Handle multiple return values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50604 91177308-0d34-0410-b5e6-96231b3b80d8

Implement operator-> for ImmutableMap iterators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50603 91177308-0d34-0410-b5e6-96231b3b80d8

Add separate intrinsics for MMX / SSE shifts with i32 integer operands. This allow us to simplify the horribly complicated matching code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50601 91177308-0d34-0410-b5e6-96231b3b80d8

Do not sink getresult.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50600 91177308-0d34-0410-b5e6-96231b3b80d8

Add assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50596 91177308-0d34-0410-b5e6-96231b3b80d8

verify builtin optimization works like gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50594 91177308-0d34-0410-b5e6-96231b3b80d8

move libcalls to the same place llvm-gcc has it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50593 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a mistake in the computation of leading zeros for udiv.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50591 91177308-0d34-0410-b5e6-96231b3b80d8

Suppress -Wshorten-64-to-32 warnings for 64-bit hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50590 91177308-0d34-0410-b5e6-96231b3b80d8

strength reduce exp2 into ldexp, rdar://5852514

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50586 91177308-0d34-0410-b5e6-96231b3b80d8

add a FIXME so we remember to eventually remove this code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50582 91177308-0d34-0410-b5e6-96231b3b80d8

Undo r50574. We are already ensuring the folded load address is 16-byte aligned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50578 91177308-0d34-0410-b5e6-96231b3b80d8

specify an arch for non-x86 hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50576 91177308-0d34-0410-b5e6-96231b3b80d8

80 column violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50575 91177308-0d34-0410-b5e6-96231b3b80d8

Not safe folding a load + FsXORPSrr into FsXORPSrm. It's loading a FR64 value but the load folding variant expects a 16-byte aligned address.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50574 91177308-0d34-0410-b5e6-96231b3b80d8

Porting r50563 from Tak to mainline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50564 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50562 91177308-0d34-0410-b5e6-96231b3b80d8

Use push_back(...) instead of resize(1, ...), per review feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50561 91177308-0d34-0410-b5e6-96231b3b80d8

Update old-style syntax in some "not grep" tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50560 91177308-0d34-0410-b5e6-96231b3b80d8

Fix uninitialized uses of the FPC variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50558 91177308-0d34-0410-b5e6-96231b3b80d8

New test for bug fixed in 50545.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50548 91177308-0d34-0410-b5e6-96231b3b80d8

Don't try to create PHIs of struct types.  Fallout
from x86-64 calling convention work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50545 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an overaggressive SimplifyDemandedBits optimization on urem. This
fixes the 254.gap regression on x86 and the 403.gcc regression on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50537 91177308-0d34-0410-b5e6-96231b3b80d8

Adding testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50536 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize a local variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50527 91177308-0d34-0410-b5e6-96231b3b80d8

don't randomly miscompile seto/setuo just because we are in
ffastmath mode.  This fixes rdar://5902801, a miscompilation
of gcc.dg/builtins-8.c.

Bill, please pull this into Tak.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50523 91177308-0d34-0410-b5e6-96231b3b80d8

1) add '-debug' output
2) Return NULL instead of false in several places for tidiness.
3) fix a bug optimizing  sprintf(p, "%c", x)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50521 91177308-0d34-0410-b5e6-96231b3b80d8

Delete the IPO simplify-libcalls and completely reimplement it as
a FunctionPass.  This makes it simpler, fixes dozens of bugs, adds
a couple of minor features, and shrinks is considerably: from
2214 to 1437 lines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50520 91177308-0d34-0410-b5e6-96231b3b80d8

fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50519 91177308-0d34-0410-b5e6-96231b3b80d8

instcombine does memset optzns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50518 91177308-0d34-0410-b5e6-96231b3b80d8

simplifylibcalls doesn't optimize llvm.memmove, instcombine does.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50517 91177308-0d34-0410-b5e6-96231b3b80d8

move some tests from libcall optimizer suite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50516 91177308-0d34-0410-b5e6-96231b3b80d8

