Extract optimization pass selection code from llvm-gcc into a separate routine.
This can be used by other stand alone tools, such as 'opt'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56229 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed Bug 2751
http://llvm.org/bugs/show_bug.cgi?id=2751

Abicall was enabled even when static code model was provided
in the command line.
The correct behavior is to disable abicall when static is
specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56228 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable SelectionDAG CSE for calls. It matters in the case of
libcalls, as in this testcase on ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56226 91177308-0d34-0410-b5e6-96231b3b80d8

Added testcase for bswap allegrexel intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56225 91177308-0d34-0410-b5e6-96231b3b80d8

End of the GlobalsModRef experiment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56222 91177308-0d34-0410-b5e6-96231b3b80d8

Update the LLVM polygen grammar for recent language changes:
x86_ssecallcc, function notes, and some whitespace adjustments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56221 91177308-0d34-0410-b5e6-96231b3b80d8

Fix documentation for these functions; they do not only modify users within the
block specified by the InstList. Patch by Stefanus Du Toit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56200 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly update kill infos after extending a live range and merge 2 val#'s; fix 56165 - do not mark val# copy field if the copy does not define the val#.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56199 91177308-0d34-0410-b5e6-96231b3b80d8

Add "sample" documentation to sample project.

This fixes part of PR2793.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56198 91177308-0d34-0410-b5e6-96231b3b80d8

Fix WriteAsOperand to not emit a leading space character. Adjust
its callers to emit a space character before calling it when a
space is needed.

This fixes several spurious whitespace issues in
ScalarEvolution's debug dumps. See the test changes for
examples.

This also fixes odd space-after-tab indentation in the output
for switch statements, and changes calls from being printed like
this:
  call void @foo( i32 %x )
to this:
  call void @foo(i32 %x)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56196 91177308-0d34-0410-b5e6-96231b3b80d8

adjust last patch per review feedback

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56194 91177308-0d34-0410-b5e6-96231b3b80d8

minor correction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56190 91177308-0d34-0410-b5e6-96231b3b80d8

Remove isImm(), isReg(), and friends, in favor of
isImmediate(), isRegister(), and friends, to avoid confusion
about having two different names with the same meaning. I'm
not attached to the longer names, and would be ok with
changing to the shorter names if others prefer it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56189 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2792: treat volatile loads as writing memory somewhere.
Treat stores as reading memory, just to play safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56188 91177308-0d34-0410-b5e6-96231b3b80d8

Fix random abort.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56184 91177308-0d34-0410-b5e6-96231b3b80d8

Define CallSDNode, an SDNode subclass for use with ISD::CALL.
Currently it just holds the calling convention and flags
for isVarArgs and isTailCall.

And it has several utility methods, which eliminate magic
5+2*i and similar index computations in several places.

CallSDNodes are not CSE'd. Teach UpdateNodeOperands to handle
nodes that are not CSE'd gracefully.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56183 91177308-0d34-0410-b5e6-96231b3b80d8

Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56182 91177308-0d34-0410-b5e6-96231b3b80d8

Rely on instruction format to determine so_reg operand for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56181 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 56176. All those instruction formats are still needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56180 91177308-0d34-0410-b5e6-96231b3b80d8

Accidentially flipped the condition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56179 91177308-0d34-0410-b5e6-96231b3b80d8

Add debug dumps.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56178 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate unnecessary instruction formats.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56176 91177308-0d34-0410-b5e6-96231b3b80d8

Addrmode 1 S bit can be dynamically set. Look for CPSR def.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56172 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite address mode 1 code emission routines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56171 91177308-0d34-0410-b5e6-96231b3b80d8

The "alias" keyword comes first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56170 91177308-0d34-0410-b5e6-96231b3b80d8

On some targets, non-move instructions can become move instructions because of coalescing. e.g.
vr2 = OR vr0, vr1
=>
vr2 = OR vr1, vr1   // after coalescing vr0 with vr1

Update the value# of the destination register with the copy instruction if that happens.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56165 91177308-0d34-0410-b5e6-96231b3b80d8

Change ConstantSDNode and ConstantFPSDNode to use ConstantInt* and
ConstantFP* instead of APInt and APFloat directly.

