minor edits

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51405 91177308-0d34-0410-b5e6-96231b3b80d8

Use 'continue' to reduce nesting in this loop.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51399 91177308-0d34-0410-b5e6-96231b3b80d8

Add a Type::isSingleValueType method. This will be used by code
that currently uses Type::isFirstClassType and depends on it
returning false for struct or array types.

This commit doesn't change the behavior of Type::isFirstClassType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51396 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2343. An *interesting* coalescer bug.

BB1:
  vr1025 = copy vr1024
  ..
BB2:
  vr1024 = op
         = op vr1025
  <loop eventually branch back to BB1>

Even though vr1025 is copied from vr1024, it's not safe to coalesced them since live range of vr1025 intersects the def of vr1024. This happens when vr1025 is assigned the value of the previous iteration of vr1024 in the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51394 91177308-0d34-0410-b5e6-96231b3b80d8

Follow-up to the reverting of r51218. This puts the checks out-of-line. Because
they aren't in the header file, systems with a <string> header file that isn't
64-bit clean shouldn't warn if #including Path.h and specifying
-Wshorten-64-to-32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51393 91177308-0d34-0410-b5e6-96231b3b80d8

update a comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51390 91177308-0d34-0410-b5e6-96231b3b80d8

adapt to new API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51387 91177308-0d34-0410-b5e6-96231b3b80d8

Rename -no-lazy to -disable-lazy-compilation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51386 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple issues with the JIT and multiple modules:

1. The "JITState" object creates a PassManager with the ModuleProvider that the
   jit is created with.  If the ModuleProvider is removed and deleted, the
   PassManager is invalid.

2. The Global maps in the JIT were not invalidated with a ModuleProvider was
   removed.  This could lead to a case where the Module would be freed, and a
   new Module with Globals at the same addresses could return invalid results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51384 91177308-0d34-0410-b5e6-96231b3b80d8

resurrect lost tests by renaming them to not end with .tr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51375 91177308-0d34-0410-b5e6-96231b3b80d8

suppress gcc3.4.6's <no value returned> warnings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51372 91177308-0d34-0410-b5e6-96231b3b80d8

When LSR is replacing an instruction, call
ScalarEvolution::deleteValueFromRecords on it before doing the
replaceAllUsesWith, because ScalarEvolution looks at the instruction's
users to find SCEV references to the instruction's SCEV object in its
internal maps.

Move all of LSR's loop-related state clearing after processing the loop
and before cleaning up dead PHI nodes. This eliminates all of LSR's SCEV
references just before the calls to ScalarEvolution::deleteValueFromRecords
so that when ScalarEvolution drops its own SCEV references, the reference
counts will reach zero and the SCEVs will be deleted immediately.

These changes fix some compiler aborts involving ScalarEvolution holding
onto and reusing SCEV objects for instructions that have been deleted.
No regression test unfortunately; because the symptoms were due to
dangling pointers, reduced testcases ended up being fairly arbitrary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51359 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting r51218 because of breakage on PPC32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51358 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate questionable syntax for stdin redirection. This probably also speeds things up a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51357 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2346 by marking vaarg as volatile so that licm doesn't try to
hoist them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51356 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, commit the version of this test that actually works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51351 91177308-0d34-0410-b5e6-96231b3b80d8

Port SelectionDAG's ComputeNumSignBits-using code to instcombine,
now that instcombine also has ComputeNumSignBits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51350 91177308-0d34-0410-b5e6-96231b3b80d8

sabre brings to my attention that the 'tr' suffix is also obsolete

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51349 91177308-0d34-0410-b5e6-96231b3b80d8

Port the fix for the select operator from instcombine's
ComputeNumSignBits to SelectionDAG's ComputeNumSignBits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51348 91177308-0d34-0410-b5e6-96231b3b80d8

Code simplification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51345 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify that and/or/xor/sdiv etc all allow vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51335 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify copyright transfer when you don't own the copyright of
the contributed code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51329 91177308-0d34-0410-b5e6-96231b3b80d8

Rename the last test with .llx extension to .ll, resolve duplicate test by renaming to isnan2. Now that no test has llx ending there is no need to search for them from dg.exp too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51328 91177308-0d34-0410-b5e6-96231b3b80d8

movsd and movq do not require 16-byte alignment. This fixes vec_set-5.ll on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51327 91177308-0d34-0410-b5e6-96231b3b80d8

"The two generated files in the AsmParser.vcproj are pointing to the
wrong directory, this is the patch to fix it (also attached)."

