In the special case, call the comparison function instead of
manually performing the comparison. This allows the special
case to work correctly even in the case where someone is
experimenting with a different comparison function :-).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49670 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade these tests for the current intrinsic prototypes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49669 91177308-0d34-0410-b5e6-96231b3b80d8

Remove -unwind-tables-optional everywhere, since
this is now the default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49667 91177308-0d34-0410-b5e6-96231b3b80d8

Fix const-correctness issues with the SrcValue handling in the
memory intrinsic expansion code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49666 91177308-0d34-0410-b5e6-96231b3b80d8

Reverse sense of unwind-tables option.  This means
stack tracebacks on Darwin x86-64 won't work by default;
nevertheless, everybody but me thinks this is a good idea.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49663 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49661 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r49614.  As Dan pointed out, some of these aren't correct.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49657 91177308-0d34-0410-b5e6-96231b3b80d8

Fix /test/CodeGen/PowerPC/big-endian-actual-args.ll for linux/ppc32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49652 91177308-0d34-0410-b5e6-96231b3b80d8

improve diagnostics in call parsing, patch suggested by
Matthijs Kooijman

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49648 91177308-0d34-0410-b5e6-96231b3b80d8

VAArg may trap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49646 91177308-0d34-0410-b5e6-96231b3b80d8

Initial libcall support for LegalizeTypes.  This is
much simpler than in LegalizeDAG because calls are
not yet expanded into call sequences: that happens
after type legalization has finished.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49634 91177308-0d34-0410-b5e6-96231b3b80d8

add a new CallGraphNode::removeCallEdgeFor method, tidy some comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49617 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for equality comparison of CallSite's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49616 91177308-0d34-0410-b5e6-96231b3b80d8

Replace calls of the form V1->setName(V2->getName()) with V1->takeName(V2),
which is significantly more efficient.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49614 91177308-0d34-0410-b5e6-96231b3b80d8

LegalizeTypes can sometimes have deleted nodes
in its maps.  Add some sanity checks that catch
this kind of thing.  Hopefully these can be
removed one day (once all problems are fixed!)
but for the moment it seems wise to have them in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49612 91177308-0d34-0410-b5e6-96231b3b80d8

Add a divided flag for the first piece of an argument divided into mulitple parts. Fixes PR1643

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49611 91177308-0d34-0410-b5e6-96231b3b80d8

The functionality being tested was removed because it was horribly unsafe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49610 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some serious logic errors that broke the jit on darwin/x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49606 91177308-0d34-0410-b5e6-96231b3b80d8

Merge LLVMBuilder and FoldingBuilder, calling
the result IRBuilder.  Patch by Dominic Hamon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49604 91177308-0d34-0410-b5e6-96231b3b80d8

Gabor points out that reserveOperandSpace takes # of values,
not # of operands as an input.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49599 91177308-0d34-0410-b5e6-96231b3b80d8

Provide option for stack alignment override

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49593 91177308-0d34-0410-b5e6-96231b3b80d8

This patch corrects the handling of byval arguments for tailcall
optimized x86-64 (and x86) calls so that they work (... at least for
my test cases).

Should fix the following problems:

Problem 1: When i introduced the optimized handling of arguments for
tail called functions (using a sequence of copyto/copyfrom virtual
registers instead of always lowering to top of the stack) i did not
handle byval arguments correctly e.g they did not work at all :).

Problem 2: On x86-64 after the arguments of the tail called function
are moved to their registers (which include ESI/RSI etc), tail call
optimization performs byval lowering which causes xSI,xDI, xCX
registers to be overwritten. This is handled in this patch by moving
the arguments to virtual registers first and after the byval lowering
the arguments are moved from those virtual registers back to
RSI/RDI/RCX.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49584 91177308-0d34-0410-b5e6-96231b3b80d8

Factor some libcall code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49583 91177308-0d34-0410-b5e6-96231b3b80d8

Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.

Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.

This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.

Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.

This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49572 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug that prevented x86-64 from using rep.movsq for
8-byte-aligned data.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49571 91177308-0d34-0410-b5e6-96231b3b80d8

80 col fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49569 91177308-0d34-0410-b5e6-96231b3b80d8

Restore code to disable crash catcher on older OS X systems

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49568 91177308-0d34-0410-b5e6-96231b3b80d8

Add debugging code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49566 91177308-0d34-0410-b5e6-96231b3b80d8

Reenable JIT symbol table.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49548 91177308-0d34-0410-b5e6-96231b3b80d8

Do not add empty live intervals to handled_. They should never be undone for backtracking.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49544 91177308-0d34-0410-b5e6-96231b3b80d8

