Fix compilation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44864 91177308-0d34-0410-b5e6-96231b3b80d8

Update credits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44861 91177308-0d34-0410-b5e6-96231b3b80d8

Add information on address space qualifiers for pointer types and global
declarations to the LangRef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44860 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44859 91177308-0d34-0410-b5e6-96231b3b80d8

Implement address space attribute for LLVM pointer types. Address spaces are
regions of memory that have a target specific relationship, as described in the
Embedded C Technical Report.

This also implements the 2007-12-11-AddressSpaces test,
which demonstrates how address space attributes can be used in LLVM IR.

In addition, this patch changes the bitcode signature for stores (in a backwards
compatible manner), such that the pointer type, rather than the pointee type, is
encoded. This permits type information in the pointer (e.g. address space) to be
preserved for stores.

LangRef updates are forthcoming.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44858 91177308-0d34-0410-b5e6-96231b3b80d8

Teach VMCore to constant fold shufflevectors with constant operands.
This allows us to compile:

#include <emmintrin.h>
typedef __m128i VSInt16;
typedef short vSInt16 __attribute__ ((__vector_size__ (16)));
VSInt16 t3() {
  return (VSInt16)((vSInt16)_mm_set1_epi16(6518));
}

into:

_t3:
movaps LCPI1_0, %xmm0
ret

instead of:
_t3:
movl $6518, %eax
movd %eax, %xmm0
pextrw $0, %xmm0, %eax
xorps %xmm0, %xmm0
pinsrw $0, %eax, %xmm0
punpcklwd %xmm0, %xmm0
pshufd $0, %xmm0, %xmm0
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44856 91177308-0d34-0410-b5e6-96231b3b80d8

Implement constant folding if vector<->vector bitcasts where the number
of source/dest elements changes.  This implements
test/Transforms/InstCombine/bitcast-vector-fold.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44855 91177308-0d34-0410-b5e6-96231b3b80d8

Don't have APInt.cpp depend upon DerivedTypes.h. This helps with splitting the
Support libraries separately into their own module.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44852 91177308-0d34-0410-b5e6-96231b3b80d8

Add first and really dirty version of generic Trie structure

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44851 91177308-0d34-0410-b5e6-96231b3b80d8

significantly simplify some code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44850 91177308-0d34-0410-b5e6-96231b3b80d8

refactor some code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44849 91177308-0d34-0410-b5e6-96231b3b80d8

Switch over to MachineLoopInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44838 91177308-0d34-0410-b5e6-96231b3b80d8

Pretty print shuffle mask operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44837 91177308-0d34-0410-b5e6-96231b3b80d8

- Improved v8i16 shuffle lowering. It now uses pshuflw and pshufhw as much as
possible before resorting to pextrw and pinsrw.
- Better codegen for v4i32 shuffles masquerading as v8i16 or v16i8 shuffles.
- Improves (i16 extract_vector_element 0) codegen by recognizing
  (i32 extract_vector_element 0) does not require a pextrw.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44836 91177308-0d34-0410-b5e6-96231b3b80d8

x86 doesn't actually want to custom lower v3i32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44835 91177308-0d34-0410-b5e6-96231b3b80d8

CollectorMetadata and Collector are rejiggered to get along with
per-function collector model. Collector is now the factory for
CollectorMetadata, so the latter may be subclassed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44827 91177308-0d34-0410-b5e6-96231b3b80d8

Project file maintenance.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44826 91177308-0d34-0410-b5e6-96231b3b80d8

Move TargetData::hostIsLittleEndian out of line, which means we
don't have to #include config.h in it.  #including config.h breaks
other projects that have their own autoconf stuff and try to #include
the llvm headers.  One obscure example is llvm-gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44825 91177308-0d34-0410-b5e6-96231b3b80d8

Adding Ocaml bindings for the bitreader as requested by Sarah
Thompson. Usage should be something like this:

open Llvm
open Llvm_bitreader

match read_bitcode_file fn with
  | Bitreader_failure msg ->
      prerr_endline msg
  | Bitreader_success m ->
      ...;
      dispose_module m

Compile with: ocamlc llvm.cma llvm_bitreader.cma
              ocamlopt llvm.cmxa llvm_bitreader.cmxa

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44824 91177308-0d34-0410-b5e6-96231b3b80d8

