Factor GC metadata table assembly generation out of Collector in preparation for splitting AsmPrinter into its own library.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54880 91177308-0d34-0410-b5e6-96231b3b80d8

Get raw_ostream.cpp to compile on MSVC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54879 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 54821. It's miscompiling 252.eon and 447.dealII

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54878 91177308-0d34-0410-b5e6-96231b3b80d8

I found a better place for this optz'n.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54877 91177308-0d34-0410-b5e6-96231b3b80d8

Xor'ing both sides of icmp by sign-bit is equivalent to swapping signedness of
the predicate.

Also, make this optz'n apply in more cases where it's safe to do so.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54876 91177308-0d34-0410-b5e6-96231b3b80d8

avoid an extraneous std::string construction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54875 91177308-0d34-0410-b5e6-96231b3b80d8

Rework the routines that convert AP[S]Int into a string.  Now, instead of
returning an std::string by value, it fills in a SmallString/SmallVector
passed in.  This significantly reduces string thrashing in some cases.

More specifically, this:
 - Adds an operator<< and a print method for APInt that allows you to
   directly send them to an ostream.
 - Reimplements APInt::toString to be much simpler and more efficient
   algorithmically in addition to not thrashing strings quite as much.

This speeds up llvm-dis on kc++ by 7%, and may also slightly speed up the
asmprinter.  This also fixes a bug I introduced into the asmwriter in a
previous patch w.r.t. alias printing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54873 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build on GCC 4.3.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54870 91177308-0d34-0410-b5e6-96231b3b80d8

remove a dead APInt ctor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54869 91177308-0d34-0410-b5e6-96231b3b80d8

avoid string thrashing when formatting names in output.  This
speeds up release-asserts llvm-dis on kc++ from 1.86s to 1.04s (~79%)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54868 91177308-0d34-0410-b5e6-96231b3b80d8

random cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54866 91177308-0d34-0410-b5e6-96231b3b80d8

add support for a cout/cerr analog (outs()/errs()) as well as
a simple adaptor class to give raw output capabilities to
something that wants to write to an ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54865 91177308-0d34-0410-b5e6-96231b3b80d8

opening "-" automatically yields stdout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54863 91177308-0d34-0410-b5e6-96231b3b80d8

rename OutputData to 'write' to match ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54857 91177308-0d34-0410-b5e6-96231b3b80d8

remove accidental comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54856 91177308-0d34-0410-b5e6-96231b3b80d8

add a new raw_ostream class which is an extremely high performance ostream that
can *only* output data (no seeking, reading, etc).  This is adapted from the
clang "-E outputter", and is roughly 10% faster than stdio on darwin and 30%
(or more) faster than std::ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54855 91177308-0d34-0410-b5e6-96231b3b80d8

PPC/Linux normally uses named section for bss

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54847 91177308-0d34-0410-b5e6-96231b3b80d8

Use proper strings section name for PPC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54846 91177308-0d34-0410-b5e6-96231b3b80d8

Use correct name for TLS address resolution routine on x86-64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54845 91177308-0d34-0410-b5e6-96231b3b80d8

Add interface for section override. Use this for Sparc, since it should use named BSS section.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54844 91177308-0d34-0410-b5e6-96231b3b80d8

Move SLEB/ULEB size calculation routines from AsmPrinter to TargetAsmInfo. This makes JIT asmprinter-free.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54843 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce heap trashing due to std::string construction / concatenation via caching of section flags string representations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54842 91177308-0d34-0410-b5e6-96231b3b80d8

Remove GCSE and LoadVN from the testsuite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54832 91177308-0d34-0410-b5e6-96231b3b80d8

Remove GCSE, ValueNumbering, and LoadValueNumbering.  These have been deprecated for almost a year; it's finally time for them to go away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54822 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply 54786. Add overflow and number of mantissa bits checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54821 91177308-0d34-0410-b5e6-96231b3b80d8

