- Sadly, this doesn't seem to give any .ll size win so far. It is possible to make this routine significantly smarter & avoid various shifting, masking, and zext/sext, but I'm not really convinced it is worth it. It is tricky, and this is really instcombine's job.
- No intended functionality change; the test case is just to increase coverage & serves as a demo file, it worked before this commit.
The new fixes from r101222 are:
1. The shift to the target position needs to occur after the value is extended to the correct size. This broke Clang bootstrap, among other things no doubt.
2. Swap the order of arguments to OR, to get a tad more constant folding.
llvm-svn: 101339
Stop multiplying constant by 8 accordingly in the header and change
intrinsic definition for what types we expect.
Add to existing palignr test to check that we're emitting the correct things.
llvm-svn: 101332
containing the time taken for the code completion to be sent to a designated server
(which is specified using a compile-time -D flag).
llvm-svn: 101326
that have reference or const scalar members, since those members can
never be initializer or modified. Fixes <rdar://problem/7804350>.
llvm-svn: 101316
ResolveAddressOfOverloadedFunction when asked to complain. Previously,
we had some weird handshake where ResolveAddressOfOverloadedFunction
expected its caller to handle some of the diagnostics but not others,
and yet there was no way for the caller to know which case we were
in. Eliminate this madness, fixing <rdar://problem/7765884>.
llvm-svn: 101312
separate count of "suppressed" errors. This way, semantic analysis
bits that depend on the error count to determine whether problems
occured (e.g., some template argument deduction failures, jump-scope
checking) will not get confused.
The actual problem here is that a missing #include (which is a fatal
error) could cause the jump-scope checker to run on invalid code,
which it is not prepared to do. Trivial fix for both
<rdar://problem/7775941> and <rdar://problem/7775709>.
llvm-svn: 101297
generally recover from typos in keywords (since we would effectively
have to mangle the token stream). However, there are still benefits to
typo-correcting with keywords:
- We don't make stupid suggestions when the user typed something
that is similar to a keyword.
- We can suggest the keyword in a diagnostic (did you mean
"static_cast"?), even if we can't recover and therefore don't have
a fix-it.
llvm-svn: 101274
than just a bool indicating that correction occurred. No actual
functionality change (it's still always used like a bool), but this
refactoring will be used to support typo correction to keywords.
llvm-svn: 101259
that adds parentheses from the main diagnostic down to a new
note. This way, when the fix-it represents a choice between two
options, each of the options is associted with a note. There is no
default option in such cases. For example:
/Users/dgregor/t.c:2:9: warning: & has lower precedence than ==; ==
will be
evaluated first [-Wparentheses]
if (x & y == 0) {
^~~~~~~~
/Users/dgregor/t.c:2:9: note: place parentheses around the &
expression to
evaluate it first
if (x & y == 0) {
^
( )
/Users/dgregor/t.c:2:9: note: place parentheses around the ==
expression to
silence this warning
if (x & y == 0) {
^
( )
llvm-svn: 101249
- Sadly, this doesn't seem to give any .ll size win so far. It is possible to make this routine significantly smarter & avoid various shifting, masking, and zext/sext, but I'm not really convinced it is worth it. It is tricky, and this is really instcombine's job.
- No intended functionality change; the test case is just to increase coverage & serves as a demo file, it worked before this commit.
llvm-svn: 101222
receiver is a mis-typed class name. Previously, we would give a non-specific
typo-correction diagnostic from the expression-parsing code, but there
was no fix-it because it was too late to recover. Now, we give a nice
diagnostic
honk.m:6:4: error: unknown receiver 'Hnk'; did you mean 'Honk'?
[Hnk method];
^~~
Honk
honk.m:1:1: note: 'Honk' declared here
@interface Honk
^
which includes a fix-it.
We still need to recover better from mis-typing "super".
llvm-svn: 101211
super message sends in Objective-C. No actual functionality change
here, but it provides a hook so that Sema can typo-correct the
receiver in some cases.
llvm-svn: 101207
- This lets the method focus slightly more on emitting clean IR to honor the policy which has been selected. On 403.gcc's combine.c, x86_64, -O0, this reduces the number of lines in the .ll file (~= # of instructions) by 2.5%.
