We don't really need to #include IPO.h into this file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14911 91177308-0d34-0410-b5e6-96231b3b80d8

Initial (outline only) draft of the System Library requirements, design
and details documentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14910 91177308-0d34-0410-b5e6-96231b3b80d8

Added a README to point users to the documentation on the design of this
library which is a prerequisite for maintaining the library.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14909 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect computation of mod/ref sets.  Do not ask for mod/ref information
for objects of size 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14908 91177308-0d34-0410-b5e6-96231b3b80d8

Print modref information in a useful way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14907 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanups: fold two loops into one
New features: -print-all-alias-modref-info option, print more info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14906 91177308-0d34-0410-b5e6-96231b3b80d8

link.h causes problems on some platforms and is not needed on solaris, where
the comment came from.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14905 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14904 91177308-0d34-0410-b5e6-96231b3b80d8

Add note about IA64 C++

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14903 91177308-0d34-0410-b5e6-96231b3b80d8

Fix warning on 64-bit machines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14902 91177308-0d34-0410-b5e6-96231b3b80d8

* Use LI(S) to copy constants into registers intead of ADDI(S) as the latter is
  a funky way to "use" R0 for a 0-valued operand
* Add IMPLICIT_DEFs for incoming function arguments via registers to help the
  register allocator not clobber those registers
* Implement comparisons with longs
* Teach emitSelectOperation() to fold the SetCC operation

Patch contributed by Nate Begeman

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14901 91177308-0d34-0410-b5e6-96231b3b80d8

* Store all non-volatile int registers R13-31 on the stack, restore on exit
* Fix comment formatting

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14900 91177308-0d34-0410-b5e6-96231b3b80d8

Fix code formatting

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14899 91177308-0d34-0410-b5e6-96231b3b80d8

Implement PowerPCInstrInfo::isMoveInstr(), patch by Nate Begeman

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14898 91177308-0d34-0410-b5e6-96231b3b80d8

Add prototype for TargetInstrInfo::isMoveInstr()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14897 91177308-0d34-0410-b5e6-96231b3b80d8

* Enable allocation of registers r2-r10
* Allocate registers 13-31 backwards (to be able to store them all at once)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14896 91177308-0d34-0410-b5e6-96231b3b80d8

Add IMPLICIT_DEFS pseudo-instruction; patch by: Nate Begeman

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14895 91177308-0d34-0410-b5e6-96231b3b80d8

The generated instruction selector isn't (yet) functional

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14894 91177308-0d34-0410-b5e6-96231b3b80d8

* Output non-lazy linking stubs for external global variables
* Get rid of dead and #if 0'd code
* Minor for loop speed-up: save end iterator instead of querying every time

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14893 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly build shared objects on MacOS X for debugging code generators

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14892 91177308-0d34-0410-b5e6-96231b3b80d8

Define double alignment as 8 bytes now that assert(DoubleAlignment == PointerSize)
has been eliminated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14891 91177308-0d34-0410-b5e6-96231b3b80d8

* Add spaces between words and numbers in comments printed out for longs/floats
* Print out IMPLICIT_DEFS as comments in the assembly, patch by Nate Begeman

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14890 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14888 91177308-0d34-0410-b5e6-96231b3b80d8

Add convinience constructor for function calls with two args.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14885 91177308-0d34-0410-b5e6-96231b3b80d8

Add a class for pseudo-instructions. Use it.
Add IMPLICIT_USE and IMPLICIT_DEF, a la X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14884 91177308-0d34-0410-b5e6-96231b3b80d8

Add what will eventually be the TSFlags. Big switch(opcode) statements are bad.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14883 91177308-0d34-0410-b5e6-96231b3b80d8

Add special handling for pseudo-instructions (print them as comments).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14882 91177308-0d34-0410-b5e6-96231b3b80d8

Add to-do list.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14881 91177308-0d34-0410-b5e6-96231b3b80d8

Do IMPLICIT_DEFs on incoming args' hard regs, to avoid confusing the regalloc.
Support single-fp incoming args.
Support single-fp outgoing args ('call' operands).
Support double-fp return values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14880 91177308-0d34-0410-b5e6-96231b3b80d8

The powerpc is now gone.  However it is now just known as the Skeleton target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14877 91177308-0d34-0410-b5e6-96231b3b80d8

Build skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14876 91177308-0d34-0410-b5e6-96231b3b80d8

Build the skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14875 91177308-0d34-0410-b5e6-96231b3b80d8

Initial checkin of the rest of the skeleton target

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14874 91177308-0d34-0410-b5e6-96231b3b80d8

Initial skeleton tablegen files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14873 91177308-0d34-0410-b5e6-96231b3b80d8

Add skeleton makefile

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14872 91177308-0d34-0410-b5e6-96231b3b80d8

Initial readme

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14871 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the regressions handling unnamed global variables

