[cfe-commits] r161650 - in /cfe/trunk: include/clang/Frontend/VerifyDiagnosticConsumer.h lib/ARCMigrate/ARCMT.cpp lib/Frontend/VerifyDiagnosticConsumer.cpp test/ARCMT/verify.m test/ASTMerge/function.c test/Frontend/verify.c test/Frontend/verify2.c test/Frontend/veri

[Andy Gibbs]

On Thursday, August 16, 2012 6:28 PM, Jordan Rose wrote:
> Hm. Now I'm having second thoughts about this:
>
>> fatal_error:
>>       llvm::report_fatal_error(Twine("presence of -verify directives 
>> cannot be "
>>                                      "verified during post-processing 
>> check of ",
>>                                StringRef(FE ? FE->getName() : 
>> "(unknown)")));
>
> This happens if we don't have a SourceManager or Preprocessor available.
> But what if we really /don't/ have a Preprocessor available?

The best, of course, is that by the time we get to the post-processing
check, there are *no* unparsed files.  Thinking about it again, what could
be done is this:

(this is a hybrid between your suggestion and my previous approach)

Store two lists:

  Map<FileID, FileEntry*> Parsed
  Map<FileID, Pair<FileEntry*, bool FoundDirectives>> Unparsed

---When we enter a file from the preprocessor---
remove from unparsed set
add to parsed set

---When we get a diagnostic---
if (is a macro expansion)
  ignore [[big culprit this one!]]
if (not in parsed set)
  if (is modules header)
    add to parsed set
  else
    run through findDirectives
    add to unparsed set, setting FoundDirectives flag appropriately

---When you check at the end---
if (unparsed set not empty)
  build set of parsed FileEntry pointers
  iterate through unparsed set:
    ignore all files with aliases
    fatal error if FoundDirective flag set


This way, SourceManager and Preprocessor instances are not needed at all
at the end check.  SourceManager *is* needed for checking diagnostics
lists, but this is another thing.  This is the big advantage of this
approach: the end check is really freed of its constraints, and
becomes a very fast loop (all the data is already gathered).

The downside, but a small one perhaps?, is that findDirectives may be
called more times than necessary.  Analysing the situation: in most
cases, a file will be in the parsed set before it produces a diagnostic.
Those that aren't are in most cases actually macro expansions and should
therefore be ignored.  Those that fall outside that are PCH, modules and
-ast-merge tests.  Of these, its the PCH and -ast-merge tests that will
call findDirectives on their included files, and therefore be handled
less efficiently (potentially) this way.

Now, for findDirectives, we can (I think) get rid of the Preprocessor
requirement since the only thing it is really needed for is the LangOpts
structure for the Lexer -- the rest is freely available elsewhere.  (And
LangOpts can be picked up from BeginSourceFile, so even this shouldn't be
a problem.)

This means we really start to be quite free from the Preprocessor.

What do you think to this approach?

Andy