This reduces the amount of time to create ConstantSDNode
and ConstantFPSDNode nodes when ConstantInt* and ConstantFP*
respectively are already available, as is the case in
SelectionDAGBuild.cpp. Also, it reduces the amount of time
to legalize constants into constant pools, and the amount of
time to add ConstantFP operands to MachineInstrs, due to
eliminating ConstantInt::get and ConstantFP::get calls.

It increases the amount of work needed to create new constants
in cases where the client doesn't already have a ConstantInt*
or ConstantFP*, such as legalize expanding 64-bit integer constants
to 32-bit constants. And it adds a layer of indirection for the
accessor methods. But these appear to be outweight by the benefits
in most cases.

It will also make it easier to make ConstantSDNode and
ConstantFPNode more consistent with ConstantInt and ConstantFP.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56162 91177308-0d34-0410-b5e6-96231b3b80d8

Pass "earlyclobber" bit through to machine
representation; coalescer and RA need to know
about it.  No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56161 91177308-0d34-0410-b5e6-96231b3b80d8

Rename ConstantSDNode::getValue to getZExtValue, for consistency
with ConstantInt. This led to fixing a bug in TargetLowering.cpp
using getValue instead of getAPIntValue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56159 91177308-0d34-0410-b5e6-96231b3b80d8

Give GlobalsModRef a whirl in the nightly testers.
I placed it just before GVN because that it is the
pass most likely to benefit from it.  Some quick
and dirty testing confirms that this is a decent
place for it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56144 91177308-0d34-0410-b5e6-96231b3b80d8

Rather than marking all internal globals "Ref"
when a readonly declaration is called, set a
flag.  This is faster and uses less memory.
In theory it is less accurate, because before
only those internal globals that were read
by someone were being marked "Ref", but now
all are.  But in practice, thanks to other
passes, all internal globals of the kind
considered here will be both read and stored
to: those only read will have been turned
into constants, and those only stored to will
have been deleted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56143 91177308-0d34-0410-b5e6-96231b3b80d8

The sequence for ppcf128 compares was not IEEE
safe in the presence of NaNs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56136 91177308-0d34-0410-b5e6-96231b3b80d8

On 64-bit targets, change 32-bit getelementptr indices to be 64-bit
getelementptr indices, inserting an explicit cast if necessary.
This helps expose the sign-extension operation to other optimizations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56133 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a vectorshuffle instcombine bug introduced by r55995.
Patch by Nicolas Capens!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56129 91177308-0d34-0410-b5e6-96231b3b80d8

Add indirect tail call (function pointer) examples.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56127 91177308-0d34-0410-b5e6-96231b3b80d8

udpate header comment: s/VP/VFP/

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56126 91177308-0d34-0410-b5e6-96231b3b80d8

When tailcallopt is enabled all fastcc calls must have an aligned argument stack size. Add a test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56119 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2748. Avoid coalescing physical register with virtual register which would create illegal extract_subreg. e.g.
vr1024 = extract_subreg vr1025, 1
...
vr1024 = mov8rr AH
If vr1024 is coalesced with AH, the extract_subreg is now illegal since AH does not have a super-reg whose sub-register 1 is AH.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56118 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in ANY_EXTEND handling that was breaking 403.gcc on X86-64 in fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56117 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56116 91177308-0d34-0410-b5e6-96231b3b80d8

Intrinsics don't touch internal global variables
(unless passed one via a parameter), even if they
are IntrWriteMem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56115 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a copy+paste bug that Duncan spotted. For several
cases it was still getting lucky and detecting overflow
but it was clearly incorrect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56113 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2783 - coalescer bug. Missing a TargetRegisterInfo::isVirtualRegister check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56112 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate some unused methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56108 91177308-0d34-0410-b5e6-96231b3b80d8

Indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56107 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/SubtargetFeature.cpp asserts that the FeatureKV[] table be sorted
by its first field, but TableGen doesn't actually enforce creating it that
way. TableGen sorts the records that will be used to create it by the names
of the records, not the Name field of those records.