If a PHI node has a single implicit_def source, replace it with an implicit_def instead of a copy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49543 91177308-0d34-0410-b5e6-96231b3b80d8

Use of implicit_def is not part of live interval. Create empty intervals for the uses when the live interval is being spilled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49542 91177308-0d34-0410-b5e6-96231b3b80d8

Change llvm tools install location.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49540 91177308-0d34-0410-b5e6-96231b3b80d8

improvements for IntrusiveRefCntPtr, patch by Mikhail Glushenkov

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49538 91177308-0d34-0410-b5e6-96231b3b80d8

The Ada front-end from the LLVM 2.2 release
cannot be build with GNAT GPL 2006, only with
GNAT GPL 2005.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49529 91177308-0d34-0410-b5e6-96231b3b80d8

detabify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49524 91177308-0d34-0410-b5e6-96231b3b80d8

Add testcase for PR2213.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49517 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2213 by simultaneously making GVN more aggressive with the return values
of calls and less aggressive with non-readnone calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49516 91177308-0d34-0410-b5e6-96231b3b80d8

New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49514 91177308-0d34-0410-b5e6-96231b3b80d8

Remove implicit_def instructions that become dead as result of coalescing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49513 91177308-0d34-0410-b5e6-96231b3b80d8

Allow registers defined by implicit_def to be clobbered.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49512 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49504 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49502 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR2190.  Memdep's non-local caching was checking dirtied blocks in the
wrong order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49499 91177308-0d34-0410-b5e6-96231b3b80d8

Make several symbols static.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49496 91177308-0d34-0410-b5e6-96231b3b80d8

Teach InstCombine's ComputeMaskedBits to handle pointer expressions
in addition to integer expressions. Rewrite GetOrEnforceKnownAlignment
as a ComputeMaskedBits problem, moving all of its special alignment
knowledge to ComputeMaskedBits as low-zero-bits knowledge.

Also, teach ComputeMaskedBits a few basic things about Mul and PHI
instructions.

This improves ComputeMaskedBits-based simplifications in a few cases,
but more noticeably it significantly improves instcombine's alignment
detection for loads, stores, and memory intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49492 91177308-0d34-0410-b5e6-96231b3b80d8

A copy instruction may use a register multiple times on some targets. Change them all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49491 91177308-0d34-0410-b5e6-96231b3b80d8

Add comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49469 91177308-0d34-0410-b5e6-96231b3b80d8

add a note, this is actually not too bad to implement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49466 91177308-0d34-0410-b5e6-96231b3b80d8

move the x86-32 part of PR2108 here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49465 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the x86-64 side of PR2108 by adding a v2f64 version of
MOVZQI2PQIrr.  This would be better handled as a dag combine
(with the goal of eliminating the bitconvert) but I don't know
how to do that safely.  Thoughts welcome.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49463 91177308-0d34-0410-b5e6-96231b3b80d8

produce an error on invalid input instead of asserting:

def : Pat<((v2f64 (vector_shuffle immAllZerosV_bc,
           ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49462 91177308-0d34-0410-b5e6-96231b3b80d8

Teach branch folding pass about implicit_def instructions. Unfortunately we can't just eliminate them since register scavenger expects every register use to be defined. However, we can delete them when there are no intra-block uses.  Carefully removing some implicit def's which enable more blocks to be optimized away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49461 91177308-0d34-0410-b5e6-96231b3b80d8

Disable an xform we've had for a long time, pow(x,0.5) -> sqrt.
This is not safe for all inputs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49458 91177308-0d34-0410-b5e6-96231b3b80d8

Move memcpy / memset optimization pass after GVN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49457 91177308-0d34-0410-b5e6-96231b3b80d8

- More aggressively coalescing away copies whose source is defined by an implicit_def.
- Added insert_subreg coalescing support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49448 91177308-0d34-0410-b5e6-96231b3b80d8

Make isVectorClearMaskLegal's operand list const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49446 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some minor errors in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49445 91177308-0d34-0410-b5e6-96231b3b80d8

Add const qualifiers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49443 91177308-0d34-0410-b5e6-96231b3b80d8

Update comments to use 2.0 syntax type names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49442 91177308-0d34-0410-b5e6-96231b3b80d8

Add XMM1 as a second return value register for f32 and f64 on x86-64. This
is needed for the x86-64-ABI handling of structs that contain floating-point
members that are returned by value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49441 91177308-0d34-0410-b5e6-96231b3b80d8

Add DX as a second return value register for i16 on x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49440 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize getUnaryFloatFunction to handle any FP unary function, automatically
figuring out the suffix to use.  implement pow(2,x) -> exp2(x).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49437 91177308-0d34-0410-b5e6-96231b3b80d8

remove capital letter from test name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49436 91177308-0d34-0410-b5e6-96231b3b80d8

use the new ConstantFP::get method to make this work with
long double and simplify the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49435 91177308-0d34-0410-b5e6-96231b3b80d8