Patch by 'OvermindDL1'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51323 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify the testing documentation a bit and solve two HTML validator warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51313 91177308-0d34-0410-b5e6-96231b3b80d8

More local spiller complexity!
If local spiller optimization turns some instruction into an identity copy, it will be removed. If the output register happens to be dead (and source is obviously killed), transfer the kill / dead information to last use / def in the same MBB.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51306 91177308-0d34-0410-b5e6-96231b3b80d8

Don't spill dead def.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51305 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51303 91177308-0d34-0410-b5e6-96231b3b80d8

Teach instcombine 4 new xforms:
  (add (sext x), cst) --> (sext (add x, cst'))
  (add (sext x), (sext y)) --> (sext (add int x, y))
  (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
  (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))

This generally reduces conversions.  For example MiBench/telecomm-gsm
gets these simplifications:

HACK2:  %tmp67.i142.i.i = sext i16 %tmp6.i141.i.i to i32 ; <i32> [#uses=1]
%tmp23.i139.i.i = sext i16 %tmp2.i138.i.i to i32 ; <i32> [#uses=1]
%tmp8.i143.i.i = add i32 %tmp67.i142.i.i, %tmp23.i139.i.i ; <i32> [#uses=3]
HACK2:  %tmp67.i121.i.i = sext i16 %tmp6.i120.i.i to i32 ; <i32> [#uses=1]
%tmp23.i118.i.i = sext i16 %tmp2.i117.i.i to i32 ; <i32> [#uses=1]
%tmp8.i122.i.i = add i32 %tmp67.i121.i.i, %tmp23.i118.i.i ; <i32> [#uses=3]
HACK2:  %tmp67.i.i190.i = sext i16 %tmp6.i.i189.i to i32 ; <i32> [#uses=1]
%tmp23.i.i187.i = sext i16 %tmp2.i.i186.i to i32 ; <i32> [#uses=1]
%tmp8.i.i191.i = add i32 %tmp67.i.i190.i, %tmp23.i.i187.i ; <i32> [#uses=3]
HACK2:  %tmp67.i173.i.i.i = sext i16 %tmp6.i172.i.i.i to i32 ; <i32> [#uses=1]
%tmp23.i170.i.i.i = sext i16 %tmp2.i169.i.i.i to i32 ; <i32> [#uses=1]
%tmp8.i174.i.i.i = add i32 %tmp67.i173.i.i.i, %tmp23.i170.i.i.i ; <i32> [#uses=3]
HACK2:  %tmp67.i152.i.i.i = sext i16 %tmp6.i151.i.i.i to i32 ; <i32> [#uses=1]
%tmp23.i149.i.i.i = sext i16 %tmp2.i148.i.i.i to i32 ; <i32> [#uses=1]
%tmp8.i153.i.i.i = add i32 %tmp67.i152.i.i.i, %tmp23.i149.i.i.i ; <i32> [#uses=3]
HACK2:  %tmp67.i.i.i.i = sext i16 %tmp6.i.i.i.i to i32 ; <i32> [#uses=1]
%tmp23.i.i5.i.i = sext i16 %tmp2.i.i.i.i to i32 ; <i32> [#uses=1]
%tmp8.i.i7.i.i = add i32 %tmp67.i.i.i.i, %tmp23.i.i5.i.i ; <i32> [#uses=3]

This also fixes a bug in ComputeNumSignBits handling select and
makes it more aggressive with and/or.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51302 91177308-0d34-0410-b5e6-96231b3b80d8

fix two issues Neil noticed, thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51296 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bool to isEdgeFeasible that tells it whether to treat unknown
value as undef or untracked.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51295 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ExecutionEngine's constant code to work properly when structs and arrays
will become first-class types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51293 91177308-0d34-0410-b5e6-96231b3b80d8

Refine the fix in r51169 to only apply when the operand val being
replaced is a PHI. This prevents it from inserting uses before defs
in the case that it isn't a PHI and it depends on other instructions
later in the block. This fixes the 447.dealII regression on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51292 91177308-0d34-0410-b5e6-96231b3b80d8

runOnMachineFunction should set IsPIC because relocation model may have been changed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51291 91177308-0d34-0410-b5e6-96231b3b80d8

Make AssociativeOpt static.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51290 91177308-0d34-0410-b5e6-96231b3b80d8

Run vortex-bug as x86-64, which is what the original bug was triggered on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51289 91177308-0d34-0410-b5e6-96231b3b80d8

Silence a warning about "*/" outside a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51287 91177308-0d34-0410-b5e6-96231b3b80d8

Do not erase induction variable increment if it is used outside the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51280 91177308-0d34-0410-b5e6-96231b3b80d8

