Make the GraphRoot edge look like a chain edge, which is more accurate,
and use the right result number, in the off chance that the graph root
has multiple result values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53923 91177308-0d34-0410-b5e6-96231b3b80d8

Provide default implementation of different small-sections related stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53920 91177308-0d34-0410-b5e6-96231b3b80d8

Tie small stuff to non-small by default on ELF platforms

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53919 91177308-0d34-0410-b5e6-96231b3b80d8

simplified small section logic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53912 91177308-0d34-0410-b5e6-96231b3b80d8

Fix encoding of atomic compare and swap for i64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53911 91177308-0d34-0410-b5e6-96231b3b80d8

Added small section asm emition logic for mips.
Fixed small bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53908 91177308-0d34-0410-b5e6-96231b3b80d8

Basic support for small sections

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53907 91177308-0d34-0410-b5e6-96231b3b80d8

Remove more tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53905 91177308-0d34-0410-b5e6-96231b3b80d8

Remove another tab.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53904 91177308-0d34-0410-b5e6-96231b3b80d8

More tab removals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53903 91177308-0d34-0410-b5e6-96231b3b80d8

Removing tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53902 91177308-0d34-0410-b5e6-96231b3b80d8

Use — because.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53901 91177308-0d34-0410-b5e6-96231b3b80d8

Remove references to llvm-gcc-4.0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53900 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53898 91177308-0d34-0410-b5e6-96231b3b80d8

Another buildbot test commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53896 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the name of MachineMemOperand's include guard.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53895 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53894 91177308-0d34-0410-b5e6-96231b3b80d8

Fix multiple-return-value-to-first-class-aggregates autoupgrade to
correctly handle the case where multiple-return-value constructs
were used to return one or zero values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53890 91177308-0d34-0410-b5e6-96231b3b80d8

Trivial check-in to test buildbot. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53889 91177308-0d34-0410-b5e6-96231b3b80d8

InsertValue and ExtractValue constant expressions are always
folded. Remove code that handled the case where they aren't
folded, and remove bitcode reader/writer support for them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53887 91177308-0d34-0410-b5e6-96231b3b80d8

Provide llvm bitcode file to native object file interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53886 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammaros in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53884 91177308-0d34-0410-b5e6-96231b3b80d8

Add the PR number to the test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53880 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in LSR's dead-PHI cleanup. If a PHI has a def-use chain that
leads into a cycle involving a different PHI, LSR got stuck running
around that cycle looking for the original PHI. To avoid this, keep
track of visited PHIs and stop searching if we see one more than once.
This fixes PR2570.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53879 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance the GraphWriter support for edge destinations, and teach the
SelectionDAG graph writer to make use of them. Now, nodes with multiple
values are displayed as such, with incoming edges pointing to the
specific value they use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53875 91177308-0d34-0410-b5e6-96231b3b80d8

After early-lowering the FORMAL_ARGUMENTS node, delete it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53874 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate a compilation warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53873 91177308-0d34-0410-b5e6-96231b3b80d8

Add titles to the various SelectionDAG viewGraph calls
that include useful information like the name of the
block being viewed and the current phase of compilation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53872 91177308-0d34-0410-b5e6-96231b3b80d8

Make the GraphWriter be more consistent about the string
used for the graph "title" and the graph "label", as there
are differences in interpretation of these strings
between viewers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53871 91177308-0d34-0410-b5e6-96231b3b80d8

Fix uses of underscore-capital names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53870 91177308-0d34-0410-b5e6-96231b3b80d8

Added initial support for small sections on Mips.
Added gp_rel relocations to support addressing small section contents.
Added command line to specify small section threshold in bytes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53869 91177308-0d34-0410-b5e6-96231b3b80d8

Now that the MachineInstr leaks are fixed, enable leak checking
in the MachineInstr clone code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53868 91177308-0d34-0410-b5e6-96231b3b80d8

Temporary hack to build with GCC 4.0 instead of 4.2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53860 91177308-0d34-0410-b5e6-96231b3b80d8

Use better variable names

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53859 91177308-0d34-0410-b5e6-96231b3b80d8

Don't use larger alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53857 91177308-0d34-0410-b5e6-96231b3b80d8

Add VerifyNode, a place to put sanity checks on
generic SDNode's (nodes with their own constructors
should do sanity checking in the constructor).  Add
sanity checks for BUILD_VECTOR and fix all the places
that were producing bogus BUILD_VECTORs, as found by
"make check".  My favorite is the BUILD_VECTOR with
only two operands that was being used to build a
vector with four elements!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53850 91177308-0d34-0410-b5e6-96231b3b80d8

Use movaps instead of movups to spill 16-byte vector values when default alignment is >= 16. This fixes some massive performance regressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53844 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r53812 -- premature. LegalizeTypes isn't actually on yet!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53816 91177308-0d34-0410-b5e6-96231b3b80d8

Switch on the use of arbitrary precision integers in scalar evolution. This will
bail after 256-bits to avoid producing code that the backends can't handle.
Previously, we capped it at 64-bits, preferring to miscompile in those cases.

