this passes now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102301 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR6921: globals were not getting correctly rounded up to their
preferred alignment unless they were common or some other special
case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102300 91177308-0d34-0410-b5e6-96231b3b80d8

silence a warning, patch by "mike".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102297 91177308-0d34-0410-b5e6-96231b3b80d8

remove #if 1's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102296 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an autoconf bug: The second argument to AC_LANG_PROGRAM is code that
gets placed inside a main function, and should not itself be a main
function. This is silently hidden in GCC-hosted builds because the
inner main looks like a nested function declaration, which GCC supports.
In builds with compilers which do not support nested functions (by default),
this was causing an error, which caused these autoconf checks to fail,
leaving their options disabled.

This fixes test/Feature/load_module.ll on x86_64-unknown-linux-gnu
llvm-gcc selfhost builds, among other things.

This also includes a regenerated configure, as the diff is small and telling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102288 91177308-0d34-0410-b5e6-96231b3b80d8

PHI nodes also do not necessarily return trap with a trap operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102278 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102276 91177308-0d34-0410-b5e6-96231b3b80d8

Use provided llvm-defined variables instead of twisting user-controlled knobs.
No visible change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102240 91177308-0d34-0410-b5e6-96231b3b80d8

Don't touch CXXFLAGS. If it's set in the environment then sub-makes will see
the variable that we set as though it came from the user. Fixes PR6835.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102239 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid promoting a i16 node if it would eliminate a (store (op (load))) opportunity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102237 91177308-0d34-0410-b5e6-96231b3b80d8

When a load operand is promoted to an extload, replace other uses with uses of extload result truncated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102236 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize LSR's OptimizeMax to handle the new kinds of max expressions
that indvars may use, now that indvars is recognizing le and ge loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102235 91177308-0d34-0410-b5e6-96231b3b80d8

ScalarEvolution support for <= and >= loops.

Also, generalize ScalarEvolutions's min and max recognition to handle
some new forms of min and max that this change makes more common.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102234 91177308-0d34-0410-b5e6-96231b3b80d8

Use SimplifyICmpOperands in isKnownPredicate too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102233 91177308-0d34-0410-b5e6-96231b3b80d8

Update isImpliedCond to use the new SimplifyICmpOperands utility.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102232 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new utility function SimplifyICmpOperands. Much of this code is
refactored out of ScalarEvolution::isImpliedCond, which will be updated
to use this new utility routine soon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102229 91177308-0d34-0410-b5e6-96231b3b80d8

Apply a fix for a vector setcc dagcombine from Jan Sjodin. No
testcase yet, as the testcase now fails downstream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102228 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed edis to tokenize instructions with no
operands correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102227 91177308-0d34-0410-b5e6-96231b3b80d8

cross-build Makefile needs to unset CFLAGS/CXXFLAGS when building the build-side utilities since the flags will be for the cross-compiler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102225 91177308-0d34-0410-b5e6-96231b3b80d8

no longer xfail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102220 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes to edis that mark x86 call targets as
memory operands rather than immediate operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102217 91177308-0d34-0410-b5e6-96231b3b80d8

Per Chris, fuse four trivial tests using grep (r102199) into one that uses FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102216 91177308-0d34-0410-b5e6-96231b3b80d8

enable my inliner change: add newly devirtualized call sites to
the worklist, making them inline candidates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102213 91177308-0d34-0410-b5e6-96231b3b80d8

Change TargetData's algorithm for computing defualt vector type
alignment to match what's used in clang and GCC for __alignof, rather
than trying to guess what Legalize is going to be doing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102206 91177308-0d34-0410-b5e6-96231b3b80d8

Code refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102202 91177308-0d34-0410-b5e6-96231b3b80d8

Add some missing x86 patterns for movdq2q.  Fixes two (LLVM-)GCC DejaGNU testcases.  Radar 6881029.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102199 91177308-0d34-0410-b5e6-96231b3b80d8

fix some failures my callgraph dump format change broke.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102197 91177308-0d34-0410-b5e6-96231b3b80d8

switch InlineInfo.DevirtualizedCalls's list to be of WeakVH.
This fixes a bug where calls inlined into an invoke would get
changed into an invoke but the array would keep pointing to
the (now dead) call.  The improved inliner behavior is still
disabled for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102196 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for the bug that required a patch to be reverted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102195 91177308-0d34-0410-b5e6-96231b3b80d8

fix callgraph dump to not print 0x0x1234 for nodes.
Add the instruction pointer value for debuggability.
We now get dump output that looks like this:

