remove some weirdness that came from the LSR code that has
nothing to do with dead instruction elimination.  No tests in
dejagnu depend on this, so I don't know what it was needed for.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60202 91177308-0d34-0410-b5e6-96231b3b80d8

rewrite a big chunk of how DSE does recursive dead operand
elimination to use more modern infrastructure.  Also do a bunch
of small cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60201 91177308-0d34-0410-b5e6-96231b3b80d8

Scrap some boilerplate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60200 91177308-0d34-0410-b5e6-96231b3b80d8

Support multiple compilation graph definitions. Not terribly useful, but makes the code more generic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60199 91177308-0d34-0410-b5e6-96231b3b80d8

Add 'hidden' and 'really_hidden' option properties.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60198 91177308-0d34-0410-b5e6-96231b3b80d8

Documentation: clarify what is meant by 'multiple edges'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60197 91177308-0d34-0410-b5e6-96231b3b80d8

delete ErasePossiblyDeadInstructionTree, replacing uses of it with
RecursivelyDeleteTriviallyDeadInstructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60196 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify LoopStrengthReduce::DeleteTriviallyDeadInstructions by
making it use RecursivelyDeleteTriviallyDeadInstructions to do
the heavy lifting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60195 91177308-0d34-0410-b5e6-96231b3b80d8

enhance RecursivelyDeleteTriviallyDeadInstructions to make
PHIs dead if they are single-value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60194 91177308-0d34-0410-b5e6-96231b3b80d8

Enhance RecursivelyDeleteTriviallyDeadInstructions to optionally
return a list of deleted instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60193 91177308-0d34-0410-b5e6-96231b3b80d8

use continue to reduce indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60192 91177308-0d34-0410-b5e6-96231b3b80d8

remove doConstantPropagation and dceInstruction, they are just
wrappers around the interesting code and use an obscure iterator
abstraction that dates back many many years.

Move EraseDeadInstructions to Transforms/Utils and name it
RecursivelyDeleteTriviallyDeadInstructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60191 91177308-0d34-0410-b5e6-96231b3b80d8

simplify code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60190 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60189 91177308-0d34-0410-b5e6-96231b3b80d8

Also update the README.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60188 91177308-0d34-0410-b5e6-96231b3b80d8

Chris prefers icmp/select over udiv!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60187 91177308-0d34-0410-b5e6-96231b3b80d8

Add a synthetic missed optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60186 91177308-0d34-0410-b5e6-96231b3b80d8

Add a couple of missed optimizations on integer vectors. Multiply and divide
by 1, as well as multiply by -1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60182 91177308-0d34-0410-b5e6-96231b3b80d8

defensive patch: if CGP is merging a block with the entry block, make sure
it ends up being the entry block.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60180 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR3138: if we merge the entry block into another block, make sure to
move the other block back up into the entry position!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60179 91177308-0d34-0410-b5e6-96231b3b80d8

Silence a warning.

Despite changing the order of evaluation, this doesn't actually change the
meaning of the statement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60177 91177308-0d34-0410-b5e6-96231b3b80d8

fix build on some machines. thanks buildbot

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60175 91177308-0d34-0410-b5e6-96231b3b80d8

fix my previous commit r60064: compare strings instead of pointers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60174 91177308-0d34-0410-b5e6-96231b3b80d8

switch InstCombine::visitLoadInst to use
FindAvailableLoadedValue

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60169 91177308-0d34-0410-b5e6-96231b3b80d8

improve const correctness.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60168 91177308-0d34-0410-b5e6-96231b3b80d8

enhance FindAvailableLoadedValue to make use of AliasAnalysis
if it has it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60167 91177308-0d34-0410-b5e6-96231b3b80d8

move FindAvailableLoadedValue from JumpThreading to Transforms/Utils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60166 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of bogus "control may reach end of non-void function ‘...’ being
inlined" message.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60165 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60164 91177308-0d34-0410-b5e6-96231b3b80d8

