Prevent looping when DenseSet is abused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70572 91177308-0d34-0410-b5e6-96231b3b80d8

HTML-escape '>' as '>' in sample C++ code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70569 91177308-0d34-0410-b5e6-96231b3b80d8

Allow CONCAT_VECTORS nodes to be legal or have custom lowering for some targets.
Changes to take advantage of this will come later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70560 91177308-0d34-0410-b5e6-96231b3b80d8

Actually insert inserted instructions into the InsertedValues map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70557 91177308-0d34-0410-b5e6-96231b3b80d8

Add an accessor method to allow clients to test if a given expression
is associated with a SCEV expansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70556 91177308-0d34-0410-b5e6-96231b3b80d8

Make RequiresTypeConversion canonicalize the types before calling the
target hooks canLosslesslyBitCastTo and isTruncateFree. This allows
targets to avoid worrying about handling all combinations of integer
and pointer types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70555 91177308-0d34-0410-b5e6-96231b3b80d8

When printing a SCEVUnknown with pointer type, don't print an
artificial "ptrtoint", as it tends to clutter up complicated
expressions. The cast operators now print both source and
destination types, which is usually sufficient.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70554 91177308-0d34-0410-b5e6-96231b3b80d8

Short-circuit inttoptr-ptrtoint constant expressions; these aren't
always folded by the regular constant folder because it doesn't have
TargetData information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70553 91177308-0d34-0410-b5e6-96231b3b80d8

Make SCEVExpander::addInsertedValue able to accept Values, not just
Instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70552 91177308-0d34-0410-b5e6-96231b3b80d8

Minor whitespace fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70551 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an 80-column violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70550 91177308-0d34-0410-b5e6-96231b3b80d8

When creating cast scevs, canonicalize the destination type. This
avoids duplicate scevs that differ only in type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70549 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a compiler warning on hosts where uint64_t isn't unsigned long long.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70548 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some code to work if TargetLowering is not available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70546 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify more code and add timer stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70539 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify more code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70537 91177308-0d34-0410-b5e6-96231b3b80d8

hasSCEV() was declared in ScalarEvolution.h, but never defined. This must have
gone lost during the pImpl conversion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70536 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify some code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70534 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace. It was confusing me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70533 91177308-0d34-0410-b5e6-96231b3b80d8

Lets install the manual page with install-clang!  Radar 6838692

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70529 91177308-0d34-0410-b5e6-96231b3b80d8

Code clean up. Bye bye PhysRegTracker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70524 91177308-0d34-0410-b5e6-96231b3b80d8

Fix compilation for some targets other than x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70522 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment to refer to the section of the programmer's manual that explains what the pointer tagging in Use is for.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70521 91177308-0d34-0410-b5e6-96231b3b80d8

Make DebugLoc independent of DwarfWriter.

-Replace DebugLocTuple's Source ID with CompileUnit's GlobalVariable*
-Remove DwarfWriter::getOrCreateSourceID
-Make necessary changes for the above (fix callsites, etc.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70520 91177308-0d34-0410-b5e6-96231b3b80d8

Set FnEnd in JITEmitter::finishFunction to point strictly to the end of function's machine code.
Don't include memory allocated for global variables during relocations resolution.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70517 91177308-0d34-0410-b5e6-96231b3b80d8

Add a mention of TypeBuilder to the programmer's manual, and clean up the class
comment a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70515 91177308-0d34-0410-b5e6-96231b3b80d8

Join cross class copies using getCommonSubClass()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70513 91177308-0d34-0410-b5e6-96231b3b80d8

getCommonSubClass() - Calculate the largest common sub-class of two register
classes.

This is implemented as a function rather than a method on TargetRegisterClass
because it is symmetric in its arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70512 91177308-0d34-0410-b5e6-96231b3b80d8

Slightly change TableGen's definition of a register subclass.

