Significantly simplify this code and make it more aggressive.  Instead of having
a special case hack for X86, make the hack more general: if an incoming argument
register is not used in any block other than the entry block, don't copy it to
a vreg.  This helps us compile code like this:

%struct.foo = type { int, int, [0 x ubyte] }
int %test(%struct.foo* %X) {
        %tmp1 = getelementptr %struct.foo* %X, int 0, uint 2, int 100
        %tmp = load ubyte* %tmp1                ; <ubyte> [#uses=1]
        %tmp2 = cast ubyte %tmp to int          ; <int> [#uses=1]
        ret int %tmp2
}

to:

_test:
        lbz r3, 108(r3)
        blr

instead of:

_test:
        lbz r2, 108(r3)
        or r3, r2, r2
        blr

The (dead) copy emitted to copy r3 into a vreg for extra-block uses was
increasing the live range of r3 past the load, preventing the coallescing.

This implements CodeGen/PowerPC/reg-coallesce-simple.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24115 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24112 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce the number of copies emitted as machine instructions by
generating results in vregs that will need them.  In the case of something
like this:  CopyToReg((add X, Y), reg1024), we no longer emit code like
this:

   reg1025 = add X, Y
   reg1024 = reg 1025

Instead, we emit:

   reg1024 = add X, Y

Whoa! :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24111 91177308-0d34-0410-b5e6-96231b3b80d8

If the module has no t-t and the host is an alpha, default to using the Alpha BE

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24110 91177308-0d34-0410-b5e6-96231b3b80d8

fix some broken comparisons, this affected the Pattern isel too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24109 91177308-0d34-0410-b5e6-96231b3b80d8

The pass was removed, but the spirit lives on

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24108 91177308-0d34-0410-b5e6-96231b3b80d8

This is implemented

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24107 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen mul by negative power of two with a shift and negate.
This implements test/Regression/CodeGen/PowerPC/mul-neg-power-2.ll,
producing:

_foo:
        slwi r2, r3, 1
        subfic r3, r2, 63
        blr

instead of:

_foo:
        mulli r2, r3, -2
        addi r3, r2, 63
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24106 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24103 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a problem that Nate noticed with LSR:

When inserting code for an addrec expression with a non-unit stride, be
more careful where we insert the multiply.  In particular, insert the multiply
in the outermost loop we can, instead of the requested insertion point.

This allows LSR to notice the mul in the right loop, reducing it when it gets
to it.  This allows it to reduce the multiply, where before it missed it.

This happens quite a bit in the test suite, for example, eliminating 2
multiplies in art, 3 in ammp, 4 in apsi, reducing from 1050 multiplies to
910 muls in galgel (!), from 877 to 859 in applu, and 36 to 30 in bzip2.

This speeds up galgel from 16.45s to 16.01s, applu from 14.21 to 13.94s and
fourinarow from 66.67s to 63.48s.

This implements Transforms/LoopStrengthReduce/nested-reduce.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24102 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24099 91177308-0d34-0410-b5e6-96231b3b80d8

Revert an accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24098 91177308-0d34-0410-b5e6-96231b3b80d8

More files deleted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24097 91177308-0d34-0410-b5e6-96231b3b80d8

Make -time-passes output prettier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24096 91177308-0d34-0410-b5e6-96231b3b80d8

Make negative immediates in patterns work correctly, silence some warnings
building the itanium backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24095 91177308-0d34-0410-b5e6-96231b3b80d8

add some FP stuff, some mix.* stuff, and constant pool support to the
DAG instruction selector, which should be destroyed one day (in the pattern
isel also) since ia64 can pack any constant in the instruction stream

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24094 91177308-0d34-0410-b5e6-96231b3b80d8

Fix <> in html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24093 91177308-0d34-0410-b5e6-96231b3b80d8

First hack-and-slash on the release notes.  Comments and improvements welcome

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24092 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the lowerconstantexprs pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24091 91177308-0d34-0410-b5e6-96231b3b80d8

This pass is now obsolete since all targets have moved to the SelectionDAG
infrastructure and the simple isels have been removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24090 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the LowerConstantExpressionsPass pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24089 91177308-0d34-0410-b5e6-96231b3b80d8

remove reference to this pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24088 91177308-0d34-0410-b5e6-96231b3b80d8

Basic updates

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24087 91177308-0d34-0410-b5e6-96231b3b80d8

Buh bye Expression.(cpp|h)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24086 91177308-0d34-0410-b5e6-96231b3b80d8

remove a dead file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24085 91177308-0d34-0410-b5e6-96231b3b80d8

remove a dead header

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24084 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24083 91177308-0d34-0410-b5e6-96231b3b80d8

