Fix trailing whitespace from ScalarEvolution::print.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73666 91177308-0d34-0410-b5e6-96231b3b80d8

ARM unified syntax is not specific to ELF; use it for Darwin, too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73665 91177308-0d34-0410-b5e6-96231b3b80d8

Teach ScalarEvolution how to recognize another xor(and(x, C), C) case.
If C is a single bit and the and gets analyzed as a truncate and
zero-extend, the xor can be represnted as an add.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73664 91177308-0d34-0410-b5e6-96231b3b80d8

Delete unnecessary braces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73663 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply r73647 in a non-broken form.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73662 91177308-0d34-0410-b5e6-96231b3b80d8

Fix asm string from MOVi16

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73661 91177308-0d34-0410-b5e6-96231b3b80d8

Thumb2 instructions are enabled only in unified assembler mode.
Emit switch directive for it. I have no idea whether this is
requirement for Darwin or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73660 91177308-0d34-0410-b5e6-96231b3b80d8

Instcombine zext(trunc(x) & mask) to x&mask, even if the trunc has
multiple users.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73656 91177308-0d34-0410-b5e6-96231b3b80d8

Use double-checked locking for this lazy initialization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73653 91177308-0d34-0410-b5e6-96231b3b80d8

Protect the GC table in Function.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73647 91177308-0d34-0410-b5e6-96231b3b80d8

Honour calling convention and attributes of Callee by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73646 91177308-0d34-0410-b5e6-96231b3b80d8

Use atomic increment here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73643 91177308-0d34-0410-b5e6-96231b3b80d8

Thread-safe (and ManagedStatic-ized) LeakDetector!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73642 91177308-0d34-0410-b5e6-96231b3b80d8

Guard mutation of the timing info global.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73639 91177308-0d34-0410-b5e6-96231b3b80d8

We need to use double-checked locking for lazy initialization in this case when running multithreaded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73636 91177308-0d34-0410-b5e6-96231b3b80d8

Fix link to BasicBlock doxygen documentation in Programmer's Manual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73635 91177308-0d34-0410-b5e6-96231b3b80d8

VNInfo cleanup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73634 91177308-0d34-0410-b5e6-96231b3b80d8

Add -disable-output to a bunch of tests that don't care about the output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73633 91177308-0d34-0410-b5e6-96231b3b80d8

This fixes a bug introduced in 72661, which can
move loads back past a check that the load address
is valid, see new testcase.  The test that went
in with 72661 has exactly this case, except that
the conditional it's moving past is checking
something else; I've settled for changing that
test to reference a global, not a pointer.  It
may be possible to scan all the tests you pass and
make sure none of them are checking any component
of the address, but it's not trivial and I'm not
trying to do that here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73632 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out some common code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73631 91177308-0d34-0410-b5e6-96231b3b80d8

Add an RAII ScopedWriter, which allows one to acquire a writer lock for the duration of a scope.  Simplify a lot of uses of
writer locks in Constants.cpp by using it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73630 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the locking on the Constants tables, and make it more efficient, by pushing it into the ValueMap from the callers.
Document those ValueMap functions that are _not_ locked, so that callers are aware that they need to do the locking themselves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73628 91177308-0d34-0410-b5e6-96231b3b80d8

Add debug message about non-local loads being clobbered.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73625 91177308-0d34-0410-b5e6-96231b3b80d8

Type safety for Constants.cpp!  Some of this is temporary, as I'm planning to push some of the R/W locking into FoldingSet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73624 91177308-0d34-0410-b5e6-96231b3b80d8

Initial support for some Thumb2 instructions.
Patch by Viktor Kutuzov and Anton Korzh from Access Softek, Inc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73622 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments to use doxygen syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73621 91177308-0d34-0410-b5e6-96231b3b80d8

Protect the ValueHandle table.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73620 91177308-0d34-0410-b5e6-96231b3b80d8

We need to guard reads of the AbstractTypeUsers list, as well as writes to it.  While it would be nice to use a R/W lock here,
we can't, because it HAS to be recursive.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73617 91177308-0d34-0410-b5e6-96231b3b80d8