Use the new MergeBasicBlockIntoOnlyPred function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60163 91177308-0d34-0410-b5e6-96231b3b80d8

move MergeBasicBlockIntoOnlyPred to Transforms/Utils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60162 91177308-0d34-0410-b5e6-96231b3b80d8

XFAil test due to reverting of patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60161 91177308-0d34-0410-b5e6-96231b3b80d8

rename ThreadBlock to ProcessBlock, since it does other things than
just simple threading.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60157 91177308-0d34-0410-b5e6-96231b3b80d8

Comment out code that isn't entirely correct.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60156 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed HTML closing tag, cleaned up some spacing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60153 91177308-0d34-0410-b5e6-96231b3b80d8

Removing redundant semicolons. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60149 91177308-0d34-0410-b5e6-96231b3b80d8

Make jump threading substantially more powerful, in the following ways:

1. Make it fold blocks separated by an unconditional branch.  This enables
   jump threading to see a broader scope.
2. Make jump threading able to eliminate locally redundant loads when they
   feed the branch condition of a block.  This frequently occurs due to
   reg2mem running.
3. Make jump threading able to eliminate *partially redundant* loads when
   they feed the branch condition of a block.  This is common in code with
   lots of loads and stores like C++ code and 255.vortex.

This implements thread-loads.ll and rdar://6402033.

Per the fixme's, several pieces of this should be moved into Transforms/Utils.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60148 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate a compile time warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60145 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid inserting noop's in the middle of a loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60141 91177308-0d34-0410-b5e6-96231b3b80d8

On x86 favors folding short immediate into some arithmetic operations (e.g. add, and, xor, etc.) because materializing an immediate in a register is expensive in turns of code size.
e.g.
movl 4(%esp), %eax
addl $4, %eax

is 2 bytes shorter than

movl $4, %eax
addl 4(%esp), %eax

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60139 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missing case in visitADD.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60137 91177308-0d34-0410-b5e6-96231b3b80d8

Add -march=x86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60135 91177308-0d34-0410-b5e6-96231b3b80d8

Add typedef to StringMapEntry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60134 91177308-0d34-0410-b5e6-96231b3b80d8

Disallow multiple edges.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60127 91177308-0d34-0410-b5e6-96231b3b80d8

Add x86-specific test for add-with-overflow intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60125 91177308-0d34-0410-b5e6-96231b3b80d8

Generate something sensible for an [SU]ADDO op when the overflow/carry flag is
the conditional for the BRCOND statement. For instance, it will generate:

    addl %eax, %ecx
    jo LOF

instead of

    addl %eax, %ecx
    ; About 10 instructions to compare the signs of LHS, RHS, and sum.
    jl LOF

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60123 91177308-0d34-0410-b5e6-96231b3b80d8

Turn on my codegen prepare heuristic by default.  It doesn't affect
performance in most cases on the Grawp tester, but does speed some
things up (like shootout/hash by 15%).  This also doesn't impact
compile time in a noticable way on the Grawp tester.

It also, of course, gets the testcase it was designed for right :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60120 91177308-0d34-0410-b5e6-96231b3b80d8

Small formatting change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60113 91177308-0d34-0410-b5e6-96231b3b80d8

Update to explain how ssp and sspreq attributes override each other.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60112 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60111 91177308-0d34-0410-b5e6-96231b3b80d8

Cosmetic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60110 91177308-0d34-0410-b5e6-96231b3b80d8

Check that running the DAG combiner between type
and operation legalization does something useful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60108 91177308-0d34-0410-b5e6-96231b3b80d8

Describe some more options in the man page.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60105 91177308-0d34-0410-b5e6-96231b3b80d8

ignore build dirs and generated files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60104 91177308-0d34-0410-b5e6-96231b3b80d8

Allow custom lowering of ADDE/ADDC/SUBE/SUBC operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60102 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the -I option (llvmc -I dir1 -I dir2 didn't work).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60101 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor Tools.td to remove repetition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60100 91177308-0d34-0410-b5e6-96231b3b80d8

Small fix: the error message was incorrect in some cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60099 91177308-0d34-0410-b5e6-96231b3b80d8

