Case sensitive not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31827 91177308-0d34-0410-b5e6-96231b3b80d8

Check for MaxAlign.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31826 91177308-0d34-0410-b5e6-96231b3b80d8

Tests to verify PowerPC ABI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31825 91177308-0d34-0410-b5e6-96231b3b80d8

1. Ignore the -disable-fp-elim when the routine is a leaf.
2. Offsets on 64-bit stores are still in bytes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31824 91177308-0d34-0410-b5e6-96231b3b80d8

Typo. Fix the nightly tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31823 91177308-0d34-0410-b5e6-96231b3b80d8

Fixing the ENABLE_OPTIMIZED=1 DISABLE_ASSERTIONS=1 build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31822 91177308-0d34-0410-b5e6-96231b3b80d8

Need <iostream> for the time being.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31821 91177308-0d34-0410-b5e6-96231b3b80d8

Need iostream to be included for the time being.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31820 91177308-0d34-0410-b5e6-96231b3b80d8

Moved definition of llvm_ostream wrappers to the Streams.cpp file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31819 91177308-0d34-0410-b5e6-96231b3b80d8

Used llvm_ostream instead of std::ostream objects. This will reduce use
of the icky <iostream> class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31818 91177308-0d34-0410-b5e6-96231b3b80d8

Added wrappers for the std::cerr/std::cout objects. The wrappers will
soon replace all uses of those objects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31817 91177308-0d34-0410-b5e6-96231b3b80d8

Needed <iostream> for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31816 91177308-0d34-0410-b5e6-96231b3b80d8

Needs the iostream include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31815 91177308-0d34-0410-b5e6-96231b3b80d8

Removed iostream #includes. Replaced std::cerr with DOUT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31814 91177308-0d34-0410-b5e6-96231b3b80d8

Removed even more std::cerr and #include <iostream> things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31813 91177308-0d34-0410-b5e6-96231b3b80d8

Replaced DEBUG(std::cerr with DOUT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31812 91177308-0d34-0410-b5e6-96231b3b80d8

Replace DEBUG(std::cerr with DOUT. Removed some iostream #includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31811 91177308-0d34-0410-b5e6-96231b3b80d8

Removed unneeded <iostream> #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31810 91177308-0d34-0410-b5e6-96231b3b80d8

If an indvar with a variable stride is used by the exit condition, go ahead
and handle it like constant stride vars.  This fixes some bad codegen in
variable stride cases.  For example, it compiles this:

void foo(int k, int i) {
  for (k=i+i; k <= 8192; k+=i)
    flags2[k] = 0;
}

to:

LBB1_1: #bb.preheader
        movl %eax, %ecx
        addl %ecx, %ecx
        movl L_flags2$non_lazy_ptr, %edx
LBB1_2: #bb
        movb $0, (%edx,%ecx)
        addl %eax, %ecx
        cmpl $8192, %ecx
        jle LBB1_2      #bb
LBB1_5: #return
        ret

or (if the array is local and we are in dynamic-nonpic or static mode):

LBB3_2: #bb
        movb $0, _flags2(%ecx)
        addl %eax, %ecx
        cmpl $8192, %ecx
        jle LBB3_2      #bb

and:

        lis r2, ha16(L_flags2$non_lazy_ptr)
        lwz r2, lo16(L_flags2$non_lazy_ptr)(r2)
        slwi r3, r4, 1
LBB1_2: ;bb
        li r5, 0
        add r6, r4, r3
        stbx r5, r2, r3
        cmpwi cr0, r6, 8192
        bgt cr0, LBB1_5 ;return

instead of:

        leal (%eax,%eax,2), %ecx
        movl %eax, %edx
        addl %edx, %edx
        addl L_flags2$non_lazy_ptr, %edx
        xorl %esi, %esi
LBB1_2: #bb
        movb $0, (%edx,%esi)
        movl %eax, %edi
        addl %esi, %edi
        addl %ecx, %esi
        cmpl $8192, %esi
        jg LBB1_5       #return

and:

        lis r2, ha16(L_flags2$non_lazy_ptr)
        lwz r2, lo16(L_flags2$non_lazy_ptr)(r2)
        mulli r3, r4, 3
        slwi r5, r4, 1
        li r6, 0
        add r2, r2, r5
LBB1_2: ;bb
        li r5, 0
        add r7, r3, r6
        stbx r5, r2, r6
        add r6, r4, r6
        cmpwi cr0, r7, 8192
        ble cr0, LBB1_2 ;bb

