offsetof correctly in the presence of anonymous structs/unions.
This could definitely use some cleanup, but I don't really want to mess
with the anonymous union/struct code.
llvm-svn: 70156
Before we emitted:
$ clang t.c -S -m64
llvm: error: Unsupported asm: input constraint with a matching output constraint of incompatible type!
Now we produce:
$ clang t.c -S -m64
t.c:5:40: error: unsupported inline asm: input with type 'unsigned long' matching output with type 'int'
asm volatile("foo " : "=a" (a) :"0" (b));
~~~ ~^~
llvm-svn: 70142
necessary and iterate until all types and declarations have been
written. This reduces the Cocoa.h PCH file size by about 4% (since we
don't write types we don't need), and fixes problems where writing a
declaration generates a new type.
This doesn't seem to have any impact on performance either way.
llvm-svn: 70109
- This can be used to supply a default value for -std=; the idea is
that this can be used in conjunction with CCC_ADD_ARGS or
QA_OVERRIDE_GCC3_OPTIONS to change the default without having to
modify the build system.
llvm-svn: 70102
most of which are ignored. Instead, move the __COUNTER__ value out to
a PCH-level record (since it is handled eagerly) and move the header
file information into the SourceManager block (which is also,
currently, loaded eagerly).
This results in another 17% performance improvement in the
Cocoa-prefixed "Hello, World" with PCH.
llvm-svn: 70097
PCH file. In the Cocoa-prefixed "Hello, World" benchmark, this takes
us from reading 503 identifiers down to 37 and from 470 macros down to
4. It also results in an 8% performance improvement.
llvm-svn: 70094
Overall, I'm not particularly happy with the current situation regarding
constant expression diagnostics, but I plan to improve it at some point.
llvm-svn: 70089
VerifyIntegerConstantExpression instead of isIntegerConstantExpr.
This makes it ext-warn but tolerate things that fold to a constant
but that are not valid i-c-e's.
There must be a bug in the i-c-e computation though, because it
doesn't catch this case even with pedantic.
This also switches the later code to use EvaluateAsInt which is
simpler and handles everything that evaluate does.
llvm-svn: 70081
identifier's ID. In this case, we know where the identifier's entry is
located in the hash table (it starts right before the identifier
string itself), so skip the hash table lookup and read the entry
directly. The performance improvement here is, gain, hard to quantify,
but it's the right thing to do.
llvm-svn: 70078
"interesting" identifiers (e.g., those where the IdentifierInfo has
some useful information) from "uninteresting" identifiers (where the
IdentifierInfo is just a name). This makes the hash table smaller (so
searching in it should be faster) and, when loading "uninteresting"
identifiers, we skip the lookup in the hash table.
PCH file size is slightly smaller than before (since we don't emit the
contents of the uninteresting IdentifierInfo structures). The
Cocoa.h-prefixed "Hello, World" doesn't show any speedup, although
we're getting to the point where system noise is a bit issue.
llvm-svn: 70075
allocating IdentifierInfos with a pointer into the string data stored
in the PCH file rather than having an entry in the identifier table's
string map. However, we don't actually get these savings at the
moment, because we go through the IdentifierTable when loading
identifiers from the on-disk hash table.
This commit is for record-keeping purposes only. I'll be reverting
this change (and the PCH layout tweak that preceded it) because it
appears that implementing this optimization will collide with another,
future optimization to reduce the size of the on-disk hash table for
identifiers. That optimization is likely to provide more benefit (with
less voodoo).
llvm-svn: 70070
that the PCH reader does not have to decode the VBR encoding at PCH
load time.
Also, reduce the size of the identifier offsets from 64 bits down to
32 bits. The identifier table itself isn't going to grow to more than
4GB :)
Overall, this results in a 13% speedup in the Cocoa-prefixed "Hello,
World" benchmark.
llvm-svn: 70063
blobs, so that we don't need to do any work to get these arrays into
memory at PCH load time.
This gives another 19% performance improvement to the Cocoa-prefixed
"Hello, World!".
llvm-svn: 70059
This results in a 10% speedup on the Cocoa-prefixed "Hello, World!",
all of which is (not surprisingly) user time. There was a tiny
reduction in the size of the PCH file for Cocoa.h, because certain
selectors aren't being written twice.