Add CreateCall3/CreateCall4 at Eric's request.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50515 91177308-0d34-0410-b5e6-96231b3b80d8

Add a spiffy little "CreateCall2" method, which can be used to make
a function call that takes two Value*'s as arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50514 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bug in my previous patch, a classic =/== bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50483 91177308-0d34-0410-b5e6-96231b3b80d8

Really commit the test checking the argument lowering behaviour on x86-64 :).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50478 91177308-0d34-0410-b5e6-96231b3b80d8

Tail call optimization improvements:

Move platform independent code (lowering of possibly overwritten
arguments, check for tail call optimization eligibility) from
target X86ISelectionLowering.cpp to TargetLowering.h and
SelectionDAGISel.cpp.

Initial PowerPC tail call implementation:

Support ppc32 implemented and tested (passes my tests and
test-suite llvm-test).
Support ppc64 implemented and half tested (passes my tests).
On ppc tail call optimization is performed if
  caller and callee are fastcc
  call is a tail call (in tail call position, call followed by ret)
  no variable argument lists or byval arguments
  option -tailcallopt is enabled
Supported:
 * non pic tail calls on linux/darwin
 * module-local tail calls on linux(PIC/GOT)/darwin(PIC)
 * inter-module tail calls on darwin(PIC)
If constraints are not met a normal call will be emitted.

A test checking the argument lowering behaviour on x86-64 was added.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50477 91177308-0d34-0410-b5e6-96231b3b80d8

set ignore property

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50476 91177308-0d34-0410-b5e6-96231b3b80d8

fcntl.h is pretty standard on unix (without the sys/)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50475 91177308-0d34-0410-b5e6-96231b3b80d8

Move this test to LoopDeletion, where it now passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50474 91177308-0d34-0410-b5e6-96231b3b80d8

This condition got inverted accidentally.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50473 91177308-0d34-0410-b5e6-96231b3b80d8

move lowering of llvm.memset -> store from simplify libcalls
to instcombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50472 91177308-0d34-0410-b5e6-96231b3b80d8

no reason for simplifylibcalls to simplify intrinsics, instcombine does
a fine job.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50470 91177308-0d34-0410-b5e6-96231b3b80d8

remove redundant check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50469 91177308-0d34-0410-b5e6-96231b3b80d8

add missing #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50468 91177308-0d34-0410-b5e6-96231b3b80d8

add a method for comparing to see if a value has a specified name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50465 91177308-0d34-0410-b5e6-96231b3b80d8

use string length computation to generalize several xforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50464 91177308-0d34-0410-b5e6-96231b3b80d8

Add comments for previous patch as requested.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50463 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixes and updates for CellSPU, syncing up with trunk. Most notable
fixes are target-specific lowering of frame indices, fix constants generated
for the FSMBI instruction, and fixing SPUTargetLowering::computeMaskedBitsFor-
TargetNode().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50462 91177308-0d34-0410-b5e6-96231b3b80d8

Fix custom target lowering for zero/any/sign_extend: make sure that
DAG.UpdateNodeOperands() is called before (not after) the call to
TLI.LowerOperation().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50461 91177308-0d34-0410-b5e6-96231b3b80d8

Make eh_frame objects by 8-byte aligned on 64-bit
targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50451 91177308-0d34-0410-b5e6-96231b3b80d8

Minor spelling and typo fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50448 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r50441.  The original code was correct.  Add some more comments so that I don't make the same mistake in the future.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50446 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in memcpyopt where the memcpy-memcpy transform was never being applied because
we were checking for it in the wrong order.  This caused a miscompilation because the
return slot optimization assumes that the call it is dealing with is NOT a memcpy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50444 91177308-0d34-0410-b5e6-96231b3b80d8

We should be returning true here since we've changed the function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50442 91177308-0d34-0410-b5e6-96231b3b80d8

