Daniel Dunbar [Sat, 10 Mar 2012 01:03:58 +0000 (01:03 +0000)]
IRgen/ABI/x86_64: Avoid passing small structs using byval sometimes.
- We do this when it is easy to determine that the backend will pass them on
the stack properly by itself.
Currently LLVM codegen is really bad in some cases with byval, for example, on
the test case here (which is derived from Sema code, which likes to pass
SourceLocations around)::
struct s47 { unsigned a; };
void f47(int,int,int,int,int,int,struct s47);
void test47(int a, struct s47 b) { f47(a, a, a, a, a, a, b); }
Richard Smith [Sat, 10 Mar 2012 00:28:11 +0000 (00:28 +0000)]
Assign APValues by swapping from a temporary. Removes a bunch of unnecessary
copy-construction, which Daniel Dunbar reports as giving a 0.75% speedup on
403.gcc/combine.c. The performance differences on my constexpr torture tests
are below the noise floor.
Richard Smith [Fri, 9 Mar 2012 22:27:51 +0000 (22:27 +0000)]
Improve diagnostics for UCNs referring to control characters and members of the
basic source character set in C++98. Add -Wc++98-compat diagnostics for same in
literals in C++11. Extend such support to cover string literals as well as
character literals, and mark N2170 as done.
This seems too minor to warrant a release note to me. Let me know if you disagree.
Anna Zaks [Fri, 9 Mar 2012 21:14:01 +0000 (21:14 +0000)]
[analyzer] Add support for NoRedundancy inlining mode.
We do not reanalyze a function, which has already been analyzed as an
inlined callee. As per PRELIMINARY testing, this gives over
50% run time reduction on some benchmarks without decreasing of the
number of bugs found.
Daniel Dunbar [Fri, 9 Mar 2012 20:41:19 +0000 (20:41 +0000)]
Revert r152430 "[AST] Mark ASTContext::getTypeInfo and some friends as
LLVM_READONLY.", getTypeInfo() is totally not READONLY, which I should have
probably noticed given that I made it so mere hours ago. Oops.
When an error made a record member invalid, the record would stay as "isBeingDefined" and
not "completeDefinition". Even easily recoverable errors ended up propagating records in
such "beingDefined" state, for example:
struct A {
~A() const; // expected-error {{'const' qualifier is not allowed on a destructor}}
};
struct B : A {}; // A & B would stay as "not complete definition" and "being defined".
This weird state was impending lookups in the records and hitting assertion in the ASTWriter.
Daniel Dunbar [Fri, 9 Mar 2012 19:35:29 +0000 (19:35 +0000)]
[AST/etc] Mark {getSourceRange(),getStartLoc(),getEndLoc()} as LLVM_READONLY.
- The theory here is that we have these functions sprinkled in all over the
place. This should allow the optimizer to at least realize it can still do
load CSE across these calls.
- I blindly marked all instances as such, even though the optimizer can infer
this attribute in some instances (some of the inline ones) as that was easier
and also, when given the choice between thinking and not thinking, I prefer
the latter.
You might think this is mere frivolity, but actually this is good for a .7 -
1.1% speedup on 403.gcc/combine.c, JSC/Interpreter.cpp,
OGF/NSBezierPath-OAExtensions.m.
lldb support: under debugger support flag, when sending message
to forward class, and assigning to an 'id' type var, message
sends default to 'id'. // rdar"//10988847
Daniel Dunbar [Fri, 9 Mar 2012 18:35:03 +0000 (18:35 +0000)]
[AST/Sema/libclang] Replace getSourceRange().getBegin() with getLocStart().
- getSourceRange().getBegin() is about as awesome a pattern as .copy().size().
I already killed the hot paths so this doesn't seem to impact performance on my
tests-of-the-day, but it is a much more sensible (and shorter) pattern.
Daniel Dunbar [Fri, 9 Mar 2012 15:39:19 +0000 (15:39 +0000)]
[AST] Reimplement Stmt::getLoc{Start,End} to dispatch to subclass overloads.
- getSourceRange() can be very expensive, we should try to avoid it if at all possible.
