Fix liveness calculation when splitting critical edges during PHI elimination.

- Edges are split before any phis are eliminated, so the code is SSA.

- Create a proper IR BasicBlock for the split edges.

- LiveVariables::addNewBlock now has same syntax as
  MachineDominatorTree::addNewBlock. Algorithm calculates predecessor live-out
  set rather than successor live-in set.

This feature still causes some miscompilations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86867 91177308-0d34-0410-b5e6-96231b3b80d8

Reenable StackTracke.cpp test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86861 91177308-0d34-0410-b5e6-96231b3b80d8

Add SetDebugLocation() variant to
add debug info location to an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86859 91177308-0d34-0410-b5e6-96231b3b80d8

Add TargetLowering::isLegalICmpImmediate. It tells LSR what immediate can be folded into target icmp instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86858 91177308-0d34-0410-b5e6-96231b3b80d8

Do jump table adjustment before constant island allocation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86857 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation level.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86856 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86855 91177308-0d34-0410-b5e6-96231b3b80d8

Prefix MBB numbers with "BB#" in debug output to make it clear what
the numbers mean.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86854 91177308-0d34-0410-b5e6-96231b3b80d8

Minor code simplification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86853 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a copy+pasto in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86852 91177308-0d34-0410-b5e6-96231b3b80d8

Set isBarrier = 1 on return instructions, as they are control barriers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86851 91177308-0d34-0410-b5e6-96231b3b80d8

Use a tab in INT3's asm string, for consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86850 91177308-0d34-0410-b5e6-96231b3b80d8

another const prop failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86848 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86847 91177308-0d34-0410-b5e6-96231b3b80d8

Reject duplicate case values in a switch, PR5450.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86846 91177308-0d34-0410-b5e6-96231b3b80d8

Don't trivially delete unused calls to llvm.invariant.start.  This allows
llvm.invariant.start to be used without necessarily being paired with a call
to llvm.invariant.end.  If you run the entire optimization pipeline then such
calls are in fact deleted (adce does it), but that's actually a good thing since
we probably do want them to be zapped late in the game.  There should really be
an integration test that checks that the llvm.invariant.start call lasts long
enough that all passes that do interesting things with it get to do their stuff
before it is deleted.  But since no passes do anything interesting with it yet
this will have to wait for later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86840 91177308-0d34-0410-b5e6-96231b3b80d8

Add nounwind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86814 91177308-0d34-0410-b5e6-96231b3b80d8

remove the now dead condprop pass, PR3906.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86810 91177308-0d34-0410-b5e6-96231b3b80d8

Fix JITTest.ModuleDeletion in -Asserts mode (which turns off JITEmitDebugInfo
by default).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86807 91177308-0d34-0410-b5e6-96231b3b80d8

remove condprop testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86804 91177308-0d34-0410-b5e6-96231b3b80d8

Add StringRef::split(StringRef), to complement StringRef::split(char).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86803 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86802 91177308-0d34-0410-b5e6-96231b3b80d8

Show command-line args and features passed into backend in debug output. Approved by Evan Cheng.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86797 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing run line. Devang, please check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86795 91177308-0d34-0410-b5e6-96231b3b80d8

Fix -Asserts warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86794 91177308-0d34-0410-b5e6-96231b3b80d8

The TBB and TBH instructions for Thumb2 are really handy for jump tables, but
can only branch forward. To best take advantage of them, we'd like to adjust
the basic blocks around a bit when reasonable. This patch puts basics in place
to do that, with a super-simple algorithm for backwards jump table targets that
creates a new branch after the jump table which branches backwards. Real
heuristics for reordering blocks or other modifications rather than inserting
branches will follow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86791 91177308-0d34-0410-b5e6-96231b3b80d8

stub out some LazyValueInfo interfaces, and have JumpThreading
start using them in a trivial way when -enable-jump-threading-lvi
is passed.  enable-jump-threading-lvi will be my playground for
awhile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86789 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test to work on every platform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86786 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test to work on every platform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86785 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86784 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure that the exception handling data has the same visibility as the
function it's generated for.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86779 91177308-0d34-0410-b5e6-96231b3b80d8

Add Triple::str() which returns the contents of the Triple as a string, as a more readable alternative to getTriple().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86773 91177308-0d34-0410-b5e6-96231b3b80d8

Do not assume first function scope seen represents current function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86771 91177308-0d34-0410-b5e6-96231b3b80d8

Add From arguments to StringRef search functions, and tweak doxyments.

