First step, handle diagnostics in StringLiteral's that are due to token pasting.
For example, we now handle:
id str2 = @"foo"
"bar"
@"baz"
" b\0larg"; // expected-warning {{literal contains NUL character}}
Correctly:
test/SemaObjC/exprs.m:17:15: warning: CFString literal contains NUL character
" b\0larg"; // expected-warning {{literal contains NUL character}}
~~~^~~~~~~
There are several other related issues still to be done.
llvm-svn: 64924
us whether there was an error in trying to parse a type-name (type-id
in C++). This allows propagation of errors further in the compiler,
suppressing more bogus error messages.
llvm-svn: 64922
any named parameters, e.g., this is accepted in C:
void f(...) __attribute__((overloadable));
although this would be rejected:
void f(...);
To do this, moved the checking of the "ellipsis without any named
arguments" condition from the parser into Sema (where it belongs anyway).
llvm-svn: 64902
to do in this area, since there are other places that reference
FunctionDecls.
Don't allow "overloadable" functions (in C) to be declared without a
prototype.
llvm-svn: 64897
the various PPTokens that are pasted together to make it. In the course
of working on this, I discovered ParseObjCStringLiteral which needs some
work. I'll tackle it next.
llvm-svn: 64892
When comparing if one Range is "less" than another, compare the actual APSInt
numeric values instead of their pointer addresses. This ensures that the
ImmutableSet in RangeSet always has a consistent ordering between Ranges. This
is critical for generating the same digest/hash for the contents of the sets.
This was a serious performance bug because it would often cause state caching
to be disabled along complicated paths.
Along the way:
- Put Range and RangeSet in the "anonymous namespace" and mark them hidden
llvm-svn: 64890
back to the summary used when evaluating the statement associated with a
simulation node. This is now being used to help improve the checker's
diagnostics. To get things started, the checker now emits a path diagnostic
indicating that 'autorelease' is a no-op in GC mode.
Some of these changes are exposing further grossness in the interface between
BugReporter and the ExplodedGraph::Trim facilities. These really need to be
cleaned up one day.
llvm-svn: 64881
1. Return of _Complex long double used wrong type.
2. va_arg of types passed in two SSE registers didn't account for
extra space in register save area.
Down to 18 failures on gcc/compat/x86_64. Combined 32/64 results are:
--
=== gcc Summary ===
# of expected passes 1292
# of unexpected failures 34
# of unsupported tests 2
--
llvm-svn: 64880
- This idiom ensures that the result will have the right width and
type.
- Tested on most of x86_64/llvm-test to satisfy my paranoia.
- This fixes at least the following bugs:
o UnaryTypeTraitExpr wasn't setting the width correctly.
o Arithmetic on _Bool wasn't setting the width correctly.
And probably a number more.
llvm-svn: 64864
keep searching for C++ headers when in C++ mode). In theory clang
should be able to find all of its own headers now. If not, the
CPATH or C_INCLUDE_PATH environment variables can be specified to
add a include path.
llvm-svn: 64862
specialization of class templates, e.g.,
template<typename T> class X;
template<> class X<int> { /* blah */ };
Each specialization is a different *Decl node (naturally), and can
have different members. We keep track of forward declarations and
definitions as for other class/struct/union types.
This is only the basic framework: we still have to deal with checking
the template headers properly, improving recovery when there are
failures, handling nested name specifiers, etc.
llvm-svn: 64848
- Renamed to getDeclAlignInBytes since most other query functions
work in bits.
- Fun to track down as isIntegerConstantExpr was getting it right,
but Evaluate() was getting it wrong. Maybe we should assert they
compute the same thing when they succeed?
llvm-svn: 64828
Zhongxing Xu. The resultant code is less than 1/2 the size of the original.
Key highlights:
- All CouldBeXXX methods have been removed. Checking for feasibility is now just
done in the AddXXX methods.
- RangeSets now represent "all possible values" explicitly as the range set {
[min, max] } instead of the empty set. The empty set now represents "no
feasible values". This change consolidated much of the core algorithm to only
have one code path instead of alternate paths that considered the empty set to
represent "all possible falues."
llvm-svn: 64787
IRgen no longer relies on isConstantInitializer, instead we just try
to emit the constant. If that fails then in C we emit an error
unsupported (this occurs when Sema accepted something that it doesn't
know how to fold, and IRgen doesn't know how to emit) and in C++ we
emit a guarded initializer.
