Restore this assert now that the livevar bug is fixed.
This verifies kill info for "ret" fp operands is right.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48656 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR2160.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48655 91177308-0d34-0410-b5e6-96231b3b80d8

A couple of kill marker maintainence bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48653 91177308-0d34-0410-b5e6-96231b3b80d8

clean up and fix the call section.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48651 91177308-0d34-0410-b5e6-96231b3b80d8

cleanups for the getresult section.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48650 91177308-0d34-0410-b5e6-96231b3b80d8

FunctionExtractorPass has been superceded by GVExtractorPass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48648 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to get an empty struct using
the new StructType::get method.  The second NULL
is to pacify the gcc warning mechanism.  This
patch compiles but is otherwise untested.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48645 91177308-0d34-0410-b5e6-96231b3b80d8

Introduce a new node for holding call argument
flags.  This is needed by the new legalize types
infrastructure which wants to expand the 64 bit
constants previously used to hold the flags on
32 bit machines.  There are two functional changes:
(1) in LowerArguments, if a parameter has the zext
attribute set then that is marked in the flags;
before it was being ignored; (2) PPC had some bogus
code for handling two word arguments when using the
ELF 32 ABI, which was hard to convert because of
the bogusness.  As suggested by the original author
(Nicolas Geoffray), I've disabled it for the moment.
Tested with "make check" and the Ada ACATS testsuite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48640 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the build for gcc-4.2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48639 91177308-0d34-0410-b5e6-96231b3b80d8

remove Evan's "ugly hack" that sorta attempted to get
x86-64 return conventions correct, but was never enabled.
We can now do the "right thing" with multiple return values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48635 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for calls that return two FP values in
ST(0)/ST(1).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48634 91177308-0d34-0410-b5e6-96231b3b80d8

disable a bogus assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48633 91177308-0d34-0410-b5e6-96231b3b80d8

Enable support for returning two long-double values in ST(0)/ST(1).
This allows us to compile fp-stack-2results.ll into:

_test:
fldz
fld1
ret

which returns 1 in ST(0) and 0 in ST(1).  This is needed for x86-64
_Complex long double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48632 91177308-0d34-0410-b5e6-96231b3b80d8

Teach masked value is zero about add and sub, and use MVIZ to
simplify things like (X & 4) >> 1 == 2  --> (X & 4) == 4.

since it is obvious that the shift doesn't remove any bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48631 91177308-0d34-0410-b5e6-96231b3b80d8

Undo 48570. Correctly match mmx shift instructions with an immediate operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48627 91177308-0d34-0410-b5e6-96231b3b80d8

ignore the Output/ directory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48626 91177308-0d34-0410-b5e6-96231b3b80d8

this directory is accidentally added

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48625 91177308-0d34-0410-b5e6-96231b3b80d8

These passes preserve CFG.
This patch fixes Benchmarks/Trimaran/enc-pc1/enc-pc1 failure reported by Grawp-PIC i386 nightly tester

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48623 91177308-0d34-0410-b5e6-96231b3b80d8

Incorporate feedback.
- Fix loop nest.
- Use RetVals.size()
- Check for null return value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48605 91177308-0d34-0410-b5e6-96231b3b80d8

C and Objective Caml bindings for mem2reg and reg2mem.

Patch by Erick Tryzelaar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48602 91177308-0d34-0410-b5e6-96231b3b80d8

enable -analyze for andersens

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48601 91177308-0d34-0410-b5e6-96231b3b80d8

Take the old function's name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48588 91177308-0d34-0410-b5e6-96231b3b80d8

ubyte and sbyte? what are those?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48585 91177308-0d34-0410-b5e6-96231b3b80d8

add a note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48583 91177308-0d34-0410-b5e6-96231b3b80d8

Check even more carefully before applying this DAGCombine transform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48580 91177308-0d34-0410-b5e6-96231b3b80d8

Restore isCFGOnly property of various analysis passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48579 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this xform: (sra (shl X, m), result_size) -> (sign_extend (trunc (shl X, result_size - n - m)))

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48578 91177308-0d34-0410-b5e6-96231b3b80d8

detabify llvm, patch by Mike Stump!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48577 91177308-0d34-0410-b5e6-96231b3b80d8

Keep track of analysis information inherited from Module pass manager.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48576 91177308-0d34-0410-b5e6-96231b3b80d8

