pass machinemoduleinfo down into getSymbolForDwarfGlobalReference,
currently unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82157 91177308-0d34-0410-b5e6-96231b3b80d8

Teach ScalarEvolution how to reason about no-wrap flags on loops
where the induction variable has a non-unit stride, such as {0,+,2}, and
there are expressions such as {1,+,2} inside the loop formed with
or or add nsw operators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82151 91177308-0d34-0410-b5e6-96231b3b80d8

grammar

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82150 91177308-0d34-0410-b5e6-96231b3b80d8

grammar

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82149 91177308-0d34-0410-b5e6-96231b3b80d8

Another try at fixing compile warnings on 4.0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82148 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize HasMetadata to zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82145 91177308-0d34-0410-b5e6-96231b3b80d8

Remove test cases using -regalloc=simple.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82130 91177308-0d34-0410-b5e6-96231b3b80d8

Remove simple regalloc. It has bit rotted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82127 91177308-0d34-0410-b5e6-96231b3b80d8

add a version of the APFloat constructor that initializes to 0.0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82110 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR4910: Broken logic in coalescer means when a physical register liveness is being shortened, the sub-registers were not. The symptom is the register allocator could not find a free register for this particular test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82108 91177308-0d34-0410-b5e6-96231b3b80d8

Some platforms may need malloc.h for alloca.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82100 91177308-0d34-0410-b5e6-96231b3b80d8

Update CMake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82097 91177308-0d34-0410-b5e6-96231b3b80d8

Added the LODS (load byte into register, usually
as part string parsing) instructions to the Intel
instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82089 91177308-0d34-0410-b5e6-96231b3b80d8

Add StringRef::{rfind, rsplit}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82087 91177308-0d34-0410-b5e6-96231b3b80d8

Added the LAR (load segment access rights)
instructions to the Intel instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82084 91177308-0d34-0410-b5e6-96231b3b80d8

Added the LOOP family of instructions to the Intel
instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82083 91177308-0d34-0410-b5e6-96231b3b80d8

Added an alternate form of register-register CMP
to the Intel instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82081 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82080 91177308-0d34-0410-b5e6-96231b3b80d8

At iSel time, update DebugLoc based on debug info attached with an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82077 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new pass for doing late hoisting of floating-point and vector
constants out of loops. These aren't covered by the regular LICM
pass, because in LLVM IR constants don't require separate
instructions. They're not always covered by the MachineLICM pass
either, because it doesn't know how to unfold folded constant-pool
loads. This is somewhat experimental at this point, and off by
default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82076 91177308-0d34-0410-b5e6-96231b3b80d8

Print debug info attached with an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82075 91177308-0d34-0410-b5e6-96231b3b80d8

Expand vector floating-point conversions not supported by NEON.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82074 91177308-0d34-0410-b5e6-96231b3b80d8

Now that llc can read .ll files directly, teach it to recognize .ll as
an extension, so that the default output filename for foo.ll is foo.s,
not foo.ll.s

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82071 91177308-0d34-0410-b5e6-96231b3b80d8

Provide a way to extract location info from DILocation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82064 91177308-0d34-0410-b5e6-96231b3b80d8

Parse debug info attached with an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82063 91177308-0d34-0410-b5e6-96231b3b80d8

Add an interface to attach debugging information with an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82062 91177308-0d34-0410-b5e6-96231b3b80d8

Add llvm::Metadata to manage metadata used in a context.
This interface will be used to attach metadata with an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82060 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed some problems with the logic of parsing line comments by adding
isAtStartOfComment and using that instead in two places where a loop
to check if the char was in MAI.getCommentString().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82059 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect assert that should be a user error for code like 'mov $0, %%eax'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82054 91177308-0d34-0410-b5e6-96231b3b80d8

Change FoldPHIArgBinOpIntoPHI to decline folding if it would introduce two
phis, similar to the FoldPHIArgGEPIntoPHI change.