This ends up handling a few more cases, because IRgen was actually
able to emit some of the constants Sema accepts but can't Evaluate().
For example, PR3398.
llvm-svn: 64780
diagnostics. I'm not sure I want to keep this, but hey, it's easy
and could be useful or something, even if guarded by a
-fshow-me-tons-of-details option. A silly example is:
#define A B
#define C A
#define D C
int y = D;
We now emit:
t.c:11:9: error: use of undeclared identifier 'B'
int y = D;
^
t.c:9:11: note: instantiated from:
#define D C
^
t.c:8:11: note: instantiated from:
#define C A
^
t.c:7:11: note: instantiated from:
#define A B
^
A more useful example is from tgmath:
t.c:4:9: error: no matching function for call to '__tg_acos'
return acos(x);
^~~~~~~
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:51:17: note: instantiated from:
#define acos(x) __tg_acos(x)
^
... candidate set follows ...
This does not yet print ranges in instantiation info, (e.g. highlighting the
range "__tg_acos(x)" in the last example), but that could be added if we
decide this is a good idea :).
Thoughts and bug reports welcome!
llvm-svn: 64761
t.c:4:9: error: invalid type 'short *' to __real operator
__tg_choose (__real__(z), C##f(z), (C)(z), C##l(z)),
^
instead of:
t.c:4:9: error: invalid type 'short *' to __real or __imag operator
__tg_choose (__real__(z), C##f(z), (C)(z), C##l(z)),
^
fixing a fixme. It would be even fancier to get the spelling of the token, but I
don't care *that* much :)
llvm-svn: 64759
as the last non-note diagnostic that preceeded them. This ensures
that diagnostics in main files which have notes with locations in
system headers get all the bits and pieces emitted or not in a
unit. This fixes PR3215.
llvm-svn: 64746
highlight the arguments to the macro as well as the identifier.
Before:
t.c:3:9: error: no matching function for call to '__tg_acos'; candidates are:
return acos(x);
^~~~
after:
t.c:3:9: error: no matching function for call to '__tg_acos'; candidates are:
return acos(x);
^~~~~~~
llvm-svn: 64743
functions, so if we're declaring a static we should implicitly declare
a library function by the same name (e.g., malloc, strdup). Fixes PR3592.
llvm-svn: 64736
possible future "data loss" warnings, and b) makes it intensely
obvious to the user what the impl of these functions do in a
ambiguity error.
llvm-svn: 64724
Now no candidates are printed because the 'notes' are in a system header. :(
#include <tgmath-sofar.h>
double foo2(short *x) {
return acos(x);
}
t.c:10:10: error: no matching function for call to '__tg_acos'; candidates are:
return acos(x);
^~~~
1 diagnostic generated.
llvm-svn: 64722
produces really horrible diagnostics when overload ambiguities
happen:
t.c:10:10: error: call to '__tg_acos' is ambiguous; candidates are:
return acos(x);
^~~~
In file included from t.c:1:
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
/Users/sabre/llvm/Debug/Headers/tgmath-sofar.h:49:1: note: candidate function
__TG_RC_1(x, acos, cacos)
^
A possible fix is to just not use macros for this, which I'll probably go for,
but it would be nice to emit the type at the call, so we know what we asked for!
llvm-svn: 64720
CXXRecordDecl that is used to represent class template
specializations. These are canonical declarations that can refer to
either an actual class template specialization in the code, e.g.,
template<> class vector<bool> { };
or to a template instantiation. However, neither of these features is
actually implemented yet, so really we're just using (and uniqing) the
declarations to make sure that, e.g., A<int> is a different type from
A<float>. Note that we carefully distinguish between what the user
wrote in the source code (e.g., "A<FLOAT>") and the semantic entity it
represents (e.g., "A<float, int>"); the former is in the sugared Type,
the latter is an actual Decl.
llvm-svn: 64716
makes -emit-html do nice things for code like:
#define FOO(X) y
int FOO(4
);
highlighting the FOO instance as well as the ) on the next line properly.
llvm-svn: 64710
- Define pow[lf]?, sqrt[lf]? as builtins.
