Add MSP430 mem-mem insts support. Patch by Brian Lucas with some my refinements

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83811 91177308-0d34-0410-b5e6-96231b3b80d8

remove some harmful code that would turn an insertelement on an undef
into a shuffle even if it was used by another insertelement.  If the
visitation order of instcombine was wrong, this would turn a chain of
insertelements into a chain of shufflevectors, which was quite painful.

Since CollectShuffleElements handles these cases, the code can just
be nuked.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83810 91177308-0d34-0410-b5e6-96231b3b80d8

Add bunch of MSP430 'feature' tests. Patch by Brian Lucas with some my refinements

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83809 91177308-0d34-0410-b5e6-96231b3b80d8

reduce vec_shuffle2 and merge into vec_shuffle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83807 91177308-0d34-0410-b5e6-96231b3b80d8

filecheckize vec_shuffle.ll and merge shuffle.ll into it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83806 91177308-0d34-0410-b5e6-96231b3b80d8

filecheckize

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83805 91177308-0d34-0410-b5e6-96231b3b80d8

rename test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83804 91177308-0d34-0410-b5e6-96231b3b80d8

remove old testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83803 91177308-0d34-0410-b5e6-96231b3b80d8

merge test into shift.ll, this also eliminates awful grepping on -stats output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83802 91177308-0d34-0410-b5e6-96231b3b80d8

convert to filecheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83801 91177308-0d34-0410-b5e6-96231b3b80d8

teach instcombine to simplify xor's harder, catching the
new testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83799 91177308-0d34-0410-b5e6-96231b3b80d8

cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83797 91177308-0d34-0410-b5e6-96231b3b80d8

convert xor2 to filecheck, merge in a random regtest

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83796 91177308-0d34-0410-b5e6-96231b3b80d8

cleanup, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83795 91177308-0d34-0410-b5e6-96231b3b80d8

generalize a transformation even more: we don't care whether the
input the the mul is a zext from bool, just that it is all zeros
other than the low bit.  This fixes some phase ordering issues
that would cause us to miss some xforms in mul.ll when the worklist
is visited differently.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83794 91177308-0d34-0410-b5e6-96231b3b80d8

simplify a transformation by making it more general.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83792 91177308-0d34-0410-b5e6-96231b3b80d8

temporarily revert previous patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83791 91177308-0d34-0410-b5e6-96231b3b80d8

populate instcombine's initial worklist more carefully, causing
it to visit instructions from the start of the function to the
end of the function in the first path.  This greatly speeds up
some pathological cases (e.g. PR5150).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83790 91177308-0d34-0410-b5e6-96231b3b80d8

Remove CleanupDbgInfo, instcombine does this and its not worth duplicating it
here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83789 91177308-0d34-0410-b5e6-96231b3b80d8

More DragonEgg verbiage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83788 91177308-0d34-0410-b5e6-96231b3b80d8

Remove spurious brackets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83787 91177308-0d34-0410-b5e6-96231b3b80d8

LICM shouldn't sink/delete debug information. Fix this and add a testcase.
For now the metadata of sinked/hoisted instructions is still wrong, but that'll
be fixed when instructions will have debug metadata directly attached.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83786 91177308-0d34-0410-b5e6-96231b3b80d8

Implement 'm' memory operand properly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83785 91177308-0d34-0410-b5e6-96231b3b80d8

Implement proper asmprinting for the globals. This eliminates bogus "call" modifier and also adds support for offsets wrt globals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83784 91177308-0d34-0410-b5e6-96231b3b80d8

Implement asm printing for inline asm memory operands

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83783 91177308-0d34-0410-b5e6-96231b3b80d8

add PR5004 as a known problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83782 91177308-0d34-0410-b5e6-96231b3b80d8

duncan points out that llvm-gcc doesn't do the right thing with -fverbose-asm yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83781 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83780 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83779 91177308-0d34-0410-b5e6-96231b3b80d8

when folding duplicate conditions, delete the
now-probably-dead instruction tree feeding it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83778 91177308-0d34-0410-b5e6-96231b3b80d8

some notes from Anton

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83777 91177308-0d34-0410-b5e6-96231b3b80d8

catch some other serial commas that my earlier grep did not spot

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83772 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate some instances of serial comma. sabre, if you feel strong about this, feel free to revert this rev

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83771 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83770 91177308-0d34-0410-b5e6-96231b3b80d8

apply some tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83769 91177308-0d34-0410-b5e6-96231b3b80d8

fix some obvious typos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83768 91177308-0d34-0410-b5e6-96231b3b80d8

