Remove a bogus piece of my AssertSext/AssertZext patch.  oops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23148 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for AssertSext and AssertZext, folding other extensions with
them.  This allows for elminination of redundant extends in the entry
blocks of functions on PowerPC.

Add support for i32 x i32 -> i64 multiplies, by recognizing when the inputs
to ISD::MUL in ExpandOp are actually just extended i32 values and not real
i64 values.  this allows us to codegen

int mulhs(int a, int b) { return ((long long)a * b) >> 32; }
as:
_mulhs:
        mulhw r3, r4, r3
        blr

instead of:
_mulhs:
        mulhwu r2, r4, r3
        srawi r5, r3, 31
        mullw r5, r4, r5
        add r2, r2, r5
        srawi r4, r4, 31
        mullw r3, r4, r3
        add r3, r2, r3
        blr

with a similar improvement on x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23147 91177308-0d34-0410-b5e6-96231b3b80d8

Add AssertSext, AssertZext nodes for targets that pass arguments in
registers, and the incoming values have already been zero or sign extended
from the appopriate type to the register width.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23146 91177308-0d34-0410-b5e6-96231b3b80d8

Name this variable to be what it really is!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23145 91177308-0d34-0410-b5e6-96231b3b80d8

Handle CopyToReg nodes with flag operands correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23144 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure the selector emits register register copies with flag operands
linking them to calls when appropriate, this prevents the scheduler from
pulling these copies away from the call.

This fixes Ptrdist/yacr2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23143 91177308-0d34-0410-b5e6-96231b3b80d8

add some method variants

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23142 91177308-0d34-0410-b5e6-96231b3b80d8

The first operand to AND does not always have more than two operands. This
fixes MediaBench/toast with the dag selector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23141 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in my patch for legalizing to fsel.  It cannot handle seteq/setne,
which I failed to include when I moved the code over.  This fixes
MallocBench/gs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23140 91177308-0d34-0410-b5e6-96231b3b80d8

emit FMR instructions to convert f64<->f32 instructions, so things like
STOREs, know the right type to store.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23139 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some really strange indentation that xcode likes to use.

no xcode, this is not right:

   if (!foo) break;
     X;

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23138 91177308-0d34-0410-b5e6-96231b3b80d8

fix a crash in cfrac

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23137 91177308-0d34-0410-b5e6-96231b3b80d8

Implement DYNAMIC_STACKALLOC, wrap some long lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23136 91177308-0d34-0410-b5e6-96231b3b80d8

Add a hack to avoid some horrible code in some cases by always emitting
token chains first.  For this C function:

int test() {
  int i;
  for (i = 0; i < 100000; ++i)
    foo();
}

Instead of emitting this (condition before call)

.LBB_test_1:    ; no_exit
        addi r30, r30, 1
        lis r2, 1
        ori r2, r2, 34464
        cmpw cr2, r30, r2
        bl L_foo$stub
        bne cr2, .LBB_test_1    ; no_exit

Emit this:

.LBB_test_1:    ; no_exit
        bl L_foo$stub
        addi r30, r30, 1
        lis r2, 1
        ori r2, r2, 34464
        cmpw cr0, r30, r2
        bne cr0, .LBB_test_1    ; no_exit

Which makes it so we don't have to save/restore cr2 in the prolog/epilog of
the function.

This also makes the code much more similar to what the pattern isel produces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23135 91177308-0d34-0410-b5e6-96231b3b80d8

improve comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23134 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a dumb bug of mine where we were mishandling the PPC ABI (undef handling).
This fixes voronoi and bh in Olden, allowing all of olden to pass!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23133 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new API for nate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23132 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new API for Nate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23131 91177308-0d34-0410-b5e6-96231b3b80d8

Some of us cared about the the promote path

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23130 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an infinite loop on x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23129 91177308-0d34-0410-b5e6-96231b3b80d8

Allow bugpoint+PPC codegen to use fsqrt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23128 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug the last patch exposed in treeadd among others

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23127 91177308-0d34-0410-b5e6-96231b3b80d8

A hack to fix a problem folding immedaites.  This fixes Olden/power.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23126 91177308-0d34-0410-b5e6-96231b3b80d8

Fix order of operands for copytoreg node when emitting calls.  This fixes
Olden/msFix order of operands for copytoreg node when emitting calls.  This fixes
Olden/mstt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23125 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in my previous patch that was using the wrong iterator.  This fixes
Olden/bisort among others.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23124 91177308-0d34-0410-b5e6-96231b3b80d8

add operands in the correct order

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23123 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in ReplaceAllUsesWith

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23122 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in FP_EXTEND, implement FP_TO_SINT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23121 91177308-0d34-0410-b5e6-96231b3b80d8

fix an assertion failure in treeadd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23120 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to member variable name change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23119 91177308-0d34-0410-b5e6-96231b3b80d8

