Fix PR8811 by teaching MachineVerifier about optional defs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122199 91177308-0d34-0410-b5e6-96231b3b80d8

StrongPHIElimination will never run before TwoAddressInstructionPass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122197 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some code to just reuse a setcc if we can instead of
going through the CSE maps to get it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122196 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new way of silencing this warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122195 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing standard headers. Patch by Joerg Sonnenberger!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122193 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing std:: prefixes to some calls. C++ doesn't require that <cfoo>
headers provide symbols outside namespace std and the LLVM coding standards
state that we should prefix all of them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122192 91177308-0d34-0410-b5e6-96231b3b80d8

teach MaskedValueIsZero how to analyze ADDE.  This is
enough to teach it that ADDE(0,0) is known 0 except the
low bit, for example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122191 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122190 91177308-0d34-0410-b5e6-96231b3b80d8

now that generic vector types aren't selected onto MMX operations,
we don't need -disable-mmx anymore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122189 91177308-0d34-0410-b5e6-96231b3b80d8

now that generic vector types aren't selected onto MMX registers, these
tests don't need -disable-mmx.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122188 91177308-0d34-0410-b5e6-96231b3b80d8

reduce copy/paste programming with the power of for loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122187 91177308-0d34-0410-b5e6-96231b3b80d8

X86 supports i8/i16 overflow ops (except i8 multiplies), we should
generate them.

Now we compile:

define zeroext i8 @X(i8 signext %a, i8 signext %b) nounwind ssp {
entry:
  %0 = tail call %0 @llvm.sadd.with.overflow.i8(i8 %a, i8 %b)
  %cmp = extractvalue %0 %0, 1
  br i1 %cmp, label %if.then, label %if.end

into:

_X:                                     ## @X
## BB#0:                                ## %entry
subl $12, %esp
movb 16(%esp), %al
addb 20(%esp), %al
jo LBB0_2

Before we were generating:

_X:                                     ## @X
## BB#0:                                ## %entry
pushl %ebp
movl %esp, %ebp
subl $8, %esp
movb 12(%ebp), %al
testb %al, %al
setge %cl
movb 8(%ebp), %dl
testb %dl, %dl
setge %ah
cmpb %cl, %ah
sete %cl
addb %al, %dl
testb %dl, %dl
setge %al
cmpb %al, %ah
setne %al
andb %cl, %al
testb %al, %al
jne LBB0_2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122186 91177308-0d34-0410-b5e6-96231b3b80d8

add a general coverage test for overflow intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122185 91177308-0d34-0410-b5e6-96231b3b80d8

move a transformation to a more logical place, simplifying it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122183 91177308-0d34-0410-b5e6-96231b3b80d8

recognize an unsigned add with overflow idiom into uadd.
This resolves a README entry and technically resolves PR4916,
but we still get poor code for the testcase in that PR because
GVN isn't CSE'ing uadd with add, filed as PR8817.

Previously we got:

_test7:                                 ## @test7
addq %rsi, %rdi
cmpq %rdi, %rsi
movl $42, %eax
cmovaq %rsi, %rax
ret

Now we get:

_test7:                                 ## @test7
addq %rsi, %rdi
movl $42, %eax
cmovbq %rsi, %rax
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122182 91177308-0d34-0410-b5e6-96231b3b80d8

optimize uadd(x, cst) into a comparison when the normal
result is dead.  This is required for my next patch to not
regress the testsuite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122181 91177308-0d34-0410-b5e6-96231b3b80d8

add a version of IRBuilder::SetInsertPoint that takes an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122180 91177308-0d34-0410-b5e6-96231b3b80d8

use IC.ReplaceInstUsesWith instead of a raw RAUW so that uses of
the old thing end up on the instcombine worklist.  Not doing this
can cause an extra top-level iteration of instcombine, burning
compile time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122179 91177308-0d34-0410-b5e6-96231b3b80d8

generalize the sadd creation code to not require that the
sadd formed is half the size of the original type. We can
now compile this into a sadd.i8:

unsigned char X(char a, char b) {
  int res = a+b;
  if ((unsigned )(res+128) > 255U)
    abort();
  return res;
}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122178 91177308-0d34-0410-b5e6-96231b3b80d8

fix another miscompile in the llvm.sadd formation logic: it wasn't
checking to see if the high bits of the original add result were dead.
Inserting a smaller add and zexting back to that size is not good enough.

