re-apply r80197, now that iterator.h is not mentioned any more

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80254 91177308-0d34-0410-b5e6-96231b3b80d8

Hopefully the final missing part :(
scalar_to_vector is fully legal now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80251 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot about actual change :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80250 91177308-0d34-0410-b5e6-96231b3b80d8

scalar_to_vector is fully legal now (implemented as subreg accesses)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80249 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an entry from ValueMap before removing elements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80247 91177308-0d34-0410-b5e6-96231b3b80d8

Ok, sometimes it's profitable to turn scalar_to_vector stuff into subreg access.
Add a testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80246 91177308-0d34-0410-b5e6-96231b3b80d8

Transform float scalar_to_vector into subreg accesses.
No idea whether this is profitable or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80245 91177308-0d34-0410-b5e6-96231b3b80d8

STRD and LDRD require ARMv5TE, not just ARMv5T.
See http://llvm.org/PR4687 for more info and links.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80244 91177308-0d34-0410-b5e6-96231b3b80d8

Inverse logic to increase portability.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80240 91177308-0d34-0410-b5e6-96231b3b80d8

To make mcc16 run correctly on mac.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80239 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Add MCCodeEmitter support, for encoding instructions.
 - No relocations yet, of course.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80235 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/X86: Implement single instruction encoding interface for MC.
 - Note, this is a gigantic hack, with the sole purpose of unblocking further
   work on the assembler (its also possible to test the mathcer more completely
   now).

 - Despite being a hack, its actually good enough to work over all of 403.gcc
   (although some encodings are probably incorrect). This is a testament to the
   beauty of X86's MachineInstr, no doubt! ;)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80234 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Print encodings after the instruction, and only when we have an asm
printer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80233 91177308-0d34-0410-b5e6-96231b3b80d8

X86: Mark EH_RETURN as code-gen-only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80232 91177308-0d34-0410-b5e6-96231b3b80d8

Add {MCInst,MCOperand}::{print,dump}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80231 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Only show instruction encodings with --show-encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80230 91177308-0d34-0410-b5e6-96231b3b80d8

the buildbots revealed one more breakage. fix. (why didn't I see this?)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80227 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up the minor mess I caused with removing iterator.h. I shall take care of 80-col violations and the FIXME later. (Thanks goodness that I live in another continent, so the monkeypox did not strike me :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80224 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a new optimization in the inliner: if inlining multiple
calls into a function and if the calls bring in arrays, try to merge
them together to reduce stack size.  For example, in the testcase
we'd previously end up with 4 allocas, now we end up with 2 allocas.

As described in the comments, this is not really the ideal solution
to this problem, but it is surprisingly effective.  For example, on
176.gcc, we end up eliminating 67 arrays at "gccas" time and another
24 at "llvm-ld" time.

One piece of concern that I didn't look into: at -O0 -g with
forced inlining this will almost certainly result in worse debug
info.  I think this is acceptable though given that this is a case
of "debugging optimized code", and we don't want debug info to
prevent the optimizer from doing things anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80215 91177308-0d34-0410-b5e6-96231b3b80d8

the inliner shouldn't crash on this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80214 91177308-0d34-0410-b5e6-96231b3b80d8

For now, only run MC tests if X86 is configured.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80213 91177308-0d34-0410-b5e6-96231b3b80d8

This is passing for PPC on Mac OS X.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80210 91177308-0d34-0410-b5e6-96231b3b80d8

Revert accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80208 91177308-0d34-0410-b5e6-96231b3b80d8

unbreak the build, yay for symlinks + makefiles. :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80205 91177308-0d34-0410-b5e6-96231b3b80d8

reduce header #include'age

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80204 91177308-0d34-0410-b5e6-96231b3b80d8

enhance InlineFunction to be able to optionally return
a the list of static allocas that it inlined.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80203 91177308-0d34-0410-b5e6-96231b3b80d8

smallvectorize the list of returns built by CloneAndPruneFunctionInto.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80202 91177308-0d34-0410-b5e6-96231b3b80d8

remove CloneTrace, which appears to be dead since 2004.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80201 91177308-0d34-0410-b5e6-96231b3b80d8

reduce inlining factor some stuff out to a static helper function,
and other code cleanups.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80199 91177308-0d34-0410-b5e6-96231b3b80d8

Revert accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80198 91177308-0d34-0410-b5e6-96231b3b80d8

--- Reverse-merging r80147 into '.':
A    include/llvm/ADT/iterator.cmake
U    autoconf/configure.ac
--- Reverse-merging r80161 into '.':
U    cmake/config-ix.cmake
--- Reverse-merging r80171 into '.':
U    Makefile
--- Reverse-merging r80173 into '.':
U    configure
U    include/llvm/Config/config.h.in
--- Reverse-merging r80180 into '.':
A    include/llvm/ADT/iterator.h.in

Despite common miscomceptions, iterator.h is alive and well. It broke the build
bots for several hours. And yet no one bothered to look at them.

