Document function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55808 91177308-0d34-0410-b5e6-96231b3b80d8

For whatever the reason, x86 CallingConv::Fast (i.e. fastcc) was not passing scalar arguments in registers. This patch defines a new fastcc CC which is slightly different from the FastCall CC. In addition to passing integer arguments in ECX and EDX, it also specify doubles are passed in 8-byte slots which are 8-byte aligned (instead of 4-byte aligned). This avoids a potential performance hazard where doubles span cacheline boundaries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55807 91177308-0d34-0410-b5e6-96231b3b80d8

A loop may be unswitched multiple times. Reconstruct dom info. at the end.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55806 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust tests to expect new math intrinsics to be
emitted only when errno is not in use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55797 91177308-0d34-0410-b5e6-96231b3b80d8

If function notes say optimize for size, then adjust alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55794 91177308-0d34-0410-b5e6-96231b3b80d8

Add an include of SmallSet.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55793 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize loop data first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55792 91177308-0d34-0410-b5e6-96231b3b80d8

Neaten this up a bit.  No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55789 91177308-0d34-0410-b5e6-96231b3b80d8

Do not unswitch if the function notes say we're optimizing this function for size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55786 91177308-0d34-0410-b5e6-96231b3b80d8

try to seperate the mechanism into something others can use

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55785 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust libcalls tests to expect intrinsic for exp2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55784 91177308-0d34-0410-b5e6-96231b3b80d8

Add intrinsic forms of pow and exp2.  The non-intrinsic
forms remain to handle older IR files, but will go away soon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55781 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up several unbeseeming casts from pointer to intptr_t.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55779 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the ordering of operands to the store (inverted relative to LLVM IR), and fix the testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55777 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up uses of TargetLowering::getTargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55769 91177308-0d34-0410-b5e6-96231b3b80d8

cleanup as per Duncan's review

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55766 91177308-0d34-0410-b5e6-96231b3b80d8

Add a first attempt at implementing stores for X86 fast isel using target hooks.
Dan or Evan, please review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55764 91177308-0d34-0410-b5e6-96231b3b80d8

Load from GV stub should be locally CSE'd.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55763 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an overly strict assertion. Source register of a copy may not be killed, it may be killed by an implicit super-register use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55762 91177308-0d34-0410-b5e6-96231b3b80d8

Remove code that pad number of bytes to pop for X86_FastCall CC. The code doesn't do the "aligning" for Cygwin, Mingw, and Windows. But aligning it on Darwin and Linux breaks gcc compatibility. That ruled out all the platforms we support!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55756 91177308-0d34-0410-b5e6-96231b3b80d8

Add intrinsics for log, log2, log10, exp, exp2.
No functional change (and no FE change to generate them).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55753 91177308-0d34-0410-b5e6-96231b3b80d8

Do trivial local CSE for constants and other non-Instruction values
in FastISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55748 91177308-0d34-0410-b5e6-96231b3b80d8

Put RegsForValue in the llvm namespace to avoid warnings about
classes in the llvm namespace having members with types from
anonymous namespaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55747 91177308-0d34-0410-b5e6-96231b3b80d8

Create HandlePHINodesInSuccessorBlocksFast, a version of
HandlePHINodesInSuccessorBlocks that works FastISel-style. This
allows PHI nodes to be updated correctly while using FastISel.

This also involves some code reorganization; ValueMap and
MBBMap are now members of the FastISel class, so they needn't
be passed around explicitly anymore. Also, SelectInstructions
is changed to SelectInstruction, and only does one instruction
at a time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55746 91177308-0d34-0410-b5e6-96231b3b80d8

Update inline threshold for current function if the notes say, optimize for size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55745 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug that prevented PRE from applying in some cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55744 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid extra comma.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55742 91177308-0d34-0410-b5e6-96231b3b80d8

Parse and print opt_size note.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55740 91177308-0d34-0410-b5e6-96231b3b80d8

Initial version of a Partial Specialization IPO pass.  It triggers a couple hundred times on 176.gcc.  I don't know the performance impact yet, the heuristic is quite simple still.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55734 91177308-0d34-0410-b5e6-96231b3b80d8

Do not emit a UsedDirective for things in the llvm.used
list that have internal linkage; the linker doesn't need
or want this.  (These objects must still be preserved
at compile time, so just removing them from the llvm.used
list doesn't work.)  Should affect only Darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55722 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55720 91177308-0d34-0410-b5e6-96231b3b80d8

Add missing decls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55719 91177308-0d34-0410-b5e6-96231b3b80d8

Add parentheses to make code more readable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55717 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55716 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for commits 55700 and 55714.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55715 91177308-0d34-0410-b5e6-96231b3b80d8

If a SCC has a node without a function, then the SCC
analysis would bail out without removing function
records for other members of the SCC (which may exist
if those functions read or wrote global variables).
Since these are initialized to "readnone", this
resulted in incorrect alias analysis results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55714 91177308-0d34-0410-b5e6-96231b3b80d8