A lot of cleanups and documentation improvements, as well as a few corner case fixes.  Most
of this was suggested by Chris.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50441 91177308-0d34-0410-b5e6-96231b3b80d8

Rename DeadLoopElimination to LoopDeletion, part 2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50437 91177308-0d34-0410-b5e6-96231b3b80d8

Rename DeadLoopElimination to LoopDeletion, part one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50436 91177308-0d34-0410-b5e6-96231b3b80d8

Don't do stupid things: doInitialization(Module&) is not applicable to ModulePass :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50433 91177308-0d34-0410-b5e6-96231b3b80d8

don't eliminate load from volatile value on paths where the load is dead.
This fixes the second half of PR2262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50430 91177308-0d34-0410-b5e6-96231b3b80d8

make this test reduced and *valid*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50429 91177308-0d34-0410-b5e6-96231b3b80d8

fix a subtle volatile handling bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50428 91177308-0d34-0410-b5e6-96231b3b80d8

Use std::set instead of std::priority_queue for the RegReductionPriorityQueue.
This removes the existing bottleneck related to the removal of elements from
the middle of the queue.

Also fixes a subtle bug in ScheduleDAGRRList::CapturePred:
It was updating the state of the SUnit before removing it. As a result, the
comparison operators were working incorrectly and this SUnit could not be removed
from the queue properly.

Reviewed by Evan and Dan. Approved by Dan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50412 91177308-0d34-0410-b5e6-96231b3b80d8

Implement more aggressive support for analyzing string length.  This
generalizes the previous code to handle the case when the string is not
an immediate to the strlen call (for example, crazy stuff like
strlen(c ? "foo" : "bart"+1) -> 3).  This implements
gcc.c-torture/execute/builtins/strlen-2.c.  I will generalize other
cases in simplifylibcalls to use the same routine later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50408 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify what we mean by a dead loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50406 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR2094.  The inline asms should not pin allocas to the
stack anymore.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50397 91177308-0d34-0410-b5e6-96231b3b80d8

don't delete the last store to an alloca if the store is volatile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50390 91177308-0d34-0410-b5e6-96231b3b80d8

make the vector conversion magic handle multiple results.
We now compile test2/test3 to:

_test2:
## InlineAsm Start
set %xmm0, %xmm1
## InlineAsm End
addps %xmm1, %xmm0
ret
_test3:
## InlineAsm Start
set %xmm0, %xmm1
## InlineAsm End
paddd %xmm1, %xmm0
ret

as expected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50389 91177308-0d34-0410-b5e6-96231b3b80d8

add support for multiple return values in inline asm. This is a step
towards PR2094.  It now compiles the attached .ll file to:

_sad16_sse2:
movslq %ecx, %rax
## InlineAsm Start
%ecx %rdx %rax %rax %r8d %rdx %rsi
## InlineAsm End
## InlineAsm Start
set %eax
## InlineAsm End
ret

which is pretty decent for a 3 output, 4 input asm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50386 91177308-0d34-0410-b5e6-96231b3b80d8

Another extract_subreg coalescing bug.
e.g.
vr1024<2> extract_subreg vr1025, 2
If vr1024 do not have the same register class as vr1025, it's not safe to coalesce this away. For example, vr1024 might be a GPR32 while vr1025 might be a GPR64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50385 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50384 91177308-0d34-0410-b5e6-96231b3b80d8

Remove debugging code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50383 91177308-0d34-0410-b5e6-96231b3b80d8

Add dead loop elimination, which removes dead loops for which we can compute
the trip count.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50382 91177308-0d34-0410-b5e6-96231b3b80d8

Add -march=x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50380 91177308-0d34-0410-b5e6-96231b3b80d8

Update and_ops.ll according to the recent dagcombiner changes.
Add a new test, and_ops_more.ll, which is XFAIL'd, to
record the parts of and_ops.ll that were affected by this
change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50379 91177308-0d34-0410-b5e6-96231b3b80d8