Gabor and Doug, please review your changes and make sure that they actually
build before resubmitting them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80197 91177308-0d34-0410-b5e6-96231b3b80d8

Add some checks for MachineCFG consistency. Use AnalyzeBranch and
do extra checking when it succeeds, as those are cases where
CodeGen will be doing particularly interesting CFG modifications.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80196 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMake

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80195 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Tweak MCCodeEmitter skeleton.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80193 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize the PoisonMemory member before initializing
members that call methods that read the PoisonMemory member.
This fixes potential spurious (though probably otherwise
harmless) poising of unused memory, and fixes the
associated valgrind error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80192 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4789. Teach eliminateFrameIndex how to handle VLDRQ and VSTRQ which cannot fold any immediate offset.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80191 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing declarations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80190 91177308-0d34-0410-b5e6-96231b3b80d8

Sketch TargetRegistry support for MCCodeEmitter abstract interface.
 - Of course, nothing actually can provide this interface yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80188 91177308-0d34-0410-b5e6-96231b3b80d8

Try to make MSVC just a little happier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80187 91177308-0d34-0410-b5e6-96231b3b80d8

X86FastISel support for loading and storing values of type i1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80186 91177308-0d34-0410-b5e6-96231b3b80d8

Expand i8 selects into control flow instead of 16-bit conditional
moves. This avoids the need to promote the operands (or implicitly
extend them, a partial register update condition), and can reduce
i8 register pressure. This substantially speeds up code such as
write_hex in lib/Support/raw_ostream.cpp.

subclass-coalesce.ll is too trivial and no longer tests what it was
originally intended to test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80184 91177308-0d34-0410-b5e6-96231b3b80d8

now that include/llvm/ADT/iterator.h.in is not attempted to be built, its prerequisite can go away too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80180 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80176 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate configure

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80173 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate references to ADT/iterator.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80171 91177308-0d34-0410-b5e6-96231b3b80d8

Make this into a static method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80170 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bad length argument to substr calls.  Apparently I'm the first one to
attempt more than 2 constraints on an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80169 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Unique sections properly, so we don't get duplicate text
sections, etc.
 - The quick and dirty way, just clone the TargetLoweringObjectFile
   code. Eventually this should be shared... somehow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80168 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Don't put assembler temporary labels in the symbol table.
 - I moved section creation back into AsmParser. I think policy decisions like
   this should be pushed higher, not lower, when possible (in addition the
   assembler has flags which change this behavior, for example).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80162 91177308-0d34-0410-b5e6-96231b3b80d8

De-bork CMake build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80161 91177308-0d34-0410-b5e6-96231b3b80d8

Rework getPersonalityIndex slightly - 0 is now a valid and not-NULL
personality function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80153 91177308-0d34-0410-b5e6-96231b3b80d8

nobody includes llvm/ADT/iterator.h any more,
so get rid of this monstrosity. iterator.h.in is scheduled for deletion in my working copy,
but I wait till I see that configure gets regenerated, as it depends on it. I'll commit
then.
There are still some AC_* tests in the configure.ac dealing with iterators, those can
be zapped probably too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80147 91177308-0d34-0410-b5e6-96231b3b80d8

If we're emitting additional CIEs due to personality functions
don't emit the default one. Explicitly check for the NULL
CIE later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80146 91177308-0d34-0410-b5e6-96231b3b80d8

Nuke trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80145 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Set .subsections_via_symbols flag properly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80144 91177308-0d34-0410-b5e6-96231b3b80d8

Don't submit test directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80139 91177308-0d34-0410-b5e6-96231b3b80d8

Allocate the module provider in the Kaleidoscope code on the heap, not the stack, so that it can be properly deleted.  Also update the tutorial with the new code.  This fixes PR4762, hopefully better than the last time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80138 91177308-0d34-0410-b5e6-96231b3b80d8

Remove all the LLVM_COMPACTIFY_SENTINELS-related macro magic as discussed with Chris on IRC. Anybody wanting to debug sentinel dereferencing problems must revert this patch and perform the indicated modifications.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80128 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Removed outdated TODO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80124 91177308-0d34-0410-b5e6-96231b3b80d8

Generate section for bss and enable weak symbols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80121 91177308-0d34-0410-b5e6-96231b3b80d8

Convert some more Neon tests to FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80120 91177308-0d34-0410-b5e6-96231b3b80d8

Alter 79292 to produce output that actually assembles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80119 91177308-0d34-0410-b5e6-96231b3b80d8

