No intended functionality change.

llvm-svn: 67478

llvm-svn: 67477

llvm-svn: 67476

to something like:
define void @bar(%struct.foo* noalias sret %agg.result) nounwind {
instead of:
define void @bar(%struct.foo* noalias sret %agg.result, ...) nounwind {

llvm-svn: 67475

llvm-svn: 67474

llvm-svn: 67472

- Make the Diagnostic::Level for PTH errors to be specified by the caller

clang (driver):
- Set the PTHManager diagnostic level to "Diagnostic::Error" for -include-pth
  (a hard error) and Diagnostic::Warning for -token-cache (we can still
  proceed).

llvm-svn: 67462

Keep track of whether a class is abstract or not. This is currently only used for the __is_abstract type trait.

llvm-svn: 67461

build.  This shaves another 3% off.

llvm-svn: 67460

in release-assert builds.  For automatic variables, explicitly set
a name with setName that does not make a temporary std::string.

This speeds up -emit-llvm-only -disable-free on PR3810 by 4.6%

llvm-svn: 67459

ctor function.  This avoids creating a temporary std::string for the name,
speeding up the testcase in PR3810 by 3.8%

llvm-svn: 67457

llvm-svn: 67456

llvm-svn: 67455

copy_helpers and dispose_helpers as necessary for them.

llvm-svn: 67453

lists. The code wasn't accounting for the distinction between the
top-level call to getStructuredSubobjectInit and later calls that
occur deeper in the hierarchy. This problem manifested itself as
over-allocation in cases where we have large arrays of small
structures (<rdar://problem/6707362>).

llvm-svn: 67452

category or protocol.

llvm-svn: 67450

 - PR3854.

I think it makes more sense to change MemoryBuffer::getSTDIN (return 0
should indicate error, not empty), but it is documented to return 0
for empty inputs, and some other code appears to rely on this.

llvm-svn: 67449

 - PR3854.

I think it makes more sense to change MemoryBuffer::getSTDIN (return 0
should indicate error, not empty), but it is documented to return 0
for empty inputs, and some other code appears to rely on this.

llvm-svn: 67448

very simple places, reimplement the deferred decl emission logic to not be O(N^2),
fixing PR3810.

llvm-svn: 67447

function definition.

llvm-svn: 67446

code path.

llvm-svn: 67445

GetAddrOfFunction.  This is simpler and more efficient.

llvm-svn: 67444

CreateFunctionPrototypeIR, though my next patch will eliminate
it entirely.

llvm-svn: 67443

1. it wasn't applying to definitions, only declarations, e.g. int x __asm("foo")
2. multiple definitions were conflicting, they weren't getting merged.
3. the code was duplicated in several places.

llvm-svn: 67442

llvm-svn: 67441

llvm-svn: 67440

llvm-svn: 67439

more optimistic that it will work (optimizing for the common case).

llvm-svn: 67438

llvm-svn: 67437

llvm-svn: 67436

by ~2%.

llvm-svn: 67434

This speeds up getAsIdentifierInfo from being a call to a function
with a big switch to a single testl instruction.  This speeds up
the example in PR3810 by 6.2%

llvm-svn: 67433

llvm-svn: 67432

speeding up the testcase in PR3810 by 60%.

llvm-svn: 67431

other code presumably) by 4.3%

llvm-svn: 67430

This speeds up a testcase in 3810 by ~16%.

llvm-svn: 67429

code; and don't return an error code when -### is present, even if
errors occur.

llvm-svn: 67425

it a hard error.

llvm-svn: 67424

llvm-svn: 67423

its vectors based on the subobject type we're initializing and the
(unstructured) initializer list. This eliminates some malloc thrashing
when parsing initializers (from 117 vector reallocations down to 0
when parsing Cocoa.h). We can't always pre-allocate the right amount
of storage, since designated initializers can cause us to initialize
in non-predictable patterns.

llvm-svn: 67421