Hey, English is not my native language :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44820 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify the need of CFI() stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44819 91177308-0d34-0410-b5e6-96231b3b80d8

Provide convenient way to disable CFI stuff for old/broken assemblers.
Use it for Darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44818 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1850 by removing an unsafe transformation from VMCore/ConstantFold.cpp.
Reimplement the xform in Analysis/ConstantFolding.cpp where we can use
targetdata to validate that it is safe.  While I'm in there, fix some const
correctness issues and generalize the interface to the "operand folder".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44817 91177308-0d34-0410-b5e6-96231b3b80d8

Added two bounds checks to the BitVector class to detect
out-of-bounds bit accesses.  The checks are only performed
in a Debug build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44815 91177308-0d34-0410-b5e6-96231b3b80d8

split isBinaryOp into a static and member version.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44814 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some wording.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44810 91177308-0d34-0410-b5e6-96231b3b80d8

Add StringPool + new CodeGen files to win32 build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44805 91177308-0d34-0410-b5e6-96231b3b80d8

Disable cfi directives for now, darwin does't support them.
These should probably be something like:

  CFI(".cfi_def_cfa_offset 16\n")

where CFI is defined to a noop on darwin and other platforms
that don't support those directives.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44803 91177308-0d34-0410-b5e6-96231b3b80d8

Make PruneEH update the nounwind/noreturn attributes
on functions as it calculates them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44802 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1836: in the interpreter, read and write apints
using the minimum possible number of bytes.  For little
endian targets run on little endian machines, apints are
stored in memory from LSB to MSB as before.  For big endian
targets on big endian machines they are stored from MSB to
LSB which wasn't always the case before (if the target and
host endianness doesn't match values are stored according
to the host's endianness).  Doing this requires knowing the
endianness of the host, which is determined when configuring -
thanks go to Anton for this.  Only having access to little
endian machines I was unable to properly test the big endian
part, which is also the most complicated...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44796 91177308-0d34-0410-b5e6-96231b3b80d8

And finally annotate X86-64 version of callback.
All bad stuff from SSE version is implicitely inherited :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44794 91177308-0d34-0410-b5e6-96231b3b80d8

Provide annotation for SSE version of callback. It's even more
broken, because doesn't mark xmm regs properly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44793 91177308-0d34-0410-b5e6-96231b3b80d8

Annotate JIT callback function with call frame infromation.
This will allow us (theoretically) to unwind through JITer.
The code wasn't verified, so I'm pretty sure offsets are wrong :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44792 91177308-0d34-0410-b5e6-96231b3b80d8

It looks like this has been broken for some time -
get it to compile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44791 91177308-0d34-0410-b5e6-96231b3b80d8

A little more progress on StrongPHIElimination, now that I have a better sense of
how the CodeGen machinery works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44786 91177308-0d34-0410-b5e6-96231b3b80d8

Improve branch folding by recgonizing that explict successor relationships impact the value of fall-through choices.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44785 91177308-0d34-0410-b5e6-96231b3b80d8

Delete the CollectorNamePool if it should become empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44775 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo spotted by Nick Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44774 91177308-0d34-0410-b5e6-96231b3b80d8

Adding a collector name attribute to Function in the IR. These
methods are new to Function:

  bool hasCollector() const;
  const std::string &getCollector() const;
  void setCollector(const std::string &);
  void clearCollector();

The assembly representation is as such:

  define void @f() gc "shadow-stack" { ...

The implementation uses an on-the-side table to map Functions to
collector names, such that there is no overhead. A StringPool is
further used to unique collector names, which are extremely
likely to be unique per process.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44769 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting dtor devirtualization patch.
_sabre_: it has a major problem: by the time ~Value is run, all of the "parts" of the derived classes have been destroyed
_sabre_: the vtable lives to fight another day

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44760 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44757 91177308-0d34-0410-b5e6-96231b3b80d8

fix some warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44756 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a big problem in the recent dtor refactoring work, now it passes 2007-11-19-InlineAsm.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44755 91177308-0d34-0410-b5e6-96231b3b80d8

Devirtualizing Value destructor (PR889). Patch by Pawel Kunio!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44747 91177308-0d34-0410-b5e6-96231b3b80d8

