Add a few more functions to TargetLibraryInfo. More of rdar://10500969.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145596 91177308-0d34-0410-b5e6-96231b3b80d8

fix broken tag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145590 91177308-0d34-0410-b5e6-96231b3b80d8

Revert commit 145449 (ddunbar) since it is breaking the dragonegg buildbots.
Original commit message:
llvm-config: Replace with C++ version (was llvm-config-2).
 - Reapply of r144300, with lots of fixes/migration easement in between.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145582 91177308-0d34-0410-b5e6-96231b3b80d8

For 64-bit the rest of the general regs are ok for the q constraint. Make
sure we can emit both the high and low versions of those registers.

Fixes rdar://10392864

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145579 91177308-0d34-0410-b5e6-96231b3b80d8

Add some missing anchors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145578 91177308-0d34-0410-b5e6-96231b3b80d8

Pass AVX vectors which are arguments to varargs functions on the stack. <rdar://problem/10463281>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145573 91177308-0d34-0410-b5e6-96231b3b80d8

Removed use of grep from test and moved it to be with other icmp tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145570 91177308-0d34-0410-b5e6-96231b3b80d8

Added instcombine pattern to spot comparing -val or val against 0.
(val != 0) == (-val != 0) so "abs(val) != 0" becomes "val != 0"

Fixes <rdar://problem/10482509>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145563 91177308-0d34-0410-b5e6-96231b3b80d8

Propagate TargetLibraryInfo throughout ConstantFolding.cpp and
InstructionSimplify.cpp.  Other fixups as needed.
Part of rdar://10500969

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145559 91177308-0d34-0410-b5e6-96231b3b80d8

Make use of "getScalarType()". No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145556 91177308-0d34-0410-b5e6-96231b3b80d8

Small fix for assembler generation on Darwin PPC64.  Patch by Michael Kostylev.  PR11437.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145553 91177308-0d34-0410-b5e6-96231b3b80d8

Replace an assert() with an actual diagnostic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145535 91177308-0d34-0410-b5e6-96231b3b80d8

make asan work at -O0, llvm part. Patch by glider@google.com

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145530 91177308-0d34-0410-b5e6-96231b3b80d8

Support for encoding all FMA4 instructions and tablegen patterns for all
remaining FMA4 instructions and intrinsics with tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145525 91177308-0d34-0410-b5e6-96231b3b80d8

Make GlobalMerge honor the preferred alignment on globals without an explicitly specified alignment.

<rdar://problem/10497732>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145523 91177308-0d34-0410-b5e6-96231b3b80d8

More cleanups. No content change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145522 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanup. No content change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145521 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the install-clang-c makefile target.

When I did this before it broke a buildbot that was testing that target, but
we've removed that buildbot now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145519 91177308-0d34-0410-b5e6-96231b3b80d8

Remove an XXX which hasn't been fixed yet. It's too late now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145518 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused variable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145517 91177308-0d34-0410-b5e6-96231b3b80d8

Add some tests for all-lanes VLD1 parsing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145512 91177308-0d34-0410-b5e6-96231b3b80d8

ARM parsing for VLD1 all lanes, with writeback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145510 91177308-0d34-0410-b5e6-96231b3b80d8

Add a few functions to TargetLibraryInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145508 91177308-0d34-0410-b5e6-96231b3b80d8

ARM parsing for VLD1 two register all lanes, no writeback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145504 91177308-0d34-0410-b5e6-96231b3b80d8

Add test arch to make it pass on non x86 targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145498 91177308-0d34-0410-b5e6-96231b3b80d8

X86: Turns out bulldozer also supports sse42 and lzcnt.

