InstCombine: When comparing two GEPs that were derived from the same base pointer but use different types, expand the offset calculation and to the compare on the offset if profitable.

This came up in SmallVector code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150962 91177308-0d34-0410-b5e6-96231b3b80d8

InstCombine: Make OptimizePointerDifference more aggressive.

- Ignore pointer casts.
- Also expand GEPs that aren't constantexprs when they have one use or only constant indices.

- We now compile "&foo[i] - &foo[j]" into "i - j".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150961 91177308-0d34-0410-b5e6-96231b3b80d8

Teach the DAGCombiner that certain loadext nodes followed by ANDs can be converted to zeroexts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150957 91177308-0d34-0410-b5e6-96231b3b80d8

Improve generated code for extending loads and some trunc stores on ARM.

Teach TargetSelectionDAG about lengthening loads for vector types and set v4i8 as legal. Allow FP_TO_UINT for v4i16 from v4i32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150956 91177308-0d34-0410-b5e6-96231b3b80d8

Make post-ra tail duplication bundle safe. No test case as recent codegen
flow changes have already hidden the bug. rdar://10893812

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150949 91177308-0d34-0410-b5e6-96231b3b80d8

Rename class Evaluate to Evaluator and put it in an anonymous namespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150947 91177308-0d34-0410-b5e6-96231b3b80d8

fold comparisons of gep'd alloca points with null to false,
implementing PR12013.  We now compile the testcase to:

__Z4testv:                              ## @_Z4testv
## BB#0:                                ## %_ZN4llvm15SmallVectorImplIiE9push_backERKi.exit
pushq %rbx
subq $64, %rsp
leaq 32(%rsp), %rbx
movq %rbx, (%rsp)
leaq 64(%rsp), %rax
movq %rax, 16(%rsp)
movl $1, 32(%rsp)
leaq 36(%rsp), %rax
movq %rax, 8(%rsp)
leaq (%rsp), %rdi
callq __Z1gRN4llvm11SmallVectorIiLj8EEE
movq (%rsp), %rdi
cmpq %rbx, %rdi
je LBB0_2
## BB#1:
callq _free
LBB0_2:                                 ## %_ZN4llvm11SmallVectorIiLj8EED1Ev.exit
addq $64, %rsp
popq %rbx
ret

instead of:

__Z4testv:                              ## @_Z4testv
## BB#0:
pushq %rbx
subq $64, %rsp
xorl %eax, %eax
leaq (%rsp), %rbx
addq $32, %rbx
movq %rbx, (%rsp)
movq %rbx, 8(%rsp)
leaq 64(%rsp), %rcx
movq %rcx, 16(%rsp)
je LBB0_2
## BB#1:
movl $1, 32(%rsp)
movq %rbx, %rax
LBB0_2:                                 ## %_ZN4llvm15SmallVectorImplIiE9push_backERKi.exit
addq $4, %rax
movq %rax, 8(%rsp)
leaq (%rsp), %rdi
callq __Z1gRN4llvm11SmallVectorIiLj8EEE
movq (%rsp), %rdi
cmpq %rbx, %rdi
je LBB0_4
## BB#3:
callq _free
LBB0_4:                                 ## %_ZN4llvm11SmallVectorIiLj8EED1Ev.exit
addq $64, %rsp
popq %rbx
ret

This doesn't shrink clang noticably though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150944 91177308-0d34-0410-b5e6-96231b3b80d8

Switch to a more idiomatic way of silencing unused variable warnings in
release builds. Silences clang's -Wself-assign.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150942 91177308-0d34-0410-b5e6-96231b3b80d8

Move constructors out-of-line and flesh out their documentation. No
functionality changed. This is in preparation for some refactoring of
how this class behaves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150941 91177308-0d34-0410-b5e6-96231b3b80d8

Move EvaluateFunction and EvaluateBlock into a class, and make the class store
the information that they pass around between them. No functionality change!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150939 91177308-0d34-0410-b5e6-96231b3b80d8

Silence operator precedence warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150921 91177308-0d34-0410-b5e6-96231b3b80d8

autoconf/m4/visibility_inlines_hidden.m4: Tweak for cygwin-g++-4.5.3. It folded conftest regardress of attributge((noinline)) with >=O1 .

