r228234 - [modules] When using -E, we may try to merge decls despite having no Sema

Richard Smith richard at metafoo.co.uk
Mon Mar 23 17:06:45 PDT 2015


On Mon, Mar 23, 2015 at 5:01 PM, Ben Langmuir <blangmuir at apple.com> wrote:

>
> On Mar 23, 2015, at 4:18 PM, Richard Smith <richard at metafoo.co.uk> wrote:
>
> On Mon, Mar 23, 2015 at 3:59 PM, Ben Langmuir <blangmuir at apple.com> wrote:
>
>>
>> On Mar 23, 2015, at 3:43 PM, Richard Smith <richard at metafoo.co.uk> wrote:
>>
>> On Mon, Mar 23, 2015 at 2:41 PM, Ben Langmuir <blangmuir at apple.com>
>> wrote:
>>
>>>
>>> On Mar 23, 2015, at 1:05 PM, Richard Smith <richard at metafoo.co.uk>
>>> wrote:
>>>
>>> On Sat, Mar 14, 2015 at 9:25 AM, Ben Langmuir <blangmuir at apple.com>
>>> wrote:
>>>
>>>> Hi Richard,
>>>>
>>>> This commit looks like it caused a regression in PCH-loading
>>>> performance when the PCH imports modules.  It doesn’t affect performance of
>>>> loading modules on their own, or of a PCH that doesn’t use modules.  If you
>>>> have access to a Darwin system, it’s easy to reproduce with:
>>>>
>>>> $ cat t.h
>>>> @import Cocoa;
>>>>
>>>> $ cat t.m
>>>> // empty
>>>>
>>>> $ clang -fmodules -fmodules-cache-path=mcp -x objective-c-header t.h -o
>>>> t.pch
>>>> $ time clang -fmodules -fmodules-cache-path=mcp -Xclang -include-pch
>>>> -Xclang t.pch t.m -fsyntax-only
>>>>
>>>> I’m seeing ~8x difference between r228233 and r228234.  In real world
>>>> code (where the .m file isn’t empty) we’ve seen up to a ~90% difference.
>>>>
>>>> Do you know what’s going on here?
>>>
>>>
>>> Hmm. I suspect this change causes the lookup table for the translation
>>> unit to be built, which in turn pulls in all the lexical contents of the
>>> translation unit (because outside of C++ we don't serialize a DeclContext
>>> lookup table for the TU, nor even build one usually).
>>>
>>> When a PCH loads I think we do it without Sema, but (at least
>>>> previously) queued up the interesting Decls we saw for when Sema was added
>>>> later.  So I’m surprised this had such a big effect.
>>>
>>>
>>> If we were loading PCHs without Sema, then I think before this change we
>>> would not have merged declarations of the same entity from two independent
>>> modules imported into the same PCH.
>>>
>>>
>>> Wouldn’t that happen in InitializeSema() when we
>>> called pushExternalDeclIntoScope() on the contents of PreloadedDeclIDs?
>>>
>>
>> No; we're talking about the case where a Decl is deserialized before we
>> have a Sema instance. If we never deserialized the Decl from its DeclID
>> before we had a Sema (which I think is what you're suggesting) we'd never
>> have reached the modified code.
>>
>>
>> If we want to support loading Decls without Sema, we need to be able to
>> merge them without Sema. If we don't, then we should find whichever
>> codepath is leading us to do so and fix it. Being able to load a serialized
>> AST without a Sema seems useful to me, in the absence of a good reason why
>> it shouldn't work.
>>
>>
>> I thought PCH intentionally wouldn’t deserialize decls until after we
>> initialized Sema.   Maybe we have a bug (which wouldn’t surprise me), or
>> maybe that wasn’t the intent (which would scare me a bit since then we have
>> some decls showing up immediately and some waiting for sema with no clear
>> split between them).
>>
>> Perhaps we should use a separate side-table for merging in this case
>>> outside C++, rather than falling back to DeclContext lookup?
>>>
>>>
>>> I don’t know this code path well enough to say.  What would go into the
>>> table (or perhaps more importantly what wouldn’t)?
>>>
>>
>> Any deserialized declarations that should be visible by name lookup in
>> the TU would be in the table. The only difference would be that we'd be
>> more careful to ensure that we never trigger the import of all lexical
>> declarations in the TU in order to build the table.
>>
>> Are you still seeing this performance problem with Clang trunk? Towards
>> the end of last week I fixed a bug where we would unnecessarily pull in all
>> lexical declarations within the TU when doing certain lookups, which might
>> have fixed this slowdown. (If not, it'd be useful to find out whether
>> that's the problem, and if so, what is causing us to load these
>> declarations.)
