Move the cloning of the function body much earlier in the inlinefunction
process.  The only optimization we did so far is to avoid creating a
PHI node, then immediately destroying it in the common case where the
callee has one return statement.  Instead, we just don't create the return
value.  This has no noticable performance impact, but paves the way for
future improvements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11108 91177308-0d34-0410-b5e6-96231b3b80d8

Give CloneBasicBlock an optional function argument to specify which function
to add the cloned block to.  This allows the block to be added to the function
immediately, and all of the instructions to be immediately added to the function
symbol table, which speeds up the inliner from 3.7 -> 3.38s on the PR209.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11107 91177308-0d34-0410-b5e6-96231b3b80d8

If changing a parent, don't add then remove the object from the leak detector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11106 91177308-0d34-0410-b5e6-96231b3b80d8

In BasicBlock::splitBasicBlock, just use islist::splice to move the instructions,
instead of a loop that is really inefficient with large basic blocks.

This speeds up the inliner pass on the testcase in PR209 from 13.8s to 2.24s
which still isn't exactly speedy, but is a lot better.  :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11105 91177308-0d34-0410-b5e6-96231b3b80d8

Revised version of a patch for PR222 that works with my version of gmake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11104 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11103 91177308-0d34-0410-b5e6-96231b3b80d8

Apply Michael Kahl's patch to fix the 'make tools-only' target when OBJDIR != SRCDIR.

See PR222.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11102 91177308-0d34-0410-b5e6-96231b3b80d8

Bunch up all locally used allocas by the block they are allocated in, and
process them all as a group.  This speeds up SRoA/mem2reg from 28.46s to
0.62s on the testcase from PR209.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11100 91177308-0d34-0410-b5e6-96231b3b80d8

Handle extremely trivial cases extremely efficiently.  This speeds up
SRoA/mem2reg from 41.2s to 27.5s on the testcase in PR209.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11099 91177308-0d34-0410-b5e6-96231b3b80d8

Generate ftst instructions for comparison against zero

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11098 91177308-0d34-0410-b5e6-96231b3b80d8

bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11097 91177308-0d34-0410-b5e6-96231b3b80d8

Add the ftst instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11095 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for one argument OneArgFP instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11094 91177308-0d34-0410-b5e6-96231b3b80d8

When an instruction like: A += B had both A and B virtual registers
spilled, A was loaded from its stack location twice. This fixes the bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11093 91177308-0d34-0410-b5e6-96231b3b80d8

Revert changes. Will implement this using a different set of primitives

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11091 91177308-0d34-0410-b5e6-96231b3b80d8

Fix debugging output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11088 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly update def/use information for modified machine operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11087 91177308-0d34-0410-b5e6-96231b3b80d8

Add MachineOperand::setDef() and MachineOperand::setUse() so that the
TwoAddressInstructionPass can correctly update use/def information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11086 91177308-0d34-0410-b5e6-96231b3b80d8

Should be more careful. The previously applied change made all counts
outside of loops = 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11085 91177308-0d34-0410-b5e6-96231b3b80d8

Floating point negates are -0.0 - X, not 0.0 - X

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11084 91177308-0d34-0410-b5e6-96231b3b80d8

Disable (x - (y - z)) => (x + (z - y)) optimization for floating point.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11083 91177308-0d34-0410-b5e6-96231b3b80d8

Update comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11082 91177308-0d34-0410-b5e6-96231b3b80d8

No need to declare implicit uses/defs of ST0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11081 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11080 91177308-0d34-0410-b5e6-96231b3b80d8

Make deadarghaX0r warning louder.

(I just love typing haX0r.   haX0r haX0r haX0r.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11079 91177308-0d34-0410-b5e6-96231b3b80d8

Generate the fchs instruction to negate a floating point number

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11078 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for OneArgFPRW instructions, fix a couple of typeos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11077 91177308-0d34-0410-b5e6-96231b3b80d8

Include PowerPC.h.  Flesh out the stub versions of addPassesToEmitAssembly()
and addPassesToJITCompile() slightly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11076 91177308-0d34-0410-b5e6-96231b3b80d8

Add a -march=powerpc option. Automatically select it if this looks like a
big-endian, 32-bit module, or if __ppc__, __POWERPC__, or __APPLE__ are
defined.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11075 91177308-0d34-0410-b5e6-96231b3b80d8

Link in the PowerPC back-end.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11074 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments describing how you would add prototypes for factory methods for
PowerPC-specific passes here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11073 91177308-0d34-0410-b5e6-96231b3b80d8

