a critical edge.

llvm-svn: 174936

live ranges should always be extended, and the only successor that should be
considered for extension of other ranges is the target of the split edge.

llvm-svn: 174935

Sorry for the lack of a test case.  I tried writing one for i386 as i know selects are illegal on this target, but they are actually considered legal by isel and expanded later.

I can't see any targets to trigger this, but checking for the legality of a node before forming it is general goodness.

llvm-svn: 174934

A reverse shuffle is lowered to a vrev and possibly a vext instruction (quad
word).

radar://13171406

llvm-svn: 174933

Check for reverse shuffles in the CostModel analysis pass and query
TargetTransform info accordingly. This allows us we can write test cases for
reverse shuffles.

radar://13171406

llvm-svn: 174932

lexical storage but not visible storage' case in C++. It's unclear whether we
even need the special-case handling for C++, since it seems to be working
around our not serializing a lookup table for the TU in C. But in any case,
the assertion is incorrect.

llvm-svn: 174931

expression.

llvm-svn: 174930

Lower reverse shuffles to a vrev64 and a vext instruction instead of the default
legalization of storing and loading to the stack. This is important because we
generate reverse shuffles in the loop vectorizer when we reverse store to an
array.

  uint8_t Arr[N];
  for (i = 0; i < N; ++i)
    Arr[N - i - 1] = ...

radar://13171760

llvm-svn: 174929

Patch by Joey Gouly!

llvm-svn: 174928

llvm-svn: 174927

Part of rdar://12470373

llvm-svn: 174926

llvm-svn: 174925

llvm-svn: 174924

llvm-svn: 174923

of immediately afterwards.

llvm-svn: 174922

llvm-svn: 174921

Fixed decode of existing 3dNow prefetchw instruction
Intel is scheduled to add a compatible prefetchw (same encoding) to future CPUs

llvm-svn: 174920

move-constructors and move-assignment operators, use memcpy to copy adjacent
POD members.

Previously, classes with one or more Non-POD members would fall back on
element-wise copies for all members, including POD members. This often
generated a lot of IR. Without padding metadata, it wasn't often possible
for the LLVM optimizers to turn the element-wise copies into a memcpy.

This code hasn't yet received any serious tuning. I didn't see any serious
regressions on a self-hosted clang build, or any of the nightly tests, but
I think it's important to get this out in the wild to get more testing.
Insights, feedback and comments welcome.

Many thanks to David Blaikie, Richard Smith, and especially John McCall for
their help and feedback on this work.

llvm-svn: 174919

llvm-svn: 174918

llvm-svn: 174916

When building the pairable-instruction dependency map, don't search
past the last pairable instruction. For large blocks that have been
divided into multiple instruction groups, searching past the last
instruction in each group is very wasteful. This gives a 32% speedup
on the csa.ll test case from PR15222 (when using 50 instructions
in each group).

No functionality change intended.

llvm-svn: 174915

This map is queried only for instructions in pairs of pairable
instructions; so make sure that only pairs of pairable
instructions are added to the map. This gives a 3.5% speedup
on the csa.ll test case from PR15222.

No functionality change intended.

llvm-svn: 174914

llvm-svn: 174913

MipsCodeEmitter.cpp.

JALR and NOP are expanded by function emitPseudoExpansionLowering, which is not
called when the old JIT is used.

This fixes the following tests which have been failing on
llvm-mips-linux builder:

LLVM :: ExecutionEngine__2003-01-04-LoopTest.ll
LLVM :: ExecutionEngine__2003-05-06-LivenessClobber.ll
LLVM :: ExecutionEngine__2003-06-04-bzip2-bug.ll
LLVM :: ExecutionEngine__2005-12-02-TailCallBug.ll
LLVM :: ExecutionEngine__2003-10-18-PHINode-ConstantExpr-CondCode-Failure.ll
LLVM :: ExecutionEngine__hello2.ll
LLVM :: ExecutionEngine__stubs.ll
LLVM :: ExecutionEngine__test-branch.ll
LLVM :: ExecutionEngine__test-call.ll
LLVM :: ExecutionEngine__test-common-symbols.ll
LLVM :: ExecutionEngine__test-loadstore.ll
LLVM :: ExecutionEngine__test-loop.ll

llvm-svn: 174912

llvm-svn: 174911

llvm-svn: 174910

llvm-svn: 174909

llvm-svn: 174907

declarations if we didn't have a lookup map when the external decls were added.

llvm-svn: 174906

llvm-svn: 174905

- variety of compare instructions,
- loops with no preheader,
- arbitrary lower and upper bounds.

llvm-svn: 174904

llvm-svn: 174903

llvm-svn: 174902

llvm-svn: 174901

*added file for test cases for i386 intel syntax

llvm-svn: 174900

of @throw statement by finding location of the ';'
correctly. // rdar://13186010

llvm-svn: 174898

llvm-svn: 174897

[NVPTX] Remove NoCapture from address space conversion intrinsics. NoCapture is not valid in this case, and was causing incorrect optimizations.

llvm-svn: 174896

[Modules] Cope better with top-level declarations loaded after being declared in the current translation unit <rdar://problem/13189985>.

These two related tweaks to keep the information associated with a
given identifier correct when the identifier has been given some
top-level information (say, a top-level declaration) and more
information is then loaded from a module. The first ensures that an
identifier that was "interesting" before being loaded from an AST is
considered to be different from its on-disk counterpart. Otherwise, we
lose such changes when writing the current translation unit as a
module.

Second, teach the code that injects AST-loaded names into the
identifier chain for name lookup to keep the most recent declaration,
so that we don't end up confusing our declaration chains by having a
different declaration in there.

llvm-svn: 174895

We need to remove one dimension. Any is correct as long as it exists. We have
choosen for whatever reason the dimension #dims - 2. This is incorrect if
there is just one dimension. For CLooG this case did never happen. For isl
however, the case can happen and causes undefined behavior including crashes.
We choose now always the last dimension #dims - 1. We could have choosen
dimension '0' but the last dimension is what we remove conceptionally in the
algorithm, so it seems better to actually program it that way.

While at it remove another piece of undefined behavior.

llvm-svn: 174894