Changes to IdentifierTable:
- High-level summary: StringMap never owns IdentifierInfos. It just
references them.
- The string map now has StringMapEntry<IdentifierInfo*> instead of
StringMapEntry<IdentifierInfo>. The IdentifierInfo object is
allocated using the same bump pointer allocator as used by the
StringMap.
Changes to IdentifierInfo:
- Added an extra pointer to point to the
StringMapEntry<IdentifierInfo*> in the string map. This pointer
will be null if the IdentifierInfo* is *only* used by the PTHLexer
(that is it isn't in the StringMap).
Algorithmic changes:
- Non-PTH case:
IdentifierInfo::get() will always consult the StringMap first to
see if we have an IdentifierInfo object. If that StringMapEntry
references a null pointer, we allocate a new one from the BumpPtrAllocator
and update the reference in the StringMapEntry.
- PTH case:
We do the same lookup as with the non-PTH case, but if we don't get
a hit in the StringMap we do a secondary lookup in the PTHManager for
the IdentifierInfo. If we don't find an IdentifierInfo we create a
new one as in the non-PTH case. If we do find and IdentifierInfo
in the PTHManager, we update the StringMapEntry to refer to it so
that the IdentifierInfo will be found on the next StringMap lookup.
This way we only do a binary search in the PTH file at most once
for a given IdentifierInfo. This greatly speeds things up for source
files containing a non-trivial amount of code.
Performance impact:
While these changes do add some extra indirection in
IdentifierTable to access an IdentifierInfo*, I saw speedups even
in the non-PTH case as well.
Non-PTH: For -fsyntax-only on Cocoa.h, we see a 6% speedup.
PTH (with Cocoa.h in token cache): 11% speedup.
I also did an experiment where we did -fsyntax-only on a source file
including a large header and Cocoa.h, but the token cache did not
contain the larger header. For this file, we were seeing a performance
*regression* when using PTH of 3% over non-PTH. Now we are seeing
a performance improvement of 9%!
Tests:
The serialization tests are now failing. I looked at this extensively,
and I my belief is that this change is unmasking a bug rather than
introducing a new one. I have disabled the serialization tests for now.
llvm-svn: 62636
the chunk ID not the file ID. This exposes problems in
TextDiagnosticPrinter where it should have been using the canonical
file ID but wasn't. Fix these along the way.
llvm-svn: 62427
method. This lets us clean up the interface and make it more obvious that
this method is *really really* _Pragma specific.
Note that _Pragma handling uglifies the Lexer in the critical path. It would
be very interesting to consider making _Pragma remapping be a new special
lexer class of its own.
llvm-svn: 62425
"FileID" a concept that is now enforced by the compiler's type checker
instead of yet-another-random-unsigned floating around.
This is an important distinction from the "FileID" currently tracked by
SourceLocation. *That* FileID may refer to the start of a file or to a
chunk within it. The new FileID *only* refers to the file (and its
#include stack and eventually #line data), it cannot refer to a chunk.
FileID is a completely opaque datatype to all clients, only SourceManager
is allowed to poke and prod it.
llvm-svn: 62407
if warnings in system headers are disabled. isIdenticalTo can end up
calling the expensive getSpelling method, and other bad stuff and is
completely unneeded if the warning will be discarded anyway. rdar://6502956
llvm-svn: 62347
the "physical" location of tokens, refer to the "spelling" location.
This is more concrete and useful, tokens aren't really physical objects!
llvm-svn: 62309
- IdentifierInfo can now (optionally) have its string data not be
co-located with itself. This is for use with PTH. This aspect is a
little gross, as getName() and getLength() now make assumptions
about a possible alternate representation of IdentifierInfo.
Perhaps we should make IdentifierInfo have virtual methods?
IdentifierTable:
- Added class "IdentifierInfoLookup" that can be used by
IdentifierTable to perform "string -> IdentifierInfo" lookups using
an auxilliary data structure. This is used by PTH.
- Perform tests show that IdentifierTable::get() does not slow down
because of the extra check for the IdentiferInfoLookup object (the
regular StringMap lookup does enough work to mitigate the impact of
an extra null pointer check).
- The upshot is that now that some IdentifierInfo objects might be
owned by the IdentiferInfoLookup object. This should be reviewed.
PTH:
- Modified PTHManager::GetIdentifierInfo to *not* insert entries in
IdentifierTable's string map, and instead create IdentifierInfo
objects on the fly when mapping from persistent IDs to
IdentifierInfos. This saves a ton of work with string copies,
hashing, and StringMap lookup and resizing. This change was
motivated because when processing source files in the PTH cache we
don't need to do any string -> IdentifierInfo lookups.
- PTHManager now subclasses IdentifierInfoLookup, allowing clients of
IdentifierTable to transparently use IdentifierInfo objects managed
by the PTH file. PTHManager resolves "string -> IdentifierInfo"
queries by doing a binary search over a sorted table of identifier
strings in the PTH file (the exact algorithm we use can be changed
as needed).
These changes lead to the following performance changes when using PTH on Cocoa.h:
- fsyntax-only: 10% performance improvement
- Eonly: 30% performance improvement
llvm-svn: 62273
lexical order of the corresponding identifier strings. This will be used for a
forthcoming optimization. This slows down PTH generation time by 7%. We can
revert this change if the optimization proves to not be valuable.
llvm-svn: 62248
- Use canonical FileID when using getSpelling() caching. This
addresses some cache misses we were seeing with -fsyntax-only on
Cocoa.h
- Added Preprocessor::getPhysicalCharacterAt() utility method for
clients to grab the first character at a specified sourcelocation.
This uses the PTH spelling cache.
- Modified Sema::ActOnNumericConstant() to use
Preprocessor::getPhysicalCharacterAt() instead of
SourceManager::getCharacterData() (to get PTH hits).
These changes cause -fsyntax-only to not page in any sources from
Cocoa.h. We see a speedup of 27%.
llvm-svn: 62193
- Refactor caching logic into a helper class PTHSpellingSearch
- Allow "random accesses" in the spelling cache, thus catching the remaining
cases where 'getSpelling' wasn't hitting the PTH cache
For -Eonly, PTH, Cocoa.h:
- This reduces wall time by 3% (user time unchanged, sys time reduced)
- This reduces the amount of paged source by 1112K.
The remaining 1112K still being paged in is from somewhere else
(investigating).
llvm-svn: 62009
performance gain. Here's what we see for -Eonly on Cocoa.h (using PTH):
- wall time decreases by 21% (26% speedup overall)
- system time decreases by 35%
- user time decreases by 6%
These reductions are due to not paging source files just to get spellings for
literals. The solution in place doesn't appear to be 100% yet, as we still see
some of the pages for source files getting mapped in. Using -print-stats, we see
that SourceManager maps in 7179K less bytes of source text (reduction of 75%).
Will investigate why the remaining 25% are getting paged in.
With these changes, here's how PTH compares to non-PTH on Cocoa.h:
-Eonly: PTH takes 64% of the time as non-PTH (54% speedup)
-fsyntax-only: PTH takes 89% of the time as non-PTH (11% speedup)
llvm-svn: 61913
Ted Kremenek