Be less stingy as to how many selects and phi nodes we
are prepared to look through.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92898 91177308-0d34-0410-b5e6-96231b3b80d8

More trivial optimizations to a function well outside the critical path

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92896 91177308-0d34-0410-b5e6-96231b3b80d8

handle ConstantVector while I'm in here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92892 91177308-0d34-0410-b5e6-96231b3b80d8

fix a globalopt crash on 'bullet' (handling evaluation of a store
to an element of a vector in a static ctor) which occurs with an
unrelated patch I'm testing.  Annoyingly, EvaluateStoreInto basically
does exactly the same stuff as InsertElement constant folding, but it
now handles vectors, and you can't insertelement into a vector.  It
would be 'really nice' if GEP into a vector were not legal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92889 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a minor regression from my dag combiner changes. One more place which needs to look pass truncates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92885 91177308-0d34-0410-b5e6-96231b3b80d8

Switch StringRef::edit_distance over to using raw pointers, since both
std::vector and llvm::SmallVector have annoying performance
tradeoffs. No, I don't expect this to matter, and now it won't.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92884 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92883 91177308-0d34-0410-b5e6-96231b3b80d8

80 column violations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92876 91177308-0d34-0410-b5e6-96231b3b80d8

Allow double defs after tail duplication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92874 91177308-0d34-0410-b5e6-96231b3b80d8

Add Target hook to duplicate machine instructions.

Some instructions refer to unique labels, and so cannot be trivially cloned
with CloneMachineInstr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92873 91177308-0d34-0410-b5e6-96231b3b80d8

Addressing mode 6 (load/store) instructions can't encode an immediate offset
for stack references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92871 91177308-0d34-0410-b5e6-96231b3b80d8

Anti-dependency breaking needs to be careful regarding instructions with
multiple register definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92864 91177308-0d34-0410-b5e6-96231b3b80d8

Move the object size intrinsic optimization to inst-combine and make
it work for any integer size return type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92853 91177308-0d34-0410-b5e6-96231b3b80d8

Comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92850 91177308-0d34-0410-b5e6-96231b3b80d8

Teach dag combine to fold the following transformation more aggressively:
(OP (trunc x), (trunc y)) -> (trunc (OP x, y))

Unfortunately this simple change causes dag combine to infinite looping. The problem is the shrink demanded ops optimization tend to canonicalize expressions in the opposite manner. That is badness. This patch disable those optimizations in dag combine but instead it is done as a late pass in sdisel.

This also exposes some deficiencies in dag combine and x86 setcc / brcond lowering. Teach them to look pass ISD::TRUNCATE in various places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92849 91177308-0d34-0410-b5e6-96231b3b80d8

Fix struct/class mismatch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92841 91177308-0d34-0410-b5e6-96231b3b80d8

When parsing function-local metadata, create a function-local MDNode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92838 91177308-0d34-0410-b5e6-96231b3b80d8

80 column and whitespace cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92837 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a README item: have functionattrs look through selects and
phi nodes when deciding which pointers point to local memory.
I actually checked long ago how useful this is, and it isn't
very: it hardly ever fires in the testsuite, but since Chris
wants it here it is!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92836 91177308-0d34-0410-b5e6-96231b3b80d8

Formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92831 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed malformed -*- lines in PBQP headers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92830 91177308-0d34-0410-b5e6-96231b3b80d8

Partially address a README by having functionattrs consider calls to
memcpy, memset and other intrinsics that only access their arguments
to be readnone if the intrinsic's arguments all point to local memory.
This improves the testcase in the README to readonly, but it could in
theory be made readnone, however this would involve more sophisticated
analysis that looks through the memcpy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92829 91177308-0d34-0410-b5e6-96231b3b80d8

This is testing a darwin specific feature, so only turn
it on for darwin (it fails on linux).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92826 91177308-0d34-0410-b5e6-96231b3b80d8

tweaks suggested by Duncan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92824 91177308-0d34-0410-b5e6-96231b3b80d8

Reenable debug info on PPC.  Works well enough to
bootstrap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92818 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure debug info hook gets called when emitting
synonyms for PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92817 91177308-0d34-0410-b5e6-96231b3b80d8

Teach instcombine's sext elimination logic to be more aggressive.
Previously, instcombine would only promote an expression tree to
the larger type if doing so eliminated two casts.  This is because
a need to manually do the sign extend after the promoted expression
tree with two shifts.  Now, we keep track of whether the result of
the computation is going to be properly sign extended already.  If
so, we can unconditionally promote the expression, which allows us
to zap more sext's.

This implements rdar://6598839 (aka gcc pr38751)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92815 91177308-0d34-0410-b5e6-96231b3b80d8

Add <imp-def> and <imp-kill> operands when replacing virtual sub-register defs and kills.

