[PATCH] #pragma vectorize

Aaron Ballman aaron at aaronballman.com
Tue May 13 06:56:40 PDT 2014


On Mon, May 12, 2014 at 7:20 PM, Tyler Nowicki <tnowicki at apple.com> wrote:
> Hi Aaron, Ben,
>
> Thank again for the review. I’ve made the changes you suggested.
>
> There was a question about why I reversed the iteration over the attribute
> list. The answer is that the attribute list built by ParsedAttributes stores
> attributes in reverse order. However, this is wrong because the correct
> order should be maintained for serialization/deserialization and error
> reporting. So we have to iterate rbegin->rend. Ideally the ParsedAttributes
> list should be fixed to store attributes in the order they appear but this
> looks like a lot of work.

Ah, yes. That problem keeps rearing its ugly head. :-/

>
> Tyler
>
>
>
>
> That part of tablegen probably shouldn’t be specialized for each type of
> pragma should it? But from your comment below it sounds like thats what you
> are thinking.
>
>
> I'd have to think about it more, but it seems like tablegen shouldn't
> have to specialize for each pragma, just all pragmas. Eg) the
> difference between printing pragmas and printing attributes is minor
> enough that it could be handled entirely by tablegen without the
> pragma authors having to write special code.
>
>
> I’m pretty sure that each type of pragma has a unique syntax that makes it
> difficult to generalize.
>
>
> +                          ["disable", "enable", "value"],
> +                          ["Disable", "Enable", "Value"]>,
>
>
> This is actually an optional argument as well, but is not marked as
> such. It should get a , 1. Also, this suggests we need a new argument
> type that represents a union of arguments, since that's really what
> you want (one of these two arguments must be used, but you don't care
> which). A FIXME would probably be appropriate (though you don't have
> to implement the functionality for this patch).
>
>
> I don’t think it is. Just specifying vectorize or interleave does not imply
> a default action. Perhaps it should?

As it's written, it implies that you must specify vectorize or
interleave, followed by disable/enable/value, and then optionally
supply a value. Eg) #pragma loop vectorize(enable, 4)

That's why I think we should have a Union argument type because you
really want it to either be enable|disable, or an integer value, not
both.

Not something that needs doing for this patch by any means. :-)

>
>
> +              ExprArgument<"Value", 1>];
>
>
> Judging by the tests, this should be a DefaultIntArgument<"Value", 1>.
> Either that, or there are tests missing where expressions are used
> (and honestly, it would strike me as slightly strange to allow general
> expressions here).
>
>
> I was thinking ahead to non-type template arguments. But that can wait. I’ll
> use an int for now.
>
>
> const char *Names[] = { "llvm.vectorizer.width", "llvm.vectorizer.unroll" };
> llvm::Value *Value;
> llvm::MDString *Name;
>
> if (Kind == LoopHintAttr::Enable) {
>  Name = llvm::MDString::get(Context, "llvm.vectorizer.enable");
>  Value = Builder.getTrue();
> } else {
>  Name = llvm::MDString::get(Context, Names[Option]);
>  Value = llvm::ConstantInt::get(Int32Ty, ValueInt); // You already
> set ValueInt to 1 by default, and overwrite when the Kind is a Value.
> }
>
>
> Good idea!
>
>
> +  }
> +
> +  // Get the next statement.
> +  MaybeParseCXX11Attributes(Attrs);
> +
> +  StmtResult S = ParseStatementOrDeclarationAfterAttributes(Stmts,
> +                  /*OnlyStatement*/ true, 0, Attrs);
>
>
> Shouldn't we be passing the OnlyStatement which was passed into the
> function? Same for passing in the TrailingElseLoc instead of 0?
>
>
> These inputs confused me, I duplicated the call made in
> ParseLabeledStatement(). I think OnlyStatement indicates that the next thing
> parsed is expected be a statement, rather than a declaration. I’ll pass the
> arguments as you suggest.
>
>
> Btw, when I test your patch locally, I get failed assertions from the
> STL. "array iterator + offset out of range" on a call to std::copy
> within ASTStmtReader::VisitAttributedStmt().
>
>
> I don’t seem to have that, also I didn’t make any changes to ASTStmtReader?
> Could you try out the attached patch and provide the stack dump if it fails
> again.

Yup, still failing.

