expose a ctor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23924 91177308-0d34-0410-b5e6-96231b3b80d8

add a proto

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23923 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new LinkAllAnalyses.h header instead of forcing passes to be
in relinked object files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23922 91177308-0d34-0410-b5e6-96231b3b80d8

new header

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23921 91177308-0d34-0410-b5e6-96231b3b80d8

implement some prototypes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23920 91177308-0d34-0410-b5e6-96231b3b80d8

add some prototypes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23919 91177308-0d34-0410-b5e6-96231b3b80d8

move this to the analyze tool

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23918 91177308-0d34-0410-b5e6-96231b3b80d8

Don't invade the system namespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23917 91177308-0d34-0410-b5e6-96231b3b80d8

Use archive versions of these libraries, using the LinkAllPasses header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23916 91177308-0d34-0410-b5e6-96231b3b80d8

There is no need for this to be VC++ only

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23915 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a nasty bug that was causing miscompilation of global variables
on big endian 32-bit targets in some cases (e.g. PPC).  This fixes several
PPC JIT failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23914 91177308-0d34-0410-b5e6-96231b3b80d8

Shrinkify to match llc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23912 91177308-0d34-0410-b5e6-96231b3b80d8

Shrinkify to make --help output look better

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23911 91177308-0d34-0410-b5e6-96231b3b80d8

shrinkify the option name a bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23910 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify this, matching changes in the tblgen emitter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23909 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify this due to changes in the tblgen side

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23908 91177308-0d34-0410-b5e6-96231b3b80d8

Add the needed #include, emit enums with the sizes of tables, remove
definitions from the LLVM namespace, since they are all static.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23907 91177308-0d34-0410-b5e6-96231b3b80d8

mark this as beta

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23906 91177308-0d34-0410-b5e6-96231b3b80d8

If a user requests help, give them help on both features and processors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23905 91177308-0d34-0410-b5e6-96231b3b80d8

Autogen subtarget information from .td files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23904 91177308-0d34-0410-b5e6-96231b3b80d8

Add subtarget feature/processor defns to the .td file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23903 91177308-0d34-0410-b5e6-96231b3b80d8

rearrange things a bit so that instructions can use subtarget features in the
future.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23902 91177308-0d34-0410-b5e6-96231b3b80d8

add a marker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23901 91177308-0d34-0410-b5e6-96231b3b80d8

add a note that Nate mentioned last week

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23898 91177308-0d34-0410-b5e6-96231b3b80d8

Put some of my random notes somewhere public

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23897 91177308-0d34-0410-b5e6-96231b3b80d8

Work around GCC's dislike of attributes on function definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23896 91177308-0d34-0410-b5e6-96231b3b80d8

Remove deleted files from Visual Studio.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23895 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the obsolete instr selector emitter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23894 91177308-0d34-0410-b5e6-96231b3b80d8

Improve help output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23893 91177308-0d34-0410-b5e6-96231b3b80d8

improve -help output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23892 91177308-0d34-0410-b5e6-96231b3b80d8

Move static functions from .h file, reduce #includes, pass strings by const&,
use LowercaseString from StringExtras.h, remove extraneous space from help
output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23891 91177308-0d34-0410-b5e6-96231b3b80d8

Move static functions to .cpp file, reduce #includes, pass strings by
const&.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23890 91177308-0d34-0410-b5e6-96231b3b80d8

Remove redundant flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23889 91177308-0d34-0410-b5e6-96231b3b80d8

When a function takes a variable number of pointer arguments, with a zero
pointer marking the end of the list, the zero *must* be cast to the pointer
type.  An un-cast zero is a 32-bit int, and at least on x86_64, gcc will
not extend the zero to 64 bits, thus allowing the upper 32 bits to be
random junk.

The new END_WITH_NULL macro may be used to annotate a such a function
so that GCC (version 4 or newer) will detect the use of un-casted zero
at compile time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23888 91177308-0d34-0410-b5e6-96231b3b80d8

Add several things.
loads
branches
setcc
working calls
Global address
External addresses

now I can manage malloc calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23887 91177308-0d34-0410-b5e6-96231b3b80d8

add TargetExternalSymbol

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23886 91177308-0d34-0410-b5e6-96231b3b80d8

Add new files to Visual Studio.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23885 91177308-0d34-0410-b5e6-96231b3b80d8

Well, the Constant matching pattern works.  Can't say much about calls or globals yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23884 91177308-0d34-0410-b5e6-96231b3b80d8

document this as experimental

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23883 91177308-0d34-0410-b5e6-96231b3b80d8

This file is entirely ifdef'd out

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23882 91177308-0d34-0410-b5e6-96231b3b80d8

BuildSDIV and BuildUDIV only work for i32/i64, but they don't check that
the input is that type, this caused a failure on gs on X86 last night.
Move the hard checks into Build[US]Div since that is where decisions like
this should be made.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23881 91177308-0d34-0410-b5e6-96231b3b80d8

Add g3 back to the mix and reorder to irritate them anal folk.  Actually, it's
to group appropriately and provide cues to maintainers that the lists don't
need to be ordered.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23880 91177308-0d34-0410-b5e6-96231b3b80d8

