Move the handling of ANY_EXTEND, SIGN_EXTEND_INREG, and TRUNCATE
out of X86ISelDAGToDAG.cpp C++ code and into tablegen code.
Among other things, using tablegen for these things makes them
friendlier to FastISel.

Tablegen can handle the case of i8 subregs on x86-32, but currently
the C++ code for that case uses MVT::Flag in a tricky way, and it
happens to schedule better in some cases. So for now, leave the
C++ code in place to handle the i8 case on x86-32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55078 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the BuildMI calls even more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55077 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify FastISel's constructor argument list, make the FastISel
class hold a MachineRegisterInfo member, and make the
MachineBasicBlock be passed in to SelectInstructions rather
than the FastISel constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55076 91177308-0d34-0410-b5e6-96231b3b80d8

Dump the instruction that foiled ISel even when -debug is not used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55075 91177308-0d34-0410-b5e6-96231b3b80d8

Don't hoist instructions that define a physical register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55074 91177308-0d34-0410-b5e6-96231b3b80d8

Make more use of the BuildMI API.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55072 91177308-0d34-0410-b5e6-96231b3b80d8

Minor code reorganization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55071 91177308-0d34-0410-b5e6-96231b3b80d8

Minor whitespace cleanup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55070 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 column violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55069 91177308-0d34-0410-b5e6-96231b3b80d8

Kill off SimpleBBISel, it's replaced by FastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55067 91177308-0d34-0410-b5e6-96231b3b80d8

Move the fast-path (<=i64) cases of various APInt methods inline
and the slow-path cases out of line.  This speeds up instcombine
a bit in real world cases.  Patch contributed by m-s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55063 91177308-0d34-0410-b5e6-96231b3b80d8

Disable DAGCombine's alignment inference in "fast" codegen mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55059 91177308-0d34-0410-b5e6-96231b3b80d8

Change the FoldingSetNodeID usage for objects which carry
alignment and volatility information, such as loads and
stores, to reduce the number of integer values added to
the FoldingSetNodeID.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55058 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up a dead return missed in r55055.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55057 91177308-0d34-0410-b5e6-96231b3b80d8

Tablegen generated code already tests the opcode value, so it's not
necessary to use dyn_cast in these predicates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55055 91177308-0d34-0410-b5e6-96231b3b80d8

Use BitVector instead of std::vector<unsigned char>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55054 91177308-0d34-0410-b5e6-96231b3b80d8

It's not necessary to check if a value is null before delete[].

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55053 91177308-0d34-0410-b5e6-96231b3b80d8

Use cast instead of dyn_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55052 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid an empty-if-body warning in release builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55050 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55049 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment spacing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55047 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed shuffle optimizations to handle non power of 2 vectors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55035 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the string for MVT::isVoid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55034 91177308-0d34-0410-b5e6-96231b3b80d8

Fix FastISel to recognize that the last block in the function does
not have a fall-through successor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55033 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a leak in the FastISel code that Chris pointed out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55031 91177308-0d34-0410-b5e6-96231b3b80d8

Add a -march line for this test, and run it on x86-64 too for fun.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55030 91177308-0d34-0410-b5e6-96231b3b80d8

Add remaining 64-bit atomic patterns for x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55029 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for running SelectionDAG if FastISel fails. This is under
a command-line option, so that the default behavior is an abort, which
is useful for exposing code that isn't supported yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55028 91177308-0d34-0410-b5e6-96231b3b80d8

Fix FastISel to recognize unhandled operands, such as constants
that aren't available as virtual registers (for now).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55026 91177308-0d34-0410-b5e6-96231b3b80d8

ll printer prints each zeroinit on the same line now, update regex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55025 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r55018 and apply the correct "fix" for the 64-bit sub_and_fetch atomic.
Just expand it like the other X-bit sub_and_fetches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55023 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of a couple of dynamic_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55022 91177308-0d34-0410-b5e6-96231b3b80d8

Add FastISel support for floating-point operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55021 91177308-0d34-0410-b5e6-96231b3b80d8

Add FastISel support for several more binary operators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55020 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for the __sync_sub_and_fetch atomics and friends for X86. The code
was already present, but not hooked up to anything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55018 91177308-0d34-0410-b5e6-96231b3b80d8

Fast-isel is now *minimally* functional. Add a testcase to
demonstrate the extent of its capabilities. Note that it
only attempts to operate on one of the blocks in this
testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55016 91177308-0d34-0410-b5e6-96231b3b80d8