Call graph node for function: 'f1'<<0x1017086b0>>  #uses=1
  CS<0x1017046f8> calls external node

Call graph node for function: '_ZNSt6vectorIdSaIdEEC1EmRKdRKS0_'<<0x1017086f0>>  #uses=1
  CS<0x0> calls external node

Call graph node for function: 'f4'<<0x1017087a0>>  #uses=1
  CS<0x101708c88> calls function 'f3'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102194 91177308-0d34-0410-b5e6-96231b3b80d8

Fix X86ISD::CMP i16 to i32 promotion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102192 91177308-0d34-0410-b5e6-96231b3b80d8

Move FastISel's HandlePHINodesInSuccessorBlocks call down into FastISel
itself too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102176 91177308-0d34-0410-b5e6-96231b3b80d8

Switch getelementptr inbounds and sdiv exact from undef to trap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102175 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed EDOperand to use the operand type, not the
flags, to determine whether or not the operand is
a memory operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102158 91177308-0d34-0410-b5e6-96231b3b80d8

Fix LSR to tolerate cases where ScalarEvolution initially
misses an opportunity to fold add operands, but folds them
after LSR has separated them out. This fixes rdar://7886751.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102157 91177308-0d34-0410-b5e6-96231b3b80d8

When it doesn't matter whether zero or sign extension is used,
use ScalarEvolutions "any" extend function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102156 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 102135, 102129, 102127, 102106, 102104, 102102, 102012, 102004,
because 102004 causes codegen to emit invalid assembly on at least
x86_64-unknown-gnu-linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102155 91177308-0d34-0410-b5e6-96231b3b80d8

reject invalid comma stuff with a message.  We reject the case in
PR6888 with:

$ llvm-as t.ll
llvm-as: t.ll:2:29: error: expected metadata or 'align'
 store <3 x i32> %x, i32 1, i32 1>, <3 x i32>* %p
                            ^

instead of:

$ llvm-as t.ll
llvm-as:

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102154 91177308-0d34-0410-b5e6-96231b3b80d8

disable my previous inliner patch, it appears to be busting self-host.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102153 91177308-0d34-0410-b5e6-96231b3b80d8

The inliner was choosing to not consider call sites
that appear in the SCC as a result of inlining as candidates
for inlining.  Change this so that it *does* consider call
sites that change from being indirect to being direct as a
result of inlining.  This allows it to completely
"devirtualize" the testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102146 91177308-0d34-0410-b5e6-96231b3b80d8

Update ARM DAGtoDAG for matching UBFX instruction for unsigned bitfield
extraction. This fixes PR5998.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102144 91177308-0d34-0410-b5e6-96231b3b80d8

Add an initial description of a new concept: trap values, and change
the definition of the nsw and nuw flags to make use of it.

nsw was introduced to help optimizers answer yes to the following:

  // Can we change i from i32 to i64 to eliminate the cast inside the loop?
  for (int i = 0; i < n; ++i) A[i] *= 0.1;

  // Can we assume that this loop will eventually terminate?
  for (int i = 0; i <= n; ++i) A[i] *= 0.1;

In its current form, it isn't truly sufficient for either.

In the first case, if the increment overflows, it'll still have some
valid i32 value; sign-extending it will produce a value which is 33
homogeneous sign bits trailed by 31 independent undef bits. If i is
promoted to i64, it won't have those same values when it reaches that
point. (The compiler could recover here by reasoning about how i is
used by the load, but that's a lot more complicated and isn't always
possible.)

In the second case, there is no value for i which will be greater than
n, so having the increment return undef on overflow doesn't help.

