Initial support for copy elimination by commuting its definition MI.
PR1877.
A3 = op A2 B0<kill>
...
B1 = A3      <- this copy
...
   = op A3   <- more uses

==>

B2 = op B0 A2<kill>
...
B1 = B2      <- now an identify copy
...
   = op B2   <- more uses

This speeds up FreeBench/neural by 29%, Olden/bh by 12%, oopack_v1p8 by 53%.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47046 91177308-0d34-0410-b5e6-96231b3b80d8

Make register scavenging happy by not using a reg (CR0) that isn't defined

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47045 91177308-0d34-0410-b5e6-96231b3b80d8

- Added removeValNo() to remove all live ranges of a particular value#.
- removeRange() can now update value# information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47044 91177308-0d34-0410-b5e6-96231b3b80d8

commuteInstr() can now commute non-ssa machine instrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47043 91177308-0d34-0410-b5e6-96231b3b80d8

Added debugging routine dumpUses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47042 91177308-0d34-0410-b5e6-96231b3b80d8

Convert SelectionDAG::ComputeMaskedBits to use APInt instead of uint64_t.
Add an overload that supports the uint64_t interface for use by clients
that haven't been updated yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47039 91177308-0d34-0410-b5e6-96231b3b80d8

__DATA not __DATA__ is the right segment name on darwin.
Spotted by Nick Kledzik.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47037 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47036 91177308-0d34-0410-b5e6-96231b3b80d8

SSE4.1 64b integer insert/extract pattern support
Move formats into the formats file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47035 91177308-0d34-0410-b5e6-96231b3b80d8

Change APInt::getBitsSet to accept a "half-open" range, where the
hiBit parameter marks the index one past the last desired set bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47032 91177308-0d34-0410-b5e6-96231b3b80d8

Note that these instructions are for x86-32 linux
(the only platform on which the Ada compiler even
builds).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47031 91177308-0d34-0410-b5e6-96231b3b80d8

Add more spacing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47028 91177308-0d34-0410-b5e6-96231b3b80d8

Add instructions for building Ada and Fortran.
Adjust mentions of gcc4 to be 4.0/4.2 agnostic.
This file should probably be renamed tor
 GCCFEBuildInstrs.html...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47027 91177308-0d34-0410-b5e6-96231b3b80d8

Re-apply the patch to improve the optimizations of memcpy's, with several
bugs fixed.  This now passes PPC bootstrap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47026 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize getCopyFromParts and getCopyToParts to
handle arbitrary precision integers and any number
of parts.  For example, on a 32 bit machine an i50
corresponds to two i32 parts.  getCopyToParts will
extend the i50 to an i64 then write half of the i64
to each part; getCopyFromParts will combine the two
i32 parts into an i64 then truncate the result to
i50.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47024 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r46916 PPCTargetAsmInfo.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47020 91177308-0d34-0410-b5e6-96231b3b80d8

Only using x86-64 rip relative addressing in non-staic mode?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47019 91177308-0d34-0410-b5e6-96231b3b80d8

Don't mask the isel bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47018 91177308-0d34-0410-b5e6-96231b3b80d8

This test assumes no SSE4.1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47017 91177308-0d34-0410-b5e6-96231b3b80d8

Add a doxygen comment for SrcValueSDNode, to make its purpose
clear and to clarify how it differs from MemOperandSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47015 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47012 91177308-0d34-0410-b5e6-96231b3b80d8

Now that ScalarEvolution::print writes to the correct stream, there is
no need to redirect stderr into stdout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47009 91177308-0d34-0410-b5e6-96231b3b80d8

Change negative grep into positive one in my yesterday's testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47008 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2002. Suppose n is the initial value for the induction
variable (with step 1) and m is its final value. Then, the correct trip
count is SMAX(m,n)-n. Previously, we used SMAX(0,m-n), but m-n may
overflow and can't in general be interpreted as signed.

