Pre-alloc splitting needs to be more careful to avoid inserting spills/restores
between call frame setup/restore points.  Unfortunately, this regresses
code size a bit, but at least it's correct now!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63837 91177308-0d34-0410-b5e6-96231b3b80d8

PR3485, document alignment on byref.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63836 91177308-0d34-0410-b5e6-96231b3b80d8

It's not obvious, but lto_module_create_from_memory doesn't need to use the
buffer after it creates the Module. Thus, we don't need to store this pointer
in claimed_file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63834 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug where we were not emitting a cvt rnd sat node for converting
between a unsigned integer and signed integer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63831 91177308-0d34-0410-b5e6-96231b3b80d8

Remove accidentally included debug message!

Reword a comment for clarity. Remove some extra whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63823 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply 63765.  Patches for clang and llvm-gcc to follow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63812 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of 3 non-DebugLoc getNode variants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63808 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore dbg intrinsics while folding switch instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63802 91177308-0d34-0410-b5e6-96231b3b80d8

Remove non-DebugLoc versions of getMergeValues, ZeroExtendInReg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63800 91177308-0d34-0410-b5e6-96231b3b80d8

Remove non-DebugLoc forms of CopyToReg and CopyFromReg.
Adjust callers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63789 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting 63765.  This broke the build of both clang
and llvm-gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63786 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore dbg intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63781 91177308-0d34-0410-b5e6-96231b3b80d8

Add test case for r63760.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63774 91177308-0d34-0410-b5e6-96231b3b80d8

Remove now-incorrect test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63772 91177308-0d34-0410-b5e6-96231b3b80d8

Alphabetize includes. Update comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63771 91177308-0d34-0410-b5e6-96231b3b80d8

80 column rule.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63768 91177308-0d34-0410-b5e6-96231b3b80d8

Remove non-DebugLoc versions of getLoad and getStore.
Adjust the many callers of those versions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63767 91177308-0d34-0410-b5e6-96231b3b80d8

New feature: add support for target intrinsics being defined in the
target directories themselves.  This also means that VMCore no longer
needs to know about every target's list of intrinsics.  Future work
will include converting the PowerPC target to this interface as an
example implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63765 91177308-0d34-0410-b5e6-96231b3b80d8

Avoids generating a legalization assert for the case where a vector type is legal
but when legalizing the operation, we split the vector type and generate a library
call whose type needs to be promoted. For example, X86 with SSE on but MMX off,
a divide v2i64 will be scalarized to 2 calls to a library using i64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63760 91177308-0d34-0410-b5e6-96231b3b80d8

mention that PIC is needed for libLTO and libLLVMgold

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63755 91177308-0d34-0410-b5e6-96231b3b80d8

Bill implemented this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63752 91177308-0d34-0410-b5e6-96231b3b80d8

add a note, this is why we're faster at SciMark-MonteCarlo with
SSE disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63751 91177308-0d34-0410-b5e6-96231b3b80d8

Skip over zero registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63748 91177308-0d34-0410-b5e6-96231b3b80d8

allow main to have any integer type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63743 91177308-0d34-0410-b5e6-96231b3b80d8

remove printf - it was there only for debugging!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63742 91177308-0d34-0410-b5e6-96231b3b80d8

add support for .a files containing LLVM IR to the gold plugin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63741 91177308-0d34-0410-b5e6-96231b3b80d8

Minor code cleanups; no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63740 91177308-0d34-0410-b5e6-96231b3b80d8

Since I'm obliged to work with a development OS that currently doesn't
support GraphViz, I've been using the foo->dump() facility.  This
patch is a minor rewrite to the SelectionDAG dump() stuff to make it a
little more helpful.  The existing foo->dump() functionality does not
change; this patch adds foo->dumpr().  All of this is only useful when
running LLVM under a debugger.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63736 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the inverse transform x86_fp80 -> i80 (also
fires during the Ada build).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63731 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR3468: a crash when constant folding a bitcast of
i80 to x86 long double (this was presumably generated
by sroa).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63730 91177308-0d34-0410-b5e6-96231b3b80d8