This patch corrects the sort to use the "Name" field of the record as the
sort key.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56106 91177308-0d34-0410-b5e6-96231b3b80d8

Intrinsics don't read these kinds of global
variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56105 91177308-0d34-0410-b5e6-96231b3b80d8

Change getSubReg semantics. It now returns zero if the specified register doesn't have a subreg of the specified index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56099 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a 80 column violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56097 91177308-0d34-0410-b5e6-96231b3b80d8

The version of AtomicSDNode::AtomicSDNode used (only) for
cmp-and-swap reversed the Cmp and Swap arguments; comments
make it clear this is unintentional.  Unfortunately, the
x86 BE had a compensating reversal, which is removed here.
PPC is OK.

From inspection of the Alpha code I think it is OK, but
if somebody has that platform please check it out.  I
cannot test on that platform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56091 91177308-0d34-0410-b5e6-96231b3b80d8

If ISD::ANY_EXTEND fails, try ISD::ZERO_EXTEND and ISD::SIGN_EXTEND before giving up.  This fixes 445.gobmk on
X86-64 in fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56088 91177308-0d34-0410-b5e6-96231b3b80d8

Succumb utterly to compatibility and implement
__sync_fetch_and_nand as ANDC, even though that's
not what nand means.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56087 91177308-0d34-0410-b5e6-96231b3b80d8

Propagate subreg index when promoting a load to a copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56085 91177308-0d34-0410-b5e6-96231b3b80d8

In my analysis for r56076 I missed the case where the original
multiplication overflows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56082 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an icmp+sdiv optimization to check for and handle an overflow
condition. This fixes PR2740.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56076 91177308-0d34-0410-b5e6-96231b3b80d8

Add more documentation advertising the -view-*-dags options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56073 91177308-0d34-0410-b5e6-96231b3b80d8

X86FastISel support for double->float and float->double casts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56070 91177308-0d34-0410-b5e6-96231b3b80d8

FastISel support for i1 PHI nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56069 91177308-0d34-0410-b5e6-96231b3b80d8

FastISel support for i1 constants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56068 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in the coalescer where it didn't check if a live interval existed before trying to manipulate it.  This
was exposed by fast isel's handling of shifts on X86-64.  With this, FreeBench/pcompress2 passes on X86-64 in fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56067 91177308-0d34-0410-b5e6-96231b3b80d8

Add X86FastISel support for static allocas, and refences
to static allocas. As part of this change, refactor the
address mode code for laods and stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56066 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2664 - spiller GetRegForReload wasn't respecting sub-register indices on machine operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56065 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a fastcc + sret bug. If fastcc and sret, callee doesn't need to pop the hidden struct ptr; Re-enable fastcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56061 91177308-0d34-0410-b5e6-96231b3b80d8

Handle new intrinsics with vector arguments.
Patch by Paul Redmond.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56059 91177308-0d34-0410-b5e6-96231b3b80d8

Add a break statement that I accidentally deleted when
I shuffled the fast-isel command-line options around. This fixes
a bunch of fast-isel failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56057 91177308-0d34-0410-b5e6-96231b3b80d8

fix white spaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56056 91177308-0d34-0410-b5e6-96231b3b80d8

Fix name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56055 91177308-0d34-0410-b5e6-96231b3b80d8

Add trampoline support for the new FastCC calling
convention (not related to recent Ada testsuite
failures).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56054 91177308-0d34-0410-b5e6-96231b3b80d8

Turn off the new FastCC for the moment.  It causes
a slew of Ada testsuite failures on x86-32 linux.
Seems to be related to the use of float.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56053 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary bit-wise AND from the limited precision work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56049 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 col violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56048 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a warning about comparing signed and unsigned values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56040 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56037 91177308-0d34-0410-b5e6-96231b3b80d8