- Add -fmath-errno option which binds to LangOptions.MathErrno
- Add new builtin flag Builtin::Context::isConstWithoutErrno for
functions which can be marked as const if errno isn't respected for
math functions. Sema automatically marks these functions as const
when they are defined, if MathErrno=0.
- IRgen uses const attribute on sqrt and pow library functions to
decide if it can use the llvm intrinsic.
llvm-svn: 64689
Doug: please verify that it is expected that LastIdx can be less that
NumInits. And perhaps add a comment so that Chris doesn't break your
code. :)
llvm-svn: 64688
- If a declaration is an invalid redeclaration of an existing name,
complain about the invalid redeclaration then avoid adding it to
the AST (we can still parse the definition or initializer, if any).
- If the declaration is invalid but there is no prior declaration
with that name, introduce the invalid declaration into the AST
(for later error recovery).
- If the declaration is an invalid redeclaration of a builtin that
starts with __builtin_, we produce an error and drop the
redeclaration. If it is an invalid redeclaration of a library
builtin (e.g., malloc, printf), warn (don't error!) and drop the
redeclaration.
If a user attempts to define a builtin, produce an error and (if it's
a library builtin like malloc) suggest -ffreestanding.
This addresses <rdar://problem/6097585> and PR2892. However, PR3588 is
still going to cause some problems when builtins are redeclared
without a prototype.
llvm-svn: 64639
would be in one place. Since, now, only DeclNodes.def needs to be modified, move things out-of-line and simplify the DeclContext class.
llvm-svn: 64630
-In DeclNodes.def, only mark as DeclContexts the top classes that directly derive from DeclContext. If the Decl has subclasses,
it should be marked with DECL_CONTEXT_BASE.
-Use DeclNodes.def to automate the DeclContext::classof and DeclContext::CastTo definitions.
llvm-svn: 64629
DiagnoseUseOfDeprecatedDecl method. This ensures that they
are treated consistently. This gets us 'unavailable' support
on a few new types of decls, and makes sure we consistently
silence deprecated when the caller is also deprecated.
llvm-svn: 64612
*end* of a macro instantiation, not the start of it. This is
really all about bug-for-bug compatibility with GCC, but not
doing this breaks the FreeBSD kernel.
llvm-svn: 64604
Now instead of just tracking the expansion history, also track the full
range of the macro that got replaced. For object-like macros, this doesn't
change anything. For _Pragma and function-like macros, this means we track
the locations of the ')'.
This is required for PR3579 because apparently GCC uses the line of the ')'
of a function-like macro as the location to expand __LINE__ to.
llvm-svn: 64601
which consequently caused a Seg fault. during meta-data
generation. It also addresses an issue related to
late binding of newly synthesize ivars (when we support it).
llvm-svn: 64563
about, whether they are builtins or not. Use this to add the
appropriate "format" attribute to NSLog, NSLogv, asprintf, and
vasprintf, and to translate builtin attributes (from Builtins.def)
into actual attributes on the function declaration.
Use the "printf" format attribute on function declarations to
determine whether we should do format string checking, rather than
looking at an ad hoc list of builtins and "known" function names.
Be a bit more careful about when we consider a function a "builtin" in
C++.
llvm-svn: 64561
ConstraintManager:
- constify getSymVal()
BasicConstraintManager:
- Pull out logic that would be common to ConstraintManagers of a similar nature
and put them in a parent class called 'SimpleConstraintManager'.
RangeConstraintManager:
- Added a new prototype ConstraintManager to track ranges of variables! This
ConstraintManager keeps tracks of ranges of concrete integers that a symbolic
integer may have.
AnalysisConsumer:
- Add driver option to use RangeConstraintManager with GRExprEngine-based
analyses.
llvm-svn: 64558
1) implement parser and sema support for reading and verifying attribute(warnunusedresult).
2) rename hasLocalSideEffect to isUnusedResultAWarning, inverting the sense
of its result.
3) extend isUnusedResultAWarning to directly return the loc and range
info that should be reported to the user. Make it substantially more
precise in some cases than what was previously reported.
