Wrong size for load

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30610 91177308-0d34-0410-b5e6-96231b3b80d8

Can't move a load node if it's chain is not used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30609 91177308-0d34-0410-b5e6-96231b3b80d8

Various random and minor code cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30608 91177308-0d34-0410-b5e6-96231b3b80d8

test that the no_dead_strip directive is emitted on darwin-x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30607 91177308-0d34-0410-b5e6-96231b3b80d8

print the preds of each MBB

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30606 91177308-0d34-0410-b5e6-96231b3b80d8

Compile:
int x __attribute__((used));

to:

        .data
.comm _x,4              ; 'x'
        .no_dead_strip  _x

on both x86 and ppc darwin targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30605 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for targets that want to do something with the llvm.used list,
because they have an aggressive linker that does dead code stripping.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30604 91177308-0d34-0410-b5e6-96231b3b80d8

order this properly to avoid warnings in TargetAsmInfo.cpp.  Add a comment
in a format that matches every other ivars in this class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30603 91177308-0d34-0410-b5e6-96231b3b80d8

Accidental enable of bad code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30601 91177308-0d34-0410-b5e6-96231b3b80d8

Fix chain dropping in load and drop unused stores in ret blocks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30600 91177308-0d34-0410-b5e6-96231b3b80d8

Regression test for PR#922.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30599 91177308-0d34-0410-b5e6-96231b3b80d8

more notes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30598 91177308-0d34-0410-b5e6-96231b3b80d8

Core antialiasing for load and store.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30597 91177308-0d34-0410-b5e6-96231b3b80d8

Fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30596 91177308-0d34-0410-b5e6-96231b3b80d8

Fix jump tables to match gcc (and the ABI and whatnot)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30594 91177308-0d34-0410-b5e6-96231b3b80d8

Add support for other relocation bases to jump tables, as well as custom asm directives

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30593 91177308-0d34-0410-b5e6-96231b3b80d8

basic jump table test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30592 91177308-0d34-0410-b5e6-96231b3b80d8

jump table note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30591 91177308-0d34-0410-b5e6-96231b3b80d8

PIC jump table entries are always 32-bit. This fixes PIC jump table support on X86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30590 91177308-0d34-0410-b5e6-96231b3b80d8

Enable 'predsimplify' optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30589 91177308-0d34-0410-b5e6-96231b3b80d8

Style changes only. Remove dead code, fix a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30588 91177308-0d34-0410-b5e6-96231b3b80d8

Be far more careful when splitting a loop header, either to form a preheader
or when splitting loops with a common header into multiple loops.  In particular
the old code would always insert the preheader before the old loop header.  This
is disasterous in cases where the loop hasn't been rotated.  For example, it can
produce code like:

        .. outside the loop...
        jmp LBB1_2      #bb13.outer
LBB1_1: #bb1
        movsd 8(%esp,%esi,8), %xmm1
        mulsd (%edi), %xmm1
        addsd %xmm0, %xmm1
        addl $24, %edi
        incl %esi
        jmp LBB1_3      #bb13
LBB1_2: #bb13.outer
        leal (%edx,%eax,8), %edi
        pxor %xmm1, %xmm1
        xorl %esi, %esi
LBB1_3: #bb13
        movapd %xmm1, %xmm0
        cmpl $4, %esi
        jl LBB1_1       #bb1

Note that the loop body is actually LBB1_1 + LBB1_3, which means that the
loop now contains an uncond branch WITHIN it to jump around the inserted
loop header (LBB1_2).  Doh.

This patch changes the preheader insertion code to insert it in the right
spot, producing this code:

        ... outside the loop, fall into the header ...
LBB1_1: #bb13.outer
        leal (%edx,%eax,8), %esi
        pxor %xmm0, %xmm0
        xorl %edi, %edi
        jmp LBB1_3      #bb13
LBB1_2: #bb1
        movsd 8(%esp,%edi,8), %xmm0
        mulsd (%esi), %xmm0
        addsd %xmm1, %xmm0
        addl $24, %esi
        incl %edi
LBB1_3: #bb13
        movapd %xmm0, %xmm1
        cmpl $4, %edi
        jl LBB1_2       #bb1

Totally crazy, no branch in the loop! :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30587 91177308-0d34-0410-b5e6-96231b3b80d8

Teach UpdateDomInfoForRevectoredPreds to handle revectored preds that are not
reachable, making it general purpose enough for use by InsertPreheaderForLoop.
Eliminate custom dominfo updating code in InsertPreheaderForLoop, using
UpdateDomInfoForRevectoredPreds instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30586 91177308-0d34-0410-b5e6-96231b3b80d8

remove misleading comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30585 91177308-0d34-0410-b5e6-96231b3b80d8

add method, correct comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30584 91177308-0d34-0410-b5e6-96231b3b80d8

