Read in expanders too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7678 91177308-0d34-0410-b5e6-96231b3b80d8

Continued evolution of the selector emitter: Represent instruction patterns
as first class objects

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7677 91177308-0d34-0410-b5e6-96231b3b80d8

Implement LLVM intrinsics `llvm.setjmp' and `llvm.longjmp' as follows:
* setjmp() simply returns 0
* longjmp() simply calls abort()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7676 91177308-0d34-0410-b5e6-96231b3b80d8

Fix assertion in MachineInstr::substituteValue().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7675 91177308-0d34-0410-b5e6-96231b3b80d8

Fix sanity-checking in 'maskUnsigned' code to be more precise:
use or def-and-use operands can be substituted after one def-only
operand has been substituted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7674 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed a segfault in gccld.
The original code does not work because the value from WorkList.end() is
invalidated once WorkList.erase() is called.  To ensure proper functionality,
we must ensure that WorkList.erase() is always called before WorkList.end().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7673 91177308-0d34-0410-b5e6-96231b3b80d8

Update tablegen interfaces

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7672 91177308-0d34-0410-b5e6-96231b3b80d8

Finish implementation of the type inference engine.
Start working on reading in nonterminals

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7671 91177308-0d34-0410-b5e6-96231b3b80d8

Allow clients to get at the pointer type

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7670 91177308-0d34-0410-b5e6-96231b3b80d8

Add new method getVAlueAsDag

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7669 91177308-0d34-0410-b5e6-96231b3b80d8

Initial checkin of tree pattern parser and type inference engine (which still needs work).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7668 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate now-dead method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7667 91177308-0d34-0410-b5e6-96231b3b80d8

Start using the CodeGeneratorWrappers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7666 91177308-0d34-0410-b5e6-96231b3b80d8

Initial checkin of useful wrappers around the Target classes, for now, only ValueType and
Target are wrapped

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7665 91177308-0d34-0410-b5e6-96231b3b80d8

Removed the support for pool allocation testing. This will be in a separate TEST file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7664 91177308-0d34-0410-b5e6-96231b3b80d8

Change implementation so that variable sized slabs are used to allow arbitrary sized array allocations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7663 91177308-0d34-0410-b5e6-96231b3b80d8

Added code for pool allocating only the pool-allocatable data structures in the presence of collapsed nodes + a couple of bug fixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7662 91177308-0d34-0410-b5e6-96231b3b80d8

This register is never used, disable it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7661 91177308-0d34-0410-b5e6-96231b3b80d8

Added a flag which is set when all data structures are not pool allocated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7660 91177308-0d34-0410-b5e6-96231b3b80d8

There was no reason for  these to be bit-fields, they just need to be unique.
Also, add an isVoid item

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7659 91177308-0d34-0410-b5e6-96231b3b80d8

Changing command-line option formats to be more consistent with LLVM style.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7658 91177308-0d34-0410-b5e6-96231b3b80d8

* Renamed option from `nopreselect' to `nopreopt' since it disables more than
  just PreSelection
* Wrapped code at 80 columns
* Added the DecomposeMultiDimRefs Pass to the JIT compilation path

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7657 91177308-0d34-0410-b5e6-96231b3b80d8

Ick, add the RIGHT accessor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7656 91177308-0d34-0410-b5e6-96231b3b80d8

Add accessor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7655 91177308-0d34-0410-b5e6-96231b3b80d8

There is something wrong with code that looks like:
if (R == 0 || ...)
  R->getName()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7654 91177308-0d34-0410-b5e6-96231b3b80d8

Use the registers g1 and g5 as temporaries for making far jumps and far calls,
because saving i1 and i2 to their ``designated'' stack slots corrupts unknown
memory in other functions, standard libraries, and worse.

