Add patterns for CALL32m and CALL64m. They aren't matched in most
cases due to an isel deficiency already noted in
lib/Target/X86/README.txt, but they can be matched in this fold-call.ll
testcase, for example.

This is interesting mainly because it exposes a tricky tblgen bug;
tblgen was incorrectly computing the starting index for variable_ops
in the case of a complex pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51706 91177308-0d34-0410-b5e6-96231b3b80d8

Remove more iostream header includes. Needed to implement a "FlushStream"
function to flush a specified std::ostream.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51705 91177308-0d34-0410-b5e6-96231b3b80d8

Remove <iostream>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51704 91177308-0d34-0410-b5e6-96231b3b80d8

Add newline at end of file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51701 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a tblgen problem handling variable_ops in tblgen instruction
definitions. This adds a new construct, "discard", for indicating
that a named node in the input matching pattern is to be discarded,
instead of corresponding to a node in the output pattern. This
allows tblgen to know where the arguments for the varaible_ops are
supposed to begin.

This fixes "rdar://5791600", whatever that is ;-).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51699 91177308-0d34-0410-b5e6-96231b3b80d8

const-ify getOpcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51698 91177308-0d34-0410-b5e6-96231b3b80d8

Prune and tidy #includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51697 91177308-0d34-0410-b5e6-96231b3b80d8

Expand small memmovs using inline code. Set the X86 threshold for expanding
memmove to a more plausible value, now that it's actually being used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51696 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1338: Rename test dirs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51695 91177308-0d34-0410-b5e6-96231b3b80d8

Revert part of my last patch that I didn't intend to commit yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51694 91177308-0d34-0410-b5e6-96231b3b80d8

Add file with warning for backward comptibility. Should be removed after 2.4

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51693 91177308-0d34-0410-b5e6-96231b3b80d8

Update guards

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51692 91177308-0d34-0410-b5e6-96231b3b80d8

Renumbering needs to account for instruction slot offsets when performing lookups in the index maps.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51691 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51688 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1338: rename include/llvm/ADT/ilist and friends to end with ".h"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51687 91177308-0d34-0410-b5e6-96231b3b80d8

Force postdom to be linked into opt and bugpoint, even though it is no longer used by any passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51686 91177308-0d34-0410-b5e6-96231b3b80d8

Move these tests into the proper directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51685 91177308-0d34-0410-b5e6-96231b3b80d8

Add a newline at the end of this file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51680 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable the newly simplified ADCE.  This fixes a regression on
Dhrystone introduced by its removal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51669 91177308-0d34-0410-b5e6-96231b3b80d8

Replace the old ADCE implementation with a new one that more simply solves
the one case that ADCE catches that normal DCE doesn't: non-induction variable
loop computations.

This implementation handles this problem without using postdominators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51668 91177308-0d34-0410-b5e6-96231b3b80d8

Implement vector shift up / down and insert zero with ps{rl}lq / ps{rl}ldq.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51667 91177308-0d34-0410-b5e6-96231b3b80d8

Remove debugging code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51666 91177308-0d34-0410-b5e6-96231b3b80d8

Add nounwind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51665 91177308-0d34-0410-b5e6-96231b3b80d8

XOR?RI instructions aren't as cheap as moves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51664 91177308-0d34-0410-b5e6-96231b3b80d8

convert more operand loops to iterator formulation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51663 91177308-0d34-0410-b5e6-96231b3b80d8

Implement "AsCheapAsAMove" for some obviously cheap instructions: xor and the
like.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51662 91177308-0d34-0410-b5e6-96231b3b80d8

convert another operand loop to iterator formulation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51661 91177308-0d34-0410-b5e6-96231b3b80d8

Add a flag to indicate that an instruction is as cheap (or cheaper) than a move
instruction to execute. This can be used for transformations (like two-address
conversion) to remat an instruction instead of generating a "move"
instruction. The idea is to decrease the live ranges and register pressure and
all that jazz.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51660 91177308-0d34-0410-b5e6-96231b3b80d8

Check the "isSafeToMove" predicate, which has a series of tests to make sure
that it's safe to remat an instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51659 91177308-0d34-0410-b5e6-96231b3b80d8

Remap VNInfo data as well when doing renumbering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51658 91177308-0d34-0410-b5e6-96231b3b80d8

Factor the numbering computation into a separate method, and add the slightest attempt at some renumbering logic, which is currently unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51652 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2289: vr defined by multiple implicit_def as result of coalescing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51648 91177308-0d34-0410-b5e6-96231b3b80d8

Teach local register allocator to deal with landing pad MBB's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51647 91177308-0d34-0410-b5e6-96231b3b80d8

Update some comments noticed in a recent checkin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51644 91177308-0d34-0410-b5e6-96231b3b80d8

Update text to point people at the right version of the tutorial for
their release.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51638 91177308-0d34-0410-b5e6-96231b3b80d8

Implement PR2370: memmove(x,x,size) -> noop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51636 91177308-0d34-0410-b5e6-96231b3b80d8