Add more patterns to match in the integer comparison test harnesses.
Fix bugs encountered, mostly due to range matching for immediates;
the CellSPU's 10-bit immediates are sign extended, covering a
larger range of unsigned values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48575 91177308-0d34-0410-b5e6-96231b3b80d8

80 col violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48573 91177308-0d34-0410-b5e6-96231b3b80d8

Add intrinsics to match mmx shift builtin's with immediate operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48569 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48567 91177308-0d34-0410-b5e6-96231b3b80d8

style and spelling

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48562 91177308-0d34-0410-b5e6-96231b3b80d8

#if 1  .. #endif markers do not add any value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48560 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48556 91177308-0d34-0410-b5e6-96231b3b80d8

PassInfo keep tracks whether a pass is an analysis pass or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48554 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for multiple return values for the PPC target by
converting call result lowering to use the CallingConvLowering
infastructure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48552 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-ld deserves disable-verify too.  opt shouldn't have all the fun options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48550 91177308-0d34-0410-b5e6-96231b3b80d8

Don't loose incoming argument registers. Fix documentation style.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48545 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48543 91177308-0d34-0410-b5e6-96231b3b80d8

Fix X86's isTruncateFree to not claim that truncate to i1 is free. This fixes Bill's testcase that failed for r48491.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48542 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48538 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48536 91177308-0d34-0410-b5e6-96231b3b80d8

add some convenience methods for creating GEP instructions and
struct types.  Patch by David Chisnall, with some tweaks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48531 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests to not use llvm-upgrade.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48530 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests to not use llvm-upgrade.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48529 91177308-0d34-0410-b5e6-96231b3b80d8

C bindings for Module-, Function-, and BasicBlock::iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48528 91177308-0d34-0410-b5e6-96231b3b80d8

Remove llvm-upgrade and update tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48527 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a coalescer bug caused by a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48526 91177308-0d34-0410-b5e6-96231b3b80d8

C and Objective Caml bindings for the various getParent methods of the IR.

Based on Erick Tryzelaar's patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48523 91177308-0d34-0410-b5e6-96231b3b80d8

Fix live variables issues:
1. If part of a register is re-defined, an implicit kill and an implicit def are added to denote read / mod / write. However, this should only be necessary if the register is actually read later. This is a performance issue.
2. If a sub-register is being defined, and it doesn't have a previous use, do not add a implicit kill to the last use of a super-register:
   = EAX, AX<imp-use,kill>
...
AX =
In this case, EAX is live but AX is killed, this is wrong and will cause the coalescer to do bad things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48521 91177308-0d34-0410-b5e6-96231b3b80d8

Do not use virtual function to identify an analysis pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48520 91177308-0d34-0410-b5e6-96231b3b80d8

On Darwin, GCC issues a ".globl" for something that has a "visibility protected"
attribute instead of ".protected".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48516 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a x86-64 isel lowering bug that's been around forever. A x86-64 varargs function implicitly reads X86::AL, don't clobber it!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48515 91177308-0d34-0410-b5e6-96231b3b80d8

Note that gcc 4.0.1 will trip internal LLVM asserts (e.g., pred_iterator) if compiling
for Release with optimization levels greater than -O0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48513 91177308-0d34-0410-b5e6-96231b3b80d8

It might be nice to have this run as x86 on non-x86 platforms...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48511 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily revert r48491. It's breaking  test/CodeGen/X86/xorl.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48510 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR 2160 by making sure arguments to external functions get marked as pointing to anything

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48509 91177308-0d34-0410-b5e6-96231b3b80d8

Do not pass -g flag when compiling tests, so remove the C.Flags. This only happens if you have a debug build of llvm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48498 91177308-0d34-0410-b5e6-96231b3b80d8

Make conversions of i8/i16 to ppcf128 work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48493 91177308-0d34-0410-b5e6-96231b3b80d8

Dial down gcc's warnings: don't use 0UL when 0U suffices (and when the
variables and methods themselves only use unsigned.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48492 91177308-0d34-0410-b5e6-96231b3b80d8

Target independent DAG transform to use truncate for field extraction + sign extend on targets where this is profitable. Passes nightly on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48491 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite code that propagate isDead information after a dead copy is coalesced. This remove some ugly spaghetti code and fixed a number of subtle bugs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48490 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests to not use llvm-upgrade.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48484 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade tests to not use llvm-upgrade.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48483 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of compilation warnings.  Per Devang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48478 91177308-0d34-0410-b5e6-96231b3b80d8

