this passes now, due to Nate's recent efforts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26140 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce this testcase a bit more, with the help of llvm-extract and some hand tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26139 91177308-0d34-0410-b5e6-96231b3b80d8

Be careful not to request or look at bits shifted in from outside the size
of the input.  This fixes the mediabench/gsm/toast failure last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26138 91177308-0d34-0410-b5e6-96231b3b80d8

Added a test case for a libcall insertion bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26137 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing patterns for andi. and andis., fixing test/Regression/CodeGen/
PowerPC/and-imm.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26136 91177308-0d34-0410-b5e6-96231b3b80d8

remove some more dead special case code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26135 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate special case hacks that are superceded by general purpose hacks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26134 91177308-0d34-0410-b5e6-96231b3b80d8

tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26133 91177308-0d34-0410-b5e6-96231b3b80d8

Three changes:
1. Teach GetConstantInType to handle boolean constants.
2. Teach instcombine to fold (compare X, CST) when X has known 0/1 bits.
   Testcase here: set.ll:test22
3. Improve the "(X >> c1) & C2 == 0" folding code to allow a noop cast
   between the shift and and.  More aggressive bitfolding for other reasons
   was turning signed shr's into unsigned shr's, leaving the noop cast in
   the way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26131 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26130 91177308-0d34-0410-b5e6-96231b3b80d8

move a failing testcase from bit-tracking.ll to narrow.ll, and move the
xfail marker with it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26129 91177308-0d34-0410-b5e6-96231b3b80d8

Revert my last patch.  It too breaks stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26128 91177308-0d34-0410-b5e6-96231b3b80d8

Make these tests fail if opt crashes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26127 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for my previously reverted patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26126 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments to be actually accurate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26124 91177308-0d34-0410-b5e6-96231b3b80d8

This is implemented by the simplify-libcalls pass, not instcombine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26123 91177308-0d34-0410-b5e6-96231b3b80d8

Port the recent innovations in ComputeMaskedBits to SimplifyDemandedBits.
This allows us to simplify on conditions where bits are not known, but they
are not demanded either!  This also fixes a couple of bugs in
ComputeMaskedBits that were exposed during this work.

In the future, swaths of instcombine should be removed, as this code
subsumes a bunch of ad-hockery.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26122 91177308-0d34-0410-b5e6-96231b3b80d8

revert my previous change, it exposed other problems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26121 91177308-0d34-0410-b5e6-96231b3b80d8

fix storing booleans (grawp missed this one)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26120 91177308-0d34-0410-b5e6-96231b3b80d8

now short immediates will get matched (previously constants were all
triggering movl 64bit imm fat instructions)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26119 91177308-0d34-0410-b5e6-96231b3b80d8

Make this check stricter.  Disallow loop exit blocks from being shared by
loops and their subloops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26118 91177308-0d34-0410-b5e6-96231b3b80d8

Prevent certain nodes that have already been selected from being folded into
X86 addressing mode. Currently we do not allow any node whose target node
produces a chain as well as any node that is at the root of the addressing
mode expression tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26117 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead expr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26116 91177308-0d34-0410-b5e6-96231b3b80d8

Reorg for integration with gcc4.  Old style debug info will not be passed though
to SelIDAG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26115 91177308-0d34-0410-b5e6-96231b3b80d8

implement unswitching of loops with switch stmts and selects in them

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26114 91177308-0d34-0410-b5e6-96231b3b80d8

Update PHI nodes in successors of exit blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26113 91177308-0d34-0410-b5e6-96231b3b80d8

Reform the unswitching code in terms of edge splitting, not block splitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26112 91177308-0d34-0410-b5e6-96231b3b80d8

Nicer code. :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26111 91177308-0d34-0410-b5e6-96231b3b80d8

Added X86 isel debugging stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26110 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a level of indirection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26109 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a case where UnswitchTrivialCondition broke critical edges with
phi's in the successors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26108 91177308-0d34-0410-b5e6-96231b3b80d8

Use the auto-generated call matcher.  Remove a broken impl of the frameaddr/returnaddr
intrinsics.