This speeds up Benchmarks/Shootout/sieve from 8.533s to 6.464s and
implements LoopStrengthReduce/var_stride_used_by_compare.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31809 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31808 91177308-0d34-0410-b5e6-96231b3b80d8

Undo removal of the runtime libraries. While this may have been a bit
premature, these libraries will be going away for the 2.0 release. Other
arrangements for profiling, gc, etc. should be made in the next few months.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31807 91177308-0d34-0410-b5e6-96231b3b80d8

More removal of std::cerr and DEBUG, replacing with DOUT instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31806 91177308-0d34-0410-b5e6-96231b3b80d8

implement a todo: change a map into a vector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31805 91177308-0d34-0410-b5e6-96231b3b80d8

Add opcode to TargetInstrDescriptor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31804 91177308-0d34-0410-b5e6-96231b3b80d8

#ifndef NDEBUG the "if (Stream) ..." stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31803 91177308-0d34-0410-b5e6-96231b3b80d8

Add opcode to TargetInstrDescriptor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31802 91177308-0d34-0410-b5e6-96231b3b80d8

Adjusted the // comments so that doxygen picks them up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31801 91177308-0d34-0410-b5e6-96231b3b80d8

Replace std::cerr uses of the "DEBUG" macro with "DOUT" instead. Removes
a #include of iostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31800 91177308-0d34-0410-b5e6-96231b3b80d8

fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31799 91177308-0d34-0410-b5e6-96231b3b80d8

Added "DOUT" macro. This is used as a replacement for the std::cerr
stream. It centralizes the use of std::cerr so that static c'tor/d'tors
aren't scattered around all over the place. The way to use it is like this:

       DOUT << "This is a status line: " << Var << "\n";

If "-debug" is specified, it will print. Otherwise, it'll not print. If
NDEBUG is defined, the DOUT does nothing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31798 91177308-0d34-0410-b5e6-96231b3b80d8

implicit_def_vrrc doesn't generate code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31797 91177308-0d34-0410-b5e6-96231b3b80d8

New test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31796 91177308-0d34-0410-b5e6-96231b3b80d8

Correct instructions for moving data between GR64 and SSE registers; also correct load i64 / store i64 from v2i64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31795 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a potential bug: MOVPDI2DI, etc. are not copy instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31794 91177308-0d34-0410-b5e6-96231b3b80d8

In LLVM 2.0 we won't use the runtime libraries as llvm-gcc3 support will
be dropped. This patch pertains to removing the runtime directory from LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31793 91177308-0d34-0410-b5e6-96231b3b80d8

This is a general clean up of the PowerPC ABI.  Address several problems and
bugs including making sure that the TOS links back to the previous frame,
that the maximum call frame size is not included twice when using frame
pointers, no longer growing the frame on calls, double storing of SP and
a cleaner/faster dynamic alloca.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31792 91177308-0d34-0410-b5e6-96231b3b80d8

fix a regression that I introduced.  stdu should scale the offset by 4
before printing it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31791 91177308-0d34-0410-b5e6-96231b3b80d8

Align stubs on 4 byte boundary. This fixes 447.dealII.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31790 91177308-0d34-0410-b5e6-96231b3b80d8

*** empty log message ***

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31789 91177308-0d34-0410-b5e6-96231b3b80d8

Allow target to specify alignment for function stub.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31788 91177308-0d34-0410-b5e6-96231b3b80d8

Match MachineCodeEmitter changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31787 91177308-0d34-0410-b5e6-96231b3b80d8

Use a release version of bugpoint, if found. Include gccld passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31786 91177308-0d34-0410-b5e6-96231b3b80d8

add a statistic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31785 91177308-0d34-0410-b5e6-96231b3b80d8

Cleaned up the document. Added "doc_code" divisions for code examples.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31784 91177308-0d34-0410-b5e6-96231b3b80d8

Don't recompute getNumOperands for each iteration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31783 91177308-0d34-0410-b5e6-96231b3b80d8

Added "removeRange", which takes and removes an entire LiveRange.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31781 91177308-0d34-0410-b5e6-96231b3b80d8