- No intended functionality change -- at -O3 this should produce equivalent if not identical output. On 403.gcc's combine.c, x86_64, -O3, this isn't quite true and some of the changes are regressions, but I'm not going to worry about that until we move to a new access policy.
- There is still some room for improvement in the generated IR, in particular we can usually fold the sign-extension of the bit-field into one of the component access. See the FIXME.
llvm-svn: 101192
- For now, these policies are computed to match the current IRgen strategy, although the new information isn't being used yet (except in -fdump-record-layouts).
- Design comments appreciated.
llvm-svn: 101178
elements with explicit zero values instead of with tail padding.
On an example like this:
struct foo { int a; int b; };
struct foo fooarray[] = {
{1, 2},
{4},
};
We now lay this out as:
@fooarray = global [2 x %struct.foo] [%struct.foo { i32 1, i32 2 }, %struct.foo { i32 4, i32 0 }]
instead of as:
@fooarray = global %0 <{ %struct.foo { i32 1, i32 2 }, %1 { i32 4, [4 x i8] zeroinitializer } }>
Preserving both the struct type of the second element, but also the array type of the entire thing.
llvm-svn: 101155
Without it, there is no reason for a compiler that supports it to
emit the dead static globals that the rewriter labels attribute(used).
llvm-svn: 101149
for reference binding. The code attempted to handle both the
computation of the ICS and the actual conversion, but the latter is an
anachronism: we now use InitializationSequence for that.
Sema::CheckReferenceInit is now a static function TryReferenceInit
that's only use within overload resolution, and has been simplified
slightly. It still needs to be updated per C++ [over.ics.ref], by
eliminating more of the lvalue/rvalue checks.
llvm-svn: 101136
from a conversion function template specialization to one of exact
match rank. We only know how to test this in C++0x with default
function template arguments, but it's also in the C++03 spec. Fixes
PR6285.
llvm-svn: 101089
copying the type location information from the conversion-type-id into
the type location information for the function type. Do something
similar for constructors and destructors, by giving their "void"
return type source-location information.
In all of these cases, we previously left this type-source information
uninitialized, which led to various unfortunate crashes.
We still aren't tracking good source-location information for the
actual names. That's PR6357.
John, please check my sanity on this.
llvm-svn: 101088
in an input file like this:
# 42
int x;
we were emitting:
# <something>
int x;
(with a space before the int) because we weren't clearing the leading
whitespace flag properly after the \n from the directive was handled.
llvm-svn: 101084
actually turned it on. If a diag is produced by a warning which
is an extension but defaults to on, and has no warning group, don't
print any option info.
llvm-svn: 101071
trailing fields may not be represented in initializer lists, they
are being handled as padding and those fields *must* be zero
initialized.
llvm-svn: 101067
name-lookup ambiguities when there are multiple base classes that are
all specializations of the same class template. This is part of a
general cleanup for ambiguities in template-name lookup. Fixes
PR6717.
llvm-svn: 101065
sure to introduce them into the current Scope (when we have one) in
addition to the DeclContext for the class, so that they can be found
by name lookup for inline members of the class. Fixes PR6570.
llvm-svn: 101047
correctly diagnose instantiation of a function parameter with Objective-C
class type (since Objective-C classes can't be passed by value).
llvm-svn: 101031
function's type is (strictly speaking) non-dependent. This ensures
that, e.g., default function arguments get instantiated properly.
And, since I couldn't resist, collapse the two implementations of
function-parameter instantiation into calls to a single, new function
(Sema::SubstParmVarDecl), since the two had nearly identical code (and
each had bugs the other didn't!). More importantly, factored out the
semantic analysis of a parameter declaration into
Sema::CheckParameter, which is called both by
Sema::ActOnParamDeclarator (when parameters are parsed) and when a
parameter is instantiated. Previously, we were missing some
Objective-C and address-space checks on instantiated function
parameters.
Fixes PR6733.
llvm-svn: 101029
LookupInObjCMethod. Doing so allows all sorts of invalid code
to slip through to codegen. This patch does not change the
AST representation of super, though that would now be a natural
thing to do since it can only be in the receiver position and
in the base of a ObjCPropertyRefExpr.
