A bunch of fixes to the BigBlock allocator improve compile-time by ~20%
and code quality by ~2% on my tests.

A big thank you to Roman Levenstein for this patch! See
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070618/050717.html
for more details.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37724 91177308-0d34-0410-b5e6-96231b3b80d8

fix Transforms/Inline/2007-06-25-WeakInline.ll by not inlining functions
with weak linkage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37723 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase, the inliner shouldn't inline this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37722 91177308-0d34-0410-b5e6-96231b3b80d8

Use the built-in postorder iterators rather than computing a postorder walk by hand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37721 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1525:
Use a better determinator for identifying constant array initializers that
are or are not zero terminated and generate code appropriately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37720 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize MVT::ValueType and associated functions to be able to represent
extended vector types. Remove the special SDNode opcodes used for pre-legalize
vector operations, and the special MVT::Vector type used with them. Adjust
lowering and legalize to work with the normal SDNode kinds instead, and to
use the normal MVT functions to work with vector types instead of using the
two special operands that the pre-legalize nodes held.

This allows pre-legalize and post-legalize DAGs, and the code that operates
on them, to be more consistent. Pre-legalize vector operators can be handled
more consistently with scalar operators. And, -view-dag-combine1-dags and
-view-legalize-dags now look prettier for vector code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37719 91177308-0d34-0410-b5e6-96231b3b80d8

Make minor adjustments to whitespace and comments to reduce differences
between SSE1 instructions and their respective SSE2 analogues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37718 91177308-0d34-0410-b5e6-96231b3b80d8

Fix loadv2i32 to be loadv4i32, though it isn't actually used anywhere yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37717 91177308-0d34-0410-b5e6-96231b3b80d8

Say AT&T instead of Intel in the comments for AT&T support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37716 91177308-0d34-0410-b5e6-96231b3b80d8

Provide hook for alloca on VCPP. Patch by Scott Graham

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37715 91177308-0d34-0410-b5e6-96231b3b80d8

1) Fix an issue with non-deterministic iteration order in phi_translate
2) Remove some maximal-set computing code that is no longer used.
3) Use a post-order CFG traversal to compute ANTIC_IN instead of a postdom traversal.
This causes the ANTIC_IN calculation to converge much faster.  Thanks to Daniel Berlin for suggesting this.

With this patch, the time to optimize 403.gcc decreased from 17.5s to 7.5s, and Anton's huge
testcase decreased from 62 minutes to 38 seconds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37714 91177308-0d34-0410-b5e6-96231b3b80d8

Fix value ranges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37713 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a silly mistake that was causing failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37712 91177308-0d34-0410-b5e6-96231b3b80d8

Rename variables to expose the fact that this test is failing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37711 91177308-0d34-0410-b5e6-96231b3b80d8

Remove tabs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37710 91177308-0d34-0410-b5e6-96231b3b80d8

Remove use of ETForest. Also cleaned up issues around unreachable basic
blocks, and optimizing within one basic block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37709 91177308-0d34-0410-b5e6-96231b3b80d8

Rework topo_sort so eliminate some behavior that scaled terribly.  This reduces the time to optimize 403.gcc from 18.2s to 17.5s,
and has an even larger effect on larger testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37708 91177308-0d34-0410-b5e6-96231b3b80d8

Perform fewer set insertions while calculating ANTIC_IN.  This reduces the amount of time to optimize 403.gcc from 21.9s to 18.2s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37707 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some code that I was using for collecting performance information that should not have been committed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37706 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37705 91177308-0d34-0410-b5e6-96231b3b80d8

