SLPVectorizer: Make it a function pass and add code for hoisting the vector-gather sequence out of loops.

llvm-svn: 179562

Instead of emitting config values in a predefined order, the code
emitter will now emit a 32-bit register index followed by the 32-bit
config value.

llvm-svn: 179546

llvm-svn: 179545

llvm-svn: 179544

llvm-svn: 179542

I will remove the isBigEndianHost function once I update clang.

The ifdef logic is designed to
* not use configure/cmake to avoid breaking -arch i686 -arch ppc.
* default to little endian
* be as small as possible

It looks like sys/endian.h is the preferred header on most modern BSD systems,
but it is better to change this in a followup patch as machine/endian.h is
available on FreeBSD, OpenBSD, NetBSD and OS X.

llvm-svn: 179527

This is a rework of the broken parts in r179373 which were subsequently reverted in r179374 due to incompatibility with C++98 compilers.  This version should be ok under C++98.

llvm-svn: 179520

Add an option -vectorize-slp-aggressive for running the BB vectorizer. Make -fslp-vectorize run the slp-vectorizer.

llvm-svn: 179508

llvm-svn: 179505

llvm-svn: 179504

Now that the CR spilling issues have been resolved, we can remove the
unmodeled-side-effect attributes from the comparison instructions (and also
mark them as isCompare). By allowing these, by default, to have unmodeled side
effects, we were hiding problems with CR spilling; but everything seems much
happier now.

llvm-svn: 179502

This fixes an ABI bug for non-Darwin PPC64. For the callee-saved condition
registers, the spill location is specified relative to the stack pointer (SP +
8). However, this is not relative to the SP after the new stack frame is
established, but instead relative to the caller's stack pointer (it is stored
into the linkage area of the parent's stack frame).

So, like with the link register, we don't directly spill the CRs with other
callee-saved registers, but just mark them to be spilled during prologue
generation.

In practice, this reverts r179457 for PPC64 (but leaves it in place for PPC32).

llvm-svn: 179500

llvm-svn: 179494

One performs: (X == 13 | X == 14) -> X-13 <u 2
The other: (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1

The problem is that there are certain values of C1 and C2 that
trigger both transforms but the first one blocks out the second,
this generates suboptimal code.

Reordering the transforms should be better in every case and
allows us to do interesting stuff like turn:
  %shr = lshr i32 %X, 4
  %and = and i32 %shr, 15
  %add = add i32 %and, -14
  %tobool = icmp ne i32 %add, 0

into:
  %and = and i32 %X, 240
  %tobool = icmp ne i32 %and, 224

llvm-svn: 179493

llvm-svn: 179483

llvm-svn: 179479

llvm-svn: 179478

Test case by llvm-stress.

llvm-svn: 179477

SLPVectorizer: Add support for trees that don't start at binary operators, and add the cost of extracting values from the roots of the tree.

llvm-svn: 179475

For when 16 TB just isn't enough.

llvm-svn: 179474

This is the default model for non-PIC 64-bit code. It supports
text+data+bss linked anywhere in the low 16 TB of the address space.

llvm-svn: 179473

64-bit code models need multiple relocations that can't be inferred from
the opcode like they can in 32-bit code.

llvm-svn: 179472

Constant pool entries are accessed exactly the same way as global
variables.

llvm-svn: 179471

llvm-svn: 179470

This fixes the pic32 code model for SPARC v9.

llvm-svn: 179469

SDNodes and MachineOperands get target flags representing the %hi() and
%lo() assembly annotations that eventually become relocations.

Also define flags to be used by the 64-bit code models.

llvm-svn: 179468

Leaving MFCR has having unmodeled side effects is not enough to prevent
unwanted instruction reordering post-RA. We could probably apply a stronger
barrier attribute, but there is a better way: Add all (not just the first) CR
to be spilled as live-in to the entry block, and add all CRs to the MFCR
instruction as implicitly killed.

Unfortunately, I don't have a small test case.

llvm-svn: 179465

Currently, only abs32 and pic32 are implemented. Add a test case for
abs32 with 64-bit code. 64-bit PIC code is currently broken.

llvm-svn: 179463

It doesn't seem like anybody is checking types this late in isel, so no
test case.

llvm-svn: 179462

There is a Constant with non-constant operands: blockaddress.

llvm-svn: 179460

This is basically the same fix in three different places. We use a set to avoid
walking the whole tree of a big ConstantExprs multiple times.

For example: (select cmp, (add big_expr 1), (add big_expr 2))
We don't want to visit big_expr twice here, it may consist of thousands of
nodes.

The testcase exercises this by creating an insanely large ConstantExprs out of
a loop. It's questionable if the optimizer should ever create those, but this
can be triggered with real C code. Fixes PR15714.

llvm-svn: 179458

For functions that need to spill CRs, and have dynamic stack allocations, the
value of the SP during the restore is not what it was during the save, and so
we need to use the FP in these cases (as for all of the other spills and
restores, but the CR restore has a special code path because its reserved slot,
like the link register, is specified directly relative to the adjusted SP).

llvm-svn: 179457

llvm-svn: 179452

MI-Sched cleanup. If an instruction has no valid sched class, do not attempt to check for a variant.

llvm-svn: 179451

The initial values were arbitrary. I want them to be more
conservative. This represents the number of latency cycles hidden by
OOO execution. In practice, I think it should be within a small factor
of the complex floating point operation latency so the scheduler can
make some attempt to hide latency even for smallish blocks.

These are by no means the best values, just a starting point for
tuning heuristics. Some benchmarks such as TSVC run faster with this
lower value for SandyBridge. I haven't run anything on Haswell, but
it's shouldn't be 2x SB.

llvm-svn: 179450

The register allocator expects minimal physreg live ranges. Schedule
physreg copies accordingly. This is slightly tricky when they occur in
the middle of the scheduling region. For now, this is handled by
rescheduling the copy when its associated instruction is
scheduled. Eventually we may instead bundle them, but only if we can
preserve the bundles as parallel copies during regalloc.

llvm-svn: 179449

I need to handle this for the test case in my following scheduler
commit.

Work is already under way to redesign the mechanism for node order
propagation because this case by case approach is unmaintainable.

llvm-svn: 179448

lowerINTRINSIC_WO_CHAIN into MipsSETargetLowering.

No functionality changes.

llvm-svn: 179444

We are now able to handle big endian macho files in llvm-readobject. Thanks to
David Fang for providing the object files.

llvm-svn: 179440

llvm-svn: 179434