58>  FAIL: Clang :: PCH/pragma-loop.cpp (3877 of 7229)
58>  ******************** TEST 'Clang :: PCH/pragma-loop.cpp' FAILED
********************
58>  Script:
58>  --
58>  E:/llvm/2013/Debug/bin/clang.EXE -cc1 -internal-isystem
E:\llvm\2013\Debug\bin\..\lib\clang\3.5.0\include -emit-pch -o
E:\llvm\2013\tools\clang\test\PCH\Output\pragma-loop.cpp.tmp.a
E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp
58>  E:/llvm/2013/Debug/bin/clang.EXE -cc1 -internal-isystem
E:\llvm\2013\Debug\bin\..\lib\clang\3.5.0\include -include-pch
E:\llvm\2013\tools\clang\test\PCH\Output\pragma-loop.cpp.tmp.a
E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp -ast-print -o - |
E:/llvm/2013/Debug/bin\FileCheck.EXE
E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp
58>  --
58>  Exit Code: 2
58>
58>  Command Output (stdout):
58>  --
58>  Command 0: "E:/llvm/2013/Debug/bin/clang.EXE" "-cc1"
"-internal-isystem"
"E:\llvm\2013\Debug\bin\..\lib\clang\3.5.0\include" "-emit-pch" "-o"
"E:\llvm\2013\tools\clang\test\PCH\Output\pragma-loop.cpp.tmp.a"
"E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp"
58>  Command 0 Result: 0
58>  Command 0 Output:
58>
58>
58>  Command 0 Stderr:
58>
58>
58>  Command 1: "E:/llvm/2013/Debug/bin/clang.EXE" "-cc1"
"-internal-isystem"
"E:\llvm\2013\Debug\bin\..\lib\clang\3.5.0\include" "-include-pch"
"E:\llvm\2013\tools\clang\test\PCH\Output\pragma-loop.cpp.tmp.a"
"E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp" "-ast-print" "-o"
"-"
58>  Command 1 Result: -2147483645
58>  Command 1 Output:
58>
58>
58>  Command 1 Stderr:
58>  Stack dump:
58>
58>  0. Program arguments: E:/llvm/2013/Debug/bin/clang.EXE -cc1
-internal-isystem E:\llvm\2013\Debug\bin\..\lib\clang\3.5.0\include
-include-pch E:\llvm\2013\tools\clang\test\PCH\Output\pragma-loop.cpp.tmp.a
E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp -ast-print -o -
58>
58>  1. <eof> parser at end of file
58>
58>  0x6C47CEE6 (0x055612B0 0x04BD89E0 0x00000618 0x00C6D504),
?_Debug_message at std@@YAXPB_W0I at Z() + 0x26 bytes(s)
58>
58>  0x032A6E98 (0x00000002 0x00C6D434 0x00C6D44C 0x032A6E0B),
std::_Array_const_iterator<clang::Attr const *,1>::operator+=() + 0x38
bytes(s), d:\program files (x86)\microsoft visual studio
12.0\vc\include\xutility, line 1560 + 0x17 byte(s)
58>
58>  0x032A6F4A (0x00000002 0xCCCCCCCC 0x00000000 0x00000000),
std::_Array_iterator<clang::Attr const *,1>::operator+=() + 0x1A
bytes(s), d:\program files (x86)\microsoft visual studio
12.0\vc\include\xutility, line 1764
58>
58>  0x032A6E0B (0x00C6D464 0x00000002 0xCCCCCCCC 0xCCCCCCCC),
std::_Array_iterator<clang::Attr const *,1>::operator+() + 0x3B
bytes(s), d:\program files (x86)\microsoft visual studio
12.0\vc\include\xutility, line 1770 + 0xC byte(s)
58>
58>  0x032A22D6 (0x00C6D504 0x00C6D564 0x00C6D56C 0x00000000),
std::_Copy_impl<clang::Attr * *,std::_Array_iterator<clang::Attr const
*,1> >() + 0x36 bytes(s), d:\program files (x86)\microsoft visual
studio 12.0\vc\include\xutility, line 2097
58>
58>  0x032A227F (0x00C6D504 0x00C6D564 0x00C6D56C 0x00000000),
std::_Copy_impl<clang::Attr * *,std::_Array_iterator<clang::Attr const
*,1> >() + 0x5F bytes(s), d:\program files (x86)\microsoft visual
studio 12.0\vc\include\xutility, line 2107 + 0x4B byte(s)
58>
58>  0x032A3313 (0x00C6D504 0x00C6D564 0x00C6D56C 0x00000000),
std::copy<clang::Attr * *,std::_Array_iterator<clang::Attr const *,1>
>() + 0xC3 bytes(s), d:\program files (x86)\microsoft visual studio
12.0\vc\include\xutility, line 2132 + 0x4F byte(s)
58>
58>  0x032A3228 (0x00C6D564 0x00C6D56C 0x00FF510C 0x00C6DEFC),
std::copy<clang::Attr * *,clang::Attr const *,1>() + 0x38 bytes(s),
d:\program files (x86)\microsoft visual studio
12.0\vc\include\xutility, line 2145 + 0x23 byte(s)
58>
58>  0x03297A73 (0x00FF50FC 0xCCCCCCCC 0x00000002 0xCCCCCCCC),
clang::ASTStmtReader::VisitAttributedStmt() + 0x133 bytes(s),
e:\llvm\llvm\tools\clang\lib\serialization\astreaderstmt.cpp, line 179
+ 0x1E byte(s)
58>
58>  0x032A7C3B (0x00FF50FC 0x00C6E3E4 0x00C6DF34 0xCCCCCCCC),
clang::StmtVisitorBase<clang::make_ptr,clang::ASTStmtReader,void>::Visit()
+ 0x3FB bytes(s),
e:\llvm\llvm\tools\clang\include\clang\ast\stmtvisitor.h, line 43 +
0xC byte(s)
58>
58>  0x03296804 (0x00FD6E30 0xCCCCCCCC 0x00FD6E30 0xCCCCCCCC),
clang::ASTReader::ReadStmtFromStream() + 0x3604 bytes(s),
e:\llvm\llvm\tools\clang\lib\serialization\astreaderstmt.cpp, line
2539
58>
58>  0x031CBE4A (0x00014648 0x00000000 0x00C6DF78 0x03BB3352),
clang::ASTReader::GetExternalDeclStmt() + 0x6A bytes(s),
e:\llvm\llvm\tools\clang\lib\serialization\astreader.cpp, line 6156
58>
58>  0x03BB42D8 (0x00FD40C4 0x00FD38C8 0x00C6DF70 0x03BB4416),
clang::LazyOffsetPtr<clang::Stmt,unsigned
__int64,{clang::ExternalASTSource::`vcall'{16}',0}>::get() + 0x88
bytes(s), e:\llvm\llvm\tools\clang\include\clang\ast\externalastsource.h,
line 351 + 0x2A byte(s)
58>
58>  0x03BA74F6 (0x00C6DF64 0xCCCCCCCC 0x00FD38C8 0xCCCCCCCC),
clang::FunctionDecl::getBody() + 0x56 bytes(s),
e:\llvm\llvm\tools\clang\lib\ast\decl.cpp, line 2246 + 0x1D byte(s)
58>
58>  0x03BB4416 (0x00FC8880 0x00000000 0x00C6E304 0x00000002),
clang::FunctionDecl::getBody() + 0x26 bytes(s),
e:\llvm\llvm\tools\clang\include\clang\ast\decl.h, line 1654
58>
58>  0x03C4F282 (0x00FD38C8 0x00C6E5B0 0x00C6E408 0x03C524AC),
`anonymous namespace'::DeclPrinter::VisitFunctionDecl() + 0xD92
bytes(s), e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 589 +
0x28 byte(s)
58>
58>  0x03C527DA (0x00FD38C8 0x0000002A 0x00C6E5B0 0x00C6E488),
clang::declvisitor::Base<clang::declvisitor::make_ptr,`anonymous
namespace'::DeclPrinter,void>::VisitCXXMethodDecl() + 0x1A bytes(s),
e:\llvm\2013\tools\clang\include\clang\ast\declnodes.inc, line 349 +
0x1A byte(s)
58>
58>  0x03C524AC (0x00FD38C8 0x00C6E4DC 0x00C6E6EC 0xCCCCCCCC),
clang::declvisitor::Base<clang::declvisitor::make_ptr,`anonymous
namespace'::DeclPrinter,void>::Visit() + 0x30C bytes(s),
e:\llvm\llvm\tools\clang\include\clang\ast\declvisitor.h, line 349 +
0xC byte(s)
58>
58>  0x03C4DD12 (0x00FD3548 0x00000001 0x00C6E570 0x00FD3548),
`anonymous namespace'::DeclPrinter::VisitDeclContext() + 0x2D2
bytes(s), e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 285
58>
58>  0x03C4FF83 (0x00FD3528 0x0000001C 0x00C6E5B0 0x00C6E570),
`anonymous namespace'::DeclPrinter::VisitCXXRecordDecl() + 0x1E3
bytes(s), e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 787
58>
58>  0x03C523BE (0x00FD3528 0x00C6E694 0x00C6E6EC 0xCCCCCCCC),
clang::declvisitor::Base<clang::declvisitor::make_ptr,`anonymous
namespace'::DeclPrinter,void>::Visit() + 0x21E bytes(s),
e:\llvm\llvm\tools\clang\include\clang\ast\declvisitor.h, line 233 +
0xC byte(s)
58>
58>  0x03C4DD12 (0x00FD30D4 0x00000000 0x00FD30D4 0x00C6E5B0),
`anonymous namespace'::DeclPrinter::VisitDeclContext() + 0x2D2
bytes(s), e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 285
58>
58>  0x03C4E03D (0x00FD30C0 0x0000003B 0x00C6E5B0 0x00C6E694),
`anonymous namespace'::DeclPrinter::VisitTranslationUnitDecl() + 0x3D
bytes(s), e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 322
58>
58>  0x03C525B8 (0x00FD30C0 0x00C6E770 0xCCCCCCCC 0x00FC8880),
clang::declvisitor::Base<clang::declvisitor::make_ptr,`anonymous
namespace'::DeclPrinter,void>::Visit() + 0x418 bytes(s),
e:\llvm\llvm\tools\clang\include\clang\ast\declvisitor.h, line 473 +
0xC byte(s)
58>
58>  0x03C4D5F6 (0x00FC8880 0x00FCF114 0x00000000 0x00000001),
clang::Decl::print() + 0x56 bytes(s),
e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 102
58>
58>  0x03C4D583 (0x00FC8880 0x00000000 0x00000001 0xCCCCCCCC),
clang::Decl::print() + 0x33 bytes(s),
e:\llvm\llvm\tools\clang\lib\ast\declprinter.cpp, line 96
58>
58>  0x023E2279 (0x00FD30C0 0x00FD30C0 0x00FC6638 0x00C6E770),
`anonymous namespace'::ASTPrinter::print() + 0x89 bytes(s),
e:\llvm\llvm\tools\clang\lib\frontend\astconsumers.cpp, line 90
58>
58>  0x023521D0 (0x00FCBA38 0x00C6E818 0x00C6E780 0x00FF2AC0),
`anonymous namespace'::ASTPrinter::HandleTranslationUnit() + 0x40
bytes(s), e:\llvm\llvm\tools\clang\lib\frontend\astconsumers.cpp, line
48 + 0xC byte(s)
58>
58>  0x03130FB8 (0x00FE4E10 0x00000000 0x00000000 0x00C6E79C),
clang::ParseAST() + 0x238 bytes(s),
e:\llvm\llvm\tools\clang\lib\parse\parseast.cpp, line 154 + 0x18
byte(s)
58>
58>  0x022E00B1 (0x00C6E7C4 0xCCCCCCCC 0xCCCCCCCC 0xCCCCCCCC),
clang::ASTFrontendAction::ExecuteAction() + 0x101 bytes(s),
e:\llvm\llvm\tools\clang\lib\frontend\frontendaction.cpp, line 501 +
0x30 byte(s)
58>
58>  0x022DFCBE (0x00C6E8BC 0x00C6F97C 0xCCCCCCCC 0xCCCCCCCC),
clang::FrontendAction::Execute() + 0x7E bytes(s),
e:\llvm\llvm\tools\clang\lib\frontend\frontendaction.cpp, line 402 +
0xF byte(s)
58>
58>  0x022AF2B1 (0x00FABE08 0x00C6EE38 0x00C6F97C 0xCCCCCCCC),
clang::CompilerInstance::ExecuteAction() + 0x281 bytes(s),
e:\llvm\llvm\tools\clang\lib\frontend\compilerinstance.cpp, line 721
58>
58>  0x023EDAE8 (0x00FA6968 0x00C6FD9C 0xCCCCCCCC 0xCCCCCCCC),
clang::ExecuteCompilerInvocation() + 0x318 bytes(s),
e:\llvm\llvm\tools\clang\lib\frontendtool\executecompilerinvocation.cpp,
line 239 + 0x11 byte(s)
58>
58>  0x012C9682 (0x00C6F984 0x00C6F9A4 0x00FA6B90 0x01201361),
cc1_main() + 0x2F2 bytes(s),
e:\llvm\llvm\tools\clang\tools\driver\cc1_main.cpp, line 112 + 0xE
byte(s)
58>
58>  0x012B80B5 (0x0000000A 0x00FA2A98 0x00F9EAF8 0xFF71AEA3), main()
+ 0x225 bytes(s), e:\llvm\llvm\tools\clang\tools\driver\driver.cpp,
line 319 + 0x45 byte(s)
58>
58>  0x03EC2F89 (0x00C6FE00 0x7655338A 0x7EFDE000 0x00C6FE40),
__tmainCRTStartup() + 0x199 bytes(s),
f:\dd\vctools\crt\crtw32\dllstuff\crtexe.c, line 626 + 0x19 byte(s)
58>
58>  0x03EC30CD (0x7EFDE000 0x00C6FE40 0x778C9F72 0x7EFDE000),
mainCRTStartup() + 0xD bytes(s),
f:\dd\vctools\crt\crtw32\dllstuff\crtexe.c, line 466
58>
58>  0x7655338A (0x7EFDE000 0x1D841147 0x00000000 0x00000000),
BaseThreadInitThunk() + 0x12 bytes(s)
58>
58>  0x778C9F72 (0x03EC30C0 0x7EFDE000 0x00000000 0x00000000),
RtlInitializeExceptionChain() + 0x63 bytes(s)
58>
58>  0x778C9F45 (0x03EC30C0 0x7EFDE000 0x00000000 0x00000000),
RtlInitializeExceptionChain() + 0x36 bytes(s)
58>
58>
58>
58>  Command 2: "E:/llvm/2013/Debug/bin\FileCheck.EXE"
"E:\llvm\llvm\tools\clang\test\PCH\pragma-loop.cpp"
58>  Command 2 Result: 2
58>  Command 2 Output:
58>
58>
58>  Command 2 Stderr:
58>CUSTOMBUILD : FileCheck error : '-' is empty.
58>
58>
58>
58>
58>  --
58>
58>  ********************