Autogen frameindex matcher

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26107 91177308-0d34-0410-b5e6-96231b3b80d8

Update to new-style flags usage, simplifying the .td file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26106 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a completed entry; add a new entry about fisttp op

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26105 91177308-0d34-0410-b5e6-96231b3b80d8

add some notes, move some code around.  Implement unswitching of loops
with branches on partially invariant computations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26104 91177308-0d34-0410-b5e6-96231b3b80d8

Move code around to be more logical, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26103 91177308-0d34-0410-b5e6-96231b3b80d8

When unswitching a trivial loop, do admit we are doing it! :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26102 91177308-0d34-0410-b5e6-96231b3b80d8

Implement unconditional unswitching of 'trivial' loops, those loops that contain
branches in their entry block that control whether or not the loop is a noop or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26101 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify control flow a bit, note that unswitch preserves canonical loop form

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26098 91177308-0d34-0410-b5e6-96231b3b80d8

Match tblgen change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26096 91177308-0d34-0410-b5e6-96231b3b80d8

Call InsertISelMapEntry rather than map insertion operator to prevent overly
aggrssive inlining. This reduces Select_store frame size from 24k to 10k.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26095 91177308-0d34-0410-b5e6-96231b3b80d8

Added SelectionDAG::InsertISelMapEntry(). This is used to workaround the gcc
problem where it inline the map insertion call too aggressively. Before this
change it was producing a frame size of 24k for Select_store(), now it's down
to 10k (by calling this method rather than calling the map insertion operator).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26094 91177308-0d34-0410-b5e6-96231b3b80d8

Make the threshold a parameter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26093 91177308-0d34-0410-b5e6-96231b3b80d8

Done

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26091 91177308-0d34-0410-b5e6-96231b3b80d8

Enable LSR by default for SPARC: it is a clear win.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26090 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the loop-unswitch pass, by not even trying to unswitch loops with
uses of loop values outside the loop.  We need loop-closed SSA form to do
this right, or to use SSA rewriting if we really care.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26089 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80-column violations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26088 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance MVIZ in three ways:
1. Teach it new tricks: in particular how to propagate through signed shr and sexts.
2. Teach it to return a bitset of known-1 and known-0 bits, instead of just zero.
3. Teach instcombine (AND X, C) to fold when we know all C bits of X.

This implements Regression/Transforms/InstCombine/bittest.ll, and allows
future things to be simplified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26087 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26086 91177308-0d34-0410-b5e6-96231b3b80d8

Match getTargetNode() changes (now return SDNode* instead of SDOperand).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26085 91177308-0d34-0410-b5e6-96231b3b80d8

Match getTargetNode() changes (now returns SDNode* instead of SDOperand).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26084 91177308-0d34-0410-b5e6-96231b3b80d8

More changes to reduce frame size.
Move all getTargetNode() out of SelectionDAG.h into SelectionDAG.cpp. This
prevents them from being inlined.
Change getTargetNode() so they return SDNode * instead of SDOperand to prevent
copying. It should also help compilation speed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26083 91177308-0d34-0410-b5e6-96231b3b80d8

this apparently passes on linux

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26082 91177308-0d34-0410-b5e6-96231b3b80d8

add an option to turn on LSR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26080 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code now that each constant pool entry is not separately allocated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26079 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to MachineConstantPool interface change: instead of keeping a
value/alignment pair for each constant, keep a value/offset pair.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26078 91177308-0d34-0410-b5e6-96231b3b80d8

instead of keeping track of Constant/alignment pairs, actually compute the
offset of each entry from the start of the constant pool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26077 91177308-0d34-0410-b5e6-96231b3b80d8

rename fields of constant pool entries

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26076 91177308-0d34-0410-b5e6-96231b3b80d8

Use a MachineConstantPoolEntry struct instead of a pair to hold
constant pool entries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26075 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code, alignment must be specified now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26074 91177308-0d34-0410-b5e6-96231b3b80d8

Assert invariants

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26073 91177308-0d34-0410-b5e6-96231b3b80d8

Require an alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26072 91177308-0d34-0410-b5e6-96231b3b80d8

Make MachineConstantPool entries alignments explicit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26071 91177308-0d34-0410-b5e6-96231b3b80d8

Always pass in an alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26070 91177308-0d34-0410-b5e6-96231b3b80d8

provide an explicit alignment for cp entries

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26069 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment: value is log2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26068 91177308-0d34-0410-b5e6-96231b3b80d8

Change Select() from
SDOperand Select(SDOperand N);
to
void Select(SDOperand &Result, SDOperand N);

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26067 91177308-0d34-0410-b5e6-96231b3b80d8

Darwin doesn't support #APP/#NO_APP

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26066 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for assembler directives that wrap inline asm

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26065 91177308-0d34-0410-b5e6-96231b3b80d8

Rename BSel -> PPCBSel for the benefit of doxygen users.
Move the methods out of line.
Remove unused Debug.h stuff.
Teach getNumBytesForInstruction to know the size of an inline asm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26064 91177308-0d34-0410-b5e6-96231b3b80d8

Disable this test for the time being as debug is brought up to speed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26063 91177308-0d34-0410-b5e6-96231b3b80d8

add a simple optimization

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26062 91177308-0d34-0410-b5e6-96231b3b80d8