Move ComputeMaskedBits, MaskedValueIsZero, and ComputeNumSignBits from
TargetLowering to SelectionDAG so that they have more convenient
access to the current DAG, in preparation for the ValueType routines
being changed from standalone functions to members of SelectionDAG for
the pre-legalize vector type changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37704 91177308-0d34-0410-b5e6-96231b3b80d8

check in the BigBlock local register allocator

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37703 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid excessive calls to find_leader when calculating AVAIL_OUT.  This reduces the time to optimize 403.gcc from 23.5s to 21.9s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37702 91177308-0d34-0410-b5e6-96231b3b80d8

std::set is really really terrible. Switch to SmallPtrSet to reduce compile time. For Duraid's example. The overall isel time is reduced from 0.6255 sec to 0.1876 sec.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37701 91177308-0d34-0410-b5e6-96231b3b80d8

Quote complex names for Darwin X86 and ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37700 91177308-0d34-0410-b5e6-96231b3b80d8

Reserve space in vectors before topologically sorting into them.  This improves the time to optimize 403.gcc from 28s to 23.5s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37699 91177308-0d34-0410-b5e6-96231b3b80d8

Make a bunch of optimizations for compile time to GVNPRE, including smarter set unions, deferring blocks rather than computing maximal sets, and smarter use of sets.  With these enhancements, the time to optimize 273.perlbmk goes from 5.3s to 2.7s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37698 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in SmallPtrSet that was causing GVNPRE to enter an infinite loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37697 91177308-0d34-0410-b5e6-96231b3b80d8

Two changes:
 1. Make SmallPtrSet::erase faster in the small case by replacing a memmove
    with a pointer copy.
 2. Fix a bug where the null terminator at the end of the array in the small
    case was not copied

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37696 91177308-0d34-0410-b5e6-96231b3b80d8

CallGraphSCCPass manager may require other passes.
Use schedulePass() to accomodate these requirement instead of directly
assigning a manager to new CallGraph PassManager.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37695 91177308-0d34-0410-b5e6-96231b3b80d8

Have internal df_iterator's use SmallPtrSet instead of std::set.  This provides compile time speedups to any pass using df_iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37694 91177308-0d34-0410-b5e6-96231b3b80d8

Fix tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37693 91177308-0d34-0410-b5e6-96231b3b80d8

Significantly improve the documentation of the instcombine divide/compare
transformation.  Also, keep track of which end of the integer interval overflows
occur on.  This fixes Transforms/InstCombine/2007-06-21-DivCompareMiscomp.ll
and rdar://5278853, a miscompilation of perl.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37692 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase miscompiled by instcombine, reduced from perl

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37691 91177308-0d34-0410-b5e6-96231b3b80d8

Change lots of sets from std::set to SmallPtrSet.  This reduces the time required to optimize 253.perlbmk from 10.9s to 5.3s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37690 91177308-0d34-0410-b5e6-96231b3b80d8

Move code to update dominator information after basic block is split
from LoopSimplify.cpp to Dominator.cpp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37689 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up ValueType names in comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37688 91177308-0d34-0410-b5e6-96231b3b80d8

Rename TargetLowering::getNumElements and friends to
TargetLowering::getNumRegisters and similar, to avoid confusion with
the actual number of elements for vector types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37687 91177308-0d34-0410-b5e6-96231b3b80d8

New tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37686 91177308-0d34-0410-b5e6-96231b3b80d8

Xforms:
(add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
(sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37685 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate a redundant check.  This speeds up optimization of 253.perlbmk from 13.5 seconds to 10.9 seconds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37683 91177308-0d34-0410-b5e6-96231b3b80d8

Comment-ize the functions in GVNPRE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37681 91177308-0d34-0410-b5e6-96231b3b80d8

refactor a bunch of code out of visitICmpInstWithInstAndIntCst into its own
routine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37679 91177308-0d34-0410-b5e6-96231b3b80d8

Split runOnFunction into many smaller functions.  This make it easier to get accurate performance analysis of GVNPRE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37678 91177308-0d34-0410-b5e6-96231b3b80d8

Modify deleting global variable with an even easier way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37676 91177308-0d34-0410-b5e6-96231b3b80d8