>>
>>
>> Cool, I’ll give ToT a shot.  I last tried with r232229.
>>
>
> r232928 onwards, or a revision in [232793, 232905), have the change I'm
> referring to.
>
>
> ToT got faster, but not even close to pre-228234.
>
> r228233: 0.0574 s <== “before”
> r228234: 0.4433 s <== “after"
> r232229: 0.4028 s
> r233028: 0.3097 s <== ToT
>

Where's ToT spending its time?


> Ben
>
> Ben
>>>>
>>>> > On Feb 4, 2015, at 3:37 PM, Richard Smith <richard-llvm at metafoo.co.uk>
>>>> wrote:
>>>> >
>>>> > Author: rsmith
>>>> > Date: Wed Feb  4 17:37:59 2015
>>>> > New Revision: 228234
>>>> >
>>>> > URL: http://llvm.org/viewvc/llvm-project?rev=228234&view=rev
>>>> > Log:
>>>> > [modules] When using -E, we may try to merge decls despite having no
>>>> Sema
>>>> > object. In such a case, use the TU's DC for merging global decls
>>>> rather than
>>>> > giving up when we find there is no TU scope.
>>>> >
>>>> > Ultimately, we should probably avoid all loading of decls when
>>>> preprocessing,
>>>> > but there are other reasonable use cases for loading an AST file with
>>>> no Sema
>>>> > object for which this is the right thing.
>>>> >
>>>> > Added:
>>>> >    cfe/trunk/test/Modules/Inputs/preprocess/
>>>> >    cfe/trunk/test/Modules/Inputs/preprocess/file.h
>>>> >    cfe/trunk/test/Modules/Inputs/preprocess/fwd.h
>>>> >    cfe/trunk/test/Modules/Inputs/preprocess/module.modulemap
>>>> > Modified:
>>>> >    cfe/trunk/include/clang/Frontend/CompilerInstance.h
>>>> >    cfe/trunk/lib/Frontend/CompilerInstance.cpp
>>>> >    cfe/trunk/lib/Serialization/ASTReaderDecl.cpp
>>>> >    cfe/trunk/test/Modules/preprocess.m
>>>> >
>>>> > Modified: cfe/trunk/include/clang/Frontend/CompilerInstance.h
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/include/clang/Frontend/CompilerInstance.h?rev=228234&r1=228233&r2=228234&view=diff
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/include/clang/Frontend/CompilerInstance.h (original)
>>>> > +++ cfe/trunk/include/clang/Frontend/CompilerInstance.h Wed Feb  4
>>>> 17:37:59 2015
>>>> > @@ -588,7 +588,7 @@ public:
>>>> >   /// Create an external AST source to read a PCH file.
>>>> >   ///
>>>> >   /// \return - The new object on success, or null on failure.
>>>> > -  static ExternalASTSource *createPCHExternalASTSource(
>>>> > +  static IntrusiveRefCntPtr<ASTReader> createPCHExternalASTSource(
>>>> >       StringRef Path, const std::string &Sysroot, bool
>>>> DisablePCHValidation,
>>>> >       bool AllowPCHWithCompilerErrors, Preprocessor &PP, ASTContext
>>>> &Context,
>>>> >       void *DeserializationListener, bool OwnDeserializationListener,
>>>> >
>>>> > Modified: cfe/trunk/lib/Frontend/CompilerInstance.cpp
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Frontend/CompilerInstance.cpp?rev=228234&r1=228233&r2=228234&view=diff
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/lib/Frontend/CompilerInstance.cpp (original)
>>>> > +++ cfe/trunk/lib/Frontend/CompilerInstance.cpp Wed Feb  4 17:37:59
>>>> 2015
>>>> > @@ -388,32 +388,30 @@ void CompilerInstance::createASTContext(
>>>> > void CompilerInstance::createPCHExternalASTSource(
>>>> >     StringRef Path, bool DisablePCHValidation, bool
>>>> AllowPCHWithCompilerErrors,
>>>> >     void *DeserializationListener, bool OwnDeserializationListener) {
>>>> > -  IntrusiveRefCntPtr<ExternalASTSource> Source;
>>>> >   bool Preamble =
>>>> getPreprocessorOpts().PrecompiledPreambleBytes.first != 0;
>>>> > -  Source = createPCHExternalASTSource(
>>>> > +  ModuleManager = createPCHExternalASTSource(
>>>> >       Path, getHeaderSearchOpts().Sysroot, DisablePCHValidation,