Add prototype for llvm::allocatePowerPCTargetMachine().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11072 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen -0.0 correctly.  Do not use fldz!  This is another -0.0 == +0.0 problem, arg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11070 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the method I just added to actually return false sometimes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11069 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new method to ConstantFP

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11068 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11067 91177308-0d34-0410-b5e6-96231b3b80d8

Create an object for tracking physical register usage. This will look
much better when I get rid of the reserved registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11066 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11065 91177308-0d34-0410-b5e6-96231b3b80d8

Second half of a fix for PR218 & test/Regression/Assembler/2004-02-01-NegativeZero.llx.
Basically we store floating point values as their integral components, instead of relying
on the semantics of floating point < to differentiate between values.  This is likely to
make the map search be faster anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11064 91177308-0d34-0410-b5e6-96231b3b80d8

The first half of a fix for PR218 & test/Regression/Assembler/2004-02-01-NegativeZero.llx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11063 91177308-0d34-0410-b5e6-96231b3b80d8

Add another test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11062 91177308-0d34-0410-b5e6-96231b3b80d8

Change weight array into a vector and make it as big as the number of
registers (not as the max number of registers).

Change toSpill from a std::set into a std::vector<bool>.

Use the reverse iterator adapter to do a reverse scan of allocatable
registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11061 91177308-0d34-0410-b5e6-96231b3b80d8

Use std::map::count() instead of std::map::find() != std::map::end()
where appropriate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11060 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for PR218

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11059 91177308-0d34-0410-b5e6-96231b3b80d8

Add MRegisterInfo::getNumRegs().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11058 91177308-0d34-0410-b5e6-96231b3b80d8

Change string for joined intervals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11057 91177308-0d34-0410-b5e6-96231b3b80d8

FpMOV is also a move instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11055 91177308-0d34-0410-b5e6-96231b3b80d8

Change xor to ^.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11054 91177308-0d34-0410-b5e6-96231b3b80d8

Now that tailduplication does not bork SSA form, run mem2reg earlier in gccas.
This tremendously improves the code generated by the LLVM optimizer, primarily
by making the inliner more aggressive.  For example, it improves the stepanov
benchmark from 55.56 mega-additions/sec to 98.04 Ma/s.  It also improves the
oopack/iterator benchmark from 338.3MFLOPS/s to 1103.4MFLOPS/s.  Less noteworthy,
it improves oopack/matrix from 573 -> 641 MFLOPS/s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11053 91177308-0d34-0410-b5e6-96231b3b80d8

Disable tail duplication in any "hard" cases, where it might break SSA form.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11052 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in a recent checkin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11050 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the count of the number of instructions removed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11049 91177308-0d34-0410-b5e6-96231b3b80d8

Missed one silly assert :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11048 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify joinIntervals() code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11047 91177308-0d34-0410-b5e6-96231b3b80d8

Remove all of the annoying statistics now that I'm finished (for the near
term) working on bytecode size stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11046 91177308-0d34-0410-b5e6-96231b3b80d8

Use MRegisterInfo::isPhysicalRegister and
MRegisterInfo::isVirtualRegister.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11045 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a crasher bug in my constant folding rewrite

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11044 91177308-0d34-0410-b5e6-96231b3b80d8

Print an error message if we can't materialize the bytecode file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11043 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11042 91177308-0d34-0410-b5e6-96231b3b80d8

Print an error message if there is an error materialize the bc file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11041 91177308-0d34-0410-b5e6-96231b3b80d8

Apply final part of Chris' patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11040 91177308-0d34-0410-b5e6-96231b3b80d8

Be a little smarter on the way we handle physical register defs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11038 91177308-0d34-0410-b5e6-96231b3b80d8

Finegrainify namespacification, use new MRegisterInfo::isVirtualRegister
method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11037 91177308-0d34-0410-b5e6-96231b3b80d8

Fix, correctly this time, the computation of the return value
Fix a spello
Tighten up the assertion checking

No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11036 91177308-0d34-0410-b5e6-96231b3b80d8

* Fix incorrect computation of the runOnMachineFunction return value
* Turn a bunch of instance variables into automatic variables

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11035 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unneeded #includes
Move Passes.h (which defines the interface to this file) to the top.
Move statistics to the top of the file.
Add a comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11034 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid referencing deleted DSgraphs when merging an SCC into a larger SCC.  This
fixes the crash in 176.gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11033 91177308-0d34-0410-b5e6-96231b3b80d8

Merge safe parts from last night's buggy commit. These do not break
any test cases :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11032 91177308-0d34-0410-b5e6-96231b3b80d8