Added a new method "CreateNewLiveInterval" which, given a list of
LiveRanges, creates a new LiveInterval from them. The LiveRanges should
have existed already in another LiveInterval, but removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31780 91177308-0d34-0410-b5e6-96231b3b80d8

make this harder

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31779 91177308-0d34-0410-b5e6-96231b3b80d8

fix broken encoding

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31778 91177308-0d34-0410-b5e6-96231b3b80d8

enable the branch folding pass for the JIT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31777 91177308-0d34-0410-b5e6-96231b3b80d8

add ppc64 r+i stores with update.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31776 91177308-0d34-0410-b5e6-96231b3b80d8

add patterns for ppc32 preinc stores.  ppc64 next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31775 91177308-0d34-0410-b5e6-96231b3b80d8

switch these back to the 'bad old way'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31774 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an incorrectly inverted condition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31773 91177308-0d34-0410-b5e6-96231b3b80d8

ChangeToRegister should clear IsImp bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31772 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ppc64 epilog bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31771 91177308-0d34-0410-b5e6-96231b3b80d8

Stop using isTwoAddress, switching to operand constraints instead.

Tell the codegen emitter that specific operands are not to be encoded, fixing
JIT regressions w.r.t. pre-inc loads and stores (e.g. lwzu, which we generate
even when general preinc loads are not enabled).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31770 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the isTwoAddress property from the CodeGenInstruction class.  It should
not be used for anything other than backwards compat constraint handling.

Add support for a new DisableEncoding property which contains a list of
registers that should not be encoded by the generated code emitter.  Convert
the codeemitter generator to use this, fixing some PPC JIT regressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31769 91177308-0d34-0410-b5e6-96231b3b80d8

add a new field needed by the code emitter generator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31768 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo pointed out by Bryan Adams

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31766 91177308-0d34-0410-b5e6-96231b3b80d8

Properly transfer kill / dead info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31765 91177308-0d34-0410-b5e6-96231b3b80d8

Kill / dead info has been moved to MI's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31764 91177308-0d34-0410-b5e6-96231b3b80d8

commuteInstruction should propagate kill / dead info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31763 91177308-0d34-0410-b5e6-96231b3b80d8

Match live variable changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31762 91177308-0d34-0410-b5e6-96231b3b80d8

Minor updates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31761 91177308-0d34-0410-b5e6-96231b3b80d8

A register def can be partially dead when the whole register has use(s) but a subreg does not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31760 91177308-0d34-0410-b5e6-96231b3b80d8

Do away with kill / dead maps. Move kill / dead info onto MI's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31759 91177308-0d34-0410-b5e6-96231b3b80d8

Add copyKillDeadInfo to copy kill / dead info; other minor updates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31758 91177308-0d34-0410-b5e6-96231b3b80d8

fix ldu/stu jit encoding. Swith 64-bit preinc load instrs to use memri
addrmodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31757 91177308-0d34-0410-b5e6-96231b3b80d8

Add run(Function &F) support in FunctionPassManager_New

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31756 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify IntrinsicLowering and clarify that it is only for use by the
CBE and interpreter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31755 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unneeded forward decls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31754 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31753 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the PPC regressions last night

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31752 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a gcc 4.2 warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31751 91177308-0d34-0410-b5e6-96231b3b80d8

Add a method to get the bit width of a packed type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31750 91177308-0d34-0410-b5e6-96231b3b80d8

Switch loads over to use memri as the operand instead of a reg/imm operand
pair for cleanliness.  Add instructions for PPC32 preinc-stores with commented
out patterns.  More improvement is needed to enable the patterns, but we're
getting close.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31749 91177308-0d34-0410-b5e6-96231b3b80d8

ADd support for adding constraints to suboperands

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31748 91177308-0d34-0410-b5e6-96231b3b80d8

Add doInitialization and doFinalization support in FunctionManager_New.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31747 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up marking of last analysis user pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31746 91177308-0d34-0410-b5e6-96231b3b80d8

Add doInitialization and doFinalization support in FunctionManager_New.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31745 91177308-0d34-0410-b5e6-96231b3b80d8