This is likely to be the fix for 8816.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122177 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some checks for StrongPHIElim. These checks make it impossible to use an
alternative register allocator that does not require LiveIntervals by specifying
it on the command-line for a target that has StrongPHIElimination enabled by
default.

These checks are pretty meaningless anyways, since StrongPHIElimination and
PHIElimination are never used at the same time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122176 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug (possibly 8816) in the sadd forming xform: it isn't
profitable (or safe) to promote code when the add-with-constant
has other uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122175 91177308-0d34-0410-b5e6-96231b3b80d8

rework the code added in r122072 to pull it out to its own
helper function, clean up comments, and reduce indentation.
No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122174 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance LICM to promote alias sets whose pointers themselves are stored,
which doesn't affect the memory address being promoted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122172 91177308-0d34-0410-b5e6-96231b3b80d8

fix PR8602, a bug in an assertion: a volatile store *of* a pointer
does not make the alias set for that pointer volatile, just stores
*to* the pointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122171 91177308-0d34-0410-b5e6-96231b3b80d8

fix PR8642: if a critical edge has a PHI value that can trap,
isel is *required* to split the edge.  PHI values get evaluated
on the edge, not in their predecessor block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122170 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead header

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122169 91177308-0d34-0410-b5e6-96231b3b80d8

revert r122164, I'm going to go with a different approach.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122168 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed version of 122160 (the previous one would fold undefined symbols).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122167 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 122160 while I debug it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122165 91177308-0d34-0410-b5e6-96231b3b80d8

first step to fixing PR8642: don't fold away empty basic blocks
which have trapping constant exprs in them due to PHI nodes.
Eliminating them can cause the constant expr to be evalutated
on new paths if the input edges are critical.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122164 91177308-0d34-0410-b5e6-96231b3b80d8

move this test into the ARM test so that it is only run when the arm backend
is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122163 91177308-0d34-0410-b5e6-96231b3b80d8

Move all folding to AttemptToFoldSymbolOffsetDifference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122160 91177308-0d34-0410-b5e6-96231b3b80d8

Support/PathV1: Clarify deprecation warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122159 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122158 91177308-0d34-0410-b5e6-96231b3b80d8

Support/PathV1: Deprecate get{Basename,Dirname,Suffix}.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122157 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122156 91177308-0d34-0410-b5e6-96231b3b80d8

Restore the behavior of frame lowering before my refactoring.
It turns out that ppc backend has really weird interdependencies
over different hooks and all stuff is fragile wrt small changes.
This should fix PR8749

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122155 91177308-0d34-0410-b5e6-96231b3b80d8

Just rename the functions, relying on matching a instruction that has the same name as a symbol is way too fragile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122154 91177308-0d34-0410-b5e6-96231b3b80d8

Test more than just label names and make test work on non-x86 hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122153 91177308-0d34-0410-b5e6-96231b3b80d8

Header warning patrol.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122152 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for lexing single quotes like 'c'.

This fixed 8615.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122150 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r122143 through r122140, which collectively broke the LLVMC tests on
the buildbots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122149 91177308-0d34-0410-b5e6-96231b3b80d8

Merge isAbsolute into IsSymbolRefDifferenceFullyResolved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122148 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the MCObjectFormat class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122147 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test that shows that we produce no fixups when computing the difference
of two symbols in the same fragment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122145 91177308-0d34-0410-b5e6-96231b3b80d8

Add a FIXME and explain a hack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122144 91177308-0d34-0410-b5e6-96231b3b80d8

Support/PathV1: Clarify deprecation warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122143 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122142 91177308-0d34-0410-b5e6-96231b3b80d8

Support/PathV1: Deprecate get{Basename,Dirname,Suffix}.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122141 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122139 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 122011, 122012, 122013, 122023 adding back an important optimization.
I added a note, but suggestions on how to add a test are really welcome.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122138 91177308-0d34-0410-b5e6-96231b3b80d8

Apparently, operandices is not a word.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122135 91177308-0d34-0410-b5e6-96231b3b80d8

Move some data to the TargetWriter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122134 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the inline spiller to attempt folding a load instruction into its single
use before rematerializing the load.