Test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50377 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in RegsForValue::getCopyToRegs() that causes cyclical scheduling units. If it's creating multiple CopyToReg nodes that are "flagged" together, it should not create a TokenFactor for it's chain outputs:

c1, f1 = CopyToReg
c2, f2 = CopyToReg
c3     = TokenFactor c1, c2
 ...
       = user c3, ..., f2

Now that the two CopyToReg's and the user are "flagged" together. They effectively forms a single scheduling unit. The TokenFactor is now both an operand and a successor of the Flagged nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50376 91177308-0d34-0410-b5e6-96231b3b80d8

Correct parameter attributes encoding for C bindings.
Patch by Anders Johnsen!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50375 91177308-0d34-0410-b5e6-96231b3b80d8

Add possibility of using arbitrary to to execute stuff from bugpoint.
Patch by Pekka Jääskeläinen!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50373 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a pointer-arithmetic bug that caused 64-bit host pointer values to
be truncated to 32 bits. This fixes the recent Benchmarks/McCat/09-vor
regression on x86-64, among other things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50372 91177308-0d34-0410-b5e6-96231b3b80d8

Fix DSE to not eliminate volatile loads with no uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50370 91177308-0d34-0410-b5e6-96231b3b80d8

Don't try to convert PPC long double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50369 91177308-0d34-0410-b5e6-96231b3b80d8

Evan pointed out that folding sext to zext may not be correct
if the zext is not legal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50368 91177308-0d34-0410-b5e6-96231b3b80d8

Delete an unused constructor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50367 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment to CreateRegForValue that clarifies the handling of
aggregate types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50366 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite the comments for RegsForValue and its members, and
reorder some of the members for clarity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50365 91177308-0d34-0410-b5e6-96231b3b80d8

Add more alignment enums.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50363 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix in BumpPtrAllocator: don't assume that all objects have the same alignment.  "Bump" of the pointer for the next allocated object to be of the specified alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50362 91177308-0d34-0410-b5e6-96231b3b80d8

Don't call size() on each iteration of the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50361 91177308-0d34-0410-b5e6-96231b3b80d8

Expose parameter attributes via C bindings.

Patch by Anders Johnsen!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50360 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the SVOffset values for loads and stores produced by
memcpy/memset expansion. It was a bug for the SVOffset value
to be used in the actual address calculations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50359 91177308-0d34-0410-b5e6-96231b3b80d8

Teach InstCombine's ComputeMaskedBits what SelectionDAG's
ComputeMaskedBits knows about cttz, ctlz, and ctpop. Teach
SelectionDAG's ComputeMaskedBits what InstCombine's knows
about SRem. And teach them both some things about high bits
in Mul, UDiv, URem, and Sub. This allows instcombine and
dagcombine to eliminate sign-extension operations in
several new cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50358 91177308-0d34-0410-b5e6-96231b3b80d8

Teach DAGCombine to convert (sext x) to (zext x) when the
sign-bit of x is known to be zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50357 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for response files to the CommandLine library.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50355 91177308-0d34-0410-b5e6-96231b3b80d8

Fix FP return for Win64 ABI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50342 91177308-0d34-0410-b5e6-96231b3b80d8

Another collection of random cleanups.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50341 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the SmallVector ctor that converts from a SmallVectorImpl.  This
conversion open the door for many nasty implicit conversion issues, and
can be easily solved by initializing with (V.begin(), V.end()) when
needed.

This patch includes many small cleanups for sdisel also.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50340 91177308-0d34-0410-b5e6-96231b3b80d8

restore the copy ctor in SmallVector.  This fixes serious
errors I introduced in my last patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50338 91177308-0d34-0410-b5e6-96231b3b80d8

switch RegsForValue::Regs to be a SmallVector to avoid
heap thrash on tiny (usually single-element) vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50335 91177308-0d34-0410-b5e6-96231b3b80d8

generalize SmallVector copy ctor, there is no requirement for
the initialization vector to have the same fixed size, just the
same element type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50334 91177308-0d34-0410-b5e6-96231b3b80d8