Micro-optimize SCEVComplexityCompare.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109267 91177308-0d34-0410-b5e6-96231b3b80d8

Add a const qualifier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109266 91177308-0d34-0410-b5e6-96231b3b80d8

Use the proper type for shift counts. This fixes a bootstrap error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109265 91177308-0d34-0410-b5e6-96231b3b80d8

IF directory name is empty then try to extract one using absolute file name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109262 91177308-0d34-0410-b5e6-96231b3b80d8

Test case to insure template function declaration refers to correct filename.  Radar 8063111.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109258 91177308-0d34-0410-b5e6-96231b3b80d8

Add AVX version of CLMUL instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109248 91177308-0d34-0410-b5e6-96231b3b80d8

DAGCombine (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
are not demanded. This often allows the anyext to be folded away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109242 91177308-0d34-0410-b5e6-96231b3b80d8

Make SDNode::dump() print a newline at the end.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109234 91177308-0d34-0410-b5e6-96231b3b80d8

fix constness warnings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109224 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r109102 for now as it's causing JIT miscompilations.

I'll try to track down why a bit later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109223 91177308-0d34-0410-b5e6-96231b3b80d8

do not (implicitly) dereference iterator many times, cache it instead

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109222 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 109220.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109221 91177308-0d34-0410-b5e6-96231b3b80d8

Simplifying use_iterators by dereferencing
is not a good idea. The codebase does not depend
in this any more, and it may introduce hidden
runtime cost. If you get compile errors, please
dereference your iterator before passing to cast<>
(and friends).

Also: please consider caching the result of
operator* and reusing that instead of dereferencing
many times.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109220 91177308-0d34-0410-b5e6-96231b3b80d8

StringRef'icate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109217 91177308-0d34-0410-b5e6-96231b3b80d8

Formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109216 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of exceptions in llvmc.

llvmc can be now compiled with llvm-gcc on Windows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109215 91177308-0d34-0410-b5e6-96231b3b80d8

give StringMap a new ctor which allows you to initialize it
with an existing allocator.  The interesting use case of this
is that it allows "StringMap<whatever, BumpPtrAllocator&>" for
when you want to allocate out of a preexisting bump pointer
allocator owned by someone else.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109213 91177308-0d34-0410-b5e6-96231b3b80d8

modernize stringset interface

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109212 91177308-0d34-0410-b5e6-96231b3b80d8

Declare CLMUL as a subtarget feature

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109207 91177308-0d34-0410-b5e6-96231b3b80d8

Add x86 CLMUL (Carry-less multiplication) cpu feature

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109206 91177308-0d34-0410-b5e6-96231b3b80d8

80-col.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109205 91177308-0d34-0410-b5e6-96231b3b80d8

Add complete assembler support for FMA3 instructions, with descriptions and encodings taken from the AVX manual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109204 91177308-0d34-0410-b5e6-96231b3b80d8

The only supported calling convention for X86-64 uses
SSE, so we can't return floating point values if this
is disabled.  Detect this error for clang.

With SSE1 only, f64 is a problem; it can be done, but
neither llvm-gcc nor clang has ever generated correct
code for it.  Since nobody noticed this I think it's
OK to treat it as an error for now.

This also handles SSE-sized vectors of floating point.
8207686, 8204109.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109201 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some AVX instructions which didnt had HasAVX prefix. And also a problem with PINSRW, which was totally wrong because of a typo I introduced previously

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109198 91177308-0d34-0410-b5e6-96231b3b80d8

Add an explicit -sdk option to xcrun command.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109196 91177308-0d34-0410-b5e6-96231b3b80d8

Warnings patrol.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109174 91177308-0d34-0410-b5e6-96231b3b80d8

missed a use of SizeRequired.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109172 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the TargetInstrInfo::GetInstSizeInBytes hook.
ARM/PPC/MSP430-specific code (which are the only targets that
implement the hook) can directly reference their target-specific
instrinfo classes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109171 91177308-0d34-0410-b5e6-96231b3b80d8

remove the dwarf sizing stuff which is now dead, it was
"yet another" copy of the dwarf EH emission code that was
copied, pasted and slightly hacked up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109169 91177308-0d34-0410-b5e6-96231b3b80d8

Add remaining AVX instructions (most of them dealing with GR64 destinations. This complete the assembler support for the general AVX ISA. But we still miss instructions from FMA3 and CLMUL specific feature flags, which are now the next step

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109168 91177308-0d34-0410-b5e6-96231b3b80d8

remove the JIT "NeedsExactSize" feature and supporting logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109167 91177308-0d34-0410-b5e6-96231b3b80d8

switch a private implementation of GetFunctionSizeInBytes.
This is probably not the best way to implement "Force LR to
be spilled if the Thumb function size is > 2048." do this,
it should use the branch shortening infrastructure, but I'm
just preserving functionality here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109165 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109164 91177308-0d34-0410-b5e6-96231b3b80d8

Try to work around the relative install-sh path problem.