Add code to call FastISel, and a command-line option to enable it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55015 91177308-0d34-0410-b5e6-96231b3b80d8

Support unconditional fall-through branches in FastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55014 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the fast-path spilling code to attempt folding, but still leaving  out remat and splitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55012 91177308-0d34-0410-b5e6-96231b3b80d8

Instantiate FastISel for X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55011 91177308-0d34-0410-b5e6-96231b3b80d8

The X86 target will soon have an implementation of createFastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55010 91177308-0d34-0410-b5e6-96231b3b80d8

Add a TargetLowering hook for creating a FastISel object.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55009 91177308-0d34-0410-b5e6-96231b3b80d8

For now, restrict FastISel to instructions that only involve one
register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55008 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out the code to scan an instruction's operands into a
helper function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55007 91177308-0d34-0410-b5e6-96231b3b80d8

Use the BuildMI overload that sets up a destination register
instead of the one that doesn't and then adding it manually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55006 91177308-0d34-0410-b5e6-96231b3b80d8

Handle the case where target-specific fastisel code doesn't have
a desired opcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55005 91177308-0d34-0410-b5e6-96231b3b80d8

Add more comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55004 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation in FastISel tablegen-emitted code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55003 91177308-0d34-0410-b5e6-96231b3b80d8

Add more checking to filter out more kinds of things that
FastISel doesn't support yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55002 91177308-0d34-0410-b5e6-96231b3b80d8

remove reference to really really old tree

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55001 91177308-0d34-0410-b5e6-96231b3b80d8

The fast-path still needs to set kill markers and spill/restore points as appropriate.
With this patch, all of MultiSource/Applications and all of SPEC2000/2006 pass with
the SimpleSpiller and this fast-path enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55000 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for 8 and 16 bit forms of __sync
builtins on X86.

Change "lock" instructions to be on a separate line.
This is needed to work around a bug in the Darwin
assembler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54999 91177308-0d34-0410-b5e6-96231b3b80d8

80 columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54998 91177308-0d34-0410-b5e6-96231b3b80d8

Add a few doxygen comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54997 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an unneeded #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54996 91177308-0d34-0410-b5e6-96231b3b80d8

Delete a dead field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54995 91177308-0d34-0410-b5e6-96231b3b80d8

[PR2327] Leverage TargetRegisterInfo to compute frame offsets for GC metadata.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54994 91177308-0d34-0410-b5e6-96231b3b80d8

Update the JIT exception writer to better mimic the codegen exception writer.
Also skip indirect encoding for platforms that ask for one: we direclty
write an address, not a pointer to the address.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54987 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54985 91177308-0d34-0410-b5e6-96231b3b80d8

more cleanups, random methods shouldn't return ostreams.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54984 91177308-0d34-0410-b5e6-96231b3b80d8

more cleanup, eliminate getLLVMName when printing out
type names at the top of the file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54983 91177308-0d34-0410-b5e6-96231b3b80d8

random cleanups, factor some printing code for linkage and visibility

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54982 91177308-0d34-0410-b5e6-96231b3b80d8

Change WriteTypeSymbolic/WriteAsOperand to return void instead of
an ostream, which is just weird.

Rename SC_DEBUG -> ST_DEBUG

Remove static indentation strangeness from WriteConstantInt.  This makes it
so that large structs are not broken down and printed on multiple lines.  If
there is demand for this to return, there are better ways to implement this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54981 91177308-0d34-0410-b5e6-96231b3b80d8

whitespace cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54980 91177308-0d34-0410-b5e6-96231b3b80d8

don't use the result of WriteAsOperand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54979 91177308-0d34-0410-b5e6-96231b3b80d8

don't use the result of WriteTypeSymbolic or WriteAsOperand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54978 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use the result of WriteAsOperand or WriteTypeSymbolic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54977 91177308-0d34-0410-b5e6-96231b3b80d8

rearrange code and make more legible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54976 91177308-0d34-0410-b5e6-96231b3b80d8

put SlotTracker in an anon namespace since it is private.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54975 91177308-0d34-0410-b5e6-96231b3b80d8

Reid had to call this class "slot machine" out of some strange reference
to the gambling device.  Name it something more meaningful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54974 91177308-0d34-0410-b5e6-96231b3b80d8

add raw_ostream method for emitting an unsigned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54972 91177308-0d34-0410-b5e6-96231b3b80d8