4) teach isUnusedResultAWarning about CallExpr to decls that are
pure/const/warnunusedresult, fixing a fixme.
5) change warn_attribute_wrong_decl_type to not pass in english strings, instead,
pass in integers and use %select.
llvm-svn: 64543
symbols from an SVal.
- Fixed a bug in EnvironmentManager::RemoveDeadBindings() where it did not mark
live all the symbols reachable from a live block-level expression.
- Fixed a bug in the retain/release checker where it did not stop tracking
symbols that 'escaped' via compound literals being assigned to something the
BasicStoreManager didn't reason about.
llvm-svn: 64534
we can define builtins such as fprintf, vfprintf, and
__builtin___fprintf_chk. Give a nice error message when we need to
implicitly declare a function like fprintf.
llvm-svn: 64526
- Added a new 'node builder' class called GRStmtNodeBuilderRef (name may
change). This is essentially a smart reference to a GRStmtNodeBuilder object
that keeps track of the current context (predecessor node, GRExprEngine
object, etc.) The idea is to gradually simplify the interface between
GRExprEngine and GRTransferFuncs using this new builder (i.e., passing 1
argument instead of 5). It also handles some of the "auto-transition" for node
creation, simplifying some of the logic in GRExprEngine itself.
- Used GRStmtBuilderRef to replace GRTransferFuncs::EvalStore with
GRTransferFuncs::EvalBind. The new EvalBind method will be used at any
arbitrary places where a binding between a location and value takes place.
Moreover, GRTransferFuncs no longer has the responsibility to request
StoreManager to do the binding; this is now in GRExprEngine::EvalBind. All
GRTransferFuncs::EvalBind does is checker-specific logic (which can be a
no-op).
llvm-svn: 64525
printf-like functions, both builtin functions and those in the
C library. The function-call checker now queries this attribute do
determine if we have a printf-like function, rather than scanning
through the list of "known functions IDs". However, there are 5
functions they are not yet "builtins", so the function-call checker
handles them specifically still:
- fprintf and vfprintf: the builtins mechanism cannot (yet)
express FILE* arguments, so these can't be encoded.
- NSLog: the builtins mechanism cannot (yet) express NSString*
arguments, so this (and NSLogv) can't be encoded.
- asprintf and vasprintf: these aren't part of the C99 standard
library, so we really shouldn't be defining them as builtins in
the general case (and we don't seem to have the machinery to make
them builtins only on certain targets and depending on whether
extensions are enabled).
llvm-svn: 64512
etc.) when we perform name lookup on them. This ensures that we
produce the correct signature for these functions, which has two
practical impacts:
1) When we're supporting the "implicit function declaration" feature
of C99, these functions will be implicitly declared with the right
signature rather than as a function returning "int" with no
prototype. See PR3541 for the reason why this is important (hint:
GCC always predeclares these functions).
2) If users attempt to redeclare one of these library functions with
an incompatible signature, we produce a hard error.
This patch does a little bit of work to give reasonable error
messages. For example, when we hit case #1 we complain that we're
implicitly declaring this function with a specific signature, and then
we give a note that asks the user to include the appropriate header
(e.g., "please include <stdlib.h> or explicitly declare 'malloc'"). In
case #2, we show the type of the implicit builtin that was incorrectly
declared, so the user can see the problem. We could do better here:
for example, when displaying this latter error message we say
something like:
'strcpy' was implicitly declared here with type 'char *(char *, char
const *)'
but we should really print out a fake code line showing the
declaration, like this:
'strcpy' was implicitly declared here as:
char *strcpy(char *, char const *)
This would also be good for printing built-in candidates with C++
operator overloading.
The set of C library functions supported by this patch includes all
functions from the C99 specification's <stdlib.h> and <string.h> that
(a) are predefined by GCC and (b) have signatures that could cause
codegen issues if they are treated as functions with no prototype
returning and int. Future work could extend this set of functions to
other C library functions that we know about.
llvm-svn: 64504
a target.
Make Preprocessor.cpp define a new __INTPTR_TYPE__ macro based on this.
On linux/32, set intptr_t to int, instead of long. This fixes PR3563.
llvm-svn: 64495
- Fix emission of static functions with constructor attribute while I
was here.