Emit declaration for globals and externs.
Custom lower AND, OR, XOR bitwise operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60098 91177308-0d34-0410-b5e6-96231b3b80d8

Fish kill flag annotations in PUSH instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60095 91177308-0d34-0410-b5e6-96231b3b80d8

LiveRanges are represented as half-open ranges. Fix the findLiveInMBBs code
and the LiveInterval.h top-level comment and accordingly. This fixes blocks
having spurious live-in registers in boundary cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60092 91177308-0d34-0410-b5e6-96231b3b80d8

teach the new heuristic how to handle inline asm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60088 91177308-0d34-0410-b5e6-96231b3b80d8

Disable -loop-index-split for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60087 91177308-0d34-0410-b5e6-96231b3b80d8

Add 'tell' method to raw_fd_ostream that clients can use to query the current location in the file the stream is writing to.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60085 91177308-0d34-0410-b5e6-96231b3b80d8

Improve ValueAlreadyLiveAtInst with a cheap and dirty, but effective
heuristic: the value is already live at the new memory operation if
it is used by some other instruction in the memop's block.  This is
cheap and simple to compute (moreso than full liveness).

This improves the new heuristic even more.  For example, it cuts two
out of three new instructions out of 255.vortex:DbmFileInGrpHdr,
which is one of the functions that the heuristic regressed.  This
overall eliminates another 40 instructions from 403.gcc and visibly
reduces register pressure in 255.vortex (though this only actually
ends up saving the 2 instructions from the whole program).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60084 91177308-0d34-0410-b5e6-96231b3b80d8

__fastcall and __stdcall are mingw extensions to gcc for windows. Use the
__attribute__ notation which is supported on more platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60083 91177308-0d34-0410-b5e6-96231b3b80d8

Start rewroking a subpiece of the profitability heuristic to be
phrased in terms of liveness instead of as a horrible hack.  :)

In pratice, this doesn't change the generated code for either
255.vortex or 403.gcc, but it could cause minor code changes in
theory.  This is framework for coming changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60082 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust indent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60081 91177308-0d34-0410-b5e6-96231b3b80d8

add a long-overdue AllocaInst::isStaticAlloca method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60080 91177308-0d34-0410-b5e6-96231b3b80d8

Add test for rdar://6394879.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60079 91177308-0d34-0410-b5e6-96231b3b80d8

add a comment, make save/restore logic more obvious.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60076 91177308-0d34-0410-b5e6-96231b3b80d8

This adds in some code (currently disabled unless you pass
-enable-smarter-addr-folding to llc) that gives CGP a better
cost model for when to sink computations into addressing modes.
The basic observation is that sinking increases register
pressure when part of the addr computation has to be available
for other reasons, such as having a use that is a non-memory
operation.  In cases where it works, it can substantially reduce
register pressure.

This code is currently an overall win on 403.gcc and 255.vortex
(the two things I've been looking at), but there are several
things I want to do before enabling it by default:

1. This isn't doing any caching of results, so it is much slower
   than it could be.  It currently slows down release-asserts llc
   by 1.7% on 176.gcc: 27.12s -> 27.60s.
2. This doesn't think about inline asm memory operands yet.
3. The cost model botches the case when the needed value is live
   across the computation for other reasons.

I'll continue poking at this, and eventually turn it on as llcbeta.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60074 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r60042. IndVarSimplify should check if APFloat is PPCDoubleDouble first before trying to convert it to an integer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60072 91177308-0d34-0410-b5e6-96231b3b80d8

Teach CodeGenPrepare to look through Bitcast instructions when attempting to
optimize addressing modes.  This allows us to optimize things like isel-sink2.ll
into:

movl 4(%esp), %eax
cmpb $0, 4(%eax)
jne LBB1_2 ## F
LBB1_1: ## TB
movl $4, %eax
ret
LBB1_2: ## F
movzbl 7(%eax), %eax
ret

instead of:

_test:
movl 4(%esp), %eax
cmpb $0, 4(%eax)
leal 4(%eax), %eax
jne LBB1_2 ## F
LBB1_1: ## TB
movl $4, %eax
ret
LBB1_2: ## F
movzbl 3(%eax), %eax
ret

This shrinks (e.g.) 403.gcc from 1133510 to 1128345 lines of .s.

