Local spiller optimization:
Turn a store folding instruction into a load folding instruction. e.g.
     xorl  %edi, %eax
     movl  %eax, -32(%ebp)
     movl  -36(%ebp), %eax
     orl   %eax, -32(%ebp)
=>
     xorl  %edi, %eax
     orl   -36(%ebp), %eax
     mov   %eax, -32(%ebp)
This enables the unfolding optimization for a subsequent instruction which will
also eliminate the newly introduced store instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43192 91177308-0d34-0410-b5e6-96231b3b80d8

Don't branch fold inline asm statements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43191 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for a few more nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43190 91177308-0d34-0410-b5e6-96231b3b80d8

Redo "last ppc long double fix" as Chris wants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43189 91177308-0d34-0410-b5e6-96231b3b80d8

Removed "-arch i386" by popular demand...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43186 91177308-0d34-0410-b5e6-96231b3b80d8

Now with RUN line!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43185 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the Foundation.h file because not everyone has that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43184 91177308-0d34-0410-b5e6-96231b3b80d8

This test now passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43183 91177308-0d34-0410-b5e6-96231b3b80d8

Update this file for 2.0 syntax. Contributed by Jan Rehders

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43182 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a really nasty vector miscompilation bill recently introduced.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43181 91177308-0d34-0410-b5e6-96231b3b80d8

Updated VC++ build system

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43180 91177308-0d34-0410-b5e6-96231b3b80d8

rename ExpandOperation to ExpandOperationResult, as suggested
by Duncan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43177 91177308-0d34-0410-b5e6-96231b3b80d8

split LowerMEMCPY into LowerMEMCPYCall and LowerMEMCPYInline in the ARM backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43176 91177308-0d34-0410-b5e6-96231b3b80d8

Support for expanding ADDE and SUBE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43175 91177308-0d34-0410-b5e6-96231b3b80d8

If the value types are equal then this routine
asserts in later checks rather than producing
the ordinary load it is supposed to.  Avoid all
such hassles by directly returning an ordinary
load in this case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43174 91177308-0d34-0410-b5e6-96231b3b80d8

Test byval with a 8 bit aligned struct

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43173 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for byval function whose argument is not 32 bit aligned.
To do this it is necessary to add a "always inline" argument to the
memcpy node. For completeness I have also added this node to memmove
and memset.  I have also added getMem* functions, because the extra
argument makes it cumbersome to use getNode and because I get confused
by it :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43172 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a few new operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43171 91177308-0d34-0410-b5e6-96231b3b80d8

Implement expansion of SINT_TO_FP and UINT_TO_FP operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43170 91177308-0d34-0410-b5e6-96231b3b80d8

implement support for custom expansion of any node type, in one place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43169 91177308-0d34-0410-b5e6-96231b3b80d8

comment fixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43168 91177308-0d34-0410-b5e6-96231b3b80d8

Make use of TLI.ExpandOperation, remove softfloat stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43167 91177308-0d34-0410-b5e6-96231b3b80d8

add expand support for bit_convert result, even allowing custom expansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43166 91177308-0d34-0410-b5e6-96231b3b80d8

add a new target hook.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43165 91177308-0d34-0410-b5e6-96231b3b80d8

Add an easy microoptimization I noticed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43164 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43162 91177308-0d34-0410-b5e6-96231b3b80d8

Negative indices aren't allowed here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43161 91177308-0d34-0410-b5e6-96231b3b80d8

More ppcf128 issues (maybe the last)?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43160 91177308-0d34-0410-b5e6-96231b3b80d8

- Added getOpcodeAfterMemoryUnfold(). It doesn't unfold an instruction, but only returns the opcode of the instruction post unfolding.
- Fix some copy+paste bugs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43153 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVectorImpl instead of SmallVector with hardcoded size in MRegister public interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43150 91177308-0d34-0410-b5e6-96231b3b80d8

Try again.
Instead of loading small global string from memory, use
integer constant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43148 91177308-0d34-0410-b5e6-96231b3b80d8

Allow GVN to eliminate redundant calls to functions without side effects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43147 91177308-0d34-0410-b5e6-96231b3b80d8

Add an uppercase conversion utility function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43146 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a misnamed parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43145 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43144 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1735 and Transforms/DeadArgElim/2007-10-18-VarargsReturn.ll by
fixing some obviously broken code :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43141 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the command line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43140 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43136 91177308-0d34-0410-b5e6-96231b3b80d8

this doesn't need dynamic_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43133 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43131 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce reliance on rtti info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43130 91177308-0d34-0410-b5e6-96231b3b80d8

fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43129 91177308-0d34-0410-b5e6-96231b3b80d8

update comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43128 91177308-0d34-0410-b5e6-96231b3b80d8