There are still several ugly areas handling super in the parser,
but this is definitely a step in the right direction.
llvm-svn: 100959
typo correction. However, now that the code has been factored out
of LookupMemberExpr, it can recurse to itself instead of to
LookupMemberExpr! Remove grossness.
llvm-svn: 100958
pointer to an objc interface out to a method in SemaExprObjC.
This is *much* uglier than it should be due to grossness in
LookupMemberExpr :(
llvm-svn: 100957
means it isn't really possible to write the test case for this code, but this is
the kind of thing that really requires testing against the installed compiler
anyway.
llvm-svn: 100935
member. Use a better diagnostic for this case. Also fix a bug with nested
anonymous structs/unions for -Wreorder; this last was PR6575.
llvm-svn: 100923
when they're instantiated. Merge the note into the -Wreorder warning; it
doesn't really contribute much, and it was splitting a thought across diagnostics
anyway. Don't crash in the parser when a constructor's initializers end in a
comma and there's no body; the recovery here is still terrible, but anything's
better than a crash.
llvm-svn: 100922
Clang is installed. This is designed to match gcc, and is important when
installed in a non-standard location.
- This is gross, but no worse than ever. It will die when we finally move to
the compiler-rt based toolchain, any day now.
llvm-svn: 100915
specializations when the explicit instantiation was... explicitly
written, i.e., not the product of an explicit instantiation of an
enclosing class. Fixes this spurious warning when Clang builds LLVM:
/Volumes/Data/dgregor/Projects/llvm/lib/CodeGen/MachineDominators.cpp:22:1:
warning: explicit instantiation of 'addRoot' that occurs after an
explicit specialization will be ignored (C++0x extension) [-pedantic]
llvm-svn: 100900
we don't have enough information to tell them how to use 'strncmp'. Instead, change the
diagnostic to indicate they should use 'strncmp'.
llvm-svn: 100890
Remove -faccess-control from -cc1; add -fno-access-control.
Make the driver pass -fno-access-control by default.
Update a bunch of tests to be correct under access control.
llvm-svn: 100880
the implicit template instantiations we need to perform. Otherwise, we
end up erroneously diagnosing static functions as used if they were
only used within an implicit template instantiation. Fixes a bunch of
spurious failures when building Clang with Clang.
llvm-svn: 100872
precompiled headers and/or when reading the contents of the file into
memory. These checks seem to be causing spurious regression-test
failures on Windows.
llvm-svn: 100866
destination type for initialization, assignment, parameter-passing,
etc. The main issue fixed here is that we used rather confusing
wording for diagnostics such as
t.c:2:9: warning: initializing 'char const [2]' discards qualifiers,
expected 'char *' [-pedantic]
char *name = __func__;
^ ~~~~~~~~
We're not initializing a 'char const [2]', we're initializing a 'char
*' with an expression of type 'char const [2]'. Similar problems
existed for other diagnostics in this area, so I've normalized them all
with more precise descriptive text to say what we're
initializing/converting/assigning/etc. from and to. The warning for
the code above is now:
t.c:2:9: warning: initializing 'char *' from an expression of type
'char const [2]' discards qualifiers [-pedantic]
char *name = __func__;
^ ~~~~~~~~
Fixes <rdar://problem/7447179>.
llvm-svn: 100832
unless they are used. I discussed this with Daniel Dunbar, and we agreed that this
provides an inconsistent warnings experience for the user and that there were
genuine cases where we wouldn't want to do this optimization.
llvm-svn: 100800
<tr1/hashtable> header, where a friend class template
std::tr1::__detail::_Map_base is declared with the wrong template
parameters. GCC doesn't catch the problem, so Clang does a little
back-flip to avoid diagnosing just this one instance of the problem.
llvm-svn: 100790
__cxxabiv1::__fundamental_type_info in every translation
unit. Previously, we would perform name lookup for
__cxxabiv1::__fundamental_type_info at the end of IRGen for a each
translation unit, to determine whether it was present. If so, we we
produce type information for all of the fundamental types. However,
this name lookup causes PCH deserialization of a significant part of the
translation unit, which has a woeful impact on performance.
With this change, we now look at each record type after we've
generated its vtable to see if it is
__cxxabiv1::__fundamental_type_info. If so, we generate type info for
all of the fundamental types. This works because
__cxxabiv1::__fundamental_type_info should always have a key function
(typically the virtual destructor), that will be defined once in the
support library. The fundamental type information will end up there.
