Address mode immediate offset has already been divided by 4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59117 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a VFP binary arithmetic instruction encoding bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59116 91177308-0d34-0410-b5e6-96231b3b80d8

Fix address mode 3 immediate offset mode encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59109 91177308-0d34-0410-b5e6-96231b3b80d8

Consolidate formats; fix FCMPED etc. encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59107 91177308-0d34-0410-b5e6-96231b3b80d8

Add ptr annotation intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59106 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VFP conversion instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59104 91177308-0d34-0410-b5e6-96231b3b80d8

Fix encoding of single-precision VFP registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59102 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the testb optimization so x86 also bootstraps.
Reenable test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59101 91177308-0d34-0410-b5e6-96231b3b80d8

VFP fld / fst immediate field is multiplied by 4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59100 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Remove removed source file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59098 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bit of lazy evaluation to PopulateCompilationGraph().

Only the tools that are mentioned in the compilation graph definition
are now inserted by PopulateCompilationGraph(). This should cut down
plugin loading time a little.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59097 91177308-0d34-0410-b5e6-96231b3b80d8

Some cosmetic changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59096 91177308-0d34-0410-b5e6-96231b3b80d8

Remove

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59093 91177308-0d34-0410-b5e6-96231b3b80d8

Undo previous check-in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59092 91177308-0d34-0410-b5e6-96231b3b80d8

This shouldn't be necessary

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59090 91177308-0d34-0410-b5e6-96231b3b80d8

Fix FMDRR encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59088 91177308-0d34-0410-b5e6-96231b3b80d8

Handle floating point constpool_entry's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59087 91177308-0d34-0410-b5e6-96231b3b80d8

Don't walk into predecessors in which the vreg is not live when doing shrinkwrapping.
This lets several failing tests get farther along, but doesn't completely fix any of them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59086 91177308-0d34-0410-b5e6-96231b3b80d8

Encode VFP load / store instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59084 91177308-0d34-0410-b5e6-96231b3b80d8

In ScheduleDAGRRList::CopyAndMoveSuccessors, create the SUnit for the load
before creating the SUnit for the operation that it was unfolded from. This
allows each SUnit to have all of its predecessor SUnits available at the time
it is created. I don't know yet if this will be absolutely required, but it
is a little tidier to do it this way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59083 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid relying on the SelectionDAG for initializing the MachineFunction and
TargetLoweringInfo variables for the scheduler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59082 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Updated library dependency info used for MSVC++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59079 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Updated list of source files for lib/Transforms/Utils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59077 91177308-0d34-0410-b5e6-96231b3b80d8

Encode VFP conversion instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59074 91177308-0d34-0410-b5e6-96231b3b80d8

last validation errors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59072 91177308-0d34-0410-b5e6-96231b3b80d8

Remove incomplete lto-bugpoint tool. If there is an interest then we can resurrect this tool later on and finish implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59071 91177308-0d34-0410-b5e6-96231b3b80d8

fix some validation errors <it> -> <i>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59070 91177308-0d34-0410-b5e6-96231b3b80d8

fix some validation errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59069 91177308-0d34-0410-b5e6-96231b3b80d8

Add utility pass to remove dbg info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59068 91177308-0d34-0410-b5e6-96231b3b80d8

remove an extraneous tag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59067 91177308-0d34-0410-b5e6-96231b3b80d8

Make this document *substantially* better and cover a lot more territory.
Document written by Mason Woo (http://www.woo.com)!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59066 91177308-0d34-0410-b5e6-96231b3b80d8

Silence unused variable warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59064 91177308-0d34-0410-b5e6-96231b3b80d8

Use actual function name in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59063 91177308-0d34-0410-b5e6-96231b3b80d8

Change the scheduler accessor methods to accept an explicit TargetMachine
argument instead of taking the SelectionDAG's TargetMachine. This is
needed for some upcoming scheduler changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59055 91177308-0d34-0410-b5e6-96231b3b80d8

Add a README entry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59052 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: forces rebuild of llvm-config's library dependency info when a
library is added or changed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59051 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: corrected split of Alpha and Sparc AsmPrinters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59050 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Removed unnecessary library path setting that was breaking the
build on OS X.

Fix by Jjgod Jiang!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59048 91177308-0d34-0410-b5e6-96231b3b80d8

Separate sparc asmprinter. This should unbreak the native build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59047 91177308-0d34-0410-b5e6-96231b3b80d8

Separate alpha asmprinter. This should unbreak native build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59046 91177308-0d34-0410-b5e6-96231b3b80d8

The 32-bit displacement field in an x86 address is signed. Arrange for it
to be sign-extended when it is promoted to 64 bits for intermediate
offset calculations. The offset calculations are done as uint64_t so that
overflow conditions are well defined.