Don't call isInvariantLoad twice.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63729 91177308-0d34-0410-b5e6-96231b3b80d8

Don't bother hoisting out a "cheap" instruction if all of its uses are PHIs. LICM "cheap" instructions are not particularly beneficial to start with. This will just end up making the copies harder to coalesce.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63728 91177308-0d34-0410-b5e6-96231b3b80d8

For now, only hoist re-materilizable instructions. LICM will increase register pressure. We want to avoid spilling more instructions if it's possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63725 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63724 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply patch from r62553, with a fix to avoid looking for an ffi.h that isn't
there.

This changes the interpreter to use libffi. After this patch, the interpreter
will barely be able to call any external functions if built on a system without
libffi installed (just enough to pass 'make check' really). But with libffi,
we can now call any function that isn't variadic or taking a struct or vector
parameter (but pointer to struct is fine). Patch by Alexei Svitkine!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63723 91177308-0d34-0410-b5e6-96231b3b80d8

Finish making AliasAnalysis aware of the fact that most atomic intrinsics only dereference their arguments, and enhance
BasicAA to make use of this fact when computing ModRef info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63718 91177308-0d34-0410-b5e6-96231b3b80d8

Remove non-DebugLoc forms of the exotic forms
of Lod and Sto; patch uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63716 91177308-0d34-0410-b5e6-96231b3b80d8

fix a pretty serious bug in StringMap where GetOrCreate(strstart,strend, z)
would not set new values to Z.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63707 91177308-0d34-0410-b5e6-96231b3b80d8

add a friend needed by a stringmap change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63706 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some more non-DebugLoc versions of construction
functions, with callers adjusted to fit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63705 91177308-0d34-0410-b5e6-96231b3b80d8

Check in file I forgot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63704 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a few non-DebugLoc versions of node creation
functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63703 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes a case where we generate an incorrect mask for pshfhw in the presence
of undefs and incorrectly determining if we have punpckldq.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63702 91177308-0d34-0410-b5e6-96231b3b80d8

While folding vallue comparison terminators ignore dbg intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63700 91177308-0d34-0410-b5e6-96231b3b80d8

Patch up omissions in DebugLoc propagation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63693 91177308-0d34-0410-b5e6-96231b3b80d8

Fill in more omissions in DebugLog propagation.
I think that's it for this directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63690 91177308-0d34-0410-b5e6-96231b3b80d8

Constify and don't return a reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63689 91177308-0d34-0410-b5e6-96231b3b80d8

Ignore dbg intrinsics while hoisting common code in the two blocks up into the branch block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63687 91177308-0d34-0410-b5e6-96231b3b80d8

DebugLoc propagation; adjustment to things omitted
from SelectionDagBuild.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63680 91177308-0d34-0410-b5e6-96231b3b80d8

Erm. Would be nice to initialize this variable. . .

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63677 91177308-0d34-0410-b5e6-96231b3b80d8

Add getDebugLocTuple to retrieve the DebugLocTuple for a given DebugLoc object.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63676 91177308-0d34-0410-b5e6-96231b3b80d8

Add getter for the index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63675 91177308-0d34-0410-b5e6-96231b3b80d8

Need this file too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63674 91177308-0d34-0410-b5e6-96231b3b80d8

Add some DL propagation to places that didn't
have it yet.  More coming.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63673 91177308-0d34-0410-b5e6-96231b3b80d8

Do not let dbg intrinsic block folding of two entry phi node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63671 91177308-0d34-0410-b5e6-96231b3b80d8

If "optimize for size" attribute is set then block non-trivial loop unswitches but allow trivial loop unswitches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63670 91177308-0d34-0410-b5e6-96231b3b80d8