Do not derive CommonPassManagerImpl from Pass.
Now BasicBlockPassManager_New is a FunctionPass,
FunctionPassManager_New is a ModulePass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31744 91177308-0d34-0410-b5e6-96231b3b80d8

restore some 'magic' code that I removed: it is needed.  Add comments explaining
why.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31743 91177308-0d34-0410-b5e6-96231b3b80d8

minimal hack to get patterns whose result type is iPTR to be selected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31742 91177308-0d34-0410-b5e6-96231b3b80d8

Copy initializeAnalysisImpl() implementation from PassManagerT.
Update LastUser to recursively walk required transitive set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31741 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31740 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for nodes that return iPTR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31739 91177308-0d34-0410-b5e6-96231b3b80d8

changes to get ptr_rc to be accepted in patterns.  This is needed for ppc preinc
stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31738 91177308-0d34-0410-b5e6-96231b3b80d8

Revert. This wasn't meant to be checked in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31737 91177308-0d34-0410-b5e6-96231b3b80d8

group load and store instructions together.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31736 91177308-0d34-0410-b5e6-96231b3b80d8

Fix predicates for unindexed stores so they don't accidentally match indexed
stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31735 91177308-0d34-0410-b5e6-96231b3b80d8

Rework PPC64 calls.  Now we have a LR8/CTR8 register which the PPC64 calls
clobber.  This allows LR8 to be save/restored correctly as a 64-bit quantity,
instead of handling it as a 32-bit quantity.  This unbreaks ppc64 codegen when
the code is actually located above the 4G boundary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31734 91177308-0d34-0410-b5e6-96231b3b80d8

remove a ton of custom selection logic no longer needed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31733 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug handling nodes with variable arguments.  The code was fixed to assume
that there were two input operands before the variable operand portion.  This
*happened* to be true for all call instructions, which took a chain and a
destination, but was not true for the PPC BCTRL instruction, whose destination
is implicit.

Making this code more general allows elimination of the custom selection logic
for BCTRL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31732 91177308-0d34-0410-b5e6-96231b3b80d8

it would be nice of ctlz were lowered to bsf etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31730 91177308-0d34-0410-b5e6-96231b3b80d8

implement InstCombine/shift-simplify.ll by transforming:
(X >> Z) op (Y >> Z)  -> (X op Y) >> Z

for all shifts and all ops={and/or/xor}.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31729 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31728 91177308-0d34-0410-b5e6-96231b3b80d8

implement InstCombine/and-compare.ll:test1.  This compiles:

typedef struct { unsigned prefix : 4; unsigned code : 4; unsigned unsigned_p : 4; } tree_common;
int foo(tree_common *a, tree_common *b) { return a->code == b->code; }

into:

_foo:
        movl 4(%esp), %eax
        movl 8(%esp), %ecx
        movl (%eax), %eax
        xorl (%ecx), %eax
        # TRUNCATE movb %al, %al
        shrb $4, %al
        testb %al, %al
        sete %al
        movzbl %al, %eax
        ret

instead of:

_foo:
        movl 8(%esp), %eax
        movb (%eax), %al
        shrb $4, %al
        movl 4(%esp), %ecx
        movb (%ecx), %cl
        shrb $4, %cl
        cmpb %al, %cl
        sete %al
        movzbl %al, %eax
        ret

saving one cycle by eliminating a shift.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31727 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31726 91177308-0d34-0410-b5e6-96231b3b80d8

lower "X = seteq Y, Z" to '(shr (ctlz (xor Y, Z)), 5)' instead of
'(shr (ctlz (sub Y, Z)), 5)'.

The use of xor better exposes the operation to bit-twiddling logic in the
dag combiner.  For example, this:

typedef struct {
  unsigned prefix : 4;
  unsigned code : 4;
  unsigned unsigned_p : 4;
} tree_common;

int foo(tree_common *a, tree_common *b) {
  return a->code == b->code;
}

Now compiles to:

_foo:
        lwz r2, 0(r4)
        lwz r3, 0(r3)
        xor r2, r3, r2
        rlwinm r2, r2, 28, 28, 31
        cntlzw r2, r2
        srwi r3, r2, 5
        blr

instead of:

_foo:
        lbz r2, 3(r4)
        lbz r3, 3(r3)
        srwi r2, r2, 4
        srwi r3, r3, 4
        subf r2, r2, r3
        cntlzw r2, r2
        srwi r3, r2, 5
        blr

saving a cycle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@31725 91177308-0d34-0410-b5e6-96231b3b80d8