Some formatting nits (you may want to run clang-format over your
patch; http://clang.llvm.org/docs/ClangFormat.html) and one question
which may require a comment to be updated. Once those are resolved,
along with the crash from above, I think you'll be all set!

> Index: include/clang/Basic/Attr.td
> ===================================================================
> --- include/clang/Basic/Attr.td (revision 208638)
> +++ include/clang/Basic/Attr.td (working copy)
> @@ -1750,3 +1750,64 @@
>  def Unaligned : IgnoredAttr {
>    let Spellings = [Keyword<"__unaligned">];
>  }
> +
> +def LoopHint : Attr {
> +  // LoopHint Vectorize:
> +  //  enable - use vector instructions.
> +  //  disable - do not use vector instructions.
> +  //  positive value - use vector instructions of the specified width.
> +
> +  // LoopHint Interleave:
> +  //  enable - interleave multiple loop iterations.
> +  //  disable - do not interleave multiple loop interactions.
> +  //  positive value - interleave the specified number of loop interations.
> +
> +  // FIXME: Add Pragma spelling to tablegen and
> +  // use it here.
> +  let Spellings = [Keyword<"loop">];
> +
> +  // State of the loop optimization specified by the spelling.
> +  let Args = [EnumArgument<"Option", "OptionType",
> +                          ["vectorize", "interleave"],
> +                          ["Vectorize", "Interleave"]>,
> +              EnumArgument<"Kind", "KindType",
> +                          ["disable", "enable", "value"],
> +                          ["Disable", "Enable", "Value"]>,
> +              DefaultIntArgument<"Value", 1>];
> +
> +  let AdditionalMembers = [{
> +  // Kinds are compatible if they are not exclusive.
> +  static bool isCompatible(int Kind1, int Kind2) {
> +    return (Kind1 == Disable) == (Kind2 == Disable);

So if Kind1 = Enable and Kind2 = Value, they are compatible? If so,
can the comment be updated to make that a bit more explicit
(otherwise, it kind of looks like this code may have a bug).