Convert several std::vectors over to SmallVector, and use reserve() as appropriate for cases where std::vector is still used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54820 91177308-0d34-0410-b5e6-96231b3b80d8

Add a value_type typedef to SmallVector, to make it more compatible with STL adapters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54819 91177308-0d34-0410-b5e6-96231b3b80d8

Inline the fastpath of PATypeHolder::get().  This is a small speedup in
instcombine among other things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54814 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 54786. It's not checking for overflows, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54813 91177308-0d34-0410-b5e6-96231b3b80d8

use smallvector instead of vector for a couple worklists.  This speeds up instcombine
by ~10% on some testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54811 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallSet instead of std::set to save allocations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54810 91177308-0d34-0410-b5e6-96231b3b80d8

Renaming LLVMC/dg.exp to LLVM/llvmc.exp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54809 91177308-0d34-0410-b5e6-96231b3b80d8

Move MachineInstr::getOpcode inline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54807 91177308-0d34-0410-b5e6-96231b3b80d8

Build the X86GenFastISel.inc file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54806 91177308-0d34-0410-b5e6-96231b3b80d8

The pass manager is not able to schedule -loop-deletion -loop-index-split.

The loop-deletion pass does not preserve dom frontier, which is required by
loop-index-split. When the PM checks dom frontier for loop-index-split, it has
already verified that lcssa is availalble. However, new dom frontier forces new
loop pass manager, which does not  have lcssa yet.

The PM should recheck availability of required analysis passes in such cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54805 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily revert r54792. It's causing an ICE during bootstrapping.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54804 91177308-0d34-0410-b5e6-96231b3b80d8

Also avoid pinsrw and pinsrb with a variable insertelement index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54803 91177308-0d34-0410-b5e6-96231b3b80d8

Convert uses of std::vector in TargetInstrInfo to SmallVector.  This change had to be propoagated down into all the targets and up into all clients of this API.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54802 91177308-0d34-0410-b5e6-96231b3b80d8

Don't try to use the insertps instruction for vector
element inserts with non-constant indices. This fixes
CodeGen/X86/vector-variable-idx.ll on machines that
have SSE4.1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54801 91177308-0d34-0410-b5e6-96231b3b80d8

Make FastISel's constructor protected, and give it a destructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54793 91177308-0d34-0410-b5e6-96231b3b80d8

Use DenseMap. Patch by Pratik Solanki.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54792 91177308-0d34-0410-b5e6-96231b3b80d8

Removed redundant ctor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54789 91177308-0d34-0410-b5e6-96231b3b80d8

Remove more uses of std::set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54787 91177308-0d34-0410-b5e6-96231b3b80d8

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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54786 91177308-0d34-0410-b5e6-96231b3b80d8

Have LeakDetector use a SmallPtrSet instead of an std::set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54785 91177308-0d34-0410-b5e6-96231b3b80d8

Improve support for vector casts in LLVM IR and CodeGen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54784 91177308-0d34-0410-b5e6-96231b3b80d8

Speed up addRegisterDead by adding more fast checks before performing the expensive
subregister query, and by increasing the size of the subregister hashtable so
that there are fewer collisions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54781 91177308-0d34-0410-b5e6-96231b3b80d8

Use empty() instead of begin() == end().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54780 91177308-0d34-0410-b5e6-96231b3b80d8

Replace two for loops with while(!X->use_empty()) loops. This prevents
invalidating the iterator by deleting the current use. This fixes a segfault on
64 bit linux reported in PR2675.

