[PATCH] D69585: PerformPendingInstatiations() already in the PCH

[Luboš Luňák via Phabricator via cfe-commits]

llunak created this revision.
llunak added reviewers: ABataev, rsmith, dblaikie.
llunak added projects: clang, OpenMP.
Herald added a reviewer: jdoerfert.
Herald added a subscriber: cfe-commits.

This patch makes template instantiations be already performed in the PCH instead of it being done in every single file that uses the PCH. I can see 20-30% build time saved with the few tests I've tried.

The delaying of template instantiations until the PCH is used comes from 7f76d11dcc9196e1fc9d1308da9ed2330a6b06c2 , which doesn't really give any useful reasoning for it, except for extending a unittest, which however now passes even with the instantiation moved back into the PCH. Since modules do not delay instantiation either, presumably whatever was wrong there has been fixed already.

I've built a complete debug build of LibreOffice with the patch and everything seems to work fine.

All Clang tests pass fine as well, with the exception of 23 tests, 22 of which are in OpenMP:

- OpenMP/declare_target_codegen.cpp fails because some virtuals of template classes are not emitted. This is fixed by the second change in the patch. I do not understand the purpose of that code (neither I have any clue about OpenMP), but it was introduced in d2c1dd5902782fb773c64dbf4e0b529aa4044b05 , which also changed this very test, and the change makes the test pass again.
- The remaining 22 tests, CodeGenCXX/vla-lambda-capturing.cpp and 21 in OpenMP/, all use the same FileCheck for normal compilation and using the source file as PCH input. Instantiating already in the PCH reorders some things, which makes the tests fail. Looking at the differences in the generated output, they all seem to be functionally equivalent to me, they just do not match exactly anymore.

That obviously makes the patch not ready, but I'd like to get feedback on this, before spending the time on adjusting the tests (which I expect will be quite a hassle[*]). So, is there any problem with this patch, besides adjusting the tests? And is there some easier way to handle those 21 OpenMP tests besides possibly (up to) doubling the checks in them?

[X] BTW, https://pastebin.com/Dg2L7K27 helped quite a lot with reviewing the differences.


Repository:
  rC Clang

https://reviews.llvm.org/D69585

Files:
  clang/lib/Sema/Sema.cpp
  clang/lib/Sema/SemaDeclCXX.cpp


Index: clang/lib/Sema/SemaDeclCXX.cpp
===================================================================
--- clang/lib/Sema/SemaDeclCXX.cpp
+++ clang/lib/Sema/SemaDeclCXX.cpp
@@ -15569,7 +15569,7 @@
     return;
   // Do not mark as used if compiling for the device outside of the target
   // region.
-  if (LangOpts.OpenMP && LangOpts.OpenMPIsDevice &&
+  if (TUKind != TU_Prefix && LangOpts.OpenMP && LangOpts.OpenMPIsDevice &&
       !isInOpenMPDeclareTargetContext() &&
       !isInOpenMPTargetExecutionDirective()) {
     if (!DefinitionRequired)
Index: clang/lib/Sema/Sema.cpp
===================================================================
--- clang/lib/Sema/Sema.cpp
+++ clang/lib/Sema/Sema.cpp
@@ -981,6 +981,12 @@
                                  LateParsedInstantiations.begin(),
                                  LateParsedInstantiations.end());
     LateParsedInstantiations.clear();
+
+    {
+      llvm::TimeTraceScope TimeScope("PerformPendingInstantiations",
+                                     StringRef(""));
+      PerformPendingInstantiations();
+    }
   }
 
   DiagnoseUnterminatedPragmaPack();


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D69585.226957.patch
Type: text/x-patch
Size: 1127 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20191029/427cfae7/attachment.bin>