This requires rtti info because tblgen uses commandline,
and tblgen requires rtti.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43127 91177308-0d34-0410-b5e6-96231b3b80d8

tblgen uses dynamic_cast heavily, so it needs rtti info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43126 91177308-0d34-0410-b5e6-96231b3b80d8

Updated VC++ build system

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43123 91177308-0d34-0410-b5e6-96231b3b80d8

Work around downrev gccs which do not inherit visibility of the
Registry<>::iterator member class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43122 91177308-0d34-0410-b5e6-96231b3b80d8

Missing 'public' keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43121 91177308-0d34-0410-b5e6-96231b3b80d8

Pointer arithmetic should be done with the index the same size as the pointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43120 91177308-0d34-0410-b5e6-96231b3b80d8

Support for ADDC/SUBC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43119 91177308-0d34-0410-b5e6-96231b3b80d8

Really fix PR1734. Carefully track which register uses are sub-register uses by
traversing inverse register coalescing map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43118 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43117 91177308-0d34-0410-b5e6-96231b3b80d8

legalizing the ret operation on f64 shouldn't introduce a new
i64 bit convert needlessly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43116 91177308-0d34-0410-b5e6-96231b3b80d8

Move Split<...>() into DomTreeBase.  This should make the #include's of DominatorInternals.h
in CodeExtractor and LoopSimplify unnecessary.

Hartmut, could you confirm that this fixes the issues you were seeing?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43115 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43111 91177308-0d34-0410-b5e6-96231b3b80d8

Changed the return type of type-specific Allocate() methods to return
void*.  This is hint that we are returning uninitialized memory rather
than a constructed object.

Patched ImutAVLTree to conform to this new interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43106 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting r43070 for now. It's causing llc test failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43103 91177308-0d34-0410-b5e6-96231b3b80d8

Test to make sure we don't generate unwind info for non-64-bit Objective-C.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43102 91177308-0d34-0410-b5e6-96231b3b80d8

ImutAVLTree now allocates tree nodes from the BumpPtrAllocator using
the new type-aligned Allocate() method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43100 91177308-0d34-0410-b5e6-96231b3b80d8

Removed inclusion of cassert, which is no longer needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43099 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cosmetic cleanups in the calculation of alignments for
StringMapEntry objects.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43097 91177308-0d34-0410-b5e6-96231b3b80d8

Added template function alignof() which provides a clean
function-based interface to getting the alignment of a type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43096 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting unnecessary commit of generated files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43095 91177308-0d34-0410-b5e6-96231b3b80d8

Switching TargetMachineRegistry to use the new generic Registry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43094 91177308-0d34-0410-b5e6-96231b3b80d8

Updated StringMap to use llvm::AlignOf to compute the alignment of map
entries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43089 91177308-0d34-0410-b5e6-96231b3b80d8

Ignoring test/Transforms/LoopIndexSplit/Output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43088 91177308-0d34-0410-b5e6-96231b3b80d8

Added member template functions to MallocAllocator and
BumpPtrAllocator that implement allocations that return a properly
typed pointer.  For BumpPtrAllocator, the allocated memory is
automatically aligned to the minimum alignment of the type (as
calculated by llvm::AlignOf::Alignment).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43087 91177308-0d34-0410-b5e6-96231b3b80d8

Added llvm::AlignOf, a template class whose purpose is to portably
compute the minimum memory alignment of arbitrary types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43086 91177308-0d34-0410-b5e6-96231b3b80d8

Do not raise free() call that is called through invoke instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43083 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed linker errors (unresolved externals: split<>(...)) when compiling with VC++. Please review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43081 91177308-0d34-0410-b5e6-96231b3b80d8

Don't mention -enable-unsafe-fp-math in LangRef.html, as that option is
assumed to globally mutate the spec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43078 91177308-0d34-0410-b5e6-96231b3b80d8

Define a helper function ConstantVector::getSplatValue for testing for
and working with broadcasted constants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43076 91177308-0d34-0410-b5e6-96231b3b80d8

Updated VC++ build system.
Silenced some VC warnings.

