switch the DomTreeNodes and IDoms maps in idom/postidom to a
DenseMap instead of an std::map.  This speeds up postdomtree
by about 25% and domtree by about 23%.  It also speeds up clients,
for example, domfrontier by 11%, mem2reg by 4% and ADCE by 6%.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40826 91177308-0d34-0410-b5e6-96231b3b80d8

rewrite the code used to construct pruned SSA form with the IDF method.
In the old way, we computed and inserted phi nodes for the whole IDF of
the definitions of the alloca, then computed which ones were dead and
removed them.

In the new method, we first compute the region where the value is live,
and use that information to only insert phi nodes that are live.  This
eliminates the need to compute liveness later, and stops the algorithm
from inserting a bunch of phis which it then later removes.

This speeds up the testcase in PR1432 from 2.00s to 0.15s (14x) in a
release build and 6.84s->0.50s (14x) in a debug build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40825 91177308-0d34-0410-b5e6-96231b3b80d8

Factor out a whole bunch of code into it's own method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40824 91177308-0d34-0410-b5e6-96231b3b80d8

Use getNumPreds(BB) instead of computing them manually. This is a very small but
measurable speedup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40823 91177308-0d34-0410-b5e6-96231b3b80d8

Change the rename pass to be "tail recursive", only adding N-1 successors
to the worklist, and handling the last one with a 'tail call'.  This speeds
up PR1432 from 2.0578s to 2.0012s (2.8%)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40822 91177308-0d34-0410-b5e6-96231b3b80d8

cache computation of #preds for a BB.  This speeds up
mem2reg from 2.0742->2.0522s on PR1432.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40821 91177308-0d34-0410-b5e6-96231b3b80d8

reserve operand space for phi nodes when we insert them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40820 91177308-0d34-0410-b5e6-96231b3b80d8

use continue to avoid nesting, no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40819 91177308-0d34-0410-b5e6-96231b3b80d8

Promoting allocas with the 'single store' fastpath is
faster than with the 'local to a block' fastpath.  This speeds
up PR1432 from 2.1232 to 2.0686s (2.6%)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40818 91177308-0d34-0410-b5e6-96231b3b80d8

When PromoteLocallyUsedAllocas promoted allocas, it didn't remember
to increment NumLocalPromoted, and didn't actually delete the
dead alloca, leading to an extra iteration of mem2reg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40817 91177308-0d34-0410-b5e6-96231b3b80d8

std::map -> DenseMap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40816 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up comments, fix up some confusing code logic.
Predsimplify fails llvm-gcc bootstrap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40815 91177308-0d34-0410-b5e6-96231b3b80d8

fix a logic bug where we wouldn't promote single store allocas if the
stored value was a non-instruction value.  Doh.

This increase the # single store allocas from 8982 to 9026, and
speeds up mem2reg on the testcase in PR1432 from 2.17 to 2.13s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40813 91177308-0d34-0410-b5e6-96231b3b80d8

When we do the single-store optimization, delete both the store
and the alloca so they don't get reprocessed.

This speeds up PR1432 from 2.20s to 2.17s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40812 91177308-0d34-0410-b5e6-96231b3b80d8

Three improvements:
  1. Check for revisiting a block before checking domination, which is faster.
  2. If the stored value isn't an instruction, we don't have to check for domination.
  3. If we have a value used in the same block more than once, make sure to remove the
     block from the UsingBlocks vector.  Not doing so forces us to go through the slow
     path for the alloca.

The combination of these improvements increases the number of allocas on the fastpath
from 8935 to 8982 on PR1432.  This speeds it up from 2.90s to 2.20s (31%)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40811 91177308-0d34-0410-b5e6-96231b3b80d8

switch from using a std::set to using a SmallPtrSet.  This speeds up the
testcase in PR1432 from 6.33s to 2.90s (2.22x)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40810 91177308-0d34-0410-b5e6-96231b3b80d8

In mem2reg, when handling the single-store case, make sure to remove
a using block from the list if we handle it.  Not doing this caused us
to not be able to promote (with the fast path) allocas which have uses (whoops).

This increases the # allocas hitting this fastpath from 4042 to 8935 on the
testcase in PR1432, speeding up mem2reg by 2.6x

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40809 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerating.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40808 91177308-0d34-0410-b5e6-96231b3b80d8

This is the patch to provide clean intrinsic function overloading support in LLVM. It cleans up the intrinsic definitions and generally smooths the process for more complicated intrinsic writing. It will be used by the upcoming atomic intrinsics as well as vector and float intrinsics in the future.

