upgrade to use new intrinsics, patch by Dan Hipschman!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110735 91177308-0d34-0410-b5e6-96231b3b80d8

Add the minimal amount of smarts necessary to instcombine of shufflevectors to recognize
patterns generated by clang for transpose of a matrix in generic vectors.  This is made
of two parts:

1) Propagating vector extracts of hi/lo half into their users
2) Recognizing an insertion of even elements followed by the odd elements as an unpack.

Testcase to come, but this shrinks the # of shuffle instructions generated on x86 from ~40 to the minimal 8.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110734 91177308-0d34-0410-b5e6-96231b3b80d8

Turn optimize compares back on with fix. We needed to test that a machine op was
a register before checking if it was defined.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110733 91177308-0d34-0410-b5e6-96231b3b80d8

Give up on register class recalculation when the register is used with subreg
operands. We don't currently have a hook to provide "the largest super class of
A where all registers' getSubReg(subidx) is valid and in B".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110730 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r110718; it broke clang-i386-darwin9.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110726 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid editing the current live interval during remat.

The live interval may be used for a spill slot as well, and that spill slot
could be shared by split registers. We cannot shrink it, even if we know the
current register won't need the spill slot in that range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110721 91177308-0d34-0410-b5e6-96231b3b80d8

More debug spew

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110720 91177308-0d34-0410-b5e6-96231b3b80d8

Turn optimize cmps on by default so that we can get some testing by the nightly
ARM testers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110718 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing argument. CreateCompositeTypeEx() users, please verify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110717 91177308-0d34-0410-b5e6-96231b3b80d8

Switch over to using ConstantRange to track integral values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110714 91177308-0d34-0410-b5e6-96231b3b80d8

Do not forget debug info for enums. Use named mdnode to keep track of these types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110712 91177308-0d34-0410-b5e6-96231b3b80d8

tests: Don't error out if HOME isn't present in t the environment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110711 91177308-0d34-0410-b5e6-96231b3b80d8

Delete some unused instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110710 91177308-0d34-0410-b5e6-96231b3b80d8

Re-apply r110655 with fixes. Epilogue must restore sp from fp if the function stack frame has a var-sized object.

Also added a test case to check for the added benefit of this patch: it's optimizing away the unnecessary restore of sp from fp for some non-leaf functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110707 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to set the flags passed to the assembler.

Nick, please review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110705 91177308-0d34-0410-b5e6-96231b3b80d8

Implement register class inflation.

When splitting a live range, the new registers have fewer uses and the
permissible register class may be less constrained. Recompute the register class
constraint from the uses of new registers created for a split. This may let them
be allocated from a larger set, possibly avoiding a spill.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110703 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r110655, "Fix ARM hasFP() semantics. It should return true whenever FP
register is", it breaks a couple test-suite tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110701 91177308-0d34-0410-b5e6-96231b3b80d8

MC/AsmParser: Fix a bug in macro argument parsing, which was dropping
parentheses from argument lists.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110692 91177308-0d34-0410-b5e6-96231b3b80d8

Install llvmCore_Sim to the simulator SDK directory.  Radar 8282845.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110687 91177308-0d34-0410-b5e6-96231b3b80d8

Recalculate the spill weight and allocation hint for virtual registers created
during live range splitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110686 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test for more architectures. Patch by Tobias Grosser.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110685 91177308-0d34-0410-b5e6-96231b3b80d8

Fix silly bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110684 91177308-0d34-0410-b5e6-96231b3b80d8

Revert "MC/MachO: Fix possible null pointer dereference."

This reverts commit 110575.

Target.isAbsolute() is true if SD would be null.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110683 91177308-0d34-0410-b5e6-96231b3b80d8

Use RunPassesOn as in the rest of bugpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110682 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify generating LLVMC_BUILTIN_PLUGIN.

Patch by NAKAMURA Takumi!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110680 91177308-0d34-0410-b5e6-96231b3b80d8

Trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110679 91177308-0d34-0410-b5e6-96231b3b80d8

Mark this variable as used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110667 91177308-0d34-0410-b5e6-96231b3b80d8

RegionInfo: Do not assert if a BB is not part of the dominance tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110665 91177308-0d34-0410-b5e6-96231b3b80d8

Fix failing testcase.

Those look like typos to me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110664 91177308-0d34-0410-b5e6-96231b3b80d8

Handle TAG_constant for integers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110656 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ARM hasFP() semantics. It should return true whenever FP register is
reserved, not available for general allocation. This eliminates all the
extra checks for Darwin.