Define LLVM_NATIVE_ARCH in CMake, so that lli can actually JIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73616 91177308-0d34-0410-b5e6-96231b3b80d8

Add RWMutex.cpp to the CMake makefiles

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73615 91177308-0d34-0410-b5e6-96231b3b80d8

Type safety for TypeSymbolTable!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73614 91177308-0d34-0410-b5e6-96231b3b80d8

make sure that JIT examples link in their appropriate target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73613 91177308-0d34-0410-b5e6-96231b3b80d8

remove two headers subsumed by TargetSelect.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73612 91177308-0d34-0410-b5e6-96231b3b80d8

switch to using llvm/Target/TargetSelect.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73611 91177308-0d34-0410-b5e6-96231b3b80d8

Add a utility header that makes it easy to link in the right set
of targets for various purposes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73610 91177308-0d34-0410-b5e6-96231b3b80d8

Improve the Win32 reader-writer lock implementation by making it just a normal
lock.  This is obviously bad, but at least it's threadsafe!  If you know how
to improve this in a pre-Vista friendly well, patches welcome!

Patch by Max Burke.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73607 91177308-0d34-0410-b5e6-96231b3b80d8

Fix libLTO by #include'ing the initializers for all targets and all asm
printers.

While I'm here, alphabetize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73606 91177308-0d34-0410-b5e6-96231b3b80d8

>> What if my global variable was into a different address space than stack?
>>
>
> It doesn't matter in terms of semantics: because AnalyzeGlobal
> returned false, we're guaranteed the address of the global is never
> taken.  I wouldn't be surprised if we end up generating invalid IR in
> some cases, though, because of the semantics of replaceAllUsesWith.
> Do you have a testcase that breaks?
>
>
The problem is replaceAllUsesWith asserts for type mismatch here. Try attached .bc with llvm-ld.

assert(New->getType() == getType() &&
        "replaceAllUses of value with new value of different type!");

Since stack is always on address space zero, I don't think that type of GV in a different address space is ever going to match.
The other way is to allow replaceAllUsesWith to ignore address spaces while comparing types. (do we have  a way to do that ?).
But then such an optimization may fail the entire idea of user wanting to place a variable into different memory space. The original idea of user might be to save on the stack space (data memory) and hence he asked the variable to be placed into different memory space (program memory). So the best bet here is to deny this optimization by checking

GV->getType()->getAddressSpace() == 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73605 91177308-0d34-0410-b5e6-96231b3b80d8

make CreateFMul forward to CreateFMul, not CreateMul.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73603 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammaro, and bad indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73602 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73601 91177308-0d34-0410-b5e6-96231b3b80d8

Formatting fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73600 91177308-0d34-0410-b5e6-96231b3b80d8

Correct an accidental duplication of the test (patch doesn't handle
creating new files very well).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73599 91177308-0d34-0410-b5e6-96231b3b80d8

PR3439: Correct a silly mistake in the SimplifyDemandedUseBits code for
SRem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73598 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73597 91177308-0d34-0410-b5e6-96231b3b80d8

Another small documentation update.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73596 91177308-0d34-0410-b5e6-96231b3b80d8

Use Doug's new LLVM_NATIVE_ARCH macro in config.h to link in the native
target so that the JIT works in LLI, not just the interpreter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73595 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ScalarEvolution's Xor handling to not assume that an And
that gets recognized with a SCEVZeroExtendExpr must be an And
with a low-bits mask. With r73540, this is no longer the case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73594 91177308-0d34-0410-b5e6-96231b3b80d8

Do not use first actual instruction's location for prologue. The debug wants to skip prologue while setting a breakpoint for the function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73592 91177308-0d34-0410-b5e6-96231b3b80d8

Update auto-generated configuration files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73591 91177308-0d34-0410-b5e6-96231b3b80d8

Define LLVM_NATIVE_ARCH in llvm/Config/config.h to be the LLVM back end that corresponds to the native executable, but only when that LLVM back end is being built

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73590 91177308-0d34-0410-b5e6-96231b3b80d8

Use atomic increment/decrement for reference counting of Type's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73588 91177308-0d34-0410-b5e6-96231b3b80d8

Add an atomic increment and decrement implementation, which will be used for
thread-safe reference counting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73587 91177308-0d34-0410-b5e6-96231b3b80d8

Add locking around the accessors for AbstractTypeUsers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73586 91177308-0d34-0410-b5e6-96231b3b80d8

Remove support for building LLVM libraries into "relinked"
object files.  Now we always build LLVM libraries into archives (.a files).

