Note to self: SAVE FILES!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12823 91177308-0d34-0410-b5e6-96231b3b80d8

Add an interface to update value numbering

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12822 91177308-0d34-0410-b5e6-96231b3b80d8

Implement InstCombine/select.ll:test13*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12821 91177308-0d34-0410-b5e6-96231b3b80d8

New testcases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12820 91177308-0d34-0410-b5e6-96231b3b80d8

Implement InstCombine/add.ll:test20
Canonicalize add of sign bit constant into a xor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12819 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12818 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite the GCSE pass to be *substantially* simpler, a bit more efficient,
and a bit more powerful

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12817 91177308-0d34-0410-b5e6-96231b3b80d8

Fix spurious warning in release mode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12816 91177308-0d34-0410-b5e6-96231b3b80d8

Silence a spurious warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12815 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code a bit, and fix a bug that was breaking perlbmk

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12814 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug in my checkin last night that was breaking programs using invoke.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12813 91177308-0d34-0410-b5e6-96231b3b80d8

Fix previous patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12811 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly update counters

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12810 91177308-0d34-0410-b5e6-96231b3b80d8

Add note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12809 91177308-0d34-0410-b5e6-96231b3b80d8

Add a couple of more functions that cannot access memory (the intrinsics) and
don't write to memory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12808 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify code a bit, and use alias analysis to allow us to delete unused
call and invoke instructions that are known to not write to memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12807 91177308-0d34-0410-b5e6-96231b3b80d8

New testcases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12806 91177308-0d34-0410-b5e6-96231b3b80d8

Clarify interface

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12805 91177308-0d34-0410-b5e6-96231b3b80d8

Create the output directory if it's not already there

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12804 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify cfg now eliminates all of the branches in this testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12803 91177308-0d34-0410-b5e6-96231b3b80d8

Another testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12802 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase to make sure we can apply the shift to the operands of the select,
eliminating the shifts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12801 91177308-0d34-0410-b5e6-96231b3b80d8

Implement select.ll:test12*

This transforms code like this:

   %C = or %A, %B
   %D = select %cond, %C, %A
into:
   %C = select %cond, %B, 0
   %D = or %A, %C

Since B is often a constant, the select can often be eliminated.  In any case,
this reduces the usage count of A, allowing subsequent optimizations to happen.

This xform applies when the operator is any of:
  add, sub, mul, or, xor, and, shl, shr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12800 91177308-0d34-0410-b5e6-96231b3b80d8

Add testcases for harder combining cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12799 91177308-0d34-0410-b5e6-96231b3b80d8

Fold code like:
  if (C)
    V1 |= V2;

into:
  Vx = V1 | V2;
  V1 = select C, V1, Vx

when the expression can be evaluated unconditionally and is *cheap* to
execute.  This limited form of if conversion is quite handy in lots of cases.
For example, it turns this testcase into straight-line code:

int in0 ; int in1 ; int in2 ; int in3 ;
int in4 ; int in5 ; int in6 ; int in7 ;
int in8 ; int in9 ; int in10; int in11;
int in12; int in13; int in14; int in15;
long output;

void mux(void) {
  output =
      (in0   ?  0x00000001 : 0) | (in1   ?  0x00000002 : 0) |
      (in2   ?  0x00000004 : 0) | (in3   ?  0x00000008 : 0) |
      (in4   ?  0x00000010 : 0) | (in5   ?  0x00000020 : 0) |
      (in6   ?  0x00000040 : 0) | (in7   ?  0x00000080 : 0) |
      (in8   ?  0x00000100 : 0) | (in9   ?  0x00000200 : 0) |
      (in10  ?  0x00000400 : 0) | (in11  ?  0x00000800 : 0) |
      (in12  ?  0x00001000 : 0) | (in13  ?  0x00002000 : 0) |
      (in14  ?  0x00004000 : 0) | (in15  ?  0x00008000 : 0) ;
}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12798 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an obvious bug in the refactoring I did a few days ago

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12797 91177308-0d34-0410-b5e6-96231b3b80d8

Expand on adding an intrinsic.  Move section to before adding an instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12796 91177308-0d34-0410-b5e6-96231b3b80d8

Reversed the order of the llvm.writeport() operands so that the value
is listed first and the address is listed second.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12795 91177308-0d34-0410-b5e6-96231b3b80d8

Fold binary operators with a constant operand into select instructions
that have a constant operand.  This implements
add.ll:test19, shift.ll:test15*, and others that are not tested

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12794 91177308-0d34-0410-b5e6-96231b3b80d8

Implement select.ll:test11

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12793 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12792 91177308-0d34-0410-b5e6-96231b3b80d8

Add definition list to each live interval.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12791 91177308-0d34-0410-b5e6-96231b3b80d8

Fixed some spelling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12790 91177308-0d34-0410-b5e6-96231b3b80d8

Modified the prototypes of the llvm.readport and llvm.writeport intrinsics
so that they do not specify specific integer sizes (as those are
constrained by the code generator, not the intrinsic itself).