Add custom inliner that handles only functions that are marked as always_inline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55713 91177308-0d34-0410-b5e6-96231b3b80d8

Handle "always inline" note during inline cost analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55712 91177308-0d34-0410-b5e6-96231b3b80d8

Check noinline note and ignore other notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55711 91177308-0d34-0410-b5e6-96231b3b80d8

Handle "noinline" note inside the simple inliner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55708 91177308-0d34-0410-b5e6-96231b3b80d8

Oops, I accidentally broke the fallback case with my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55704 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an issue where we were reusing materializations of constants in blocks not dominated by the materialization.  This is
the simple fix, materializing the constant before every use.  It might be better to either track domination of uses or
to materialize all constants and the beginning of the function and let remat sort when to do materialization at uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55703 91177308-0d34-0410-b5e6-96231b3b80d8

Update test to check call instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55702 91177308-0d34-0410-b5e6-96231b3b80d8

Split the SelectionDAG-building code, including the FunctionLoweringInfo
and SelectionDAGLowering classes, out of SelectionDAGISel.cpp and put
it in a separate file, SelectionDAGBuild.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55701 91177308-0d34-0410-b5e6-96231b3b80d8

Fix maxo bado thinko.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55700 91177308-0d34-0410-b5e6-96231b3b80d8

Separate MachineInstr-emitting routines from actual scheduling
routines and move them into a separate file, ScheduleDAGEmit.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55699 91177308-0d34-0410-b5e6-96231b3b80d8

Fix addRegisterDead and addRegisterKilled to be more thorough
when searching for redundant subregister dead/kill bits.

Previously it was common to see instructions marked like this:
  "RET %EAX<imp-use,kill>, %AX<imp-use,kill>"

With this change, addRegisterKilled continues scanning after
finding the %EAX operand, so it proceeds to discover the
redundant %AX kill and eliminates it, producing this:
  "RET %EAX<imp-use,kill>"

This currently has no effect on the generated code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55698 91177308-0d34-0410-b5e6-96231b3b80d8

Since onlyReadsMemory returns true if in fact
doesNotAccessMemory, check doesNotAccessMemory
first, since otherwise functions may be
marked readonly rather than readnone.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55697 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup GlobalsModRef a bit.  When analysing the
callgraph, when one member of a SCC calls another
then the analysis would drop to mod-ref because
there is (usually) no function info for the callee
yet; fix this.  Teach the analysis about function
attributes, in particular the readonly attribute
(which requires being careful about globals).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55696 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for PR1678.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55695 91177308-0d34-0410-b5e6-96231b3b80d8

Try to fold each element of a vector. This is needed to maintain structural
equivalence.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55694 91177308-0d34-0410-b5e6-96231b3b80d8

Add X86 target hook to implement load (even from GlobalAddress).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55693 91177308-0d34-0410-b5e6-96231b3b80d8

If TargetSelectInstruction returns true, move to next instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55692 91177308-0d34-0410-b5e6-96231b3b80d8

Make UpdateValueMap, createResultReg, etc. protected instead of private so they can used by target hooks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55691 91177308-0d34-0410-b5e6-96231b3b80d8

Don't apply this transform to vectors. Fixes PR2756.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55690 91177308-0d34-0410-b5e6-96231b3b80d8

Don't crash when trying to constant fold a vector with some elements that can't
be folded. Instead, fail to fold the entire vector.

We could also return a vector with some elements folded and some not. If anyone
thinks that's a better approach, please speak up!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55689 91177308-0d34-0410-b5e6-96231b3b80d8

Fix capitalization in #include of FastISel.h.  This unbreaks the build on case-sensitive filesystems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55687 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak fast isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55685 91177308-0d34-0410-b5e6-96231b3b80d8

Add additional check to ensure that iv is canonicalized.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55682 91177308-0d34-0410-b5e6-96231b3b80d8

Check iteration count.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55680 91177308-0d34-0410-b5e6-96231b3b80d8

Let tblgen only generate fastisel routines, not the class definition. This makes it easier for targets to define its own fastisel class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55679 91177308-0d34-0410-b5e6-96231b3b80d8

While removing PHI, use basicblock to identify incoming value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55678 91177308-0d34-0410-b5e6-96231b3b80d8

s/FP_AlwaysInline/FN_NOTE_AlwaysInline/g

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55676 91177308-0d34-0410-b5e6-96231b3b80d8

If all IV uses are extending integer IV then change the type of IV itself, if possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55674 91177308-0d34-0410-b5e6-96231b3b80d8

respect inline=never and inline=always notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55673 91177308-0d34-0410-b5e6-96231b3b80d8

80 col violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55668 91177308-0d34-0410-b5e6-96231b3b80d8

Read and write function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55657 91177308-0d34-0410-b5e6-96231b3b80d8