An instruction like this:

  %reg1097:1<def> = VMOVSR %R3<kill>, 14, %reg0

Must be replaced with this when substituting physical registers:

  %S0<def> = VMOVSR %R3<kill>, 14, %reg0, %D0<imp-def>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92812 91177308-0d34-0410-b5e6-96231b3b80d8

The previous code could potentially cause a cycle. Allow ordering w.r.t. a 0 order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92810 91177308-0d34-0410-b5e6-96231b3b80d8

Only check the ordering if there is an ordering for each nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92807 91177308-0d34-0410-b5e6-96231b3b80d8

Add a semi-primitive form of scheduling via the "SDNode ordering" to the
bottom-up scheduler. We prefer the lower order number.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92806 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92800 91177308-0d34-0410-b5e6-96231b3b80d8

Move this test from test/Transforms/IndVarSimplify to
test/CodeGen/X86, as doesn't use -indvars, and it does use
llc -march=x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92799 91177308-0d34-0410-b5e6-96231b3b80d8

Don't assign the shift the same type as the variable being shifted. This could
result in illegal types for the SHL operator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92797 91177308-0d34-0410-b5e6-96231b3b80d8

Undo r92785, it caused test failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92796 91177308-0d34-0410-b5e6-96231b3b80d8

make this a static function instead of a method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92795 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use the ISD::NodeType enum for SDNode opcodes, as CodeGen
uses several kinds of opcode values which are not declared within
that enum. This fixes PR5946.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92794 91177308-0d34-0410-b5e6-96231b3b80d8

Re-add parsing of function-local metadata; this time with testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92793 91177308-0d34-0410-b5e6-96231b3b80d8

more rearrangement and cleanup, fix my test failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92792 91177308-0d34-0410-b5e6-96231b3b80d8

cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92790 91177308-0d34-0410-b5e6-96231b3b80d8

remove two trunc xforms that are subsumed by EvaluateInDifferentType.
The only difference is that EvaluateInDifferentType checks to ensure
they are profitable before doing them :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92788 91177308-0d34-0410-b5e6-96231b3b80d8

merge some tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92786 91177308-0d34-0410-b5e6-96231b3b80d8

Add Rt2 to the asm format string for 32-bit Thumb load/store register dual
instructions.  Thumb does not have the restriction that t2 = t+1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92785 91177308-0d34-0410-b5e6-96231b3b80d8

merge cast2 into cast.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92784 91177308-0d34-0410-b5e6-96231b3b80d8

Allow null to be an element of NamedMDNode. e.g. !llvm.stuff = !{!0, !1, null}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92783 91177308-0d34-0410-b5e6-96231b3b80d8

remove useless test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92782 91177308-0d34-0410-b5e6-96231b3b80d8

another example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92781 91177308-0d34-0410-b5e6-96231b3b80d8

Don't specify CR sub-registers as implicit defs of BL instructions.

It is enough to give the super registers CR0, CR1, ..., and specifying the
sub-registers as well causes confusion in the liveness computations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92778 91177308-0d34-0410-b5e6-96231b3b80d8

remove a useless negative test, add a rdar # to an xfail that I'm working on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92777 91177308-0d34-0410-b5e6-96231b3b80d8

clean up tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92776 91177308-0d34-0410-b5e6-96231b3b80d8

just remove this xform which is subsumed by others.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92775 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment for CheckDebugInfoIntrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92774 91177308-0d34-0410-b5e6-96231b3b80d8

move a trunc-specific transform out of commonIntCastTransforms into visitTrunc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92773 91177308-0d34-0410-b5e6-96231b3b80d8

Restore dump() methods to Loop and MachineLoop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92772 91177308-0d34-0410-b5e6-96231b3b80d8

Move remaining stuff to the isInteger predicate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92771 91177308-0d34-0410-b5e6-96231b3b80d8

move a zext specific xform out of commonIntCastTransforms into visitZExt and modernize it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92770 91177308-0d34-0410-b5e6-96231b3b80d8

Remove livein checks from machine code verifier.

A phi operand that is implicitly defined in a predecessor becomes an undefined
register after phi elimination. This causes a lot of false positives when the
verifier is checking if live-in registers are live-out from all predecessors.

Removing the verifier checks seems like a better solution than insisting on
IMPLICIT_DEF instructions in predecessor blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92769 91177308-0d34-0410-b5e6-96231b3b80d8

move a trunc-specific xform out of commonIntCastTransforms into visitTrunc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92768 91177308-0d34-0410-b5e6-96231b3b80d8

reduce indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92766 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92763 91177308-0d34-0410-b5e6-96231b3b80d8

Proper deduce z/System LLVM target from target triple when
--enable-targets=host is specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92762 91177308-0d34-0410-b5e6-96231b3b80d8

NamedMDNode is a collection MDNodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92761 91177308-0d34-0410-b5e6-96231b3b80d8

Convert a ton of simple integer type equality tests to the new predicate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92760 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new predicate for integer type equality tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92759 91177308-0d34-0410-b5e6-96231b3b80d8

