Fix GetMainExecutable. Patch by Sam Bishop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52847 91177308-0d34-0410-b5e6-96231b3b80d8

Looks like this condition is inverted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52841 91177308-0d34-0410-b5e6-96231b3b80d8

Add a NextPowerOf2 function to calculate the next power of two greater than a given integer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52839 91177308-0d34-0410-b5e6-96231b3b80d8

implement some fixme's by making "autorenaming" in the value symbol table not
thrash the heap with string stuff (e.g. utostr).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52838 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some code to avoid string thrashing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52837 91177308-0d34-0410-b5e6-96231b3b80d8

Use StringSet instead of std::set<std::string>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52836 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52835 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new version of Module::getFunction that takes a const char* instead
of a std::string.  This avoids copying the string to the heap in common
cases.  Patch by Pratik Solanki!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52834 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak IRBuilder to explicitly call the setName(const char*) method on Value
instead of passing the name into the instruction ctors.  Since most
instruction ctors take their name as an std::string, this avoids copying the
string to the heap and a malloc and free.

Patch by Pratik Solanki!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52832 91177308-0d34-0410-b5e6-96231b3b80d8

Add an include file needed for VS.
Add two new file to codegen project (VS).
This unbreak the build for VS.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52831 91177308-0d34-0410-b5e6-96231b3b80d8

Add dominator info printer pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52829 91177308-0d34-0410-b5e6-96231b3b80d8

Provide correct encoding for PPC LWARX instructions.
Patch by Gary Benson!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52828 91177308-0d34-0410-b5e6-96231b3b80d8

Since we are using GCC to assemble the program, make sure the assembly syntax is AT&T.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52827 91177308-0d34-0410-b5e6-96231b3b80d8

Regression test for PR2443.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52826 91177308-0d34-0410-b5e6-96231b3b80d8

Use the c modifier to tell llvm-ar not to issue a
warning when creating the archive (the warning
causes the test to fail).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52824 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce number of times .size() is called on a vector. Rename some variables to
match normal naming scheme.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52820 91177308-0d34-0410-b5e6-96231b3b80d8

Use a SmallSet when we can to reduce memory allocations.
This speeds up a particular testcase from 0.0302s to 0.0222s in LiveVariables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52819 91177308-0d34-0410-b5e6-96231b3b80d8

Cache subregister relationships in a set in TargetRegisterInfo to allow faster lookups.
This speeds up LiveVariables from 0.6279s to 0.6165s on kimwitu++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52818 91177308-0d34-0410-b5e6-96231b3b80d8

fix the regressions from Eric's patch by making GetConstantStringInfo
tolerate a non-nul-terminated string, and handling a direct global
reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52813 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this check, GetConstantStringInfo validates that a
global is constant already.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52812 91177308-0d34-0410-b5e6-96231b3b80d8

test doesn't need eh info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52811 91177308-0d34-0410-b5e6-96231b3b80d8

when linking globals, make sure to preserve the address space of the global.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52810 91177308-0d34-0410-b5e6-96231b3b80d8

More cruft from revert...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52809 91177308-0d34-0410-b5e6-96231b3b80d8

Cruft left from patch revert...sorry. :-(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52808 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting broken patch r52803.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52806 91177308-0d34-0410-b5e6-96231b3b80d8

Don't perform expensive queries checking for super and sub registers when we know that there aren't any.
This speed up LiveVariables on instcombine at -O0 -g from 0.3855s to 0.3503s.  Look for more improvements in this area soon!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52804 91177308-0d34-0410-b5e6-96231b3b80d8

- Remove a use of std::vector.
- Make sure that we're not recalculating the size of a vector
  that never changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52803 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor the DebugInfoDesc stuff out of the MachineModuleInfo file. Clean up
some uses of std::vector, where it's return std::vector by value. Yuck!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52800 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note for Thunderbird users to the Developer Policy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52797 91177308-0d34-0410-b5e6-96231b3b80d8

allow updating the MPM, so that you can use one FunctionPAssManager with
multiple ModuleProviders, e.g. with the JIT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52796 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52795 91177308-0d34-0410-b5e6-96231b3b80d8

Remove warnings about shadowed and unused variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52791 91177308-0d34-0410-b5e6-96231b3b80d8

Update VS project files.  The clang executable now depends on the LLVM Analysis library.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52789 91177308-0d34-0410-b5e6-96231b3b80d8

Use a DenseMap instead of an std::map for mangled name lookup.  This is
improves AsmPrinter runtime on instcombine from 0.3920s to 0.3836s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52787 91177308-0d34-0410-b5e6-96231b3b80d8

duncan points out that isOperationLegal includes a check for
type legality.  Thanks Duncan!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52786 91177308-0d34-0410-b5e6-96231b3b80d8

Don't create a whole new string just to copy the elements into it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52785 91177308-0d34-0410-b5e6-96231b3b80d8

Use the -enable-pre flag so this test doesn't fail.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52784 91177308-0d34-0410-b5e6-96231b3b80d8

Make LLVM compile on DragonFly BSD (PR2499).