This change also fixes the use of FP to access frame indices in leaf
functions and cleaned up some confusing code in epilogue emission.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110655 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMake...sorry for the breakage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110654 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110653 91177308-0d34-0410-b5e6-96231b3b80d8

Drop "const". It does not add value here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110652 91177308-0d34-0410-b5e6-96231b3b80d8

Add AVX movnt{pd,ps,dq} 256-bit intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110650 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing null check reported by Amaury Pouly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110649 91177308-0d34-0410-b5e6-96231b3b80d8

Add AVX movmsk 256-bit intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110648 91177308-0d34-0410-b5e6-96231b3b80d8

Remove replicate intrinsics, clang will generate shufflevector for those. The shuffles can't be matched by x86 codegen yet, but will soon

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110647 91177308-0d34-0410-b5e6-96231b3b80d8

Support AVX 256-bit load and store intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110645 91177308-0d34-0410-b5e6-96231b3b80d8

Use i32 instead of i8 for dot product intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110643 91177308-0d34-0410-b5e6-96231b3b80d8

Do not include file static variable in pubnames list.
Refactor and simplify code to avoid redundant checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110642 91177308-0d34-0410-b5e6-96231b3b80d8

Expand uses of python 2.6's "A if B else C" syntax into regular
if-else statements, to hopefully support older pythons (PR7850).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110638 91177308-0d34-0410-b5e6-96231b3b80d8

remove code setting rw locks to PTHREAD_PROCESS_PRIVATE, which
is the default.  Patch by NAKAMURA Takumi!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110636 91177308-0d34-0410-b5e6-96231b3b80d8

Patterns to match AVX cmp instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110633 91177308-0d34-0410-b5e6-96231b3b80d8

Transpose the calculation of spill weights such that we are calculating one
register at a time. This turns out to be slightly faster than iterating over
instructions, but more importantly, it allows us to compute spill weights for
new registers created after the spill weight pass has run.

Also compute the allocation hint at the same time as the spill weight. This
allows us to use the spill weight as a cost metric for copies, and choose the
most profitable hint if there is more than one possibility.

The new hints provide a very small (< 0.1%) but universal code size improvement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110631 91177308-0d34-0410-b5e6-96231b3b80d8

Add matching patterns for vblend AVX intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110630 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the last argument type of AVX vblend intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110628 91177308-0d34-0410-b5e6-96231b3b80d8

Merge the OptimizeExts and OptimizeCmps passes into one PeepholeOptimizer
pass. This pass should expand with all of the small, fine-grained optimization
passes to reduce compile time and increase happiment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110627 91177308-0d34-0410-b5e6-96231b3b80d8

Undo accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110623 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify. Avoid redundant checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110621 91177308-0d34-0410-b5e6-96231b3b80d8

Wording.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110618 91177308-0d34-0410-b5e6-96231b3b80d8

Next bit of support for the dwarf .file directive.  This patch takes the
previously collected info from the .file directives and outputs the encoded
bytes for it.  For now this is only in the Mach-O streamer but at some point
will move to a more generic place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110617 91177308-0d34-0410-b5e6-96231b3b80d8

ARMBaseRegisterInfo::hasFP() has been broken for a while now. :-(
This will always be false before PEI:
(DisableFramePointerElim(MF) && MFI->adjustsStack())
Which means it's going to make r11 available as a general purpose register even
if -disable-fp-elim is specified. It's working on Darwin only because r7 is
always reserved. But it's obviously broken for other targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110614 91177308-0d34-0410-b5e6-96231b3b80d8

Add VCVTPD2PS, VCVTPS2DQ, VCVTPS2PDY, VCVTTPD2DQY, VCVTTPS2DQ and VCVTPD2DQ 256-bit conversion intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110608 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110607 91177308-0d34-0410-b5e6-96231b3b80d8

Add patterns to AVX conversions instructions. Do that instead of declaring more intructions whenever is possible, more coming

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110605 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to set the target triple and expose that with an option in the
gold plugin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110604 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a use after free error caught by the valgrind builders.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110601 91177308-0d34-0410-b5e6-96231b3b80d8

Refactoring. Update DbgVarible to handle queries itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110600 91177308-0d34-0410-b5e6-96231b3b80d8

Add ConstantRange information to the debugging output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110598 91177308-0d34-0410-b5e6-96231b3b80d8

PR7853: fix a silly mistake introduced in r101899, and add a test to make sure
it doesn't regress again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110597 91177308-0d34-0410-b5e6-96231b3b80d8

Update cmake library dependencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110594 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: eliminated unnecessary target_link_libraries.