Also put these intrinsics into their own "Operating System" section.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12789 91177308-0d34-0410-b5e6-96231b3b80d8

Added the llvm.readport and llvm.writeport intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12788 91177308-0d34-0410-b5e6-96231b3b80d8

Changed assertions to error messages.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12787 91177308-0d34-0410-b5e6-96231b3b80d8

Changes recommended by Chris:
InstSelectSimple.cpp:
  Change the checks for proper I/O port address size into an exit() instead
  of an assertion.  Assertions aren't used in Release builds, and handling
  this error should be graceful (not that this counts as graceful, but it's
  more graceful).

  Modified the generation of the IN/OUT instructions to have 0 arguments.
X86InstrInfo.td:
  Added the OpSize attribute to the 16 bit IN and OUT instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12786 91177308-0d34-0410-b5e6-96231b3b80d8

Implement InstCombine/cast-propagate.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12784 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12783 91177308-0d34-0410-b5e6-96231b3b80d8

Added the llvm.readport and llvm.writeport intrinsics for x86.  These do
I/O port instructions on x86.  The specific code sequence is tailored to
the parameters and return value of the intrinsic call.
Added the ability for implicit defintions to be printed in the Instruction
Printer.
Added the ability for RawFrm instruction to print implict uses and
defintions with correct comma output.  This required adjustment to some
methods so that a leading comma would or would not be printed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12782 91177308-0d34-0410-b5e6-96231b3b80d8

Added the llvm.readport and llvm.writeport intrinsics.
The Verifier ensures that their parameters are of integral types and have
the correct sign, but it does not enforce any size restrictions because
such restrictions are platform dependent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12781 91177308-0d34-0410-b5e6-96231b3b80d8

Added the llvm.readport and llvm.writeport intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12780 91177308-0d34-0410-b5e6-96231b3b80d8

Implement ScalarRepl/select_promote.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12779 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase, corresponding to phi_promote.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12778 91177308-0d34-0410-b5e6-96231b3b80d8

This file does not appear to be included anywhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12777 91177308-0d34-0410-b5e6-96231b3b80d8

Disable strict alias analysis in the backend c compiler, as the code we
generate is not TBAA safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12774 91177308-0d34-0410-b5e6-96231b3b80d8

Revert previous patch, I'm a moron :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12773 91177308-0d34-0410-b5e6-96231b3b80d8

Right, we break strict aliasing requirements.  Make sure to disable strict
aliasing in the C compiler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12772 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the "really gross hacks" that are there to deal with recursive functions.
Now we collect all of the call sites we are interested in inlining, then inline
them.  This entirely avoids issues with trying to inline a call site we got by
inlining another call site.  This also eliminates iterator invalidation issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12770 91177308-0d34-0410-b5e6-96231b3b80d8

Implement InstCombine/select.ll:test[7-10]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12769 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more cases we should combine

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12768 91177308-0d34-0410-b5e6-96231b3b80d8

This file is no longer used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12767 91177308-0d34-0410-b5e6-96231b3b80d8

Don't include InstrSelectionSupport.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12766 91177308-0d34-0410-b5e6-96231b3b80d8

Move ChooseRegOrImmed() prototype here, from InstrSelectionSupport.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12765 91177308-0d34-0410-b5e6-96231b3b80d8

Don't include InstrSelectionSupport.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12764 91177308-0d34-0410-b5e6-96231b3b80d8

Implement test/Regression/Transforms/InstCombine/getelementptr_index.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12762 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12761 91177308-0d34-0410-b5e6-96231b3b80d8

Fix insertion of SelectInsts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12760 91177308-0d34-0410-b5e6-96231b3b80d8

Don't print [%reg + 0], just print [%reg]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12759 91177308-0d34-0410-b5e6-96231b3b80d8

First version of code to handle loads. Stub function for handling stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12758 91177308-0d34-0410-b5e6-96231b3b80d8

Support loading arguments from %I0...%I5 into virtual registers in
function prologues, and fix an off-by-one in visitCallInst that was
putting call args into the wrong registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12757 91177308-0d34-0410-b5e6-96231b3b80d8

It's setting up the call args right now, but on the callee side, it's
trying to get incoming args off the stack, instead of the %i0...%i6 regs,
which is wrong.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12756 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that crashed scalarevolution on sparc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12755 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug Brian found.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12754 91177308-0d34-0410-b5e6-96231b3b80d8