While at it remove the barcelona/instanbul/shanghai subtargets, they're
unsupported by GCC and look pretty broken.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145494 91177308-0d34-0410-b5e6-96231b3b80d8

X86: Add subtargets for AMD's bulldozer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145493 91177308-0d34-0410-b5e6-96231b3b80d8

Add a tripple to the test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145489 91177308-0d34-0410-b5e6-96231b3b80d8

X86: PerformOrCombine introduced a vselect node with a wrong order of operands. This bug was introduced when a dedicated blend sdnode was replaced with the vselect node (in 139479).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145488 91177308-0d34-0410-b5e6-96231b3b80d8

Add instruction selection support for AVX2 horizontal add/sub instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145487 91177308-0d34-0410-b5e6-96231b3b80d8

Mention that -O4 does result in more optimization when used with
-fplugin-arg-dragonegg-enable-gcc-optzns, though it usually isn't
a win.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145486 91177308-0d34-0410-b5e6-96231b3b80d8

Merge VPERM2F128/VPERM2I128 ISD node types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145485 91177308-0d34-0410-b5e6-96231b3b80d8

Better test case found in duplicate PR10570.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145484 91177308-0d34-0410-b5e6-96231b3b80d8

Merge decoding of VPERMILPD and VPERMILPS shuffle masks. Merge X86ISD node type for VPERMILPD/PS. Add instruction selection support for VINSERTI128/VEXTRACTI128.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145483 91177308-0d34-0410-b5e6-96231b3b80d8

LSR: handle the expansion of phi operands that use postinc forms of the IV.

Fixes PR11431: SCEVExpander::expandAddRecExprLiterally(const llvm::SCEVAddRecExpr*): Assertion `(!isa<Instruction>(Result) || SE.DT->dominates(cast<Instruction>(Result), Builder.GetInsertPoint())) && "postinc expansion does not dominate use"' failed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145482 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145470 91177308-0d34-0410-b5e6-96231b3b80d8

Alphabetize TargetLibraryInfo enum and fix doxygen comments.  No functional
change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145468 91177308-0d34-0410-b5e6-96231b3b80d8

llvm_unreachable() is not for user diagnostics....

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145465 91177308-0d34-0410-b5e6-96231b3b80d8

ARM parsing aliases for VLD1 single register all lanes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145464 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for sqrt, sqrtl, and sqrtf in TargetLibraryInfo.  Disable
(fptrunc (sqrt (fpext x))) -> (sqrtf x) transformation if -fno-builtin is
specified.
rdar://10466410

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145460 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145458 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145456 91177308-0d34-0410-b5e6-96231b3b80d8

ARM parsing aliases for data-size suffices on VST1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145454 91177308-0d34-0410-b5e6-96231b3b80d8

FileCheckize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145452 91177308-0d34-0410-b5e6-96231b3b80d8

Change names for MIPS "generic" processors defined in Mips.td to match what GNU
tools use. Patch by Simon Atanasyan.

"mips32r1" => "mips32"
"4ke" => mips32r2"
"mips64r1" => "mips64"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145451 91177308-0d34-0410-b5e6-96231b3b80d8

ARM assembly parsing and encoding for four-register VST1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145450 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-config: Replace with C++ version (was llvm-config-2).
 - Reapply of r144300, with lots of fixes/migration easement in between.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145449 91177308-0d34-0410-b5e6-96231b3b80d8

Add another missing pattern. llvm-gcc likes f64 but clang likes i64 so it was generating poor code for some SSE builtins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145448 91177308-0d34-0410-b5e6-96231b3b80d8

Enable some VST1 tests and add a few more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145443 91177308-0d34-0410-b5e6-96231b3b80d8

ARM assembly parsing and encoding for three-register VST1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145442 91177308-0d34-0410-b5e6-96231b3b80d8

Make X86::FsFLD0SS / FsFLD0SD real pseudo-instructions.

Like V_SET0, these instructions are expanded by ExpandPostRA to xorps /
vxorps so they can participate in execution domain swizzling.

This also makes the AVX variants redundant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145440 91177308-0d34-0410-b5e6-96231b3b80d8

Potential bug in RewriteLoopBodyWithConditionConstant: use iterator should not be changed inside the uses enumeration loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145432 91177308-0d34-0410-b5e6-96231b3b80d8

Update the docs for some of the test-suite configure changes and
be more clear about what to do and how to do it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145426 91177308-0d34-0410-b5e6-96231b3b80d8

If fast-isel fails, remove dead instructions generated during the failed
attempt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145425 91177308-0d34-0410-b5e6-96231b3b80d8

grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145423 91177308-0d34-0410-b5e6-96231b3b80d8

comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145422 91177308-0d34-0410-b5e6-96231b3b80d8

build/CMake: Finish removal of add_llvm_library_dependencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145420 91177308-0d34-0410-b5e6-96231b3b80d8