Add an !eq() operator to TableGen.  It operates on strings only.
Use !cast<string>() to compare other types of objects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92754 91177308-0d34-0410-b5e6-96231b3b80d8

Add an entry on SmallBitVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92747 91177308-0d34-0410-b5e6-96231b3b80d8

optimize comparisons against cttz/ctlz/ctpop, patch by Alastair Lynn!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92745 91177308-0d34-0410-b5e6-96231b3b80d8

Delete useless trailing semicolons.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92740 91177308-0d34-0410-b5e6-96231b3b80d8

Nick Lewycky pointed out that this code makes changes unconditionally.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92739 91177308-0d34-0410-b5e6-96231b3b80d8

Have TableGen emit code that uses dbgs() rather than errs().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92738 91177308-0d34-0410-b5e6-96231b3b80d8

If a scope has only one instruction then first instruction is also the last instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92736 91177308-0d34-0410-b5e6-96231b3b80d8

Set Changed properly after calling DeleteDeadPHIs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92735 91177308-0d34-0410-b5e6-96231b3b80d8

Use do+while instead of while for loops which obviously have a
non-zero trip count. Use SmallVector's pop_back_val().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92734 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92733 91177308-0d34-0410-b5e6-96231b3b80d8

Make RecursivelyDeleteTriviallyDeadInstructions,
RecursivelyDeleteDeadPHINode, and DeleteDeadPHIs return a flag
indicating whether they made any changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92732 91177308-0d34-0410-b5e6-96231b3b80d8

Some versions of GCC don't like non-static data members in sizeof
in this context.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92731 91177308-0d34-0410-b5e6-96231b3b80d8

Add a SmallBitVector class, which mimics BitVector but uses only
a single pointer (PointerIntPair) member. In "small" mode, the
pointer field is reinterpreted as a set of bits. In "large" mode,
the pointer points to a heap-allocated object.

Also, give BitVector empty and swap functions.

And, add some simple unittests for BitVector and SmallBitVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92730 91177308-0d34-0410-b5e6-96231b3b80d8

Add newline at EOF.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92727 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid going through the LLVMContext for type equality where it's safe to dereference the type pointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92726 91177308-0d34-0410-b5e6-96231b3b80d8

prune some #includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92712 91177308-0d34-0410-b5e6-96231b3b80d8

split and/or/xor out into one overly-large (2000LOC) file.  However, I think
it does make sense to keep them together, at least for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92711 91177308-0d34-0410-b5e6-96231b3b80d8

missed file with previous commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92710 91177308-0d34-0410-b5e6-96231b3b80d8

split instcombine of shifts out to its own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92709 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate getBitCastOperand and simplify some over-complex inbounds stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92708 91177308-0d34-0410-b5e6-96231b3b80d8

split call handling out to InstCombineCalls.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92707 91177308-0d34-0410-b5e6-96231b3b80d8

optimize cttz and ctlz when we can prove something about the
leading/trailing bits.  Patch by Alastair Lynn!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92706 91177308-0d34-0410-b5e6-96231b3b80d8

this inline function moved to addsub

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92705 91177308-0d34-0410-b5e6-96231b3b80d8

split add/sub out to its own file.  Eliminate use of
dyn_castNotVal in the X+~X transform.  dyn_castNotVal is
dramatic overkill for what the xform needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92704 91177308-0d34-0410-b5e6-96231b3b80d8

all the places we use hasOneUse() we know are instructions, so inline
and simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92700 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate AssociativeOpt and its last uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92697 91177308-0d34-0410-b5e6-96231b3b80d8

inline the FoldICmpLogical functor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92695 91177308-0d34-0410-b5e6-96231b3b80d8

Code refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92694 91177308-0d34-0410-b5e6-96231b3b80d8

inline the 'AddRHS' transformation, simplifying things significantly.
Eliminate the 'AddMaskingAnd' transformation, it is redundant with this
more general code right below it:
  // A+B --> A|B iff A and B have no bits set in common.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92693 91177308-0d34-0410-b5e6-96231b3b80d8

remove massive over-genality manifested as a big template
that got instantiated.  There is no reason for instcombine
to try this hard for simple associative optimizations.  Next
up, eliminate the template completely.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92692 91177308-0d34-0410-b5e6-96231b3b80d8

split mul/div/rem instructions out to their own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92689 91177308-0d34-0410-b5e6-96231b3b80d8

clean up header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92688 91177308-0d34-0410-b5e6-96231b3b80d8

split select out to its own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92687 91177308-0d34-0410-b5e6-96231b3b80d8

split out load/store/alloca.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92685 91177308-0d34-0410-b5e6-96231b3b80d8

reduce indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92684 91177308-0d34-0410-b5e6-96231b3b80d8

split vector stuff out to InstCombineVectorOps.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@92683 91177308-0d34-0410-b5e6-96231b3b80d8