Patch by Hasso Tepper!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52781 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing semicolumn in my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52780 91177308-0d34-0410-b5e6-96231b3b80d8

Improve comments wrt createInternalize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52779 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup for unitialized types. Patch by Jean-Daniel Dupas!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52775 91177308-0d34-0410-b5e6-96231b3b80d8

"An improved Mach-O file type detection for sys::IdentifyFileType()
This patch add supports for single architecture mach-o files (the current implementation only support Universal Binary), and solve the signature conflict between java class and Universal Binary magics.

Note that this function will always returned dynamic library for Universal Binaries (like the current implementation) because the binary type is not include in the file header."

Patch by Jean-Daniel Dupas!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52766 91177308-0d34-0410-b5e6-96231b3b80d8

Reserve the size we'll need in advance.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52763 91177308-0d34-0410-b5e6-96231b3b80d8

Allow for rounding up of stack frame.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52751 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes the last x86-64 test failure in compat.exp:
<16 x float> is 64-byte aligned (for some reason),
which gets us into the stack realignment code.  The
computation changing FP-relative offsets to SP-relative
was broken, assiging a spill temp to a location
also used for parameter passing.  This
fixes it by rounding up the stack frame to a multiple
of the largest alignment (I concluded it wasn't fixable
without doing this, but I'm not very sure.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52750 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52749 91177308-0d34-0410-b5e6-96231b3b80d8

Move GetConstantStringInfo to lib/Analysis. Remove
string output routine from Constant. Update all
callers. Change debug intrinsic api slightly to
accomodate move of routine, these now return values
instead of strings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52748 91177308-0d34-0410-b5e6-96231b3b80d8

when we know the signbit of an input to uint_to_fp is zero,
change it to sint_to_fp on targets where that is cheaper (and
visaversa of course).  This allows us to compile uint_to_fp to:

_test:
movl 4(%esp), %eax
shrl $23, %eax
cvtsi2ss %eax, %xmm0
movl 8(%esp), %eax
movss %xmm0, (%eax)
ret

instead of:

.align 3
LCPI1_0: ##  double
.long 0 ## double least significant word 4.5036e+15
.long 1127219200 ## double most significant word 4.5036e+15
.text
.align 4,0x90
.globl _test
_test:
subl $12, %esp
movl 16(%esp), %eax
shrl $23, %eax
movl %eax, (%esp)
movl $1127219200, 4(%esp)
movsd (%esp), %xmm0
subsd LCPI1_0, %xmm0
cvtsd2ss %xmm0, %xmm0
movl 20(%esp), %eax
movss %xmm0, (%eax)
addl $12, %esp
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52747 91177308-0d34-0410-b5e6-96231b3b80d8

Remember which MachineOperand we were processing, so we don't have to scan the list to find it again later.
This speeds up live intervals from 0.37s to 0.30s on instcombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52745 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the text in an assert string.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52744 91177308-0d34-0410-b5e6-96231b3b80d8

- Fix a x86 vector isel bug: illegal transformation of a vector_shuffle into a
  shift.
- Add a readme entry for a missing vector_shuffle optimization that results in
  awful codegen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52740 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for expanding PPC 128 bit floats.
For this it is convenient to permit floats to
be used with EXTRACT_ELEMENT, so I tweaked
things to allow that.  I also added libcalls
for ppcf128 to i32 forms of FP_TO_XINT, since
they exist in libgcc and this case can certainly
occur (and does occur in the testsuite) - before
the i64 libcall was being used.  Also, the
XINT_TO_FP result seemed to be wrong when
the argument is an i128: the wrong fudge
factor was added (the i32 and i64 cases were
handled directly, but the i128 code fell
through to some generic softening code which
seemed to think it was i64 to f32!).  So I
fixed it by adding a fudge factor that I
found in my breakfast cereal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52739 91177308-0d34-0410-b5e6-96231b3b80d8

Implement JIT support for global aliases, patch by David Chisnall!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52738 91177308-0d34-0410-b5e6-96231b3b80d8

Restore DeadArgElim back to 52570. It's breaking 447.dealII.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52736 91177308-0d34-0410-b5e6-96231b3b80d8

Switch the PPC backend and target-independent JIT to use the libsystem
InvalidateInstructionCache method instead of calling through
a hook on the JIT.  This is a host feature, not a target feature.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52734 91177308-0d34-0410-b5e6-96231b3b80d8

fix compilation errors in my previous patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52733 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new InvalidateInstructionCache method to sys::Memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52731 91177308-0d34-0410-b5e6-96231b3b80d8

jump threading can introduce allocas, make sure we promote
them back to registers!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52729 91177308-0d34-0410-b5e6-96231b3b80d8

run mem2reg after the second jump threading pass in llvm-ld.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52728 91177308-0d34-0410-b5e6-96231b3b80d8

remove a bunch of dead options that never did anything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52727 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52726 91177308-0d34-0410-b5e6-96231b3b80d8

SimpleInstructionSelector is here no more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52725 91177308-0d34-0410-b5e6-96231b3b80d8

Add/complete support for integer and float
select_cc and friends.  This code could be
factorized a bit but I'm not sure that it's
worth it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52724 91177308-0d34-0410-b5e6-96231b3b80d8