This also changes the syntax for llvm.bswap, llvm.part.set, llvm.part.select, and llvm.ct* intrinsics. They are automatically upgraded by both the LLVM ASM reader and the bitcode reader. The test cases have been updated, with special tests added to ensure the automatic upgrading is supported.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40807 91177308-0d34-0410-b5e6-96231b3b80d8

split rewriting of single-store allocas into its own
method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40806 91177308-0d34-0410-b5e6-96231b3b80d8

refactor some code to shrink PromoteMem2Reg::run a bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40805 91177308-0d34-0410-b5e6-96231b3b80d8

add a typedef, no other change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40804 91177308-0d34-0410-b5e6-96231b3b80d8

avoid an unneeded vector copy.  This speeds up mem2reg on the testcase
in PR1432 by 6%

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40803 91177308-0d34-0410-b5e6-96231b3b80d8

make RenamePassWorkList a local var instead of an ivar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40802 91177308-0d34-0410-b5e6-96231b3b80d8

Make x86 long double alignment 32 for everything but
Darwin (which makes size within a struct==96)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40796 91177308-0d34-0410-b5e6-96231b3b80d8

long double patch 3 of N.  Add to MVT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40793 91177308-0d34-0410-b5e6-96231b3b80d8

long double patch 2 of N.  Handle it in TargetData.
(I've tried to get the info right for all targets,
but I'm not expert on all of them - check yours.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40792 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a subtle miscompilation.  This allows 197.parser to be compiled correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40791 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a subtle iterator invalidation bug in a recursive algorithm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40776 91177308-0d34-0410-b5e6-96231b3b80d8

Prepare for "core" website.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40775 91177308-0d34-0410-b5e6-96231b3b80d8

Long double, part 1 of N.  Support in IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40774 91177308-0d34-0410-b5e6-96231b3b80d8

add an observation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40772 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an accidental commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40758 91177308-0d34-0410-b5e6-96231b3b80d8

More explicit keywords.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40757 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the alignment requirements of several unpck and shuf instructions.
Generalize isPSHUFDMask and add a unary SHUFPD pattern so that SHUFPD's
memory operand alignment can be tested as well, with a fix to avoid
breaking MMX's use of isPSHUFDMask.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40756 91177308-0d34-0410-b5e6-96231b3b80d8

Fix pastos in vector arithmetic intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40754 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 col. violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40751 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 col. violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40750 91177308-0d34-0410-b5e6-96231b3b80d8

Fix 80 col. violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40749 91177308-0d34-0410-b5e6-96231b3b80d8

don't redefine a parameter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40748 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug that was causing several miscompilations on SPEC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40746 91177308-0d34-0410-b5e6-96231b3b80d8

Implement review feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40745 91177308-0d34-0410-b5e6-96231b3b80d8

I don't have time to restore this functionality right now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40743 91177308-0d34-0410-b5e6-96231b3b80d8

Replacing a cast with another one does not reduce the number of
casts in the input.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40741 91177308-0d34-0410-b5e6-96231b3b80d8

Reduced testcase for PR1594

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40740 91177308-0d34-0410-b5e6-96231b3b80d8

Disable an xform that causes an infinite loop.  This fixes PR1594

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40739 91177308-0d34-0410-b5e6-96231b3b80d8

wrap some long lines.  Major offenders that are left include
gvn, gvnpre, dse, and predsimplify.  To see these, use:

  make check-line-length

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40738 91177308-0d34-0410-b5e6-96231b3b80d8

Update dominator info for the middle blocks created while spliting
exit edge to preserve LCSSA.

Fix dominance frontier update during loop unswitch. This fixes PR 1589, again

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40737 91177308-0d34-0410-b5e6-96231b3b80d8

Mark the SSE and MMX load instructions that
X86InstrInfo::isReallyTriviallyReMaterializable knows how to handle
with the isReMaterializable flag so that it is given a chance to handle
them. Without hoisting constant-pool loads from loops this isn't very
visible, though it does keep CodeGen/X86/constant-pool-remat-0.ll from
making a copy of the constant pool on the stack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40736 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance instcombine to be more aggressive about folding casts of
operations of casts.  This implements InstCombine/zext-fold.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40726 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug spotted by Chris.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40725 91177308-0d34-0410-b5e6-96231b3b80d8

Add extend and extOrTrunc methods that do sign or zero extension depending on whether the integer is signed or not

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40724 91177308-0d34-0410-b5e6-96231b3b80d8

Switch some multiplication instructions over to the new scheme for testing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40723 91177308-0d34-0410-b5e6-96231b3b80d8

Do not emit copies for physical register output if it's not used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40722 91177308-0d34-0410-b5e6-96231b3b80d8

Fix test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40721 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1575 and test/Transforms/CondProp/2007-08-01-InvalidRead.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40720 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust for new CallInst constructor interface.
This fixes test/Feature/llvm2cpp.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40714 91177308-0d34-0410-b5e6-96231b3b80d8

Style police: Expand the tabs to spaces!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40712 91177308-0d34-0410-b5e6-96231b3b80d8

Teach BasicAA about noalias parameter attributes, but do it correctly this time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40711 91177308-0d34-0410-b5e6-96231b3b80d8

Instead of adding copyfromreg's to handle physical definitions. Now isel can
simply specify them as results and let scheduledag handle them. That
is, instead of
SDOperand Flag = DAG.getTargetNode(Opc, MVT::i32, MVT::Flag, ...)
SDOperand Result = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, Flag)

Just write:
SDOperand Result = DAG.getTargetNode(Opc, MVT::i32, MVT::i32, ...)