I'm getting linker errors, though: unresolved externals:

llvm::Split<class llvm::BasicBlock *,struct llvm::GraphTraits<class llvm::BasicBlock *> >(class llvm::DominatorTreeBase<class llvm::BasicBlock> &,class llvm::BasicBlock *)

and

llvm::Split<struct llvm::Inverse<class llvm::BasicBlock *>,struct llvm::GraphTraits<struct llvm::Inverse<class llvm::BasicBlock *> > >(class llvm::DominatorTreeBase<class llvm::BasicBlock> &,class llvm::BasicBlock *)

Where are these defined?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43073 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for ISD::SELECT in SplitVectorOp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43072 91177308-0d34-0410-b5e6-96231b3b80d8

Return Expand from getOperationAction for all extended
types.  This is needed for SIGN_EXTEND_INREG at least.
It is not clear if this is correct for other operations.
On the other hand, for the various load/store actions
it seems to correct to return the type action, as is
currently done.
Also, it seems that SelectionDAG::getValueType can be
called for extended value types; introduce a map for
holding these, since we don't really want to extend
the vector to be 2^32 pointers long!
Generalize DAGTypeLegalizer::PromoteResult_TRUNCATE
and DAGTypeLegalizer::PromoteResult_INT_EXTEND to handle
the various funky possibilities that apints introduce,
for example that you can promote to a type that needs
to be expanded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43071 91177308-0d34-0410-b5e6-96231b3b80d8

Apply "Instead of loading small c string constant, use integer constant directly" transformation while processing load instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43070 91177308-0d34-0410-b5e6-96231b3b80d8

Apply Chris' suggestions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43069 91177308-0d34-0410-b5e6-96231b3b80d8

Change fp to sint legalization on x86-32 to do 2 x i32
loads instead of 1 x i64 loads.  This doesn't change any functionality yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43068 91177308-0d34-0410-b5e6-96231b3b80d8

fix some funny indentation, add comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43066 91177308-0d34-0410-b5e6-96231b3b80d8

One more extract_subreg coalescing bug fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43065 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43064 91177308-0d34-0410-b5e6-96231b3b80d8

Yet another test case for extract_subreg coalescing crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43063 91177308-0d34-0410-b5e6-96231b3b80d8

Fix MergeValueInAsValue(). It allows overlapping live ranges but should replace
their value numbers with the specified value number.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43062 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up code that calculate MBB live-in's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43061 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up code that calculate MBB live-in's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43060 91177308-0d34-0410-b5e6-96231b3b80d8

Move splitBlock into DomTreeBase from DomTree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43059 91177308-0d34-0410-b5e6-96231b3b80d8

Use immediate stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43055 91177308-0d34-0410-b5e6-96231b3b80d8

Disable attempts to constant fold PPC f128.
Remove the assumption that this will happen from
various places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43053 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43049 91177308-0d34-0410-b5e6-96231b3b80d8

Some clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43043 91177308-0d34-0410-b5e6-96231b3b80d8

Template DominatorTreeBase by node type.  This is the next major step towards
having dominator information on MBB's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43036 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1734.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43035 91177308-0d34-0410-b5e6-96231b3b80d8

New test for svn rev 43033, radar 5538745.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43034 91177308-0d34-0410-b5e6-96231b3b80d8

Check for invalid cc's in f80 select.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43033 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug handling frame references in ppc inline asm when the frame offset
doesn't fit into 16 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43032 91177308-0d34-0410-b5e6-96231b3b80d8

Document the fact that the verifier currently requires the
top bit of a ValueType to be zero.  Enforce this by ensuring
an assertion failure if someone tries to create a ValueType
without this property.  I chose this minimal approach rather
than a more official integration of the notion of reserved
bits into ValueType because I'm hoping that the verifier will
be changed to no longer require this :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43031 91177308-0d34-0410-b5e6-96231b3b80d8

Initial infrastructure for arbitrary precision integer
codegen support.  This should have no effect on codegen
for other types.  Debatable bits: (1) the use (abuse?)
of a set in SDNode::getValueTypeList; (2) the length of
getTypeToTransformTo, which maybe should be refactored
with a non-inline part for extended value types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43030 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes due to lack of type-safety for ValueType: (1) ValueType
being passed instead of an opcode; (2) ValueType being passed
for isVolatile (!) in getLoad.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43028 91177308-0d34-0410-b5e6-96231b3b80d8

Correction to tail call optimization code. The new return address
was stored to the acutal stack slot before the parameters were
lowered to their stack slot. This could cause arguments to be
overwritten by the return address if the called function had less
parameters than the caller function. The update should remove the
last failing test case of llc-beta: SPASS.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43027 91177308-0d34-0410-b5e6-96231b3b80d8

Code clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43026 91177308-0d34-0410-b5e6-96231b3b80d8