rearrange some code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75113 91177308-0d34-0410-b5e6-96231b3b80d8

make direct calls set MO_PLT or MO_DARWIN_STUB as appropriate with fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75112 91177308-0d34-0410-b5e6-96231b3b80d8

Don't create undefined symbols for aliases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75111 91177308-0d34-0410-b5e6-96231b3b80d8

reduce nesting by rearranging branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75110 91177308-0d34-0410-b5e6-96231b3b80d8

make isel use MO_PIC_BASE_OFFSET when lowering globalvalues on darwin in pic
mode, instead of having asmprinter just "know" to print them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75109 91177308-0d34-0410-b5e6-96231b3b80d8

various minor cleanups, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75108 91177308-0d34-0410-b5e6-96231b3b80d8

make isel decide whether to emit $stub's on darwin instead of asmprinter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75107 91177308-0d34-0410-b5e6-96231b3b80d8

pc-relative references are *always* to functions, never to globals.  ISel
is incapable of lowering direct references to globals like this, zap the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75106 91177308-0d34-0410-b5e6-96231b3b80d8

make cygwin use its own stubs set, instead of overloading a darwin one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75105 91177308-0d34-0410-b5e6-96231b3b80d8

Make isel determine where to emit PLT-relative calls instead of having
asmprinter do it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75104 91177308-0d34-0410-b5e6-96231b3b80d8

add missing predicate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75103 91177308-0d34-0410-b5e6-96231b3b80d8

isPICStyleStub() is now never true in -static mode, so simplify code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75102 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some code based on the fact that picstyles != none are only valid
in pic or dynamic-no-pic mode. Also, x86-64 never used picstylegot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75101 91177308-0d34-0410-b5e6-96231b3b80d8

all this logic always returns true because GOT mode is never active in x86-64 mode.
Simplify it away, someone should evaluate this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75100 91177308-0d34-0410-b5e6-96231b3b80d8

isPICStyleRIPRel() and friends are never true in -static mode.
Simplify code based on this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75099 91177308-0d34-0410-b5e6-96231b3b80d8

Move PR4517 to README.txt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75098 91177308-0d34-0410-b5e6-96231b3b80d8

Improved tracking of value number kills. VN kills are now represented
as an (index,bool) pair. The bool flag records whether the kill is a
PHI kill or not. This code will be used to enable splitting of live
intervals containing PHI-kills.

A slight change to live interval weights introduced an extra spill
into lsr-code-insertion (outside the critical sections). The test
condition has been updated to reflect this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75097 91177308-0d34-0410-b5e6-96231b3b80d8

.o file writing shouldn't mess around with pic/relo models like the JIT does.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75096 91177308-0d34-0410-b5e6-96231b3b80d8

move a hack out of the asm-printer specific path to the main target selection path.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75095 91177308-0d34-0410-b5e6-96231b3b80d8

many more cleanups, for example if in the "none" pic-style,
make sure we're set to static codegen.  Simplify the decision
tree of target->picstyle/picmode settings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75094 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce indentation in GVRequiresExtraLoad.  Return true for windows
with DLLImport symbols even when in -static mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75093 91177308-0d34-0410-b5e6-96231b3b80d8

When in -static mode, force the PIC style to none.  Doing this requires fixing
code which conflated RIPRel PIC with x86-64.  Fix these to just check for X86-64
directly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75092 91177308-0d34-0410-b5e6-96231b3b80d8

merge two identical functions and simplify things that are GOT specific

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75091 91177308-0d34-0410-b5e6-96231b3b80d8

hoist check for IsTailCall to callers.  Eliminate redundant check for
x86-64: GOT-style PIC is never used on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75090 91177308-0d34-0410-b5e6-96231b3b80d8

change a few methods to be static functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75089 91177308-0d34-0410-b5e6-96231b3b80d8

one more added assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75088 91177308-0d34-0410-b5e6-96231b3b80d8

remove eh, convert to FileCheck style

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75087 91177308-0d34-0410-b5e6-96231b3b80d8

move handling of dllimport linkage in isel, not in asmprinter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75086 91177308-0d34-0410-b5e6-96231b3b80d8

we have no tests for dllimport/export.  Add one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75085 91177308-0d34-0410-b5e6-96231b3b80d8