Also, add unittests for find_first_of and find_first_not_of.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86770 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-gcc/clang don't (won't?) need this hack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86769 91177308-0d34-0410-b5e6-96231b3b80d8

oops, didn't mean to commit this, no harm, but add a todoops, didn't mean to commit this, no harm, but add a todoo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86768 91177308-0d34-0410-b5e6-96231b3b80d8

Stub out a new lazy value info pass, which will eventually
vend value constraint information to the optimizer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86767 91177308-0d34-0410-b5e6-96231b3b80d8

add a fixme

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86766 91177308-0d34-0410-b5e6-96231b3b80d8

remove redundant foward declaration.  This function is already in
Analysis/Passes.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86765 91177308-0d34-0410-b5e6-96231b3b80d8

While creating DbgScopes, do not forget parent scope.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86763 91177308-0d34-0410-b5e6-96231b3b80d8

Block terminator may be a switch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86761 91177308-0d34-0410-b5e6-96231b3b80d8

jump threading does everything that condprop does any more.  This passes
bootstrap on darwin i386.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86758 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86756 91177308-0d34-0410-b5e6-96231b3b80d8

I did this a week or two ago

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86754 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore variable if scope info is not available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86753 91177308-0d34-0410-b5e6-96231b3b80d8

Test this on Darwin only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86752 91177308-0d34-0410-b5e6-96231b3b80d8

Emit correct code when making a ConstantPool entry for a vector
constant whose component type is not a legal type for the target.
(If the target ConstantPool cannot handle this type either, it has
an opportunity to merge elements.  In practice any target with
8-bit bytes must support i8 *as data*).  7320806 (partial).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86751 91177308-0d34-0410-b5e6-96231b3b80d8

Implement support to debug inlined functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86748 91177308-0d34-0410-b5e6-96231b3b80d8

in -dot-cfg and -dot-cfg-only, when rendering switch instructions,
put the switch value in the successor boxes like we put T/F for branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86747 91177308-0d34-0410-b5e6-96231b3b80d8

implement a TODO by teaching jump threading about "xor x, 1".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86739 91177308-0d34-0410-b5e6-96231b3b80d8

move some generally useful functions out of jump threading
into libanalysis and transformutils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86735 91177308-0d34-0410-b5e6-96231b3b80d8

Don't mark conditional branch instructions as control barriers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86732 91177308-0d34-0410-b5e6-96231b3b80d8

Modify how the prologue encoded the "move" information for the FDE. GCC
generates a sequence similar to this:

__Z4funci:
LFB2:
        mflr r0
LCFI0:
        stmw r30,-8(r1)
LCFI1:
        stw r0,8(r1)
LCFI2:
        stwu r1,-80(r1)
LCFI3:
        mr r30,r1
LCFI4:

where LCFI3 and LCFI4 are used by the FDE to indicate what the FP, LR, and other
things are. We generated something more like this:

Leh_func_begin1:
        mflr r0
        stw r31, 20(r1)
        stw r0, 8(r1)
Llabel1:
        stwu r1, -80(r1)
Llabel2:
        mr r31, r1

Note that we are missing the "mr" instruction. This patch makes it more like the
GCC output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86729 91177308-0d34-0410-b5e6-96231b3b80d8

Process InlinedAt location info.
Update InsertDeclare to return newly inserted llvm.dbg.declare intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86727 91177308-0d34-0410-b5e6-96231b3b80d8

fix a crash in SCCP handling extractvalue of an array, pointed out and
tracked down by Stephan Reiter!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86726 91177308-0d34-0410-b5e6-96231b3b80d8

Teach PHIElimination to split critical edges when -split-phi-edges is enabled.

Critical edges leading to a PHI node are split when the PHI source variable is
live out from the predecessor block. This help the coalescer eliminate more
PHI joins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86725 91177308-0d34-0410-b5e6-96231b3b80d8

Refactoring: Extract method PHIElimination::isLiveOut().
Clean up some whitespace.
No functional changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86724 91177308-0d34-0410-b5e6-96231b3b80d8

improve comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86723 91177308-0d34-0410-b5e6-96231b3b80d8

Make jump threading eliminate blocks that just contain phi nodes,
debug intrinsics, and an unconditional branch when possible.  This
reuses the TryToSimplifyUncondBranchFromEmptyBlock function split
out of simplifycfg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86722 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize lsr code that optimize loop to count down towards zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86715 91177308-0d34-0410-b5e6-96231b3b80d8

Optimize test more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86714 91177308-0d34-0410-b5e6-96231b3b80d8

make this handle redefinition of malloc function with different prototype correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86712 91177308-0d34-0410-b5e6-96231b3b80d8

Change Thumb1 address mode printing, instead of
[r0, #2 * 4]
Now
[r0, #8]

This makes Thumb2 assembly more uniform and frankly the scale doesn't add much.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86707 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86706 91177308-0d34-0410-b5e6-96231b3b80d8

Add defensive break.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86705 91177308-0d34-0410-b5e6-96231b3b80d8

Add a monstrous hack to improve X86ISelDAGToDAG compile time.
 - Force NDEBUG on in any Release build. This drops the compile time to ~100s
   from ~600s, in Release mode.

