Legalize ConstantFP into TargetConstantFP when the target allows.  Implement
custom expansion of ConstantFP nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25772 91177308-0d34-0410-b5e6-96231b3b80d8

Targets all now request ConstantFP to be legalized into TargetConstantFP.
'fpimm' in .td files is now TargetConstantFP.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25771 91177308-0d34-0410-b5e6-96231b3b80d8

Update alpha to reflect recent constantfp legalize changes.  It's not clear
why all this code isn't autogenerated. :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25770 91177308-0d34-0410-b5e6-96231b3b80d8

Make ConstantFP legalize into TargetConstantFP like other leaf nodes do.  Allow
targets to register custom legalizers for ConstantFP in case there isn't a
fixed list of constants that can be generated.  On some architectures (ia64?)
all fp immediates are legal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25769 91177308-0d34-0410-b5e6-96231b3b80d8

remove the getBR2Way_CC method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25768 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate uses of SelectionDAG::getBR2Way_CC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25767 91177308-0d34-0410-b5e6-96231b3b80d8

it is nice not to chop off bits for those blessed with lots of bits

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25766 91177308-0d34-0410-b5e6-96231b3b80d8

make the casts actually cast to the variable type

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25765 91177308-0d34-0410-b5e6-96231b3b80d8

start of the 64bit safety cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25764 91177308-0d34-0410-b5e6-96231b3b80d8

Emit series of conditionals with &&, emitting stuff like this:

    if (N1.getOpcode() == ISD::LOAD &&
        N1.hasOneUse() &&
        !CodeGenMap.count(N1.getValue(0)) &&
        !CodeGenMap.count(N1.getValue(1))) {

instead of this:

    if (N1.getOpcode() == ISD::LOAD) {
      if (N1.hasOneUse()) {
        if (!CodeGenMap.count(N1.getValue(0))) {
          if (!CodeGenMap.count(N1.getValue(1))) {

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25763 91177308-0d34-0410-b5e6-96231b3b80d8

Factor matching code that is common between patterns.  This works around
GCC not jump-threading across this common code, and produces far nicer
output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25762 91177308-0d34-0410-b5e6-96231b3b80d8

cmovle != cmovlt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25761 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25760 91177308-0d34-0410-b5e6-96231b3b80d8

Split out code generation from analysis from emission

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25759 91177308-0d34-0410-b5e6-96231b3b80d8

move some code around, no change in the generated code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25758 91177308-0d34-0410-b5e6-96231b3b80d8

now that we have control over emission of the code, emit the code using nested
"if" statements (indenting it appropriately, of course) instead of using goto's.

This inverts the logic for all of the if statements, which makes things simpler
to understand in addition to making the generated code easier to read.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25757 91177308-0d34-0410-b5e6-96231b3b80d8

Change PatternCodeEmitter to emit code into a buffer instead of emitting it
directly to the output file.  This makes things simple because the code doesn't
have to worry about indentation or the case when there is no goto.  It also
allows us to indent the code better without touching everything :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25756 91177308-0d34-0410-b5e6-96231b3b80d8

Flesh out AMD family/models.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25755 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly determine CPU vendor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25754 91177308-0d34-0410-b5e6-96231b3b80d8

Make long warnings wrap and have less indent so other portions of the page
are not skewed out of view.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25753 91177308-0d34-0410-b5e6-96231b3b80d8

Fit to 80 columns, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25752 91177308-0d34-0410-b5e6-96231b3b80d8

Use union instead of reinterpret_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25751 91177308-0d34-0410-b5e6-96231b3b80d8

Fix recognition of Intel CPUs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25750 91177308-0d34-0410-b5e6-96231b3b80d8

Is64Bit reflects the capability of the chip, not an aspect of the target os

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25749 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bunch of JIT failures with the new isel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25748 91177308-0d34-0410-b5e6-96231b3b80d8

Improve X86 subtarget support for Windows and AMD.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25747 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new "UpdateNodeOperands" method to simplify LegalizeDAG and make it
faster.  This cuts about 120 lines of code out of the legalizer (mostly code
checking to see if operands have changed).

It also fixes an ugly performance issue, where the legalizer cloned the entire
graph after any change.  Now the "UpdateNodeOperands" method gives it a chance
to reuse nodes if the operands of a node change but not its opcode or valuetypes.

