Implement the first half of PR3290: if there is a store of an
integer to a (transitive) bitcast the alloca and if that integer
has the full size of the alloca, then it clobbers the whole thing.
Handle this by extracting pieces out of the stored integer and
filing them away in the SROA'd elements.

This triggers fairly frequently because the CFE uses integers to
pass small structs by value and the inliner exposes these.  For
example, in kimwitu++, I see a bunch of these with i64 stores to
"%struct.std::pair<std::_Rb_tree_const_iterator<kc::impl_abstract_phylum*>,bool>"

In 176.gcc I see a few i32 stores to "%struct..0anon".

In the testcase, this is a difference between compiling test1 to:

_test1:
subl $12, %esp
movl 20(%esp), %eax
movl %eax, 4(%esp)
movl 16(%esp), %eax
movl %eax, (%esp)
movl (%esp), %eax
addl 4(%esp), %eax
addl $12, %esp
ret

vs:

_test1:
movl 8(%esp), %eax
addl 4(%esp), %eax
ret

The second half of this will be to handle loads of the same form.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61853 91177308-0d34-0410-b5e6-96231b3b80d8

Factor a bunch of code out into a helper method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61852 91177308-0d34-0410-b5e6-96231b3b80d8

use continue to simplify code and reduce nesting, no functionality
change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61851 91177308-0d34-0410-b5e6-96231b3b80d8

Get TargetData once up front and cache as an ivar instead of
requerying it all over the place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61850 91177308-0d34-0410-b5e6-96231b3b80d8

Use the hasAllZeroIndices predicate to simplify some
code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61849 91177308-0d34-0410-b5e6-96231b3b80d8

The coalescer does not coalesce a virtual register to a physical register if any of the physical register's sub-register live intervals overlaps with the virtual register. This is overly conservative. It prevents a extract_subreg from being coalesced away:

v1024 = EDI  // not killed
      =
      = EDI

One possible solution is for the coalescer to examine the sub-register live intervals in the same manner as the physical register. Another possibility is to examine defs and uses (when needed) of sub-registers. Both solutions are too expensive. For now, look for "short virtual intervals" and scan instructions to look for conflict instead.

This is a small win on x86-64. e.g. It shaves 403.gcc by ~80 instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61847 91177308-0d34-0410-b5e6-96231b3b80d8

add a testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61845 91177308-0d34-0410-b5e6-96231b3b80d8

Add patterns to match conditional moves with loads folded
into their left operand, rather than their right. Do this
by commuting the operands and inverting the condition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61842 91177308-0d34-0410-b5e6-96231b3b80d8

Add load-folding table entries for cmovno too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61841 91177308-0d34-0410-b5e6-96231b3b80d8

Define instructions for cmovo and cmovno.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61836 91177308-0d34-0410-b5e6-96231b3b80d8

X86_COND_C and X86_COND_NC are alternate mnemonics for
X86_COND_B and X86_COND_AE, respectively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61835 91177308-0d34-0410-b5e6-96231b3b80d8

Improve support for type-generic vector intrinsics by teaching TableGen how
to handle LLVMMatchType intrinsic parameters, and by adding new subclasses
of LLVMMatchType to match vector types with integral elements that are
either twice as wide or half as wide as the elements of the matched type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61834 91177308-0d34-0410-b5e6-96231b3b80d8

Now that fold-pcmpeqd-0.ll is effectively testing that scheduling helps
avoid the need for spilling, add a new testcase that tests that the
pcmpeqd used for V_SETALLONES is changed to a constant-pool load as
needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61831 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r42653 and forward-port the code that lets INC64_32r be
converted to LEA64_32r in x86's convertToThreeAddress. This
replaces code like this:
   movl  %esi, %edi
   inc   %edi
with this:
   lea   1(%rsi), %edi
which appears to be beneficial.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61830 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
- Add preliminary support for v2i32; load/store generates the right code but
  there's a lot work to be done to make this vector type operational.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61829 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in ComputeLinearIndex computation handling multi-level
aggregate types. Don't increment the current index after reaching
the end of a struct, as it will already be pointing at
one-past-the end. This fixes PR3288.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61828 91177308-0d34-0410-b5e6-96231b3b80d8

Set up DwarfDebug using DebugInfo API.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61822 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot that this was needed for Linux. This should fix the builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61819 91177308-0d34-0410-b5e6-96231b3b80d8

The phi construction algorithm used for interval reconstruction is complicated by
two address instructions.  We need to keep track of things we've processed AS USES
independetly of whether we've processed them as defs.