<rdar://problem/6140899> [codegen] "static" and attribute-constructor interact poorly
llvm-svn: 64488
wine sources. This was happening because HighlightMacros was
calling EnterMainFile multiple times on the same preprocessor
object and getting an assert due to the new #line stuff (the
file in question was bison output with #line directives).
The fix for this is to not reenter the file. Instead,
relex the tokens in raw mode, swizzle them a bit and repreprocess
the token stream. An added bonus of this is that rewrite macros
will now hilight the macro definition as well as its uses. Woo.
llvm-svn: 64480
by DeclContexts (always) rather than by statements.
DeclContext currently goes out of its way to avoid destroying any
Decls that might be owned by a DeclGroupOwningRef. However, in an
error-recovery situation, a failure in a declaration statement can
cause all of the decls in a DeclGroupOwningRef to be destroyed after
they've already be added into the DeclContext. Hence, DeclContext is
left with already-destroyed declarations, and bad things happen. This
problem was causing failures that showed up as assertions on x86 Linux
in test/Parser/objc-forcollection-neg-2.m.
llvm-svn: 64474
Currently only used for 128-bit integers.
Note that we can't use the fixed-width integer types for other integer
modes without other changes because glibc headers redefines (u)int*_t
and friends using the mode attribute. For example, this means that uint64_t
has to be compatible with unsigned __attribute((mode(DI))), and
uint64_t is currently defined to long long. And I have a feeling we'll
run into issues if we try to define uint64_t as something which isn't
either long or long long.
This doesn't get the alignment right in most cases, including
the 128-bit integer case; I'll file a PR shortly. The gist of the issue
is that the targets don't really expose the information necessary to
figure out the alignment outside of the target description, so there's a
non-trivial amount of work involved in getting it working right. That
said, the alignment used is conservative, so the only issue with the
current implementation is ABI compatibility.
This makes it trivial to add some sort of "bitwidth" attribute to make
arbitrary-width integers; I'll do that in a followup.
We could also use this for stuff like the following for compatibility
with gcc, but I have a feeling it would be a better idea for clang to be
consistent between C and C++ modes rather than follow gcc's example for
C mode.
struct {unsigned long long x : 33;} x;
unsigned long long a(void) {return x.x+1;}
llvm-svn: 64434
- Add 'EvalBind', which will be used by 'EvalStore' to pull much of the value binding logic out of GRTransferFuncs.
- Rename many cases of 'St' to 'state'.
llvm-svn: 64426
only insert spaces between tokens if the code had them or if they
are actually required to avoid pasting. This reuses the same
logic as -E mode.
llvm-svn: 64421
given name in a given scope is marked as "overloadable", every
function declaration and definition with that same name and in that
same scope needs to have the "overloadable" attribute. Essentially,
the "overloadable" attribute is not part of attribute merging, so it
must be specified even for redeclarations. This keeps users from
trying to be too sneaky for their own good:
double sin(double) __attribute__((overloadable)); // too sneaky
#include <math.h>
Previously, this would have made "sin" overloadable, and therefore
given it a mangled name. Now, we get an error inside math.h when we
see a (re)declaration of "sin" that doesn't have the "overloadable"
attribute.
llvm-svn: 64414
ABI to the CodeGen library. Since C++ code-generation is so
incomplete, we can't exercise much of this mangling code. However, a
few smoke tests show that it's doing the same thing as GCC. When C++
codegen matures, we'll extend the ABI tester to verify name-mangling
as well, and complete the implementation here.
At this point, the major client of name mangling is in the uses of the
new "overloadable" attribute in C, which allows overloading. Any
"overloadable" function in C (or in an extern "C" block in C++) will
be mangled the same way that the corresponding C++ function would be
mangled.
llvm-svn: 64413
_GCC_LIMITS_H_ is defined, when __GNUC__ is defined.
Also, we need to stay away from possible conflicts with header guards.