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150920 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code. Improve llvm_unreachable text. Simplify some control flow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150918 91177308-0d34-0410-b5e6-96231b3b80d8

StringRef'ize EmitSourceFileHeader().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150917 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some unneeded includes and fix ordering in X86ISelLowering.cpp. Remove unneeded 'using namespace'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150916 91177308-0d34-0410-b5e6-96231b3b80d8

Add machinery for pushing live ranges onto bundle starts while bundling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150915 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify moveEnteringDownFrom rules.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150914 91177308-0d34-0410-b5e6-96231b3b80d8

Unify all shuffle mask checking functions take a mask and VT instead of VectorShuffleSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150913 91177308-0d34-0410-b5e6-96231b3b80d8

Skip through instructions rather than operands when looking for last use slot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150912 91177308-0d34-0410-b5e6-96231b3b80d8

APFloat::toString(): Fix overrun at scanning.

FYI, clang/test/SemaTemplate/template-id-printing.cpp had been failing due to it on cygwin-clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150911 91177308-0d34-0410-b5e6-96231b3b80d8

Fix TODO and trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150910 91177308-0d34-0410-b5e6-96231b3b80d8

Defer sanity checks on live intervals until after all have been updated. Hold (LiveInterval, LiveRange) pairs to update, rather than vregs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150909 91177308-0d34-0410-b5e6-96231b3b80d8

Make a bunch of X86ISelLowering shuffle functions static now that they are no longer needed by isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150908 91177308-0d34-0410-b5e6-96231b3b80d8

comment fix ARM.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150904 91177308-0d34-0410-b5e6-96231b3b80d8

some comment fix for X86 and ARM

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150902 91177308-0d34-0410-b5e6-96231b3b80d8

Add vmfunc instruction to X86 assembler and disassembler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150899 91177308-0d34-0410-b5e6-96231b3b80d8

Fix issue with bitwise and precedence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150897 91177308-0d34-0410-b5e6-96231b3b80d8

Hashing.h - utilities for hashing various data types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150890 91177308-0d34-0410-b5e6-96231b3b80d8

White space fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150886 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily disable this assert. Looks like it found a similar issue when
building bullet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150885 91177308-0d34-0410-b5e6-96231b3b80d8

Don't skip debug instructions when looking for the insertion point of
the cast. If we do, we can end up with

   inst1
   ---------------  < Insertion point
   dbg inst
   new inst

instead of the desired

   inst1
   new inst
   ---------------  < Insertion point
   dbg inst

Another option would be for InsertNoopCastOfTo (or its callers) to move the
insertion point and we would end up with

   inst1
   dbg inst
   new inst
   ---------------  < Insertion point

but that complicates the callers. This fixes PR12018 (and firefox's build).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150884 91177308-0d34-0410-b5e6-96231b3b80d8

Emacs-tag and some comment fix for all ARM, CellSPU, Hexagon, MBlaze, MSP430, PPC, PTX, Sparc, X86, XCore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150878 91177308-0d34-0410-b5e6-96231b3b80d8

On Cygwin/MingW, add SharedLibDir and LLVMToolDir to the library search path, since shared libraries are placed in 'bin'. (static libraries are still in 'lib').

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150876 91177308-0d34-0410-b5e6-96231b3b80d8

Add X86InstrSVM.td that I forgot to add in r150873.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150874 91177308-0d34-0410-b5e6-96231b3b80d8

Add X86 assembler and disassembler support for AMD SVM instructions. Original patch by Kay Tiong Khoo. Few tweaks by me for code density and to reduce replication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150873 91177308-0d34-0410-b5e6-96231b3b80d8

Trivial cleanup to group the generic 'armvN' cases with the 'arm' case,
etc. No functionality changed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150867 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a rather nasty regression from r150690: LHS != RHS does not imply LHS->stripPointerCasts() != RHS->stripPointerCasts().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150863 91177308-0d34-0410-b5e6-96231b3b80d8

Fix documentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150860 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for the previous commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150852 91177308-0d34-0410-b5e6-96231b3b80d8

Bring HMEditor into line with LLVM coding standards.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150851 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore the lifetime intrinsics in fast-isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150848 91177308-0d34-0410-b5e6-96231b3b80d8

Don't print out pointer values in SUnit::dump().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150842 91177308-0d34-0410-b5e6-96231b3b80d8

Sink variable into assert

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150841 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for regmask slots to HMEditor. Also fixes a comment error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150840 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in comment ldopen() -> dlopen().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150836 91177308-0d34-0410-b5e6-96231b3b80d8

Handle regmask operands in ARMInstrInfo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150833 91177308-0d34-0410-b5e6-96231b3b80d8

Put back the initializing the targets in the disassembler API with a comment as
to why this is needed.  This broke the darwin's otool(1) program.  This change
was made in r144385.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150832 91177308-0d34-0410-b5e6-96231b3b80d8

Fix ARMBaseInstrInfo::getInstrLatency for calls.