This fixes all known miscompilations when reconstruction is turned on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61802 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU: Update the README

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61785 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
- Fix bugs 3194, 3195: i128 load/stores produce correct code (although, we
  need to ensure that i128 is 16-byte aligned in real life), and 128 zero-
  extends are supported.
- New td file: SPU128InstrInfo.td: this is where all new i128 support should
  be put in the future.
- Continue to hammer on i64 operations and test cases; ensure that the only
  remaining problem will be i64 mul.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61784 91177308-0d34-0410-b5e6-96231b3b80d8

Delete this test; it's a duplicate of 2006-07-03-schedulers.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61781 91177308-0d34-0410-b5e6-96231b3b80d8

Update these argument lists for the isNormalMemory
argument. This doesn't affect current functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61779 91177308-0d34-0410-b5e6-96231b3b80d8

Use a latency value of 0 for the artificial edges inserted by
AddPseudoTwoAddrDeps. This lets the scheduling infrastructure
avoid recalculating node heights. In very large testcases this
was a major bottleneck. Thanks to Roman Levenstein for finding
this!

As a side effect, fold-pcmpeqd-0.ll is now scheduled better
and it no longer requires spilling on x86-32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61778 91177308-0d34-0410-b5e6-96231b3b80d8

no need to negate the APInt for 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61777 91177308-0d34-0410-b5e6-96231b3b80d8

Change m_ConstantInt and m_SelectCst to take their constant integers
as template arguments instead of as instance variables, exposing more
optimization opportunities to the compiler earlier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61776 91177308-0d34-0410-b5e6-96231b3b80d8

make m_ConstantInt(int64_t) safely match ConstantInt's that are larger than i64.
This fixes an instcombine crash on PR3235.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61775 91177308-0d34-0410-b5e6-96231b3b80d8

Construct subprogram DIEs using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61772 91177308-0d34-0410-b5e6-96231b3b80d8

Construct global variable DIEs using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61771 91177308-0d34-0410-b5e6-96231b3b80d8

Construct compile unit dies using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61768 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a thinko in the grammar for thread_local variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61767 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r61415 and r61484. Duncan was correct that these weren't needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61765 91177308-0d34-0410-b5e6-96231b3b80d8

Don't call setDepthDirty/setHeightDirty when adding an edge
with latency 0, since it doesn't affect the depth or height.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61762 91177308-0d34-0410-b5e6-96231b3b80d8

Extract source location info from DebugInfo.
Add methods to add source location info in a DIE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61761 91177308-0d34-0410-b5e6-96231b3b80d8

Add type DIEs using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61757 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note about passing MVT::Other to getSetCCResultType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61756 91177308-0d34-0410-b5e6-96231b3b80d8

Strength test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61755 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the internalize pass to also internalize
global aliases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61754 91177308-0d34-0410-b5e6-96231b3b80d8

When checking if an Argument escapes, check if
the argument is marked nocapture - no need to
analyze the argument if the answer is already
known!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61753 91177308-0d34-0410-b5e6-96231b3b80d8

Find loop back edges only after empty blocks are eliminated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61752 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for bill's patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61751 91177308-0d34-0410-b5e6-96231b3b80d8

Not having an aliasee is a theoretical possibility.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61745 91177308-0d34-0410-b5e6-96231b3b80d8

Format more neatly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61744 91177308-0d34-0410-b5e6-96231b3b80d8

Remove trailing spaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61743 91177308-0d34-0410-b5e6-96231b3b80d8

Delete unused global aliases with internal linkage.
In fact this also deletes those with linkonce linkage,
however this is currently dead because for the moment
aliases aren't allowed to have this linkage type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61742 91177308-0d34-0410-b5e6-96231b3b80d8

Construct composite type DIE using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61741 91177308-0d34-0410-b5e6-96231b3b80d8

Add classof() methods so that dwarf writer can decide what DIDescriptor is in its hand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61740 91177308-0d34-0410-b5e6-96231b3b80d8

Add <climits>, to get the definition of CHAR_BIT. This should fix
build errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61736 91177308-0d34-0410-b5e6-96231b3b80d8

TargetLowering.h #includes SelectionDAGNodes.h, so it doesn't need its
own OpActionsCapacity magic number; it can just use ISD::BUILTIN_OP_END,
as long as it takes care to round up when needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61733 91177308-0d34-0410-b5e6-96231b3b80d8