DebugLoc propgation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63664 91177308-0d34-0410-b5e6-96231b3b80d8

More whitespace fixin'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63663 91177308-0d34-0410-b5e6-96231b3b80d8

Whitespace and comment changes. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63660 91177308-0d34-0410-b5e6-96231b3b80d8

teach "convert from scalar" to handle loads of fca's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63659 91177308-0d34-0410-b5e6-96231b3b80d8

refactor the interface to ConvertUsesOfLoadToScalar,
renaming it to ConvertScalar_ExtractValue

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63658 91177308-0d34-0410-b5e6-96231b3b80d8

DebugLoc propagation.  done with file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63656 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63653 91177308-0d34-0410-b5e6-96231b3b80d8

convert ConvertUsesOfLoadToScalar to use IRBuilder,
no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63652 91177308-0d34-0410-b5e6-96231b3b80d8

switch ConvertScalar_InsertValue to use an IRBuilder, no
functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63651 91177308-0d34-0410-b5e6-96231b3b80d8

DebugLoc propagation.  2/3 through file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63650 91177308-0d34-0410-b5e6-96231b3b80d8

make scalar conversion handle stores of first class
aggregate values.  loads are not yet handled (coming
soon to an sroa near you).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63649 91177308-0d34-0410-b5e6-96231b3b80d8

Change the post-RA scheduler to iterate through the
basic-block segments bottom-up instead of top down. This
is the first step in a general restructuring of the way
register liveness is tracked in the post-RA scheduler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63643 91177308-0d34-0410-b5e6-96231b3b80d8

Tevert part of the x86 subtarget logic changes: when -march=x86-64
is given, override the subtarget settings and enable 64-bit support.
This restores the earlier behavior, and fixes regressions on
Non-64-bit-capable x86-32 hosts.

This isn't necessarily the best approach, but the most obvious
alternative is to require -mcpu=x86-64 or -mattr=+64bit to be used
with -march=x86-64 when the host doesn't have 64-bit support. This
makes things little more consistent, but it's less convenient, and
it has the practical drawback of requiring lots of test changes, so
I opted for the above approach for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63642 91177308-0d34-0410-b5e6-96231b3b80d8

First initialize DAG otherwise dwarf writer is used uninitialized.
Duncan spotted this. Thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63641 91177308-0d34-0410-b5e6-96231b3b80d8

Make SROA produce a vector only when the alloca is actually
accessed at least once as a vector.  This prevents it from
compiling the example in not-a-vector into:

define double @test(double %A, double %B) {
%tmp4 = insertelement <7 x double> undef, double %A, i32 0
%tmp = insertelement <7 x double> %tmp4, double %B, i32 4
%tmp2 = extractelement <7 x double> %tmp, i32 4
ret double %tmp2
}

instead, producing the integer code.  Producing vectors when they
aren't otherwise in the program is dangerous because a lot of other
code treats them carefully and doesn't want to break them down.
OTOH, many things want to break down tasty i448's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63638 91177308-0d34-0410-b5e6-96231b3b80d8

this produces an undefined result, just check that the alloca is gone
and that sroa doesn't crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63637 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR3411.  When replacing values, nodes are analyzed
in any old order.  Since analyzing a node analyzes its
operands also, this can mean that when we pop a node
off the list of nodes to be analyzed, it may already
have been analyzed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63632 91177308-0d34-0410-b5e6-96231b3b80d8

APInt'fy SimplifyDemandedVectorElts so it can analyze vectors with more than 64 elements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63631 91177308-0d34-0410-b5e6-96231b3b80d8

add a method to BumpPtrAllocator that allows allocating elements
with a specified alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63629 91177308-0d34-0410-b5e6-96231b3b80d8

Add LLVM plugin for gold.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63623 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate configure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63622 91177308-0d34-0410-b5e6-96231b3b80d8