Also, delete some comments that don't reflect the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82053 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the comment in this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82051 91177308-0d34-0410-b5e6-96231b3b80d8

Make cmake generated llvm-config output correct JIT backend for non X86 targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82049 91177308-0d34-0410-b5e6-96231b3b80d8

Don't sort the vector when it is empty. This should fix some expensive checking
failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82040 91177308-0d34-0410-b5e6-96231b3b80d8

Reapplied r81355 with the problems fixed.
(See http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20090907/086737.html and
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20090907/086746.html)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82039 91177308-0d34-0410-b5e6-96231b3b80d8

updated lib/CodeGen/CMakeLists.txt to unbreak cmake build after r82018

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82038 91177308-0d34-0410-b5e6-96231b3b80d8

Preserve ProfileInfo during CodeGenPrepare.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82034 91177308-0d34-0410-b5e6-96231b3b80d8

move FnStubs/GVSTubs/HiddenGVStub handling out of the X86 asmprinter
and use MachineModuleInfoMachO instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82022 91177308-0d34-0410-b5e6-96231b3b80d8

revert a hunk of r82018 that wasn't supposed to go in yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82020 91177308-0d34-0410-b5e6-96231b3b80d8

add a new MachineModuleInfoMachO class, which is the per-module
stuff common across all macho targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82018 91177308-0d34-0410-b5e6-96231b3b80d8

apparently russians are really hard to sort or something!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82016 91177308-0d34-0410-b5e6-96231b3b80d8

I can sort, no really.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82015 91177308-0d34-0410-b5e6-96231b3b80d8

make more clear since it is sorted by last name now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82014 91177308-0d34-0410-b5e6-96231b3b80d8

Doug is now the code owner for most of the Clang frontend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82013 91177308-0d34-0410-b5e6-96231b3b80d8

the pointer MMI keeps will start out with object-file format specific stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82012 91177308-0d34-0410-b5e6-96231b3b80d8

tidy up

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82011 91177308-0d34-0410-b5e6-96231b3b80d8

rearrange X86ATTAsmPrinter::doFinalization, making a scan of
the global variable list only happen for COFF targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82010 91177308-0d34-0410-b5e6-96231b3b80d8

Ted is christened as the owner of the clang static analyzer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82008 91177308-0d34-0410-b5e6-96231b3b80d8

remove the AsmPrinter::printMCInst hook hack now that
we have MCInstPrinter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82006 91177308-0d34-0410-b5e6-96231b3b80d8

fix cmake build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81999 91177308-0d34-0410-b5e6-96231b3b80d8

use an accessor to simplify code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81997 91177308-0d34-0410-b5e6-96231b3b80d8

Do not try and sink a load whose chain result has more than one use, when
trying to create RMW opportunities in the x86 backend.  This can cause a
cycle to appear in the graph, since the other uses may eventually feed into
the TokenFactor we are sinking the load below.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81996 91177308-0d34-0410-b5e6-96231b3b80d8

Added the ENTER instruction, which sets up a stack
frame, to the Intel instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81995 91177308-0d34-0410-b5e6-96231b3b80d8

Added the definitions for one-bit left shifts to
the Intel instruction tables.

The patterns will stay blank because ADD reg, reg
is faster, but having the encoding available is
useful for the disassembler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81994 91177308-0d34-0410-b5e6-96231b3b80d8

Don't sink gep operators through phi nodes if the result would require
more than one phi, since that leads to higher register pressure on
entry to the phi. This is especially problematic when the phi is in
a loop header, as it increases register pressure throughout the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81993 91177308-0d34-0410-b5e6-96231b3b80d8

Removed a few instructions that were already
covered by other definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81992 91177308-0d34-0410-b5e6-96231b3b80d8

Big change #1 for personality function references:
Eliminate the PersonalityPrefix/Suffix & NeedsIndirectEncoding
fields from MAI: they aren't part of the asm syntax, they are
related to the structure of the object file.

To replace their functionality, add a new
TLOF::getSymbolForDwarfGlobalReference method which asks targets
to decide how to reference a global from EH in a pc-relative way.

The default implementation just returns the symbol.  The default
darwin implementation references the symbol through an indirect
$non_lazy_ptr stub.  The bizarro x86-64 darwin specialization
handles the weird "foo@GOTPCREL+4" hack.