A subclass is allowed to have a larger spill size than the superclass, and the
spill alignment must be a multiple of the superclass alignment. This causes
the following new subclass relations:

=== Alpha ===
F4RC -> F8RC

=== PPC ===
F4RC -> F8RC

=== SPU ===
R8C -> R16C -> R32C/R32FP -> R64C/R64FP -> GPRC/VECREG

=== X86 ===
FR32  -> FR64  -> VR128
RFP32 -> RFP64 -> RFP80

These subclass relations are consistent with the behaviour of -join-cross-class-copies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70511 91177308-0d34-0410-b5e6-96231b3b80d8

Add some comments, and tidy up some whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70510 91177308-0d34-0410-b5e6-96231b3b80d8

Extend ScalarEvolution's getBackedgeTakenCount to be able to
compute an upper-bound value for the trip count, in addition to
the actual trip count. Use this to allow getZeroExtendExpr and
getSignExtendExpr to fold casts in more cases.

This may eventually morph into a more general value-range
analysis capability; there are certainly plenty of places where
more complete value-range information would allow more folding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70509 91177308-0d34-0410-b5e6-96231b3b80d8

Add a smarter heuristic to determine when to coalesce a virtual register with a physical one. More specifically, it avoid tying a virtual register in the loop with a physical register defined / used outside the loop. When it determines it's not profitable, it will use the physical register as the allocation preference instead.

This is *not* turned on by default. Testing indicates this is just as likely to pessimize code. The main issue seems to be allocation preference doesn't work effectively. That will change once I've taught register allocator "swapping".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70503 91177308-0d34-0410-b5e6-96231b3b80d8

Rename the CurMultiClass formal parameter of TGParser::AddSubMultiClass
so that it doesn't shadow the instance variable of the same name.
Make the parameter names in method declarations match the definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70502 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary "class" keywords.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70499 91177308-0d34-0410-b5e6-96231b3b80d8

Change forward declaration of MultiClass to use the "struct" keyword instead
of "class", so that it matches the subsequent definition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70498 91177308-0d34-0410-b5e6-96231b3b80d8

Make dataflow iteration possible on Value*, not only on User*: df_ext_iterator<Value*, SmallPtrSet<const Value*, 16> >

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70496 91177308-0d34-0410-b5e6-96231b3b80d8

Don't try to mix integers and pointers in an icmp instruction
in getSCEVAtScope.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70495 91177308-0d34-0410-b5e6-96231b3b80d8

Oops! Missed a file in my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70491 91177308-0d34-0410-b5e6-96231b3b80d8

Allow a user of libLTO to specify the full pathname of the gcc executable to
run when assembling.

Wire this up to the gold plugin. You can now pass --plugin-opt gcc=/foo/bar/gcc
and it will run that gcc instead of looking for it on the path.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70490 91177308-0d34-0410-b5e6-96231b3b80d8

Move helper functions for optimizing division by constant into the APInt
class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70488 91177308-0d34-0410-b5e6-96231b3b80d8

Set mayLoad on MOVZX32_NOREXrm8 too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70466 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ScalarEvolution::print to print a value for any Instruction with
a SCEVable type, not just integer types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70463 91177308-0d34-0410-b5e6-96231b3b80d8

Mark MOV8mr_NOREX and MOV8rm_NOREX as mayStore / mayLoad respectively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70461 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused flags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70459 91177308-0d34-0410-b5e6-96231b3b80d8

fix a regression handling indirect results: these need to be considered
memory operands otherwise the writebacks get lost when the inline asm
doesn't otherwise have side effects.  This fixes rdar://6839427, though
clang really shouldn't generate these anymore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70455 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the JIT bindings for ocaml.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70454 91177308-0d34-0410-b5e6-96231b3b80d8

remove progname which is never set. PR4085

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70453 91177308-0d34-0410-b5e6-96231b3b80d8

Error out with bad optimization level specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70449 91177308-0d34-0410-b5e6-96231b3b80d8

Remove LTO optimization level.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70445 91177308-0d34-0410-b5e6-96231b3b80d8

Instead of passing in an unsigned value for the optimization level, use an enum,
which better identifies what the optimization is doing. And is more flexible for
future uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70440 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for a character after a command line option. Like '-Os'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70437 91177308-0d34-0410-b5e6-96231b3b80d8