Add a link to Bill's blog post.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145419 91177308-0d34-0410-b5e6-96231b3b80d8

Release notes for segmented stacks.
Patch by Sanjoy Das.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145416 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a theoretical problem (not seen in the wild): if different instances of a
weak variable are compiled by different compilers, such as GCC and LLVM, while
LLVM may increase the alignment to the preferred alignment there is no reason to
think that GCC will use anything more than the ABI alignment.  Since it is the
GCC version that might end up in the final program (as the linkage is weak), it
is wrong to increase the alignment of loads from the global up to the preferred
alignment as the alignment might only be the ABI alignment.

Increasing alignment up to the ABI alignment might be OK, but I'm not totally
convinced that it is.  It seems better to just leave the alignment of weak
globals alone.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145413 91177308-0d34-0410-b5e6-96231b3b80d8

MC/X86/COFF: Allow quotes in names when targeting MS/Windows,
as MC is the only assembler we support.

This splits MS/Windows and GNU/Windows ASM infos into two seperate classes.
While there is currently only one difference, full MS C++ ABI support will
require many more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145409 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed ObjectFile functions:
- getSymbolOffset() renamed as getSymbolFileOffset()
- getSymbolFileOffset(), getSymbolAddress(), getRelocationAddress() returns same result for ELFObjectFile, MachOObjectFile and COFFObjectFile.
- added getRelocationOffset()
- fixed MachOObjectFile::getSymbolSize()
- fixed MachOObjectFile::getSymbolSection()
- fixed MachOObjectFile::getSymbolOffset() for symbols without section data.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145408 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed vsqrt.ss intrinsic usage - order of input operands was wrong.
Added a test.
Thanks Bruno for reviewing the patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145403 91177308-0d34-0410-b5e6-96231b3b80d8

Fix shuffle decoding for memory forms for (V)SHUFPS/D.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145392 91177308-0d34-0410-b5e6-96231b3b80d8

Fix issues in shuffle decoding around VPERM* instructions. Fix shuffle decoding for VSHUFPS/D for 256-bit types. Add pattern matching for memory forms of VPERMILPS/VPERMILPD.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145390 91177308-0d34-0410-b5e6-96231b3b80d8

[Win32] Catch exceptions (eg. segfault) on waiting for invoked clang from the driver.

clang/lib/Driver/Driver.cpp: Don't pass through negative exit status, or parent would be confused.

llvm::sys::Program::Wait(): Suppose 0x8000XXXX and 0xC000XXXX as abnormal exit code and pass it as negative value.

Win32 Exception Handler: Exit with ExceptionCode on an unhandle exception.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145389 91177308-0d34-0410-b5e6-96231b3b80d8

lit/TestRunner.py: Try to catch ERROR_FILE_NOT_FOUND, too.

Thanks to Francois, to let me know.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145381 91177308-0d34-0410-b5e6-96231b3b80d8

Install llvmCore to /usr/local.  <rdar://problem/10390708>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145378 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VINSERTF128/VEXTRACTF128 to be marked as FP instructions. Allow execution dependency fix pass to convert them to their integer equivalents when AVX2 is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145376 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly mark VPERM2F128 as being an FP instruction and add execution domain fixing support to convert it to VPERM2I128 for AVX2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145370 91177308-0d34-0410-b5e6-96231b3b80d8

MachO doesn't support the protected visibility. Don't default to 'global' here. <rdar://problem/10396775>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145368 91177308-0d34-0410-b5e6-96231b3b80d8

SCEV fix. In general, Add/Mul expressions should not inherit NSW/NUW.