Identify Analysis pass.
Do not run analysis pass again if analysis info is still available.
This fixes PR1441.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48476 91177308-0d34-0410-b5e6-96231b3b80d8

Update heuritics that estimates cost of call instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48474 91177308-0d34-0410-b5e6-96231b3b80d8

Make Complex long long/double/long double work
in ppc64 mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48459 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify the unwind attribute.  Add assumption
about sret made by the optimizers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48452 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for "make install" of ocaml docs. Patch by Erick Tryzelaar!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48451 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for escaping {} in asm strings, based on patch from Nick Burns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48450 91177308-0d34-0410-b5e6-96231b3b80d8

Check in some #ifdef'd out code switching call argument
lowering over to SparcCallingConv.td.  We can't make the switch
yet because we can't say to pass f64 registers in 2 x i32 registers
with the td file yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48449 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48448 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak JIT. Ignore TargetInstrInfo::IMPLICIT_DEF.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48447 91177308-0d34-0410-b5e6-96231b3b80d8

A first attempt at updating live intervals, with code lifted from
the coalescer.  This doesn't really work, but gets us farther than
before.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48446 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48445 91177308-0d34-0410-b5e6-96231b3b80d8

Switch sparc from using LowerCallTo to using LowerOperation(CALL) like
other targets.  Use autogenerated calling conv to lower result of
calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48444 91177308-0d34-0410-b5e6-96231b3b80d8

Start moving sparc to use SparcCallingConv.td, switching over
return lowering first.  This fixes a bug where the top and bottom
of i64 values were returned in the wrong registers before.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48443 91177308-0d34-0410-b5e6-96231b3b80d8

split sparc lowering out into SparcISelLowering.{cpp|h} to follow
best practices.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48442 91177308-0d34-0410-b5e6-96231b3b80d8

refactor the LowerOperation code out to individual functions for
each lowering, which is 'best practice'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48441 91177308-0d34-0410-b5e6-96231b3b80d8

Next round of PPC32 ABI changes.  Allow for gcc
behavior where a callee thinks a param will be
present in memory, even though the ABI doc says
it doesn't have to be.  Handle complex long long
and complex double (4 and 8 return regs).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48439 91177308-0d34-0410-b5e6-96231b3b80d8

various rotate fun.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48438 91177308-0d34-0410-b5e6-96231b3b80d8

ensure we continue matching x86-64 rotates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48437 91177308-0d34-0410-b5e6-96231b3b80d8

__builtin_ia32_movntdqa reads memory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48431 91177308-0d34-0410-b5e6-96231b3b80d8

Add a couple missing SSE4 instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48430 91177308-0d34-0410-b5e6-96231b3b80d8

C and Objective Caml bindings for the TargetData class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48422 91177308-0d34-0410-b5e6-96231b3b80d8

Add assert for non-hexadecimal radixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48421 91177308-0d34-0410-b5e6-96231b3b80d8

Housekeeping.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48420 91177308-0d34-0410-b5e6-96231b3b80d8

C and Objective Caml bindings for several scalar transforms.

Patch originally by Erick Tryzelaar, but has been modified somewhat.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48419 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48418 91177308-0d34-0410-b5e6-96231b3b80d8

All of these tests had out of date syntax and were never even running through
llvm-upgrade because nobody noticed them failing.

Update to use new syntax and actually check for the right failure by looking at
the error message.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48417 91177308-0d34-0410-b5e6-96231b3b80d8

Functions are allowed to return structures. (Note that this test never failed.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48416 91177308-0d34-0410-b5e6-96231b3b80d8

Regressions/ is long gone.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48415 91177308-0d34-0410-b5e6-96231b3b80d8

Commit works on regular functions too. Fix the syntax to allow @foo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48414 91177308-0d34-0410-b5e6-96231b3b80d8

C and Objective Caml bindings for PassManagers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48413 91177308-0d34-0410-b5e6-96231b3b80d8

Make insert_subreg a two-address instruction, vastly simplifying LowerSubregs pass. Add a new TII, subreg_to_reg, which is like insert_subreg except that it takes an immediate implicit value to insert into rather than a register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48412 91177308-0d34-0410-b5e6-96231b3b80d8

Minor documentation fix.

Patch by Erick Tryzelaar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48382 91177308-0d34-0410-b5e6-96231b3b80d8

Remove isImplicitDef TargetInstrDesc flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48381 91177308-0d34-0410-b5e6-96231b3b80d8