Note that the 2008-10-16-SpillerBug.ll testcase is dubious at best, I doubt
it is really testing what it thinks it is.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60068 91177308-0d34-0410-b5e6-96231b3b80d8

fix an over-reduced test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60067 91177308-0d34-0410-b5e6-96231b3b80d8

this doesn't need EH

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60066 91177308-0d34-0410-b5e6-96231b3b80d8

ignore tests output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60065 91177308-0d34-0410-b5e6-96231b3b80d8

change AnnotationManager to use 'const char*' instead of std::string. this fixes the leakage of those strings and avoids the creation of such strings in static cosntructors (should result in a little improvement of startup time)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60064 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: Removed tools/llvmc2 directory hierarchy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60054 91177308-0d34-0410-b5e6-96231b3b80d8

CMake: llvmc2 is now known as llvmc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60052 91177308-0d34-0410-b5e6-96231b3b80d8

Add a man page for llvmc. Really basic for now, will be updated later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60049 91177308-0d34-0410-b5e6-96231b3b80d8

Since the old llvmc was removed, rename llvmc2 to llvmc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60048 91177308-0d34-0410-b5e6-96231b3b80d8

Make -fsyntax-only, -include and -emit-llvm work for C++ and Objective-C/C++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60047 91177308-0d34-0410-b5e6-96231b3b80d8

docs: Add author info + fix incorrect code example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60046 91177308-0d34-0410-b5e6-96231b3b80d8

Small documentation update.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60045 91177308-0d34-0410-b5e6-96231b3b80d8

Document the plugin priority feature.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60044 91177308-0d34-0410-b5e6-96231b3b80d8

A simplification for checking whether the signs of the operands and sum differ. Thanks, Duncan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60043 91177308-0d34-0410-b5e6-96231b3b80d8

convertToSignExtendedInteger should return opInvalidOp instead of asserting if sematics of float does not allow arithmetics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60042 91177308-0d34-0410-b5e6-96231b3b80d8

Suppress warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60041 91177308-0d34-0410-b5e6-96231b3b80d8

This method got renamed, thanks to Mattias Holm for pointing this out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60039 91177308-0d34-0410-b5e6-96231b3b80d8

CellSPU:
(a) Remove conditionally removed code in SelectXAddr. Basically, hope for the
    best that the A-form and D-form address predicates catch everything before
    the code decides to emit a X-form address.
(b) Expand vector store test cases to include the usual suspects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60034 91177308-0d34-0410-b5e6-96231b3b80d8

add info about how to run the tests with valgrind

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60030 91177308-0d34-0410-b5e6-96231b3b80d8

Now with the correct type for the 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60016 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60015 91177308-0d34-0410-b5e6-96231b3b80d8

Hacker's Delight says, "Signed integer overflow of addition occurs if and only
if the operands have the same sign and the sum has sign opposite to that of the
operands."

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60014 91177308-0d34-0410-b5e6-96231b3b80d8

Teach MatchScaledValue to handle Scales by 1 with MatchAddr (which
can recursively match things) and scales by 0 by ignoring them.
This triggers once in 403.gcc, saving 1 (!!!!) instruction in the
whole huge app.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60013 91177308-0d34-0410-b5e6-96231b3b80d8

significantly refactor all the addressing mode matching logic
into a new AddressingModeMatcher class.  This makes it easier
to reason about and reduces passing around of stuff, but has
no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60012 91177308-0d34-0410-b5e6-96231b3b80d8

refactor all the constantexpr/instruction handling code out into a
new FindMaximalLegalAddressingModeForOperation helper method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60011 91177308-0d34-0410-b5e6-96231b3b80d8

another minor tweak

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60010 91177308-0d34-0410-b5e6-96231b3b80d8

minor cleanups no functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60009 91177308-0d34-0410-b5e6-96231b3b80d8