Fixes <rdar://problem/7840011>.
llvm-svn: 100772
parameter, explicitly ask the user to give it arguments. We used to
complain that it wasn't a type and expect the user to figure it out.
llvm-svn: 100729
emitting diagnostics after it has produced that many errors. Give this a
default value of 20 which produces plenty of errors for people to fix before
recompiling but not so many that their entire console scrolls away when the
compiler gets confused. The experience looks like this:
$ clang foo.c
<tons of crap>
foo.c:102:3: error: unknown type name 'somethingbad'
somethingbad x;
^
fatal error: too many errors emitted, stopping now
36 warnings and 20 errors generated.
llvm-svn: 100689
isn't any extra work to perform. Also, don't check for unused
parameters when the warnings will be suppressed anyway. Improves
performance of -fsyntax-only on 403.gcc's combine.c by ~2.5%.
<rdar://problem/7836787>
llvm-svn: 100686
of errors and warnings. This allows us to emit something like this:
2 warnings and 1 error generated.
instead of:
3 diagnostics generated.
This also stops counting 'notes' because they are just follow-on information
about the previous diag, not a diagnostic in themselves.
llvm-svn: 100675
- When instantiating a friend type template, perform semantic
analysis on the resulting type.
- Downgrade the errors concerning friend type declarations that do
not refer to classes to ExtWarns in C++98/03. C++0x allows
practically any type to be befriended, and ignores the friend
declaration if the type is not a class.
llvm-svn: 100635
semantic analysis) and Sema::ActOnFriendTypeDecl (the action
callback). This is a prerequisite for improving template instantiation
of friend type declarations.
llvm-svn: 100633
have the code generate slap a srcloc metadata on inline asm nodes.
This allows us to diagnose invalid inline asms with such nice
diagnostics as:
<inline asm>:1:2: error: unrecognized instruction
abc incl %eax
^
asm.c:2:12: note: generated from here
__asm__ ("abc incl %0" : "+r" (X));
^
2 diagnostics generated.
llvm-svn: 100608
definitions, e.g., after
-
or
- (id)
we'll find all of the "likely" instance methods that one would want to
declare or define at this point. In the latter case, we only produce
results whose return types match "id".
llvm-svn: 100587
them the same way as fields. This fixes a regression in RegionStore::RemoveDeadbindings()
that emerged from going to the cluster-based analysis.
llvm-svn: 100570
that protected members be used on objects of types which derive from the
naming class of the lookup. My first N attempts at this were poorly-founded,
largely because the standard is very badly worded here.
llvm-svn: 100562
while we're completing in the middle of a function call), also produce
"ordinary" name results that show what can be typed at that point.
llvm-svn: 100558
presence of precompiled headers by forcibly loading all of the
methods we know about from the PCH file before constructing our
code-completion list.
llvm-svn: 100535
deciding when we need to emit an extra "command failed" diagnostic.
- This also fixes the case where we were emitting that extra diagnostics, even
when using clang w/ the integrated assembler, which has good diagnostics.
llvm-svn: 100529
"id" or an expression of type "id". In these cases, we produce a list
of all of the (class or instance) methods, respectively, that we know about.
Note that this implementation does not yet work well with precompiled
headers; that's coming soon.
llvm-svn: 100528
- Unfortunately, this requires some horrible code in CGObjCMac which always
allocats a CGBitFieldInfo because we don't currently build a proper layout
for Objective-C classes. It needs to be cleaned up, but I don't want the
bit-field cleanups to be blocked on that.
llvm-svn: 100474
ASTUnit. Previously, we would end up with use-after-free errors
because the Diagnostic object would be creating in one place (say,
CIndex) and its ownership would not be transferred into the
ASTUnit. Fixes <rdar://problem/7818608>.
llvm-svn: 100464
- Rename "Diagnostics" and related to "StoredDiagnostics", to better
capture what we're actually storing.
- Move SourceManager and FileManager to the heap.
llvm-svn: 100441
assembly for the global "d7". We were previously testing for alignment
3, which seems to happen for some builders and not for others. I've
eliminated the alignment check and added a FIXME to unbreak the
buildbots.
llvm-svn: 100205
poor (and wrong) approximation of the actual rules governing when to
build a copy and when it can be elided.