This speeds up instruction selection time on kimwitu++ by about 8.2% with a
release build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25746 91177308-0d34-0410-b5e6-96231b3b80d8

silence a warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25745 91177308-0d34-0410-b5e6-96231b3b80d8

add another method variant

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25744 91177308-0d34-0410-b5e6-96231b3b80d8

add another variant

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25743 91177308-0d34-0410-b5e6-96231b3b80d8

add some methods for updating nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25742 91177308-0d34-0410-b5e6-96231b3b80d8

Add some methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25741 91177308-0d34-0410-b5e6-96231b3b80d8

minor tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25740 91177308-0d34-0410-b5e6-96231b3b80d8

move a bunch of code, no other change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25739 91177308-0d34-0410-b5e6-96231b3b80d8

remove a couple more now-extraneous legalizeop's

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25738 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25737 91177308-0d34-0410-b5e6-96231b3b80d8

Several major changes:

1. Pull out the expand cases for BSWAP and CT* into a separate function,
   reducing the size of LegalizeOp.
2. Fix a bug where expand(bswap i64) was wrong when i64 is legal.
3. Changed LegalizeOp/PromoteOp so that the legalizer never needs to be
   iterative.  It now operates in a single pass over the nodes.
4. Simplify a LOT of code, with a net reduction of ~280 lines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25736 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in my elimination of ISD::CALL this morning.  PPC now has to
provide the expansion for i64 calls itself

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25735 91177308-0d34-0410-b5e6-96231b3b80d8

make this work on non-native hosts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25734 91177308-0d34-0410-b5e6-96231b3b80d8

add a note about how we should implement this FIXME from the legalizer:

    // FIXME: revisit this when we have some kind of mechanism by which targets
    // can decided legality of vector constants, of which there may be very
    // many.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25733 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate the need for ExpandOp to set 'needsanotheriteration', as it already
relegalizes the stuff it returns.

Add the ability to custom expand ADD/SUB, so that targets don't need to deal
with ADD_PARTS/SUB_PARTS if they don't want.

Fix some obscure potential bugs and simplify code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25732 91177308-0d34-0410-b5e6-96231b3b80d8

Instead of making callers of ExpandLibCall legalize the result, make
ExpandLibCall do it itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25731 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate the need to do another iteration of the legalizer after inserting
a libcall.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25730 91177308-0d34-0410-b5e6-96231b3b80d8

remove this method I just added, now is not the time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25729 91177308-0d34-0410-b5e6-96231b3b80d8

remove method I just added

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25728 91177308-0d34-0410-b5e6-96231b3b80d8

add a new callback

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25727 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Promote for VAARG, and allow it to be custom promoted for people
who don't want the default behavior (Alpha).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25726 91177308-0d34-0410-b5e6-96231b3b80d8

Add a couple more things to the readme.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25724 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missing case to the dag combiner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25723 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the ISD::CALL and ISD::TAILCALL nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25721 91177308-0d34-0410-b5e6-96231b3b80d8

The 'target-independent' ISD::CALL isn't.  Nuke it, making way for Nate's
future work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25720 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25719 91177308-0d34-0410-b5e6-96231b3b80d8

Switch to AlphaISD::CALL instead of ISD::CALL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25718 91177308-0d34-0410-b5e6-96231b3b80d8

Use PPCISD::CALL instead of ISD::CALL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25717 91177308-0d34-0410-b5e6-96231b3b80d8

Use V8ISD::CALL instead of ISD::CALL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25716 91177308-0d34-0410-b5e6-96231b3b80d8

A bit of wisdom from Chris on the last entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25715 91177308-0d34-0410-b5e6-96231b3b80d8

AT&T assembly convention: registers are in lower case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25714 91177308-0d34-0410-b5e6-96231b3b80d8

Use Intel assembly syntax to look for ST

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25713 91177308-0d34-0410-b5e6-96231b3b80d8

initialize member vars

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25712 91177308-0d34-0410-b5e6-96231b3b80d8

initialize all instance vars

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25711 91177308-0d34-0410-b5e6-96231b3b80d8

Make llvm.frame/returnaddr not crash on ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25710 91177308-0d34-0410-b5e6-96231b3b80d8

Teach tablegen to generate code that is VC++ warning-free.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25709 91177308-0d34-0410-b5e6-96231b3b80d8

(store (op (load ...)), ...)
If store's chain operand is load, then use load's chain operand instead. If
it isn't (likely a TokenFactor), then do not allow the folding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25708 91177308-0d34-0410-b5e6-96231b3b80d8