In conjunction with the previous commit I measured a ~2% speedup on 403.gcc/combine.c and a 3% speedup on OmniGroupFrameworks/NSBezierPath-OAExtensions.m.
Daniel Dunbar [Fri, 9 Mar 2012 15:39:15 +0000 (15:39 +0000)]
[AST] Add {DeclRefExpr,MemberExpr,ImplicitCastExpr}::{getLocStart,getLocEnd} methods.
- There are probably a lot more of these worth adding, but these are a start at hitting some of the exprs for which getSourceRange().getBegin() is a poor substitute for getLocStart().
Hans Wennborg [Fri, 9 Mar 2012 10:10:54 +0000 (10:10 +0000)]
-Wformat-non-iso: warn about positional arguments (pr12017)
This renames the -Wformat-non-standard flag to -Wformat-non-iso,
rewords the current warnings a bit (pointing out that a format string
is not supported by ISO C rather than being "non standard"),
and adds a warning about positional arguments.
Daniel Dunbar [Fri, 9 Mar 2012 04:12:54 +0000 (04:12 +0000)]
[AST] Memoize ASTContext::getTypeInfo().
- On -emit-llvm-only of 403.gcc/combine.c, for example, we make 160k calls to
getTypeInfo but only ever deal with 680 some distinct types.
I saw these speedups (user time):
403.gcc/combine.c -- 3.1%
OmniGroupFrameworks/NSBezierPath-OAExtensions.m -- 3.6%
JavaScriptCore/Interpreter.cpp -- 1.4%
which seems pretty sweet.
I ran some histograms on those compiles and we end up doing a ton of
getTypeInfo() on 'char' and 'int'. I tried splitting out a fast path for builtin
types, but this wasn't a win. Still kinda seems like we could be doing better
here.
Daniel Dunbar [Fri, 9 Mar 2012 01:51:51 +0000 (01:51 +0000)]
[AST] Reduce Decl::getASTContext() calls.
- This function is not at all free; pass it around along some hot paths instead
of recomputing it deep inside various VarDecl methods.
Ted Kremenek [Fri, 9 Mar 2012 01:13:14 +0000 (01:13 +0000)]
[analyzer] Implement basic path diagnostic pruning based on "interesting" symbols and regions.
Essentially, a bug centers around a story for various symbols and regions. We should only include
the path diagnostic events that relate to those symbols and regions.
The pruning is done by associating a set of interesting symbols and regions with a BugReporter, which
can be modified at BugReport creation or by BugReporterVisitors.
This patch reduces the diagnostics emitted in several of our test cases. I've vetted these as
having desired behavior. The only regression is a missing null check diagnostic for the return
value of realloc() in test/Analysis/malloc-plist.c. This will require some investigation to fix,
and I have added a FIXME to the test case.
Anna Zaks [Thu, 8 Mar 2012 23:16:38 +0000 (23:16 +0000)]
[analyzer] Use call graph to determine order in which functions are
analyzed.
The CallGraph is used when inlining is on, which is the current default.
This alone does not bring any performance improvement. It's a
stepping stone for the upcoming optimization in which we do not
re-analyze a function that has already been analyzed while inlined in
other functions. Using the call graph makes it easier to play with
the order of functions to minimize redundant analyzes.
Anna Zaks [Thu, 8 Mar 2012 23:16:35 +0000 (23:16 +0000)]
[analyzer] Rework inlining related command line options.
- Remove -analyzer-inline-call.
- Add -analyzer-ipa=[none|inlining]
- Add -analyzer-inlining-mode to allow experimentation for
different performance tuning methods.
Richard Smith [Thu, 8 Mar 2012 23:06:02 +0000 (23:06 +0000)]
Fix a couple of issues with literal-operator-id parsing, and provide recovery
for a few kinds of error. Specifically:
Since we're after translation phase 6, the "" token might be formed by multiple
source-level string literals. Checking the token width is not a correct way of
detecting empty string literals, due to escaped newlines. Diagnose and recover
from a missing space between "" and suffix, and from string literals other than
"", which are followed by a suffix.