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14870 91177308-0d34-0410-b5e6-96231b3b80d8

For 64-bit platform support, we now use the clEnumValEnd macro to terminate
cl::values lists

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14869 91177308-0d34-0410-b5e6-96231b3b80d8

IA64 compat

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14867 91177308-0d34-0410-b5e6-96231b3b80d8

Fix IA64 compatibility

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14866 91177308-0d34-0410-b5e6-96231b3b80d8

Use new macro

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14865 91177308-0d34-0410-b5e6-96231b3b80d8

Be compatible with IA64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14864 91177308-0d34-0410-b5e6-96231b3b80d8

Passing integer 0 in for a pointer value doesn't work on IA64.  Fix this
by using a new macro.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14863 91177308-0d34-0410-b5e6-96231b3b80d8

Add a new macro to be used for the end of enum lists.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14862 91177308-0d34-0410-b5e6-96231b3b80d8

This patch was contributed by Daniel Berlin!

Speed up SCCP substantially by processing overdefined values quickly.  This
patch speeds up SCCP by about 30-40% on large testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14861 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR404 try #2

This version takes about 1s longer than the previous one (down to 2.35s),
but on the positive side, it actually works :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14856 91177308-0d34-0410-b5e6-96231b3b80d8

Give SetVector range support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14855 91177308-0d34-0410-b5e6-96231b3b80d8

Nuke a clearly bogus assertion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14854 91177308-0d34-0410-b5e6-96231b3b80d8

Revert previous patch until I get a bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14853 91177308-0d34-0410-b5e6-96231b3b80d8

QOI issue fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14852 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR404: Loop simplify is really slow on 252.eon

This eliminates an N*N*logN algorithm from the loop simplify pass, replacing
it with a much simpler and faster alternative.  In a debug build, this reduces
gccas time on eon from 85s to 42s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14851 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14850 91177308-0d34-0410-b5e6-96231b3b80d8

Final fix for PR341: eliminate operator<<(ostream, Value*).  Clients should
now send references to ostreams instead of pointers.  Sending pointers to
ostreams will print their addresses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14849 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14848 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14847 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14846 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14845 91177308-0d34-0410-b5e6-96231b3b80d8

Revert stuff that I didn't mean to checkin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14844 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14843 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14842 91177308-0d34-0410-b5e6-96231b3b80d8

Patches towards fixing PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14841 91177308-0d34-0410-b5e6-96231b3b80d8

Progress on PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14840 91177308-0d34-0410-b5e6-96231b3b80d8

Fixes working towards PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14839 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fixes for PR341

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14838 91177308-0d34-0410-b5e6-96231b3b80d8

Implement folding of expressions like 'uint cast (int* getelementptr (int*
null, uint 1) to uint)' to a constant integer.  We can only do this with
primitive LLVM types, because other types have target-specific sizes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14837 91177308-0d34-0410-b5e6-96231b3b80d8

Yeah, how about those 'primative' types huh?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14836 91177308-0d34-0410-b5e6-96231b3b80d8

Now that we codegen the portable "sizeof" efficiently, we can use it for
malloc lowering.  This means that lowerallocations doesn't need targetdata
anymore.  yaay.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14835 91177308-0d34-0410-b5e6-96231b3b80d8

Improve codegen for the LLVM offsetof/sizeof "operator".  Before we compiled
this LLVM function:

int %foo() {
        ret int cast (int** getelementptr (int** null, int 1) to int)
}

into:

foo:
        mov %EAX, 0
        lea %EAX, DWORD PTR [%EAX + 4]
        ret

now we compile it into:

foo:
        mov %EAX, 4
        ret

This sequence is frequently generated by the MSIL front-end, and soon the malloc lowering pass and
Java front-ends as well..

-Chris

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14834 91177308-0d34-0410-b5e6-96231b3b80d8

You cannot pass functions by value, what does that even mean??

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14833 91177308-0d34-0410-b5e6-96231b3b80d8

this "feature" is gone

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14832 91177308-0d34-0410-b5e6-96231b3b80d8

This is logically part of the last patch.  Just more really horrible code
that is made unnecessary by it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14831 91177308-0d34-0410-b5e6-96231b3b80d8

** Finally DeclareNewGlobalValue is dead!
* Simplify a lot of code because type's cannot be in function symbol tables
* Fix memory leaks in handling of redefined function prototypes
* Don't use SymbolTable directly for stuff that we can go through the Module
  for.
* Fix some minor bugs on obscure testcases like:
      test/Feature/globalredefinition.ll
* Do not create GlobalVariable objects for forward referenced Functions!
* When forward referencing a function in a constant expression, do not create
  a placeholder, add a bunch of references to it, then turn around and
  replaceAllUsesOfWith on it with a new global, deleting the placeholder.
  Instead, when we find the real definition of the global, just use the
  placeholder instead of creating a new object.

