new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21784 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR557 and basictest[34].ll.

This makes reassociate realize that loads should be treated as unmovable, and
gives distinct ranks to distinct values defined in the same basic block, allowing
reassociate to do its thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21783 91177308-0d34-0410-b5e6-96231b3b80d8

new testcases for PR557

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21782 91177308-0d34-0410-b5e6-96231b3b80d8

Add debugging information

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21781 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate gotos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21780 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap long lines.  Fix "warning: conflicting types for built-in function 'memset'"
warning from the CBE+GCC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21779 91177308-0d34-0410-b5e6-96231b3b80d8

Improve reassociation handling of inverses, implementing inverses.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21778 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21777 91177308-0d34-0410-b5e6-96231b3b80d8

clean up and modernize this pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21776 91177308-0d34-0410-b5e6-96231b3b80d8

Strength reduce SAR into SHR if there is no way sign bits could be shifted
in.  This tends to get cases like this:

  X = cast ubyte to int
  Y = shr int X, ...

Tested by: shift.ll:test24

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21775 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21774 91177308-0d34-0410-b5e6-96231b3b80d8

Update my info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21773 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor some code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21772 91177308-0d34-0410-b5e6-96231b3b80d8

Handle some simple cases where we can see that values get annihilated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21771 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a miscompilation of crafty by clobbering the "A" variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21770 91177308-0d34-0410-b5e6-96231b3b80d8

Move licm after reassociate, add some cond-propagate passes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21769 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite the guts of the reassociate pass to be more efficient and logical.  Instead
of trying to do local reassociation tweaks at each level, only process an expression
tree once (at its root).  This does not improve the reassociation pass in any real way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21768 91177308-0d34-0410-b5e6-96231b3b80d8

* Break tests into individual cases so its easier to see what's going on.
* Add four new cases to handle the new optimizations for strlen, dealing
  with:
    strlen(x) != 0 -> *x != 0
    strlen(x) == 0 -> *x == 0
  for each of two cases: strlen(x) != 0 and strlen(x) == 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21767 91177308-0d34-0410-b5e6-96231b3b80d8

* Add two strlen optimizations:
    strlen(x) != 0 -> *x != 0
    strlen(x) == 0 -> *x == 0
* Change nested statistics to use style of other LLVM statistics so that
  only the name of the optimization (simplify-libcalls) is used as the
  statistic name, and the description indicates which specific all is
  optimized. Cuts down on some redundancy and saves a few bytes of space.
* Make note of stpcpy optimization that could be done.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21766 91177308-0d34-0410-b5e6-96231b3b80d8

Don't increment the counter unless the debug flag is set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21762 91177308-0d34-0410-b5e6-96231b3b80d8

Convert shifts to muls to assist reassociation.  This implements
Reassociate/shifttest.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21761 91177308-0d34-0410-b5e6-96231b3b80d8

test that reassociate can handle shifts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21760 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify the code and rearrange it.  No major functionality changes here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21759 91177308-0d34-0410-b5e6-96231b3b80d8

Silence VC++ warnings about unsafe mixing of ints and bools with the | operator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21758 91177308-0d34-0410-b5e6-96231b3b80d8

add note about bit builtins newly supported

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21757 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21756 91177308-0d34-0410-b5e6-96231b3b80d8

fix bugs in my previous commit clarify the return type of the bit ops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21755 91177308-0d34-0410-b5e6-96231b3b80d8

add a couple of notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21754 91177308-0d34-0410-b5e6-96231b3b80d8

document calling convention extensions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21753 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead (always dynamically false) flags

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21752 91177308-0d34-0410-b5e6-96231b3b80d8

encode calling conventions for call/invoke instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21751 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21750 91177308-0d34-0410-b5e6-96231b3b80d8

encode function calling convs in the bytecode file.  invoke and call are
still to come.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21749 91177308-0d34-0410-b5e6-96231b3b80d8

parse new calling conv specifiers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21748 91177308-0d34-0410-b5e6-96231b3b80d8

wrap a longline

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21747 91177308-0d34-0410-b5e6-96231b3b80d8

add support for explicit calling conventions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21746 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for explicit calling conventions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21745 91177308-0d34-0410-b5e6-96231b3b80d8

New file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21744 91177308-0d34-0410-b5e6-96231b3b80d8

use splice instead of remove/insert for a minor speedup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21743 91177308-0d34-0410-b5e6-96231b3b80d8

remove some ugly hacks that are no longer needed since andrew removed the
varargs munging code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21742 91177308-0d34-0410-b5e6-96231b3b80d8

BAD typeo which caused many testsuite failures last night.  Note to self, do
not change code after testing it without retesting!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21741 91177308-0d34-0410-b5e6-96231b3b80d8

clean up the CBE output a bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21740 91177308-0d34-0410-b5e6-96231b3b80d8

add tail marker as a comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21739 91177308-0d34-0410-b5e6-96231b3b80d8