DwarfException.cpp now uses this to emit the reference to the
symbol in the right way, and this also eliminates another
horrible hack from DwarfException.cpp:

-    if (strcmp(MAI->getPersonalitySuffix(), "+4@GOTPCREL"))
-      O << "-" << MAI->getPCSymbol();

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81991 91177308-0d34-0410-b5e6-96231b3b80d8

lit: Add a custom test format for use in clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81987 91177308-0d34-0410-b5e6-96231b3b80d8

remove a dead variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81985 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper method for creating MCSymbol and MCSymbolRefExpr at
the same time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81984 91177308-0d34-0410-b5e6-96231b3b80d8

Added a variety of floating-point and SSE instructions.
All of these do not have patterns (they're for the
disassembler).

Many of the floating-point instructions will probably
be rolled into definitions that have patterns, and may
eventually be superseded by mdefs.  So I put them
together and left a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81979 91177308-0d34-0410-b5e6-96231b3b80d8

inline AsmPrinter::getCurrentFunctionEHName into its only caller.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81970 91177308-0d34-0410-b5e6-96231b3b80d8

Expand some more vector operations not supported by Neon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81969 91177308-0d34-0410-b5e6-96231b3b80d8

remove a dead bool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81968 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate AsmPrinter::EmitExternalGlobal, inlining its (now)
one implementation into its one caller.  This eliminates a totally
awesome and gratuitous hack where we casted a Function* to
GlobalVariable*.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81967 91177308-0d34-0410-b5e6-96231b3b80d8

Neon does not support vector divide or remainder.  Expand them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81966 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the PPC backend's implementation of EmitExternalGlobal
and use PersonalityPrefix/Suffix to achieve the same effect (like
the x86 backend).

This changes the code generated for ppc static mode, but guess what,
we were generating this before:

.byte 0x9B                                        ; Personality (indirect pcrel sdata4)
.long ___gxx_personality_v0-.                     ; Personality

which is not correct! (it is not an 'indirect' reference).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81965 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate the horrid AsmPrinter::getGlobalLinkName method, inlining
it into all of its call sites and simplifying them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81962 91177308-0d34-0410-b5e6-96231b3b80d8

simplify some code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81961 91177308-0d34-0410-b5e6-96231b3b80d8

Expand all v2f64 arithmetic operations for Neon.
Radar 7200803.  (This should also fix the
SingleSource/UnitTests/Vector/sumarray-dbl test.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81959 91177308-0d34-0410-b5e6-96231b3b80d8

Put back non-obsolete -f sections for 'opt'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81954 91177308-0d34-0410-b5e6-96231b3b80d8

Added far return instructions (that is, returns to
code in other segments) to the Intel instruction
tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81953 91177308-0d34-0410-b5e6-96231b3b80d8

remove some horrible MAI hooks which fortunately turn out to be always empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81946 91177308-0d34-0410-b5e6-96231b3b80d8

strength reduce a call to PrintRelDirective(true).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81942 91177308-0d34-0410-b5e6-96231b3b80d8

add hooks to hang target-specific goop off MachineModuleInfo,
move MachineFunctionInfo virtual method out of line to give it
a home.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81940 91177308-0d34-0410-b5e6-96231b3b80d8

Do not add the SVOffset to the Node CSE ID.  The same pointer argument cannot have different
SVOffsets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81937 91177308-0d34-0410-b5e6-96231b3b80d8

Expand on comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81928 91177308-0d34-0410-b5e6-96231b3b80d8

Updated comments per Eli's suggestion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81923 91177308-0d34-0410-b5e6-96231b3b80d8

Convert more tests to FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81915 91177308-0d34-0410-b5e6-96231b3b80d8

Added register-to-register ADD instructions to the
Intel tables, where the source operand is
specified by the R/M field and the destination
operand by the Reg field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81914 91177308-0d34-0410-b5e6-96231b3b80d8

Drop the raw_ostream required buffer size to 1.