I'm using two new tricks here that I'd like to replicate elsewhere:
(1) The selectors not used in the global method pool are packed into
the blob after the global method pool's on-disk hash table and
stored as keys, so that all selectors are in the same blob.
(2) We record the offsets of each selector key when we write it into
the global method pool (or after it, in the same blob). The offset
table is written as a blob, so that we don't need to pack/unpack a
SmallVector with its contents.
llvm-svn: 70055
always return a non-null QualType + error bit. This fixes a bunch of
cases that didn't check for null result (and could thus crash) and eliminates
some crappy code scattered throughout sema.
This also improves the diagnostics in the recursive struct case to eliminate
a bogus second error. It also cleans up the case added to function.c by forming
a proper function type even though the declarator is erroneous, allowing the
parameter to be added to the function. Before:
t.c:2:1: error: unknown type name 'unknown_type'
unknown_type f(void*P)
^
t.c:4:3: error: use of undeclared identifier 'P'
P+1;
^
After:
t.c:2:1: error: unknown type name 'unknown_type'
unknown_type f(void*P)
^
llvm-svn: 70023
This gets rid of a bunch of random InvalidDecl bools in sema, changing
us to use the following approach:
1. When analyzing a declspec or declarator, if an error is found, we
set a bit in Declarator saying that it is invalid.
2. Once the Decl is created by sema, we immediately set the isInvalid
bit on it from what is in the declarator. From this point on, sema
consistently looks at and sets the bit on the decl.
This gives a very clear separation of concerns and simplifies a bunch
of code. In addition to this, this patch makes these changes:
1. it renames DeclSpec::getInvalidType() -> isInvalidType().
2. various "merge" functions no longer return bools: they just set the
invalid bit on the dest decl if invalid.
3. The ActOnTypedefDeclarator/ActOnFunctionDeclarator/ActOnVariableDeclarator
methods now set invalid on the decl returned instead of returning an
invalid bit byref.
4. In SemaType, refering to a typedef that was invalid now propagates the
bit into the resultant type. Stuff declared with the invalid typedef
will now be marked invalid.
5. Various methods like CheckVariableDeclaration now return void and set the
invalid bit on the decl they check.
There are a few minor changes to tests with this, but the only major bad
result is test/SemaCXX/constructor-recovery.cpp. I'll take a look at this
next.
llvm-svn: 70020
parameters in a functiondecl, even if the decl is invalid and has a confusing
Declarator. On the testcase, we now emit one beautiful diagnostic:
t.c:2:1: error: unknown type name 'unknown_type'
unknown_type f(void*)
^
GCC 4.0 produces:
t.c:2: error: syntax error before ‘f’
t.c: In function ‘f’:
t.c:2: error: parameter name omitted
and GCC 4.2:
t.c:2: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘f’
llvm-svn: 70016
remove a special case that was apparently for typeof() and
generalize the code in SemaDecl that handles typedefs to
handle any sugar type (including typedef, typeof, etc).
Improve comment to make it more clear what is going on.
llvm-svn: 70015
typedef void foo(void);
We get a typedef for a functiontypeproto with no arguments, not
one with one argument and type void. This means the code being
removed in SemaDecl is dead.
llvm-svn: 70013
types.
- I broke this in the switch to representing interfaces with opaque
types.
- <rdar://problem/6822660> clang crashes on subscript of interface in
32-bit mode
llvm-svn: 70009
to the checker yet, but essentially it allows a user to specify that an
Objective-C method or C function increments the reference count of a passed
object.
llvm-svn: 70005
up to the checker yet, but essentially it allows a user to specify that an
Objective-C method or C function returns an owned an Objective-C object.
llvm-svn: 70001
by correctly propagating the fact that the type was invalid up to the
attributeRuns decl, then returning an ExprError when attributeRuns is
formed (like we do for normal declrefexprs).
llvm-svn: 69998
(1) Make sure to pad on-disk hash tables with 4 bytes, not 2, since
the reader assumes that bucket data is aligned on 4-byte
boundaries.