Patch by Nick Lewycky.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47007 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for bug 1996: optimize out loads of undef.  This code basically just
checks for a malloc/alloca immediately followed by a load.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47006 91177308-0d34-0410-b5e6-96231b3b80d8

Update comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47002 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak various insert_vector_elt and extract_vector_elt tests in presence of SSE4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@47001 91177308-0d34-0410-b5e6-96231b3b80d8

Test case for annotate builtin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46999 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some test cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46998 91177308-0d34-0410-b5e6-96231b3b80d8

More violation fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46997 91177308-0d34-0410-b5e6-96231b3b80d8

Fix validation errors: di -> dt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46996 91177308-0d34-0410-b5e6-96231b3b80d8

Reformatting. Adding "doc_code" divisions for code examples. Updated some
of the examples to reflect the current .TD files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46995 91177308-0d34-0410-b5e6-96231b3b80d8

rip out llvm 2.2 stuff in preparation for llvm 2.3

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46993 91177308-0d34-0410-b5e6-96231b3b80d8

heavily  refactor this to have less duplication between BUILD_FOR_WEBSITE mode and not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46991 91177308-0d34-0410-b5e6-96231b3b80d8

remove unknown details

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46985 91177308-0d34-0410-b5e6-96231b3b80d8

Add 2.2 release tag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46981 91177308-0d34-0410-b5e6-96231b3b80d8

Stuff noticed while grepping code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46979 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the order of the arguments in the examples in the comments
for APInt::getBitsSet. And fix an off-by-one bug in "wrapping" mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46978 91177308-0d34-0410-b5e6-96231b3b80d8

The factories for ImutAVLTree/ImmutableSet/ImmutableMap now take an (optional)
BumpPtrAllocator argument to their constructors.  This BumpPtrAllocator
will be used to allocate trees.  If no BumpPtrAllocator is provided, one
is created (as before).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46975 91177308-0d34-0410-b5e6-96231b3b80d8

Added blurb for CellSPU progress

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46968 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize the handling of call and return arguments,
in preparation for apint support.  These changes are
intended to have no functional effect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46967 91177308-0d34-0410-b5e6-96231b3b80d8

From Chris' review: use isa instead of explicitly using classof.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46964 91177308-0d34-0410-b5e6-96231b3b80d8

From Chris' review: minor corrections in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46963 91177308-0d34-0410-b5e6-96231b3b80d8

From Chris' review: use cast instead of dyn_cast with an assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46962 91177308-0d34-0410-b5e6-96231b3b80d8

From Chris' review: fix 80 column violations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46961 91177308-0d34-0410-b5e6-96231b3b80d8

From Chris' review: change MemOperandSDNode's constructor to pass its
argument by reference, rather than by value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46960 91177308-0d34-0410-b5e6-96231b3b80d8

If the LHS of the comparison is a loop-invariant we also want to move it
to the RHS. This simple change allows to compute loop iteration count
for loops with condition similar to the one in the testcase (which seems
to be quite common).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46959 91177308-0d34-0410-b5e6-96231b3b80d8

Added "Profile" method to APFloat for use with FoldingSet.

Added member template "Add" to FoldingSetNodeID that allows "adding" arbitrary
objects to a profile via dispatch to FoldingSetTrait<T>::Profile().

Removed FoldingSetNodeID::AddAPFloat and FoldingSetNodeID::APInt, as their
functionality is now replaced using the above mentioned member template.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46957 91177308-0d34-0410-b5e6-96231b3b80d8

Add arbitrary integer support to getRegisterType and
getNumRegisters.  This is needed for calling functions
with apint parameters or return values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46956 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1798 - an error in the evaluation of SCEVAddRecExpr at an
arbitrary iteration.

The patch:
1) changes SCEVSDivExpr into SCEVUDivExpr,
2) replaces PartialFact() function with BinomialCoefficient(); the
computations (essentially, the division) in BinomialCoefficient() are
performed with the apprioprate bitwidth necessary to avoid overflow;
unsigned division is used instead of the signed one.