Calls always clobber CPSR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150831 91177308-0d34-0410-b5e6-96231b3b80d8

Transfer regmasks to MRI.

MRI keeps track of which physregs have been used. Make sure it gets
updated with all the regmask-clobbered registers.

Delete the closePhysRegsUsed() function which isn't necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150830 91177308-0d34-0410-b5e6-96231b3b80d8

Calls and invokes with the new clang.arc.no_objc_arc_exceptions
metadata may still unwind, but only in ways that the ARC
optimizer doesn't need to consider. This permits more
aggressive optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150829 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor 'handleMove' code in live intervals. Clients of LiveIntervals won't see
any changes.

Internally this adds a private inner class HMEditor, to LiveIntervals. HMEditor provides
an API for updating live intervals when code is moved or bundled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150826 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a comment about an alternative approach that wouldn't
actually work, at least as described. LLVM Metadata is not
intended to suppress LLVM IR rules, as it can be stripped at
any time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150821 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150820 91177308-0d34-0410-b5e6-96231b3b80d8

Generate the correct EH frame section types on Solaris, this time without breaking other platforms...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150819 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r150814.  It turns out that there is a good reason for this after all...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150818 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r150288, "Allow Post-RA LICM to hoist reserved register reads."

This caused miscompilations on out-of-tree targets, and possibly i386 as
well.

I'll find some other way of hoisting %rip-relative loads from loops
containing calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150816 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150815 91177308-0d34-0410-b5e6-96231b3b80d8

Don't lazily allocate eh_frame.  We're not lazily allocating things like the LSDA, which are only used when the eh frame is used, so this lazy allocation doesn't really make sense.

Fix the type of eh_frame on Solaris so that Sun ld doesn't fail to combine them (thus making it impossible for the unwind library to find them and breaking exceptions).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150814 91177308-0d34-0410-b5e6-96231b3b80d8

... and it's probably best to use the correct alignment, rather than just guessing that it's the same as the size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150813 91177308-0d34-0410-b5e6-96231b3b80d8

It turns out that putting an 8-byte symbol in a 4-byte section makes Solaris ld sulk.  GNU ld is perfectly happy with it, which is worrying for a whole other set of reasons...

Thanks to Anton, Duncan and Rafael for helping me track this down.
Pointy hat to Rafael for introducing the bug in the first place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150811 91177308-0d34-0410-b5e6-96231b3b80d8

Remove question.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150809 91177308-0d34-0410-b5e6-96231b3b80d8

remove Emacs-tag form .cpp files in Mips Backend, and fix some typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150805 91177308-0d34-0410-b5e6-96231b3b80d8

Typo in variable name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150796 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the last of the old vector_shuffle patterns from X86 isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150795 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for invariant.start inside the static constructor evaluator. This is
useful to represent a variable that is const in the source but can't be constant
in the IR because of a non-trivial constructor. If globalopt evaluates the
constructor, and there was an invariant.start with no matching invariant.end
possible, it will mark the global constant afterwards.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150794 91177308-0d34-0410-b5e6-96231b3b80d8

Do not promote i32 arguments to i64. This was causing unnecessary sign extension
instructions to be emitted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150782 91177308-0d34-0410-b5e6-96231b3b80d8

Remove redundant comment. Use a more efficient datatype.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150780 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some grammar-os and formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150779 91177308-0d34-0410-b5e6-96231b3b80d8

Reverse iterator - should be incrementing rather than decrementing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150778 91177308-0d34-0410-b5e6-96231b3b80d8

add Emacs tag and fix some comment error in file headers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150775 91177308-0d34-0410-b5e6-96231b3b80d8

[fast-isel] Add support for returning non-legal types with no sign- or zero-
entend flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150774 91177308-0d34-0410-b5e6-96231b3b80d8

MachineScheduler shouldn't use/preserve LiveDebugVariables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150773 91177308-0d34-0410-b5e6-96231b3b80d8

Oops - isRegLiveIntoSuccessor is used in non-assert builds now. Remove NDEBUG guards.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150771 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable 150652 and 150654 - Make FPSCR non-reserved, and make MachineCSE bail on reserved registers. This *should* be safe as of r150786.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150769 91177308-0d34-0410-b5e6-96231b3b80d8