Delete dead code; fix 80 col violations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30583 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30581 91177308-0d34-0410-b5e6-96231b3b80d8

Use "%llvmgxx" instead of "llvm-g++".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30580 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for better rotate left and mask support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30579 91177308-0d34-0410-b5e6-96231b3b80d8

Regression testcase for PR855.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30578 91177308-0d34-0410-b5e6-96231b3b80d8

Fold AND and ROTL more often

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30577 91177308-0d34-0410-b5e6-96231b3b80d8

remove extra white spaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30576 91177308-0d34-0410-b5e6-96231b3b80d8

Use iterative algorith to assign DFS number. This reduces
call stack depth.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30575 91177308-0d34-0410-b5e6-96231b3b80d8

don't allow 'imm' or specific imms, like '1' on the LHS of a binop.
This shrinks X86GenDAGISel by ~330 lines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30574 91177308-0d34-0410-b5e6-96231b3b80d8

Make it work for DAG combine of multi-value nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30573 91177308-0d34-0410-b5e6-96231b3b80d8

Fit to 80 columns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30572 91177308-0d34-0410-b5e6-96231b3b80d8

core corrections

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30570 91177308-0d34-0410-b5e6-96231b3b80d8

Use abstract class to facilitate dlopen() interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30569 91177308-0d34-0410-b5e6-96231b3b80d8

Basic "in frame" alias analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30568 91177308-0d34-0410-b5e6-96231b3b80d8

more condition codes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30567 91177308-0d34-0410-b5e6-96231b3b80d8

if a constant can't be an immediate, add it to the constant pool

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30566 91177308-0d34-0410-b5e6-96231b3b80d8

fold (aext (and (trunc x), cst)) -> (and x, cst).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30561 91177308-0d34-0410-b5e6-96231b3b80d8

Check the right value type.  This fixes 186.crafty on x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30560 91177308-0d34-0410-b5e6-96231b3b80d8

implemented

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30559 91177308-0d34-0410-b5e6-96231b3b80d8

Compile:

int %test(ulong *%tmp) {
        %tmp = load ulong* %tmp         ; <ulong> [#uses=1]
        %tmp.mask = shr ulong %tmp, ubyte 50            ; <ulong> [#uses=1]
        %tmp.mask = cast ulong %tmp.mask to ubyte
        %tmp2 = and ubyte %tmp.mask, 3          ; <ubyte> [#uses=1]
        %tmp2 = cast ubyte %tmp2 to int         ; <int> [#uses=1]
        ret int %tmp2
}

to:

_test:
        movl 4(%esp), %eax
        movl 4(%eax), %eax
        shrl $18, %eax
        andl $3, %eax
        ret

instead of:

_test:
        movl 4(%esp), %eax
        movl 4(%eax), %eax
        shrl $18, %eax
        # TRUNCATE movb %al, %al
        andb $3, %al
        movzbl %al, %eax
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30558 91177308-0d34-0410-b5e6-96231b3b80d8

Generalize (zext (truncate x)) and (sext (truncate x)) folding to work when
the src/dst are not the same size.  This catches things like "truncate
32-bit X to 8 bits, then zext to 16", which happens a bit on X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30557 91177308-0d34-0410-b5e6-96231b3b80d8

Fit in 80-cols

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30556 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/IndVarsSimplify/2006-09-20-LFTR-Crash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30555 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30554 91177308-0d34-0410-b5e6-96231b3b80d8

Fix compile error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30553 91177308-0d34-0410-b5e6-96231b3b80d8

Don't rewrite ConstantExpr::get.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30552 91177308-0d34-0410-b5e6-96231b3b80d8

Once we're down to "setcc type constant1, constant2", at least come up
with the right answer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30550 91177308-0d34-0410-b5e6-96231b3b80d8

Adding codegeneration for StdCall & FastCall calling conventions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30549 91177308-0d34-0410-b5e6-96231b3b80d8

Account for pseudo-ops correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30548 91177308-0d34-0410-b5e6-96231b3b80d8

The DarwinAsmPrinter need not check for isDarwin.  createPPCAsmPrinterPass
should create the right asmprinter subclass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30542 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap some darwin'isms with isDarwin checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30541 91177308-0d34-0410-b5e6-96231b3b80d8