This makes the 'make' build work more like the cmake build, among other
things.  Doing this exposed some latent circular library dependencies, so
I think that llvm-config wasn't quite right for .o files anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73579 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot this file in my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73578 91177308-0d34-0410-b5e6-96231b3b80d8

Use a reader-writer lock to guard large portions of the Type infrastructure, including abstract type refinement.
There's still some more work to be done here, such as guarding removeAbstractTypeUser() and the printers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73575 91177308-0d34-0410-b5e6-96231b3b80d8

fix a circular dependency between the mips code generator
and its asmprinter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73573 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the EE bindings to use the proper EE interfaces for creating a JIT.
Calling into createJIT directly creates a circular dependency between libjit and lib ee.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73567 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Only include LLVM example executables in the build if requested by the user

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73566 91177308-0d34-0410-b5e6-96231b3b80d8

Test for llvm-gcc patch 73564.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73565 91177308-0d34-0410-b5e6-96231b3b80d8

Accidentally broke this file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73552 91177308-0d34-0410-b5e6-96231b3b80d8

Add newline at end of file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73551 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the Win32 implementation, since it doesn't compile pre-Vista.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73550 91177308-0d34-0410-b5e6-96231b3b80d8

Fix #include guards.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73548 91177308-0d34-0410-b5e6-96231b3b80d8

Make the test target-neutral

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73547 91177308-0d34-0410-b5e6-96231b3b80d8

Fix/cleanup trailing newlines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73546 91177308-0d34-0410-b5e6-96231b3b80d8

Add a portable wrapper for reader-writer locks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73545 91177308-0d34-0410-b5e6-96231b3b80d8

Introduce new headers whose inclusion forces linking and
initialization of all targets (InitializeAllTargets.h) or assembler
printers (InitializeAllAsmPrinters.h). This is a step toward the
elimination of relinked object files, so that we can build normal
archives.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73543 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize a few more instcombines to be vector/scalar-independent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73541 91177308-0d34-0410-b5e6-96231b3b80d8

Instcombine's ShrinkDemandedConstant may strip bits out of constants,
obscuring what would otherwise be a low-bits mask.  Use ComputeMaskedBits
to compute what ShrinkDemandedConstant knew about to reconstruct a
low-bits mask value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73540 91177308-0d34-0410-b5e6-96231b3b80d8

CALL64pcrel32 immediate field is 32-bit. Patch by Abhinav Duggal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73536 91177308-0d34-0410-b5e6-96231b3b80d8

Address review comments: add 3 ARM calling conventions.
Dispatch C calling conv. to one of these conventions based on
target triple and subtarget features.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73530 91177308-0d34-0410-b5e6-96231b3b80d8

GNU as refuses to assemble "pop {}" instruction. Do not emit such
(this is the case when we have thumb vararg function with single
callee-saved register, which is handled separately).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73529 91177308-0d34-0410-b5e6-96231b3b80d8

Add a configure check for pthread_rwlock_init.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73523 91177308-0d34-0410-b5e6-96231b3b80d8

Update the threading section to reflect current plans/implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73521 91177308-0d34-0410-b5e6-96231b3b80d8

Add Threading.cpp to the CMake project files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73516 91177308-0d34-0410-b5e6-96231b3b80d8

Fix validation errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73515 91177308-0d34-0410-b5e6-96231b3b80d8

Split the thread-related APIs out into their own file, and add a few more
calls for convenience.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73512 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for r73506