Richard Smith [Thu, 8 Mar 2012 21:59:28 +0000 (21:59 +0000)]
When checking the encoding of an 8-bit string literal, don't just check the
first codepoint! Also, don't reject empty raw string literals for spurious
"encoding" issues. Also, don't rely on undefined behavior in ConvertUTF.c.
Sebastian Redl [Thu, 8 Mar 2012 21:05:45 +0000 (21:05 +0000)]
Turn explicit construction of temporaries using initializer list syntax into CXXTemporaryObjectExprs, not just CXXConstructExprs, which have a worrying tendency to vanish. Fixes PR12167.
Daniel Dunbar [Thu, 8 Mar 2012 20:28:55 +0000 (20:28 +0000)]
[AST] APValue: Split the fast path of MakeUninit to be inline.
- This change seems to be a tiny loss on 403.gcc/combine.c (.2%), but I think
it is the right thing to do.
[utils/FuzzTest] The help message says that if you don't specify --max-tests
"the script will run forever continually picking new tests to run", make it so.
Daniel Dunbar [Thu, 8 Mar 2012 18:43:59 +0000 (18:43 +0000)]
[Basic] PartialDiagnostic: Manual split out the cold part of freeStorage(), the
hot path will typically profitably get inlined (based on what I see from the
users of PartialDiagnostic).
Taken into account Duncan's comments for r149481 dated by 2nd Feb 2012:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120130/136146.html
Implemented CaseIterator and it solves almost all described issues: we don't need to mix operand/case/successor indexing anymore. Base iterator class is implemented as a template since it may be initialized either from "const SwitchInst*" or from "SwitchInst*".
ConstCaseIt is just a read-only iterator.
CaseIt is read-write iterator; it allows to change case successor and case value.
Usage of iterator allows totally remove resolveXXXX methods. All indexing convertions done automatically inside the iterator's getters.
Main way of iterator usage looks like this:
SwitchInst *SI = ... // intialize it somehow
for (SwitchInst::CaseIt i = SI->caseBegin(), e = SI->caseEnd(); i != e; ++i) {
BasicBlock *BB = i.getCaseSuccessor();
ConstantInt *V = i.getCaseValue();
// Do something.
}
If you want to convert case number to TerminatorInst successor index, just use getSuccessorIndex iterator's method.
If you want initialize iterator from TerminatorInst successor index, use CaseIt::fromSuccessorIndex(...) method.
There are also related changes in llvm-clients: klee and clang.
Richard Smith [Thu, 8 Mar 2012 02:39:21 +0000 (02:39 +0000)]
Implement C++11 [lex.ext]p10 for string and character literals: a ud-suffix not
starting with an underscore is ill-formed.
Since this rule rejects programs that were using <inttypes.h>'s macros, recover
from this error by treating the ud-suffix as a separate preprocessing-token,
with a DefaultError ExtWarn. The approach of treating such cases as two tokens
is under discussion for standardization, but is in any case a conforming
extension and allows existing codebases to keep building while the committee
makes up its mind.
Reword the warning on the definition of literal operators not starting with
underscores (which are, strangely, legal) to more explicitly state that such
operators can't be called by literals. Remove the special-case diagnostic for
hexfloats, since it was both triggering in the wrong cases and incorrect.
Douglas Gregor [Thu, 8 Mar 2012 01:00:17 +0000 (01:00 +0000)]
Streamline BalancedDelimiterTracker, by eliminating the duplicate
paren/brace/bracket tracking (the Consume* functions already did it),
removing the use of ConsumeAnyToken(), and moving the hot paths inline
with the error paths out-of-line.
Anna Zaks [Thu, 8 Mar 2012 00:42:23 +0000 (00:42 +0000)]
Add a basic CallGraph to Analysis.
The final graph contains a single root node, which is a parent of all externally available functions(and 'main'). As well as a list of Parentless/Unreachable functions, which are either truly unreachable or are unreachable due to our analyses imprecision.
The analyzer checkers debug.DumpCallGraph or debug.ViewGraph can be used to look at the produced graph.
Currently, the graph is not very precise, for example, it entirely skips edges resulted from ObjC method calls.
Basically the current design is:
-for an implementation method, show as overridden the interface method.