Make the stub functions be tail calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21738 91177308-0d34-0410-b5e6-96231b3b80d8

Preserve tail marker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21737 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/Inline/inline-tail.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21736 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure to not accidentally get a bogus tail marker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21735 91177308-0d34-0410-b5e6-96231b3b80d8

preserve the tail marker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21734 91177308-0d34-0410-b5e6-96231b3b80d8

add a default arg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21733 91177308-0d34-0410-b5e6-96231b3b80d8

Don't forget these are calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21730 91177308-0d34-0410-b5e6-96231b3b80d8

lex tail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21729 91177308-0d34-0410-b5e6-96231b3b80d8

test that tail call survives bytecodification

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21728 91177308-0d34-0410-b5e6-96231b3b80d8

add bytecode reader support for tail calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21727 91177308-0d34-0410-b5e6-96231b3b80d8

Hilight tail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21726 91177308-0d34-0410-b5e6-96231b3b80d8

These are legal for tail calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21723 91177308-0d34-0410-b5e6-96231b3b80d8

Add a 'tail' marker for call instructions, patch contributed by
Alexander Friedman.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21722 91177308-0d34-0410-b5e6-96231b3b80d8

Add info on new 'tail' marker

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21721 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap long lines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21720 91177308-0d34-0410-b5e6-96231b3b80d8

DCE intrinsic instructions without side effects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21719 91177308-0d34-0410-b5e6-96231b3b80d8

These intrinsics do not access memory

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21718 91177308-0d34-0410-b5e6-96231b3b80d8

Teach instcombine propagate zeroness through shl instructions, implementing
and.ll:test31

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21717 91177308-0d34-0410-b5e6-96231b3b80d8

new testcases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21716 91177308-0d34-0410-b5e6-96231b3b80d8

Implement shift.ll:test23.  If we are shifting right then immediately truncating
the result, turn signed shift rights into unsigned shift rights if possible.

This leads to later simplification and happens *often* in 176.gcc.  For example,
this testcase:

struct xxx { unsigned int code : 8; };
enum codes { A, B, C, D, E, F };
int foo(struct xxx *P) {
  if ((enum codes)P->code == A)
     bar();
}

used to be compiled to:

int %foo(%struct.xxx* %P) {
        %tmp.1 = getelementptr %struct.xxx* %P, int 0, uint 0           ; <uint*> [#uses=1]
        %tmp.2 = load uint* %tmp.1              ; <uint> [#uses=1]
        %tmp.3 = cast uint %tmp.2 to int                ; <int> [#uses=1]
        %tmp.4 = shl int %tmp.3, ubyte 24               ; <int> [#uses=1]
        %tmp.5 = shr int %tmp.4, ubyte 24               ; <int> [#uses=1]
        %tmp.6 = cast int %tmp.5 to sbyte               ; <sbyte> [#uses=1]
        %tmp.8 = seteq sbyte %tmp.6, 0          ; <bool> [#uses=1]
        br bool %tmp.8, label %then, label %UnifiedReturnBlock

Now it is compiled to:

        %tmp.1 = getelementptr %struct.xxx* %P, int 0, uint 0           ; <uint*> [#uses=1]
        %tmp.2 = load uint* %tmp.1              ; <uint> [#uses=1]
        %tmp.2 = cast uint %tmp.2 to sbyte              ; <sbyte> [#uses=1]
        %tmp.8 = seteq sbyte %tmp.2, 0          ; <bool> [#uses=1]
        br bool %tmp.8, label %then, label %UnifiedReturnBlock

which is the difference between this:

foo:
        subl $4, %esp
        movl 8(%esp), %eax
        movl (%eax), %eax
        shll $24, %eax
        sarl $24, %eax
        testb %al, %al
        jne .LBBfoo_2

and this:

foo:
        subl $4, %esp
        movl 8(%esp), %eax
        movl (%eax), %eax
        testb %al, %al
        jne .LBBfoo_2

This occurs 3243 times total in the External tests, 215x in povray,
6x in each f2c'd program, 1451x in 176.gcc, 7x in crafty, 20x in perl,
25x in gap, 3x in m88ksim, 25x in ijpeg.