>>>> >       AllowPCHWithCompilerErrors, getPreprocessor(), getASTContext(),
>>>> >       DeserializationListener, OwnDeserializationListener, Preamble,
>>>> >       getFrontendOpts().UseGlobalModuleIndex);
>>>> > -  ModuleManager = static_cast<ASTReader*>(Source.get());
>>>> > -  getASTContext().setExternalSource(Source);
>>>> > }
>>>> >
>>>> > -ExternalASTSource *CompilerInstance::createPCHExternalASTSource(
>>>> > +IntrusiveRefCntPtr<ASTReader>
>>>> CompilerInstance::createPCHExternalASTSource(
>>>> >     StringRef Path, const std::string &Sysroot, bool
>>>> DisablePCHValidation,
>>>> >     bool AllowPCHWithCompilerErrors, Preprocessor &PP, ASTContext
>>>> &Context,
>>>> >     void *DeserializationListener, bool OwnDeserializationListener,
>>>> >     bool Preamble, bool UseGlobalModuleIndex) {
>>>> >   HeaderSearchOptions &HSOpts =
>>>> PP.getHeaderSearchInfo().getHeaderSearchOpts();
>>>> >
>>>> > -  std::unique_ptr<ASTReader> Reader;
>>>> > -  Reader.reset(new ASTReader(PP, Context,
>>>> > -                             Sysroot.empty() ? "" : Sysroot.c_str(),
>>>> > -                             DisablePCHValidation,
>>>> > -                             AllowPCHWithCompilerErrors,
>>>> > -                             /*AllowConfigurationMismatch*/false,
>>>> > -                             HSOpts.ModulesValidateSystemHeaders,
>>>> > -                             UseGlobalModuleIndex));
>>>> > +  IntrusiveRefCntPtr<ASTReader> Reader(
>>>> > +      new ASTReader(PP, Context, Sysroot.empty() ? "" :
>>>> Sysroot.c_str(),
>>>> > +                    DisablePCHValidation, AllowPCHWithCompilerErrors,
>>>> > +                    /*AllowConfigurationMismatch*/ false,
>>>> > +                    HSOpts.ModulesValidateSystemHeaders,
>>>> UseGlobalModuleIndex));
>>>> > +
>>>> > +  // We need the external source to be set up before we read the
>>>> AST, because
>>>> > +  // eagerly-deserialized declarations may use it.
>>>> > +  Context.setExternalSource(Reader.get());
>>>> >
>>>> >   Reader->setDeserializationListener(
>>>> >       static_cast<ASTDeserializationListener
>>>> *>(DeserializationListener),
>>>> > @@ -427,7 +425,7 @@ ExternalASTSource *CompilerInstance::cre
>>>> >     // Set the predefines buffer as suggested by the PCH reader.
>>>> Typically, the
>>>> >     // predefines buffer will be empty.
>>>> >     PP.setPredefines(Reader->getSuggestedPredefines());
>>>> > -    return Reader.release();
>>>> > +    return Reader;
>>>> >
>>>> >   case ASTReader::Failure:
>>>> >     // Unrecoverable failure: don't even try to process the input
>>>> file.
>>>> > @@ -442,6 +440,7 @@ ExternalASTSource *CompilerInstance::cre
>>>> >     break;
>>>> >   }
>>>> >
>>>> > +  Context.setExternalSource(nullptr);
>>>> >   return nullptr;
>>>> > }
>>>> >
>>>> >
>>>> > Modified: cfe/trunk/lib/Serialization/ASTReaderDecl.cpp
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Serialization/ASTReaderDecl.cpp?rev=228234&r1=228233&r2=228234&view=diff
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/lib/Serialization/ASTReaderDecl.cpp (original)
>>>> > +++ cfe/trunk/lib/Serialization/ASTReaderDecl.cpp Wed Feb  4 17:37:59
>>>> 2015
>>>> > @@ -2591,6 +2591,11 @@ DeclContext *ASTDeclReader::getPrimaryCo
>>>> >     return ED->getASTContext().getLangOpts().CPlusPlus?
>>>> ED->getDefinition()
>>>> >                                                       : nullptr;
>>>> >
>>>> > +  // We can see the TU here only if we have no Sema object. In that
>>>> case,
>>>> > +  // there's no TU scope to look in, so using the DC alone is
>>>> sufficient.