Sort the features and processor lists for the sake of search (and maintainers.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23879 91177308-0d34-0410-b5e6-96231b3b80d8

64-bit reg support should not be enabled by default, as support isn't complete.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23878 91177308-0d34-0410-b5e6-96231b3b80d8

add a case missing from the dag combiner that exposed the failure on
2005-10-21-longlonggtu.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23875 91177308-0d34-0410-b5e6-96231b3b80d8

Instead of aborting if not a case we can handle specially, break out and
let the generic code handle it.  This fixes CodeGen/Generic/2005-10-21-longlonggtu.ll on ppc.

also, reindent this code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23874 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashes the ppc backend

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23871 91177308-0d34-0410-b5e6-96231b3b80d8

Plugin new subtarget backend into the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23870 91177308-0d34-0410-b5e6-96231b3b80d8

New TableGen backends for subtarget information.  Only command line stuff
active now.  Scheduling itinerary next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23869 91177308-0d34-0410-b5e6-96231b3b80d8

silence a release mode warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23868 91177308-0d34-0410-b5e6-96231b3b80d8

add missing prototype

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23867 91177308-0d34-0410-b5e6-96231b3b80d8

Make the coallescer a bit smarter, allowing it to join more live ranges.
For example, we can now join things like [0-30:0)[31-40:1)[52-59:2)
with [40:60:0) if the 52-59 range is defined by a copy from the 40-60 range.
The resultant range ends up being [0-30:0)[31-60:1).

This fires a lot through-out the test suite (e.g. shrinking bc from
19492 -> 18509 machineinstrs) though most gains are smaller (e.g. about
50 copies eliminated from crafty).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23866 91177308-0d34-0410-b5e6-96231b3b80d8

Fix LiveInterval::getOverlapingRanges to take things in the right order
(an unused method).

Fix the merger so that it can merge ranges like this  [10:12)[16:40) with
[12:38) into [10:40) instead of bogus ranges.  This sort of input will be
possible for the merger coming shortly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23865 91177308-0d34-0410-b5e6-96231b3b80d8

Match rotate.  This does actually match the rotates in an rc5 cipher, but I
haven't seen it fire on our testsuite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23863 91177308-0d34-0410-b5e6-96231b3b80d8

My previous patch was too conservative.  Reject FP and void types, but do
allow pointer types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23859 91177308-0d34-0410-b5e6-96231b3b80d8

Don't generate operations that aren't yet supported

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23858 91177308-0d34-0410-b5e6-96231b3b80d8

Kill some now-dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23857 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in the dag combiner, so that this can work on i64 targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23856 91177308-0d34-0410-b5e6-96231b3b80d8

byte zap not immediate goodness

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23855 91177308-0d34-0410-b5e6-96231b3b80d8

Make tblgen emit:
tblgen: In ZAPNOTi: Cannot use 'IZAPX' in an input pattern!
for a bad pattern, instead of an ugly assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23854 91177308-0d34-0410-b5e6-96231b3b80d8

Invert the TargetLowering flag that controls divide by consant expansion.
Add a new flag to TargetLowering indicating if the target has really cheap
  signed division by powers of two, make ppc use it.  This will probably go
  away in the future.
Implement some more ISD::SDIV folds in the dag combiner
Remove now dead code in the x86 backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23853 91177308-0d34-0410-b5e6-96231b3b80d8

Inst cleanup.  As a bonus, operands are in the correct order for cmovs.  Expect new stuff to pass in the JIT tonight

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23852 91177308-0d34-0410-b5e6-96231b3b80d8

Use a literal to define ineg instead of immzero

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23851 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a conditional so we don't access past the end of the range.  Thanks to
Andrew for bringing this to my attn.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23850 91177308-0d34-0410-b5e6-96231b3b80d8

added a few 1 operand form stuff.  Seems to break regalloc on alpha.  sigh

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23849 91177308-0d34-0410-b5e6-96231b3b80d8

add cttz and ctpop

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23848 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple bugs in the const div stuff where we'd generate MULHS/MULHU
for types that aren't legal, and fail a divisor is less than zero
comparison, which would cause us to drop a subtract.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23846 91177308-0d34-0410-b5e6-96231b3b80d8

don't use llabs with apparently VC++ doesn't have

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23845 91177308-0d34-0410-b5e6-96231b3b80d8

Fix order of eval problem from when I refactored this into a function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23844 91177308-0d34-0410-b5e6-96231b3b80d8

Sounds good, finish the intop conversion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23843 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more patterns for i64 on ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23842 91177308-0d34-0410-b5e6-96231b3b80d8

add a new method, play around with some code.