Change the names of member variables per Chris' instructions, and document
them more clearly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23118 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23117 91177308-0d34-0410-b5e6-96231b3b80d8

Implement PR614:
These changes modify the makefiles so that the output of flex and bison are
placed in the SRC directory, not the OBJ directory. It is intended that they
be checked in as any other LLVM source so that platforms without convenient
access to flex/bison can be compiled. From now on, if you change a .y or
.l file you *must* also commit the generated .cpp and .h files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23115 91177308-0d34-0410-b5e6-96231b3b80d8

Disable this code, which broke many tests last night

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23114 91177308-0d34-0410-b5e6-96231b3b80d8

fix PHI node emission for basic blocks that have select_cc's in them on ppc32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23113 91177308-0d34-0410-b5e6-96231b3b80d8

The condition register being branched on may not be cr0, as such, print it.
This fixes: UnitTests/2005-07-17-INT-To-FP.c

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23112 91177308-0d34-0410-b5e6-96231b3b80d8

Propagate cr# from COND_BRANCH to the actual branch instruction as appropriate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23111 91177308-0d34-0410-b5e6-96231b3b80d8

Nate noticed that Andrew never did this.  This fixes PR600

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23110 91177308-0d34-0410-b5e6-96231b3b80d8

Don't copy regs that are only used in the entry block into a vreg.  This
changes the code generated for:

short %test(short %A) {
  %B = xor short %A, -32768
  ret short %B
}

to:

_test:
        xori r2, r3, 32768
        xoris r2, r2, 65535
        extsh r3, r2
        blr

instead of:

_test:
        rlwinm r2, r3, 0, 16, 31
        xori r2, r3, 32768
        xoris r2, r2, 65535
        extsh r3, r2
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23109 91177308-0d34-0410-b5e6-96231b3b80d8

Make this code safe for when loadRegFromStackSlot inserts multiple instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23108 91177308-0d34-0410-b5e6-96231b3b80d8

allow code using mtcrf to assemble

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23107 91177308-0d34-0410-b5e6-96231b3b80d8

Remove operand type 'crbit', since it is no longer used

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23106 91177308-0d34-0410-b5e6-96231b3b80d8

teach getClass what a condition reg is

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23105 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanups:
  * avoid calling getClass() multiple times (it is relatively expensive)
  * Allow -disable-fp-elim to turn of frame pointer elimination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23104 91177308-0d34-0410-b5e6-96231b3b80d8

Checking types here is not safe, because multiple types can map to the same
register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23103 91177308-0d34-0410-b5e6-96231b3b80d8

implement SELECT_CC fully for the DAG->DAG isel!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23101 91177308-0d34-0410-b5e6-96231b3b80d8

add some forward defs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23100 91177308-0d34-0410-b5e6-96231b3b80d8

spell this right

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23099 91177308-0d34-0410-b5e6-96231b3b80d8

Call the InsertAtEndOfBasicBlock hook if the usesCustomDAGSchedInserter
flag is set on an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23098 91177308-0d34-0410-b5e6-96231b3b80d8

spell this right!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23097 91177308-0d34-0410-b5e6-96231b3b80d8

Add a hook

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23096 91177308-0d34-0410-b5e6-96231b3b80d8

spell this variable right

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23095 91177308-0d34-0410-b5e6-96231b3b80d8

Expose a new flag to TargetInstrInfo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23094 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new instruction flag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23093 91177308-0d34-0410-b5e6-96231b3b80d8

Add a flag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23092 91177308-0d34-0410-b5e6-96231b3b80d8

Make fsel emission work with both the pattern and dag-dag selectors, by
giving it a non-instruction opcode.  The dag->dag selector used to not
select the operands of the fsel, because it thought that whole tree was
already selected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23091 91177308-0d34-0410-b5e6-96231b3b80d8

add a marker at the end of the instruction enum list

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23090 91177308-0d34-0410-b5e6-96231b3b80d8

implement the fold for:

bool %test(int %X, int %Y) {
        %C = setne int %X, 0
        ret bool %C
}

to:

_test:
        addic r2, r3, -1
        subfe r3, r2, r3
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23089 91177308-0d34-0410-b5e6-96231b3b80d8

Changes to adjust to new ReplaceAllUsesWith syntax.  Change FP_EXTEND to
just return its input, instead of emitting an explicit copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23088 91177308-0d34-0410-b5e6-96231b3b80d8