Add a ComputeNumSignBits function for use by instcombine, based on the
code in SelectionDAG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51279 91177308-0d34-0410-b5e6-96231b3b80d8

Handle quoted names when constructing $stub's,
$non_lazy_ptr's and $lazy_ptr's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51277 91177308-0d34-0410-b5e6-96231b3b80d8

switch to Type::getFPMantissaWidth instead of reinventing it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51275 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51274 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanups, teach instcombine that sitofp/uitofp cannot
produce a negative zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51272 91177308-0d34-0410-b5e6-96231b3b80d8

convert fptosi(sitofp x) -> x  if the fp value has enough bits in its mantissa
to accurately represent the integer.  This triggers 9 times in 471.omnetpp,
though 8 of those seem to be inlined from the same place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51271 91177308-0d34-0410-b5e6-96231b3b80d8

Fold FP comparisons where one operand is converted from an integer
type and the other operand is a constant into integer comparisons.
This happens surprisingly frequently (e.g. 10 times in 471.omnetpp),
which are things like this:

%tmp8283 = sitofp i32 %tmp82 to double
%tmp1013 = fcmp ult double %tmp8283, 0.0

Clearly comparing tmp82 against i32 0 is cheaper here.

this also triggers 8 times in gobmk, including this one:

%tmp375376 = sitofp i32 %tmp375 to double
%tmp377 = fcmp ogt double %tmp375376, 8.150000e+01

which is comparing an integer against 81.5 :).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51268 91177308-0d34-0410-b5e6-96231b3b80d8

Remove warnings about unused parameters and shadowed variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51266 91177308-0d34-0410-b5e6-96231b3b80d8

remove debug output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51264 91177308-0d34-0410-b5e6-96231b3b80d8

be more aggressive about transforming add -> or when the operands have no
intersecting bits.  This triggers all over the place, for example in lencode,
with adds of stuff like:

%tmp580 = mul i32 %tmp579, 2
%tmp582 = and i32 %b8, 1
and

%tmp28 = shl i32 %abs.i, 1
%sign.0 = select i1 %tmp23, i32 1, i32 0
and
%tmp344 = shl i32 %tmp343, 2
%tmp346 = and i32 %tmp96, 3

etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51263 91177308-0d34-0410-b5e6-96231b3b80d8

Check that always_inline functions are inlined
whether or not -funit-at-a-time is used (C++ uses
it, C doesn't) - it was working before only when
not doing unit-at-a-time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51258 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2341 - when the length is 4 use an i32 not
an i16!  Cleaned up trailing whitespace while there.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51240 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting accidental commit of generated files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51239 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a duplicative binding. Patch by Mahadevan R.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51238 91177308-0d34-0410-b5e6-96231b3b80d8

Update VS project files (VS 2005) so that VS builds with TOT LLVM/clang.

Patch by Dmitri Makarov!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51237 91177308-0d34-0410-b5e6-96231b3b80d8

Grammar fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51236 91177308-0d34-0410-b5e6-96231b3b80d8

Delete module.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51235 91177308-0d34-0410-b5e6-96231b3b80d8

Convert non-ASCII apostrophes into ASCII apostrophes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51234 91177308-0d34-0410-b5e6-96231b3b80d8

Delete the Module object.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51233 91177308-0d34-0410-b5e6-96231b3b80d8

s/insure/ensure/

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51232 91177308-0d34-0410-b5e6-96231b3b80d8

ignore Output dir

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51231 91177308-0d34-0410-b5e6-96231b3b80d8

Teach GVN to not assert on vector comparisons

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51230 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a backwards check in the JIT symbol table code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51229 91177308-0d34-0410-b5e6-96231b3b80d8

add missing #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51227 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2339

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51226 91177308-0d34-0410-b5e6-96231b3b80d8

remove empty file?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51225 91177308-0d34-0410-b5e6-96231b3b80d8

Revert constant-folding change that will miscompile in some cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51223 91177308-0d34-0410-b5e6-96231b3b80d8

trip count computation deficiency

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51222 91177308-0d34-0410-b5e6-96231b3b80d8

On Darwin, the string header file isn't 64-bit clean. The use of
"-Wshorten-64-to-32 -Werror" will cause a failure when compiling this complex
program:

#include <string>

class Path {
  mutable std::string path;
public:
  bool operator == (const Path &that) {
    return path == that.path;
  }
};

Using strcmp gets us past this annoying error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51218 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51217 91177308-0d34-0410-b5e6-96231b3b80d8

Constant fold inttoptr and ptrtoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51216 91177308-0d34-0410-b5e6-96231b3b80d8