Use bitwise AND.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55656 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for targets without 64-bit atomics;
xfail old test for ppc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55651 91177308-0d34-0410-b5e6-96231b3b80d8

Function notes tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55648 91177308-0d34-0410-b5e6-96231b3b80d8

Print function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55647 91177308-0d34-0410-b5e6-96231b3b80d8

Parse function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55646 91177308-0d34-0410-b5e6-96231b3b80d8

Initialize function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55645 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some bugs in the code sequences for atomics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55643 91177308-0d34-0410-b5e6-96231b3b80d8

Ensure that HandlePHINodesInSuccessorBlocks is run for all blocks,
even in FastISel mode in the case where FastISel successfully
selects all the instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55641 91177308-0d34-0410-b5e6-96231b3b80d8

Getter and setter for function notes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55638 91177308-0d34-0410-b5e6-96231b3b80d8

Reapply majority of r55557 but with the changes to compilation flags
disabled until issues with gcc 4.1 on linux 32-bit are resolved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55636 91177308-0d34-0410-b5e6-96231b3b80d8

plug memleak: destroy internal buffer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55632 91177308-0d34-0410-b5e6-96231b3b80d8

plug a little memleak in verifyFunction()
# first commit to llvm, so whatch out :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55631 91177308-0d34-0410-b5e6-96231b3b80d8

Revert r55557, it is causing linking failures on 32bit linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55628 91177308-0d34-0410-b5e6-96231b3b80d8

MMI may be null.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55626 91177308-0d34-0410-b5e6-96231b3b80d8

Add Mac OS X compatible JIT callback routine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55625 91177308-0d34-0410-b5e6-96231b3b80d8

Revamp ARM JIT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55624 91177308-0d34-0410-b5e6-96231b3b80d8

Change getBinaryCodeForInstr prototype. First operand MachineInstr& should be const. Make corresponding changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55623 91177308-0d34-0410-b5e6-96231b3b80d8

Provide two overloads of AnalyzeNewNode.

The first can update the SDNode in an SDValue
while the second is called with SDNode* and
returns a possibly updated SDNode*.

This patch has no intended functional impact,
but helps eliminating ugly temporary SDValues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55608 91177308-0d34-0410-b5e6-96231b3b80d8

Even though no caller actually uses the new value
(what matters is that it is added to the worklist),
it seems more logical to return it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55606 91177308-0d34-0410-b5e6-96231b3b80d8

Turn this legalize types test on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55605 91177308-0d34-0410-b5e6-96231b3b80d8

Add a small pass that sets the readnone/readonly
attributes on functions, based on the result of
alias analysis.  It's not hardwired to use
GlobalsModRef even though this is the only (AFAIK)
alias analysis that results in this pass actually
doing something.  Enable as follows:
  opt ... -globalsmodref-aa -markmodref ...
Advantages of this pass: (1) records the result
of globalsmodref in the bitcode, meaning it is
available for use by later passes (currently
the pass manager isn't smart enough to magically
make an advanced alias analysis available to all
later passes), which may expose more optimization
opportunities; (2) hopefully speeds up compilation
when code is optimized twice, for example when a
file is compiled to bitcode, then later LTO is done
on it: marking functions readonly/readnone when
producing the initial bitcode should speed up alias
analysis during LTO; (3) good for discovering that
globalsmodref doesn't work very well :)
Not currently turned on by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55604 91177308-0d34-0410-b5e6-96231b3b80d8

Control flow instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55601 91177308-0d34-0410-b5e6-96231b3b80d8

ldm / stm instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55599 91177308-0d34-0410-b5e6-96231b3b80d8

AXI2 and AXI3 instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55598 91177308-0d34-0410-b5e6-96231b3b80d8

Reorganize instruction formats again; AXI1 encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55597 91177308-0d34-0410-b5e6-96231b3b80d8

addrmode3 instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55596 91177308-0d34-0410-b5e6-96231b3b80d8

Reorganize some instruction format definitions. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55594 91177308-0d34-0410-b5e6-96231b3b80d8

Rest of addrmode2 instruction encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55593 91177308-0d34-0410-b5e6-96231b3b80d8

Addr2 word / byte load encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55591 91177308-0d34-0410-b5e6-96231b3b80d8

Addr1 instructions opcodes are encoded in bits 21-24; encode S bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55590 91177308-0d34-0410-b5e6-96231b3b80d8

fix a bunch of 80-col violations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55588 91177308-0d34-0410-b5e6-96231b3b80d8

Revert the "XFAIL" for the rotate_ops.ll testcase. Instead, mark ISD::ROTR
instructions in CellSPU as "Expand" so that they won't be generated. I added a
"FIXME" so that this hack can be addressed and reverted once ISD::ROTR is
supported in the .td files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55582 91177308-0d34-0410-b5e6-96231b3b80d8

Expand for ROTR with MVT::i64.

Dale, Could you please review this?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@55581 91177308-0d34-0410-b5e6-96231b3b80d8