Computations in BinomialCoefficient() require support from the code
generator for APInts. Currently, we use a hack rounding up the
neccessary bitwidth to the nearest power of 2. The hack is easy to turn
off in future.

One remaining issue: we assume the divisor of the binomial coefficient
formula can be computed accurately using 16 bits. It means we can handle
AddRecs of length up to 9. In future, we should use APInts to evaluate
the divisor.

Thanks to Nicholas for cooperation!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46955 91177308-0d34-0410-b5e6-96231b3b80d8

Add a isBigEndian method to complement isLittleEndian.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46954 91177308-0d34-0410-b5e6-96231b3b80d8

Determine whether a spill kills the register it's spilling before insertion rather than trying to undo the kill marker afterwards.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46953 91177308-0d34-0410-b5e6-96231b3b80d8

tweaks from dale.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46952 91177308-0d34-0410-b5e6-96231b3b80d8

updates from Evan

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46951 91177308-0d34-0410-b5e6-96231b3b80d8

Enable SSE4 codegen and pattern matching.
Add some notes to the README.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46949 91177308-0d34-0410-b5e6-96231b3b80d8

additional missing feature

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46948 91177308-0d34-0410-b5e6-96231b3b80d8

corrected some spellings

maybe c380004 can be hyperlinked too?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46934 91177308-0d34-0410-b5e6-96231b3b80d8

Various updates from Sam Bishop:

"I have been working my way through the JIT and Kaleidoscope tutorials in my
(minuscule) spare time.  Thanks again for writing them!  I have attached a
patch containing some minor changes, ranging from spelling and grammar fixes
to adding a "Next: <next tutorial section>" hyperlink to the bottom of each
page.

Every page has been given the "next link" treatment, but otherwise I'm only
half way through the Kaleidoscope tutorial.  I will send a follow-on patch
if time permits."

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46933 91177308-0d34-0410-b5e6-96231b3b80d8

Fix scalarrepl to not 'miscompile' undefined code, part #2.
This fixes the store case, my previous patch just fixed the load
case.  rdar://5707076.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46932 91177308-0d34-0410-b5e6-96231b3b80d8

xmm0 variable blends

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46931 91177308-0d34-0410-b5e6-96231b3b80d8

Rename MRegisterInfo to TargetRegisterInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46930 91177308-0d34-0410-b5e6-96231b3b80d8

close an li tag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46929 91177308-0d34-0410-b5e6-96231b3b80d8

Add some Ada info and correct a few buglets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46928 91177308-0d34-0410-b5e6-96231b3b80d8

Add truncate and AssertZext result expansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46926 91177308-0d34-0410-b5e6-96231b3b80d8

remove the in progress marker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46925 91177308-0d34-0410-b5e6-96231b3b80d8

validation fixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46924 91177308-0d34-0410-b5e6-96231b3b80d8

final set of major updates

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46923 91177308-0d34-0410-b5e6-96231b3b80d8

Return "(c1 + c2)" instead of yet another ADD node (which made this a
no-op).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46922 91177308-0d34-0410-b5e6-96231b3b80d8

more prose, some organization improvements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46921 91177308-0d34-0410-b5e6-96231b3b80d8

random notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46920 91177308-0d34-0410-b5e6-96231b3b80d8

add anote

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46918 91177308-0d34-0410-b5e6-96231b3b80d8

Match GCC's behaviour for these sections.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46916 91177308-0d34-0410-b5e6-96231b3b80d8

memopv16i8 had wrong alignment requirement, would have broken pabsb
pabs{b,w,d} are not two address
fix extract-to-mem sse4 ops
add sse4 vector sign extend nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46915 91177308-0d34-0410-b5e6-96231b3b80d8

add some more notes, still not ready for review :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46914 91177308-0d34-0410-b5e6-96231b3b80d8

scoped_ptr.h was removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46913 91177308-0d34-0410-b5e6-96231b3b80d8

explain that NumElements in alloca and malloc defaults to one

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46912 91177308-0d34-0410-b5e6-96231b3b80d8