Add an outline of the DragonEgg gcc plugin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83765 91177308-0d34-0410-b5e6-96231b3b80d8

implement rdar://7293527, a trivial instcombine that llvm-gcc
gets but clang doesn't, because it is implemented in GCC's
fold routine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83761 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper for matching "1".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83760 91177308-0d34-0410-b5e6-96231b3b80d8

implement a transformation in jump threading that is currently
done by condprop, but do it in a much more general form.  The
basic idea is that we can do a limited form of tail duplication
in the case when we have a branch on a phi.  Moving the branch
up in to the predecessor block makes instruction selection
much easier and encourages chained jump threadings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83759 91177308-0d34-0410-b5e6-96231b3b80d8

another testcase jump threading shouldn't crash on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83758 91177308-0d34-0410-b5e6-96231b3b80d8

rename a file, remove a poorly reduced testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83757 91177308-0d34-0410-b5e6-96231b3b80d8

restructure some code, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83756 91177308-0d34-0410-b5e6-96231b3b80d8

factor some code better and move a function, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83755 91177308-0d34-0410-b5e6-96231b3b80d8

make jump threading on a phi with undef inputs happen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83754 91177308-0d34-0410-b5e6-96231b3b80d8

there is no need to run mem2reg after jump threading at LTO time now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83753 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bunch of bad formatting, delete the dead
ConstantInt::TheTrueVal/TheFalseVal members.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83752 91177308-0d34-0410-b5e6-96231b3b80d8

merge two tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83751 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some run lines, convert a test to filecheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83750 91177308-0d34-0410-b5e6-96231b3b80d8

Update release notes blurb on the static analyzer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83749 91177308-0d34-0410-b5e6-96231b3b80d8

rewrite LCSSA to use SSAUpdate, to only return true if it modifies
the IR, and to implement the FIXME'd optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83748 91177308-0d34-0410-b5e6-96231b3b80d8

clean up and simplify some code.  Don't use setvector when things will be
inserted only once, just use vector.  Don't compute ExitBlocks unless we
need it, change std::sort to array_pod_sort.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83747 91177308-0d34-0410-b5e6-96231b3b80d8

switch GVN to use SSAUpdater.  Besides removing a lot of complexity
from GVN, this also speeds it up, inserts fewer PHI nodes (see the
testcase) and allows it to remove more loads (due to fewer PHI nodes
standing in the way).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83746 91177308-0d34-0410-b5e6-96231b3b80d8

add a simple helper method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83745 91177308-0d34-0410-b5e6-96231b3b80d8

add ability for clients of SSAUpdater to find out about the
PHI nodes inserted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83744 91177308-0d34-0410-b5e6-96231b3b80d8

clarify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83743 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83742 91177308-0d34-0410-b5e6-96231b3b80d8

add the ability to get a rewritten value from the middle of a block,
not just at the end.  Add a big comment explaining when this could
be useful (which never happens for jump threading).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83741 91177308-0d34-0410-b5e6-96231b3b80d8

rename GetValueInBlock -> GetValueAtEndOfBlock to better reflect
what it does.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83740 91177308-0d34-0410-b5e6-96231b3b80d8

It seems that OR operation does not affect status reg at all.
Remove impdef of SRW. This fixes PR4779

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83739 91177308-0d34-0410-b5e6-96231b3b80d8

fix broken link

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83738 91177308-0d34-0410-b5e6-96231b3b80d8

final changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83737 91177308-0d34-0410-b5e6-96231b3b80d8

down to 'major changes'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83736 91177308-0d34-0410-b5e6-96231b3b80d8

more updates

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83735 91177308-0d34-0410-b5e6-96231b3b80d8

add a link to the GSG for info on how to check out svn trunk

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83734 91177308-0d34-0410-b5e6-96231b3b80d8

x86 uses 5 operands for most memory refs now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83733 91177308-0d34-0410-b5e6-96231b3b80d8

More spelling and grammar tweaks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83728 91177308-0d34-0410-b5e6-96231b3b80d8

More spelling fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83724 91177308-0d34-0410-b5e6-96231b3b80d8

Spelling fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83722 91177308-0d34-0410-b5e6-96231b3b80d8

more tweaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83721 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an inappropriate line in the description of the
clang static analyser.  Decrease duplication in the text.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83720 91177308-0d34-0410-b5e6-96231b3b80d8

continue decoding chris scribble.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83719 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead passes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83717 91177308-0d34-0410-b5e6-96231b3b80d8

checkpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83716 91177308-0d34-0410-b5e6-96231b3b80d8

fix broken anchors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83715 91177308-0d34-0410-b5e6-96231b3b80d8

use a typedef instead of spelling out an insane type.  Yay for auto someday.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83707 91177308-0d34-0410-b5e6-96231b3b80d8

Change jump threading to use the new SSAUpdater class instead of
DemoteRegToStack.  This makes it more efficient (because it isn't
creating a ton of load/stores that are eventually removed by a later
mem2reg), and more slightly more effective (because those load/stores
don't get in the way of threading).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83706 91177308-0d34-0410-b5e6-96231b3b80d8