Add two static methods to avoid having client code explicitly compare against
FirstVirtualRegister

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11031 91177308-0d34-0410-b5e6-96231b3b80d8

Optimize liveAt() and overlaps(). We now use a binary search instead
of a linear search to find the first range for comparisons. This cuts
down the linear scan register allocator running time by a factor of 3
in 254.perlbmk and by a factor of 2.2 in 176.gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11030 91177308-0d34-0410-b5e6-96231b3b80d8

Revert last night's changes as they broke some tests. Will remerge parts of the patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11029 91177308-0d34-0410-b5e6-96231b3b80d8

Several performance enhancements and cleanups from Chris.

Simplification of LiveIntervals::Interval::overlaps() and addition of
examples to overlaps() and liveAt() to make them clearer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11028 91177308-0d34-0410-b5e6-96231b3b80d8

Fix thinko

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11027 91177308-0d34-0410-b5e6-96231b3b80d8

Add some comments sketching out how this is to work eventually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11026 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new flag, which is only used for symmetry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11025 91177308-0d34-0410-b5e6-96231b3b80d8

Forward method request to chained aa implementation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11024 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for better mod/ref information that basicaa can provide

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11023 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the pointsToConstantMemory() method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11022 91177308-0d34-0410-b5e6-96231b3b80d8

Improve mod/ref information based on the pointsToConstantMemory method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11021 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new pointsToConstantMemory method to the AliasAnalysis interface
which can be implemented to improve the quality of mod-ref information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11020 91177308-0d34-0410-b5e6-96231b3b80d8

Add (currently disabled) support to the instruction selector to only insert
FP_REG_KILL instructions at the end of blocks involved with critical edges.

Fix a bug where FP_REG_KILL instructions weren't inserted in fall through
unconditional branches.  Perhaps this will fix some linscan problems?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11019 91177308-0d34-0410-b5e6-96231b3b80d8

Finegrainify namespacification
Implement LiveVariables::getIndexMachineBasicBlock

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11018 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new lazily constructed mapping from Idx's the MBB they represent

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11017 91177308-0d34-0410-b5e6-96231b3b80d8

Give clients of MachineFunctionPrinter the ability to specify a banner and
choose an ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11016 91177308-0d34-0410-b5e6-96231b3b80d8

Order #includes alphabetically, per style guide.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11015 91177308-0d34-0410-b5e6-96231b3b80d8

Doxygenify comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11014 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug aflicting 265.gap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11006 91177308-0d34-0410-b5e6-96231b3b80d8

Minor bugfixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11005 91177308-0d34-0410-b5e6-96231b3b80d8

I think this is a more robust fix for the Solaris wchar problems (PR206).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11004 91177308-0d34-0410-b5e6-96231b3b80d8

Hyphenate `target-dependent'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11003 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new (static inline) std::ostream& << AllocInfo& method. Use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11002 91177308-0d34-0410-b5e6-96231b3b80d8

Rename DSGraph::ScalarMapTy -> DSScalarMap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11001 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11000 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate the call to removeTriviallyDeadNodes from updateFromGlobals graph,
moving it to the start of removeDeadNodes.  This speeds up DSA by 2s on perlbmk
from 41s

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10999 91177308-0d34-0410-b5e6-96231b3b80d8

In the TD pass, iterate over globals directly instead of through the whole scalar
map.  This saves 5s in the TD pass, from 22->17s on perlbmk

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10998 91177308-0d34-0410-b5e6-96231b3b80d8

In the TD pass, don't iterate over the scalar map to find the globals, iterate over
the globals directly.  This doesn't save any substantial time, however, because the
globals graph only contains globals!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10997 91177308-0d34-0410-b5e6-96231b3b80d8

In updateFromGlobalsGraph, instead of iterating over all of the scalars in the
function to find the globals, iterate over all of the globals directly.  This
speeds the function up from 14s to 6.3s on perlbmk, reducing DSA time from
53->46s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10996 91177308-0d34-0410-b5e6-96231b3b80d8

Keep track of all of the globals inserted into the scalar map

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10995 91177308-0d34-0410-b5e6-96231b3b80d8

Pull the ScalarMap out into something that is more structured than what we had
before.  This allows us to have a place to implement optimizations in a
structured way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10994 91177308-0d34-0410-b5e6-96231b3b80d8

Minor tweaks, eliminate useless integer pruning optimziation, turn on
timers by default

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10993 91177308-0d34-0410-b5e6-96231b3b80d8