And let scheduledag emit the move from X86::EAX to a virtual register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40710 91177308-0d34-0410-b5e6-96231b3b80d8

Added TargetInstrDescriptor::numDefs - num of results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40709 91177308-0d34-0410-b5e6-96231b3b80d8

Can't handle offset and scale if rip-relative addressing is to be used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40703 91177308-0d34-0410-b5e6-96231b3b80d8

Mac OS X X86-64 low 4G address not available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40702 91177308-0d34-0410-b5e6-96231b3b80d8

Mac OS X X86-64 low 4G address not available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40701 91177308-0d34-0410-b5e6-96231b3b80d8

Undo previous check-in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40698 91177308-0d34-0410-b5e6-96231b3b80d8

Some out operands were incorrectly specified as input operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40697 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure these tests pass for the right reasons (verifier error, rather than
failure to assemble).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40696 91177308-0d34-0410-b5e6-96231b3b80d8

Update dominator info for the middle blocks created while spliting
exit edge to preserve LCSSA.

Fix dominance frontier update during loop unswitch. This fixes PR 1589.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40695 91177308-0d34-0410-b5e6-96231b3b80d8

Make non-local memdep not be recursive, and fix a bug on 403.gcc that this exposed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40692 91177308-0d34-0410-b5e6-96231b3b80d8

Missing Requires.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40691 91177308-0d34-0410-b5e6-96231b3b80d8

New test. Bogus implicit-def prevented a copy from being coalesced.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40690 91177308-0d34-0410-b5e6-96231b3b80d8

Be more precise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40689 91177308-0d34-0410-b5e6-96231b3b80d8

Bugs: missing partial uses and redundant partial defs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40688 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test for the load/store alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40687 91177308-0d34-0410-b5e6-96231b3b80d8

Expand unaligned loads/stores when the target doesn't support them. (PR1548)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40682 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up C++ restrict test cases and add a test for restrict qualified methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40681 91177308-0d34-0410-b5e6-96231b3b80d8

these tests aren't xfailed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40680 91177308-0d34-0410-b5e6-96231b3b80d8

we're now handling this right :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40675 91177308-0d34-0410-b5e6-96231b3b80d8

Forgot to update these files for the FastDSE changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40674 91177308-0d34-0410-b5e6-96231b3b80d8

More explicit keywords.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40673 91177308-0d34-0410-b5e6-96231b3b80d8

Change a .size directive to use a tab instead of a space, for consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40672 91177308-0d34-0410-b5e6-96231b3b80d8

Make ImmutablePass::runOnModule non-virtual, since it is not
intended to be overridden.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40671 91177308-0d34-0410-b5e6-96231b3b80d8

Rename FastDSE to just DSE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40668 91177308-0d34-0410-b5e6-96231b3b80d8

Move FastDSE in to DeadStoreElimination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40667 91177308-0d34-0410-b5e6-96231b3b80d8

Remove old DSE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40666 91177308-0d34-0410-b5e6-96231b3b80d8

Update generated files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40663 91177308-0d34-0410-b5e6-96231b3b80d8

New CallInst interface to address GLIBCXX_DEBUG errors caused by
indexing an empty std::vector.

Updates to all clients.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40660 91177308-0d34-0410-b5e6-96231b3b80d8

Indexed loads each has 2 outputs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40658 91177308-0d34-0410-b5e6-96231b3b80d8

Requires SSE2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40657 91177308-0d34-0410-b5e6-96231b3b80d8

Don't let the memory allocator outsmart GVN. ;-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40655 91177308-0d34-0410-b5e6-96231b3b80d8

simpleregistercoalescing -> regcoalescing. It's too long for me to handle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40654 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a failure I accidentally caused in my last commit by mishandling the
removal of redundant phis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40650 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in GetKnownAlignment of packed structs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40649 91177308-0d34-0410-b5e6-96231b3b80d8

Change the x86 assembly output to use tab characters to separate the
mnemonics from their operands instead of single spaces. This makes the
assembly output a little more consistent with various other compilers
(f.e. GCC), and slightly easier to read. Also, update the regression
tests accordingly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40648 91177308-0d34-0410-b5e6-96231b3b80d8

Fix GLIBCXX_DEBUG error owing to dereference of end iterator.  There's
no guarantee that an instruction returned by getDependency exists in
the maps.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40647 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a misoptimization in aha.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40642 91177308-0d34-0410-b5e6-96231b3b80d8

Use SCEVExpander::InsertCastOfTo instead of calling new IntToPtrInst
directly, because the insert point used by the SCEVExpander may vary
from what LSR originally computes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40641 91177308-0d34-0410-b5e6-96231b3b80d8

Add note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40638 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40636 91177308-0d34-0410-b5e6-96231b3b80d8

Revert overly aggressive interpretation of noalias

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40635 91177308-0d34-0410-b5e6-96231b3b80d8