Update AnalysisGroup documentation to document restriction that allows
only one ImmutablePass in a group.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37675 91177308-0d34-0410-b5e6-96231b3b80d8

Add blurb on deleting global variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37674 91177308-0d34-0410-b5e6-96231b3b80d8

Make GVNPRE accurate report whether it modified the function or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37673 91177308-0d34-0410-b5e6-96231b3b80d8

Added some if-conversion tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37672 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of an unneeded helper function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37670 91177308-0d34-0410-b5e6-96231b3b80d8

Be more conservative of duplicating blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37669 91177308-0d34-0410-b5e6-96231b3b80d8

Use a DenseMap instead of an std::map for the value numbering.  This reduces the time to optimize lencod on a PPC Debug build from ~300s to ~140s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37668 91177308-0d34-0410-b5e6-96231b3b80d8

Make dependsOnInvoke much more specific in what it tests, which in turn make it much faster to run.  This reduces the time to optimize lencondwith a debug build on PPC from ~450s to ~300s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37667 91177308-0d34-0410-b5e6-96231b3b80d8

Moved Inliner.h to include/llvm/Transforms/IPO/InlinerPass.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37666 91177308-0d34-0410-b5e6-96231b3b80d8

Inliner pass header file was moved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37665 91177308-0d34-0410-b5e6-96231b3b80d8

Move inliner pass header file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37664 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid if-converting simple block that ends with unconditional branch or fallthrough unless it branches / falls to the 'false' block. Not profitable, may end up increasing code size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37660 91177308-0d34-0410-b5e6-96231b3b80d8

Allow predicated immediate ARM to ARM calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37659 91177308-0d34-0410-b5e6-96231b3b80d8

Fix pr1448

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37658 91177308-0d34-0410-b5e6-96231b3b80d8

Rename ScalarEvolution::deleteInstructionFromRecords to
deleteValueFromRecords and loosen the types to all it to accept
Value* instead of just Instruction*, since this is what
ScalarEvolution uses internally anyway. This allows more flexibility
for future uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37657 91177308-0d34-0410-b5e6-96231b3b80d8

Pass a SelectionDAG into SDNode::dump everywhere it's used, in prepration
for needing the DAG node to print pre-legalize extended value types, and
to get better debug messages with target-specific nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37656 91177308-0d34-0410-b5e6-96231b3b80d8

Handle constants in phi nodes properly.  This fixes test/Transforms/GVNPRE/2007-06-18-ConstantInPhi.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37655 91177308-0d34-0410-b5e6-96231b3b80d8

silence warning when assertions are disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37654 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for instances where a constant only occurs as an operand to a phi node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37653 91177308-0d34-0410-b5e6-96231b3b80d8

don't use binutils 2.17

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37652 91177308-0d34-0410-b5e6-96231b3b80d8

document and hide two options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37651 91177308-0d34-0410-b5e6-96231b3b80d8

describe an argument, hide it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37650 91177308-0d34-0410-b5e6-96231b3b80d8

silence a bogus warning Duraid ran into.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37649 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new testcase for memory corruption issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37648 91177308-0d34-0410-b5e6-96231b3b80d8

Be careful to erase values from all of the appropriate sets when they're not needed anymore.  This fixes a few more memory-related issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37647 91177308-0d34-0410-b5e6-96231b3b80d8

Remember to clear the maximal sets between functions.