Alignment of struct containing vectors depends on
whether SSE is present, on Darwin anyway.  Make it
explicit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46909 91177308-0d34-0410-b5e6-96231b3b80d8

We should check that existing cast operation has the appropriate opcode before we reuse it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46908 91177308-0d34-0410-b5e6-96231b3b80d8

Add a check if the initial value of the induction variable is 0 (the method comment says it should be).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46907 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused hidden option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46903 91177308-0d34-0410-b5e6-96231b3b80d8

Skeleton of insert and extract matching, more to come

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46902 91177308-0d34-0410-b5e6-96231b3b80d8

Tablegen support for insert & extract element matching

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46901 91177308-0d34-0410-b5e6-96231b3b80d8

check that terminators only occur at the end of a block.  This catches the
common problem of putting two terminators in the same block.  I can't write
a testcase for this because the .ll parser rejects this before the verifier
can, but this can occur when generating IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46900 91177308-0d34-0410-b5e6-96231b3b80d8

Change ConstantSDNode to store an APInt instead of a uint64_t, and
begin adding some methods to use it this way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46899 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR 1995.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46898 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot these files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46896 91177308-0d34-0410-b5e6-96231b3b80d8

Also print alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46895 91177308-0d34-0410-b5e6-96231b3b80d8

Add an isSignedIntN, like isIntN but for signed integer values instead of
unsigned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46894 91177308-0d34-0410-b5e6-96231b3b80d8

It's not always safe to fold movsd into xorpd, etc. Check the alignment of the load address first to make sure it's 16 byte aligned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46893 91177308-0d34-0410-b5e6-96231b3b80d8

64-bit (MMX) vectors do not need restrictive alignment.
128-bit vectors need it only when SSE is on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46890 91177308-0d34-0410-b5e6-96231b3b80d8

Constified operator<< in APSInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46882 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid needlessly casting away const qualifiers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46877 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid needlessly casting away const qualifiers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46876 91177308-0d34-0410-b5e6-96231b3b80d8

Added missing entries in X86 load / store folding tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46866 91177308-0d34-0410-b5e6-96231b3b80d8

Begin setting features for target based on target triple string.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46863 91177308-0d34-0410-b5e6-96231b3b80d8

Remove remnant of load folding in local register allocator. Patch by Holger Schurig.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46861 91177308-0d34-0410-b5e6-96231b3b80d8

Follow Chris' suggestion; change the PseudoSourceValue accessors
to return pointers instead of references, since this is always what
is needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46857 91177308-0d34-0410-b5e6-96231b3b80d8

Add SourceValue information for outgoing argument stores on x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46854 91177308-0d34-0410-b5e6-96231b3b80d8

Don't abort if a MemOperand is missing a SourceValue; just print it
as <unknown>. And make some minor adjustments to the MemOperand
dump format.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46853 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a x86-64 codegen deficiency. Allow gv + offset when using rip addressing mode.
Before:
_main:
        subq    $8, %rsp
        leaq    _X(%rip), %rax
        movsd   8(%rax), %xmm1
        movss   _X(%rip), %xmm0
        call    _t
        xorl    %ecx, %ecx
        movl    %ecx, %eax
        addq    $8, %rsp
        ret
Now:
_main:
        subq    $8, %rsp
        movsd   _X+8(%rip), %xmm1
        movss   _X(%rip), %xmm0
        call    _t
        xorl    %ecx, %ecx
        movl    %ecx, %eax
        addq    $8, %rsp
        ret

Notice there is another idiotic codegen issue that needs to be fixed asap:
xorl    %ecx, %ecx
movl    %ecx, %eax

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46850 91177308-0d34-0410-b5e6-96231b3b80d8