Print correct instruction in dump.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70427 91177308-0d34-0410-b5e6-96231b3b80d8

Fix infinite recursion in the C++ code which handles movddup by making it unnecessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70425 91177308-0d34-0410-b5e6-96231b3b80d8

Implement getSCEVAtScope for SCEV cast expressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70422 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize the cast-of-addrec folding to handle folding of SCEVs like
(sext i8 {-128,+,1} to i64) to i64 {-128,+,1}, where the iteration
crosses from negative to positive, but is still safe if the trip
count is within range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70421 91177308-0d34-0410-b5e6-96231b3b80d8

Reword and tidy up some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70416 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use 'false' for 'fast isel' here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70411 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test to match the new output from scalar-evolution.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70410 91177308-0d34-0410-b5e6-96231b3b80d8

MachineInstr::isRegTiedTo{Use,Def}Operand can safely be made const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70408 91177308-0d34-0410-b5e6-96231b3b80d8

Include the source type in SCEV cast expression debug output, and
print sext, zext, and trunc, instead of signextend, zeroextend,
and truncate, respectively, for consistency with the main IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70405 91177308-0d34-0410-b5e6-96231b3b80d8

Update comment, replace theoretically impossible check with an assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70391 91177308-0d34-0410-b5e6-96231b3b80d8

Fix recent regression in gcc.dg/pr26719.c (6835035).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70386 91177308-0d34-0410-b5e6-96231b3b80d8

Add directive to declare external globals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70379 91177308-0d34-0410-b5e6-96231b3b80d8

spillPhysRegAroundRegDefsUses() may have invalidated iterators stored in fixed_ IntervalPtrs. Reset them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70378 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for PR4082

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70375 91177308-0d34-0410-b5e6-96231b3b80d8

Implement review feedback for vector shuffle work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70372 91177308-0d34-0410-b5e6-96231b3b80d8

Add a public method called getAddressSpace() to the GlobalAddressSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70366 91177308-0d34-0410-b5e6-96231b3b80d8

Disable the load-shrinking optimization from looking at
anything larger than 64-bits, avoiding a crash.  This should
really be fixed to use APInts, though type legalization happens
to help us out and we get good code on the attached testcase at
least.

This fixes rdar://6836460

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70360 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments to reflect the current code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70357 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more doxygen comments to SCEVAddRec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70354 91177308-0d34-0410-b5e6-96231b3b80d8

Determine allocation 'preference' with right register class. I haven't seen this changing codegen so no test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70351 91177308-0d34-0410-b5e6-96231b3b80d8

The second part of the change from -fast to -O#. This changes the JIT to accept
an optimization level instead of a simple boolean telling it to generate code
"fast" or the other type of "fast".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70347 91177308-0d34-0410-b5e6-96231b3b80d8

Second attempt:

Massive check in. This changes the "-fast" flag to "-O#" in llc. If you want to
use the old behavior, the flag is -O0. This change allows for finer-grained
control over which optimizations are run at different -O levels.

Most of this work was pretty mechanical. The majority of the fixes came from
verifying that a "fast" variable wasn't used anymore. The JIT still uses a
"Fast" flag. I'll change the JIT with a follow-up patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70343 91177308-0d34-0410-b5e6-96231b3b80d8

Correct comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70340 91177308-0d34-0410-b5e6-96231b3b80d8

As with r70333, give the primary induction variable a use so
that it can't be trivially eliminated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70334 91177308-0d34-0410-b5e6-96231b3b80d8

Make this testcase slightly less trivial, so that it doesn't fail
if indvars happens to optimize away the unused primary induction
variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70333 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a grammaro in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70331 91177308-0d34-0410-b5e6-96231b3b80d8

Properly print 'P' modifier on inline asm memory operands.
This should fix PR3379 and PR4064.
Patch inspired by Edwin Török!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70328 91177308-0d34-0410-b5e6-96231b3b80d8