* Eliminate FAKE_SOURCES
* Make runtimes and projects build with the new front-end by not relying on
  'llvm-gcc -c' to build a .bc file.  Instead, use llvm-gcc -S -emit-llvm,
  then an explicit invocation of gccas.  Also, don't use llvm-gcc to link
  .bc files together, use gccld directly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25707 91177308-0d34-0410-b5e6-96231b3b80d8

Added notes about a x86 isel deficiency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25706 91177308-0d34-0410-b5e6-96231b3b80d8

Remove out of date comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25705 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the only use of FAKE_SOURCES

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25704 91177308-0d34-0410-b5e6-96231b3b80d8

Remove FAKE_SOURCES

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25703 91177308-0d34-0410-b5e6-96231b3b80d8

Added a temporary option -enable-x86-sse to enable sse support. It is used by
llc-beta.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25701 91177308-0d34-0410-b5e6-96231b3b80d8

Bye bye Pattern ISel, hello DAG ISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25700 91177308-0d34-0410-b5e6-96231b3b80d8

X86 dag isel is now (soon) the default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25699 91177308-0d34-0410-b5e6-96231b3b80d8

These are fp stack test cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25698 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL tailcall test cases until it's implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25697 91177308-0d34-0410-b5e6-96231b3b80d8

Remove TLI.LowerReturnTo, and just let targets custom lower ISD::RET for
the same functionality.  This addresses another piece of bug 680.  Next,
on to fixing Alpha VAARG, which I broke last time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25696 91177308-0d34-0410-b5e6-96231b3b80d8

Using bit size of integers instead of ambiguous "long" et all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25694 91177308-0d34-0410-b5e6-96231b3b80d8

A better workaround

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25692 91177308-0d34-0410-b5e6-96231b3b80d8

Sorry - really folowing convention.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25691 91177308-0d34-0410-b5e6-96231b3b80d8

force sse/3dnow off until they work.  This fixes all the x86 failures last night

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25690 91177308-0d34-0410-b5e6-96231b3b80d8

Following convention.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25689 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the JIT with SSE

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25688 91177308-0d34-0410-b5e6-96231b3b80d8

fix build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25687 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build error that is apparently only a warning with some compilers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25686 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot the version number.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25685 91177308-0d34-0410-b5e6-96231b3b80d8

Improve visibility/correctness of operand indices in "llvm.db" objects.
Handle 64 in DIEs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25684 91177308-0d34-0410-b5e6-96231b3b80d8

Catch up on some of the file addition/deletions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25683 91177308-0d34-0410-b5e6-96231b3b80d8

Fix auto-upgrade of intrinsics to work properly with both assembly and
bytecode reading. This code is crufty, the result of much hacking to get things
working correctly. Cleanup patches will follow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25682 91177308-0d34-0410-b5e6-96231b3b80d8

A test case for ensuring that conversion of old-format intrinsics with
signed arguments get auto-upgraded correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25681 91177308-0d34-0410-b5e6-96231b3b80d8

This now passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25680 91177308-0d34-0410-b5e6-96231b3b80d8

x86 CPU detection and proper subtarget support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25679 91177308-0d34-0410-b5e6-96231b3b80d8

Subtarget feature can now set any variable to any value

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25678 91177308-0d34-0410-b5e6-96231b3b80d8

Use printInlineAsm to, well, print inline asm's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25677 91177308-0d34-0410-b5e6-96231b3b80d8

Stub out a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25676 91177308-0d34-0410-b5e6-96231b3b80d8

add a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25675 91177308-0d34-0410-b5e6-96231b3b80d8

PHI and INLINEASM are now built-in instructions provided by Target.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25674 91177308-0d34-0410-b5e6-96231b3b80d8

PHI and INLINEASM are now builtin instructions provided by Target.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25673 91177308-0d34-0410-b5e6-96231b3b80d8

If we want to emit things in enum order, use getInstructionsByEnumValue to
get the order, don't compute it ourselves.

Don't emit stuff like (14<<0), emit 14 instead.

Don't attempt to get target properties for builtin instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25672 91177308-0d34-0410-b5e6-96231b3b80d8

There is at least a 'noitinerary' itinerary now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25671 91177308-0d34-0410-b5e6-96231b3b80d8

Add a default NoItinerary class for targets to use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25670 91177308-0d34-0410-b5e6-96231b3b80d8

Don't emit JIT code for these instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25669 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the scheduler to emit the appropriate INLINEASM MachineInstr for an
ISD::INLINEASM node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25668 91177308-0d34-0410-b5e6-96231b3b80d8