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73508 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize instcombine's isSafeToLoadUnconditionally() function
to ignore readonly calls, and factor it out of instcombine so
that it can be used by other passes.  Patch by Frits van Bommel!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73506 91177308-0d34-0410-b5e6-96231b3b80d8

Added missing libcalls for floating point to int conversions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73491 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the "llvm.umul.with.overflow is broken" warning; it should be
essentially working now, at least for common cases, and it shouldn't
silently generate bad code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73484 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed names of libcalls checked in r73480.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73483 91177308-0d34-0410-b5e6-96231b3b80d8

Code Restructuring. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73481 91177308-0d34-0410-b5e6-96231b3b80d8

Added required libcalls for PIC16 (mostly floating points to integer casting operations).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73480 91177308-0d34-0410-b5e6-96231b3b80d8

Rename RemoveCopiesFromValNo to TurnCopiesFromValNoToImpDefs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73479 91177308-0d34-0410-b5e6-96231b3b80d8

If a val# is defined by an implicit_def and it is being removed, all of the copies off the val# were removed. This causes problem later since the scavenger will see uses of registers without defs. The proper solution is to change the copies into implicit_def's instead.

TurnCopyIntoImpDef turns a copy into implicit_def and remove the val# defined by it. This causes an scavenger assertion later if the def reaches other blocks. Disable the transformation if the value live interval extends beyond its def block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73478 91177308-0d34-0410-b5e6-96231b3b80d8

Add some generic expansion logic for SMULO and UMULO.  Fixes UMULO
support for x86, and UMULO/SMULO for many architectures, including PPC
(PR4201), ARM, and Cell. The resulting expansion isn't perfect, but it's
not bad.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73477 91177308-0d34-0410-b5e6-96231b3b80d8

Misc accumulated tweaks to legalization logic for various targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73476 91177308-0d34-0410-b5e6-96231b3b80d8

another xform that is target-independent (should be done in instcombine).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73472 91177308-0d34-0410-b5e6-96231b3b80d8

I think instcombine should unconditionally do this xform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73471 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4336: Iterating over use-def chains doesn't seem to be deterministic.

The problem was that BitcodeReader::materializeModule would read functions
from the bc file in densemap pointer key order (doubly non-deterministic!),
which would cause the use-def chains to be set up for globals in
non-determinstic order.  Non-determinstic use/def chains can cause
nondeterminism in many places down-stream.

Many thanks to Julien Lerouge for putting together the pass in the PR that
shows the issue!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73470 91177308-0d34-0410-b5e6-96231b3b80d8

There doesn't seem to be a reason to move the save FP stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73468 91177308-0d34-0410-b5e6-96231b3b80d8

The DWARF to compact encoding converter assumes that the DW_CFA_def_cfa_offset
comes after the DW_CFA_def_cfa_register, because the CFA is really ESP from the
start of the function and only gets an offset when the "subl $xxx,%esp"
instruction happens, not the other way around.

And reapply r72898:

The DWARF unwind info was incorrect. While compiling with
`-fomit-frame-pointer', we would lack the DW_CFA_advance_loc information for a
lot of function, and then they would be `0'. The linker (at least on Darwin)
needs to encode the stack size. In some cases, the stack size is too large to
directly encode. So the linker checks to see if there is a "subl $xxx,%esp"
instruction at the point where the `DW_CFA_def_cfa_offset' says the pc was. If
so, the compact encoding records the offset in the function to where the stack
size is embedded. But because the `DW_CFA_advance_loc' instructions are missing,
it looks before the function and dies.

So, instead of emitting the EH debug label before the stack adjustment
operations, emit it afterwards, right before the frame move stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73465 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73464 91177308-0d34-0410-b5e6-96231b3b80d8

Use MainCU if it is available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73457 91177308-0d34-0410-b5e6-96231b3b80d8

Add initial stab at documenting the use of LLVM with threaded clients.

Comments welcome!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73456 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments to ConstantInt::get and ConstantFP::get to more fully
describe their behavior.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73454 91177308-0d34-0410-b5e6-96231b3b80d8

Use Type::getScalarType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@73451 91177308-0d34-0410-b5e6-96231b3b80d8