Duncan points out that the subtraction is unneeded since hte code
knows the vector is not pow2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44740 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrading this test to 2.0 .ll syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44738 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1782, patch by Wojtek Matyjewicz!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44733 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a very silly typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44732 91177308-0d34-0410-b5e6-96231b3b80d8

Fix accidental commit by Bill.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44729 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for splitting the operand of a return instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44728 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting 44702. It wasn't correct to rename them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44727 91177308-0d34-0410-b5e6-96231b3b80d8

add many new cases to SplitResult.  SplitResult now handles all the cases that LegalizeDAG does.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44726 91177308-0d34-0410-b5e6-96231b3b80d8

Implement splitting support for store, allowing us to compile:

%f8 = type <8 x float>

define void @test_f8(%f8* %P, %f8* %Q, %f8* %S) {
%p = load %f8* %P ; <%f8> [#uses=1]
%q = load %f8* %Q ; <%f8> [#uses=1]
%R = add %f8 %p, %q ; <%f8> [#uses=1]
store %f8 %R, %f8* %S
ret void
}

into:

_test_f8:
movaps 16(%rdi), %xmm0
addps 16(%rsi), %xmm0
movaps (%rdi), %xmm1
addps (%rsi), %xmm1
movaps %xmm0, 16(%rdx)
movaps %xmm1, (%rdx)
ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44725 91177308-0d34-0410-b5e6-96231b3b80d8

implement vector splitting of load, undef, and binops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44724 91177308-0d34-0410-b5e6-96231b3b80d8

implement some methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44723 91177308-0d34-0410-b5e6-96231b3b80d8

add scaffolding for splitting of vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44722 91177308-0d34-0410-b5e6-96231b3b80d8

Project cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44721 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead file and directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44720 91177308-0d34-0410-b5e6-96231b3b80d8

reorganize header to separate into functional blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44719 91177308-0d34-0410-b5e6-96231b3b80d8

split scalarization out to its own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44718 91177308-0d34-0410-b5e6-96231b3b80d8

Split expansion out into its own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44717 91177308-0d34-0410-b5e6-96231b3b80d8

Split promotion support out to its own file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44716 91177308-0d34-0410-b5e6-96231b3b80d8

Rename LegalizeDAGTypes.cpp -> LegalizeTypes.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44715 91177308-0d34-0410-b5e6-96231b3b80d8

Split the class definition of DAGTypeLegalizer out into a header.
Leave it visibility hidden, but not in an anon namespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44714 91177308-0d34-0410-b5e6-96231b3b80d8

Incorporating review feedback from Reid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44713 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44711 91177308-0d34-0410-b5e6-96231b3b80d8

add #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44710 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate dependency on Bitcode headers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44709 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead #include, APInt.h already has the needed forward decls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44708 91177308-0d34-0410-b5e6-96231b3b80d8

proper #include order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44707 91177308-0d34-0410-b5e6-96231b3b80d8

Ignoring generated files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44706 91177308-0d34-0410-b5e6-96231b3b80d8

Adding a StringPool data structure, which GC will use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44705 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug in constructing Ocaml option types in the bindings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44704 91177308-0d34-0410-b5e6-96231b3b80d8

aesthetic changes, no functionality change.  Evan, it's not clear
what 'Available' is, please add a comment near it and rename it
if appropriate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44703 91177308-0d34-0410-b5e6-96231b3b80d8

Renaming:

  isTriviallyReMaterializable -> hasNoSideEffects
  isReallyTriviallyReMaterializable -> isTriviallyReMaterializable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44702 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a significant code quality regression I introduced on PPC64 quite
a while ago.  We now produce:

_foo:
mflr r0
std r0, 16(r1)
ld r2, 16(r1)
std r2, 0(r3)
ld r0, 16(r1)
mtlr r0
blr

instead of:

_foo:
mflr r0
std r0, 16(r1)
lis r0, 0
ori r0, r0, 16
ldx r2, r1, r0
std r2, 0(r3)
ld r0, 16(r1)
mtlr r0
blr

for:

void foo(void **X) {
  *X = __builtin_return_address(0);
}

on ppc64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44701 91177308-0d34-0410-b5e6-96231b3b80d8

implement __builtin_return_addr(0) on ppc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44700 91177308-0d34-0410-b5e6-96231b3b80d8

refactor some code to avoid overloading the name 'usesLR' in
different places to mean different things.  Document what the
one in PPCFunctionInfo means and when it is valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44699 91177308-0d34-0410-b5e6-96231b3b80d8