Check that both operands are f32 before attempting to lower.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56036 91177308-0d34-0410-b5e6-96231b3b80d8

Implement "visitPow". This is mainly used to see if we have a pow() call of this
form:

          powf(10.0f, x);

If this is the case, and also we want limited precision floating-point
calculations, then lower to do the limited-precision stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56035 91177308-0d34-0410-b5e6-96231b3b80d8

A few more places where FPOW is being ignored.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56032 91177308-0d34-0410-b5e6-96231b3b80d8

Change -fast-isel-no-abort to -fast-isel-abort, which now defaults
to being off by default. Also, add assertion checks to check that
the various fast-isel-related command-line options are only used
when -fast-isel itself is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56029 91177308-0d34-0410-b5e6-96231b3b80d8

Legalizer was missing code that expand fpow to a libcall.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56028 91177308-0d34-0410-b5e6-96231b3b80d8

Adding 6-, 12-, and 18-bit limited-precision floating-point support for exp2
function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56025 91177308-0d34-0410-b5e6-96231b3b80d8

Move the uglier parts of deciding not to emit a
UsedDirective for some symbols in llvm.used into
Darwin-specific code.  I've decided LessPrivateGlobal
is potentially a useful abstraction and left it in
the target-independent area, with improved comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56024 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for 6-, 12-, and 18-bit limited precision calculations of exp for
floating-point numbers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56023 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a minor wording ambiguity in the Developer Policy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56022 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new option, -fast-isel-verbose, that can be used with
-fast-isel-no-abort to get a dump of all unhandled instructions,
without an abort.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56021 91177308-0d34-0410-b5e6-96231b3b80d8

Clear preference when it no longer makes sense.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56019 91177308-0d34-0410-b5e6-96231b3b80d8

Remove.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56018 91177308-0d34-0410-b5e6-96231b3b80d8

if loop induction variable is always sign or zero extended then
extend the type of induction variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56017 91177308-0d34-0410-b5e6-96231b3b80d8

Add assertion check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56016 91177308-0d34-0410-b5e6-96231b3b80d8

fix overflow check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56011 91177308-0d34-0410-b5e6-96231b3b80d8

Clean this up, based on Evan's suggestions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56009 91177308-0d34-0410-b5e6-96231b3b80d8

- Add support for 6-, 12-, and 18-bit limited precision floating-point "log"
  values.
- Refactored some of the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56008 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2757. Ignore liveinterval register allocation preference if the preference register is not in the right register class. This can happen due to sub-register coalescing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56006 91177308-0d34-0410-b5e6-96231b3b80d8

Make safer variant of alias resolution routine to be default

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56005 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify this some more.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56003 91177308-0d34-0410-b5e6-96231b3b80d8

Resolve aliases, when possible

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56001 91177308-0d34-0410-b5e6-96231b3b80d8

Add limited precision floating-point conversions of log10 for 6- and 18-bit
precisions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56000 91177308-0d34-0410-b5e6-96231b3b80d8

Mark IMPLICIT_DEF as being rematerializable and cheap-as-a-move.
It's already special-cased and treated as rematerializable within
LiveIntervals; this allows it to be handled by other passes
such as TwoAddressInstrctionPass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55999 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55998 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect linker behaviour: we shouldn't resolve weak aliases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55997 91177308-0d34-0410-b5e6-96231b3b80d8

Replace explicit pointer-size constants to TargetData query.
No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55996 91177308-0d34-0410-b5e6-96231b3b80d8

Make SimplifyDemandedVectorElts simplify vectors with multiple
users, and teach it about shufflevector instructions.

Also, fix a subtle bug in SimplifyDemandedVectorElts'
insertelement code.

This is a patch that was originally written by Eli Friedman,
with some fixes and cleanup by me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55995 91177308-0d34-0410-b5e6-96231b3b80d8