Revampt ReplaceAllUsesWith to be more efficient and easier to use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23087 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new version of ReplaceAllUsesWith, make the comments more useful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23086 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some code made dead by the fsel patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23085 91177308-0d34-0410-b5e6-96231b3b80d8

now that fsel is formed during legalization, this code is dead

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23084 91177308-0d34-0410-b5e6-96231b3b80d8

implement the other half of the select_cc -> fsel lowering, which handles
when the RHS of the comparison is 0.0.  Turn this on by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23083 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in my previous checkin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23082 91177308-0d34-0410-b5e6-96231b3b80d8

Change ConstantPoolSDNode to actually hold the Constant itself instead of
putting it into the constant pool.  This allows the isel machinery to
create constants that it will end up deciding are not needed, without them
ending up in the resultant function constant pool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23081 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some warnings in an optimized build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23080 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a huge annoyance: SelectNodeTo took types before the opcode unlike
every other SD API.  Fix it to take the opcode before the types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23079 91177308-0d34-0410-b5e6-96231b3b80d8

Changed unsigned long to unsigned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23078 91177308-0d34-0410-b5e6-96231b3b80d8

Change unsigned lng to unsigned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23077 91177308-0d34-0410-b5e6-96231b3b80d8

Fix JIT encoding of conditional branches

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23076 91177308-0d34-0410-b5e6-96231b3b80d8

add initial support for converting select_cc -> fsel in the legalizer
instead of in the backend.  This currently handles fsel cases with registers,
but doesn't have the 0.0 and -0.0 optimization enabled yet.

Once this is finished, special hack for fp immediates can go away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23075 91177308-0d34-0410-b5e6-96231b3b80d8

the 5th operand is the 4th number

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23074 91177308-0d34-0410-b5e6-96231b3b80d8

SUBFIC produces two results, not one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23073 91177308-0d34-0410-b5e6-96231b3b80d8

Implement SHL_PARTS and SRL_PARTS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23072 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for targets that want to custom expand select_cc in some cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23071 91177308-0d34-0410-b5e6-96231b3b80d8

Allow LowerOperation to return a null SDOperand in case it wants to lower
some things given to it, but not all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23070 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a nasty bug from a previous patch of mine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23069 91177308-0d34-0410-b5e6-96231b3b80d8

Emit the lo/hi parts in the right order :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23068 91177308-0d34-0410-b5e6-96231b3b80d8

Added cl::bits option type (bit vectors).  See "Collecting options as a set
of flags" in the Command Line doc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23067 91177308-0d34-0410-b5e6-96231b3b80d8

implement support for 64-bit add/sub, fix a broken assertion for 64-bit
return.  Allow the udiv breaker-upper to work with any non-zero constant
operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23066 91177308-0d34-0410-b5e6-96231b3b80d8

simplify the add/sub_parts code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23065 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak of multiple occurance paragraph.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23064 91177308-0d34-0410-b5e6-96231b3b80d8

Documentation updated to include upcoming support for bit vector support
(flags.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23063 91177308-0d34-0410-b5e6-96231b3b80d8

Finish implementing SDIV/UDIV by copying over the majik constant code from
ISelPattern

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23062 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify some code.  It's not clear why the UDIV expanded sequence
doesn't work for large uint constants, but we'll keep the current behavior

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23061 91177308-0d34-0410-b5e6-96231b3b80d8

Implement setcc correctly for G5 and non-G5 systems

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23060 91177308-0d34-0410-b5e6-96231b3b80d8

implement setcc on the G5.  We're still missing the non-g5 specific bits, but
they will come later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23059 91177308-0d34-0410-b5e6-96231b3b80d8

New fold for SELECT_CC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23058 91177308-0d34-0410-b5e6-96231b3b80d8

Remove option to make SetCC illegal on PowerPC after long discussion with
Chris.  This will be accomplished through correctly modeling CR's and
subregs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23056 91177308-0d34-0410-b5e6-96231b3b80d8

Don't auto-cse nodes that return flags

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23055 91177308-0d34-0410-b5e6-96231b3b80d8

add printer support for flag operands

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23054 91177308-0d34-0410-b5e6-96231b3b80d8

simplify the code a bit using isOperationLegal

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23053 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for sdiv by 2^k and -2^k.  Producing code like:

_test:
        srawi r2, r3, 2
        addze r3, r2
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23052 91177308-0d34-0410-b5e6-96231b3b80d8

fit in 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23051 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for flag operands

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23050 91177308-0d34-0410-b5e6-96231b3b80d8

add a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23049 91177308-0d34-0410-b5e6-96231b3b80d8

add an enum value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23048 91177308-0d34-0410-b5e6-96231b3b80d8

Remove Support/Search.h entry

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23047 91177308-0d34-0410-b5e6-96231b3b80d8

Recommended to use std::algorithms instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23046 91177308-0d34-0410-b5e6-96231b3b80d8