Add chain inputs for loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51635 91177308-0d34-0410-b5e6-96231b3b80d8

Fix CodeGen/Generic/2005-10-21-longlonggtu.ll on ia64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51634 91177308-0d34-0410-b5e6-96231b3b80d8

loads should get chains.  THis helps but does not solve CodeGen/Generic/2003-05-27-phifcmpd.ll
on ia64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51633 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 2006-04-28-Sign-extend-bool.ll  for ia64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51632 91177308-0d34-0410-b5e6-96231b3b80d8

reindent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51631 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the encoding for two more "rm" instructions that were using MRMSrcReg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51630 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed X86 encoding error CVTPS2PD and CVTPD2PS when the source operand
is a memory location

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51626 91177308-0d34-0410-b5e6-96231b3b80d8

Don't attempt to create VZEXT_LOAD out of an extload.  This an issue where the
code generator would do something like this:

f64 = load f32 <anyext>, f32mem
v2f64 = insertelt undef, %0, 0
v2f64 = insertelt %1, 0.0, 1

into

v2f64 = vzext_load f32mem

which on x86 is movsd, when you really wanted a cvtss2sd/movsd pair.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51624 91177308-0d34-0410-b5e6-96231b3b80d8

use space insted of tabs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51623 91177308-0d34-0410-b5e6-96231b3b80d8

fix infinite recursion if a global's initializer references the global

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51617 91177308-0d34-0410-b5e6-96231b3b80d8

Make Use::init() private, it breaks invariants, clients should assign or use set().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51615 91177308-0d34-0410-b5e6-96231b3b80d8

Use Function::getEntryBlock instead of Function::begin, for clarity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51613 91177308-0d34-0410-b5e6-96231b3b80d8

Print debug output when any edge becomes executable, including
the first visited edge.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51612 91177308-0d34-0410-b5e6-96231b3b80d8

Add instcombine after global optimizations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51611 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a word-o and add an explicit keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51610 91177308-0d34-0410-b5e6-96231b3b80d8

Incorporated feedback: Check that the implicitly defined operands aren't used
before deleting the instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51609 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up whitespace in ConstantRange::print output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51606 91177308-0d34-0410-b5e6-96231b3b80d8

Use IPSCCPPass instead of IPConstantPropagationPass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51605 91177308-0d34-0410-b5e6-96231b3b80d8

Specify a target so that this tests tests what it's intended to test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51600 91177308-0d34-0410-b5e6-96231b3b80d8

Make this test independent of the target-triple; the stack alignment
is specifically what this test depends on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51599 91177308-0d34-0410-b5e6-96231b3b80d8

back out last commit: The .cpp file for a module should include its corresponding header first, even if redundant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51598 91177308-0d34-0410-b5e6-96231b3b80d8

Use an explicit llvm:: prefix in the STATISTIC macro, so STATISTIC can still be
used when "using namespace llvm" is not in effect.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51592 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some constructs that gcc-4.4 warns about.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51591 91177308-0d34-0410-b5e6-96231b3b80d8

prune unneeded #includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51590 91177308-0d34-0410-b5e6-96231b3b80d8

remove unneeded reinterpret_casts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51589 91177308-0d34-0410-b5e6-96231b3b80d8

We have the correct headers included to know that BB isa Value. No reinterpret_cast necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51588 91177308-0d34-0410-b5e6-96231b3b80d8

Remove ADCE from the optimization pipeline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51581 91177308-0d34-0410-b5e6-96231b3b80d8

InequalityGraph::node() can create new nodes, invalidating iterators across
the set of nodes. Fix makeEqual to handle this by creating the new node first
then iterating across them second.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51573 91177308-0d34-0410-b5e6-96231b3b80d8

Grammaro.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51572 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate calls to deprecated Use::init() interface

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51570 91177308-0d34-0410-b5e6-96231b3b80d8

Whoops -- forgot PR reference on this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51569 91177308-0d34-0410-b5e6-96231b3b80d8

The Linux ABI emits an extra "movl %esp, %ebp" in function prologue and
sometimes a "mov %ebp, %esp" in the epilogue.

Force these tests that rely on counting 'mov' to use i686-apple-darwin8.8.0
where they were written.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51568 91177308-0d34-0410-b5e6-96231b3b80d8

Factor code to copy global value attributes like
the section or the visibility from one global
value to another: copyAttributesFrom.  This is
particularly useful for duplicating functions:
previously this was done by explicitly copying
each attribute in turn at each place where a
new function was created out of an old one, with
the result that obscure attributes were regularly
forgotten (like the collector or the section).
Hopefully now everything is uniform and nothing
is forgotten.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51567 91177308-0d34-0410-b5e6-96231b3b80d8

Use a DenseMap instead of an std::map, speeding up the testcase in PR2368 by about a third.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51565 91177308-0d34-0410-b5e6-96231b3b80d8

The enabling of remat in 2-address conversion breaks this test:

Running /Users/void/llvm/llvm.src/test/CodeGen/X86/dg.exp ...
FAIL: /Users/void/llvm/llvm.src/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll
Failed with exit(1) at line 1
while running: llvm-as < /Users/void/llvm/llvm.src/test/CodeGen/X86/2007-11-30-LoadFolding-Bug.ll | llc -march=x86 -mattr=+sse2 -stats |&  grep {1 .*folded into instructions}
child process exited abnormally

Make this conditional for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51563 91177308-0d34-0410-b5e6-96231b3b80d8

A problem that's exposed when machine LICM is enabled. Consider this code:

LBB1_3:   # bb
...
        xorl    %ebp, %ebp
        subl    (%ebx), %ebp
...
        incl    %ecx
        cmpl    %edi, %ecx
        jl      LBB1_3  # bb

Whe using machine LICM, LLVM converts it into:

        xorl %esi, %esi
LBB1_3: # bb
...
        movl    %esi, %ebp
        subl    (%ebx), %ebp
...
        incl    %ecx
        cmpl    %edi, %ecx
        jl      LBB1_3  # bb

Two address conversion inserts the copy instruction. However, it's cheaper to
rematerialize it, and remat helps reduce register pressure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51562 91177308-0d34-0410-b5e6-96231b3b80d8

Use {} instead of "" in RUN lines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51561 91177308-0d34-0410-b5e6-96231b3b80d8

Don't treat values as signed when looking at loop steppings in HowForToNonZero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51560 91177308-0d34-0410-b5e6-96231b3b80d8

"ret (constexpr)" can't be folded into a Constant. Add a method to
Analysis/ConstantFolding to fold ConstantExpr's, then make instcombine use it
to try to use targetdata to fold constant expressions on void instructions.

Also extend the icmp(inttoptr, inttoptr) folding to handle the case where
int size != ptr size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51559 91177308-0d34-0410-b5e6-96231b3b80d8

Create archives with the same permissions are ar.

Patch by Mikael Lepistö.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51540 91177308-0d34-0410-b5e6-96231b3b80d8

Fix win32 when no bison is around, PR2331, patch by 'hume'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51539 91177308-0d34-0410-b5e6-96231b3b80d8

Add FreeBSD/PPC support, patch by Marcel Moolenaar!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51538 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a serious brain-o.  Obviously no-one reviewed my patch :(

This fixes PR2359

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51536 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR2358 by resolving calls with undef arguments to overdefined.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51535 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate x86.sse2.punpckh.qdq and x86.sse2.punpckl.qdq.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51533 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate x86.sse2.movs.d, x86.sse2.shuf.pd, x86.sse2.unpckh.pd, and x86.sse2.unpckl.pd intrinsics. These will be lowered into shuffles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51531 91177308-0d34-0410-b5e6-96231b3b80d8

Tweak how ConstantFP80Ty constants are output
so that gcc doesn't warn about them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51529 91177308-0d34-0410-b5e6-96231b3b80d8

Put initialized const weak objects into correct
sections on ppc32 darwin.  g++.dg/abi/key2.C

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51527 91177308-0d34-0410-b5e6-96231b3b80d8

This is done.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51526 91177308-0d34-0410-b5e6-96231b3b80d8

New loadl_pd and loadh_pd tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51525 91177308-0d34-0410-b5e6-96231b3b80d8

Autoupgrade x86.sse2.loadh.pd and x86.sse2.loadl.pd.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51523 91177308-0d34-0410-b5e6-96231b3b80d8

Remove x86.sse2.loadh.pd and x86.sse2.loadl.pd. These will be lowered into load and shuffle instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51522 91177308-0d34-0410-b5e6-96231b3b80d8

Remove x86.sse2.loadh.pd and x86.sse2.loadl.pd. These will be lowered into load and shuffle instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51521 91177308-0d34-0410-b5e6-96231b3b80d8

Document common linkage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51517 91177308-0d34-0410-b5e6-96231b3b80d8

Revert 51440 as it breaks a bunch of PIC tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51513 91177308-0d34-0410-b5e6-96231b3b80d8

label is a first-class type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51512 91177308-0d34-0410-b5e6-96231b3b80d8

It turns out there are only 3 non-first-class type kinds left now, so
it's simpler for isFirstClassType to use a negative test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51511 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a bogus svn:mime-type property.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51510 91177308-0d34-0410-b5e6-96231b3b80d8

Update the description of first-class types to reflect that
structs and arrays are now first-class. And fix a sentance
fragment in the insertvalue description. Thanks to Chris
for pointing these out!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51506 91177308-0d34-0410-b5e6-96231b3b80d8

Don't silently truncate array extents to 32 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51505 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missed CommonLinkage check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51503 91177308-0d34-0410-b5e6-96231b3b80d8

Use movlps / movhps to modify low / high half of 16-byet memory location.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51501 91177308-0d34-0410-b5e6-96231b3b80d8

Remove lingering references to .llx and .tr in the tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51500 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up BasicBlock::getFirstNonPHI, and change a bunch of places to
use it instead of duplicating its functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51499 91177308-0d34-0410-b5e6-96231b3b80d8