This allows us to produce:

    addps LCPI0_1(%rip), %xmm2

Instead of:

    movaps LCPI0_1(%rip), %xmm3
    addps %xmm3, %xmm2

Saving a register and an instruction. The standard spiller already knows how to
do this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122133 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak debug spew.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122132 91177308-0d34-0410-b5e6-96231b3b80d8

r120333 changed the opcode for the Thumb1 stuff from ARM::tMOVr to
ARM::tMOVgpr2gpr. But this check didn't change. As a result, we were getting
misaligned references to the jump table from an ADR instruction.

There is a test case, but unfortunately it's sensitive to random code changes.

<rdar://problem/8782223>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122131 91177308-0d34-0410-b5e6-96231b3b80d8

RemoveUnusedCPEntries can change things. Track it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122129 91177308-0d34-0410-b5e6-96231b3b80d8

Test for push being relaxed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122124 91177308-0d34-0410-b5e6-96231b3b80d8

Check that the register is live-in to the loop header before inserting copies in
the loop predecessors.

The register can be live-out from a predecessor without being live-in to the
loop header if there is a critical edge from the predecessor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122123 91177308-0d34-0410-b5e6-96231b3b80d8

Fix GCC warning:
  lib/CodeGen/RegAllocGreedy.cpp:311: error: unused variable 'PhysReg' [-Wunused-variable]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122122 91177308-0d34-0410-b5e6-96231b3b80d8

Relax push instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122121 91177308-0d34-0410-b5e6-96231b3b80d8

Make LazyValueInfo non-recursive.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122120 91177308-0d34-0410-b5e6-96231b3b80d8

Rearrange some Neon multiclasses.  No functional changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122119 91177308-0d34-0410-b5e6-96231b3b80d8

Support/PathV1: Deprecate getLast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122116 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Cleanup and document MSVC warning flags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122115 91177308-0d34-0410-b5e6-96231b3b80d8

Add support to CallbackVH to receive notification when a Value's use-list changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122114 91177308-0d34-0410-b5e6-96231b3b80d8

Pass a Banner argument to the machine code verifier both from
createMachineVerifierPass and MachineFunction::verify.

The banner is printed before the machine code dump, just like the printer pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122113 91177308-0d34-0410-b5e6-96231b3b80d8

Fix result type of Neon floating-point comparisons against zero.
The result vector elements are always integers.  Radar 8782191.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122112 91177308-0d34-0410-b5e6-96231b3b80d8

Add some missing entries in ARMTargetLowering::getTargetNodeName.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122111 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace fixes. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122110 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid dereferencing end() in collectInterferingVRegs() when there is no
interference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122108 91177308-0d34-0410-b5e6-96231b3b80d8

Make the -verify-regalloc command line option available to base classes as
RegAllocBase::VerifyEnabled.

Run the machine code verifier in a few interesting places during RegAllocGreedy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122107 91177308-0d34-0410-b5e6-96231b3b80d8

Enable loop splitting in RegAllocGreedy.

The heuristics split around the largest loop where the current register may be
allocated without interference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122106 91177308-0d34-0410-b5e6-96231b3b80d8

Add vector versions of some existing scalar transforms to aid codegen in matching psign & pblend operations to the IR produced by clang/gcc for their C idioms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122105 91177308-0d34-0410-b5e6-96231b3b80d8

During local stack slot allocation, the materializeFrameBaseRegister function
may be called. If the entry block is empty, the insertion point iterator will be
the "end()" value. Calling ->getParent() on it (among others) causes problems.