This change also reverts much of r52248 because the invariants the code was
expecting are now being met.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53812 91177308-0d34-0410-b5e6-96231b3b80d8

This header isn't necessary now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53811 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2088. Use modulo linear equation solver to compute loop iteration
count.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53810 91177308-0d34-0410-b5e6-96231b3b80d8

Updated VC++ projects

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53809 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for first part of PR2562. Generate the "pinsrw" instruction for inserts
into v4i16 vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53807 91177308-0d34-0410-b5e6-96231b3b80d8

Pull r53795 from Gaz into mainline:

If .loc and .file aren't used, always emit the "debug_line" section. This
requires at least one entry in the line matrix. So if there's nothing to emit
into the matrix, emit an end of matrix value anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53803 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak build: 'DarwinTargetAsmInfo' was already taken as PPC TAI flavour.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53801 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure custom lowering for LegalizeTypes
returns a node with the right number of
return values.  This fixes codegen of
Generic/cast-fp.ll, Generic/fp_to_int.ll
and PowerPC/multiple-return-values.ll
when using -march=ppc32 -mattr=+64bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53794 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53793 91177308-0d34-0410-b5e6-96231b3b80d8

While testing particular algorithms to compute loop iteration count the brute
force evaluation (ComputeIterationCountExhaustively) should be turned off.

It doesn't apply to trip-count2.ll because this file tests the brute force
evaluation.

The test for PR2364 (2008-05-25-NegativeStepToZero.ll) currently fails
showing that the patch for this bug doesn't work. I'll fix it in a few hours
with a patch for PR2088.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53792 91177308-0d34-0410-b5e6-96231b3b80d8

Use chars, where possible

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53791 91177308-0d34-0410-b5e6-96231b3b80d8

Switch MIPS to new ELFTargetAsmInfo. Add few FIXMEs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53790 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a FIXME :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53789 91177308-0d34-0410-b5e6-96231b3b80d8

Use generic ELFTargetAsmInfo and DarwinTargetAsmInfo for X86 code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53788 91177308-0d34-0410-b5e6-96231b3b80d8

Add TargetAsmInfo stuff for all darwin-based targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53787 91177308-0d34-0410-b5e6-96231b3b80d8

Add TargetAsmInfo for all ELF-based targets

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53786 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR2549

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53785 91177308-0d34-0410-b5e6-96231b3b80d8

Use aligned stack spills, where possible. This fixes PR2549.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53784 91177308-0d34-0410-b5e6-96231b3b80d8

Remove bogus assertion. This unbreaks mingw, where ConstantSDNode
has alignment 8 and LoadSDNode (used as LargestT template) - 4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53782 91177308-0d34-0410-b5e6-96231b3b80d8

tentative definitions are default visibility

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53780 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a memory leak in LiveIntervalAnalysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53779 91177308-0d34-0410-b5e6-96231b3b80d8

typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53778 91177308-0d34-0410-b5e6-96231b3b80d8

If all linker input files are native object files then lto-bugpoint is not useful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53777 91177308-0d34-0410-b5e6-96231b3b80d8

Start writing lto-bugpoint tool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53774 91177308-0d34-0410-b5e6-96231b3b80d8

Softfloat support for FDIV.  Patch by
Richard Pennington.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53773 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53772 91177308-0d34-0410-b5e6-96231b3b80d8

Supress a gcc-4.3 warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53771 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 53729, after waking up in the middle of
the night realising that it was wrong :)  I
think the reason the same type was being used
for the shufflevec of indices as for the actual
indices is so that if one of them needs splitting
then so does the other.  After my patch it might
be that the indices need splitting but not the
rest, yet there is no good way of handling that.
I think the right solution is to not have the
shufflevec be an operand at all: just have it
be the list of numbers it actually is, stored
as extra info in the node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53768 91177308-0d34-0410-b5e6-96231b3b80d8

In the CBackend, use casts to force integer add, subtract, and
multiply to be done as unsigned, so that they have well defined
behavior on overflow. This fixes PR2408.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53767 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a LocalSpiller leak. This fixes tramp3d-v4.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53766 91177308-0d34-0410-b5e6-96231b3b80d8

Make PRE actually handle critical edges (by splitting them).  Confirmed that bootstrap passes with this change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53762 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply r53735.  My last patch fixed the failures Dan observed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53761 91177308-0d34-0410-b5e6-96231b3b80d8

Add some checks that got lost in the shuffle.  This fixes 464.h264ref.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53760 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r53735. It broke SPEC 464.h264ref.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53757 91177308-0d34-0410-b5e6-96231b3b80d8

verify limits-fndefn.c from the GCC testsuite before the
heat death of the universe, fixing an O(N^2) problem in the
size of a basic block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53749 91177308-0d34-0410-b5e6-96231b3b80d8

Re-introduce LeakDetector support for MachineInstrs and MachineBasicBlocks.
Fix a leak that this turned up in LowerSubregs.cpp.
And, comment a leak in LiveIntervalAnalysis.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53746 91177308-0d34-0410-b5e6-96231b3b80d8