Compress manpages.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54971 91177308-0d34-0410-b5e6-96231b3b80d8

Make SimpleSpiller respect subregister indices.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54968 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54964 91177308-0d34-0410-b5e6-96231b3b80d8

Add a flag to enable the fast spilling path.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54958 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the FastISel class' doxygen comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54957 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a few more bugs:
  1) Assign stack slots to new temporaries.
  2) Don't insert an interval into the return vector more than once.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54956 91177308-0d34-0410-b5e6-96231b3b80d8

remove empty file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54950 91177308-0d34-0410-b5e6-96231b3b80d8

Fix several bugs in the new fast-path:
  1) Remove an incorrect assertion.
  2) Set the stack slot weight properly.
  3) Resize the VirtRegMap when needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54949 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak cpp backend: upgrade output due to change in APInt API

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54942 91177308-0d34-0410-b5e6-96231b3b80d8

Clients of addIntervalForSpills expect the added intervals to be returned sorted by starting index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54939 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug daniel pointed out to me, where asmprinter started
printing ascii code for hex numbers instead of the hex numbers
themselves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54936 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the fast-patch interval spilling by using MachineRegisterInfo::reg_iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54930 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL this test for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54929 91177308-0d34-0410-b5e6-96231b3b80d8

Resurrect some ancient code to add spill ranges without attempting folding, remat, or splitting.  This code has been updated to current APIs
in so far as it compiles and, in theory, works, but does not take advantage of recent advancements.  For instance, it could be improved by using
MachineRegisterInfo::use_iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54924 91177308-0d34-0410-b5e6-96231b3b80d8

Register the frame register function when allocating the JIT,
so that lli works out of the box with -enable-eh.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54920 91177308-0d34-0410-b5e6-96231b3b80d8

ARM asm printer can't handle dwarf info yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54913 91177308-0d34-0410-b5e6-96231b3b80d8

Don't hoist instructions that have implicit uses or defines. Before, it was
hoisting out some "test" and "cmp" instructions. This was obvious badness.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54908 91177308-0d34-0410-b5e6-96231b3b80d8

Make this comment clearer. Instead of using an ambiguous ~ (not) on an icmp
predicate, swap the order of the operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54907 91177308-0d34-0410-b5e6-96231b3b80d8

Consider the case where xor by -1 and xor by 128 have been combined already to
produce an xor by 127.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54906 91177308-0d34-0410-b5e6-96231b3b80d8

Update VC++ project.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54905 91177308-0d34-0410-b5e6-96231b3b80d8

Didn't mean to change this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54904 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a (u)comiss intrinsic lowering bug. It was using anyext which can return junk in higher bits. Patch by Nate Begeman.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54903 91177308-0d34-0410-b5e6-96231b3b80d8

Don't require Registry specializations to define random static variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54902 91177308-0d34-0410-b5e6-96231b3b80d8

Drop an unnecessary include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54901 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54900 91177308-0d34-0410-b5e6-96231b3b80d8

Rename some GC classes so that their roll will hopefully be clearer.

In particular, Collector was confusing to implementors. Several
thought that this compile-time class was the place to implement
their runtime GC heap. Of course, it doesn't even exist at runtime.
Specifically, the renames are:

  Collector               -> GCStrategy
  CollectorMetadata       -> GCFunctionInfo
  CollectorModuleMetadata -> GCModuleInfo
  CollectorRegistry       -> GCRegistry
  Function::getCollector  -> getGC (setGC, hasGC, clearGC)

Several accessors and nested types have also been renamed to be
consistent. These changes should be obvious.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54899 91177308-0d34-0410-b5e6-96231b3b80d8

Make it compile on VC2005:
- update VC projects.
- Add an overload to llvm::Stream for <<, since std::hex and std::dec have type std::ios_base& (*)(std::ios_base&) in VC++. (templating the function don't work, due to ambiguities)
- add ../ on several include in X86/AsmPrinter/

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54898 91177308-0d34-0410-b5e6-96231b3b80d8

getLLVMName is only used for types now, which always pass in LocalPrefix.  Specialize on it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54897 91177308-0d34-0410-b5e6-96231b3b80d8

switch valuemap's from std::map to densemap.  This speeds up llvm-dis
on a stripped kc++ .bc file from 0.83 to 0.77s (8%)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54896 91177308-0d34-0410-b5e6-96231b3b80d8