Fix and encoding error in the psrad xmm, imm8 instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51020 91177308-0d34-0410-b5e6-96231b3b80d8

On x86, it's safe to treat i32 load anyext as a normal i32 load. Ditto for i8 anyext load to i16.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51019 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up the use of static and anonymous namespaces. This turned up
several things that were neither in an anonymous namespace nor static
but not intended to be global.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51017 91177308-0d34-0410-b5e6-96231b3b80d8

Initial documentation for first-class aggregates changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51013 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a malformed %gt;

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51011 91177308-0d34-0410-b5e6-96231b3b80d8

Teach Legalize how to scalarize VSETCC
Teach X86 a few more vsetcc patterns.  Custom lowering for unsupported ones is next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51009 91177308-0d34-0410-b5e6-96231b3b80d8

Xform bitconvert(build_pair(load a, load b)) to a single load if the load locations are at the right offset from each other.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51008 91177308-0d34-0410-b5e6-96231b3b80d8

New test for tail merging

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51007 91177308-0d34-0410-b5e6-96231b3b80d8

Be more aggressive about tail-merging small blocks
if those blocks consist entirely of common instructions;
merging will not add an extra branch in this case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51006 91177308-0d34-0410-b5e6-96231b3b80d8

Constify isSourceDefinedByImplicitDef function. Otherwise, just formatting
changes that don't change functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51004 91177308-0d34-0410-b5e6-96231b3b80d8

Constify the machine instruction passed into the
"is{Trivially,Really}ReMaterializable" methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51001 91177308-0d34-0410-b5e6-96231b3b80d8

Initial X86 codegen support for VSETCC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51000 91177308-0d34-0410-b5e6-96231b3b80d8

Further rework of tail merge algorithm.  Not quite
semantically identical, but little difference in
either results or execution speed; but it's much
easier to read, at least IMO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50999 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify some checks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50998 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a copy+paste bug; pseudo-instructions shouldn't have
encoding information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50997 91177308-0d34-0410-b5e6-96231b3b80d8

Pointer comparisons should use icmp, not vicmp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50996 91177308-0d34-0410-b5e6-96231b3b80d8

Go back to passing the analyses around as parameters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50995 91177308-0d34-0410-b5e6-96231b3b80d8

Pointer comparisons should be handled by icmp, not vicmp :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50994 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50993 91177308-0d34-0410-b5e6-96231b3b80d8

Hard code CmpInst back to i1 for now while I go track down what in the bitcode reader/writer is assuming i1
This was breaking a bunch of tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50992 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor isConsecutiveLoad from X86 to TargetLowering so DAG combiner can make use of it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50991 91177308-0d34-0410-b5e6-96231b3b80d8

Revert the previous commit. Go ahead and hoist rematerializable instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50990 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for vicmp/vfcmp codegen, more legalize support coming.
This is necessary to unbreak the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50988 91177308-0d34-0410-b5e6-96231b3b80d8

One real change - don't hoist something that's trivially rematerializable. It's
possible for it to produce worse code than before.

The rest of this patch is code cleanup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50987 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build breakage

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50986 91177308-0d34-0410-b5e6-96231b3b80d8

Add two new instructions to the llvm IR, vicmp and vfcmp.  see updated LangRef
for details.  CodeGen support coming in a follow up patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50985 91177308-0d34-0410-b5e6-96231b3b80d8

Make firstEightPowers const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50975 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50974 91177308-0d34-0410-b5e6-96231b3b80d8

Filter option names to escape symbols not allowed as C++ identifiers.

Makes it possible to use options with names like "Wa,".
Also fixes the -Wall option handling as a side-effect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50973 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to choose between different compilation graph definitions at compile-time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50972 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some error messages; Make LLVMC pass through the exit code of a failed tool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50971 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a compile error on compilers that still want a return value
in a non-void function that calls abort.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50969 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a missing break in the ISD::FLT_ROUNDS_ handling. Patch by giuma!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50967 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR2264.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50965 91177308-0d34-0410-b5e6-96231b3b80d8

Move the various analyses used by GVN into static variables so we don't have to keep passing them around or refetching them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50963 91177308-0d34-0410-b5e6-96231b3b80d8

prune #includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50962 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new SparsePropagation analysis utility, which allows you to do
SCCP like sparse lattice analysis with relative ease.  Just pick your
lattice function and implement the transfer function and you're good.
Just make sure you don't break monotonicity ;-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50961 91177308-0d34-0410-b5e6-96231b3b80d8

