Make TargetInstrInfo::copyRegToReg return a bool indicating whether the copy requested
was inserted or not.  This allows bitcast in fast isel to properly handle the case
where an appropriate reg-to-reg copy is not available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55375 91177308-0d34-0410-b5e6-96231b3b80d8

If IV is used in a int-to-float cast inside the loop then try to eliminate the cast operation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55374 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for fast isel of non-constant fptosi instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55373 91177308-0d34-0410-b5e6-96231b3b80d8

fix doxygen comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55372 91177308-0d34-0410-b5e6-96231b3b80d8

Grammar fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55371 91177308-0d34-0410-b5e6-96231b3b80d8

Added "str()" method to raw_string_ostream.
str() mirrors the same method in std::ostringstream.
str() flushes the buffered stream contents to string and returns a reference to the string.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55370 91177308-0d34-0410-b5e6-96231b3b80d8

code simplification, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55363 91177308-0d34-0410-b5e6-96231b3b80d8

stabilize more printing, this doesn't cause a problem
in the example attached to PR2590, but is a problem in general.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55361 91177308-0d34-0410-b5e6-96231b3b80d8

stablize SubRegsSet printing, part of PR2590

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55360 91177308-0d34-0410-b5e6-96231b3b80d8

Stabilize 'getDwarfRegNumFull' output to not depend on random memory
orders, part of PR2590

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55359 91177308-0d34-0410-b5e6-96231b3b80d8

If an xmm register is referenced explicitly in an inline asm, make sure to
assign it to a version of the xmm register with the regclass that matches its
type.  This fixes PR2715, a bug handling some crazy xpcom case in mozilla.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55358 91177308-0d34-0410-b5e6-96231b3b80d8

typo fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55355 91177308-0d34-0410-b5e6-96231b3b80d8

Actually recycle SDNode allocations. SelectionDAG is using
RecyclingAllocator, but this change is needed for the nodes
to actually be recycled. This cuts SelectionDAG's memory
usage high-water-mark in half in some cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55351 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid a warning about isTargetNullPtr being unused in release builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55350 91177308-0d34-0410-b5e6-96231b3b80d8

Throw the switch to allow FastISel to emit instructions whose return types different from their inputs.  Next step: adding lowering pattens in FastISel that actually use these newly available opcodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55349 91177308-0d34-0410-b5e6-96231b3b80d8

This is done.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55348 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance TableGen to emit code for FastISel of opcodes with variadic return types without slowing down opcodes that are not variadic.  No such opcodes are currently generated, but in theory it should be a matter of just hitting the switch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55347 91177308-0d34-0410-b5e6-96231b3b80d8

Add a RetVT parameter to emitted FastISel methods, so that we will be able to pass the desired return
type down.  This is not currently used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55345 91177308-0d34-0410-b5e6-96231b3b80d8

Deepen the map structure tablegen uses to compute FastISel patterns, in preparation for having patterns
with return types that differ from their input types.  This is not yet used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55344 91177308-0d34-0410-b5e6-96231b3b80d8

Implement 32 & 64 bit versions of PPC atomic
binary primitives.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55343 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55342 91177308-0d34-0410-b5e6-96231b3b80d8

80 col. violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55341 91177308-0d34-0410-b5e6-96231b3b80d8

Expand bitcast support in fast isel to support bitcasts of non-constant values by emitting reg-reg copies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55340 91177308-0d34-0410-b5e6-96231b3b80d8

Try approach to moving call address load inside of callseq_start. Now it's done during the preprocess of x86 isel. callseq_start's chain is changed to load's chain node; while load's chain is the last of callseq_start or the loads or copytoreg nodes inserted to move arguments to the right spot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55338 91177308-0d34-0410-b5e6-96231b3b80d8

Remove PPC-specific lowering for atomics; the
generic stuff works fine.