 - As best I can tell, we have eliminated all the code which used to require a
   larger buffer size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81912 91177308-0d34-0410-b5e6-96231b3b80d8

Remove references to obsolete -f option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81911 91177308-0d34-0410-b5e6-96231b3b80d8

Update llc/opt PODs to clarify they support .ll input.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81910 91177308-0d34-0410-b5e6-96231b3b80d8

Fix -Asserts warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81909 91177308-0d34-0410-b5e6-96231b3b80d8

lit: When finding nested test suites, check first in the execpath in case there
is a site configuration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81902 91177308-0d34-0410-b5e6-96231b3b80d8

Better solution for tracking both the original alignment of the access, and the current alignment based
on the source value offset.  This avoids increasing the size of mem nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81897 91177308-0d34-0410-b5e6-96231b3b80d8

Correct comment pasto

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81896 91177308-0d34-0410-b5e6-96231b3b80d8

Added a new register class for segment registers
to the Intel register table.
Added 16- and 64-bit MOVs to and from the segment
registers to the Intel instruction tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81895 91177308-0d34-0410-b5e6-96231b3b80d8

Change the marker byte for stubs from 0xcd to 0xce (another form of
interrupt instruction, which shouldn't arise any other way).  0xcd is
also used by JITMemoryManager to initialize the buffer to garbage,
which means it could appear following a noreturn call even when
that is not a stub, confusing X86CompilationCallback2.  PR 4929.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81888 91177308-0d34-0410-b5e6-96231b3b80d8

fix PR4984 by ensuring that fastisel adds properly sign extended GEP displacement
values to machineinstrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81886 91177308-0d34-0410-b5e6-96231b3b80d8

rename test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81884 91177308-0d34-0410-b5e6-96231b3b80d8

convert to filecheck

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81882 91177308-0d34-0410-b5e6-96231b3b80d8

add missing file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81881 91177308-0d34-0410-b5e6-96231b3b80d8

Handle AddrMode4 for Thumb2 in rewriteT2FrameIndex.  This occurs for
VLDM/VSTM instructions, and without this check, the code assumes that an
offset is allowed, as it would be with VLDR/VSTR.  The asm printer,
however, silently drops the offset, producing incorrect code.  Since the
address register in this case is either the stack or frame pointer, the
spill location ends up conflicting with some other stack slot or with
outgoing arguments on the stack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81879 91177308-0d34-0410-b5e6-96231b3b80d8

Fix superreg use in ARMAsmPrinter. Approved by Anton Korobeynikov.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81878 91177308-0d34-0410-b5e6-96231b3b80d8

several major improvements to the sparc backend: support for weak linkage
and PIC codegen.  Patch by Venkatraman Govindaraju!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81877 91177308-0d34-0410-b5e6-96231b3b80d8

Teach ValueTracking how to look through GlobalAliases. GlobalAliases are
not folded in the constant folder because the constant folder doesn't
simplify ConstantExpr operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81864 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an accidental inversion of the inbounds flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81862 91177308-0d34-0410-b5e6-96231b3b80d8

When a constant's type is refined, update the constant in place
instead of cloning and RAUWing it.

 - Make AbstractTypeUser a friend of Value so that it can offer
   its subclasses a way to update a Value's type in place. This
   is better than a universally visible setType method on Value,
   and it's sufficient for the immediate need.

 - Eliminate the constant "convert" functions. This eliminates a
   lot of logic duplication, and fixes a complicated bug where a
   constant can't actually be cloned during the type refinement
   process because some of the types that its folder needs are
   half-destroyed, being in the middle of refinement themselves.

 - Move the getValType functions from being static overloaded
   functions in Constants.cpp to be members of class template
   specializations in ConstantsContext.h. This means that the
   code ends up getting instantiated twice, however it also
   makes it possible to eliminate all "convert" functions, so
   it's not a big net code size increase. And if desired, the
   duplicate instantiations could be eliminated with some
   reorganization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81861 91177308-0d34-0410-b5e6-96231b3b80d8

Use llvm-link -S instead of using llvm-dis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81860 91177308-0d34-0410-b5e6-96231b3b80d8

Give llvm-link a -S option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@81859 91177308-0d34-0410-b5e6-96231b3b80d8