Add the ability to pass the path to binutils source to configure. This is
needed to build the LLVM gold plugin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63621 91177308-0d34-0410-b5e6-96231b3b80d8

add another case of undefined behavior without crashing, PR3466.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63620 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r63600. It didn't fix the bug, it just moved it a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63618 91177308-0d34-0410-b5e6-96231b3b80d8

Teach AliasAnalysis that a bunch of the atomic intrinsics only dereference their arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63616 91177308-0d34-0410-b5e6-96231b3b80d8

Update the callgraph when replacing InvokeInst with CallInst when inlining.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63600 91177308-0d34-0410-b5e6-96231b3b80d8

Explicitly pass in debug location information to BuildMI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63599 91177308-0d34-0410-b5e6-96231b3b80d8

Pass in something sensible for the debug location information when creating the
initial PHI nodes of the machine function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63598 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bitcode reader bug where it can't handle extractelement correctly:
the index of the value being extracted is always an i32.  This fixes PR3465

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63597 91177308-0d34-0410-b5e6-96231b3b80d8

Teach ConvertUsesToScalar to handle memset, allowing it to handle
crazy cases like:

struct f {  int A, B, C, D, E, F; };
short test4() {
  struct f A;
  A.A = 1;
  memset(&A.B, 2, 12);
  return A.C;
}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63596 91177308-0d34-0410-b5e6-96231b3b80d8

Fill in some missing DL propagation in getNode()s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63595 91177308-0d34-0410-b5e6-96231b3b80d8

Use SDL->getCurDebugLoc() instead of unknown loc for landing pads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63594 91177308-0d34-0410-b5e6-96231b3b80d8

Explicitly pass in the "unknown" debug location. This is probably not
correct. We need more infrastructure before we can get the DebugLoc info for
these instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63593 91177308-0d34-0410-b5e6-96231b3b80d8

Delete these two tests. They are specific to x86-64, and there's no
reliable way to do this with the current dejagnu infrastructure.
If someone can figure out how to fix these tests so that they test
what they are intended to test without spuriously failing on any
popular platforms, they are invited to reinstate them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63592 91177308-0d34-0410-b5e6-96231b3b80d8

Alphabetize includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63591 91177308-0d34-0410-b5e6-96231b3b80d8

rearrange how SRoA handles promotion of allocas to vectors.
With the new world order, it can handle cases where the first
store into the alloca is an element of the vector, instead of
requiring the first analyzed store to have the vector type
itself.  This allows us to un-xfail
test/CodeGen/X86/vec_ins_extract.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63590 91177308-0d34-0410-b5e6-96231b3b80d8

Move isIdentifiedObject and isNoAliasCall into AliasAnalysis.cpp since
they are useful to analyses other than BasicAliasAnalysis.cpp. Include
the full comment for isIdentifiedObject in the header file. Thanks to
Chris for suggeseting this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63589 91177308-0d34-0410-b5e6-96231b3b80d8

this test produces an undefined value, we don't care
what it is, but we do want the alloca promoted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63587 91177308-0d34-0410-b5e6-96231b3b80d8

Propagate debug loc info during SDNode -> machine instr creation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63585 91177308-0d34-0410-b5e6-96231b3b80d8

Create DebugLoc information in FastISel. Several temporary methods were
created. Specifically, those BuildMIs which use
"DebugLoc::getUnknownLoc()". I'll remove them soon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63584 91177308-0d34-0410-b5e6-96231b3b80d8

Propagation in TargetLowering.  Includes passing a DL
into SimplifySetCC which gets called elsewhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63583 91177308-0d34-0410-b5e6-96231b3b80d8

It fails on Linux. XFAIL that machine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63582 91177308-0d34-0410-b5e6-96231b3b80d8

This is passing for us. Should it have been reenabled?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63580 91177308-0d34-0410-b5e6-96231b3b80d8

Add explicit -march=x86 to these tests so that they don't
default to -march=x86-64 on 64-bit hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@63579 91177308-0d34-0410-b5e6-96231b3b80d8