Turn off assertion, conservatively compute liveness for live-in un-allocatable registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150768 91177308-0d34-0410-b5e6-96231b3b80d8

Remove comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150739 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary assignment to temporary, ResultReg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150737 91177308-0d34-0410-b5e6-96231b3b80d8

Use –mcpu=generic, so that the test will not fail when run on an Intel Atom
processor, due to the Atom scheduler producing an instruction sequence that is
different from that which is expected.
Patch by Michael Spencer!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150736 91177308-0d34-0410-b5e6-96231b3b80d8

Update credits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150715 91177308-0d34-0410-b5e6-96231b3b80d8

Note x86 regmask operands in release notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150712 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the YMM_HI_6_15 hack.

Call clobbers are now represented with register mask operands.  The
regmask can easily represent the fact that xmm6 is call-preserved while
ymm6 isn't.  This is automatically computed by TableGen from the
CalleeSavedRegs containing xmm6.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150709 91177308-0d34-0410-b5e6-96231b3b80d8

Use the same CALL instructions for Windows as for everything else.

The different calling conventions and call-preserved registers are
represented with regmask operands that are added dynamically.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150708 91177308-0d34-0410-b5e6-96231b3b80d8

Remove trailing whitespace. Add newline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150706 91177308-0d34-0410-b5e6-96231b3b80d8

Disable machine copy propagation for now. It's known to be buggy (PR11940) and introduces subtle miscompiles in many places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150703 91177308-0d34-0410-b5e6-96231b3b80d8

Revert "InstSimplify: Strip pointer casts early."

Turns out this isn't safe, because the code below depends on LHS and RHS having
the same type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150695 91177308-0d34-0410-b5e6-96231b3b80d8

InstSimplify: Strip pointer casts early.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150694 91177308-0d34-0410-b5e6-96231b3b80d8

InstSimplify: Ignore pointer casts when constant folding compares between pointers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150690 91177308-0d34-0410-b5e6-96231b3b80d8

s/ModAttrBehavior/ModFlagBehavior/g to be consistent with how module flags are named elsewhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150679 91177308-0d34-0410-b5e6-96231b3b80d8

Add 'exception handling' to my description.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150678 91177308-0d34-0410-b5e6-96231b3b80d8

Add blurb about module flags and reformat a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150677 91177308-0d34-0410-b5e6-96231b3b80d8

Remove extraneous #include and spelling mistake introduced in r150669.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150670 91177308-0d34-0410-b5e6-96231b3b80d8

Modify the algorithm when traversing the DAGCombiner's worklist to be O(log N) for all operations. This fixes a horrible worst case with lots of nodes where 99% of the time was being spent in std::remove.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150669 91177308-0d34-0410-b5e6-96231b3b80d8

VMCore/AsmWriter.cpp: Tweak to check #INF and #NAN earlier.

With MSVCRT, prior checker missed emission of #INF and #NAN.

FIXME: Checking should be simpler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150667 91177308-0d34-0410-b5e6-96231b3b80d8

Replace all instances of dg.exp file with lit.local.cfg, since all tests are run with LIT now and now Dejagnu. dg.exp is no longer needed.

Patch reviewed by Daniel Dunbar. It will be followed by additional cleanup patches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150664 91177308-0d34-0410-b5e6-96231b3b80d8

Template specialize SmallVector::push_back based on POD-ness of the type.  Reduces clang binary by 188KB

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150662 91177308-0d34-0410-b5e6-96231b3b80d8

VMCore/AsmWriter.cpp: Use APFloat instead of atof(3).

atof(3) might behave differently among platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150661 91177308-0d34-0410-b5e6-96231b3b80d8

Oop - r150653 + r150654 broke one of my test cases. Backing out for now...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150655 91177308-0d34-0410-b5e6-96231b3b80d8

FPSCR shouldn't be reserved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150654 91177308-0d34-0410-b5e6-96231b3b80d8

MachineCSE shouldn't extend the live ranges of reserved or allocatable registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150653 91177308-0d34-0410-b5e6-96231b3b80d8

Give a description of the Objective-C garbage collection module flags.

The rule governing the flags is this:

  no-gc + no-gc   = no-gc
  no-gc +    gc   = no-gc
  no-gc + gc-only = error
     gc +    gc   = gc
     gc + gc-only = gc-only
gc-only + gc-only = gc-only

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@150646 91177308-0d34-0410-b5e6-96231b3b80d8