> +  }
> +
> +  static StringRef getOptionName(int Option) {
> +    switch (Option) {
> +    case Vectorize: return "vectorize";
> +    case Interleave: return "interleave";
> +    }
> +    llvm_unreachable("Unhandled LoopHint option.");
> +  }
> +
> +  static StringRef getKindName(int Kind) {
> +    switch (Kind) {
> +    case Disable: return "disable";
> +    case Enable: return "enable";
> +    case Value: return "value";
> +    }
> +    llvm_unreachable("Unhandled LoopHint kind.");
> +  }
> +
> +  // FIXME: Modify pretty printer to print this pragma
> +  void print(raw_ostream &OS, const PrintingPolicy &Policy) const {
> +    OS << getOptionName(option) << "(";
> +    if (getKind() == Value)
> +      OS << value;
> +    else
> +      OS << getKindName(kind);
> +    OS << ")\n";
> +  }
> +  }];
> +
> +  let Documentation = [Undocumented];
> +}
> Index: include/clang/Basic/DiagnosticParseKinds.td
> ===================================================================
> --- include/clang/Basic/DiagnosticParseKinds.td (revision 208638)
> +++ include/clang/Basic/DiagnosticParseKinds.td (working copy)
> @@ -884,6 +884,14 @@
>    "unexpected OpenMP clause '%0' in directive '#pragma omp %1'">;
>  def err_omp_more_one_clause : Error<
>    "directive '#pragma omp %0' cannot contain more than one '%1' clause">;
> +
> +// Pragma loop support.
> +def err_pragma_loop_invalid_option : Error<
> +  "%select{invalid|missing}0 option%select{ '%1'|}0 in directive "
> +  "'#pragma loop'; expected either vectorize or interleave">;
> +def err_pragma_loop_invalid_type : Error<
> +  "invalid value '%0' in directive '#pragma loop %1'; expected either "
> +  "'enable', 'disable', or a positive integer">;
>  } // end of Parse Issue category.
>
>  let CategoryName = "Modules Issue" in {
> Index: include/clang/Basic/DiagnosticSemaKinds.td
> ===================================================================
> --- include/clang/Basic/DiagnosticSemaKinds.td (revision 208638)
> +++ include/clang/Basic/DiagnosticSemaKinds.td (working copy)
> @@ -537,6 +537,13 @@
>    "#pragma visibility pop with no matching #pragma visibility push">;
>  def note_surrounding_namespace_starts_here : Note<
>    "surrounding namespace with visibility attribute starts here">;
> +def err_pragma_loop_invalid_value : Error<
> +  "expected a positive integer in directive '#pragma loop %0'">;
> +def err_pragma_loop_incompatible : Error<
> +  "'%0' and '%1' directive option types are incompatible in '#pragma loop %2'">;
> +def err_pragma_loop_precedes_nonloop : Error<
> +  "expected a for, while, or do-while loop to follow the '#pragma loop' "
> +  "directive">;
>
>  /// Objective-C parser diagnostics
>  def err_duplicate_class_def : Error<
> Index: include/clang/Basic/TokenKinds.def
> ===================================================================
> --- include/clang/Basic/TokenKinds.def (revision 208638)
> +++ include/clang/Basic/TokenKinds.def (working copy)
> @@ -701,6 +701,11 @@
>  ANNOTATION(pragma_openmp)
>  ANNOTATION(pragma_openmp_end)
>
> +// Annotations for loop pragma directives #pragma loop ...
> +// The lexer produces these so that they only take effect when the parser
> +// handles #pragma loop ... directives.
> +ANNOTATION(pragma_loop_hint)
> +
>  // Annotations for module import translated from #include etc.
>  ANNOTATION(module_include)
>  ANNOTATION(module_begin)
> Index: include/clang/Parse/Parser.h
> ===================================================================
> --- include/clang/Parse/Parser.h (revision 208638)
> +++ include/clang/Parse/Parser.h (working copy)
> @@ -20,6 +20,7 @@
>  #include "clang/Lex/CodeCompletionHandler.h"
>  #include "clang/Lex/Preprocessor.h"
>  #include "clang/Sema/DeclSpec.h"
> +#include "clang/Sema/LoopHint.h"
>  #include "clang/Sema/Sema.h"
>  #include "llvm/ADT/SmallVector.h"
>  #include "llvm/Support/Compiler.h"
> @@ -160,6 +161,7 @@
>    std::unique_ptr<PragmaHandler> MSConstSeg;
>    std::unique_ptr<PragmaHandler> MSCodeSeg;
>    std::unique_ptr<PragmaHandler> MSSection;
> +  std::unique_ptr<PragmaHandler> LoopHintHandler;
>
>    std::unique_ptr<CommentHandler> CommentSemaHandler;
>
> @@ -518,6 +520,10 @@
>    /// #pragma clang __debug captured
>    StmtResult HandlePragmaCaptured();
>
> +  /// \brief Handle the annotation token produced for
> +  /// #pragma vectorize...
> +  LoopHint HandlePragmaLoopHint();
> +
>    /// GetLookAheadToken - This peeks ahead N tokens and returns that token
>    /// without consuming any tokens.  LookAhead(0) returns 'Tok', LookAhead(1)
>    /// returns the token after Tok, etc.
> @@ -1600,6 +1606,10 @@
>    StmtResult ParseReturnStatement();
>    StmtResult ParseAsmStatement(bool &msAsm);
>    StmtResult ParseMicrosoftAsmStatement(SourceLocation AsmLoc);
> +  StmtResult ParsePragmaLoopHint(StmtVector &Stmts,
> +                                 bool OnlyStatement,
> +                                 SourceLocation *TrailingElseLoc,
> +                                 ParsedAttributesWithRange &Attrs);
>
>    /// \brief Describes the behavior that should be taken for an __if_exists
>    /// block.
> Index: include/clang/Sema/LoopHint.h
> ===================================================================
> --- include/clang/Sema/LoopHint.h (revision 0)
> +++ include/clang/Sema/LoopHint.h (working copy)
> @@ -0,0 +1,32 @@
> +//===--- LoopHint.h - Types for LoopHint ------------------------*- C++ -*-===//
> +//
> +//                     The LLVM Compiler Infrastructure
> +//
> +// This file is distributed under the University of Illinois Open Source
> +// License. See LICENSE.TXT for details.
> +//
> +//===----------------------------------------------------------------------===//
> +
> +
> +#ifndef LLVM_CLANG_SEMA_LOOPHINT_H
> +#define LLVM_CLANG_SEMA_LOOPHINT_H
> +
> +#include "clang/Basic/IdentifierTable.h"
> +#include "clang/Basic/SourceLocation.h"
> +#include "clang/Sema/AttributeList.h"
> +#include "clang/Sema/Ownership.h"
> +
> +namespace clang {
> +
> +/// \brief Loop hint specified by a pragma loop directive.
> +struct LoopHint {
> +  SourceRange Range;
> +  Expr *ValueExpr;
> +  IdentifierLoc *LoopLoc;
> +  IdentifierLoc *ValueLoc;
> +  IdentifierLoc *OptionLoc;
> +};
> +
> +} // end namespace clang
> +
> +#endif // LLVM_CLANG_SEMA_LOOPHINT_H
> Index: lib/AST/StmtPrinter.cpp
> ===================================================================
> --- lib/AST/StmtPrinter.cpp (revision 208638)
> +++ lib/AST/StmtPrinter.cpp (working copy)
> @@ -168,19 +168,32 @@
>  }
>
>  void StmtPrinter::VisitAttributedStmt(AttributedStmt *Node) {
> -  OS << "[[";
>    bool first = true;
> -  for (ArrayRef<const Attr*>::iterator it = Node->getAttrs().begin(),
> -                                       end = Node->getAttrs().end();
> -                                       it != end; ++it) {
> -    if (!first) {
> -      OS << ", ";
> -      first = false;
> +  std::string raw_attr_os;
> +  llvm::raw_string_ostream AttrOS(raw_attr_os);
> +  for (auto I = Node->getAttrs().rbegin(), E = Node->getAttrs().rend();
> +       I != E; ++I) {
> +    // FIXME: This hack will be removed when printPretty
> +    // has been modified to print pretty pragmas.
> +    if (dyn_cast<LoopHintAttr>(*I)) {
> +      const LoopHintAttr *LHA = cast<LoopHintAttr>(*I);
> +      OS << "#pragma loop ";
> +      LHA->print(OS, Policy);
> +    } else {
> +      if (!first) {
> +        AttrOS << ", ";
> +        first = false;
> +      }
> +      // TODO: check this
> +      (*I)->printPretty(AttrOS, Policy);
>      }
> -    // TODO: check this
> -    (*it)->printPretty(OS, Policy);
>    }
> -  OS << "]] ";
> +
> +  // Check to see if any attributes were printed.
> +  StringRef AttrStr = AttrOS.str();
> +  if (!AttrStr.empty())
> +    OS << "[[" << AttrStr << "]] ";
> +
>    PrintStmt(Node->getSubStmt(), 0);
>  }
>
> Index: lib/CodeGen/CGStmt.cpp
> ===================================================================
> --- lib/CodeGen/CGStmt.cpp (revision 208638)
> +++ lib/CodeGen/CGStmt.cpp (working copy)
> @@ -18,6 +18,7 @@
>  #include "clang/AST/StmtVisitor.h"
>  #include "clang/Basic/PrettyStackTrace.h"
>  #include "clang/Basic/TargetInfo.h"
> +#include "clang/Sema/LoopHint.h"
>  #include "clang/Sema/SemaDiagnostic.h"
>  #include "llvm/ADT/StringExtras.h"
>  #include "llvm/IR/CallSite.h"
> @@ -396,7 +397,23 @@
>  }
>
>  void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
> -  EmitStmt(S.getSubStmt());
> +  const Stmt *SubStmt = S.getSubStmt();
> +  switch (SubStmt->getStmtClass()) {
> +  case Stmt::DoStmtClass:
> +    EmitDoStmt(cast<DoStmt>(*SubStmt), S.getAttrs());
> +    break;
> +  case Stmt::ForStmtClass:
> +    EmitForStmt(cast<ForStmt>(*SubStmt), S.getAttrs());
> +    break;
> +  case Stmt::WhileStmtClass:
> +    EmitWhileStmt(cast<WhileStmt>(*SubStmt), S.getAttrs());
> +    break;
> +  case Stmt::CXXForRangeStmtClass:
> +    EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*SubStmt), S.getAttrs());
> +    break;
> +  default:
> +    EmitStmt(SubStmt);
> +  }
>  }
>
>  void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
> @@ -502,7 +519,57 @@
>    EmitBlock(ContBlock, true);
>  }
>
> -void CodeGenFunction::EmitWhileStmt(const WhileStmt &S) {
> +void CodeGenFunction::EmitCondBrHints(llvm::LLVMContext &Context,
> +                                      llvm::BranchInst *CondBr,
> +                                      ArrayRef<const Attr *> &Attrs) {
> +  // Do not continue if there are not hints.
> +  if (Attrs.empty())
> +    return;
> +
> +  // Add vectorize hints to the metadata on the conditional branch.
> +  // Iterate in reverse so hints are put in the same order they appear.
> +  SmallVector<llvm::Value*, 2> Metadata(1);
> +  for (auto I = Attrs.rbegin(), E = Attrs.rend(); I != E; ++I) {
> +    const LoopHintAttr *LH = dyn_cast<LoopHintAttr>(*I);
> +
> +    // Skip non loop hint attributes
> +    if (!LH)
> +      continue;
> +
> +    LoopHintAttr::OptionType Option = LH->getOption();
> +    int ValueInt = LH->getValue();
> +    int Kind = LH->getKind();
> +
> +    llvm::Value *Value;
> +    llvm::MDString *Name;
> +    const char *MetadataNames[] = { "llvm.vectorizer.width",
> +                                    "llvm.vectorizer.unroll" };
> +    if (Kind == LoopHintAttr::Enable) {
> +      Name = llvm::MDString::get(Context, "llvm.vectorizer.enable");
> +      Value = Builder.getTrue();
> +    } else {
> +      // No need for the disable case because ValueInt is 1 if Kind is disable.
> +      Name = llvm::MDString::get(Context, MetadataNames[Option]);
> +      Value = llvm::ConstantInt::get(Int32Ty, ValueInt);
> +    }
> +
> +    SmallVector<llvm::Value*, 2> OpValues;
> +    OpValues.push_back(Name);
> +    OpValues.push_back(Value);
> +
> +    // Set or overwrite metadata indicated by Name.
> +    Metadata.push_back(llvm::MDNode::get(Context, OpValues));
> +  }
> +
> +  // Add llvm.loop MDNode to CondBr.
> +  llvm::MDNode *LoopID = llvm::MDNode::get(Context, Metadata);
> +  LoopID->replaceOperandWith(0, LoopID); // First op points to itself.
> +
> +  CondBr->setMetadata("llvm.loop", LoopID);
> +}
> +
> +void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
> +                                    ArrayRef<const Attr *> WhileAttrs) {
>    RegionCounter Cnt = getPGORegionCounter(&S);
>
>    // Emit the header for the loop, which will also become
> @@ -547,13 +614,17 @@
>      llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
>      if (ConditionScope.requiresCleanups())
>        ExitBlock = createBasicBlock("while.exit");
> -    Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock,
> -                         PGO.createLoopWeights(S.getCond(), Cnt));
> +    llvm::BranchInst *CondBr = Builder.CreateCondBr(
> +                               BoolCondVal, LoopBody, ExitBlock,
> +                               PGO.createLoopWeights(S.getCond(), Cnt));

Formatting.