ConstantFP::get should be static.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49434 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug where an incorrect bit mask would be generated if a target's last asm
string began at a power of 2 in the string index.  For example, if "ret" started
at position 16, the ret instruction would be assigned code 16, but the mask would be AsmChars[] + Code & 15, not Code & 31.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49433 91177308-0d34-0410-b5e6-96231b3b80d8

Be conservative if getresult operand is neither call nor invoke.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49430 91177308-0d34-0410-b5e6-96231b3b80d8

Factor a bunch of functionality related to memcpy and memset transforms out of
GVN and into its own pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49419 91177308-0d34-0410-b5e6-96231b3b80d8

Remove accidentally duplicated code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49418 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak teh build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49417 91177308-0d34-0410-b5e6-96231b3b80d8

add a simplified accessor for creating an fp constant of a
particular value but variable type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49416 91177308-0d34-0410-b5e6-96231b3b80d8

Missed a hasInterval check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49415 91177308-0d34-0410-b5e6-96231b3b80d8

add a version of ConstantFP::get that doesn't take a redundant Type* value,
start migrating code over to use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49413 91177308-0d34-0410-b5e6-96231b3b80d8

many cleanups to the pow optimizer.  Allow it to handle powf,
add support for  pow(x, 2.0) -> x*x.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49411 91177308-0d34-0410-b5e6-96231b3b80d8

make ConstantFP::isExactlyValue work for long double as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49410 91177308-0d34-0410-b5e6-96231b3b80d8

Add CreateGetResult()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49398 91177308-0d34-0410-b5e6-96231b3b80d8

Check that bodies and calls but not declarations
are marked nounwind when compiling without
-fexceptions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49393 91177308-0d34-0410-b5e6-96231b3b80d8

Rename -disable-required-unwind-tables to -unwind-tables-optional.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49391 91177308-0d34-0410-b5e6-96231b3b80d8

Rename -disable-required-unwind-tables to unwind-tables-optional.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49389 91177308-0d34-0410-b5e6-96231b3b80d8

merge r48768 from branches/ggreif/parallelized-test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49382 91177308-0d34-0410-b5e6-96231b3b80d8

Add multiple value return instruction constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49374 91177308-0d34-0410-b5e6-96231b3b80d8

Convenience method for setting the nounwind
attribute for a function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49373 91177308-0d34-0410-b5e6-96231b3b80d8

silence warning when assertions are disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49371 91177308-0d34-0410-b5e6-96231b3b80d8

Fix insert point handling for multiple return values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49367 91177308-0d34-0410-b5e6-96231b3b80d8

Handle the situation in 2008-01-25-EmptyFunction.ll
correctly when unwind info is being generated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49366 91177308-0d34-0410-b5e6-96231b3b80d8

Missed one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49365 91177308-0d34-0410-b5e6-96231b3b80d8

Add -disable-required-unwind-tables to tests
that need it (usually, grepping for some string
found in unwind info)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49364 91177308-0d34-0410-b5e6-96231b3b80d8

Implement new llc flag -disable-required-unwind-tables.
Corresponds to -fno-unwind-tables (usually default in gcc).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49361 91177308-0d34-0410-b5e6-96231b3b80d8

Make getDirnameSep a static method (not part of Path's interface).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49354 91177308-0d34-0410-b5e6-96231b3b80d8

Added method Path::getDirname().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49352 91177308-0d34-0410-b5e6-96231b3b80d8

Added support for Create() calls that take an argument besides the
deserializer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49350 91177308-0d34-0410-b5e6-96231b3b80d8

Rename MemOperand to MachineMemOperand. This was suggested by
review feedback from Chris quite a while ago. No functionality
change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49348 91177308-0d34-0410-b5e6-96231b3b80d8

Remove llvm-upgrade.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49347 91177308-0d34-0410-b5e6-96231b3b80d8

Added note to getting started guide to use llvm-gcc4.2.
Removed llvm-upgrade from documentations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49346 91177308-0d34-0410-b5e6-96231b3b80d8

Add operator= implementations to SparseBitVector, allowing it to be used in GVN.  This results
in both time and memory savings for GVN.  For example, one testcase went from 10.5s to 6s with
this patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49345 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for pr2169.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49344 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49343 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR 2169.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49339 91177308-0d34-0410-b5e6-96231b3b80d8

Use Intrinsic::getDeclaration in more places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49338 91177308-0d34-0410-b5e6-96231b3b80d8