Delete an unused variable and simplify the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61732 91177308-0d34-0410-b5e6-96231b3b80d8

s/ConstructType/ConstructTypeDIE/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61731 91177308-0d34-0410-b5e6-96231b3b80d8

make llvm-ld smart enough to link against native libraries that are
not in system library directories by checking -L paths as well.
Patch by Axel Naumann!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61730 91177308-0d34-0410-b5e6-96231b3b80d8

Construct stuct field DIEs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61729 91177308-0d34-0410-b5e6-96231b3b80d8

fix wordo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61728 91177308-0d34-0410-b5e6-96231b3b80d8

Remove redundant ValID::ValID:: scoping (doesn't compile on Windows).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61727 91177308-0d34-0410-b5e6-96231b3b80d8

Construct enumerator DIE using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61726 91177308-0d34-0410-b5e6-96231b3b80d8

Reject PR3281:accepted03.ll with:

llvm-as: accepted03.ll:1:35: invalid unresolved type up reference
declare void @r({ \7, opaque, \10 } %su)
                                  ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61725 91177308-0d34-0410-b5e6-96231b3b80d8

Construct array/vector type DIEs using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61724 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of sentinel insertion in interval reconstruction.  It just masked the
problem, rather than fixing it.  The problem has now been fixed the right way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61723 91177308-0d34-0410-b5e6-96231b3b80d8

reject PR3281:crash11.ll with:
llvm-as: crash11.ll:2:27: function may not return return opaque type
  "xw" = tail call opaque @608(label %31)
                          ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61722 91177308-0d34-0410-b5e6-96231b3b80d8

reject PR3281:crash10.ll with:

llvm-as: crash10.ll:3:35: floating point constant does not have type 'ppc_fp128'
  "dumy" = fcmp ult ppc_fp128 "j",9209.4
                                  ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61721 91177308-0d34-0410-b5e6-96231b3b80d8

reject PR3281:crash09.ll with this diagnostic:

llvm-as: crash09.ll:3:1: self referential type is invalid
type %0
^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61720 91177308-0d34-0410-b5e6-96231b3b80d8

allow opaque undefs.  This resolves PR3282 and fixes
test/Assembler/2005-05-05-OpaqueUndefValues.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61719 91177308-0d34-0410-b5e6-96231b3b80d8

Now, getTag() is used by DwarfWriter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61718 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up #includes, deleting a bunch of unnecessary #includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61715 91177308-0d34-0410-b5e6-96231b3b80d8

Construct basic and derived type DIEs using DebugInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61714 91177308-0d34-0410-b5e6-96231b3b80d8

Handle iAny and fAny types in TreePatternNode::UpdateNodeType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61713 91177308-0d34-0410-b5e6-96231b3b80d8

subsume ConstructPointerType()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61711 91177308-0d34-0410-b5e6-96231b3b80d8

Add the keyword 'default'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61710 91177308-0d34-0410-b5e6-96231b3b80d8

subsume ConstructBasicType().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61709 91177308-0d34-0410-b5e6-96231b3b80d8

Renamed Google Test license file from COPYING to LICENSE.TXT to match LLVM
conventions, per John Criswell.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61708 91177308-0d34-0410-b5e6-96231b3b80d8

squash warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61707 91177308-0d34-0410-b5e6-96231b3b80d8

A few more polygen grammar updates.
 - After GlobalAssign, emit addrspace before global/constant, to follow
   the new syntax.
 - Eliminate "type void", which is now invalid.
 - Fix invalid liblists like [, "foo"].
 - Tweak whitespace in a few places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61706 91177308-0d34-0410-b5e6-96231b3b80d8

set standard ignores

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61703 91177308-0d34-0410-b5e6-96231b3b80d8

Fix spelling in some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61702 91177308-0d34-0410-b5e6-96231b3b80d8

another fix to my previous commit:
* some picky <g> compilers get insulted by const-incorrectness
* respect 80-char limit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61701 91177308-0d34-0410-b5e6-96231b3b80d8

Fix misplaced right parentheses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61699 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate tabs from my previous commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61695 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of the tagging functions and use PointerIntPair.
This means that we have to include an additional header.