This fixes a problem which is currently hidden by the x86 AsmPrinter but
which was exposed by r58917 (which is temporarily reverted).  See PR3027
for details.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59044 91177308-0d34-0410-b5e6-96231b3b80d8

Add svn:ignore properties for srcdir = objdir builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59043 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMakeLists.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59039 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR3040:

The CC was changed, but wasn't checked to see if it was legal if the DAG
combiner was being run after legalization. Threw in a couple of checks just to
make sure that it's okay. As far as the PR is concerned, no back-end target
actually exhibited this problem, so there isn't an associated testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59035 91177308-0d34-0410-b5e6-96231b3b80d8

fix another libgcc blocker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59026 91177308-0d34-0410-b5e6-96231b3b80d8

Cleaned up and fix bugs in convert_rndsat node

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59025 91177308-0d34-0410-b5e6-96231b3b80d8

Un-XFAIL tests now that they're fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59023 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the buildbot and back out (inadvertant) casting edits in CellSPU
backend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59018 91177308-0d34-0410-b5e6-96231b3b80d8

Encode VFP arithmetic instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59016 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup debug info. assocated with deleted instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59012 91177308-0d34-0410-b5e6-96231b3b80d8

Add utility routines to remove dead debug info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59011 91177308-0d34-0410-b5e6-96231b3b80d8

r59009 broke these tests. XFAIL for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59010 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU: Fix bug 3606, as well as some ongoing work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59009 91177308-0d34-0410-b5e6-96231b3b80d8

Small simplification. Use the iterator already present as the insertion point.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59008 91177308-0d34-0410-b5e6-96231b3b80d8

Comments and indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59007 91177308-0d34-0410-b5e6-96231b3b80d8

Correct PIC function stub codegen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59006 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Fixed cross-compiling.

`clean' makefile target removes add-hoc directory created for building
the native tablegen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59005 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59004 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily revert r58979 and related patch. It's causing a failure in X86 bootstrap:

Comparing stages 2 and 3
warning: ./cc1-checksum.o differs
warning: ./cc1obj-checksum.o differs
warning: ./cc1objplus-checksum.o differs
warning: ./cc1plus-checksum.o differs
Bootstrap comparison failure!
./alias.o differs
./alloc-pool.o differs
./attribs.o differs
./bb-reorder.o differs
./bitmap.o differs
./build/errors.o differs
./build/genattrtab.o differs
./build/genautomata.o differs
./build/genemit.o differs
./build/genextract.o differs
...

-bw

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59003 91177308-0d34-0410-b5e6-96231b3b80d8

- Make sure that we don't over-increment the iterator when going through the
  basic blocks.
- Minor code clean-up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59002 91177308-0d34-0410-b5e6-96231b3b80d8

Added CONVERT_RNDSAT (conversion with rounding and saturation) SDNode to
support targets that support these conversions. Users should avoid using
this node as the current targets don't generating code for it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@59001 91177308-0d34-0410-b5e6-96231b3b80d8

Update VC++ project file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58997 91177308-0d34-0410-b5e6-96231b3b80d8

If the sign of exit condition and split condition does not match
then do not split loop index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58995 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2667: add soft float support for sint_to_fp/uint_to_fp
where the argument is an apint, or smaller than the minimum
size for which there is a libcall (i32).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58994 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58993 91177308-0d34-0410-b5e6-96231b3b80d8

Small cleanups.  No functionality change intended!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58992 91177308-0d34-0410-b5e6-96231b3b80d8

When promoting the result of fp_to_uint/fp_to_sint,
inform the optimizers that the result must be zero/
sign extended from the smaller type.  For example,
if a fp to unsigned i16 is promoted to fp to i32,
then we are allowed to assume that the extra 16 bits
are zero (because the result of fp to i16 is undefined
if the result does not fit in an i16).  This is
quite aggressive, but should help the optimizers
produce better code.  This requires correcting a
test which thought that fp_to_uint is some kind
of truncation, which it is not: in the testcase
(which does fp to i1), either the fp value converts
to 0 or 1 or the result is undefined, which is
quite different to truncation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58991 91177308-0d34-0410-b5e6-96231b3b80d8

Work around PR1000.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58984 91177308-0d34-0410-b5e6-96231b3b80d8

Proper way of doing llvm canadian-cross compilation.
Patch by Jim Grosbach!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58981 91177308-0d34-0410-b5e6-96231b3b80d8

Reenable test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58980 91177308-0d34-0410-b5e6-96231b3b80d8

Really fix testb optimization on big-endian.
Fixes ppc32 bootstrap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58979 91177308-0d34-0410-b5e6-96231b3b80d8

Third time's a charm.