Next time the build is broken due to wrong library dependencies, just
try building again (if you are on some Unix and are building all LLVM
targets) or ask someone to commit the regenerated LLVMLibDeps.cmake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110593 91177308-0d34-0410-b5e6-96231b3b80d8

It is ok, and convenient, to pass descriptors by value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110590 91177308-0d34-0410-b5e6-96231b3b80d8

A REG_SEQUENCE instruction may use the same register twice.

If we are emitting COPY instructions for the REG_SEQUENCE, make sure the kill
flag goes on the last COPY. Otherwise we may be using a killed register.

<rdar://problem/8287792>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110589 91177308-0d34-0410-b5e6-96231b3b80d8

Explicitly initialize SlowFPBrcc and Pref32BitThumb to false.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110587 91177308-0d34-0410-b5e6-96231b3b80d8

Rename a method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110586 91177308-0d34-0410-b5e6-96231b3b80d8

Change -prefer-32bit-thumb to attribute -mattr=+32bit instead to disable more 32-bit to 16-bit optimizations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110584 91177308-0d34-0410-b5e6-96231b3b80d8

Memory version of vcvtdq2pd intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110582 91177308-0d34-0410-b5e6-96231b3b80d8

Patterns to match vinsert, vbroadcast, vmovmask and vcvtdq2pd AVX intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110580 91177308-0d34-0410-b5e6-96231b3b80d8

Add an option to disable 32 -> 16-bit Thumb2 size reduction pass for experimentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110579 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the ValueMap operator=, which was wrong (it did't correct the
Map pointers of any contained ValueMapCallbackVH's) and unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110577 91177308-0d34-0410-b5e6-96231b3b80d8

Have SPU handle halfvec stores aligned by 8 bytes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110576 91177308-0d34-0410-b5e6-96231b3b80d8

MC/MachO: Fix possible null pointer dereference.

Discovered by Microsoft Visual Studio 2010 Code Analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110575 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL for mingw that has no plugins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110574 91177308-0d34-0410-b5e6-96231b3b80d8

Don't try to build a plugin on windows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110573 91177308-0d34-0410-b5e6-96231b3b80d8

Stop the JIT from refusing to work just because the program it was compiled into
was built with -static.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110564 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: llvm_map_components_to_libraries now returns system libs too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110563 91177308-0d34-0410-b5e6-96231b3b80d8

Change the Interval type to signed in ImmutableIntervalMap.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110562 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: document how to incorporate LLVM into the cmake project of an
external project.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110561 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: system for providing llvm-config-like features to the user.

The user can use a cmake function for obtaining the LLVM libraries
corresponding to a list of LLVM components.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110560 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: documented testing procedure

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110559 91177308-0d34-0410-b5e6-96231b3b80d8

llc: Fix help typo as pointed out by Nick Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110556 91177308-0d34-0410-b5e6-96231b3b80d8

Add a opt-args option that can be used to pass arguments to every opt
invocation. Fixes PR7793:

bugpoint -debug test.ll --opt-args -unroll-count=4

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110555 91177308-0d34-0410-b5e6-96231b3b80d8

s/libLLVMgold/LLVMgold/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110552 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110551 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the ValueMap copy constructor.  It's not used anywhere,
and removing it catches the mistake of passing a ValueMap by
copy rather than by reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110549 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110548 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bit of a hack to avoid multiple defines of variables in
config.h and llvm-config.h.  This could probably be improved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110547 91177308-0d34-0410-b5e6-96231b3b80d8

Add optimization to Target/README.txt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110543 91177308-0d34-0410-b5e6-96231b3b80d8

Reject unrepresentable pointer types in intrinsics. Fixes PR7316.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110541 91177308-0d34-0410-b5e6-96231b3b80d8

Use the "isCompare" machine instruction attribute instead of calling the
relatively expensive comparison analyzer on each instruction. Also rename the
comparison analyzer method to something more in line with what it actually does.

This pass is will eventually be folded into the Machine CSE pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110539 91177308-0d34-0410-b5e6-96231b3b80d8

Do more to modernize MergeFunctions. Refactor in response to Chris' code review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110538 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: set configure macro LLVM_PREFIX

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110536 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: configure header llvm-config.h

This is the cmake equivalent of r110532, which fixed bug #7809.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110535 91177308-0d34-0410-b5e6-96231b3b80d8

Most of bugpoint now only needs to know the pass names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110534 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110533 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR7809 by creating a header for just llvm variables that can be
included in exported interfaces.  Update a couple of exported interfaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110532 91177308-0d34-0410-b5e6-96231b3b80d8

Add back in r109901, which adds a Compare flag to the target instructions. It's
useful after all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110531 91177308-0d34-0410-b5e6-96231b3b80d8

Use %shlibext instead of .so

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110529 91177308-0d34-0410-b5e6-96231b3b80d8