The correct implementation is actually simpler than the
approximation. When we only enumerate constructors as part of
initialization (e.g., for direct initialization or when we're copying
from a class type or one of its derived classes), we don't create a
copy. When we enumerate all conversion functions, we do create a
copy. Before, we created some extra copies and missed some
others. The new test copy-initialization.cpp shows a case where we
missed creating a (required, non-elidable) copy as part of a
user-defined conversion, which resulted in a miscompile. This commit
also fixes PR6757, where the missing copy made us reject well-formed
code in the ternary operator.
This commit also cleans up our handling of copy elision in the case
where we create an extra copy of a temporary object, which became
necessary now that we produce the right copies. The code that seeks to
find the temporary object being copied has moved into
Expr::getTemporaryObject(); it used to have two different
not-quite-the-same implementations, one in Sema and one in CodeGen.
Note that we still do not attempt to perform the named return value
optimization, so we miss copy elisions for return values and throw
expressions.
llvm-svn: 100196
PartialDiagnostic objects, and hoisting the stream operator for QualType into
Type.h with the operator for the normal DiagnosticBuilder. Also remove the no
longer needed friend declaration for the DeclarationName stream operator.
llvm-svn: 100169
an object or function. Our previous checking was too lax, and ended up
allowing missing or extraneous address-of operators, among other
evils. The new checking provides better diagnostics and adheres more
closely to the standard.
Fixes PR6563 and PR6749.
llvm-svn: 100125
nested-name-specifier (e.g., "class T::foo") fails to find a tag
member in the scope nominated by the
nested-name-specifier. Previously, we gave a bland
error: 'Nested' does not name a tag member in the specified scope
which didn't actually say where we were looking, which was rather
horrible when the nested-name-specifier was instantiated. Now, we give
something a bit better:
error: no class named 'Nested' in 'NoDepBase<T>'
llvm-svn: 100060
(such as "class T::foo") from an ElaboratedType of a TypenameType to a
DependentNameType, which more accurately models the underlying
concept.
Improve template instantiation for DependentNameType nodes that
represent nested-name-specifiers, by performing tag name lookup and
checking the resulting tag appropriately. Fixes PR5681.
There is still much testing and cleanup to do in this area.
llvm-svn: 100054
This more likely matches with the ideal order the blocks should be visited.
This shaves another 1% off the -fsyntax-only time for compare.c (403.gcc).
llvm-svn: 100030
and a DenseSet for caching instead of using a single SmallPtrSet.
This makes the behavior of the DataFlowSolver more deterministic, and
reduces the -fsyntax-only time on compare.c (403.gcc) by 1%.
llvm-svn: 100026
the existing (and already well-tested) linkage computation for types,
with minor tweaks for dynamic classes and (pointers to) incomplete
types. Fixes PR6597.
llvm-svn: 99968
that will interfere (they will be parsed as if they are after the class' '}') and a crash will occur because
the CachedTokens that holds them will be deleted while the lexer is still using them.
Make sure that the tokens of default args are removed from the token stream.
Fixes PR6647.
llvm-svn: 99939
name in the outermost block of a if/else that declares the same name
is ill-formed. Turns out that Clang and MSVC were right about PR6739;
GCC is too lax.
llvm-svn: 99937
the underlying/instantiated decl) through a lot of API, including "intermediate"
MemberExprs required for (e.g.) template instantiation. This is necessary
because of the access semantics of member accesses to using declarations:
only the base class *containing the using decl* need be accessible from the
naming class.
This allows us to complete an access-controlled selfhost, if there are no
recent regressions.
llvm-svn: 99936
involving substitution of deduced template arguments into a class
template partial specialization or function template, or when
substituting explicitly-specific template arguments into a function
template. We now print the actual deduced argument bindings so the
user can see what got deduced.
llvm-svn: 99923
This introduces FunctionType::ExtInfo to hold the calling convention and the
noreturn attribute. The next patch will extend it to include the regparm
attribute and fix the bug.