Mention that delta can be used to reduce some Front-end problems.
Patch by Marco Matthies, thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26061 91177308-0d34-0410-b5e6-96231b3b80d8

Add SRoA to the lexicon.  Patch by Marco Matthies!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26060 91177308-0d34-0410-b5e6-96231b3b80d8

Added options -cflag, -cxxflags, and -ldflags to override the default C
compilation, C++ compilation, and linker options.
e.g. This is the options I use for testing on my x86 iMac:
nice ./NightlyTest.pl -release -cflags "-Os -DNDEBUG -fomit-frame-pointer" -cxxflags "-Os -DNDEBUG -finline-functions -felide-constructors -fomit-frame-pointer"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26057 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify some code, reducing calls to MaskedValueIsZero.  Implement a minor
optimization where we reduce the number of bits in AND masks when possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26056 91177308-0d34-0410-b5e6-96231b3b80d8

Remove -pedantic. It no longer works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26055 91177308-0d34-0410-b5e6-96231b3b80d8

more email -> README moving

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26054 91177308-0d34-0410-b5e6-96231b3b80d8

Emit the 'mr' pseudoop for easier reading.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26053 91177308-0d34-0410-b5e6-96231b3b80d8

Add some random notes, not high-prio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26052 91177308-0d34-0410-b5e6-96231b3b80d8

Move emails from nate into public places

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26051 91177308-0d34-0410-b5e6-96231b3b80d8

Use EraseInstFromFunction in a few cases to put the uses of the removed
instruction onto the worklist (in case they are now dead).

Add a really trivial local DSE implementation to help out bitfield code.
We now fold this:

struct S {
    unsigned char a : 1, b : 1, c : 1, d : 2, e : 3;
    S();
};

S::S() : a(0), b(0), c(1), d(0), e(6) {}

to this:

void %_ZN1SC1Ev(%struct.S* %this) {
entry:
        %tmp.1 = getelementptr %struct.S* %this, int 0, uint 0
        store ubyte 38, ubyte* %tmp.1
        ret void
}

much earlier (in gccas instead of only in gccld after DSE runs).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26050 91177308-0d34-0410-b5e6-96231b3b80d8

Implement some more interesting select sccp cases.  This implements:
test/Regression/Transforms/SCCP/select.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26049 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for more interesting select sccp cases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26048 91177308-0d34-0410-b5e6-96231b3b80d8

Compile this:

        xori r6, r2, 1
        rlwinm r6, r6, 0, 31, 31
        cmpwi cr0, r6, 0
        bne cr0, LBB1_3 ; endif

to this:

        rlwinm r6, r2, 0, 31, 31
        cmpwi cr0, r6, 0
        beq cr0, LBB1_3 ; endif

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26047 91177308-0d34-0410-b5e6-96231b3b80d8

Add some happy helper methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26046 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a problem in my patch yesterday, causing a miscompilation of 176.gcc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26045 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a local common symbol bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26044 91177308-0d34-0410-b5e6-96231b3b80d8

For ELF, .comm takes alignment value as the optional 3rd argument. It must be
specified in bytes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26043 91177308-0d34-0410-b5e6-96231b3b80d8

Implement getConstraintType for PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26042 91177308-0d34-0410-b5e6-96231b3b80d8

getConstraintType should be virtual.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26041 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/InstCombine/2006-02-07-SextZextCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26040 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase that caused instcombine to crash on 176.gcc last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26039 91177308-0d34-0410-b5e6-96231b3b80d8

Darwin ABI issues: weak, linkonce, etc. dynamic-no-pic support is complete.
Also fixed a function stub bug. Added weak and linkonce support for
x86 Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26038 91177308-0d34-0410-b5e6-96231b3b80d8

Remind myself to add PIC and static asm printer support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26037 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize MaskedValueIsZero into a ComputeMaskedNonZeroBits function, which
is just as efficient as MVIZ and is also more general.

Fix a few minor bugs introduced in recent patches

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26036 91177308-0d34-0410-b5e6-96231b3b80d8

Make MaskedValueIsZero take a uint64_t instead of a ConstantIntegral as a
mask.  This allows the code to be simpler and more efficient.

Also, generalize some of the cases in MVIZ a bit, making it slightly more aggressive.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26035 91177308-0d34-0410-b5e6-96231b3b80d8

Use Type::getIntegralTypeMask() to simplify some code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26034 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the beginnings of a facility for simplifying expressions based on
'demanded bits', inspired by Nate's work in the dag combiner.  This isn't
complete, but needs to unrelated instcombiner changes to continue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26033 91177308-0d34-0410-b5e6-96231b3b80d8

add a new Type::getIntegralTypeMask() method, which is useful for clients that
want to do bitwise inspection of integer types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26032 91177308-0d34-0410-b5e6-96231b3b80d8