Make constructors target-specific. This fixes problems where the path would
include backslashes on Windows. This should fix llvm-ld problems on win32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50960 91177308-0d34-0410-b5e6-96231b3b80d8

Add note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50959 91177308-0d34-0410-b5e6-96231b3b80d8

Fix various DOUTs to not call the extremely expensive Value::getName()
method.  DOUT statements are disabled when assertions are off, but the
side effects of getName() are still evaluated.  Just call getNameSTart,
which is close enough and doesn't cause heap traffic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50958 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code by using SwitchInst::findCaseValue instead of reimplementing it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50957 91177308-0d34-0410-b5e6-96231b3b80d8

Remove warnings when using -Wshorten-64-to-32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50952 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR2303.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50951 91177308-0d34-0410-b5e6-96231b3b80d8

Fix linking of internal aliases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50950 91177308-0d34-0410-b5e6-96231b3b80d8

Prevent -W64-to-32-shortened warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50947 91177308-0d34-0410-b5e6-96231b3b80d8

Prevent warnings from the -Wshorten-64-to-32 flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50946 91177308-0d34-0410-b5e6-96231b3b80d8

remove commented-out code, it is subsumed by DECLARE_TRANSPARENT_OPERAND_ACCESSORS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50944 91177308-0d34-0410-b5e6-96231b3b80d8

merge of use-diet branch to trunk

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50943 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable loop deletion by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50941 91177308-0d34-0410-b5e6-96231b3b80d8

When transforming a vector_shuffle to a load, the base address must not be an undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50940 91177308-0d34-0410-b5e6-96231b3b80d8

Set to 2.4 and regenerate configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50935 91177308-0d34-0410-b5e6-96231b3b80d8

Add nounwind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50931 91177308-0d34-0410-b5e6-96231b3b80d8

For now, abort when an ISD::VAARG is encountered on x86-64, rather
than silently generate invalid code.

llvm-gcc does not currently use VAArgInst; it lowers va_arg in the
front-end.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50930 91177308-0d34-0410-b5e6-96231b3b80d8

Some clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50929 91177308-0d34-0410-b5e6-96231b3b80d8

If movl top bits are undef, let it be selected to movlps, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50928 91177308-0d34-0410-b5e6-96231b3b80d8

If all sources of a PHI node are defined by an implicit_def, just emit an implicit_def instead of a copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50927 91177308-0d34-0410-b5e6-96231b3b80d8

Cosmetic changes:
 - Comment fixes.
 - Moar whitespace.
 - Made ivars "private" by default.
No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50926 91177308-0d34-0410-b5e6-96231b3b80d8

Add a pattern to do move the low element of a v4f32 and zero extend the rest.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50922 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an evil vector bool.  Cosmetic refactoring,
no functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50921 91177308-0d34-0410-b5e6-96231b3b80d8

Handle a few more cases of folding load i64 into xmm and zero top bits.
Note, some of the code will be moved into target independent part of DAG combiner in a subsequent patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50918 91177308-0d34-0410-b5e6-96231b3b80d8

Make OpActionsCapacity multiple of 4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50917 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite tail merging algorithm to handle the
case where there are multiple blocks with a large
number of common tail instructions more efficiently
(compile time optimization).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50916 91177308-0d34-0410-b5e6-96231b3b80d8

put LibCallAliasAnalysis into anonymous namespace to avoid
clashes on ::X

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50912 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50911 91177308-0d34-0410-b5e6-96231b3b80d8

Rename Example.td to Graph.td.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50910 91177308-0d34-0410-b5e6-96231b3b80d8

use doxygen comments for makeBuffer()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50908 91177308-0d34-0410-b5e6-96231b3b80d8

don't sink invokes, even if they are readonly.  This fixes a
crash on kimwitu++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50901 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a type and formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50900 91177308-0d34-0410-b5e6-96231b3b80d8

ignore Output dirs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50899 91177308-0d34-0410-b5e6-96231b3b80d8

ignore AutoGenerated.inc, it is created by TableGen

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50898 91177308-0d34-0410-b5e6-96231b3b80d8

ignore Output dir

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50897 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply 50867: A small refactoring (extract method) + some comment fixes.
Fixed the build breakage, sorry for that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50895 91177308-0d34-0410-b5e6-96231b3b80d8

add support for pattern matching 'neg'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50883 91177308-0d34-0410-b5e6-96231b3b80d8