Pacify gcc-4.3.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52723 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the OrigVT member from AtomicSDNode, as it is redundant with
the base SDNode's VTList.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52722 91177308-0d34-0410-b5e6-96231b3b80d8

simplify shell syntax to work better on solaris, patch by
Nathan Keynes!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52721 91177308-0d34-0410-b5e6-96231b3b80d8

Added MemOperands to Atomic operations since Atomics touches memory.
Added abstract class MemSDNode for any Node that have an associated MemOperand
Changed atomic.lcs => atomic.cmp.swap, atomic.las => atomic.load.add, and
atomic.lss => atomic.load.sub

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52706 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a (false) warning on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52705 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some cosmetics in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52704 91177308-0d34-0410-b5e6-96231b3b80d8

- Use O(1) check of basic block size limit.
- Avoid speculatively execute vector ops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52703 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2488, a case where we deleted stack restores too aggressively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52702 91177308-0d34-0410-b5e6-96231b3b80d8

Enable two-address remat by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52701 91177308-0d34-0410-b5e6-96231b3b80d8

Use push_back rather than operator[], which is incorrect in this cases. Unfortunately, this slow the testcase down a little bit,
but only marginally.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52700 91177308-0d34-0410-b5e6-96231b3b80d8

Use present tense when talking about User layout. It is implemented now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52699 91177308-0d34-0410-b5e6-96231b3b80d8

Append to the ActiveTimers std::vector before looking at the timer instead
of after, so that any reallocation it does doesn't get counted for the pass
being timed.  This probably doesn't account for a timing discrepancy I was
looking into, but I'm fixing it anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52693 91177308-0d34-0410-b5e6-96231b3b80d8

v2f32 is now a valid (MMX) type which breaks this
test (doesn't work for any MMX vector types, it's
not me).  Rewritten to use v2i16 which is generic
and going to stay that way; I think that preserves
the point of the test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52692 91177308-0d34-0410-b5e6-96231b3b80d8

Add v2f32 (MMX) type to X86.  Support is primitive:
load,store,call,return,bitcast.  This is enough to
make call and return work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52691 91177308-0d34-0410-b5e6-96231b3b80d8

In ConstantArray::getAsString(), we know the size of the resultant string in advance so we can pre-allocate it and just fill in
the entries.  This improves the time for the AsmPrinter on InstructionCombining.cpp from 0.4248s to 0.3370s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52690 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVector instead of std::vector for a minor compile time improvement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52689 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 52645, the loop unroller changes. It caused a regression in 252.eon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52688 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52687 91177308-0d34-0410-b5e6-96231b3b80d8

Make Allocate<T>() return a T* instead of a void*. And use
static_cast instead of reinterpret_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52686 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52685 91177308-0d34-0410-b5e6-96231b3b80d8

Use const_cast instead of a C-style cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52684 91177308-0d34-0410-b5e6-96231b3b80d8

Correct a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52683 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for building on solaris, working around namespace
polution problems from system headers.  Patch by Nathan Keynes!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52682 91177308-0d34-0410-b5e6-96231b3b80d8

Pass std::string by reference. Thanks Chris!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52678 91177308-0d34-0410-b5e6-96231b3b80d8

Commit the new DeadArgElim pass again, this time with the gcc bootstrap failures fixed.

Also add a testcase to reproduce the gcc bootstrap failure in very much reduced form.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52677 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the test suite to be checked out into projects/test-suite.

We will keep the old projects/llvm-test working for existing installs.

The changes to configure are made manually, since I lack autoconf-2.6. Someone
might want to run AutoGen.sh to see if that changes anything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52675 91177308-0d34-0410-b5e6-96231b3b80d8

Update testing documentation with more consistent terminology.

All of LLVM's testing features are now referred to as the "Testing
infrastructure", the DejaGNU tests are just that, and the whole program tests
are referred to as the "test suite".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52674 91177308-0d34-0410-b5e6-96231b3b80d8

Rename a few variables to be more consistent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52672 91177308-0d34-0410-b5e6-96231b3b80d8

If it's determined safe, remat MOV32r0 (i.e. xor r, r) and others as it is instead of using the longer MOV32ri instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52670 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some signed vs. unsigned issues in array and vector handling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52664 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note about a potential PIC optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52663 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes for being compiled PIC on Linux. This isn't the most
general solution possible, but it's a fairly simple one.
Based on a patch from the OpenGTL project!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52662 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an incorrect comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52660 91177308-0d34-0410-b5e6-96231b3b80d8

Use InstrSlots::NUM rather than pre-dividing by four.  Also, mark this const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52659 91177308-0d34-0410-b5e6-96231b3b80d8

Comment fixes, and make Schedule() pure virtual.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52658 91177308-0d34-0410-b5e6-96231b3b80d8

A brief survey of priority_queue usage in the tree turned this up
as a questionable case, but the code isn't actually needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52657 91177308-0d34-0410-b5e6-96231b3b80d8

Add a clear() method to PriorityQueue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52656 91177308-0d34-0410-b5e6-96231b3b80d8