
+    test_class(){++x;}
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_classC1Ev
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_classC2Ev
+  // Windows ignores calling convention on constructor/destructors.
+  // CHECK-WIN64-DAG: define linkonce_odr %class.test_class* @"\01??0test_class@@QEAA@XZ"
+  // CHECK-WIN32-DAG: define linkonce_odr x86_thiscallcc %class.test_class* @"\01??0test_class@@QAE@XZ"
+
+#ifndef WIN_TEST
+  __regcall 
+#endif
+  ~test_class(){--x;}
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_classD2Ev
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_classD1Ev
+  // Windows ignores calling convention on constructor/destructors.
+  // CHECK-WIN64-DAG: define linkonce_odr void @"\01??_Dtest_class@@QEAA@XZ"
+  // CHECK-WIN32-DAG: define linkonce_odr x86_thiscallcc void @"\01??_Dtest_class@@QAE@XZ"
+  
+  test_class& __regcall operator+=(const test_class&){
+    return *this;
+  }
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc dereferenceable(4) %class.test_class* @_ZN10test_classpLERKS_
+  // CHECK-WIN64-DAG: define linkonce_odr x86_regcallcc dereferenceable(4) %class.test_class* @"\01??Ytest_class@@QEAwAEAV0@AEBV0@@Z"
+  // CHECK-WIN32-DAG: define linkonce_odr x86_regcallcc dereferenceable(4) %class.test_class* @"\01??Ytest_class@@QAwAAV0@ABV0@@Z"
+  void __regcall do_thing(){}
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_class20__regcall3__do_thingEv
+  // CHECK-WIN64-DAG: define linkonce_odr x86_regcallcc void @"\01?do_thing@test_class@@QEAwXXZ"
+  // CHECK-WIN32-DAG: define linkonce_odr x86_regcallcc void @"\01?do_thing@test_class@@QAwXXZ"
+  
+  template<typename T>
+  void __regcall tempFunc(T i){}
+  // CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_ZN10test_class20__regcall3__tempFuncIiEEvT_
+  // CHECK-WIN64-DAG: define linkonce_odr x86_regcallcc void @"\01??$freeTempFunc@H@@YwXH@Z"
+  // CHECK-WIN32-DAG: define linkonce_odr x86_regcallcc void @"\01??$freeTempFunc@H@@YwXH@Z"
+};
+
+bool __regcall operator ==(const test_class&, const test_class&){ --x; return false;}
+// CHECK-LIN-DAG: define x86_regcallcc zeroext i1 @_ZeqRK10test_classS1_
+// CHECK-WIN64-DAG: define x86_regcallcc zeroext i1 @"\01??8@Yw_NAEBVtest_class@@0@Z"
+// CHECK-WIN32-DAG: define x86_regcallcc zeroext i1 @"\01??8@Yw_NABVtest_class@@0@Z"
+
+test_class __regcall operator""_test_class (unsigned long long) { ++x; return test_class{};} 
+// CHECK-LIN64-DAG: define x86_regcallcc %class.test_class @_Zli11_test_classy(i64)
+// CHECK-LIN32-DAG: define x86_regcallcc void @_Zli11_test_classy(%class.test_class* inreg noalias sret %agg.result, i64)
+// CHECK-WIN64-DAG: \01??__K_test_class@@Yw?AVtest_class@@_K@Z"
+// CHECK-WIN32-DAG: \01??__K_test_class@@Yw?AVtest_class@@_K@Z"
+
+template<typename T>
+void __regcall freeTempFunc(T i){}
+// CHECK-LIN-DAG: define linkonce_odr x86_regcallcc void @_Z24__regcall3__freeTempFuncIiEvT_
+// CHECK-WIN64-DAG: define linkonce_odr x86_regcallcc void @"\01??$freeTempFunc@H@@YwXH@Z"
+// CHECK-WIN32-DAG: define linkonce_odr x86_regcallcc void @"\01??$freeTempFunc@H@@YwXH@Z"
+
+// class to force generation of functions
+void force_gen() {
+  test_class t;
+  test_class t2 = 12_test_class;
+  t += t2;
+  auto t3 = 100_test_class;
+  t3.tempFunc(1);
+  freeTempFunc(1);
+  t3.do_thing();
+}

cfe/trunk/tools/libclang/CXType.cpp

CXCallingConv clang_getFunctionTypeCallingConv(CXType X) {
   if (const FunctionType *FD = T->getAs<FunctionType>()) {
 #define TCALLINGCONV(X) case CC_##X: return CXCallingConv_##X
     switch (FD->getCallConv()) {
       TCALLINGCONV(C);
       TCALLINGCONV(X86StdCall);
       TCALLINGCONV(X86FastCall);
       TCALLINGCONV(X86ThisCall);
       TCALLINGCONV(X86Pascal);
+      TCALLINGCONV(X86RegCall);
       TCALLINGCONV(X86VectorCall);
       TCALLINGCONV(X86_64Win64);
       TCALLINGCONV(X86_64SysV);
       TCALLINGCONV(AAPCS);
       TCALLINGCONV(AAPCS_VFP);
       TCALLINGCONV(IntelOclBicc);
       TCALLINGCONV(Swift);
       TCALLINGCONV(PreserveMost);