Implement PR2298.  This transforms:
   ~x < ~y --> y < x
   -x == -y --> x == y

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50882 91177308-0d34-0410-b5e6-96231b3b80d8

restore doxygen comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50881 91177308-0d34-0410-b5e6-96231b3b80d8

<rdar://problem/5917641> use getMemBufferCopy if supplied buffer is not already zero terminated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50880 91177308-0d34-0410-b5e6-96231b3b80d8

Check for validity of aliasee pointer before dereference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50878 91177308-0d34-0410-b5e6-96231b3b80d8

Use movq to move low half of XMM register and zero-extend the rest.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50874 91177308-0d34-0410-b5e6-96231b3b80d8

conservatively say that volatile stores read memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50872 91177308-0d34-0410-b5e6-96231b3b80d8

Revertin 50867 since it was breaking the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50871 91177308-0d34-0410-b5e6-96231b3b80d8

store can't read from memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50869 91177308-0d34-0410-b5e6-96231b3b80d8

Add -E and -S options

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50868 91177308-0d34-0410-b5e6-96231b3b80d8

A small refactoring (extract method) + some comment fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50867 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead return.  Thanks to Bill for the review!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50863 91177308-0d34-0410-b5e6-96231b3b80d8

Improve pass documentation and comments.

Patch by Matthijs Kooijman!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50861 91177308-0d34-0410-b5e6-96231b3b80d8

More than just loads can read from memory: readonly calls like strlen
also need to be checked for memory modifying instructions before we
can sink them.  THis fixes the second half of PR2297.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50860 91177308-0d34-0410-b5e6-96231b3b80d8

Make instcombine's DSE respect loads as well as stores.  It is not safe to
delete the first store in:

store x -> p
load p
store y -> p

This is for PR2297.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50859 91177308-0d34-0410-b5e6-96231b3b80d8

add a new Instruction::mayReadFromMemory predicate, make
Instruction::mayWriteToMemory stronger for invokes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50858 91177308-0d34-0410-b5e6-96231b3b80d8

Check linkage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50851 91177308-0d34-0410-b5e6-96231b3b80d8

Get exception handling working again on 64 bit
Darwin.  This is a hack of course, but it does
at least look at the right thing: gotpcrel means
that this is already an offset, so an explicit
offset is not needed (and wrong).  I think this
is good enough for the moment: Anton is working
on something better.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50850 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50841 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for constant folding the 'offsetof' pattern even if the
base is not zero.  This fixes test/C++Frontend/2008-05-07-CrazyOffsetOf.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50840 91177308-0d34-0410-b5e6-96231b3b80d8

Handle vector move / load which zero the destination register top bits (i.e. movd, movq, movss (addr), movsd (addr)) with X86 specific dag combine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50838 91177308-0d34-0410-b5e6-96231b3b80d8

Add nounwind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50837 91177308-0d34-0410-b5e6-96231b3b80d8

Turn StripPointerCast() into a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50836 91177308-0d34-0410-b5e6-96231b3b80d8

Use Doxygen-style comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50833 91177308-0d34-0410-b5e6-96231b3b80d8

Make the interface of CStrInCStrNoCase be the same as strcasestr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50828 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new LibCallAliasAnalysis pass, which is parameterized
by an instance of LibCallInfo to provide mod/ref info of
standard library functions.  This is powerful enough to
say that 'sqrt' is readonly except that it modifies errno,
or that "printf doesn't store to memory unless the %n
constraint is present" etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50827 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some serious logical errors in CStrInCStrNoCase pointed out by Bill.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50826 91177308-0d34-0410-b5e6-96231b3b80d8

Output correct exception handling and frame info
on x86-64 linux.  This causes no regressions on
32 bit linux and 32 bit ppc.  More tests pass
on 64 bit ppc with no regressions.  I didn't
turn on eh on 64 bit linux because the intrinsics
needed to compile the eh runtime aren't done
yet.  But if you turn it on and link with the
mainline runtime then eh seems to work fine
on x86-64 linux with this patch.  Thanks to
Dale for testing.  The main point of the patch
is that if you output that some object is
encoded using 4 bytes you had better not output
8 bytes for it: the patch makes everything
consistent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50825 91177308-0d34-0410-b5e6-96231b3b80d8