Also remove an unneeded if.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54778 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of a use of std::map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54770 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bogus srem rule - a negative value srem'd by a power-of-2
can have a non-negative result; for example, -16%16 is 0. Also,
clarify the related comments. This fixes PR2670.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54767 91177308-0d34-0410-b5e6-96231b3b80d8

Expunge the last uses of std::map from LiveIntervals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54766 91177308-0d34-0410-b5e6-96231b3b80d8

Move r2iMap_ over to DenseMap from std::map.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54765 91177308-0d34-0410-b5e6-96231b3b80d8

Allow SelectionDAG to create EXTRACT_VECTOR_ELT nodes with
non-constant indices. Only a few of the peephole checks require
a constant index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54764 91177308-0d34-0410-b5e6-96231b3b80d8

Make the allocation of LiveIntervals explicit, rather than holding them in the r2iMap_ by value.  This will prevent references to them from being invalidated
if the map is changed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54763 91177308-0d34-0410-b5e6-96231b3b80d8

Switch this from std::map to DenseMap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54761 91177308-0d34-0410-b5e6-96231b3b80d8

Fix SCCP's handling of struct value loads and stores. SCCP doesn't
track individual leaf values in such cases, so it needs to treat
struct values as normal values in this case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54760 91177308-0d34-0410-b5e6-96231b3b80d8

Add svn:ignore on several Release-Asserts directories

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54758 91177308-0d34-0410-b5e6-96231b3b80d8

Add default constructor to APSInt
 - Creates uninitialized APInt.
 - Prevents need for embedding arbitrary constants when used as an out
   parameter, for example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54757 91177308-0d34-0410-b5e6-96231b3b80d8

Update makellvm to return correct result code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54756 91177308-0d34-0410-b5e6-96231b3b80d8

Rename. s/FindIVForUser/FindIVUserForCond/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54754 91177308-0d34-0410-b5e6-96231b3b80d8

Teach constant folding that an inttoptr of a
ptrtoint can be turned into a bitcast if the
integer is at least as wide as a pointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54752 91177308-0d34-0410-b5e6-96231b3b80d8

Initial checkin of the new "fast" instruction selection support. See
the comments in FastISelEmitter.cpp for details on what this is.
This is currently experimental and unusable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54751 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, check in these files too, for the FastISel -> Fast rename.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54750 91177308-0d34-0410-b5e6-96231b3b80d8

Rename SelectionDAGISel's FastISel to Fast, to begin to make
room for the new FastISel instruction selection code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54749 91177308-0d34-0410-b5e6-96231b3b80d8

Generated files for 54744.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54745 91177308-0d34-0410-b5e6-96231b3b80d8

Add read/write support for X86's sseregparm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54744 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54742 91177308-0d34-0410-b5e6-96231b3b80d8

1) Merge entire live intervals instead of parts of them.
2) Conditionalize temporary insertion if we don't need it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54741 91177308-0d34-0410-b5e6-96231b3b80d8

Rename this, in case people think that NullFolder
has something to do with folding null values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54725 91177308-0d34-0410-b5e6-96231b3b80d8

Removed SELECT_CC custom lowering. This is not needed anymore, the SELECT node
is lowered properly and covers everything LowerSELECT_CC did.
Added method printUnsignedImm in AsmPrinter to print uimm16 operands. This
avoid the ugly instruction by instruction checking in printOperand.
Added a swap instruction present in the allegrex core.
Added two conditional instructions present in the allegrex core : MOVZ and MOVN.
They both allow a more efficient SELECT operation for integers.
Also added SELECT patterns to optimize MOVZ and MOVN usage.
The brcond and setcc patterns were cleaned: redundant and suboptimal patterns
were
removed. The suboptimals were replaced by more efficient ones.
Fixed some instructions that were using immZExt16 instead of immSExt16.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54724 91177308-0d34-0410-b5e6-96231b3b80d8

Check sign to detect overflow before changing compare stride.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54710 91177308-0d34-0410-b5e6-96231b3b80d8

When resolving a stub in x86-64 JIT, use a PC-relative branch
rather than the absolute address if the target is within range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54708 91177308-0d34-0410-b5e6-96231b3b80d8

Remove tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54707 91177308-0d34-0410-b5e6-96231b3b80d8

Update. Remove bogus webpage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54705 91177308-0d34-0410-b5e6-96231b3b80d8