We have a global reg. allocator now -- thanks to Alkis.
Fix a typo.
Add a project I've always thought would be cool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12747 91177308-0d34-0410-b5e6-96231b3b80d8

New file that can be useful for hand inspection of assembly required for certain
kinds of instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12746 91177308-0d34-0410-b5e6-96231b3b80d8

This is a start on handling setcc instructions.  As the comment notes, we
have no good way of handling this until the code generator is improved.
We should probably just emit V9 instructions in the meantime.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12745 91177308-0d34-0410-b5e6-96231b3b80d8

andd subcc instructions which is used to create the 'cmp' pseudo instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12744 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid emitting an extra copy on each 32-bit operation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12743 91177308-0d34-0410-b5e6-96231b3b80d8

Make generation of stack-slot loads and copies less ugly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12742 91177308-0d34-0410-b5e6-96231b3b80d8

Fix bug in printing loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12741 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for shift instructions, wrap some long lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12740 91177308-0d34-0410-b5e6-96231b3b80d8

Fix encoding of existing shift instructions, add rr shifts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12739 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12738 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bunch more instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12737 91177308-0d34-0410-b5e6-96231b3b80d8

Merge my changes with brians

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12736 91177308-0d34-0410-b5e6-96231b3b80d8

Add in some things I forgot, which Chris helpfully reminded me of...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12735 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for the "Y" register, used by MUL & DIV.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12734 91177308-0d34-0410-b5e6-96231b3b80d8

Add UDIV, SDIV, and a few variants of WR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12733 91177308-0d34-0410-b5e6-96231b3b80d8

Preliminary support for getting 64-bit integer constants into registers.
Preliminary support for division. It's gross because you have to initialize
the "Y" register, which is the top 32 bits of the thing you're dividing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12732 91177308-0d34-0410-b5e6-96231b3b80d8

Prune unnecessary #includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12731 91177308-0d34-0410-b5e6-96231b3b80d8

Simple delay slot filler pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12730 91177308-0d34-0410-b5e6-96231b3b80d8

Add references to delay slot filler pass.
Fill in addPassesToJITCompile method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12729 91177308-0d34-0410-b5e6-96231b3b80d8

First attempt at handling frame index elimination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12728 91177308-0d34-0410-b5e6-96231b3b80d8

First attempt at special-casing printing of [%reg + offset] for
ld/st instructions - doesn't seem to work yet, but I think it's
just a typo or something somewhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12727 91177308-0d34-0410-b5e6-96231b3b80d8

Delete reference to "the Mach-O Runtime ABI".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12726 91177308-0d34-0410-b5e6-96231b3b80d8

Deal with call return values.
Don't put NOPs in delay slots at all. We'll have a fix-up pass later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12725 91177308-0d34-0410-b5e6-96231b3b80d8

Adding kimwitu++ license.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12719 91177308-0d34-0410-b5e6-96231b3b80d8

Bugs fixed new features implemented

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12716 91177308-0d34-0410-b5e6-96231b3b80d8

file based off InstSelectSimple.cpp, slowly being replaced by generated code from the really simple X86 instruction selector tablegen backend

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12715 91177308-0d34-0410-b5e6-96231b3b80d8

Tablgen files for really simple instruction selector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12714 91177308-0d34-0410-b5e6-96231b3b80d8

Tablegen backend for really simple instruction selector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12713 91177308-0d34-0410-b5e6-96231b3b80d8

add tablgen backend for really simple instruction selector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12712 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR313: [x86] JIT miscompiles unsigned short to floating point

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12711 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect encoding of some ADC and SBB instuctions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12710 91177308-0d34-0410-b5e6-96231b3b80d8

Added licensing information for treecc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12703 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a minor bug in previous checking
Enable folding of long seteq/setne comparisons into branches and select instructions
Implement unfolded long relational comparisons against a constants a bit more efficiently

Folding comparisons changes code that looks like this:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %EDX, DWORD PTR [%ESP + 8]
        mov %ECX, %EAX
        or %ECX, %EDX
        sete %CL
        test %CL, %CL
        je .LBB2 # PC rel: F

into code that looks like this:
        mov %EAX, DWORD PTR [%ESP + 4]
        mov %EDX, DWORD PTR [%ESP + 8]
        mov %ECX, %EAX
        or %ECX, %EDX
        jne .LBB2 # PC rel: F

This speeds up 186.crafty by 6% with llc-ls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12702 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap at 80 cols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12701 91177308-0d34-0410-b5e6-96231b3b80d8

Minor cleanups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12700 91177308-0d34-0410-b5e6-96231b3b80d8

Document new option

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12699 91177308-0d34-0410-b5e6-96231b3b80d8