Implement correct isa<UnaryInstruction>, problem reported by "ST".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44697 91177308-0d34-0410-b5e6-96231b3b80d8

Incorporated comments from Evan and Chris:

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20071203/056043.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20071203/056048.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44696 91177308-0d34-0410-b5e6-96231b3b80d8

Fix several cache coherence bugs in MemDep/GVN that were found.  Also add some (disabled) debugging code
to make such problems easier to diagnose in the future, written by Duncan Sands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44695 91177308-0d34-0410-b5e6-96231b3b80d8

Doh

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44694 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a compilation warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44692 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a compilation warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44691 91177308-0d34-0410-b5e6-96231b3b80d8

Adding MachineLoopInfo.h,.cpp and MachineLICP.cpp to VStudio build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44690 91177308-0d34-0410-b5e6-96231b3b80d8

Initial commit of the machine code LICM pass. It successfully hoists this:

_foo:
        li r2, 0
LBB1_1: ; bb
        li r5, 0
        stw r5, 0(r3)
        addi r2, r2, 1
        addi r3, r3, 4
        cmplw cr0, r2, r4
        bne cr0, LBB1_1 ; bb
LBB1_2: ; return
        blr

to:

_foo:
        li r2, 0
        li r5, 0
LBB1_1: ; bb
        stw r5, 0(r3)
        addi r2, r2, 1
        addi r3, r3, 4
        cmplw cr0, r2, r4
        bne cr0, LBB1_1 ; bb
LBB1_2: ; return
        blr

ZOMG!! :-)

Moar to come...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44687 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44686 91177308-0d34-0410-b5e6-96231b3b80d8

Much improved v8i16 shuffles. (Step 1).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44676 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a stupid error that Chris pointed out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44674 91177308-0d34-0410-b5e6-96231b3b80d8

New test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44672 91177308-0d34-0410-b5e6-96231b3b80d8

Add an option to control this heuristic tweak so I can test it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44671 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a bogus optimization. It's not possible to do a move to low element to a <8 x i16> or <16 x i8> vector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44669 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bogus test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44668 91177308-0d34-0410-b5e6-96231b3b80d8

Redo previous patch so optimization only done for i1.
Simpler and safer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44663 91177308-0d34-0410-b5e6-96231b3b80d8

Turning simple splitting on. Start testing new coalescer heuristics as new llcbeta.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44660 91177308-0d34-0410-b5e6-96231b3b80d8

third time around: instead of disabling this completely,
only disable it if we don't know it will be obviously profitable.
Still fixme, but less so. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44658 91177308-0d34-0410-b5e6-96231b3b80d8

Actually, disable this code for now.  More analysis and improvements to
the X86 backend are needed before this should be enabled by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44657 91177308-0d34-0410-b5e6-96231b3b80d8

implement a readme entry, compiling the code into:

_foo:
movl $12, %eax
andl 4(%esp), %eax
movl _array(%eax), %eax
ret

instead of:

_foo:
movl 4(%esp), %eax
shrl $2, %eax
andl $3, %eax
movl _array(,%eax,4), %eax
ret

As it turns out, this triggers all the time, in a wide variety of
situations, for example, I see diffs like this in various programs:

-       movl    8(%eax), %eax
-       shll    $2, %eax
-       andl    $1020, %eax
-       movl    (%esi,%eax), %eax
+       movzbl  8(%eax), %eax
+       movl    (%esi,%eax,4), %eax

-       shll    $2, %edx
-       andl    $1020, %edx
-       movl    (%edi,%edx), %edx
+       andl    $255, %edx
+       movl    (%edi,%edx,4), %edx

Unfortunately, I also see stuff like this, which can be fixed in the
X86 backend:

-       andl    $85, %ebx
-       addl    _bit_count(,%ebx,4), %ebp
+       shll    $2, %ebx
+       andl    $340, %ebx
+       addl    _bit_count(%ebx), %ebp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44656 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44655 91177308-0d34-0410-b5e6-96231b3b80d8

implement the rest of the functionality from SelectionDAGLegalize::ScalarizeVectorOp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44654 91177308-0d34-0410-b5e6-96231b3b80d8