(2) Don't emit the number of factory methods twice. This was
throwing off the data counts and therefore causing lookups to
fail. I've added asserts so that this class of error cannot happen
again.
llvm-svn: 69991
pools, combined). The methods in the global method pool are lazily
loaded from an on-disk hash table when Sema looks into its version of
the hash tables.
llvm-svn: 69989
- Fix summary lookup for class methods to now use the (optional)
ObjCInterfaceDecl associated with a message expression. This removes a
long-standing FIXME.
- Partial fix for <rdar://problem/6062730> by stop tracking objects that
are passed to [NSObject performSelector]. These methods are often used
for delegates, which the analyzer doesn't reason about well yet.
llvm-svn: 69982
- Otherwise, we will end up with stray .dSYM files which don't get
lipo'ed or removed.
- Ideally we would run dsymutil on the result, but we don't have the
infrastructure for that yet. Note that gcc doesn't handle this case
either.
- <rdar://problem/6809621> [driver] clang leaves .dSYM files lying
around in tmp.
llvm-svn: 69951
Several changes here:
1. We change Type::isIncompleteType to realize that forward declared
interfaces are incomplete. This eliminate special case code for this
from the sizeof path, and starts us rejecting P[4] when P is a pointer
to an incomplete interface.
2. Explicitly reject P[4] when P points to an interface in non-fragile ABI
mode.
3. Switch the sizeof(interface) diagnostic back to an error instead of a
warning in non-fragile abi mode.
llvm-svn: 69943
As part of this, make ObjCImplDecl inherit from NamedDecl (since
ObjCImplementationDecls now need to have names so that they can be
found). This brings ObjCImplDecl very, very close to
ObjCContainerDecl; we may be able to merge them soon.
llvm-svn: 69941
their own namespace (IDNS_Protocol) and use the normal name-lookup
routines to find them. Aside from the simplification this provides
(one less DenseMap!), it means that protocols will be lazily
deserialized from PCH files.
Make the code size of the selector table block match the code size of
the type and decl blocks.
llvm-svn: 69939
SEL, Class, Protocol, CFConstantString, and
__objcFastEnumerationState. With this, we can now run the Objective-C
methods and properties PCH tests.
llvm-svn: 69932
extend the number of objects tracked by the retain/release checker by assuming
that all class and instance methods should follow Cocoa object "getter" and
"alloc/new" conventions.
llvm-svn: 69908
This enables class recognition to work with PCH. I believe this means we can remove Sema::ObjCInterfaceDecls and it's usage within Sema::LookupName(). Will investigate.
llvm-svn: 69891
the type assigned by sema (and is visible with sizeof(__func__) for
example) has nothing to do with what codegen ends up producing.
We should eventually add a method on PredefinedExpr to handle this.
In the meantime, just set up some framework and add some fixme's.
llvm-svn: 69872
methods, class methods, and property implementations) and instead
place all of these entities into the DeclContext.
This eliminates more linear walks when looking for class or instance
methods and should make PCH (de-)serialization of ObjCDecls trivial
(and lazy).
llvm-svn: 69849
in a bunch of declarations from the PCH file. We're down to loading
very few declarations in Carbon-prefixed "Hello, World!":
*** PCH Statistics:
6/20693 types read (0.028995%)
7/59230 declarations read (0.011818%)
50/44914 identifiers read (0.111324%)
0/32954 statements read (0.000000%)
5/6187 macros read (0.080815%)
llvm-svn: 69825
file needs to store. CodeGen needs to see these definitions (via
HandleTopLevelDecl), otherwise it won't be able to generate code for
them.
This patch notifies the consumer (e.g., CodeGen) about function
definitions and variable definitions when the corresponding
declarations are deserialized. Hence, we don't eagerly deserialize the
declarations for every variable or function that has a definition in
the PCH file. This gives another 5% speedup for the Carbon-prefixed
"Hello, World!", and brings our PCH statistics down to something far
more reasonable:
*** PCH Statistics:
13/20693 types read (0.062823%)
17/59230 declarations read (0.028702%)
54/44914 identifiers read (0.120230%)
0/32954 statements read (0.000000%)
5/6187 macros read (0.080815%)
llvm-svn: 69820
PCH files now contain complete information about builtins, including
any declarations that have been synthesized as part of building the
PCH file. When using a PCH file, we do not initialize builtins at all;
when needed, they'll be found in the PCH file.