Patch from Eli.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109163 91177308-0d34-0410-b5e6-96231b3b80d8

X86MCInstLower now depends on AsmPrinter being around.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109154 91177308-0d34-0410-b5e6-96231b3b80d8

instead of migrating it to the MC instruction encoder, just
rip out the implementation of X86InstrInfo::GetInstSizeInBytes.
The code being ripped out just implemented a copy and hacked up
version of the (old) instruction encoder, and is buggy and
terrible in other ways.  Since "GetInstSizeInBytes" is really
only there to support the JIT's "NeedsExactSize" hook (which
noone is using), just rip out the code.  I will rip out the
NeedsExactSize hook next.

This resolves rdar://7617809 - switch X86InstrInfo::GetInstSizeInBytes to use X86MCCodeEmitter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109149 91177308-0d34-0410-b5e6-96231b3b80d8

Speculatively revert 109117

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109132 91177308-0d34-0410-b5e6-96231b3b80d8

ARMv4 JIT forgets to set the lr register when making a indirect function call. Fixes PR7608

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109125 91177308-0d34-0410-b5e6-96231b3b80d8

keep in 80 cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109122 91177308-0d34-0410-b5e6-96231b3b80d8

Map MDNode correctly.
A non function local MDNode can have an operand which is cloned by MapValue().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109117 91177308-0d34-0410-b5e6-96231b3b80d8

add dyn_cast_or_null tests, exclude invalid dyn_cast test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109111 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109110 91177308-0d34-0410-b5e6-96231b3b80d8

add dyn_cast tests and beef up others a bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109109 91177308-0d34-0410-b5e6-96231b3b80d8

fit in 80 columns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109108 91177308-0d34-0410-b5e6-96231b3b80d8

use cascading operator-> feature

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109104 91177308-0d34-0410-b5e6-96231b3b80d8

mass elimination of reliance on automatic iterator dereferencing

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109103 91177308-0d34-0410-b5e6-96231b3b80d8

Increase the max physreg size.  Patch by Pekka Jääskeläinen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109102 91177308-0d34-0410-b5e6-96231b3b80d8

simplify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109101 91177308-0d34-0410-b5e6-96231b3b80d8

do not access arguments via low-level interface, do not multiply dereference use_iterators

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109100 91177308-0d34-0410-b5e6-96231b3b80d8

pass dereferenced iterator to dyn_cast

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109099 91177308-0d34-0410-b5e6-96231b3b80d8

pass dereferenced iterator to dyn_cast

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109098 91177308-0d34-0410-b5e6-96231b3b80d8

use -> instead of (*).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109094 91177308-0d34-0410-b5e6-96231b3b80d8

cache dereferenced iterator

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109093 91177308-0d34-0410-b5e6-96231b3b80d8

undo 80 column trespassing I caused

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109092 91177308-0d34-0410-b5e6-96231b3b80d8

Mark an assert-only variable as used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109091 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the generated file name for CMake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109090 91177308-0d34-0410-b5e6-96231b3b80d8

Add new RegionInfo pass.

The RegionInfo pass detects single entry single exit regions in a function,
where a region is defined as any subgraph that is connected to the remaining
graph at only two spots.
Furthermore an hierarchical region tree is built.
Use it by calling "opt -regions analyze" or "opt -view-regions".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109089 91177308-0d34-0410-b5e6-96231b3b80d8

Attempt to fix linking issues with CMake. Please review other CMake users,
especially on other platforms. Is there a better way to fix this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109084 91177308-0d34-0410-b5e6-96231b3b80d8

Re-apply r109079 with fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109083 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r109079, which broke a lot of CodeGen tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109082 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMake files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109081 91177308-0d34-0410-b5e6-96231b3b80d8

Initial modifications to MCAssembler and TargetMachine for the MCJIT.

Patch by Olivier Meurant!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109080 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize RegLimit only when register pressure is being tracked.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109079 91177308-0d34-0410-b5e6-96231b3b80d8

Custom lower the memory barrier instructions and add support
for lowering without sse2.  Add a couple of new testcases.