The previous patches didn't match correctly. Also, we need to make sure that
the conditional is the same before doing the transformation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58978 91177308-0d34-0410-b5e6-96231b3b80d8

final edits

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58966 91177308-0d34-0410-b5e6-96231b3b80d8

Added support for the following definition of shufflevector
<result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58964 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate a couple more uses of utohexstr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58963 91177308-0d34-0410-b5e6-96231b3b80d8

Use utohex_buffer instead of utohexstr to avoid creating a temporary
string in the .ll and .s printers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58962 91177308-0d34-0410-b5e6-96231b3b80d8

split out the functionality of utohexstr into a new utohex_buffer
helper.  This allows us to convert numbers to hex without necessarily
needing to make a std::string to hold the result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58961 91177308-0d34-0410-b5e6-96231b3b80d8

Move getCastToEmpty out of DIDescriptor into DIFactory.  It is an
implementation detail of DIFactory anyway, and this allows it to avoid
recomputing the same type over and over.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58960 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an over aggressive assumption that struct decls will have a pointer to
an array that is empty.  Instead of requiring this array, allow a null pointer.
This shrinks all forward references of structs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58959 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug with default arguments that apple gcc doesn't notice that llvmbb does.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58958 91177308-0d34-0410-b5e6-96231b3b80d8

move some cases around to silence these sorts of warnings in
release-asserts build:

llvm/CodeGen/SelectionDAGNodes.h:1298: warning: control may reach end of non-void function 'unsigned int llvm::MVT::getSizeInBits() const' being inlined

what an unhelpful warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58957 91177308-0d34-0410-b5e6-96231b3b80d8

add new file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58955 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new set of helper classes for creating and reading debug
information.  This logically replaces the "Desc" classes in
MachineModuleInfo.  Nice features of these classes are that they:

1. Are much more efficient than MMI because they don't create a
   temporary parallel data structure for debug info that has to be
   'serialized' and 'deserialized' into/out of the module.
2. These provide a much cleaner abstraction for debug info than
   MMI, which will make it easier to change the implementation in
   the future (to be MDNode-based).
3. These are much easier to use than the MMI interfaces, requiring
   a lot less code in the front-ends.
4. These can be used to both create (for frontends) and read (for
   codegen) debug information.  DebugInfoBuilder can only be used
   to create the nodes.

So far, this is implemented just enough to support the debug info
generation needs of clang.  This can and should be extended to
support the full set of debug info constructs, and we should switch
llvm-gcc and llc over to using this in the near future.

This code also has a ton of FIXMEs in it, because the way we
currently represent debug info in LLVM IR is basically insane in a
variety of details.  This sort of issue should be fixed when we
eventually reimplement debug info on top of MDNodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58954 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Builds a native tblgen when cross-compiling and the user didn't
set LLVM_TABLEGEN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58953 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot these.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58952 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Implement magic word `all' for selecting all targets. Check
that specified targets are known.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58951 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Non-working code for auto-building the native tblgen util when
cross-compiling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58950 91177308-0d34-0410-b5e6-96231b3b80d8

Rename isGVNonLazyPtr to isIndirectSym to reflect how it will be used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58949 91177308-0d34-0410-b5e6-96231b3b80d8

Correction for the last patch. Should match the conditional in the first part
of the select match, not the select instruction itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58947 91177308-0d34-0410-b5e6-96231b3b80d8

The method of doing the matching with a 'select' instruction was wrong. The
original code was matching like this:

if (match(A, m_Not(m_Value(B))))

B was already matched as a 'select' instruction. However, this isn't matching
what we think it's matching. It would match B as a 'Value', so basically
anything would match to it. In this case, a Constant matched. B was replaced
with a constant representation. And then the wrong value would be used in the
SelectInst::Create statement, causing a crash.

After thinking on this for a moment, and after Nick L. told me how the pattern
matching stuff was supposed to work, the solution was to match NOT an m_Value,
but an m_Select.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58946 91177308-0d34-0410-b5e6-96231b3b80d8

suppress build products

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58943 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Corrected detection of `nm'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58941 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Support for cross-compiling. For now, requires a previously
built native tblgen which is passed to cmake in the variable
LLVM_TABLEGEN.

See

http://www.cmake.org/Wiki/CmakeMingw

for a quick example on how to cross-compile with CMake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58939 91177308-0d34-0410-b5e6-96231b3b80d8

fix validation problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58938 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL this while waiting for a fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58934 91177308-0d34-0410-b5e6-96231b3b80d8

fix leakage of ValueNumbering

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58933 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily revert 58825, which breaks PPC bootstrap.
xs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58930 91177308-0d34-0410-b5e6-96231b3b80d8

If the LHS of the FCMP is coming from a UIToFP instruction, then we don't want
to generate signed ICMP instructions to replace the FCMP. This would violate
the following:

define i1 @test1(i32 %val) {
  %1 = uitofp i32 %val to double
  %2 = fcmp ole double %1, 0.000000e+00
  ret i1 %2
}

would be transformed into:

define i1 @test1(i32 %val) {
  %1 = icmp slt i33 %val, 1
  ret i1 %1
}

which is obviously wrong. This patch modifes InstCombiner::FoldFCmp_IntToFP_Cst
to handle when the LHS comes from UIToFP.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58929 91177308-0d34-0410-b5e6-96231b3b80d8

Typo fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58928 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU: Update expected counts on expected patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58927 91177308-0d34-0410-b5e6-96231b3b80d8