>>>> > +  if (auto *TU = dyn_cast<TranslationUnitDecl>(DC))
>>>> > +    return TU;
>>>> > +
>>>> >   return nullptr;
>>>> > }
>>>> >
>>>> >
>>>> > Added: cfe/trunk/test/Modules/Inputs/preprocess/file.h
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/test/Modules/Inputs/preprocess/file.h?rev=228234&view=auto
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/test/Modules/Inputs/preprocess/file.h (added)
>>>> > +++ cfe/trunk/test/Modules/Inputs/preprocess/file.h Wed Feb  4
>>>> 17:37:59 2015
>>>> > @@ -0,0 +1,3 @@
>>>> > +struct __FILE;
>>>> > +#include "fwd.h"
>>>> > +typedef struct __FILE FILE;
>>>> >
>>>> > Added: cfe/trunk/test/Modules/Inputs/preprocess/fwd.h
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/test/Modules/Inputs/preprocess/fwd.h?rev=228234&view=auto
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/test/Modules/Inputs/preprocess/fwd.h (added)
>>>> > +++ cfe/trunk/test/Modules/Inputs/preprocess/fwd.h Wed Feb  4
>>>> 17:37:59 2015
>>>> > @@ -0,0 +1 @@
>>>> > +struct __FILE;
>>>> >
>>>> > Added: cfe/trunk/test/Modules/Inputs/preprocess/module.modulemap
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/test/Modules/Inputs/preprocess/module.modulemap?rev=228234&view=auto
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/test/Modules/Inputs/preprocess/module.modulemap (added)
>>>> > +++ cfe/trunk/test/Modules/Inputs/preprocess/module.modulemap Wed
>>>> Feb  4 17:37:59 2015
>>>> > @@ -0,0 +1,2 @@
>>>> > +module fwd { header "fwd.h" export * }
>>>> > +module file { header "file.h" export * }
>>>> >
>>>> > Modified: cfe/trunk/test/Modules/preprocess.m
>>>> > URL:
>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/test/Modules/preprocess.m?rev=228234&r1=228233&r2=228234&view=diff
>>>> >
>>>> ==============================================================================
>>>> > --- cfe/trunk/test/Modules/preprocess.m (original)
>>>> > +++ cfe/trunk/test/Modules/preprocess.m Wed Feb  4 17:37:59 2015
>>>> > @@ -1,9 +1,14 @@
>>>> > // RUN: rm -rf %t
>>>> > -// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs
>>>> -include %S/Inputs/preprocess-prefix.h -E %s | FileCheck -strict-whitespace
>>>> %s
>>>> > -// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -x
>>>> objective-c-header -emit-pch %S/Inputs/preprocess-prefix.h -o %t.pch
>>>> > -// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs
>>>> -include-pch %t.pch -E %s | FileCheck -strict-whitespace %s
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -include %S/Inputs/preprocess-prefix.h -E %s |
>>>> FileCheck -strict-whitespace %s
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -x objective-c-header -emit-pch
>>>> %S/Inputs/preprocess-prefix.h -o %t.pch
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -include-pch %t.pch -E %s | FileCheck
>>>> -strict-whitespace %s
>>>> > +//
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -x objective-c++ -include
>>>> %S/Inputs/preprocess-prefix.h -E %s | FileCheck -strict-whitespace %s
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -x objective-c++-header -emit-pch
>>>> %S/Inputs/preprocess-prefix.h -o %t.pch
>>>> > +// RUN: %clang_cc1 -fmodules -fmodules-cache-path=%t -I %S/Inputs -I
>>>> %S/Inputs/preprocess -x objective-c++ -include-pch %t.pch -E %s | FileCheck
>>>> -strict-whitespace %s
>>>> > #import "diamond_right.h"
>>>> > #import "diamond_right.h" // to check that imports get their own line
>>>> > +#include "file.h"
>>>> > void test() {
>>>> >   top_left_before();
>>>> >   left_and_right();
>>>> > @@ -15,6 +20,7 @@ void test() {
>>>> >
>>>> > // CHECK: @import diamond_right; /* clang -E: implicit import for
>>>> "{{.*}}diamond_right.h" */{{$}}
>>>> > // CHECK: @import diamond_right; /* clang -E: implicit import for
>>>> "{{.*}}diamond_right.h" */{{$}}
>>>> > +// CHECK: @import file; /* clang -E: implicit import for
>>>> "{{.*}}file.h" */{{$}}
>>>> > // CHECK-NEXT: void test() {{{$}}
>>>> > // CHECK-NEXT:    top_left_before();{{$}}
>>>> > // CHECK-NEXT:    left_and_right();{{$}}
>>>> >
>>>> >
>>>> > _______________________________________________
>>>> > cfe-commits mailing list
>>>> > cfe-commits at cs.uiuc.edu
>>>> > http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150323/443836bd/attachment.html>


More information about the cfe-commits mailing list