Work around an ICE in gcc-4.2.4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75084 91177308-0d34-0410-b5e6-96231b3b80d8

one more assertion!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75083 91177308-0d34-0410-b5e6-96231b3b80d8

add some more assertions.  Remove code to handle dllimport on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75082 91177308-0d34-0410-b5e6-96231b3b80d8

Make SelectionDAG::getVectorShuffle work properly for VECTOR_SHUFFLE
nodes with operand types that differ from the result type. (This
doesn't normally happen right now, because
SelectionDAGLowering::visitShuffleVector normalizes vector shuffles.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75081 91177308-0d34-0410-b5e6-96231b3b80d8

remove "asmcall", using print_pcrel_imm instead of printOperand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75080 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some logic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75079 91177308-0d34-0410-b5e6-96231b3b80d8

* add some assertions for sanity checking.
* remove some old code that was needed when we'd put ESP in the scale instead of
  the base of some instructions.
* Fix a bug with the P modifier in inline asm that caused us to drop it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75077 91177308-0d34-0410-b5e6-96231b3b80d8

stop on the first file mismatch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75076 91177308-0d34-0410-b5e6-96231b3b80d8

add a test for dale's recent change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75074 91177308-0d34-0410-b5e6-96231b3b80d8

Make SmallPtrSet iterators real iterators

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75073 91177308-0d34-0410-b5e6-96231b3b80d8

LLVMContext-ification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75072 91177308-0d34-0410-b5e6-96231b3b80d8

switch test to FileCheck-style and test the P and non-P cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75071 91177308-0d34-0410-b5e6-96231b3b80d8

rename a test to make it a feature test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75070 91177308-0d34-0410-b5e6-96231b3b80d8

Use common code for both ARM and Thumb-2 instruction and register info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75067 91177308-0d34-0410-b5e6-96231b3b80d8

* manually concatenate two string literals
* remove some dead code: darwin doesn't support dllimport linkage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75066 91177308-0d34-0410-b5e6-96231b3b80d8

- Add some NEON ld / st instruction static encoding.
- Make bits 25-27 for ldrh, etc. explicitly zero. Previously only the JIT uses the encoding information and it's assuming anything not specified to be zero. Making them explicit so the disassembler is happy.
Patch by Sean Callanan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75065 91177308-0d34-0410-b5e6-96231b3b80d8

Fix thinko.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75061 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the AssertingVH unittests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75060 91177308-0d34-0410-b5e6-96231b3b80d8

Add an option to allocate JITed global data separately from code.  By
default, this option is not enabled to support clients who rely on
this behavior.

Fixes http://llvm.org/PR4483

A patch to allocate additional memory for globals after we run out is
forthcoming.

Patch by Reid Kleckner!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75059 91177308-0d34-0410-b5e6-96231b3b80d8

Reformat.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75058 91177308-0d34-0410-b5e6-96231b3b80d8

Drop "constant" from
  !0 = constant metadata !{...}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75057 91177308-0d34-0410-b5e6-96231b3b80d8

Allow users of GraphWriter to display graphs asynchronously.  This
provides a way to quickly dump a bunch of graph information to dot files
and display them.  It's a timesaver when working on large systems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75056 91177308-0d34-0410-b5e6-96231b3b80d8

Add an ExecuteNoWait interface to support asynchronous process spawning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75055 91177308-0d34-0410-b5e6-96231b3b80d8

Update SLotTracker to handle MDNode slots.
Simplify MDNode printing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75053 91177308-0d34-0410-b5e6-96231b3b80d8

Nowadays vectors are only split if they have an even
number of elements.  Make some simplifications based
on this (in particular SplitVecRes_SETCC).  Tighten
up some checking while there.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75050 91177308-0d34-0410-b5e6-96231b3b80d8

Change how so_imm and t2_so_imm are handled. At instruction selection time, the immediates are no longer encoded in the imm8 + rot format, that are left as it is. The encoding is now done in ams printing and code emission time instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75048 91177308-0d34-0410-b5e6-96231b3b80d8