Now that instcombine does this xform, remove it from the -raise pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24082 91177308-0d34-0410-b5e6-96231b3b80d8

Pull some code out into a function, give it the ability to see through +.
This allows us to turn code like malloc(4*x+4) -> malloc int, (x+1)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24081 91177308-0d34-0410-b5e6-96231b3b80d8

add shladd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24080 91177308-0d34-0410-b5e6-96231b3b80d8

test for the formation of shladd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24077 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a special case, allowing the general case to handle it.  No functionality
change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24076 91177308-0d34-0410-b5e6-96231b3b80d8

New case to handle someday

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24075 91177308-0d34-0410-b5e6-96231b3b80d8

Switch more code over to using getValueAsListOfDefs.  Look at all the -'s. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24074 91177308-0d34-0410-b5e6-96231b3b80d8

Don't emit "32" for unordered comparison

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24073 91177308-0d34-0410-b5e6-96231b3b80d8

Rename Record::getValueAsListDef to getValueAsListOfDefs, to more accurately
reflect what it is.
Convert some more code over to use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24072 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new interface Jim added

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24071 91177308-0d34-0410-b5e6-96231b3b80d8

Removed Mr. Smith from the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24070 91177308-0d34-0410-b5e6-96231b3b80d8

Added method to return a vector of records for a ListInit of Def field.  This
simplifies using list of records.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24069 91177308-0d34-0410-b5e6-96231b3b80d8

add a hack to get code with ordered comparisons working.  This hack is
tracked as PR642

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24068 91177308-0d34-0410-b5e6-96231b3b80d8

add support for branch on ordered/unordered.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24067 91177308-0d34-0410-b5e6-96231b3b80d8

add the xfail lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24066 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase.  Probably many targets don't support this, so they should probably
add themselves as xfails until they do (at least for the release).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24065 91177308-0d34-0410-b5e6-96231b3b80d8

Do not globalize internal symbols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24064 91177308-0d34-0410-b5e6-96231b3b80d8

These are autogenerated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24063 91177308-0d34-0410-b5e6-96231b3b80d8

DAG->DAG instruction selection for ia64! "hello world" works, not much else.
use -enable-ia64-dag-isel to turn this on

TODO: delete lowering stuff from the pattern isel
    : get operations on predicate bits working
    : get other bits of pseudocode going
    : use sampo's mulh/mull-using divide-by-constant magic
    : *so* many patterns ("extr", "tbit" and "dep" will be fun :)
    : add FP
    : add a JIT!
    : get it working 100%

in short: this'll be happier in a couple of weeks, but it's here now so
the tester can make me feel guilty sooner.

OTHER: there are a couple of fixes to the pattern isel, in particular
making the linker happy with big blobs of fun like pypy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24058 91177308-0d34-0410-b5e6-96231b3b80d8

The nightly tester report doesn't report JIT code size anymore, remove it
from the olden graph.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24057 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bit of backwards logic that broke exptree and smg2000

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24056 91177308-0d34-0410-b5e6-96231b3b80d8

Add some commentary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24055 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24054 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate getClass, it is not needed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24053 91177308-0d34-0410-b5e6-96231b3b80d8

Keep Visual Studio happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24052 91177308-0d34-0410-b5e6-96231b3b80d8

a bad case for bitfield insert

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24051 91177308-0d34-0410-b5e6-96231b3b80d8

Now generating instruction itineraries for scheduling.  Not my best work, but...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24050 91177308-0d34-0410-b5e6-96231b3b80d8

Structures used to hold scheduling information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24049 91177308-0d34-0410-b5e6-96231b3b80d8

update project file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24048 91177308-0d34-0410-b5e6-96231b3b80d8

Do not sink any instruction with side effects, including vaarg.  This fixes
PR640

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24046 91177308-0d34-0410-b5e6-96231b3b80d8

Fix #include order

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24044 91177308-0d34-0410-b5e6-96231b3b80d8

unbreak the build again

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24041 91177308-0d34-0410-b5e6-96231b3b80d8

Move some constant folding functions into LLVMAnalysis since they are used
by Analysis and Transformation passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24038 91177308-0d34-0410-b5e6-96231b3b80d8

Move some constant folding code shared by Analysis and Transform passes
into the LLVMAnalysis library.
This allows LLVMTranform and LLVMTransformUtils to be archives and linked
with LLVMAnalysis.a, which provides any missing definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24036 91177308-0d34-0410-b5e6-96231b3b80d8