This substantially simplifies the asmwriter and should even speed it up on
cases heavy in constantexprs (like C++, Java, MSIL)...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14830 91177308-0d34-0410-b5e6-96231b3b80d8

New, horrible, testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14829 91177308-0d34-0410-b5e6-96231b3b80d8

* Fairly substantial change.  Instead of creating new globalvariables, then
  replaceAllUsesWith'ing any forward references, just use the forward
  reference if it exists.

This introduces GetForwardRefForGlobal, which will eventually completely
replace the horrible DeclareNewGlobalValue function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14828 91177308-0d34-0410-b5e6-96231b3b80d8

Fold setValueNameMergingDuplicates into ParseGlobalVariable, allowing us
to substantially simplify the result.  In particular, we no longer create
GlobalVariables and then immediately destroy them when they are duplciate
definitions.

The real point of this patch though is that it gets us closer to the
DeclareNewGlobalValue calls...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14827 91177308-0d34-0410-b5e6-96231b3b80d8

Actually set the endian/pointersize flags on the module being read in!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14826 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14825 91177308-0d34-0410-b5e6-96231b3b80d8

Make PATypeHolder and friends return non-const pointers to the types they
hold.  Because types are basically immutable anyway, they should not be
referenced as "const Type*" everywhere.  Just "Type*" should suffice!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14824 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a regression from last night.  Apparently the CFE is broken and outputs
functions multiple times, expecting them to be merged.  This should be fixed
in the CFE, then here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14823 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure MTSPR instruction is inserted into the BasicBlock

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14822 91177308-0d34-0410-b5e6-96231b3b80d8

Implement new helper methods for creating two-index GEP instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14821 91177308-0d34-0410-b5e6-96231b3b80d8

* Change code to use a static_cast instead of reinterpret_cast
* Add comments
* Add two new gep ctors for the very common case of creating a two operand
  GEP instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14820 91177308-0d34-0410-b5e6-96231b3b80d8

Don't define the same register twice when loading a ConstantPointerRef to a reg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14819 91177308-0d34-0410-b5e6-96231b3b80d8

* Fix multiplication by powers of two and otherwise
* Clarify variable name (StoreInst SI instead of LI)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14818 91177308-0d34-0410-b5e6-96231b3b80d8

Pull out code shared between GV forward-decl and definition processing.
This gives us only a single call site for setValueNameMergingDuplicates.
The next stage is the start merging them together.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14811 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code.  Do not allow functions to be redefined more than once.
Since the stupid '%X = const int 4' thing is gone, we can now simplify
setValueNameMergingDuplicates a bit more.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14810 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a gross and crufty "feature" that was never documented and doesn't work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14809 91177308-0d34-0410-b5e6-96231b3b80d8

Finegrainify namespacification

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14808 91177308-0d34-0410-b5e6-96231b3b80d8

Revamp handling of labels.  In particular, if we create a forward reference
for a basic block, use it when the block is defined instead of deleting it
and creating a new one.  Also, only create at most ONE forward reference
for any block, instead of one for each forward reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14807 91177308-0d34-0410-b5e6-96231b3b80d8

Add Machine-CFG edges to SparcV9 MachineBasicBlocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14806 91177308-0d34-0410-b5e6-96231b3b80d8

Split the basic block handling case out of getVal into getBBVal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14805 91177308-0d34-0410-b5e6-96231b3b80d8

Make Argument::print more resilient to non-verifiable IR

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14801 91177308-0d34-0410-b5e6-96231b3b80d8

Catch aggregates passed by value sooner rather than later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14800 91177308-0d34-0410-b5e6-96231b3b80d8

* Specify that FP arith options have 3 operands
* Correctly load FP constants from the constant pool, should be refactored

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14799 91177308-0d34-0410-b5e6-96231b3b80d8

Actually, use an exception to stop verification.  This gives us much better
error messages because verification stops at the first error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14794 91177308-0d34-0410-b5e6-96231b3b80d8

There is no reason to abort and print a stack trace if there is a verification
error.  Just print the message like a good little tool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14793 91177308-0d34-0410-b5e6-96231b3b80d8

Fine-grainify namespacification, prune #include

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14792 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typeo and refactor bb productions to make it possible for us to reuse any
forward reference blocks if they have been created (instead of creating a new
block, replaceAllUsesOfWith, then nuking the placeholder).   This is not yet
implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14791 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate some mega-cruft here.  There is no reason to DERIVE FROM IR CLASSES
just to keep track of some per-object state!  Gaah!  Whoever wrote this stuff...
oh wait, that would be me.  Never mind.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14790 91177308-0d34-0410-b5e6-96231b3b80d8

Inline the now trivial setValueNameInternal function into both callers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@14789 91177308-0d34-0410-b5e6-96231b3b80d8