>
>      if (ExitBlock != LoopExit.getBlock()) {
>        EmitBlock(ExitBlock);
>        EmitBranchThroughCleanup(LoopExit);
>      }
> +
> +    // Attach metadata to loop body conditional branch.
> +    EmitCondBrHints(LoopBody->getContext(), CondBr, WhileAttrs);
>    }
>
>    // Emit the loop body.  We have to emit this in a cleanup scope
> @@ -582,7 +653,8 @@
>      SimplifyForwardingBlocks(LoopHeader.getBlock());
>  }
>
> -void CodeGenFunction::EmitDoStmt(const DoStmt &S) {
> +void CodeGenFunction::EmitDoStmt(const DoStmt &S,
> +                                 ArrayRef<const Attr *> DoAttrs) {
>    JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
>    JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
>
> @@ -619,10 +691,15 @@
>        EmitBoolCondBranch = false;
>
>    // As long as the condition is true, iterate the loop.
> -  if (EmitBoolCondBranch)
> -    Builder.CreateCondBr(BoolCondVal, LoopBody, LoopExit.getBlock(),
> -                         PGO.createLoopWeights(S.getCond(), Cnt));
> +  if (EmitBoolCondBranch) {
> +    llvm::BranchInst *CondBr = Builder.CreateCondBr(
> +                                 BoolCondVal, LoopBody, LoopExit.getBlock(),
> +                                 PGO.createLoopWeights(S.getCond(), Cnt));

Formatting.

>
> +    // Attach metadata to loop body conditional branch.
> +    EmitCondBrHints(LoopBody->getContext(), CondBr, DoAttrs);
> +  }
> +
>    // Emit the exit block.
>    EmitBlock(LoopExit.getBlock());
>
> @@ -632,7 +709,8 @@
>      SimplifyForwardingBlocks(LoopCond.getBlock());
>  }
>
> -void CodeGenFunction::EmitForStmt(const ForStmt &S) {
> +void CodeGenFunction::EmitForStmt(const ForStmt &S,
> +                                  ArrayRef<const Attr *> ForAttrs) {
>    JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
>
>    RunCleanupsScope ForScope(*this);
> @@ -686,9 +764,13 @@
>      // C99 6.8.5p2/p4: The first substatement is executed if the expression
>      // compares unequal to 0.  The condition must be a scalar type.
>      llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
> -    Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock,
> -                         PGO.createLoopWeights(S.getCond(), Cnt));
> +    llvm::BranchInst *CondBr = Builder.CreateCondBr(
> +                                 BoolCondVal, ForBody, ExitBlock,
> +                                 PGO.createLoopWeights(S.getCond(), Cnt));
>
> +    // Attach metadata to loop body conditional branch.
> +    EmitCondBrHints(ForBody->getContext(), CondBr, ForAttrs);
> +
>      if (ExitBlock != LoopExit.getBlock()) {
>        EmitBlock(ExitBlock);
>        EmitBranchThroughCleanup(LoopExit);
> @@ -728,7 +810,8 @@
>    EmitBlock(LoopExit.getBlock(), true);
>  }
>
> -void CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S) {
> +void CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
> +                                          ArrayRef<const Attr *> ForAttrs) {
>    JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
>
>    RunCleanupsScope ForScope(*this);
> @@ -761,9 +844,13 @@
>    // The body is executed if the expression, contextually converted
>    // to bool, is true.
>    llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
> -  Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock,
> -                       PGO.createLoopWeights(S.getCond(), Cnt));
> +  llvm::BranchInst *CondBr = Builder.CreateCondBr(BoolCondVal, ForBody,
> +                                      ExitBlock,
> +                                      PGO.createLoopWeights(S.getCond(), Cnt));

Formatting.

>
> +  // Attach metadata to loop body conditional branch
> +  EmitCondBrHints(ForBody->getContext(), CondBr, ForAttrs);
> +
>    if (ExitBlock != LoopExit.getBlock()) {
>      EmitBlock(ExitBlock);
>      EmitBranchThroughCleanup(LoopExit);
> Index: lib/CodeGen/CodeGenFunction.h
> ===================================================================
> --- lib/CodeGen/CodeGenFunction.h (revision 208638)
> +++ lib/CodeGen/CodeGenFunction.h (working copy)
> @@ -1846,9 +1846,16 @@
>    void EmitGotoStmt(const GotoStmt &S);
>    void EmitIndirectGotoStmt(const IndirectGotoStmt &S);
>    void EmitIfStmt(const IfStmt &S);
> -  void EmitWhileStmt(const WhileStmt &S);
> -  void EmitDoStmt(const DoStmt &S);
> -  void EmitForStmt(const ForStmt &S);
> +
> +  void EmitCondBrHints(llvm::LLVMContext &Context,
> +                       llvm::BranchInst *CondBr,
> +                       ArrayRef<const Attr *> &Attrs);
> +  void EmitWhileStmt(const WhileStmt &S,
> +                     ArrayRef<const Attr *> Attrs = ArrayRef<const Attr *>());
> +  void EmitDoStmt(const DoStmt &S,
> +                  ArrayRef<const Attr *> Attrs = ArrayRef<const Attr *>());
> +  void EmitForStmt(const ForStmt &S,
> +                   ArrayRef<const Attr *> Attrs = ArrayRef<const Attr *>());
>    void EmitReturnStmt(const ReturnStmt &S);
>    void EmitDeclStmt(const DeclStmt &S);
>    void EmitBreakStmt(const BreakStmt &S);
> @@ -1872,7 +1879,8 @@
>
>    void EmitCXXTryStmt(const CXXTryStmt &S);
>    void EmitSEHTryStmt(const SEHTryStmt &S);
> -  void EmitCXXForRangeStmt(const CXXForRangeStmt &S);
> +  void EmitCXXForRangeStmt(const CXXForRangeStmt &S,
> +                       ArrayRef<const Attr *> Attrs = ArrayRef<const Attr *>());

Formatting.