Modify materializeFrameBaseRegister to take the machine basic block and insert
the frame base register at the beginning of that block. (It's very similar to
what the code does all ready. The only difference is that it will always insert
at the beginning of the entry block instead of after a previous materialization
of the frame base register. I doubt that that matters here.)

<rdar://problem/8782198>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122104 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a DAGCombiner crash when folding binary vector operations with constant
BUILD_VECTOR operands where the element type is not legal.  I had previously
changed this code to insert TRUNCATE operations, but that was just wrong.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122102 91177308-0d34-0410-b5e6-96231b3b80d8

Combine several vector-related DAGCombiner tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122101 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid report_fatal_error in ARM's PrintAsmOperand method.
The standard error handling in AsmPrinter::EmitInlineAsm handles this much
better, so just use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122100 91177308-0d34-0410-b5e6-96231b3b80d8

Remove trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122099 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for matching psign & plendvb to the x86 target
Remove unnecessary pandn patterns, 'vnot' patfrag looks through bitcasts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122098 91177308-0d34-0410-b5e6-96231b3b80d8

Add more checks to IntervalMapOverlaps::advance() to ensure that advanceTo sees
monotonic keys.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122093 91177308-0d34-0410-b5e6-96231b3b80d8

It is allowed to call IntervalMap::const_iterator::advanceTo() with a key that
moves the iterator to end(), and it is valid to call it on end().

That means it is valid to call advanceTo() with any monotonic key sequence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122092 91177308-0d34-0410-b5e6-96231b3b80d8

Thumb's forced-PC-alignment requirement applies to the _total_ displacement, not just to the fragment relative
portion.  While the fragment boundary is usually already aligned, it is possible for it not to be, which
would lead to a non-aligned final displacement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122091 91177308-0d34-0410-b5e6-96231b3b80d8

Add a transform to DAG Combiner.  This improves the
code for the case where 32-bit divide by constant is
turned into 64-bit multiply by constant.  8771012.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122090 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 col.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122089 91177308-0d34-0410-b5e6-96231b3b80d8

Support/Path: Deprecate PathV1::isAbsolute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122086 91177308-0d34-0410-b5e6-96231b3b80d8

PathV2: Use StringRef::substr to simplify substring creation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122085 91177308-0d34-0410-b5e6-96231b3b80d8

Allow missing kill flags on an untied operand of a two-address instruction when
the operand uses the same register as a tied operand:

  %r1 = add %r1, %r1

If add were a three-address instruction, kill flags would be required on at
least one of the uses. Since it is a two-address instruction, the tied use
operand must not have a kill flag.

This change makes the kill flag on the untied use operand optional.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122082 91177308-0d34-0410-b5e6-96231b3b80d8

Fix crash when IntervalMapOverlaps::advanceTo moves past the last overlap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122081 91177308-0d34-0410-b5e6-96231b3b80d8

If The ARM WriteNopData() gets an unaligned byte count to pad out, fill in with
a partial value. rdar://8782954

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122078 91177308-0d34-0410-b5e6-96231b3b80d8

Missed some StringRefRefs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122077 91177308-0d34-0410-b5e6-96231b3b80d8

Add bits 31-28 to the Thumb2 encoding of TBB/TBH.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122076 91177308-0d34-0410-b5e6-96231b3b80d8

Handle 2 and 4 byte data blob fixup values for ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122075 91177308-0d34-0410-b5e6-96231b3b80d8

Pass StringRefs by value, for consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122074 91177308-0d34-0410-b5e6-96231b3b80d8

Add MachineLoopRange comparators for sorting loop lists by number and by area.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122073 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply r121905 (automatic synthesis of @llvm.sadd.with.overflow) with a fix for a bug that manifested itself
on the DragonEgg self-host bot.  Unfortunately, the testcase is pretty messy and doesn't reduce well due to
interactions with other parts of InstCombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122072 91177308-0d34-0410-b5e6-96231b3b80d8

Store and free the TargetObjectWriter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122070 91177308-0d34-0410-b5e6-96231b3b80d8

Stub out explicit MCELFObjectTargetWriter interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122067 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup and document .gitignore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122066 91177308-0d34-0410-b5e6-96231b3b80d8