 - This may just be a temporary workaround, I don't know the true nature of the
   gcc-4.2 compile time performance problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86695 91177308-0d34-0410-b5e6-96231b3b80d8

Fix obvious typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86694 91177308-0d34-0410-b5e6-96231b3b80d8

clarify logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86689 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Add Darwin-specific linker flags for building loadable modules

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86684 91177308-0d34-0410-b5e6-96231b3b80d8

Teach DSE to eliminate useless trampolines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86683 91177308-0d34-0410-b5e6-96231b3b80d8

Add brackets to make gcc-4.4 happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86681 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen support for the llvm.invariant/lifetime.start/end intrinsics:
just throw them away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86678 91177308-0d34-0410-b5e6-96231b3b80d8

Update computeArraySize() to use ComputeMultiple() to determine the array size associated with a malloc; also extend PerformHeapAllocSRoA() to check if the optimized malloc's arg had its highest bit set, so that it is safe for ComputeMultiple() to look through sext instructions while determining the optimized malloc's array size

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86676 91177308-0d34-0410-b5e6-96231b3b80d8

Add ComputeMultiple() analysis function that recursively determines if a Value V is a multiple of unsigned Base

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86675 91177308-0d34-0410-b5e6-96231b3b80d8

optimize test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86672 91177308-0d34-0410-b5e6-96231b3b80d8

unify the code that determines whether it is a good idea to change the type
of a computation.  This fixes some infinite loops when dealing with TD that
has no native types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86670 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86668 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply r86359, "Teach dead store elimination that certain intrinsics write to
memory just like a store" with bug fixed (partial-overwrite.ll is the
regression test).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86667 91177308-0d34-0410-b5e6-96231b3b80d8

refactor TryToSimplifyUncondBranchFromEmptyBlock out of SimplifyCFG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86666 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Support for building llvm loadable modules.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86656 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Start documentation testing architecture.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86655 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Add ExampleTests, for testing lit and demonstrating test suite features.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86654 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Fix bug in --show-suites which accidentally override the list of tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86653 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR5445

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86651 91177308-0d34-0410-b5e6-96231b3b80d8

I misread the parens, not so redundant after all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86648 91177308-0d34-0410-b5e6-96231b3b80d8

make jump threading recursively simplify expressions instead of doing it
just one level deep.  On the testcase we go from getting this:

F1:                                               ; preds = %T2
  %F = and i1 true, %cond                         ; <i1> [#uses=1]
  br i1 %F, label %X, label %Y

to a fully threaded:

F1:                                               ; preds = %T2
  br label %Y

This changes gets us to the point where we're forming (too many) switch
instructions on doug's strswitch testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86646 91177308-0d34-0410-b5e6-96231b3b80d8

remove some redundant parens.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86645 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Remove unnecessary `unset' which was not supported by old cmake
releases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86644 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86642 91177308-0d34-0410-b5e6-96231b3b80d8

Minor code simplification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86641 91177308-0d34-0410-b5e6-96231b3b80d8

Trim a bunch of unneeded code from this testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86640 91177308-0d34-0410-b5e6-96231b3b80d8

don't invalidate PN, rewrite of this code is in progress anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86639 91177308-0d34-0410-b5e6-96231b3b80d8

add a new SimplifyInstruction API, which is like ConstantFoldInstruction,
except that the result may not be a constant.  Switch jump threading to
use it so that it gets things like (X & 0) -> 0, which occur when phi preds
are deleted and the remaining phi pred was a zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86637 91177308-0d34-0410-b5e6-96231b3b80d8

Fix DenseMap iterator constness.

This patch forbids implicit conversion of DenseMap::const_iterator to
DenseMap::iterator which was possible because DenseMapIterator inherited
(publicly) from DenseMapConstIterator. Conversion the other way around is now
allowed as one may expect.

The template DenseMapConstIterator is removed and the template parameter
IsConst which specifies whether the iterator is constant is added to
DenseMapIterator.

Actually IsConst parameter is not necessary since the constness can be
determined from KeyT but this is not relevant to the fix and can be addressed
later.

Patch by Victor Zverovich!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86636 91177308-0d34-0410-b5e6-96231b3b80d8

factor simplification logic for AND and OR out to InstSimplify from instcombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86635 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed to address code review. No functional changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86634 91177308-0d34-0410-b5e6-96231b3b80d8

Fix MemoryBuffer::getSTDIN to *not* return null if stdin is empty, this is a lame API.

Also, Stringrefify some more MemoryBuffer functions, and add two performance FIXMEs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86630 91177308-0d34-0410-b5e6-96231b3b80d8

Allow targets to specify register classes whose member registers should not be renamed to break anti-dependencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86628 91177308-0d34-0410-b5e6-96231b3b80d8

pull a bunch of logic out of instcombine into instsimplify for compare
simplification, this handles the foldable fcmp x,x cases among many others.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86627 91177308-0d34-0410-b5e6-96231b3b80d8

Pass the (optional) TargetData object to ConstantFoldInstOperands
and ConstantFoldCompareInstOperands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@86626 91177308-0d34-0410-b5e6-96231b3b80d8