Fixes a few libgomp tests and latent bugs.  Remove a few todos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109078 91177308-0d34-0410-b5e6-96231b3b80d8

Fix constant island pass's handling of tBR_JTr. The offset of the instruction does not have to be 4-byte aligned. Rather, it's the offset + 2 that must be aligned since the instruction expands into:
        mov     pc, r1
        .align  2
LJTI0_0_0:
        .long    LBB0_14

This fixes rdar://8213383. No test case since it's not possible to come up with a suitable small one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109076 91177308-0d34-0410-b5e6-96231b3b80d8

If 'other' was empty 'overlapsFrom(other, other.begin());' will segfault. This avoids that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109075 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Add some example tests for previous commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109071 91177308-0d34-0410-b5e6-96231b3b80d8

80-columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109070 91177308-0d34-0410-b5e6-96231b3b80d8

Make fast isel win64-aware w.r.t. call-clobbered regs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109069 91177308-0d34-0410-b5e6-96231b3b80d8

More register pressure aware scheduling work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109064 91177308-0d34-0410-b5e6-96231b3b80d8

Add more 256-bit forms for a bunch of regular AVX instructions
Add 64-bit (GR64) versions of some instructions (which are not
described in their SSE forms, but are described in AVX)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109063 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Add support for 'REQUIRES: feature-one, feature-two, ...' in the
integrated-test formats (sh and tcl style). The particular features which get
recognized are up to the test suite itself to define.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109062 91177308-0d34-0410-b5e6-96231b3b80d8

Make NamedMDNode not be a subclass of Value, and simplify the interface
for creating and populating NamedMDNodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109061 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes win64. It was broken by a previous patch where I missed the !isWin64
and then forced every register to be a vr128 on win64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109060 91177308-0d34-0410-b5e6-96231b3b80d8

Add INSTANTIATE_AG_PASS, which combines RegisterPass<> with RegisterAnalysisGroup<> for pass registration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109058 91177308-0d34-0410-b5e6-96231b3b80d8

For ARM/Darwin, add a dwarf entry indicating whether a function is arm or thumb
rdar://8202967

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109057 91177308-0d34-0410-b5e6-96231b3b80d8

add some rough support for making mcinst lowering work without an
asmprinter or mangler around.  This is option #B for killing off
X86InstrInfo::GetInstSizeInBytes.  Option #A (killing
"needsexactsize") was sent for consideration to llvmdev.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109056 91177308-0d34-0410-b5e6-96231b3b80d8

First stab at updating the documentation for INITIALIZE_PASS().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109055 91177308-0d34-0410-b5e6-96231b3b80d8

Baby steps towards ARM fast-isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109047 91177308-0d34-0410-b5e6-96231b3b80d8

Fix batch of converting RegisterPass<> to INTIALIZE_PASS().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109045 91177308-0d34-0410-b5e6-96231b3b80d8

Hack around extracts that aren't easy to process.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109043 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109042 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing AVX convert instructions. Those instructions are not described in their SSE forms (although they exist), but add the AVX forms anyway, so the assembler can benefit from it

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109039 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109038 91177308-0d34-0410-b5e6-96231b3b80d8

80 column and trailing whitespace cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109037 91177308-0d34-0410-b5e6-96231b3b80d8

Add some debug output to help diagnose PR7689.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109036 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple issues with Win64 ABI

1) all registers were spilled as xmm, regardless of actual size
2) win64 abi doesn't do the varargs-size-in-%al thing

Still to look into:

xmm6-15 are marked as clobbered by call instructions on win64 even though they aren't.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109035 91177308-0d34-0410-b5e6-96231b3b80d8

Add the INITIALIZE_PASS macro.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109034 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid AVX instructions to be selected instead of its SSE form

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109032 91177308-0d34-0410-b5e6-96231b3b80d8

NamedMDNode is never an operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109031 91177308-0d34-0410-b5e6-96231b3b80d8

Disallow null as a named metadata operand.
Make MDNode::destroy private.
Fix the one thing that used MDNode::destroy, outside of MDNode itself.

One should never delete or destroy an MDNode explicitly. MDNodes
implicitly go away when there are no references to them (implementation
details aside).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109028 91177308-0d34-0410-b5e6-96231b3b80d8

Use System/DataTypes.h instead of stdint.h, which isn't portable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109026 91177308-0d34-0410-b5e6-96231b3b80d8

Use TrackingVH instead of WeakVH for NamedMDNode's operands, since nodes
referenced by NamedMDNodes shouldn't be deleted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109021 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109020 91177308-0d34-0410-b5e6-96231b3b80d8

Move the smarts of AnalysisGroup registration into PassRegistry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109019 91177308-0d34-0410-b5e6-96231b3b80d8

Delete an obsolete comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109017 91177308-0d34-0410-b5e6-96231b3b80d8

strip out the 2.7 release notes, this really is the 2.8 release notes
now. Add a few items like the fpstackifier improvements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109013 91177308-0d34-0410-b5e6-96231b3b80d8

Fix calling convention on ARM if vfp2+ is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109009 91177308-0d34-0410-b5e6-96231b3b80d8

appease Validator

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109007 91177308-0d34-0410-b5e6-96231b3b80d8

add info on operand rotation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109006 91177308-0d34-0410-b5e6-96231b3b80d8