Mark rewritten cmp-and-swap as not using CR1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55336 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for fast isel of (integer) immediate materialization pattens, and use them to support
bitcast of constants in fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55325 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust grep's for new code sequence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55320 91177308-0d34-0410-b5e6-96231b3b80d8

It's important for the cmp-and-swap to balance
loads and stores but it's even more important for
it to store the right value.:(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55319 91177308-0d34-0410-b5e6-96231b3b80d8

Nevermind. This broke the bootstrap (?!).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55318 91177308-0d34-0410-b5e6-96231b3b80d8

MOVQ2DQ and MOVQ2DQ use SSE2. We should conditionalize the use of these
instructions on having SSE2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55317 91177308-0d34-0410-b5e6-96231b3b80d8

Fix pasto (values -> types) noticed by Josh Conner

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55315 91177308-0d34-0410-b5e6-96231b3b80d8

flush stream after dumping.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55301 91177308-0d34-0410-b5e6-96231b3b80d8

Fix asm printing of MOVSDto64mr and MOV64toSDrm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55300 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this test. Don't null out the file, just XFAIL it until patch can be fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55296 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily reverting r55292. It's causing a bootstraping failure:

/Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm-gcc.obj/./gcc/xgcc ... src/libiberty/make-temp-file.c -o make-temp-file.o
Assertion failed: (Node2Index[SU->NodeNum] > Node2Index[I->Dep->NodeNum] && "Wrong topological sorting"), function InitDAGTopologicalSorting, file /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.src/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp, line 508.
../../../../llvm-gcc.src/libiberty/hashtab.c:955: internal compiler error: Abort trap
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://developer.apple.com/bugreporter> for instructions.
make[4]: *** [hashtab.o] Error 1
make[4]: *** Waiting for unfinished jobs....
make[3]: *** [multi-do] Error 1
make[2]: *** [all] Error 2
make[1]: *** [all-target-libiberty] Error 2
make: *** [all] Error 2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55295 91177308-0d34-0410-b5e6-96231b3b80d8

remove the type checking logic already done by tblgen, just keep the
parts tblgen doesn't do (checking for constant, alloca, enclosing
function has gc marker).  This passes dj! :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55294 91177308-0d34-0410-b5e6-96231b3b80d8

simplify PseudoSourceValue printing a bit.  Unnest all of PseudoSourceValue.cpp from the llvm namespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55293 91177308-0d34-0410-b5e6-96231b3b80d8

Move callseq_start above the call address load to allow load to be folded into the call node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55292 91177308-0d34-0410-b5e6-96231b3b80d8

Print PseudoSourceValue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55291 91177308-0d34-0410-b5e6-96231b3b80d8

we should use 'class' for non-pod types instead of 'struct' to
make it easier to deal with the VC++ struct/class bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55290 91177308-0d34-0410-b5e6-96231b3b80d8

add a hack to temporarily allow pseudo source  values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55289 91177308-0d34-0410-b5e6-96231b3b80d8

make sure to flush the stream after dumping, to make sure it goes out immediately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55288 91177308-0d34-0410-b5e6-96231b3b80d8

Use additionnal include directory instead of ../ in #include.
Suggested by aKor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55282 91177308-0d34-0410-b5e6-96231b3b80d8

Updating VC++ project.
Adding one include file and correct one declaration from class to struct in order to make llvm compile on VC2005.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55279 91177308-0d34-0410-b5e6-96231b3b80d8

Update GC docs for clarified naming and AsmWriter refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55275 91177308-0d34-0410-b5e6-96231b3b80d8

improve encapsulation of the BBExecutable set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55271 91177308-0d34-0410-b5e6-96231b3b80d8

Switch an assortment of maps, sets and vectors to more efficient versions,
patch contributed by m-s!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55270 91177308-0d34-0410-b5e6-96231b3b80d8

get MachineConstantPool off std::ostream, onto raw_ostream.  It would be
really nice if someone converted MachineFunction::print to raw_ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55268 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55267 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55266 91177308-0d34-0410-b5e6-96231b3b80d8

Add raw_stream adaptors that write into an std::string and SmallVector/SmallString.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55265 91177308-0d34-0410-b5e6-96231b3b80d8

Switch the asmprinter (.ll) and all the stuff it requires over to
use raw_ostream instead of std::ostream.  Among other goodness,
this speeds up llvm-dis of kc++ with a release build from 0.85s
to 0.49s (88% faster).

Other interesting changes:
 1) This makes Value::print be non-virtual.
 2) AP[S]Int and ConstantRange can no longer print to ostream directly,
    use raw_ostream instead.
 3) This fixes a bug in raw_os_ostream where it didn't flush itself
    when destroyed.
 4) This adds a new SDNode::print method, instead of only allowing "dump".