This is not useful, and is inconsistent with the C++ side
-for an interface method, show as overridden the protocols methods (this is desirable)
and the methods from the categories; methods from categories are not useful
since they are considered the same method (same USR).
-If there is a protocol method or category method reported, it does not check the
super class for overridden methods. This is really problematic since
overridden methods from super class is what we want to give back.
Change clang_getOverriddenCursors to show as overridden any method in the class's
base class, its protocols, or its categories' protocols, that has the same
selector and is of the same kind (class or instance).
If no such method exists, the search continues to the class's superclass,
its protocols, and its categories, and so on. A method from an Objective-C
implementation is considered to override the same methods as its
corresponding method in the interface.
Bob Wilson [Wed, 7 Mar 2012 23:50:05 +0000 (23:50 +0000)]
Workaround module test failures by removing the version info from module hashes.
PR12196: The module hash strings are not actually hashing the compiler version
string; the entire version string is being included in the hash. Depending on
the module cache directory name, that can lead to failures where the path
names become too long. As a temporary workaround, just remove the version
string from the hash.
Richard Smith [Wed, 7 Mar 2012 08:35:16 +0000 (08:35 +0000)]
AST representation for user-defined literals, plus just enough of semantic
analysis to make the AST representation testable. They are represented by a
new UserDefinedLiteral AST node, which is a sugared CallExpr. All semantic
properties, including full CodeGen support, are achieved for free by this
representation.
UserDefinedLiterals can never be dependent, so no custom instantiation
behavior is required. They are mangled as if they were direct calls to the
underlying literal operator. This matches g++'s apparent behavior (but not its
actual mangling, which is broken for literal-operator-ids).
User-defined *string* literals are now fully-operational, but the semantic
analysis is quite hacky and needs more work. No other forms of user-defined
literal are created yet, but the AST support for them is present.
This patch committed after midnight because we had already hit the quota for
new kinds of literal yesterday.
Richard Smith [Wed, 7 Mar 2012 03:13:00 +0000 (03:13 +0000)]
Add -Wc++11-compat warning for string and character literals followed by
identifiers, in cases where those identifiers would be treated as
user-defined literal suffixes in C++11.
Richard Smith [Wed, 7 Mar 2012 01:58:44 +0000 (01:58 +0000)]
Don't even try to directly emit the value of a DeclRefExpr if that declaration
is not usable in a constant expression. ~2.5% speedup on 403.gcc / combine.c.
[PCH] Mark a PCH file with a flag to indicate if the serialized AST had
compiler errors or not.
-Control whether ASTReader should reject such a PCH by a boolean flag at ASTReader's creation time.
By default, such a PCH file will be rejected with an error when trying to load it.
[libclang] Allow clang_saveTranslationUnit to create a PCH file even if compiler errors
occurred.
-Have libclang API calls accept a PCH that had compiler errors.
The general idea is that we want libclang to stay functional even if a PCH had a compiler error.
rdar://10976363.
Andrew Trick [Wed, 7 Mar 2012 00:44:24 +0000 (00:44 +0000)]
More git-svn compatible version string, by request.
If you're using git-svn, the clang and llvm repository will typically
map to a different revision.
Before we had:
clang version 3.1 (trunk 152167 trunk 152162)
After this change:
clang version 3.1 (trunk 152167) (llvm/trunk 152162)
So it's self-descriptive with an extra parens group. Which is more
compatible with version string parsers is probably debatable, but this
style was requested.
Michael Han [Wed, 7 Mar 2012 00:12:16 +0000 (00:12 +0000)]
Refactor Clang sema attribute handling.
This submission improves Clang sema handling by using Clang tablegen
to generate common boilerplate code. As a start, it implements AttributeList
enumerator generation and case statements for AttributeList::getKind.
A new field "SemaHandler" is introduced in Attr.td and by default set to 1
as most of attributes in Attr.td have semantic checking in Sema. For a small
number of attributes that don't appear in Sema, the value is set to 0.
Also there are a small number of attributes that only appear in Sema but not
in Attr.td. Currently these attributes are still hardcoded in Sema AttributeList.
projects / clang / log