Make x86-64 JIT changes Darwin-specific.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54700 91177308-0d34-0410-b5e6-96231b3b80d8

Add a NullFolder class that doesn't fold constants.
This may be used as the second IRBuilder template
parameter, the idea being that people learning LLVM
may find it helpful (several people asked on IRC
if it was possible to turn off constant folding
because it made it hard for them to see what was
going on).  Compiles, but otherwise completely
untested.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54698 91177308-0d34-0410-b5e6-96231b3b80d8

Extend ScalarEvolution's executesAtLeastOnce logic to be able to
continue past the first conditional branch when looking for a
relevant test. This helps it avoid using MAX expressions in
loop trip counts in more cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54697 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace cleanup. Test commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54695 91177308-0d34-0410-b5e6-96231b3b80d8

In the absence of a linker to build the GOT, use the 32-bit
non_lazy_ptr mechanism on x86-64 Darwin JIT.  Fixes a bunch
of last night's failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54692 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the filename in the top-of-file comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54688 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid repeatedly reallocating the FoldingSetNodeID when searching
through multiple nodes in a bucket.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54687 91177308-0d34-0410-b5e6-96231b3b80d8

Use SmallVector instead of std::vector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54685 91177308-0d34-0410-b5e6-96231b3b80d8

Point people to ConstantExpr and ConstantFolding,
in case they get the wrong idea.  Fit in 80 columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54676 91177308-0d34-0410-b5e6-96231b3b80d8

Use DenseMap to keep track of last users.
Use inversed map for faster queries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54662 91177308-0d34-0410-b5e6-96231b3b80d8

Some fixes for x86-64 JIT.  Make it use small code
model, except for external calls; this makes
addressing modes PC-relative.  Incomplete.

The assertion at the top of Emitter::runOnMachineFunction
was obviously bogus (always true) so I removed it.
If someone knows what the correct test should be to cover
all the various targets, please fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54656 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper method to sys::Path for clang, patch by
Kovarththanan Rajaratnam!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54655 91177308-0d34-0410-b5e6-96231b3b80d8

Implement support for simplifying vector comparisons by 0.0 and 1.0 like we
do for scalars.  Patch contributed by Nicolas Capens

This also generalizes the previous xforms to work on long double, now that
isExactlyValue works for long double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54653 91177308-0d34-0410-b5e6-96231b3b80d8

Keep track of analysis usage information for passes. Avoid invoking
getAnalysisUsage() repeatedly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54650 91177308-0d34-0410-b5e6-96231b3b80d8

Improve the grep commands for this test to be tolerant of ABI
differences, and to be more specific.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54648 91177308-0d34-0410-b5e6-96231b3b80d8

Take the FrameOffset into account when computing the alignment
of stack objects. This fixes PR2656.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54646 91177308-0d34-0410-b5e6-96231b3b80d8

Implement ISD::TRAP support on PPC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54644 91177308-0d34-0410-b5e6-96231b3b80d8

Make it possible to use different constant
folding policies with IRBuilder.  The default,
provided by ConstantFolder, is to do minimal
folding like now: what ConstantExpr provides.
An alternative is to use TargetFolder, which
uses target information to fold constants more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54640 91177308-0d34-0410-b5e6-96231b3b80d8

the stacker doc is way out of date.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54631 91177308-0d34-0410-b5e6-96231b3b80d8

remove obsolete files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54630 91177308-0d34-0410-b5e6-96231b3b80d8

move some more stuff out of my email into readme.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54603 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54602 91177308-0d34-0410-b5e6-96231b3b80d8

"This patch adds a virtual call to AbstractLatticeFunction to derive a
type lattice value for an Argument*, giving clients the opportunity to
use something other than Top for it if they choose to."

Patch by John McCall!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54589 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak the llvm-gcc build: recent changes to
IRBuilder are not expecting null names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54581 91177308-0d34-0410-b5e6-96231b3b80d8

update VS project.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54580 91177308-0d34-0410-b5e6-96231b3b80d8