Implement an efficient and fully general SSA update mechanism that
works on unstructured CFGs.  This implements PR217, our oldest open PR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83705 91177308-0d34-0410-b5e6-96231b3b80d8

add some WeakVH::operator='s.  Without these, assigning
a Value* to a WeakVH was constructing a temporary WeakVH
(due to the implicit assignment operator).  This avoids
that cost.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83704 91177308-0d34-0410-b5e6-96231b3b80d8

change some static_cast into cast, pointed out by Gabor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83703 91177308-0d34-0410-b5e6-96231b3b80d8

add a version of PHINode::getIncomingBlock that takes a raw
Use, to complement the version that takes a use_iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83702 91177308-0d34-0410-b5e6-96231b3b80d8

random tidying

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83701 91177308-0d34-0410-b5e6-96231b3b80d8

Create a new InstrEmitter class for translating SelectionDAG nodes
into MachineInstrs. This is mostly just moving the code from
ScheduleDAGSDNodesEmit.cpp into a new class. This decouples MachineInstr
emitting from scheduling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83699 91177308-0d34-0410-b5e6-96231b3b80d8

Make getMachineNode return a MachineSDNode* instead of a generic SDNode*
since it won't do any folding. This will help avoid some inconvenient
casting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83698 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a no-longer-necessary #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83697 91177308-0d34-0410-b5e6-96231b3b80d8

Replace X86's CanRematLoadWithDispOperand by calling the target-independent
MachineInstr::isInvariantLoad instead, which has the benefit of being
more complete.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83696 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a missing initialization of PostRAScheduler's AA member.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83695 91177308-0d34-0410-b5e6-96231b3b80d8

The ScheduleDAG framework now requires an AliasAnalysis argument, though
it isn't needed in the ScheduleDAGSDNodes schedulers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83691 91177308-0d34-0410-b5e6-96231b3b80d8

Update this test; the code is the same but it gets counted as one
fewer remat.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83690 91177308-0d34-0410-b5e6-96231b3b80d8

Mark the LDR instruction with isReMaterializable, as it is rematerializable
when loading from an invariant memory location.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83688 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out LiveIntervalAnalysis' code to determine whether an instruction
is trivially rematerializable and integrate it into
TargetInstrInfo::isTriviallyReMaterializable. This way, all places that
need to know whether an instruction is rematerializable will get the
same answer.

This enables the useful parts of the aggressive-remat option by
default -- using AliasAnalysis to determine whether a memory location
is invariant, and removes the questionable parts -- rematting operations
with virtual register inputs that may not be live everywhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83687 91177308-0d34-0410-b5e6-96231b3b80d8

Extract scope information from the variable itself, instead of relying on alloca or llvm.dbg.declare location.

While recording beginning of a function, use scope info from the first location entry instead of just relying on first location entry itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83684 91177308-0d34-0410-b5e6-96231b3b80d8

ExecutionEngine::clearGlobalMappingsFromModule failed to remove reverse
mappings, which could cause errors and assert-failures.  This patch fixes that,
adds a test, and refactors the global-mapping-removal code into a single place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83678 91177308-0d34-0410-b5e6-96231b3b80d8

Add a const qualifier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83677 91177308-0d34-0410-b5e6-96231b3b80d8

Use names instead of numbers for some of the magic
constants used in inlining heuristics (especially
those used in more than one file).  No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83675 91177308-0d34-0410-b5e6-96231b3b80d8

Added another bit of the ARM target assembler to llvm-mc to parse register
lists.  Changed ARMAsmParser::MatchRegisterName to return -1 instead of 0 on
errors so 0-15 values could be returned as register numbers.  Also added the
rest of the arm register names to the currently hacked up version to allow more
testing.  Some changes to ARMAsmParser::ParseOperand to give different errors
for things not yet supported and some additions to the hacked
ARMAsmParser::MatchInstruction to allow more testing for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83673 91177308-0d34-0410-b5e6-96231b3b80d8

isTriviallyReMaterializable checks the
TargetInstrDesc::isRematerializable flag, so it isn't necessary to do
this check in its callers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83671 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the x86 test-shrink optimization so that it doesn't shrink comparisons
when one of the bits being tested would end up being the sign bit in the
narrower type, and a signed comparison is being performed, since this would
change the result of the signed comparison. This fixes PR5132.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83670 91177308-0d34-0410-b5e6-96231b3b80d8

Merge a bunch of NEON tests into larger files so they run faster.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83667 91177308-0d34-0410-b5e6-96231b3b80d8

Add basic infrastructure and x86 support for preserving MachineMemOperand
information when unfolding memory references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83656 91177308-0d34-0410-b5e6-96231b3b80d8