Move isTrueWhenEqual to ICmpInst.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51215 91177308-0d34-0410-b5e6-96231b3b80d8

Less conservative verison of previous patch,
suggested by Duncan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51211 91177308-0d34-0410-b5e6-96231b3b80d8

Add CommonLinkage, in a way that preserves the
behavior on old .bc files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51210 91177308-0d34-0410-b5e6-96231b3b80d8

Add CommonLinkage to lto (treated same as weak AFAICT)
and llvm-nm (prints as C).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51209 91177308-0d34-0410-b5e6-96231b3b80d8

Add CommonLinkage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51208 91177308-0d34-0410-b5e6-96231b3b80d8

Weak functions not declared non-throwing might be
replaced at linktime with a body that throws, even
if the body in this file does not.  Make PruneEH
be more conservative in this case.
g++.dg/eh/weak1.C

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51207 91177308-0d34-0410-b5e6-96231b3b80d8

Add functions to enable adding a single attribute to a function and
its associated call site.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51204 91177308-0d34-0410-b5e6-96231b3b80d8

Record weak external linkage in a case where we were
missing it. gcc.dg/darwin-weakimport-2.c.
Handle common and weak differently for darwin ppc32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51201 91177308-0d34-0410-b5e6-96231b3b80d8

API change for {BinaryOperator|CmpInst|CastInst}::create*() --> Create. Legacy interfaces will be in place for some time.  (Merge from use-diet branch.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51200 91177308-0d34-0410-b5e6-96231b3b80d8

If the result of a BIT_CONVERT is a v1* vector, it doesn't mean its source is a v1* vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51192 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51191 91177308-0d34-0410-b5e6-96231b3b80d8

Bill pointed out that system headers should be
included after local headers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51187 91177308-0d34-0410-b5e6-96231b3b80d8

Silence the compiler warning differently.  The
original method caused gcc-4.2 to complain.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51186 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable tail duplication pass (now with default threshold down to 1 instruction).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51184 91177308-0d34-0410-b5e6-96231b3b80d8

Do not dup malloc, vector instructions, etc. Throttle the default theshold way down.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51183 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ADCE's ability to delete loops.  This ability is now implemented in a
safer manner by loop deletion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51182 91177308-0d34-0410-b5e6-96231b3b80d8

Move this test from ADCE to loop deletion, where it is more appropriate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51181 91177308-0d34-0410-b5e6-96231b3b80d8

Use loop deletion instead of ADCE in these tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51180 91177308-0d34-0410-b5e6-96231b3b80d8

Clean ups for loop deletion based on Chris' feedback.

Also, use SCEV to determine the trip count of the loop, which is more powerful
and accurate that Loop::getTripCount.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51179 91177308-0d34-0410-b5e6-96231b3b80d8

Use loop deletion instead of ADCE for removing loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51178 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test.  It was testing broken behavior in that it required ADCE to eliminate
a potentially infinite loop, which is undesirable.  Instead, test the LICM behavior
that we're really interested in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51177 91177308-0d34-0410-b5e6-96231b3b80d8

implement PR2328.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51176 91177308-0d34-0410-b5e6-96231b3b80d8

Use common where we mean common, not weak.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51173 91177308-0d34-0410-b5e6-96231b3b80d8

Treat common as distinct from weak global on Darwin x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51172 91177308-0d34-0410-b5e6-96231b3b80d8

Revert the change from r51157 in
test/Verifier/2002-11-05-GetelementptrPointers.ll, which was incorrect.
Instead, fix getIndexedType to not follow pointer types, as
PointerType is a subclass of CompositeType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51171 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments based on Duncan's feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51170 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in LoopStrengthReduce that caused it to emit IR with
use-before-def. The problem comes up in code with multiple PHIs where
one PHI is being rewritten in terms of the other, but the other needs
to be casted first. LLVM rules requre the cast instruction to be
inserted after any PHI instructions, but when instructions were
inserted to replace the second PHI value with a function of the first,
they were ended up going before the cast instruction. Avoid this
problem by remembering the location of the cast instruction, when one
is needed, and inserting the expansion of the new value after it.

This fixes a bug that surfaced in 255.vortex on x86-64 when
instcombine was removed from the middle of the loop optimization
passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51169 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the S92 code, which really has nothing to do
with what the test is testing; makes it pass again on ppc32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51167 91177308-0d34-0410-b5e6-96231b3b80d8

Evan has implemented this on ppc, so run the test there.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51166 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typos and comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51165 91177308-0d34-0410-b5e6-96231b3b80d8