We should use CLANG_ to prefix all header guards.
llvm-svn: 64408
union subobject initialization before checking whether the next
initiailizer was actually a designated initializer. This led to
spurious "excess elements in union initializer" errors. Thanks to
rdivacky for reporting the bug!
llvm-svn: 64392
tried to put FIXMEs on the most important things to fix up. Lots left
to do including more codegen, more documentation and cleaning code and
style cleanups.
llvm-svn: 64390
- rename isObjCIdType/isObjCClassType -> isObjCIdStructType/isObjCClassStructType. The previous name didn't do what you would expect.
- add back isObjCIdType/isObjCClassType to do what you would expect. Not currently used, however many of the isObjCIdStructType/isObjCClassStructType clients could be converted over time.
- move static Sema function areComparableObjCInterfaces to ASTContext (renamed to areComparableObjCPointerTypes, since it now operates on pointer types).
llvm-svn: 64385
to use this stat information in the PTH file using a 'StatSysCallCache' object.
Performance impact (Cocoa.h, PTH):
- number of stat calls reduces from 1230 to 425
- fsyntax-only: time improves by 4.2%
We can reduce the number of stat calls to almost zero by caching negative stat
calls and directory stat calls in the PTH file as well.
llvm-svn: 64353
- set the 'StatSysCallCache' object using a setter method instead of
FileManager's constructor. This allows the cache to be installed after the
FileManager object is created.
- Add 'file mode' to FileEntry (useful for stat caching)
llvm-svn: 64351
for use by FileManager. FileManager now takes a StatSysCallCache* in its
constructor (which defaults to NULL). This will be used for evaluating whether
or not caching 'stat' system calls in PTH is a performance win. This shim adds
no observable performance impact in the case where the 'StatSysCallCache*' is
null.
llvm-svn: 64345
complex conversions where the conversion between the real types is an
integral promotion. This is how G++ handles complex promotions for its
complex integer extension.
llvm-svn: 64344
system. Since C99 doesn't have overloading and C++ doesn't have
_Complex, there is no specification for this. Here's what I think
makes sense.
Complex conversions come in several flavors:
- Complex promotions: a complex -> complex conversion where the
underlying real-type conversion is a floating-point promotion. GCC
seems to call this a promotion, EDG does something else. This is
given "promotion" rank for determining the best viable function.
- Complex conversions: a complex -> complex conversion that is
not a complex promotion. This is given "conversion" rank for
determining the best viable function.
- Complex-real conversions: a real -> complex or complex -> real
conversion. This is given "conversion" rank for determining the
best viable function.
These rules are the same for C99 (when using the "overloadable"
attribute) and C++. However, there is one difference in the handling
of floating-point promotions: in C99, float -> long double and double
-> long double are considered promotions (so we give them "promotion"
rank), while C++ considers these conversions ("conversion" rank).
llvm-svn: 64343
This commit adds a new attribute, "overloadable", that enables C++
function overloading in C. The attribute can only be added to function
declarations, e.g.,
int *f(int) __attribute__((overloadable));
If the "overloadable" attribute exists on a function with a given
name, *all* functions with that name (and in that scope) must have the
"overloadable" attribute. Sets of overloaded functions with the
"overloadable" attribute then follow the normal C++ rules for
overloaded functions, e.g., overloads must have different
parameter-type-lists from each other.
When calling an overloaded function in C, we follow the same
overloading rules as C++, with three extensions to the set of standard
conversions:
- A value of a given struct or union type T can be converted to the
type T. This is just the identity conversion. (In C++, this would
go through a copy constructor).
- A value of pointer type T* can be converted to a value of type U*
if T and U are compatible types. This conversion has Conversion
rank (it's considered a pointer conversion in C).
- A value of type T can be converted to a value of type U if T and U
are compatible (and are not both pointer types). This conversion
has Conversion rank (it's considered to be a new kind of
conversion unique to C, a "compatible" conversion).
Known defects (and, therefore, next steps):
1) The standard-conversion handling does not understand conversions
involving _Complex or vector extensions, so it is likely to get
these wrong. We need to add these conversions.
2) All overloadable functions with the same name will have the same
linkage name, which means we'll get a collision in the linker (if
not sooner). We'll need to mangle the names of these functions.
llvm-svn: 64336
actually *slightly* slower than the binary search. Since this is algorithmically
better, further performance tuning should be able to make this faster.
llvm-svn: 64326