Thanks to Nicholas for valgrinding this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37646 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor GVNPRE to use a much smart method of uniquing value sets, and centralize a lot of the value numbering information.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37645 91177308-0d34-0410-b5e6-96231b3b80d8

Replace M_REMATERIALIZIBLE and the newly-added isOtherReMaterializableLoad
with a general target hook to identify rematerializable instructions. Some
instructions are only rematerializable with specific operands, such as loads
from constant pools, while others are always rematerializable. This hook
allows both to be identified as being rematerializable with the same
mechanism.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37644 91177308-0d34-0410-b5e6-96231b3b80d8

Replace TargetInstrInfo::CanBeDuplicated() with a M_NOT_DUPLICABLE bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37643 91177308-0d34-0410-b5e6-96231b3b80d8

Replace CanBeDuplicated() with a M_NOT_DUPLICABLE bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37642 91177308-0d34-0410-b5e6-96231b3b80d8

If a function is vararg, never pass inreg arguments in registers.  Thanks to
Anton for half of this patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37641 91177308-0d34-0410-b5e6-96231b3b80d8

add isVarArg to CCState

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37640 91177308-0d34-0410-b5e6-96231b3b80d8

add isvararg tracking to CCState

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37639 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for PR1513

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37638 91177308-0d34-0410-b5e6-96231b3b80d8

New test case: identity operation of RHS / LHS of a VECTOR_SHUFFLE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37637 91177308-0d34-0410-b5e6-96231b3b80d8

Look for VECTOR_SHUFFLE that's identity operation on either LHS or RHS. This can happen before DAGCombiner catches it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37636 91177308-0d34-0410-b5e6-96231b3b80d8

Update docs for llvm.var.annotation intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37635 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some fragile code wrt CFG edge updating.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37634 91177308-0d34-0410-b5e6-96231b3b80d8

Move CorrectExtraCFGEdges() from BranchFolding.cpp to a MachineBasicBlock method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37633 91177308-0d34-0410-b5e6-96231b3b80d8

Modify annotate intrinsic to take 2 additional args: file and line number.
Fix up Annotate attribute test case.
Fix DebugFilename.c test case to look for right file name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37632 91177308-0d34-0410-b5e6-96231b3b80d8

add helper

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37631 91177308-0d34-0410-b5e6-96231b3b80d8

Fix quotes in debug messages.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37630 91177308-0d34-0410-b5e6-96231b3b80d8

make ComputeTopDownOrdering significantly faster and use less stack space
by making it non-recursive

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37629 91177308-0d34-0410-b5e6-96231b3b80d8

In SCEVAddExpr::get, skip over any cast operands before looking for nested
add operands after constant operands. The recent change to recognize
sign-extend expressions caused this to be exposed more often.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37628 91177308-0d34-0410-b5e6-96231b3b80d8

Make BasicAliasAnalysis correctly register itself. Patch by Devang

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37627 91177308-0d34-0410-b5e6-96231b3b80d8

Make chain dependencies blue, in addition to being dashed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37626 91177308-0d34-0410-b5e6-96231b3b80d8

Define the pushq instruction for x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37625 91177308-0d34-0410-b5e6-96231b3b80d8

Add keyword explicit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37624 91177308-0d34-0410-b5e6-96231b3b80d8

Properly remove duplicate instructions as result of diamond if-conversion. Other bug fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37623 91177308-0d34-0410-b5e6-96231b3b80d8

Cache the results of dependsOnInvoke()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37622 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37621 91177308-0d34-0410-b5e6-96231b3b80d8

Don't perform an expensive check if it's not necessary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37620 91177308-0d34-0410-b5e6-96231b3b80d8

ensure we don't regress on these tests.  We generate aweful code in x86-32 for
these though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37619 91177308-0d34-0410-b5e6-96231b3b80d8

XFAILing until I can fix properly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37618 91177308-0d34-0410-b5e6-96231b3b80d8

Revert patch. It regresses:

define double @test2(i64 %A) {
   %B = bitcast i64 %A to double
   ret double %B
}

$ llvm-as < t.ll | llc -march=x86-64

before:

         .align  4
         .globl  _test2
_test2:
         movd %rdi, %xmm0
         ret

after:

_test2:
         subq $8, %rsp
         movq %rdi, (%rsp)
         movsd (%rsp), %xmm0
         addq $8, %rsp
         ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37617 91177308-0d34-0410-b5e6-96231b3b80d8