Allow unsigned long long DenseMapInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80118 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unneeded ARM-specific DAG nodes for VLD* and VST* Neon operations.
The instructions can be selected directly from the intrinsics.  We will need
to add some ARM-specific nodes for VLD/VST of 3 and 4 128-bit vectors, but
those are not yet implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80117 91177308-0d34-0410-b5e6-96231b3b80d8

Add extload expansion for f128

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80116 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak FP128 stuff in cbe

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80115 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore -fPIC test on Windows. Suggested by Yonggang Luo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80111 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak CMake build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80109 91177308-0d34-0410-b5e6-96231b3b80d8

Expand scalar_to_vector - we don't have any isel logic for it now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80107 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments detailing a known bug, so that people writing other
backends don't use it as an example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80105 91177308-0d34-0410-b5e6-96231b3b80d8

-fast is now -O0. -fast-isel is no longer experimental.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80104 91177308-0d34-0410-b5e6-96231b3b80d8

Move ProfileInfo::Edge's operator<< out of line. Among other benefits,
this eliminates the ATTRIBUTE_USED, which wasn't being used in a manner
acceptable to some GCC versions, according to the buildbots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80103 91177308-0d34-0410-b5e6-96231b3b80d8

Bugfix for r80100, forgot include. Sorry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80101 91177308-0d34-0410-b5e6-96231b3b80d8

Implemented comments from Daniel Dunbar.
(See http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20090817/084958.html)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80100 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use INSERT_SUBREG to model anyext operations on x86-64, as it
leads to partial-register definitions. To help avoid redundant
zero-extensions, also teach the h-register matching patterns that
use movzbl to match anyext as well as zext.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80099 91177308-0d34-0410-b5e6-96231b3b80d8

Create a ScalarEvolution-based AliasAnalysis implementation.

This is a simple AliasAnalysis implementation which works by making
ScalarEvolution queries. ScalarEvolution has a more complete understanding
of arithmetic than BasicAA's collection of ad-hoc checks, so it handles
some cases that BasicAA misses, for example p[i] and p[i+1] within the
same iteration of a loop.

This is currently experimental. It may be that the main use for this pass
will be to help find cases where BasicAA can be profitably extended, or
to help in the development of the overall AliasAnalysis infrastructure,
however it's also possible that it could grow up to become a directly
useful pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80098 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a missing newline, now that Value's operator<< doesn't add one of its own.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80096 91177308-0d34-0410-b5e6-96231b3b80d8

Use SetVector instead of std::set so that alias relations are tested and
printed in a deterministic order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80095 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Add support for relocations.
 - I haven't really tried to find the "right" way to store the fixups or apply
   them, yet. This works, but isn't particularly elegant or fast.

 - Still no evaluation support, so we don't actually ever not turn a fixup into
   a relocation entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80089 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Move symbol indices into the MCSymbolData structure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80088 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Dump relocations and section data (optionally) in my Mach-O dumper.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80087 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Add symbol entries for undefined symbols used in .fill and .org.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80086 91177308-0d34-0410-b5e6-96231b3b80d8

Add dummy inline asm handling for 'r' constraint. This fixes PR4778

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80085 91177308-0d34-0410-b5e6-96231b3b80d8

Moved isDeclaration() check further down to allow for function counts for
declarations if necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80084 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Change MCContext value table to take const MCSymbol*s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80079 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Make MCValue take const MCSymbol*s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80078 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Make non-sensical max bytes to .align an error.

Also, warn about overflow in alignment values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80077 91177308-0d34-0410-b5e6-96231b3b80d8

Changed std::cout to outs(), retaining formating.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80076 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 79977. It causes llvm-gcc bootstrap failures on some platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80073 91177308-0d34-0410-b5e6-96231b3b80d8

some mips and some sparc compilers apparently
predefine mips and sparc respectively.  Just overrule them :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80072 91177308-0d34-0410-b5e6-96231b3b80d8

test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80070 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Fix tests for python variations in int printing, sigh.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80069 91177308-0d34-0410-b5e6-96231b3b80d8

fix some funky indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80068 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Add section padding where needed (to align the next section).

Also, simplify some of Mach-O writer code which can now use section addresses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80067 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc/Mach-O: Set addresses for symbols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80065 91177308-0d34-0410-b5e6-96231b3b80d8

Add an 'inline hint' attribute to represent source
code hints that it would be a good idea to inline
a function ("inline" keyword).  No functional change
yet; FEs do not emit this and inliner does not use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80063 91177308-0d34-0410-b5e6-96231b3b80d8

Add isClosure() predicate. This is used to add DW_AT_APPLE_block attribute.
Patch by Caroline Tice.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80061 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Improve indirect symbol support (add the indirect index table).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80059 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80058 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-mc: Add MCSection::isDefined()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@80057 91177308-0d34-0410-b5e6-96231b3b80d8