This optimization translations into a 9% speedup for "Hello, World!"
with Carbon.h as a prefix header and roughly a 5% speedup for 403.gcc
with its prefix header. We're also reading less of the PCH file for
"Hello, World!":
*** PCH Statistics:
286/20693 types read (1.382110%)
1630/59230 declarations read (2.751984%)
764/44914 identifiers read (1.701029%)
1/32954 statements read (0.003035%)
5/6187 macros read (0.080815%)
down from
*** PCH Statistics:
411/20693 types read (1.986179%)
2553/59230 declarations read (4.310316%)
1093/44646 identifiers read (2.448148%)
1/32954 statements read (0.003035%)
21/6187 macros read (0.339421%)
llvm-svn: 69815
- Remove stale assertion that was breaking the test suite.
- When popping location contexts, only add a control-flow piece for fileID
locations.
llvm-svn: 69814
character instead of the entire range for the IfStmt, ForStmt, etc. We may
gradually refine these ranges later, but basically terminator ranges just refer
to the first keyword.
llvm-svn: 69812
Rework the shadow struct that is layed out for Objective-C classes.
- Superclasses are now always laid out in their shadow structure at
the first field.
- Prior to this, the entire class heirarchy was flattened into a
single structure which meant that alignment, padding, and bitfields
were incorrect (the ASTRecordLayout was correct however, which
meant our debug info didn't coincide with ivar offsets, for
example).
- This is still very suboptimal (for example, ivar are looked up
recursively, but I believe the ivar layout itself is now at least
close to correct.
- <rdar://problem/6773388> error: objc[29823]: layout bitmap sliding
backwards
llvm-svn: 69811
- This has pros and cons, but for now the pros seem to significantly
outway the con.
The con is that we will always need to cast in the runtime
implementation to a struct type, if we wish to access an interface
directly.
The pros are:
- Avoid the cost of generating types which are used. Most
manipulation of Objective-C objects is done through messages, and
only the implementation of a class will directly access
memory. Previously, we would convert the type even if it only
appear as a function parameter, for example.
- We don't need to worry about incomplete types, and
UpdateCompletedType for interfaces is gone.
- It becomes easier to narrow the interface to the shadow struct for
Objective-C interfaces (so it can be eliminated).
Currently the runtimes still use the CodeGenTypes machinery to
generate the LLVM structure they need via ConvertTagDecl, but this can
eventually be replaced.
llvm-svn: 69797
- If we don't find a file looking relative to the current working
directory, fall back to header search. This is closer to what would
happen if the lookup was starting from right directory in the first
place (except it will find files in the directory of the main
source file, which I *think* should not be found).
- PR3992.
llvm-svn: 69794
- For now, this means we are always doing the address computations by
hand instead of constructing a proper GEP. Right now, however, this
is less important than having fewer entry points to dealing with
Objective-C interface layout.
llvm-svn: 69787
start of the declspec. The fixit still goes there, and we underline
the declspec. This helps when the start of the declspec came from a
macro that expanded from a system header. For example, we now produce:
t.c:2:8: warning: type specifier missing, defaults to 'int' [-Wimplicit-int]
static x;
~~~~~~ ^
llvm-svn: 69777
- Superclasses are now always laid out their shadow structure at the
first field.
- Prior to this, the entire class heirarchy was flattened into a
single structure which meant that alignment, padding, and bitfields
weren't packed correctly (the ASTRecordLayout was correct however,
which meant our debug info didn't coincide with ivar offsets, for
example).
- This is still very suboptimal, but I believe the ivar layout itself
is now at least close to correct.
- <rdar://problem/6773388> error: objc[29823]: layout bitmap sliding
backwards
llvm-svn: 69771
by marking the predefines buffer as a system header. The problem
with stdint is that it was getting problems like this:
/Volumes/Projects/cvs/llvm/Debug/lib/clang/1.0/include/stdint.h:43:9: warning: 'long long' is an extension when C99 mode is not enabled
typedef __INT64_TYPE__ int64_t;
^
<built-in>:73:29: note: instantiated from:
#define __INT64_TYPE__ long long
^
We correctly silence warnings in system headers, but only if the
spelling location of the token came from the system header. This is
designed so that if you use a system macro in your code that you don't
get punished for its definition. This is all cool except that the
predefines buffer wasn't considered a system header.
llvm-svn: 69770
for scoping to match C99 even when in C89 mode. This patch fixes this
(eliminating a "redefinition of thisKey" error), and also prevents non-sensical
diagnostics in -pedantic mode like this:
t.m:7:8: warning: variable declaration in for loop is a C99-specific feature
for (id thisKey in keys) ;
^
llvm-svn: 69760
extern. Previously we would warn about it and ignore the attribute.