Fix trailing whitespace and 80-col. violations in recent TableGen changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70319 91177308-0d34-0410-b5e6-96231b3b80d8

Move getMatchingSuperReg() out of coalescer and into TargetRegisterInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70309 91177308-0d34-0410-b5e6-96231b3b80d8

Test for llvm-gcc bug fixed by 70301.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70302 91177308-0d34-0410-b5e6-96231b3b80d8

GlobalValue is always pointer type, so an assert isn't required.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70300 91177308-0d34-0410-b5e6-96231b3b80d8

Fix choice of version of Windows callback to use to consider not only the Visual Studio version, but also the Windows SDK version.

Patch by Tareq Siraj.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70299 91177308-0d34-0410-b5e6-96231b3b80d8

Don't coalesce a physical register with an incompatible virtual register.

If the physical register does not belong to the virtual register's regclass,
don't coalesce. The physical register could be an invalid operand for an
instruction using the vreg.

The regclass matching is done after determining the actual subregisters being copied.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70298 91177308-0d34-0410-b5e6-96231b3b80d8

Initialized arrays can be in any address space.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70297 91177308-0d34-0410-b5e6-96231b3b80d8

Move getSubRegisterRegClass from ScheduleDagSDNodesEmit.cpp to a TargetRegisterClass method.
Also make the method non-asserting. It will return NULL when given an invalid subreg index.

The method is needed by an upcoming patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70296 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4034. Bug in LiveInterval::join when it's compacting new valno's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70291 91177308-0d34-0410-b5e6-96231b3b80d8

Initial test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70290 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR4051. When 2address pass delete an instruction, update kill info when necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70279 91177308-0d34-0410-b5e6-96231b3b80d8

r70270 isn't ready yet. Back this out. Sorry for the noise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70275 91177308-0d34-0410-b5e6-96231b3b80d8

Massive check in. This changes the "-fast" flag to "-O#" in llc. If you want to
use the old behavior, the flag is -O0. This change allows for finer-grained
control over which optimizations are run at different -O levels.

Most of this work was pretty mechanical. The majority of the fixes came from
verifying that a "fast" variable wasn't used anymore. The JIT still uses a
"Fast" flag. I'm not 100% sure if it's necessary to change it there...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70270 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an obsolete comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70262 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR 4086, a bug in FP IV elimination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70247 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4076. Correctly create live interval of physical register with two-address update.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70245 91177308-0d34-0410-b5e6-96231b3b80d8

Permit ChangeCompareStride to rewrite a comparison when the factor
between the comparison's iv stride and the candidate stride is
exactly -1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70244 91177308-0d34-0410-b5e6-96231b3b80d8

Add tests for WeakVH and AssertingVH.  These pointed out that the overloads for
the comparison operators were not only unnecessary in the presence of the
implicit conversion; they caused ambiguous overload errors. So I deleted them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70243 91177308-0d34-0410-b5e6-96231b3b80d8

Teach getZeroExtendExpr and getSignExtendExpr to use trip-count
information to simplify [sz]ext({a,+,b}) to {zext(a),+,[zs]ext(b)},
as appropriate.

These functions and the trip count code each call into the other, so
this requires careful handling to avoid infinite recursion. During
the initial trip count computation, conservative SCEVs are used,
which are subsequently discarded once the trip count is actually
known.

Among other benefits, this change lets LSR automatically eliminate
some unnecessary zext-inreg and sext-inreg operation where the
operand is an induction variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70241 91177308-0d34-0410-b5e6-96231b3b80d8

give bitstreamreader an API to ignore names for blocks/records,
only llvm-bcanalyzer wants this info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70239 91177308-0d34-0410-b5e6-96231b3b80d8

Don't skip the CopyMI when removing kill markers.
This should have no effect on generated code, but makes the intermediate state
of the coalescer more sane.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70238 91177308-0d34-0410-b5e6-96231b3b80d8

Now that PR2957 is resolved, remove a bunch of
no-longer needed workarounds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70234 91177308-0d34-0410-b5e6-96231b3b80d8