In addition, this has the benefit of improving JIT performance because we
eliminate writing out 4 instructions in CompilationCallback() and 2 loads and 2
stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7653 91177308-0d34-0410-b5e6-96231b3b80d8

Removed Makefile.common.
Since we are including Makefile.test, we automatically get Makefile.common.
Furthermore, the double inclusion of Makefile.common causes the test suite to
be executed twice per invocation of the top level make.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7652 91177308-0d34-0410-b5e6-96231b3b80d8

Export the register classes so that the instruction selector can get at them as needed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7651 91177308-0d34-0410-b5e6-96231b3b80d8

Makefile.common: Remove commented-out and duplicate rules.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7650 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bunch of new Alpha Intrinsics for Rahul Joshi

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7646 91177308-0d34-0410-b5e6-96231b3b80d8

1. Bug fix: was using SLL instead of SLLX for ULongTy.  Chump.
2. Handle fp-to-uint conversions directly here instead of relying on
   a pre-transformation to replace them with the 2-step conversion.
3. Use size rather than explicitly checking types when deciding what
   opcodes to use, wherever possible.  This is less error prone (the
   bug fix above was not the first time!).
4. Float-to-pointer casts shd now work though this hasn't been tested.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7645 91177308-0d34-0410-b5e6-96231b3b80d8

Remove conversion of fp-to-uint cast into a multi-step cast:
this is not an optional transformation on SPARC and is now handled
directly by instruction selection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7644 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7643 91177308-0d34-0410-b5e6-96231b3b80d8

The HAVE_JUMP code is dead, these intrinsics should _never_ be expanded

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7642 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unnecessary use of NonCopyable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7641 91177308-0d34-0410-b5e6-96231b3b80d8

Added asserts to prevent negative shift amounts from being generated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7640 91177308-0d34-0410-b5e6-96231b3b80d8

SparcV9CodeEmitter.cpp:
* Doxygen-ified comments
* Added capability to make far calls (i.e., beyond 30 bits in CALL instr)
  which implies that we need to delete function references that were added by
  the call to addFunctionReference() because the actual call instruction is 10
  instructions away (thanks to 64-bit address construction)
* Cleaned up code that generates far jumps by using an array+loop

SparcV9CodeEmitter.h:
* Explained more of the side-effects of emitFarCall()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7639 91177308-0d34-0410-b5e6-96231b3b80d8

Sort the file lists

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7638 91177308-0d34-0410-b5e6-96231b3b80d8

Loop unroller doesn't necessarily need to be profile driven

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7637 91177308-0d34-0410-b5e6-96231b3b80d8

Completely eliminate the isVoid TSFlag, shifting over all other fields

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7636 91177308-0d34-0410-b5e6-96231b3b80d8

add a pattern for RET, immediates no longer need to be explicitly typed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7635 91177308-0d34-0410-b5e6-96231b3b80d8

Start adding usefulness to the DAG node definitions, add a new Expander
class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7634 91177308-0d34-0410-b5e6-96231b3b80d8

Add error checking code to the node type parser.  Start the instruction pattern
reader

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7632 91177308-0d34-0410-b5e6-96231b3b80d8

Initial support for an instruction selector emitter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7631 91177308-0d34-0410-b5e6-96231b3b80d8

All good classes with virtual functions should have virtual dtors...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7630 91177308-0d34-0410-b5e6-96231b3b80d8

Add an instruction selector emitter skeleton

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7629 91177308-0d34-0410-b5e6-96231b3b80d8

convert over to using TableGen backends

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7628 91177308-0d34-0410-b5e6-96231b3b80d8

Switch code over to being a TableGenBackend

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7627 91177308-0d34-0410-b5e6-96231b3b80d8

Add more helper methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7626 91177308-0d34-0410-b5e6-96231b3b80d8

New common interface for backends to use

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7625 91177308-0d34-0410-b5e6-96231b3b80d8

Added POOLFLAGS option to run pool allocation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7624 91177308-0d34-0410-b5e6-96231b3b80d8

"fix" coding style stuff
Change some <>'s into &lt;&gt;'s

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7623 91177308-0d34-0410-b5e6-96231b3b80d8

Completely eliminate the per-machine-instruction regsUsed set.