>
>    llvm::Function *EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K);
>    llvm::Function *GenerateCapturedStmtFunction(const CapturedDecl *CD,
> Index: lib/Parse/ParsePragma.cpp
> ===================================================================
> --- lib/Parse/ParsePragma.cpp (revision 208638)
> +++ lib/Parse/ParsePragma.cpp (working copy)
> @@ -15,6 +15,7 @@
>  #include "clang/Lex/Preprocessor.h"
>  #include "clang/Parse/ParseDiagnostic.h"
>  #include "clang/Parse/Parser.h"
> +#include "clang/Sema/LoopHint.h"
>  #include "clang/Sema/Scope.h"
>  #include "llvm/ADT/StringSwitch.h"
>  using namespace clang;
> @@ -131,6 +132,12 @@
>                      Token &FirstToken) override;
>  };
>
> +struct PragmaLoopHintHandler : public PragmaHandler {
> +  PragmaLoopHintHandler() : PragmaHandler("loop") {}
> +  void HandlePragma(Preprocessor &PP, PragmaIntroducerKind Introducer,
> +                    Token &FirstToken) override;
> +};
> +
>  }  // end namespace
>
>  void Parser::initializePragmaHandlers() {
> @@ -195,6 +202,9 @@
>      MSSection.reset(new PragmaMSPragma("section"));
>      PP.AddPragmaHandler(MSSection.get());
>    }
> +
> +  LoopHintHandler.reset(new PragmaLoopHintHandler());
> +  PP.AddPragmaHandler(LoopHintHandler.get());
>  }
>
>  void Parser::resetPragmaHandlers() {
> @@ -249,6 +259,10 @@
>
>    PP.RemovePragmaHandler("STDC", FPContractHandler.get());
>    FPContractHandler.reset();
> +
> +  PP.RemovePragmaHandler(LoopHintHandler.get());
> +  LoopHintHandler.reset();
> +
>  }
>
>  /// \brief Handle the annotation token produced for #pragma unused(...)
> @@ -570,6 +584,40 @@
>        DiagnosticsEngine::Error, "'#pragma %0' not implemented.");
>  }
>
> +struct PragmaLoopHintInfo {
> +  Token Loop;
> +  Token Value;
> +  Token Option;
> +};
> +
> +LoopHint Parser::HandlePragmaLoopHint() {
> +  assert(Tok.is(tok::annot_pragma_loop_hint));
> +  PragmaLoopHintInfo *Info =
> +    static_cast<PragmaLoopHintInfo *>(Tok.getAnnotationValue());
> +
> +  LoopHint Hint;
> +  Hint.LoopLoc = IdentifierLoc::create(Actions.Context,
> +                                       Info->Loop.getLocation(),
> +                                       Info->Loop.getIdentifierInfo());
> +  Hint.OptionLoc = IdentifierLoc::create(Actions.Context,
> +                                         Info->Option.getLocation(),
> +                                         Info->Option.getIdentifierInfo());
> +  Hint.ValueLoc = IdentifierLoc::create(Actions.Context,
> +                                        Info->Value.getLocation(),
> +                                        Info->Value.getIdentifierInfo());
> +  Hint.Range = SourceRange(Info->Option.getLocation(),
> +                           Info->Value.getLocation());
> +
> +  // FIXME: We should support template parameters for the loop hint value.
> +  // See bug report #19610
> +  if (Info->Value.is(tok::numeric_constant))
> +    Hint.ValueExpr = Actions.ActOnNumericConstant(Info->Value).get();
> +  else
> +    Hint.ValueExpr = nullptr;
> +
> +  return Hint;
> +}
> +
>  // #pragma GCC visibility comes in two variants:
>  //   'push' '(' [visibility] ')'
>  //   'pop'
> @@ -1531,3 +1579,108 @@
>
>    Actions.ActOnPragmaMSComment(Kind, ArgumentString);
>  }
> +
> +/// \brief Handle the \#pragma vectorize hint.
> +///  #pragma 'loop' loop-hints
> +///
> +///  loop-hints:
> +///    loop-hint loop-hints[opt]
> +///
> +///  loop-hint:
> +///    'vectorize' '(' loop-hint-value ')'
> +///    'interleave' '(' loop-hint-value ')'
> +///
> +///  loop-hint-value:
> +///    'enable'
> +///    'disable'
> +///    constant-expression
> +///
> +/// Specifying vectorize(enable) or vectorize(_value_) instructs llvm to
> +/// try vectorizing the instructions of the loop it precedes. Specifying
> +/// interleave(enable) or interleave(_value_) instructs llvm to try interleaving
> +/// multiple iterations of the loop it precedes. The _value_ indicates the
> +/// width of the vector instructions or the number of iterations of the loop
> +/// that should be interleaved. Consequently, a value of 1 or disable prevents
> +/// the optimization, even if it is possible and profitable, and 0 is invalid.
> +/// The loop vectorizer currently only works on inner loops.
> +///
> +void PragmaLoopHintHandler::HandlePragma(Preprocessor &PP,
> +                                         PragmaIntroducerKind Introducer,
> +                                         Token &Tok) {
> +  Token Loop = Tok;
> +  SmallVector <Token, 1> TokenList;
> +
> +  // Lex the optimization option and verify it is an identifier.
> +  PP.Lex(Tok);
> +  if (Tok.isNot(tok::identifier)) {
> +    PP.Diag(Tok.getLocation(), diag::err_pragma_loop_invalid_option)
> +      << /*InvalidOption*/true << "";
> +    return;
> +  }
> +
> +  while(Tok.is(tok::identifier)) {
> +    Token Option = Tok;
> +    StringRef OptionName = Tok.getIdentifierInfo()->getName();
> +
> +    if (OptionName != "vectorize" && OptionName != "interleave") {
> +      PP.Diag(Tok.getLocation(), diag::err_pragma_loop_invalid_option)
> +        << /*InvalidOption*/false << OptionName;
> +      return;
> +    }
> +
> +    // Read '('
> +    PP.Lex(Tok);
> +    if (Tok.isNot(tok::l_paren)) {
> +      PP.Diag(Tok.getLocation(), diag::err_expected) << tok::l_paren;
> +      return;
> +    }
> +
> +    PP.Lex(Tok);
> +    if (Tok.isNot(tok::identifier) && Tok.isNot(tok::numeric_constant)) {
> +      PP.Diag(Tok.getLocation(), diag::err_pragma_loop_invalid_type)
> +        << Tok.getName();
> +      return;
> +    }
> +
> +    Token Value = Tok;
> +
> +    // Read ')'
> +    PP.Lex(Tok);
> +    if (Tok.isNot(tok::r_paren)) {
> +      PP.Diag(Tok.getLocation(), diag::err_expected) << tok::r_paren;
> +      return;
> +    }
> +
> +    // Get next optimization option.
> +    PP.Lex(Tok);
> +
> +    PragmaLoopHintInfo *Info =
> +      (PragmaLoopHintInfo*) PP.getPreprocessorAllocator().Allocate(
> +        sizeof(PragmaLoopHintInfo), llvm::alignOf<PragmaLoopHintInfo>());
> +
> +    Info->Loop = Loop;
> +    Info->Option = Option;
> +    Info->Value = Value;
> +
> +    // Generate the vectorization hint token.
> +    Token LoopHintTok;
> +    LoopHintTok.startToken();
> +    LoopHintTok.setKind(tok::annot_pragma_loop_hint);
> +    LoopHintTok.setLocation(Loop.getLocation());
> +    LoopHintTok.setAnnotationValue(static_cast<void*>(Info));
> +    TokenList.push_back(LoopHintTok);
> +  }
> +
> +  if (Tok.isNot(tok::eod)) {
> +    PP.Diag(Tok.getLocation(), diag::warn_pragma_extra_tokens_at_eol)
> +      << "loop";
> +    return;
> +  }
> +
> +  Token *TokenArray = new Token[TokenList.size()];
> +  std::copy(TokenList.begin(), TokenList.end(), TokenArray);
> +
> +  PP.EnterTokenStream(TokenArray, TokenList.size(),
> +                      /*DisableMacroExpansion=*/false,
> +                      /*OwnsTokens=*/true);
> +}
> Index: lib/Parse/ParseStmt.cpp
> ===================================================================
> --- lib/Parse/ParseStmt.cpp (revision 208638)
> +++ lib/Parse/ParseStmt.cpp (working copy)
> @@ -15,11 +15,13 @@
>  #include "clang/Parse/Parser.h"
>  #include "RAIIObjectsForParser.h"
>  #include "clang/AST/ASTContext.h"
> +#include "clang/Basic/Attributes.h"
>  #include "clang/Basic/Diagnostic.h"
>  #include "clang/Basic/PrettyStackTrace.h"
>  #include "clang/Basic/SourceManager.h"
>  #include "clang/Basic/TargetInfo.h"
>  #include "clang/Sema/DeclSpec.h"
> +#include "clang/Sema/LoopHint.h"
>  #include "clang/Sema/PrettyDeclStackTrace.h"
>  #include "clang/Sema/Scope.h"
>  #include "clang/Sema/TypoCorrection.h"
> @@ -357,6 +359,10 @@
>      ProhibitAttributes(Attrs);
>      HandlePragmaMSPragma();
>      return StmtEmpty();
> +
> +  case tok::annot_pragma_loop_hint:
> +    ProhibitAttributes(Attrs);
> +    return ParsePragmaLoopHint(Stmts, OnlyStatement, TrailingElseLoc, Attrs);
>    }
>
>    // If we reached this code, the statement must end in a semicolon.
> @@ -1751,6 +1757,43 @@
>    return Actions.ActOnReturnStmt(ReturnLoc, R.take(), getCurScope());
>  }
>
> +StmtResult Parser::ParsePragmaLoopHint(StmtVector &Stmts,
> +                                       bool OnlyStatement,
> +                                       SourceLocation *TrailingElseLoc,
> +                                       ParsedAttributesWithRange &Attrs) {
> +  // Create temporary attribute list.
> +  ParsedAttributesWithRange TempAttrs(AttrFactory);
> +
> +  // Get vectorize hints and consume annotated token.
> +  while (Tok.is(tok::annot_pragma_loop_hint)) {
> +    LoopHint Hint = HandlePragmaLoopHint();
> +    ConsumeToken();
> +
> +    if (!Hint.LoopLoc ||
> +        !Hint.OptionLoc ||
> +        !Hint.ValueLoc)
> +      continue;
> +
> +    ArgsUnion ArgHints[] = {Hint.OptionLoc,
> +                            Hint.ValueLoc,
> +                            ArgsUnion(Hint.ValueExpr)};
> +    // FIXME: Replace AS_Keyword with Pragma spelling AS_Pragma.
> +    TempAttrs.addNew(Hint.LoopLoc->Ident, Hint.Range,
> +                     0, Hint.LoopLoc->Loc,
> +                     ArgHints, 3,
> +                     AttributeList::AS_Keyword);
> +  }
> +
> +  // Get the next statement.
> +  MaybeParseCXX11Attributes(Attrs);
> +
> +  StmtResult S = ParseStatementOrDeclarationAfterAttributes(Stmts,
> +                  OnlyStatement, TrailingElseLoc, Attrs);
> +
> +  Attrs.takeAllFrom(TempAttrs);
> +  return S;
> +}
> +
>  namespace {
>    class ClangAsmParserCallback : public llvm::MCAsmParserSemaCallback {
>      Parser &TheParser;
> Index: lib/Sema/SemaStmtAttr.cpp
> ===================================================================
> --- lib/Sema/SemaStmtAttr.cpp (revision 208638)
> +++ lib/Sema/SemaStmtAttr.cpp (working copy)
> @@ -16,6 +16,7 @@
>  #include "clang/Basic/SourceManager.h"
>  #include "clang/Sema/DelayedDiagnostic.h"
>  #include "clang/Sema/Lookup.h"
> +#include "clang/Sema/LoopHint.h"
>  #include "clang/Sema/ScopeInfo.h"
>  #include "llvm/ADT/StringExtras.h"
>
> @@ -42,7 +43,83 @@
>                                             A.getAttributeSpellingListIndex());
>  }
>
> +static Attr *handleLoopHintAttr(Sema &S, Stmt *St, const AttributeList &A,
> +                                SourceRange Range) {
> +  if (St->getStmtClass() != Stmt::DoStmtClass &&
> +      St->getStmtClass() != Stmt::ForStmtClass &&
> +      St->getStmtClass() != Stmt::CXXForRangeStmtClass &&
> +      St->getStmtClass() != Stmt::WhileStmtClass) {
> +    S.Diag(St->getLocStart(), diag::err_pragma_loop_precedes_nonloop);
> +    return nullptr;
> +  }
>
> +  IdentifierLoc *OptionLoc = A.getArgAsIdent(0);
> +  IdentifierInfo *OptionInfo = OptionLoc->Ident;
> +  IdentifierLoc *ValueLoc = A.getArgAsIdent(1);
> +  IdentifierInfo *ValueInfo = ValueLoc->Ident;
> +  Expr *ValueExpr = A.getArgAsExpr(2);
> +
> +  assert(OptionInfo && "Attribute must have valid option info.");
> +
> +  LoopHintAttr::OptionType Option = LoopHintAttr::Vectorize;
> +  if (OptionInfo->getName() == "vectorize")
> +    Option = LoopHintAttr::Vectorize;
> +  else if (OptionInfo->getName() == "interleave")
> +    Option = LoopHintAttr::Interleave;
> +
> +  LoopHintAttr::KindType Kind = LoopHintAttr::Value;
> +  if (ValueInfo && ValueInfo->getName() == "enable")
> +    Kind = LoopHintAttr::Enable;
> +  else if (ValueInfo && ValueInfo->getName() == "disable")
> +    Kind = LoopHintAttr::Disable;
> +
> +  // FIXME: We should support template parameters for the loop hint value.
> +  // See bug report #19610
> +  int ValueInt = 1; // No vectorization/interleaving when kind set to disable.
> +  if (Kind == LoopHintAttr::Value) {
> +    llvm::APSInt ValueAPS;
> +    if(!ValueExpr ||
> +       !ValueExpr->isIntegerConstantExpr(ValueAPS, S.Context) ||
> +       (ValueInt = ValueAPS.getSExtValue()) < 1) {
> +      S.Diag(ValueLoc->Loc, diag::err_pragma_loop_invalid_value)
> +        << LoopHintAttr::getOptionName(Option);
> +      return nullptr;
> +    }
> +  }
> +
> +  return LoopHintAttr::CreateImplicit(S.Context,
> +                                      Option, Kind, ValueInt, A.getRange());
> +}
> +
> +static void CheckForIncompatibleAttributes(Sema &S,
> +                                          SmallVectorImpl<const Attr*> &Attrs) {