This is incorrect, it should be handled as a c89 "extern inline"
function. Many thanks to Matthieu Castet for pointing this out and
beating me over the head until I got it.
PR3988: extern inline function are not externally visible
llvm-svn: 69756
This optimization improves performance on the Carbon-prefixed "Hello,
World!" example by 57%. For reference, we're now about 2.25x faster
than GCC PCH. We're also pulling in far less of the PCH file:
*** PCH Statistics:
411/20693 types read (1.986179%)
2553/59230 declarations read (4.310316%)
1093/44646 identifiers read (2.448148%)
1/32954 statements read (0.003035%)
21/6187 macros read (0.339421%)
llvm-svn: 69755
When the StoreManager doesn't reason well about pointer-arithmetic, propagate
the non-nullness constraint on a pointer value when performing pointer
arithmetic uisng ++/--.
llvm-svn: 69741
mark exactly the blocks which have references that are "live through".
This fixes a rejects valid:
rdar://6808730 - [sema] [blocks] block rejected at global scope
llvm-svn: 69738
identifiers from a precompiled header.
This patch changes the primary name lookup method for entities within
a precompiled header. Previously, we would load all of the names of
declarations at translation unit scope into a large DenseMap (inside
the TranslationUnitDecl's DeclContext), and then perform a special
"last resort" lookup into this DeclContext when we knew there was a
PCH file (see Sema::LookupName). Now, when we see an identifier named
for the first time, we load all of the declarations with that name
that are visible from the translation unit into the IdentifierInfo's
chain of declarations. Thus, the explicit "look into the translation
unit's DeclContext" code is gone, and Sema effectively uses the same
IdentifierInfo-based name lookup mechanism whether we are using a PCH
file or not.
This approach should help PCH scale with the size of the input program
rather than the size of the PCH file. The "Hello, World!" application
with Carbon.h as a PCH file now loads 20% of the identifiers in the
PCH file rather than 85% of the identifiers.
90% of the 20% of identifiers loaded are actually loaded when we
deserialize the preprocessor state. The next step is to make the
preprocessor load macros lazily, which should drastically reduce the
number of types, declarations, and identifiers loaded for "Hello,
World".
llvm-svn: 69737
into StoreManager::CastRegion. Both methods were practically identical, and this
is core logic that is common to all StoreManagers since it defines the basic
invariants of the abstract memory model.
llvm-svn: 69730
@implementation that closes a @class delcaration.
- I don't know how to make a test case for this, but this strengthens
the invariants that hold internally. The functionality change here
is the edit to SemaDeclObjC.cpp.
llvm-svn: 69728
in cases like this:
typedef struct {
short instance;
char name[0];
} ATTR_LIST_ENTRY2;
void test() {
ATTR_LIST_ENTRY2 X = (ATTR_LIST_ENTRY2) { .instance = 7, };
}
While it is safe to emit them, it is pretty silly.
llvm-svn: 69687
tentative definitions off to the ASTConsumer at the end of the
translation unit.
Eliminate CodeGen's internal tracking of tentative definitions, and
instead hook into ASTConsumer::CompleteTentativeDefinition. Also,
tweak the definition-deferal logic for C++, where there are no
tentative definitions.
Fixes <rdar://problem/6808352>, and will make it much easier for
precompiled headers to cope with tentative definitions in the future.
llvm-svn: 69681
addRecordToClass.
- Among other things, this fixes a crash when applying sizeof to an
interface with synthesized ivars, although things still aren't
"correct" here.
llvm-svn: 69675
that also includes the contents of the IdentifierInfo itself (the
various fields and flags, along with the chain of identifiers visible
at the top level that have that name).
We don't make any use of the hash table yet, except that our
identifier ID -> string mapping points into the hash table now.
llvm-svn: 69625
Sebastian Redl