Trap values are a formalization of some existing concepts that we have
about LLVM IR, and give the optimizers a better basis for answering yes
to both questions above.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102140 91177308-0d34-0410-b5e6-96231b3b80d8

refactor the interface to InlineFunction so that most of the in/out
arguments are handled with a new InlineFunctionInfo class.  This
makes it easier to extend InlineFunction to return more info in the
future.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102137 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the test for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102135 91177308-0d34-0410-b5e6-96231b3b80d8

add macruby back

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102132 91177308-0d34-0410-b5e6-96231b3b80d8

when inlining something like this:

define void @f3(void (i8*)* %__f) ssp {
entry:
  call void %__f(i8* undef)
  unreachable
}

define void @f4(i8* %this) ssp align 2 {
entry:
  call void @f3(void (i8*)* @f2) ssp
  ret void
}

The inliner is turning the indirect call to %__f into a direct
call to F2.  Make the call graph more precise when this happens.

The inliner doesn't revisit call sites introduced by inlining,
so there isn't an easy way to test for this, but a more precise
callgraph is a good thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102131 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102129 91177308-0d34-0410-b5e6-96231b3b80d8

Sink SelectionDAGBuilder's HandlePHINodesInSuccessorBlocks down
into SelectionDAGBuilder itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102128 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust debug range offsets for isWeakForLinker() functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102127 91177308-0d34-0410-b5e6-96231b3b80d8

Add a blurb about the new LSR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102126 91177308-0d34-0410-b5e6-96231b3b80d8

The -condprop pass has been deleted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102125 91177308-0d34-0410-b5e6-96231b3b80d8

The post-RA scheduler is currently off by default on x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102124 91177308-0d34-0410-b5e6-96231b3b80d8

Move HandlePHINodesInSuccessorBlocks functions out of SelectionDAGISel
and into SelectionDAGBuilder and FastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102123 91177308-0d34-0410-b5e6-96231b3b80d8

Add some verbiage about Clang to the 2.7 release notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102121 91177308-0d34-0410-b5e6-96231b3b80d8

add a DEBUG call so that -debug lists when CGSCCPM iterates.

Fix RefreshCallGraph to use CGN->replaceCallEdge instead of hand
rolling its own loop.  replaceCallEdge properly maintains the
reference counts of the nodes, fixing a crash exposed by the
iterative callgraph stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102120 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate dead #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102119 91177308-0d34-0410-b5e6-96231b3b80d8

- It's not safe to promote rotates (at least not trivially).
- Some code refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102111 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102110 91177308-0d34-0410-b5e6-96231b3b80d8

Move PHINodesToUpdate out of SelectionDAGBuilder and into
FunctionLoweringInfo, as it isn't SelectionDAG-specific. This isn't
completely natural, as PHI node state is not per-function but rather
per-basic-block, however there's currently no other convenient
per-basic-block state to group it with.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102109 91177308-0d34-0410-b5e6-96231b3b80d8

Rename InsnAfterLabelMap and InsnBeforeLabelMap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102106 91177308-0d34-0410-b5e6-96231b3b80d8

Keep track of MCSymbol used to mark beginning of a function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102104 91177308-0d34-0410-b5e6-96231b3b80d8

Add mention of ARM atomic builtin support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102103 91177308-0d34-0410-b5e6-96231b3b80d8

At this point Start and End are not null.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102102 91177308-0d34-0410-b5e6-96231b3b80d8

Update ARM frame index scavenging description

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102101 91177308-0d34-0410-b5e6-96231b3b80d8

remove dupe

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102100 91177308-0d34-0410-b5e6-96231b3b80d8

reword my incoherent babble into something that makes sense
from Wesley Peck!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102096 91177308-0d34-0410-b5e6-96231b3b80d8

two more projects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102095 91177308-0d34-0410-b5e6-96231b3b80d8

add some notes from Matthieu Moy!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102093 91177308-0d34-0410-b5e6-96231b3b80d8

improve tblgen doc, pointed out by Zhongxing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102072 91177308-0d34-0410-b5e6-96231b3b80d8

Fixing a typo in the "Clang project" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102064 91177308-0d34-0410-b5e6-96231b3b80d8

Small fix to the "Other Improvements and New Features" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102063 91177308-0d34-0410-b5e6-96231b3b80d8

Fixing small typos in the "New Useful APIs" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102062 91177308-0d34-0410-b5e6-96231b3b80d8