Fix DSE to not nuke dead stores unless they redundant store is the same
VT as the killing one.  Fix fixes PR491

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24034 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24033 91177308-0d34-0410-b5e6-96231b3b80d8

Teach instcombine to promote stuff like (cast (malloc sbyte, 8*X) to int*)
into: malloc int, (2*X)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24032 91177308-0d34-0410-b5e6-96231b3b80d8

Promote cases like cast (malloc sbyte, 100) to int* into
(malloc [25 x int]) directly without having to convert to
(malloc [100 x sbyte]) first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24031 91177308-0d34-0410-b5e6-96231b3b80d8

Minor change to this file to support obscure cases with constant array amounts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24030 91177308-0d34-0410-b5e6-96231b3b80d8

Add a simple xform that is useful for bitfield operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24029 91177308-0d34-0410-b5e6-96231b3b80d8

Fine tune Visual Studio's use of bison/flex.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24025 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure to build things in the right order, build with the .o file, not the
.a file to unbreak the build after john's change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24024 91177308-0d34-0410-b5e6-96231b3b80d8

1. Remove libraries no longer created from the list of libraries linked into the
   SparcV9 JIT.
2. Make LLVMTransformUtils a relinked object file and always link it before
   LLVMAnalysis.a.  These two libraries have circular dependencies on each
   other which creates problem when building the SparcV9 JIT.  This change
   fixes the dependency on all platforms problems with a minimum of fuss.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24023 91177308-0d34-0410-b5e6-96231b3b80d8

int comparison patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24020 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some spello's pointed out by Gabor Greif

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24019 91177308-0d34-0410-b5e6-96231b3b80d8

Typo made worse x 2 - take 2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24018 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an assert compiling MallocBench/gs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24017 91177308-0d34-0410-b5e6-96231b3b80d8

Typo x 2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24016 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24015 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify instinfo, set random bits on more fp insts, and fix 1 opcode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24014 91177308-0d34-0410-b5e6-96231b3b80d8

Give full control of subtarget features over to table generated code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24013 91177308-0d34-0410-b5e6-96231b3b80d8

Add attribute name and type to SubtargetFeatures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24012 91177308-0d34-0410-b5e6-96231b3b80d8

fold nested and's early to avoid inefficiencies in MaskedValueIsZero.  This
fixes a very slow compile in PR639.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24011 91177308-0d34-0410-b5e6-96231b3b80d8

Condcodes are in the ISD namespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24010 91177308-0d34-0410-b5e6-96231b3b80d8

Add nodes for CondCodeSDNode and setcc, and add a bunch of pattern fragments
to make it easy to use them.  This lets you write patterns like:

(set PRRC:$rd, (setne GPRC:$rS, imm:$SH))

and stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24009 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for CondCode's

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24008 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24004 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate use of sed in Visual Studio builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24003 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate need for bison/flex in Visual Studio builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23999 91177308-0d34-0410-b5e6-96231b3b80d8

Update Visual Studio projects to reflect moved file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23998 91177308-0d34-0410-b5e6-96231b3b80d8

Add a regression test for the recent fix for FP_TO_UINT lowering in the ppc
backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23995 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note about some bitfield stuff we could be doing better.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23994 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly Expand or Promote FP_TO_UINT based on the capabilities of the
machine.  This allows us to generate great code for i32 FP_TO_UINT now on
targets with 64 bit extensions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23993 91177308-0d34-0410-b5e6-96231b3b80d8

Allow custom lowered FP_TO_SINT ops in the check for whether a larger
FP_TO_SINT is preferred to a larger FP_TO_UINT.  This seems to be begging
for a TLI.isOperationCustom() helper function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23992 91177308-0d34-0410-b5e6-96231b3b80d8

autogen undef

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23991 91177308-0d34-0410-b5e6-96231b3b80d8

Add undef

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23990 91177308-0d34-0410-b5e6-96231b3b80d8

Add a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23989 91177308-0d34-0410-b5e6-96231b3b80d8

Allow pseudos to have patterns, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23988 91177308-0d34-0410-b5e6-96231b3b80d8

Autogen fsel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23987 91177308-0d34-0410-b5e6-96231b3b80d8

Expose the fextend on the DAG instead of doing it in the matcher

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23986 91177308-0d34-0410-b5e6-96231b3b80d8

Autogen a few new ppc-specific nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23985 91177308-0d34-0410-b5e6-96231b3b80d8

The dag isel generator generates this now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23984 91177308-0d34-0410-b5e6-96231b3b80d8

Emit some boilerplate for targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23983 91177308-0d34-0410-b5e6-96231b3b80d8