Fix a *bug* in the extendIntervalEndTo method.  In particular, if adding
[2:10) to an interval containing [0:2),[10:30), we produced [0:10),[10,30).
Which is not the most smart thing to do.  Now produce [0:30).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23841 91177308-0d34-0410-b5e6-96231b3b80d8

add a new method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23840 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor some code, pulling it out into a function.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23839 91177308-0d34-0410-b5e6-96231b3b80d8

Do NOT touch FP ops with LSR.  This fixes a testcase Nate sent me from an
inner loop like this:

LBB_RateConvertMono8AltiVec_2:  ; no_exit
        lis r2, ha16(.CPI_RateConvertMono8AltiVec_0)
        lfs f3, lo16(.CPI_RateConvertMono8AltiVec_0)(r2)
        fmr f3, f3
        fadd f0, f2, f0
        fadd f3, f0, f3
        fcmpu cr0, f3, f1
        bge cr0, LBB_RateConvertMono8AltiVec_2  ; no_exit

to an inner loop like this:

LBB_RateConvertMono8AltiVec_1:  ; no_exit
        fsub f2, f2, f1
        fcmpu cr0, f2, f1
        fmr f0, f2
        bge cr0, LBB_RateConvertMono8AltiVec_1  ; no_exit

Doh! good catch!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23838 91177308-0d34-0410-b5e6-96231b3b80d8

Add some pattern fragments to simplify the repetitive parts of the patterns
for some common ops and use them for a few examples.  Andrew, if you like
this, feel free to convert the rest over, if you hate it, feel free to
revert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23837 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this a bit by using immediates

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23836 91177308-0d34-0410-b5e6-96231b3b80d8

Move the target constant divide optimization up into the dag combiner, so
that the nodes can be folded with other nodes, and we can not duplicate
code in every backend.  Alpha will probably want this too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23835 91177308-0d34-0410-b5e6-96231b3b80d8

Enable targets to say that integer divide is expensive, which will trigger
an upcoming optimization in the DAG Combiner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23834 91177308-0d34-0410-b5e6-96231b3b80d8

forgot this one

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23833 91177308-0d34-0410-b5e6-96231b3b80d8

ret 0; works, not much else

still lots of uglyness.
Maybe calls will come soon.
Fixing the return value of things will be necessary to make alpha work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23832 91177308-0d34-0410-b5e6-96231b3b80d8

This fixes PR638:
Regression/CodeGen/Generic/2004-02-08-UnwindSupport.llx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23831 91177308-0d34-0410-b5e6-96231b3b80d8

Added InstrSchedClass to each of the PowerPC Instructions.

Note that when adding new instructions that you should refer to the table at the
bottom of PPCSchedule.td.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23830 91177308-0d34-0410-b5e6-96231b3b80d8

Moved to Regression/Codegen/Generic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23829 91177308-0d34-0410-b5e6-96231b3b80d8

This is a generic test for all code generators.  It originally came from
Regression/Codegen/X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23826 91177308-0d34-0410-b5e6-96231b3b80d8

Write patterns for the various shl and srl patterns that don't involve
doing something clever.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23824 91177308-0d34-0410-b5e6-96231b3b80d8

Force i386 code generation for an i386 specific test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23821 91177308-0d34-0410-b5e6-96231b3b80d8

Push processor descriptions to the top of target and add command line info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23820 91177308-0d34-0410-b5e6-96231b3b80d8

add support for literal immediates in patterns to match, allowing us to
write things like this:

def : Pat<(add GPRC:$in, 12),
          (ADD12 GPRC:$in)>;

Andrew: if this isn't enough or doesn't work for you, please lemme know.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23819 91177308-0d34-0410-b5e6-96231b3b80d8

now that tblgen is smarter, use integers directly.  This should help Andrew too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23818 91177308-0d34-0410-b5e6-96231b3b80d8

Add basic support for integer constants in pattern results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23817 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some checking that was causing duraid to get a perplexing assertion
instead of a happy error message

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23816 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for patterns that have physical registers in them.  Testcase:

def : Pat<(trunc G8RC:$in),
          (OR8To4 G8RC:$in, X0)>;

Even though this doesn't make any sense on PPC :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23815 91177308-0d34-0410-b5e6-96231b3b80d8

Asserting here is to violent

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23814 91177308-0d34-0410-b5e6-96231b3b80d8

teach ppc backend these are copies

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23813 91177308-0d34-0410-b5e6-96231b3b80d8

Make this work with the internalize change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23812 91177308-0d34-0410-b5e6-96231b3b80d8

Convert these cases to patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23811 91177308-0d34-0410-b5e6-96231b3b80d8

Nate wants to define 'Pat's which turn into instructions that don't have
patterns.  Certainly a logical request.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23810 91177308-0d34-0410-b5e6-96231b3b80d8

Woo, it kinda works.  We now generate this atrociously bad, but correct,
code for long long foo(long long a, long long b) { return a + b; }

_foo:
        or r2, r3, r3
        or r3, r4, r4
        or r4, r5, r5
        or r5, r6, r6
        rldicr r2, r2, 32, 31
        rldicl r3, r3, 0, 32
        rldicr r4, r4, 32, 31
        rldicl r5, r5, 0, 32
        or r2, r3, r2
        or r3, r5, r4
        add r4, r3, r2
        rldicl r2, r4, 32, 32
        or r4, r4, r4
        or r3, r2, r2
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23809 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase that crashes llvmgcc in type layout

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23806 91177308-0d34-0410-b5e6-96231b3b80d8