Recommit r74952 with a bug fix:

DWARF requires frame moves be specified at specific times. If you have a
prologue like this:

__Z3fooi:
Leh_func_begin1:
LBB1_0: ## entry
       pushl   %ebp
Llabel1:
       movl    %esp, %ebp
Llabel2:
       pushl   %esi
Llabel3:
       subl    $20, %esp
       call    "L1$pb"
"L1$pb":
       popl    %esi

The "pushl %ebp" needs a table entry specifying the offset. The "movl %esp,
%ebp" makes %ebp the new stack frame register, so that needs to be specified in
DWARF. And "pushl %esi" saves the callee-saved %esi register, which also needs
to be specified in DWARF.

Before, all of this logic was in one method. This didn't work too well, because
as you can see there are multiple FDE line entries that need to be created.

This fix creates the "MachineMove" objects directly when they're needed; instead
of waiting until the end, and losing information.

There is some ugliness where we generate code like this:

LBB22_0: ## entry
pushl %ebp
Llabel280:
movl %esp, %ebp
Llabel281:
Llabel284:
pushl %ebp  <----------
pushl %ebx
pushl %edi
pushl %esi
Llabel282:
subl $328, %esp

Notice the extra "pushl %ebp". If we generate a "machine move" instruction in
the FDE for that pushl, the linker may get very confused about what value %ebp
should have when exitting the function. I.e., it'll give it the value %esp
instead of the %ebp value from the first "pushl". Not to mention that, in this
case, %ebp isn't modified in the function (that's a separate bug). I put a small
hack in to get it to work. It might be the only solution, but should be
revisited once the above case is fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75047 91177308-0d34-0410-b5e6-96231b3b80d8

Use interators instead of counters for loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75046 91177308-0d34-0410-b5e6-96231b3b80d8

Missed an exit during the conversion.
Will convert assert(0) that don't have abort() to LLVM_UNREACHABLE in a later
commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75045 91177308-0d34-0410-b5e6-96231b3b80d8

Implement changes from Chris's feedback.
Finish converting lib/Target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75043 91177308-0d34-0410-b5e6-96231b3b80d8

Push LLVMContext _back_ through IRBuilder.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75040 91177308-0d34-0410-b5e6-96231b3b80d8

Switch all the MC tests to use FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75039 91177308-0d34-0410-b5e6-96231b3b80d8

convert to FileCheck style.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75038 91177308-0d34-0410-b5e6-96231b3b80d8

Implement NEON vst1 instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75037 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize opcode selection in ARMBaseRegisterInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75036 91177308-0d34-0410-b5e6-96231b3b80d8

Fix cmake build.
Added ARMBaseRegisterInfo.cpp to lib/Target/ARM/CMakeLists.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75035 91177308-0d34-0410-b5e6-96231b3b80d8

Support MDNode forward reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75031 91177308-0d34-0410-b5e6-96231b3b80d8

Make the code that updates ScalarEvolution's internal state in response
to a loop deletion more thorough. Don't prune the def-use tree search at
instructions that don't have SCEVs computed, because an instruction with
a user that has a computed SCEV may itself lack a computed SCEV.  Also,
remove loop-related values from the ValuesAtScopes and
ConstantEvolutionLoopExitValues maps as well.

This fixes a regression in 483.xalancbmk.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75030 91177308-0d34-0410-b5e6-96231b3b80d8

Tell ScalarEvolution to forget a loop before starting to delete it.
This way ScalarEvolution can examine the loop to determine what state
it needs to update, if it chooses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75029 91177308-0d34-0410-b5e6-96231b3b80d8

Convert more abort() calls to llvm_report_error().
Also remove trailing semicolon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75027 91177308-0d34-0410-b5e6-96231b3b80d8

Switch GlobalVariable ctors to a sane API, where *either* a context or a module is required.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75025 91177308-0d34-0410-b5e6-96231b3b80d8

add some more check for vector compares.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75024 91177308-0d34-0410-b5e6-96231b3b80d8

convert a test to "FileCheck" style.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75023 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new little "FileCheck" utility for regression testing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75022 91177308-0d34-0410-b5e6-96231b3b80d8