This patch should be functionally equivalent. I cannot outrule any performance
degradation, though I do not expect any.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61694 91177308-0d34-0410-b5e6-96231b3b80d8

Don't spew bitcode to standard out if this test
fails, like it is right now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61690 91177308-0d34-0410-b5e6-96231b3b80d8

This test passes again, unXFAIL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61688 91177308-0d34-0410-b5e6-96231b3b80d8

Atom and Core i7 do not have same model number after all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61686 91177308-0d34-0410-b5e6-96231b3b80d8

produce the same diagnostics for vicmp constant exprs as vicmp instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61685 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR3281:crash08.ll with this diagnostic:
llvm-as: crash08.ll:3:15: invalid operand type for instruction
  "qp" = sdiv fp128 0x1, %30
              ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61684 91177308-0d34-0410-b5e6-96231b3b80d8

reject PR3281:crash07.ll with:

llvm-as: crash07.ll:2:32: va_arg requires operand with first class type
  %y = va_arg [52 x <{}>] %43, double (...) sspreq
                               ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61683 91177308-0d34-0410-b5e6-96231b3b80d8

alignment of 0 is not valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61682 91177308-0d34-0410-b5e6-96231b3b80d8

reject undef/zero labels.  This fixes PR3281:crash0[56].ll with these
diagnostics:
llvm-as: crash05.ll:1:14: invalid type for null constant
global label zeroinitializer addrspace (75), section "c"
             ^
llvm-as: crash06.ll:2:14: invalid type for null constant
  udiv label zeroinitializer, @0
             ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61681 91177308-0d34-0410-b5e6-96231b3b80d8

add checking intentionally elided for vfcmp/vicmp since they should really
just be removed.  However, this fixes PR3281:crash04.ll, diagnosing it with:

lvm-as: crash04.ll:2:13: vfcmp requires vector floating point operands
  vfcmp uno double* undef, undef
            ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61680 91177308-0d34-0410-b5e6-96231b3b80d8

diagnose PR3281:crash02.ll with:
llvm-as: crash02.ll:1:62: invalid function return type
declare { <{ <{}>, void ([1898 x { void ()* }], opaque, ...) (), fp128 * }>, opaque } @t ()
                                                             ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61679 91177308-0d34-0410-b5e6-96231b3b80d8

reject PR3281:crash01.ll with:

llvm-as: crash01.ll:1:9: invalid function return type
declare opaque @t()
        ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61678 91177308-0d34-0410-b5e6-96231b3b80d8

tighten up return type check

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61677 91177308-0d34-0410-b5e6-96231b3b80d8

fix PR3281:accepted0[02].ll: represent empty arrays distinctly, and
diagnose attempts to initialize non-empty arrays with them.  This
produces:
llvm-as: accepted02.ll:1:28: invalid empty array initializer
@"o" = global [5 x double] []
                           ^
llvm-as: accepted00.ll:1:32: invalid empty array initializer
@"za" = thread_local global {} []
                               ^
[

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61676 91177308-0d34-0410-b5e6-96231b3b80d8

PR3281:crash00.ll: produce this diagnostic instead of crashing:

@t = global i8 0, align 3
                        ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61675 91177308-0d34-0410-b5e6-96231b3b80d8

Handle weak_extern in the JIT. This fixes
SingleSource/UnitTests/2007-04-25-weak.c in JIT mode. The test
now passes on systems which are able to produce a correct
reference output to compare with.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61674 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
- Teach SPU64InstrInfo.td about the remaining signed comparisons, update tests
  accordingly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61672 91177308-0d34-0410-b5e6-96231b3b80d8

gc is not an obsolete keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61670 91177308-0d34-0410-b5e6-96231b3b80d8

Update polygen grammar for recent language changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61669 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
- Add an 8-bit operation test, which doesn't do much at this point.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61665 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
- Fix (brcond (setq ...)) bug, where BRNZ should have been used vice BRZ.
- Kill unused/unnecessary nodes in SPUNodes.td
- Beef out the i64operations.c test harness to use a lot of unaligned
  loads, test loops and LLVM loop/basic block optimizations; run the
  test harness successfully on real Cell hardware.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61664 91177308-0d34-0410-b5e6-96231b3b80d8

Move the libcall annotating part from doFinalization to doInitialization.
Finalization occurs after all the FunctionPasses in the group have run, which
is clearly not what we want.

This also means that we have to make sure that we apply the right param
attributes when creating a new function.

Also, add a missed optimization: strdup and strndup. NoCapture and
NoAlias return!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@61658 91177308-0d34-0410-b5e6-96231b3b80d8