When printing MemOperand nodes, only use print() for
PseudoSourceValue values, which never have names. Use getName()
for all other values, because we want to print just a short summary
of the value, not the entire instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53738 91177308-0d34-0410-b5e6-96231b3b80d8

Use MergeBlockIntoPredecessor to simplify some code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53735 91177308-0d34-0410-b5e6-96231b3b80d8

Subreg live interval valno may not have a corresponding def machineinstr since it's less precise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53734 91177308-0d34-0410-b5e6-96231b3b80d8

Add nounwind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53733 91177308-0d34-0410-b5e6-96231b3b80d8

test commit access

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53732 91177308-0d34-0410-b5e6-96231b3b80d8

Make MergeBlockIntoPredecessor more aggressive when the same successor appears
more than once.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53731 91177308-0d34-0410-b5e6-96231b3b80d8

Enable PRE.  My last batch of changes fixed the miscompile.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53730 91177308-0d34-0410-b5e6-96231b3b80d8

Use a legal type for elements of the vector_shuffle
mask.  These are just indices into the shuffled vector
so their type is unrelated to the type of the
shuffled elements (which is what was being used before).
This fixes vec_shuffle-11.ll when using LegalizeTypes.
What seems to have happened is that Dan's recent change
r53687, which corrected the result type of the shuffle,
somehow caused LegalizeTypes to notice that the mask
operand was a BUILD_VECTOR with a legal type but elements
of an illegal type (i64).  LegalizeTypes legalized this
by introducing a new BUILD_VECTOR of i32 and bitcasting
it to the old type.  But the mask operand is not supposed
to be a bitcast but a straight BUILD_VECTOR of constants,
causing a crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53729 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new function, ReplaceAllUsesOfValuesWith, which handles bulk
replacement of multiple values. This is slightly more efficient
than doing multiple ReplaceAllUsesOfValueWith calls, and theoretically
could be optimized even further. However, an important property of this
new function is that it handles the case where the source value set and
destination value set overlap. This makes it feasible for isel to use
SelectNodeTo in many very common cases, which is advantageous because
SelectNodeTo avoids a temporary node and it doesn't require CSEMap
updates for users of values that don't change position.

Revamp MorphNodeTo, which is what does all the work of SelectNodeTo, to
handle operand lists more efficiently, and to correctly handle a number
of corner cases to which its new wider use exposes it.

This commit also includes a change to the encoding of post-isel opcodes
in SDNodes; now instead of being sandwiched between the target-independent
pre-isel opcodes and the target-dependent pre-isel opcodes, post-isel
opcodes are now represented as negative values. This makes it possible
to test if an opcode is pre-isel or post-isel without having to know
the size of the current target's post-isel instruction set.

These changes speed up llc overall by 3% and reduce memory usage by 10%
on the InstructionCombining.cpp testcase with -fast and -regalloc=local.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53728 91177308-0d34-0410-b5e6-96231b3b80d8

Indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53727 91177308-0d34-0410-b5e6-96231b3b80d8

LegalizeTypes support for what seems to be the
only missing ppc long double operations: FNEG
and FP_EXTEND.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53723 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary readme entry

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53722 91177308-0d34-0410-b5e6-96231b3b80d8

Turn LegalizeTypes back off again for the moment:
it is breaking Darwin bootstrap due to missing
functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53721 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo in last commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53720 91177308-0d34-0410-b5e6-96231b3b80d8

SSE codegen for vsetcc nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53719 91177308-0d34-0410-b5e6-96231b3b80d8

Make GlobalOpt preserve address spaces when scalar replacing aggregate globals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53716 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2553

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53715 91177308-0d34-0410-b5e6-96231b3b80d8

When lowering certain atomics, we need to copy the memoperand from the old
atomic operation to the new one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53714 91177308-0d34-0410-b5e6-96231b3b80d8

Factorize some code for determining which libcall to use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53713 91177308-0d34-0410-b5e6-96231b3b80d8

Inliner tweak. Function calls should cost more than one instruction!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53712 91177308-0d34-0410-b5e6-96231b3b80d8

Factor MergeBlockIntoPredecessor out into BasicBlockUtils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53705 91177308-0d34-0410-b5e6-96231b3b80d8

Name string length is end position marker - begin position marker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53697 91177308-0d34-0410-b5e6-96231b3b80d8

Do not forget global definitions from inline asm code block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53693 91177308-0d34-0410-b5e6-96231b3b80d8

Mark function used by asm block as used, otherwise optimizer may not see the use and may delete the function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53692 91177308-0d34-0410-b5e6-96231b3b80d8

There's no need to iterate block merging and PRE.  In fact, iterating the latter
could cause problems for memdep when it breaks critical edges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53691 91177308-0d34-0410-b5e6-96231b3b80d8

Somehow, custom lowering of i64 multiplications got dropped along the way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@53689 91177308-0d34-0410-b5e6-96231b3b80d8