Push methods into base class in preparation for sharing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75020 91177308-0d34-0410-b5e6-96231b3b80d8

Implement NEON vld1 instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75019 91177308-0d34-0410-b5e6-96231b3b80d8

Start converting to new error handling API.
cerr+abort -> llvm_report_error
assert(0)+abort -> LLVM_UNREACHABLE (assert(0)+llvm_unreachable-> abort() included)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75018 91177308-0d34-0410-b5e6-96231b3b80d8

Conform...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75017 91177308-0d34-0410-b5e6-96231b3b80d8

Start breaking out common base functionality for register info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75016 91177308-0d34-0410-b5e6-96231b3b80d8

convert comments to doxygen style

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75015 91177308-0d34-0410-b5e6-96231b3b80d8

Add rev16 test... xfail for now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75012 91177308-0d34-0410-b5e6-96231b3b80d8

Add myself to blame file...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75011 91177308-0d34-0410-b5e6-96231b3b80d8

Checkpoint Thumb2 Instr info work. Generalized base code so that it can be shared between ARM and Thumb2. Not yet activated because register information must be generalized first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75010 91177308-0d34-0410-b5e6-96231b3b80d8

Remove trailing whitespace.  Reorder some methods
and cases alphabetically.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75001 91177308-0d34-0410-b5e6-96231b3b80d8

Update the example to show that an archive can contain llvm bitcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@75000 91177308-0d34-0410-b5e6-96231b3b80d8

Trial first commit
Fixed build, requires std header before using std::string

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74995 91177308-0d34-0410-b5e6-96231b3b80d8

Mark sublw_cc and subfw_cc as Terminator insns so that they are part of the terminator insns for a basic block alongwith branch insns. This way a copy is not getting inserted between cmp and branch during PHIElimination disturbing the status flags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74992 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the vicmp and vfcmp instructions. Because we never had a release with
these instructions, no autoupgrade or backwards compatibility support is
provided.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74991 91177308-0d34-0410-b5e6-96231b3b80d8

Add a Thumb2 instruction flag to that indicates whether the instruction can be transformed to 16-bit variant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74988 91177308-0d34-0410-b5e6-96231b3b80d8

Push LLVMContext through GlobalVariables and IRBuilder.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74985 91177308-0d34-0410-b5e6-96231b3b80d8

remove two methods that no longer exist.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74982 91177308-0d34-0410-b5e6-96231b3b80d8

more getting windows to build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74981 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the v[if]cmp versions of these tests, now that [if]cmp+sext works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74980 91177308-0d34-0410-b5e6-96231b3b80d8

Change these tests to use [fi]cmp+sext instead of v[fi]cmp.  No
functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74979 91177308-0d34-0410-b5e6-96231b3b80d8

dag combine sext(setcc) -> vsetcc before legalize.  To make this safe,
VSETCC must define all bits, which is different than it was documented
to before.  Since all targets that implement VSETCC already have this
behavior, and we don't optimize based on this, just change the
documentation.  We now get nice code for vec_compare.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74978 91177308-0d34-0410-b5e6-96231b3b80d8

hopefully fix the build on windows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74977 91177308-0d34-0410-b5e6-96231b3b80d8

Add a todo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74976 91177308-0d34-0410-b5e6-96231b3b80d8

Also statically set bit 25 for BR_JT instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74974 91177308-0d34-0410-b5e6-96231b3b80d8

LLVMContext-ification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74973 91177308-0d34-0410-b5e6-96231b3b80d8

Statically encode bit 25 to indicate immediate form of data processing instructions. Patch by Sean Callanan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74972 91177308-0d34-0410-b5e6-96231b3b80d8

--- Reverse-merging (from foreign repository) r74952 into '.':
U    lib/Target/X86/X86RegisterInfo.cpp
U    lib/Target/X86/X86RegisterInfo.h

Temporarily revert. This was causing an infinite loop in the linker on Leopard.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74970 91177308-0d34-0410-b5e6-96231b3b80d8