Formatting.

> +  int PrevKind[2] = {-1, -1};
> +
> +  // FIXME: Attrs are stored in the reverse order than they appear.
> +  for (auto I = Attrs.rbegin(), E = Attrs.rend(); I != E; ++I) {
> +    const LoopHintAttr *LH = dyn_cast<LoopHintAttr>(*I);
> +
> +    // Skip non loop hint attributes
> +    if (!LH)
> +      continue;
> +
> +    int Option = LH->getOption();
> +    int Kind = LH->getKind();
> +    SourceLocation ValueLoc = LH->getRange().getEnd();
> +
> +    // We only need to check that a loop hint is compatible with the
> +    // previous loop hint to ensure that all hints are compatible.
> +    if (PrevKind[Option] != -1 &&
> +        !LoopHintAttr::isCompatible(PrevKind[Option], Kind)) {
> +      S.Diag(ValueLoc, diag::err_pragma_loop_incompatible)
> +        << LoopHintAttr::getKindName(PrevKind[Option])
> +        << LoopHintAttr::getKindName(Kind)
> +        << LoopHintAttr::getOptionName(Option);
> +    }
> +    PrevKind[Option] = Kind;
> +  }
> +}
> +
>  static Attr *ProcessStmtAttribute(Sema &S, Stmt *St, const AttributeList &A,
>                                    SourceRange Range) {
>    switch (A.getKind()) {
> @@ -53,6 +130,8 @@
>      return 0;
>    case AttributeList::AT_FallThrough:
>      return handleFallThroughAttr(S, St, A, Range);
> +  case AttributeList::AT_LoopHint:
> +    return handleLoopHintAttr(S, St, A, Range);
>    default:
>      // if we're here, then we parsed a known attribute, but didn't recognize
>      // it as a statement attribute => it is declaration attribute
> @@ -65,11 +144,14 @@
>  StmtResult Sema::ProcessStmtAttributes(Stmt *S, AttributeList *AttrList,
>                                         SourceRange Range) {
>    SmallVector<const Attr*, 8> Attrs;
> +  // FIXME: Attrs are stored in the reverse order than they appear.
>    for (const AttributeList* l = AttrList; l; l = l->getNext()) {
>      if (Attr *a = ProcessStmtAttribute(*this, S, *l, Range))
>        Attrs.push_back(a);
>    }
>
> +  CheckForIncompatibleAttributes(*this, Attrs);
> +
>    if (Attrs.empty())
>      return S;
>
> Index: test/CodeGen/pragma-loop.cpp
> ===================================================================
> --- test/CodeGen/pragma-loop.cpp (revision 0)
> +++ test/CodeGen/pragma-loop.cpp (working copy)
> @@ -0,0 +1,120 @@
> +// RUN: %clang_cc1 -std=c++11 -emit-llvm -o - %s | FileCheck %s
> +
> +// CHECK: br i1 %cmp, label %while.body, label %while.end, !llvm.loop !1
> +// CHECK: br i1 %cmp, label %do.body, label %do.end, !llvm.loop !5
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !7
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !8
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !11
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !13
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !14
> +// CHECK: br i1 %cmp, label %for.body, label %for.end, !llvm.loop !16
> +
> +// CHECK: !1 = metadata !{metadata !1, metadata !2, metadata !3, metadata !4}
> +// CHECK: !2 = metadata !{metadata !"llvm.vectorizer.enable", i1 true}
> +// CHECK: !3 = metadata !{metadata !"llvm.vectorizer.unroll", i32 4}
> +// CHECK: !4 = metadata !{metadata !"llvm.vectorizer.width", i32 4}
> +// CHECK: !5 = metadata !{metadata !5, metadata !6, metadata !3}
> +// CHECK: !6 = metadata !{metadata !"llvm.vectorizer.width", i32 8}
> +// CHECK: !7 = metadata !{metadata !7, metadata !2, metadata !3}
> +// CHECK: !8 = metadata !{metadata !8, metadata !9, metadata !10}
> +// CHECK: !9 = metadata !{metadata !"llvm.vectorizer.width", i32 2}
> +// CHECK: !10 = metadata !{metadata !"llvm.vectorizer.unroll", i32 2}
> +// CHECK: !11 = metadata !{metadata !11, metadata !12}
> +// CHECK: !12 = metadata !{metadata !"llvm.vectorizer.width", i32 1}
> +// CHECK: !13 = metadata !{metadata !13, metadata !9, metadata !10}
> +// CHECK: !14 = metadata !{metadata !14, metadata !6, metadata !15}
> +// CHECK: !15 = metadata !{metadata !"llvm.vectorizer.unroll", i32 8}
> +// CHECK: !16 = metadata !{metadata !16, metadata !9, metadata !10}
> +
> +// Verify while loop is recognized after sequence of pragma loop directives.
> +void while_test(int *List, int Length) {
> +  int i = 0;
> +
> +  #pragma loop vectorize(enable)
> +  #pragma loop interleave(4)
> +  #pragma loop vectorize(4)
> +  while(i < Length) {
> +    List[i] = i*2;
> +    i++;
> +  }
> +}
> +
> +// Verify do loop is recognized after multi-option pragma loop directive.
> +void do_test(int *List, int Length) {
> +  int i = 0;
> +
> +  #pragma loop vectorize(8) interleave(4)
> +  do {
> +    List[i] = i*2;
> +    i++;
> +  } while (i < Length);
> +}
> +
> +// Verify for loop is recognized after sequence of pragma loop directives.
> +void for_test(int *List, int Length) {
> +  #pragma loop interleave(enable)
> +  #pragma loop interleave(4)
> +  for(int i = 0; i < Length; i++) {
> +    List[i] = i*2;
> +  }
> +}
> +
> +// Verify c++11 for range loop is recognized after
> +// sequence of pragma loop directives.
> +void for_range_test() {
> +  double List[100];
> +
> +  #pragma loop vectorize(2) interleave(2)
> +  for (int i : List) {
> +    List[i] = i;
> +  }
> +}
> +
> +// Verify disable pragma loop directive generates correct metadata
> +void disable_test(int *List, int Length) {
> +  #pragma loop vectorize(disable)
> +  for(int i = 0; i < Length; i++) {
> +    List[i] = i*2;
> +  }
> +}
> +
> +#define VECWIDTH 2
> +#define INTERLEAVE 2
> +
> +// Verify defines are correctly resolved in pragma loop directive
> +void for_define_test(int *List, int Length, int Value) {
> +  #pragma loop vectorize(VECWIDTH) interleave(INTERLEAVE)
> +  for(int i = 0; i < Length; i++) {
> +    List[i] = i*Value;
> +  }
> +}
> +
> +// Verify metadata is generated when template is used.
> +template <typename A>
> +void for_template_test(A *List, int Length, A Value) {
> +
> +  #pragma loop vectorize(8) interleave(8)
> +  for(int i = 0; i < Length; i++) {
> +    List[i] = i*Value;
> +  }
> +}
> +
> +// Verify define is resolved correctly when template is used.
> +template <typename A>
> +void for_template_define_test(A *List, int Length, A Value) {
> +  #pragma loop vectorize(VECWIDTH) interleave(INTERLEAVE)
> +  for(int i = 0; i < Length; i++) {
> +    List[i] = i*Value;
> +  }
> +}
> +
> +#undef VECWIDTH
> +#undef INTERLEAVE
> +
> +// Use templates defined above. Test verifies metadata is generated correctly.
> +void template_test(double *List, int Length) {
> +  double Value = 10;
> +
> +  for_template_test<double>(List, Length, Value);
> +  for_template_define_test<double>(List, Length, Value);
> +}
> Index: test/PCH/pragma-loop.cpp