A lot of APIs have both std::ostream and raw_ostream versions, it would
be useful to go through and systematically anihilate the std::ostream
versions.

This passes dejagnu, but there may be minor fallout, plz let me know if
so and I'll fix it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55263 91177308-0d34-0410-b5e6-96231b3b80d8

use proper namespace qualifications

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55259 91177308-0d34-0410-b5e6-96231b3b80d8

Add #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55258 91177308-0d34-0410-b5e6-96231b3b80d8

add #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55257 91177308-0d34-0410-b5e6-96231b3b80d8

add #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55252 91177308-0d34-0410-b5e6-96231b3b80d8

Split format() out to its own file, add support for formatting
up to three values in one call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55251 91177308-0d34-0410-b5e6-96231b3b80d8

conditionalize this to work on windows, untested but it should work
according to the intarnetz

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55248 91177308-0d34-0410-b5e6-96231b3b80d8

add a simple mechanism for formatted output.  This gives raw_ostream's
all the power and risk of fprintf format strings.  Use them like this:

  OS << format("%10.4f", 42.0) << "\n" << format("%x", 42) << '\n';

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55246 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for 64bit maskmovq

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55239 91177308-0d34-0410-b5e6-96231b3b80d8

Provide a 64 bit variant of mmx.maskmovq intrinsic lowering.
Is there way to avoid explicit target check?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55238 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting r55227. This was causing the following failures in the regression
tests:

Running /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/dg.exp ...
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcread-ptrptr.ll for PR1633
Failed with exit(1) at line 1
while running: not llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcread-ptrptr.ll >& /dev/null
child process exited abnormally
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-alloca.ll for PR1633
Failed with exit(1) at line 1
while running: not llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-alloca.ll >& /dev/null
child process exited abnormally
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-meta.ll for PR1633
Failed with exit(1) at line 1
while running: not llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-meta.ll >& /dev/null
child process exited abnormally
FAIL: ndbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-ptrptr.ll for PR1633
Failed with exit(1) at line 1
while running: not llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcroot-ptrptr.ll >& /dev/null
child process exited abnormally
FAIL: /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcwrite-ptrptr.ll for PR1633
Failed with exit(1) at line 1
while running: not llvm-as < /Volumes/Sandbox/Buildbot/llvm/full-llvm/build/llvm.HEAD.src/test/Verifier/gcwrite-ptrptr.ll >& /dev/null
child process exited abnormally

===  Summary ===

# of expected passes 3021
# of unexpected failures 6
# of expected failures 16
make[1]: *** [check-local] Error 1
make: *** [check] Error 2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55233 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new trivial -inst-namer pass which makes it possible to diff the
before/after effects of a pass, crazy!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55230 91177308-0d34-0410-b5e6-96231b3b80d8

Verify that the alignment argument to llvm.memcpy is a constant
integer, PR2318.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55228 91177308-0d34-0410-b5e6-96231b3b80d8

remove verification for gc intrinsic argument types.  it was indented wrong
and tblgen now knows intrinsic prototypes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55227 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2423 by checking all indices for out of range access, not only
indices that start with an array subscript.  x->field[10000] is just
as bad as (*X)[14][10000].

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55226 91177308-0d34-0410-b5e6-96231b3b80d8

the bug was apparently fixed long ago, reenable the assertion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55224 91177308-0d34-0410-b5e6-96231b3b80d8

Make MBBMap a DenseMap instead of a std::map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55220 91177308-0d34-0410-b5e6-96231b3b80d8

Move the point at which FastISel taps into the SelectionDAGISel
process up to a higher level. This allows FastISel to leverage
more of SelectionDAGISel's infastructure, such as updating Machine
PHI nodes.

Also, implement transitioning from SDISel back to FastISel in
the middle of a block, so it's now possible to go back and
forth. This allows FastISel to hand individual CallInsts and other
complicated things off to SDISel to handle, while handling the rest
of the block itself.