This substantially shrinks the size of each machine instruction, which should
make allocation faster and the cache footprint of the machine code lighter.

Here are some timings for code generation of the larger benchmarks we have.
This are timings of code generation phases of the X86 JIT, when compiled in
debug mode:

Before After Diff
164.gzip:
  InstSel 0.0878 0.0722 -21.6%
  RegAlloc 0.2031 0.1757 -15.6%
  TOTAL 0.5585 0.4999 -11.7%
Ptrdist-bc:
  InstSel 0.0878 0.0722 -21.6%
  RegAlloc 0.2070 0.1933 - 7.1%
  TOTAL 0.6972 0.6464 - 7.9%
197.parser:
  InstSel 0.2148 0.2148 - 0.0%
  RegAlloc 0.4941 0.4277 -15.5%
  TOTAL 1.3749 1.2851 - 7.0%
175.vpr:
  InstSel 0.2519 0.2109 -19.4%
  RegAlloc 0.5976 0.5663 - 5.5%
  TOTAL 1.6933 1.6347 - 3.5%
254.gap:
  InstSel 1.1328 0.9921 -14.2%
  RegAlloc 2.6933 2.4804 - 8.6%
  TOTAL 7.7871 7.2499 - 7.4%

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7622 91177308-0d34-0410-b5e6-96231b3b80d8

Use a new local data structure instead of the MachineInstr::regsUsed set

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7621 91177308-0d34-0410-b5e6-96231b3b80d8

Add a map

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7620 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7619 91177308-0d34-0410-b5e6-96231b3b80d8

Physical registers no longer live in the regsUsed set for each machine instr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7618 91177308-0d34-0410-b5e6-96231b3b80d8

Do not insert physical regsiters into the regsUsed set

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7617 91177308-0d34-0410-b5e6-96231b3b80d8

Reverted back to using OR for cmp/mv operations for lex/yacc output.
The shell AND/OR operators short-circuit on command success/failure, which is
the inverse of exit status (i.e. 0 means success, non-zero means failure).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7616 91177308-0d34-0410-b5e6-96231b3b80d8

Switched from using diff to cmp for two reasons:
o Not all versions of diff have the -q option
o The cmp program is probably faster than diff
Fixed the logic that only copies the file over if no differences are found.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7615 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed minor bug in SafeToHoist and made some changes suggested by Chris.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7614 91177308-0d34-0410-b5e6-96231b3b80d8

Added LICM test cases to:

1) Check that trapping instructionns that are not guaranteed to execute are not hoisted.
2) Check that trapping instructions that are guaranteed to execute are hoisted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7613 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed LICM bug that hoists trapping instructions that are not guaranteed to execute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7612 91177308-0d34-0410-b5e6-96231b3b80d8

Minor changes:
  * Expand most tabs into spaces
  * Move #define DEBUG_TYPE to top of file to avoid warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7611 91177308-0d34-0410-b5e6-96231b3b80d8

Add more verbose comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7610 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment to the method decl

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7609 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7608 91177308-0d34-0410-b5e6-96231b3b80d8

Added declaration of mergeInGlobalsGraph

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7607 91177308-0d34-0410-b5e6-96231b3b80d8

Added function mergeInGlobalsGraph which merges in the entire globals graph with the graph of a function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7606 91177308-0d34-0410-b5e6-96231b3b80d8

Major bug fixes including a memory leak and tracking some exceptional conditions. Also added support for including global and indirect call information in the DS graphs used by the pool allocation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7605 91177308-0d34-0410-b5e6-96231b3b80d8

Specify DEBUG_TYPE's for the JIT debug messages

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7604 91177308-0d34-0410-b5e6-96231b3b80d8

This method has now been changed to preserve flags for us!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7603 91177308-0d34-0410-b5e6-96231b3b80d8

All callers of these methods actually wanted them to preserve the flags,
so get rid of the def/use parameters that were getting passed in.