> ===================================================================
> --- test/PCH/pragma-loop.cpp (revision 0)
> +++ test/PCH/pragma-loop.cpp (working copy)
> @@ -0,0 +1,59 @@
> +// RUN: %clang_cc1 -emit-pch -o %t.a %s
> +// RUN: %clang_cc1 -include-pch %t.a %s -ast-print -o - | FileCheck %s
> +
> +// CHECK: #pragma loop vectorize(8)
> +// CHECK: #pragma loop interleave(4)
> +// CHECK: #pragma loop vectorize(disable)
> +// CHECK: #pragma loop interleave(enable)
> +// CHECK: #pragma loop vectorize(enable)
> +// CHECK: #pragma loop interleave(disable)
> +
> +#ifndef HEADER
> +#define HEADER
> +
> +class pragma_test {
> +public:
> +  inline void run1(int *List, int Length) {
> +    int i = 0;
> +    #pragma loop vectorize(8)
> +    #pragma loop interleave(4)
> +    while (i < Length) {
> +      List[i] = i;
> +      i++;
> +    }
> +  }
> +
> +  inline void run2(int *List, int Length) {
> +    int i = 0;
> +    #pragma loop vectorize(disable)
> +    #pragma loop interleave(enable)
> +    while (i-1 < Length) {
> +      List[i] = i;
> +      i++;
> +    }
> +  }
> +
> +  inline void run3(int *List, int Length) {
> +    int i = 0;
> +    #pragma loop vectorize(enable)
> +    #pragma loop interleave(disable)
> +    while (i-3 < Length) {
> +      List[i] = i;
> +      i++;
> +    }
> +  }
> +};
> +
> +#else
> +
> +void test() {
> +  int List[100];
> +
> +  pragma_test pt;
> +
> +  pt.run1(List, 100);
> +  pt.run2(List, 100);
> +  pt.run3(List, 100);
> +}
> +
> +#endif
> Index: test/Parser/pragma-loop-ast.cpp
> ===================================================================
> --- test/Parser/pragma-loop-ast.cpp (revision 0)
> +++ test/Parser/pragma-loop-ast.cpp (working copy)
> @@ -0,0 +1,33 @@
> +// RUN: %clang_cc1 -ast-print %s | FileCheck %s
> +
> +// CHECK: #pragma loop vectorize(4)
> +// CHECK: #pragma loop interleave(8)
> +// CHECK: #pragma loop vectorize(enable)
> +// CHECK: #pragma loop interleave(disable)
> +// CHECK: #pragma loop vectorize(disable)
> +// CHECK: #pragma loop interleave(enable)
> +
> +void test(int *List, int Length)
> +{
> +  int i = 0;
> +  #pragma loop vectorize(4)
> +  #pragma loop interleave(8)
> +  while (i < Length) {
> +    List[i] = i * 2;
> +    i++;
> +  }
> +
> +  #pragma loop vectorize(enable)
> +  #pragma loop interleave(disable)
> +  while (i-1 < Length) {
> +    List[i] = i * 2;
> +    i++;
> +  }
> +
> +  #pragma loop vectorize(disable)
> +  #pragma loop interleave(enable)
> +  while (i-2 < Length) {
> +    List[i] = i * 2;
> +    i++;
> +  }
> +}
> Index: test/Parser/pragma-loop.cpp
> ===================================================================
> --- test/Parser/pragma-loop.cpp (revision 0)
> +++ test/Parser/pragma-loop.cpp (working copy)
> @@ -0,0 +1,105 @@
> +// RUN: %clang_cc1 -fsyntax-only -std=c++11 -verify %s
> +
> +// Note that this puts the expected lines before the directives to work around
> +// limitations in the -verify mode.
> +
> +void test(int *List, int Length)
> +{
> +  int i = 0;
> +
> +  #pragma loop vectorize(4)
> +  #pragma loop interleave(8)
> +  while (i+1 < Length) {
> +    List[i] = i;
> +  }
> +
> +  #pragma loop vectorize(enable)
> +  #pragma loop interleave(enable)
> +  while (i < Length) {
> +    List[i] = i;
> +  }
> +
> +  #pragma loop vectorize(disable)
> +  #pragma loop interleave(disable)
> +  while (i-1 < Length) {
> +    List[i] = i;
> +  }
> +
> +  #pragma loop vectorize(4) interleave(8)
> +  while (i-2 < Length) {
> +    List[i] = i;
> +  }
> +
> +  #pragma loop interleave(16)
> +  while (i-3 < Length) {
> +    List[i] = i;
> +  }
> +
> +  int VList[Length];
> +  #pragma loop vectorize(disable) interleave(disable)
> +  for (int j : VList) {
> +    VList[j] = List[j];
> +  }
> +
> +/* expected-error {{expected '('}} */ #pragma loop vectorize
> +/* expected-error {{expected '('}} */ #pragma loop interleave
> +
> +/* expected-error {{expected ')'}} */ #pragma loop vectorize(4
> +/* expected-error {{expected ')'}} */ #pragma loop interleave(4
> +
> +/* expected-error {{missing option in directive '#pragma loop'}} */ #pragma loop
> +/* expected-error {{invalid option 'badkeyword' in directive '#pragma loop'}} */ #pragma loop badkeyword
> +/* expected-error {{invalid option 'badkeyword' in directive '#pragma loop'}} */ #pragma loop badkeyword(2)
> +/* expected-error {{invalid option 'badkeyword' in directive '#pragma loop'}} */ #pragma loop vectorize(4) badkeyword(4)
> +/* expected-warning {{extra tokens at end of '#pragma loop'}} */ #pragma loop vectorize(4) ,
> +
> +  while (i-4 < Length) {
> +    List[i] = i;
> +  }
> +
> +/* expected-error {{expected a positive integer in directive '#pragma loop vectorize'}} */ #pragma loop vectorize(0)
> +/* expected-error {{expected a positive integer in directive '#pragma loop interleave'}} */ #pragma loop interleave(0)
> +  while (i-5 < Length) {
> +    List[i] = i;
> +  }
> +
> +/* expected-error {{expected a positive integer in directive '#pragma loop vectorize'}} */ #pragma loop vectorize(3000000000)
> +/* expected-error {{expected a positive integer in directive '#pragma loop interleave'}} */ #pragma loop interleave(3000000000)
> +  while (i-6 < Length) {
> +    List[i] = i;
> +  }
> +
> +/* expected-error {{expected a positive integer in directive '#pragma loop vectorize'}} */ #pragma loop vectorize(badvalue)
> +/* expected-error {{expected a positive integer in directive '#pragma loop interleave'}} */ #pragma loop interleave(badvalue)
> +  while (i-7 < Length) {
> +    List[i] = i;
> +  }
> +
> +#pragma loop vectorize(enable)
> +/* expected-error {{expected a for, while, or do-while loop to follow the '#pragma loop' directive}} */ int j = Length;
> +  List[0] = List[1];
> +
> +  while (j-1 < Length) {
> +    List[j] = j;
> +  }
> +
> +#pragma loop vectorize(4)
> +/* expected-error {{'value' and 'disable' directive option types are incompatible in '#pragma loop vectorize'}} */ #pragma loop vectorize(disable)
> +  while (i-8 < Length) {
> +    List[i] = i;
> +  }
> +
> +#pragma loop interleave(enable)
> +/* expected-error {{'enable' and 'disable' directive option types are incompatible in '#pragma loop interleave'}} */ #pragma loop interleave(disable)
> +  while (i-9 < Length) {
> +    List[i] = i;
> +  }
> +
> +#pragma loop vectorize(disable)
> +/* expected-error {{'disable' and 'value' directive option types are incompatible in '#pragma loop vectorize'}} */ #pragma loop vectorize(4)
> +  while (i-10 < Length) {
> +    List[i] = i;
> +  }
> +
> +#pragma loop interleave(enable)
> +/* expected-error {{expected statement}} */ }
>

~Aaron




More information about the cfe-commits mailing list