To help support this, reorganize the SelectionDAG class so that it
is allocated once and reused throughout a function, instead of
being completely reallocated for each block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55219 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid creating shift-by-zero SDNodes in the common case of
i8* getelementptr. DAGCombine eliminates these, but this is
a fairly common case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55214 91177308-0d34-0410-b5e6-96231b3b80d8

Move SelectionDAG's constructor out of line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55212 91177308-0d34-0410-b5e6-96231b3b80d8

Add a clear() method to FoldingSet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55210 91177308-0d34-0410-b5e6-96231b3b80d8

Test all currently supported atomic builtins on x86-{32,64}.
These just test that they go through the BE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55208 91177308-0d34-0410-b5e6-96231b3b80d8

Comprehensive test of FE handling of __sync builtins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55206 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply r55191 and r55192.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55205 91177308-0d34-0410-b5e6-96231b3b80d8

Make option variables static, so they won't cause nameclash

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55203 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting r55190, r55191, and r55192. They broke the build with this error message:

{standard input}:17:bad register name `%sil'
make[4]: *** [libgcc/./_addvsi3.o] Error 1
make[4]: *** Waiting for unfinished jobs....
{standard input}:23:bad register name `%dil'
{standard input}:28:bad register name `%dil'
make[4]: *** [libgcc/./_addvdi3.o] Error 1
{standard input}:18:bad register name `%sil'
make[4]: *** [libgcc/./_subvsi3.o] Error 1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55200 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the InsertBranch call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55192 91177308-0d34-0410-b5e6-96231b3b80d8

Support non-fallthrough unconditional branches in FastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55191 91177308-0d34-0410-b5e6-96231b3b80d8

Anyext tweaks for x86. When extloading a value to i32 or i64, choose
instructions that define the full 32 or 64-bit value. When anyexting
from i8 to i16 or i32, it's not necessary to zero out the high
portion of the register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55190 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55189 91177308-0d34-0410-b5e6-96231b3b80d8

Add FastISel support for PHINodes. Machine PHI nodes
are not yet updated properly, but that's a separate
task.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55187 91177308-0d34-0410-b5e6-96231b3b80d8

Implement __sync_synchronize on ppc32.  Patch by Gary Benson.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55186 91177308-0d34-0410-b5e6-96231b3b80d8

Add an empty() member to FoldingSet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55182 91177308-0d34-0410-b5e6-96231b3b80d8

Fix SmallVector's size calculation so that a size of 0 is
handled correctly, and change a few SmallVector uses to use
size 0 to more clearly reflect their intent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55181 91177308-0d34-0410-b5e6-96231b3b80d8

improve support for systems that need unistd.h to get STDOUT_FILENO.
Patch contributed by Bjorn Reese!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55179 91177308-0d34-0410-b5e6-96231b3b80d8

Also overload for char, since the "char" type depends on the host.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55173 91177308-0d34-0410-b5e6-96231b3b80d8

Overload for both signed and unsigned char.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55171 91177308-0d34-0410-b5e6-96231b3b80d8

consolidate DenseMapInfo implementations, and add one for std::pair.
Patch contributed by m-s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55167 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite ppc code generated for __sync_{bool|val}_compare_and_swap
so that lwarx and stwcx are always executed the same number of times.
This is important for performance, I'm told.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55163 91177308-0d34-0410-b5e6-96231b3b80d8

Add a few comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55157 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out the predicate check code from DAGISelEmitter.cpp
and use it in FastISelEmitter.cpp, and make FastISel
subtarget aware. Among other things, this lets it work
properly on x86 targets that don't have SSE, where it
successfully selects x87 instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55156 91177308-0d34-0410-b5e6-96231b3b80d8

Remove tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55154 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR2585.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55151 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily reverting r55137. This was causing the bootstrap to go into an
infinite loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55149 91177308-0d34-0410-b5e6-96231b3b80d8

Fix write() when the string being written is larger than the buffer.  This broke various ObjC testcases
with very long symbol names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55148 91177308-0d34-0410-b5e6-96231b3b80d8

If part of the mask is "undef", then ignore it as we don't care what goes into it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55147 91177308-0d34-0410-b5e6-96231b3b80d8

Fix whitespace. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55146 91177308-0d34-0410-b5e6-96231b3b80d8

Add -mattr=sse2 so this test doesn't fail on non-x86 hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55145 91177308-0d34-0410-b5e6-96231b3b80d8

Make x86 and sse2 explicit for non-x86 hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55141 91177308-0d34-0410-b5e6-96231b3b80d8