This reverts r139450, fixes r139453, and adds much needed comments and a
unit test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145367 91177308-0d34-0410-b5e6-96231b3b80d8

Make SCEV print <nsw><nuw> for Add/MulExpr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145364 91177308-0d34-0410-b5e6-96231b3b80d8

Filecheckize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145363 91177308-0d34-0410-b5e6-96231b3b80d8

Remove content that has been moved to Clang release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145362 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145361 91177308-0d34-0410-b5e6-96231b3b80d8

On MachO, the pointer to the personality function should always be in the
non_lazy_symbol_pointers section (__IMPORT,__pointers). Ignore the 'hidden' part
since that will place it in the wrong section.
<rdar://problem/10443720>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145356 91177308-0d34-0410-b5e6-96231b3b80d8

build/cmake: Switch to using llvm-build computed dependencies.
 - I verified locally that the current dependency lists are identical.
 - This makes add_llvm_library_dependencies() a no-op. I'll remove it once this
   change passes the bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145355 91177308-0d34-0410-b5e6-96231b3b80d8

Zap some completely ridiculous code.  There's probably a miscompile here, but I don't really want to try to write a testcase involving an invoke returning a pointer to a varargs function...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145347 91177308-0d34-0410-b5e6-96231b3b80d8

Better fix for ARM MOVT relocation encoding of thumb bit.

Replaces r145318 with a more targetted fix for the relocation handling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145346 91177308-0d34-0410-b5e6-96231b3b80d8

Reenable this IndVars unit test.

SCEV can't optimize undef in all cases, which is a separate issue from this test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145343 91177308-0d34-0410-b5e6-96231b3b80d8

build: Update CMakeLists.txt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145341 91177308-0d34-0410-b5e6-96231b3b80d8

Add a link from the LLVM release notes to the Clang release notes.
I suspect we could profitably remove/move some of the bullet points
under Clang here to the Clang notes in order to keep things clean on
both sides. Unless I hear objections I'll start doing that once folks
have read over the Clang notes a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145340 91177308-0d34-0410-b5e6-96231b3b80d8

edis: Sink EDMain.cpp into lib/MC/MCDisassembler.
 - This fixes some layering violations and matches how we handle the llvm-c lib,
   for example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145338 91177308-0d34-0410-b5e6-96231b3b80d8

edis: Don't do the target initialization in EDGetDisassembler, this is contrary
to the way we currently expect target selection to work -- clients are supposed
to have control over what targets are available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145331 91177308-0d34-0410-b5e6-96231b3b80d8

llvm-c: Add a few missing InitializeAll* functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145330 91177308-0d34-0410-b5e6-96231b3b80d8

build/Make: edis isn't built as a shared library anymore, remove related cruft from the Makefile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145329 91177308-0d34-0410-b5e6-96231b3b80d8

llvmbuild/CMake: Update CMake output fragment to include explicit library
dependency information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145328 91177308-0d34-0410-b5e6-96231b3b80d8

Expand the part about CFI a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145324 91177308-0d34-0410-b5e6-96231b3b80d8

Add documentation for llvm-cov.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145319 91177308-0d34-0410-b5e6-96231b3b80d8

Thumb2 only force the fixup thumb bit for data relocations.

rdar://10493453

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145318 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missing safety check to ProcessUGT_ADDCST_ADD.  Fixes PR11438.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145316 91177308-0d34-0410-b5e6-96231b3b80d8

Remove obsolete FIXME.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145313 91177308-0d34-0410-b5e6-96231b3b80d8

Make SelectionDAG::InferPtrAlignment use llvm::ComputeMaskedBits instead of duplicating the logic for globals.  Make llvm::ComputeMaskedBits handle GlobalVariables slightly more aggressively, to match what InferPtrAlignment knew how to do.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145304 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r145273 and fix in SelectionDAG::InferPtrAlignment() instead.
Conservatively returns zero when the GV does not specify an alignment nor is it
initialized. Previously it returns ABI alignment for type of the GV. However, if
the type is a "packed" type, then the under-specified alignments is attached to
the load / store instructions. In that case, the alignment of the type cannot be
trusted.
rdar://10464621

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145300 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some possible gcc-4.2 may be used uninitialized warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145292 91177308-0d34-0410-b5e6-96231b3b80d8

Don't define these unless we plan to use them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145289 91177308-0d34-0410-b5e6-96231b3b80d8

Merging two bullet points into one

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145287 91177308-0d34-0410-b5e6-96231b3b80d8

DAG combine should not increase alignment of loads / stores with alignment less
than ABI alignment. These are loads / stores from / to "packed" data structures.
Their alignments are intentionally under-specified.

rdar://10301431

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145273 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing avx pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145272 91177308-0d34-0410-b5e6-96231b3b80d8