Maybe this will cause a little jump on gcc tommorow :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21715 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21714 91177308-0d34-0410-b5e6-96231b3b80d8

Implement xor.ll:test22

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21713 91177308-0d34-0410-b5e6-96231b3b80d8

implement and.ll:test30 and set.ll:test21

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21712 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for unneeded and

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21711 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21710 91177308-0d34-0410-b5e6-96231b3b80d8

implement or.ll:test20

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21709 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21708 91177308-0d34-0410-b5e6-96231b3b80d8

* Order #includes alphabetically
* Remove commented-out debug printouts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21707 91177308-0d34-0410-b5e6-96231b3b80d8

Remove extra blank line

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21706 91177308-0d34-0410-b5e6-96231b3b80d8

* Use consistent spacing for function arguments
* Output single-character strings as chars

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21705 91177308-0d34-0410-b5e6-96231b3b80d8

Remove vim settings from source code; people should use llvm/utils/vim/vimrc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21704 91177308-0d34-0410-b5e6-96231b3b80d8

Convert tabs to spaces

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21703 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR541

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21702 91177308-0d34-0410-b5e6-96231b3b80d8

add support for undef values of opaque type, addressing PR541

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21701 91177308-0d34-0410-b5e6-96231b3b80d8

Add some extra checks.  Opaque types don't have a null marker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21700 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21699 91177308-0d34-0410-b5e6-96231b3b80d8

When hitting an unsupported intrinsic, actually print it
Lower debug info to noops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21698 91177308-0d34-0410-b5e6-96231b3b80d8

ctpop lowering in legalize

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21697 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug compimling Ruby, fixing this testcase:
LowerSetJmp/2005-05-05-OldUses.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21696 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for a bug Jim found

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21695 91177308-0d34-0410-b5e6-96231b3b80d8

fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21693 91177308-0d34-0410-b5e6-96231b3b80d8

Well, add support for ct* for 21264 only.
21164 is broken until expand works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21692 91177308-0d34-0410-b5e6-96231b3b80d8

Make promoteOp work for CT*

Proof?

ubyte %bar(ubyte %x) {
entry:
        %tmp.1 = call ubyte %llvm.ctlz( ubyte %x )
        ret ubyte %tmp.1
}

==>

zapnot $16,1,$0
CTLZ $0,$0
subq $0,56,$0
zapnot $0,1,$0
ret $31,($26),1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21691 91177308-0d34-0410-b5e6-96231b3b80d8

Instcombine: cast (X != 0) to int, cast (X == 1) to int  -> X iff X has only the low bit set.

This implements set.ll:test20.

This triggers 2x on povray, 9x on mesa, 11x on gcc, 2x on crafty, 1x on eon,
6x on perlbmk and 11x on m88ksim.

It allows us to compile these two functions into the same code:

struct s { unsigned int bit : 1; };
unsigned foo(struct s *p) {
  if (p->bit)
    return 1;
  else
    return 0;
}
unsigned bar(struct s *p) { return p->bit; }

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21690 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21689 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the IsDigitOptimization for simplifying calls to the isdigit
library function:
  isdigit(chr) -> 0 or 1 if chr is constant
  isdigit(chr) -> chr - '0' <= 9 otherwise

Although there are many calls to isdigit in llvm-test, most of them are
compiled away by macros leaving only this:

2 MultiSource/Applications/hexxagon

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21688 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure both optimization cases get tested.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21687 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test case for the IsDigitOptimization class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21686 91177308-0d34-0410-b5e6-96231b3b80d8

comment fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21685 91177308-0d34-0410-b5e6-96231b3b80d8

see if the legalize code propery compensates for the additional zeros introduced

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21684 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the descriptions of WHILE...END and RROT. Thanks to Lalo Martins
for pointing out these errors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21683 91177308-0d34-0410-b5e6-96231b3b80d8

WHILE does not "pop" a value, it "examines" to top of stack. Make this
clear in the applicable comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21682 91177308-0d34-0410-b5e6-96231b3b80d8

added a use of ctpop to debug stuff with

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21681 91177308-0d34-0410-b5e6-96231b3b80d8

fixup argument

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21680 91177308-0d34-0410-b5e6-96231b3b80d8

* Correct the function prototypes for some of the functions to match the
  actual spec (int -> uint)
* Add the ability to get/cache the strlen function prototype.
* Make sure generated values are appropriately named for debugging purposes
* Add the SPrintFOptimiation for 4 casts of sprintf optimization:
    sprintf(str,cstr) -> llvm.memcpy(str,cstr) (if cstr has no %)
    sprintf(str,"")   -> store sbyte 0, str
    sprintf(str,"%s",src) -> llvm.memcpy(str,src) (if src is constant)
    sprintf(str,"%c",chr) -> store chr, str   ; store sbyte 0, str+1

The sprintf optimization didn't fire as much as I had hoped:

  2 MultiSource/Applications/SPASS
  5 MultiSource/Benchmarks/McCat/18-imp
 22 MultiSource/Benchmarks/Prolangs-C/TimberWolfMC
  1 MultiSource/Benchmarks/Prolangs-C/assembler
  6 MultiSource/Benchmarks/Prolangs-C/unix-smail
  2 MultiSource/Benchmarks/mediabench/mpeg2/mpeg2dec

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21679 91177308-0d34-0410-b5e6-96231b3b80d8

Minor clean-ups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21678 91177308-0d34-0410-b5e6-96231b3b80d8

initial descriptions of count intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21677 91177308-0d34-0410-b5e6-96231b3b80d8