llvm-svn: 99920
dyn_cast) invocations for C++ and Objective-C types, declarations,
expressions, and statements. The statistics will be printed when
-print-stats is provided to Clang -cc1, with results such as:
277073 clang - Number of checks for C++ declaration nodes
13311 clang - Number of checks for C++ expression nodes
18 clang - Number of checks for C++ statement nodes
174182 clang - Number of checks for C++ type nodes
92300 clang - Number of checks for Objective-C declaration nodes
9800 clang - Number of checks for Objective-C expression nodes
7 clang - Number of checks for Objective-C statement nodes
65733 clang - Number of checks for Objective-C type nodes
The statistics are only gathered when NDEBUG is not defined, since
they introduce potentially-expensive operations into very low-level
routines (isa).
llvm-svn: 99912
storage for that partial diagnostic via the ASTContext's
BumpPtrAllocator rather than using up slots in the ASTContext's
cache. Now that we do this, we don't have to worry about destroying
dependent diagnostics when destroying a DependentStoredDeclsMap.
llvm-svn: 99854
cache of PartialDiagnostic::Storage objects into an allocator within
the ASTContext. This eliminates a significant amount of malloc
traffic, for a 10% performance improvement in -fsyntax-only wall-clock
time with 403.gcc's combine.c.
Also, eliminate the RequireNonAbstractType hack I put in earlier,
which was but a symptom of this larger problem.
Fixes <rdar://problem/7806091>.
llvm-svn: 99849
check deduced non-type template arguments and template template
arguments against the template parameters for which they were deduced,
performing conversions as appropriate so that deduced template
arguments get the same treatment as explicitly-specified template
arguments. This is the bulk of PR6723.
Also keep track of whether deduction of a non-type template argument
came from an array bound (vs. anywhere else). With this information,
we enforce C++ [temp.deduct.type]p17, which requires exact type
matches when deduction deduces a non-type template argument from
something that is not an array bound.
Finally, when in a SFINAE context, translate the "zero sized
arrays are an extension" extension diagnostic into a hard error (for
better standard conformance), which was a minor part of PR6723.
llvm-svn: 99734
nested within, and suddenly local classes start working. Wouldn't be
necessary if I hadn't used local classes in Clang in the first place.
Or, well, wouldn't be necessary yet. :)
llvm-svn: 99709
through the bases and fields of the definition of the class in which
the constructor is declared, rather than some other declaration of
that class.
llvm-svn: 99661
What happens here is that we actually turn the first declaration into a
definition, regardless of whether it was actually originally a definition,
and furthermore we do this all after we've instantiated all the declarations.
This exposes a bug in my DefinitionData patch where it was only setting the
DefinitionData for previous declarations, not future declarations.
Fortunately, there's an iterator for that.
llvm-svn: 99657
expression can be converted to the type of another, only apply the
lvalue-to-rvalue conversion to the type of the expression we're
converting, *not* the array-to-pointer or function-to-pointer
conversions. Fixes PR6595.
llvm-svn: 99652
parameter, keep the integral value exactly as it was in the source
code rather than trying to convert it to the type of the non-type
template parameter (which may still be dependent!). The value will
then be converted to the appropriate type once we check the resulting
template arguments. Fixes PR6707.
llvm-svn: 99611
that we extend/truncate then correct the sign to convert the non-type
template argument to the template parameter's type. Previously, we
reported an error when the non-type template argument was out of
range; now we just warn.
llvm-svn: 99600
the type of its corresponding non-type template parameter changes the
value. Previously, we were diagnosing this as an error, which was
wrong. We give reasonably nice warnings like:
test/SemaTemplate/temp_arg_nontype.cpp💯10: warning: non-type template
argument value '256' truncated to '0' for template parameter of type
'unsigned char'
Overflow<256> *overflow3; // expected-warning{{non-type template ...
^~~
test/SemaTemplate/temp_arg_nontype.cpp:96:24: note: template parameter is
declared here
template<unsigned char C> struct Overflow;
^
llvm-svn: 99561
how to handle a diagnostic during template argument deduction, which
may be "substitution failure", "suppress", or "report". This keeps us
from, e.g., emitting warnings while performing template argument
deduction.
llvm-svn: 99560
friendship for a derived class if the base class specifier was non-public,
and thus not considering friendship for non-public members of public bases.
llvm-svn: 99554
gcc, and the common expectation seems to be that they are unused. If and when
someone cares we can add them back with well documented demantics.
llvm-svn: 99522
Daniel Dunbar