Fixing small typos in the "ARM Target Improvements" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102061 91177308-0d34-0410-b5e6-96231b3b80d8

Rewording the "generated debug information" paragraph in the "Interpreter and JIT Improvements" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102060 91177308-0d34-0410-b5e6-96231b3b80d8

Reworking the wording of the "target data string" paragraph in the "Optimizer Improvments" section of the release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102059 91177308-0d34-0410-b5e6-96231b3b80d8

<pre> seems to close <p>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102057 91177308-0d34-0410-b5e6-96231b3b80d8

the validator coughed up some problems, most of them fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102056 91177308-0d34-0410-b5e6-96231b3b80d8

shorten link

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102055 91177308-0d34-0410-b5e6-96231b3b80d8

static analyzer has its own section

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102054 91177308-0d34-0410-b5e6-96231b3b80d8

add mc and static analyzer blurbs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102053 91177308-0d34-0410-b5e6-96231b3b80d8

another checkpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102052 91177308-0d34-0410-b5e6-96231b3b80d8

various updates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102051 91177308-0d34-0410-b5e6-96231b3b80d8

fix file header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102049 91177308-0d34-0410-b5e6-96231b3b80d8

Don't attempt to analyze values which are obviously undef. This fixes some
assertion failures in extreme cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102042 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102041 91177308-0d34-0410-b5e6-96231b3b80d8

Run LiveVariables instead of computing liveness locally in -regalloc=fast.

This actually makes everything slower, but the plan is to have isel add <kill>
flags the way it is already adding <dead> flags. Then LiveVariables can be
removed again.

When ignoring the time spent in LiveVariables, -regalloc=fast is now twice as
fast as -regalloc=local.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102034 91177308-0d34-0410-b5e6-96231b3b80d8

Do not try to optimize a copy that has already been marked for deletion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102027 91177308-0d34-0410-b5e6-96231b3b80d8

Add command line option to disable debug info printing in .s file. This option does not impact debug info generation and preservation through earlier compile starges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102012 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a performance problem with the new SSAUpdater.  This showed up in the
GCCAS time for MultiSource/Benchmarks/ASCI_Purple/SMG2000.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102009 91177308-0d34-0410-b5e6-96231b3b80d8

Modified some assert() msg strings; no other functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102008 91177308-0d34-0410-b5e6-96231b3b80d8

Add fast register allocator, enabled with -regalloc=fast.

So far this is just a clone of -regalloc=local that has been lobotomized to run
25% faster. It drops the least-recently-used calculations, and is just plain
stupid when it runs out of registers.

The plan is to make this go even faster for -O0 by taking advantage of the short
live intervals in unoptimized code. It should not be necessary to calculate
liveness when most virtual registers are killed 2-3 instructions after they are
born.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102006 91177308-0d34-0410-b5e6-96231b3b80d8

Identify when a lexical scope is split in to multiple instruction ranges. Emit such ranges using DW_AT_ranges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102004 91177308-0d34-0410-b5e6-96231b3b80d8

Make ScalarEvolution::getConstant support pointer types, for consistency
with ScalarEvolution's overall approach to pointer types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102003 91177308-0d34-0410-b5e6-96231b3b80d8

Dragonegg will be released along side llvm-2.7.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102001 91177308-0d34-0410-b5e6-96231b3b80d8

final hacking for tonight, still more to go.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101995 91177308-0d34-0410-b5e6-96231b3b80d8

continue the process of detangling this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101992 91177308-0d34-0410-b5e6-96231b3b80d8

rough pass moving stuff into relevant sections, still much
editing to do.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101987 91177308-0d34-0410-b5e6-96231b3b80d8

remove ldc, rubinious, macruby, icedtea, llvm-lua, which
don't have updates for 2.7.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101985 91177308-0d34-0410-b5e6-96231b3b80d8

Implement -disable-non-leaf-fp-elim which disable frame pointer elimination
optimization for non-leaf functions. This will be hooked up to gcc's
-momit-leaf-frame-pointer option. rdar://7886181

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101984 91177308-0d34-0410-b5e6-96231b3b80d8

isel (i32 anyext i16) as insert_subreg when 16-bit ops are being promoted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101979 91177308-0d34-0410-b5e6-96231b3b80d8