**** This now changes the semantics of these methods to preserve the flags,
     not clobber them!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7602 91177308-0d34-0410-b5e6-96231b3b80d8

Added the declaration of InlineIndirectCalls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7601 91177308-0d34-0410-b5e6-96231b3b80d8

Factor shared code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7600 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug: SimplifyCFG/2003-08-05-InvokeCrash.ll
Fix bug: SimplifyCFG/2003-08-05-MishandleInvoke.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7599 91177308-0d34-0410-b5e6-96231b3b80d8

new testcases that simplifycfg breaks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7598 91177308-0d34-0410-b5e6-96231b3b80d8

* Removed `using' declaration, now use full namespace qualifier std::string
* Simplified code by using an inline function instead of copy-pasted code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7597 91177308-0d34-0410-b5e6-96231b3b80d8

If we're debugging the SimplifyCFG pass, we _REALLY_ don't want to use it for
narrowing, no matter what.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7596 91177308-0d34-0410-b5e6-96231b3b80d8

Implement TODO: print out short form of Invoke if possible

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7595 91177308-0d34-0410-b5e6-96231b3b80d8

Only test the vector of functions if it is non-empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7594 91177308-0d34-0410-b5e6-96231b3b80d8

Stop special-casing annul and predict bits (which are Sparc-specific anyway)
since those bits are now hard-coded in Sparc*.td files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7593 91177308-0d34-0410-b5e6-96231b3b80d8

* Set annul bit to be 0, because the Sparc backend currently does not use it.
* Use the name of the predict field instead of just the const 1 in the
  Instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7592 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bugs handling ESP in alloca references

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7591 91177308-0d34-0410-b5e6-96231b3b80d8

No functional changes, comment the fix I just put in

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7590 91177308-0d34-0410-b5e6-96231b3b80d8

The CodeEmitterGenerator used to consider ANY uninitialized field as being an
operand (unless it's annul or predict).  Now we only consider fields to be
operands if they are uninitialized AND used in the "Inst" field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7589 91177308-0d34-0410-b5e6-96231b3b80d8

Revert previous change, and be really anal about what physical registers can do.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7588 91177308-0d34-0410-b5e6-96231b3b80d8

This is the real fix for the previous register allocator problem.
Physical registers should not float around.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7587 91177308-0d34-0410-b5e6-96231b3b80d8

Reformatted code to match the prevalent LLVM style; fit code into 80 columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7586 91177308-0d34-0410-b5e6-96231b3b80d8

ElectricFence found this bug where we were reading past the vector boundary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7585 91177308-0d34-0410-b5e6-96231b3b80d8

Don't bother calculating info unless its needed.  May reduce number of stack slots created.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7584 91177308-0d34-0410-b5e6-96231b3b80d8

* Fix spelling of 'necessary'
* Add a lot more DEBUG output, which is better structured than before
* Fix bug: Jello/2003-08-04-PhysRegLiveFailure.llx

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7583 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, identified by Vikram, reduced by Brian

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7582 91177308-0d34-0410-b5e6-96231b3b80d8

Add patterns for (mov R, R) (mov R, I) and subtracts.  The moves are to enable
testing, the subtracts are because I was in the neighborhood.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7581 91177308-0d34-0410-b5e6-96231b3b80d8

Change comments into something that TableGen can read!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7580 91177308-0d34-0410-b5e6-96231b3b80d8

Allow instructions to have a DAG pattern associated with them.
Define a few preliminary node types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7579 91177308-0d34-0410-b5e6-96231b3b80d8

Rename register classes to be upper case to make it obvious that they are X86
specific in the tree patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7578 91177308-0d34-0410-b5e6-96231b3b80d8

Parse DAG patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7577 91177308-0d34-0410-b5e6-96231b3b80d8

add support for DagInit initializers, which represent DAG patterns

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7576 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7575 91177308-0d34-0410-b5e6-96231b3b80d8