Use a total ordering to compare instructions.
Fixes infinite loop in resolve().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30540 91177308-0d34-0410-b5e6-96231b3b80d8

simplify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30535 91177308-0d34-0410-b5e6-96231b3b80d8

catch constants more often

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30534 91177308-0d34-0410-b5e6-96231b3b80d8

catch another constant

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30533 91177308-0d34-0410-b5e6-96231b3b80d8

clarify with test case

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30531 91177308-0d34-0410-b5e6-96231b3b80d8

Add Note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30530 91177308-0d34-0410-b5e6-96231b3b80d8

Trim the home directory from the dejagnu test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30519 91177308-0d34-0410-b5e6-96231b3b80d8

item done

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30518 91177308-0d34-0410-b5e6-96231b3b80d8

Compile:
int test3(int a, int b) { return (a < 0) ? a : 0; }

to:

_test3:
        srawi r2, r3, 31
        and r3, r2, r3
        blr

instead of:

_test3:
        cmpwi cr0, r3, 1
        li r2, 0
        blt cr0, LBB2_2 ;entry
LBB2_1: ;entry
        mr r3, r2
LBB2_2: ;entry
        blr

This implements: PowerPC/select_lt0.ll:seli32_a_a

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30517 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30516 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30515 91177308-0d34-0410-b5e6-96231b3b80d8

Fold the full generality of (any_extend (truncate x))

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30514 91177308-0d34-0410-b5e6-96231b3b80d8

Two things:
1. teach SimplifySetCC that '(srl (ctlz x), 5) == 0' is really x != 0.
2. Teach visitSELECT_CC to use SimplifySetCC instead of calling it and
   ignoring the result.  This allows us to compile:

bool %test(ulong %x) {
  %tmp = setlt ulong %x, 4294967296
  ret bool %tmp
}

to:

_test:
        cntlzw r2, r3
        cmplwi cr0, r3, 1
        srwi r2, r2, 5
        li r3, 0
        beq cr0, LBB1_2 ;
LBB1_1: ;
        mr r3, r2
LBB1_2: ;
        blr

instead of:

_test:
        addi r2, r3, -1
        cntlzw r2, r2
        cntlzw r3, r3
        srwi r2, r2, 5
        cmplwi cr0, r2, 0
        srwi r2, r3, 5
        li r3, 0
        bne cr0, LBB1_2 ;
LBB1_1: ;
        mr r3, r2
LBB1_2: ;
        blr

This isn't wonderful, but it's an improvement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30513 91177308-0d34-0410-b5e6-96231b3b80d8

This is already done

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30512 91177308-0d34-0410-b5e6-96231b3b80d8

We went through all that trouble to compute whether it was safe to transform
this comparison, but never checked it.  Whoops, no wonder we miscompiled
177.mesa!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30511 91177308-0d34-0410-b5e6-96231b3b80d8

Improve PPC64 equality comparisons like PPC32 comparisons.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30510 91177308-0d34-0410-b5e6-96231b3b80d8

Two improvements:

1. Codegen this comparison:
     if (X == 0x8000)

as:

        cmplwi cr0, r3, 32768
        bne cr0, LBB1_2 ;cond_next

instead of:

        lis r2, 0
        ori r2, r2, 32768
        cmpw cr0, r3, r2
        bne cr0, LBB1_2 ;cond_next

2. Codegen this comparison:
      if (X == 0x12345678)

as:

        xoris r2, r3, 4660
        cmplwi cr0, r2, 22136
        bne cr0, LBB1_2 ;cond_next

instead of:

        lis r2, 4660
        ori r2, r2, 22136
        cmpw cr0, r3, r2
        bne cr0, LBB1_2 ;cond_next

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30509 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note that we should match rlwnm better

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30508 91177308-0d34-0410-b5e6-96231b3b80d8

Legalize is no longer limited to cleverness with just constant shift amounts.
Allow it to be clever when possible and fall back to the gross code when needed.

This allows us to compile:

long long foo1(long long X, int C) {
  return X << (C|32);
}
long long foo2(long long X, int C) {
  return X << (C&~32);
}

to:
_foo1:
        rlwinm r2, r5, 0, 27, 31
        slw r3, r4, r2
        li r4, 0
        blr

        .globl  _foo2
        .align  4
_foo2:
        rlwinm r2, r5, 0, 27, 25
        subfic r5, r2, 32
        slw r3, r3, r2
        srw r5, r4, r5
        or r3, r3, r5
        slw r4, r4, r2
        blr

instead of:

_foo1:
        ori r2, r5, 32
        subfic r5, r2, 32
        addi r6, r2, -32
        srw r5, r4, r5
        slw r3, r3, r2
        slw r6, r4, r6
        or r3, r3, r5
        slw r4, r4, r2
        or r3, r3, r6
        blr

        .globl  _foo2
        .align  4
_foo2:
        rlwinm r2, r5, 0, 27, 25
        subfic r5, r2, 32
        addi r6, r2, -32
        srw r5, r4, r5
        slw r3, r3, r2
        slw r6, r4, r6
        or r3, r3, r5
        slw r4, r4, r2
        or r3, r3, r6
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30507 91177308-0d34-0410-b5e6-96231b3b80d8

Expand 64-bit shifts more optimally if we know that the high bit of the
shift amount is one or zero.  For example, for:

long long foo1(long long X, int C) {
  return X << (C|32);
}

long long foo2(long long X, int C) {
  return X << (C&~32);
}

we get:

_foo1:
        movb $31, %cl
        movl 4(%esp), %edx
        andb 12(%esp), %cl
        shll %cl, %edx
        xorl %eax, %eax
        ret
_foo2:
        movb $223, %cl
        movl 4(%esp), %eax
        movl 8(%esp), %edx
        andb 12(%esp), %cl
        shldl %cl, %eax, %edx
        shll %cl, %eax
        ret

instead of:

_foo1:
        subl $4, %esp
        movl %ebx, (%esp)
        movb $32, %bl
        movl 8(%esp), %eax
        movl 12(%esp), %edx
        movb %bl, %cl
        orb 16(%esp), %cl
        shldl %cl, %eax, %edx
        shll %cl, %eax
        xorl %ecx, %ecx
        testb %bl, %bl
        cmovne %eax, %edx
        cmovne %ecx, %eax
        movl (%esp), %ebx
        addl $4, %esp
        ret
_foo2:
        subl $4, %esp
        movl %ebx, (%esp)
        movb $223, %cl
        movl 8(%esp), %eax
        movl 12(%esp), %edx
        andb 16(%esp), %cl
        shldl %cl, %eax, %edx
        shll %cl, %eax
        xorl %ecx, %ecx
        xorb %bl, %bl
        testb %bl, %bl
        cmovne %eax, %edx
        cmovne %ecx, %eax
        movl (%esp), %ebx
        addl $4, %esp
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30506 91177308-0d34-0410-b5e6-96231b3b80d8

Back out Chris' last set of changes. This breaks 177.mesa and povray somehow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30505 91177308-0d34-0410-b5e6-96231b3b80d8

80 col.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30504 91177308-0d34-0410-b5e6-96231b3b80d8

Allow PatFrag to be a leaf node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30498 91177308-0d34-0410-b5e6-96231b3b80d8

Add result of a Xform to isel queue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30497 91177308-0d34-0410-b5e6-96231b3b80d8

If we have an add, do it in the pointer realm, not the int realm.  This is critical in the linux kernel for pointer analysis correctness

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30496 91177308-0d34-0410-b5e6-96231b3b80d8

Inspired by the linux kernel, the more we keep adds in the pointer realm, the better pointer analysis works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30495 91177308-0d34-0410-b5e6-96231b3b80d8

Fix UnitTests/2005-05-12-Int64ToFP.c with llc-beta.  In particular, do not
allow it to go into an infinite loop, filling up the disk!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30494 91177308-0d34-0410-b5e6-96231b3b80d8

fix header
add comments
untabify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30486 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a MachineFunctionPass to fix the mul instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30485 91177308-0d34-0410-b5e6-96231b3b80d8

number test right

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30484 91177308-0d34-0410-b5e6-96231b3b80d8

item done

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30483 91177308-0d34-0410-b5e6-96231b3b80d8

implement select.ll:test19-22

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30482 91177308-0d34-0410-b5e6-96231b3b80d8

make this harder

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30481 91177308-0d34-0410-b5e6-96231b3b80d8

new testcases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30480 91177308-0d34-0410-b5e6-96231b3b80d8

Fold the PPCISD shifts when presented with 0 inputs.  This occurs for code
like:
long long test(long long X, int Y) {
  return 1ULL << Y;
}
long long test2(long long X, int Y) {
  return -1LL << Y;
}

which we used to compile to:

_test:
        li r2, 1
        subfic r3, r5, 32
        li r4, 0
        addi r6, r5, -32
        srw r3, r2, r3
        slw r4, r4, r5
        slw r6, r2, r6
        or r3, r4, r3
        slw r4, r2, r5
        or r3, r3, r6
        blr
_test2:
        li r2, -1
        subfic r3, r5, 32
        addi r6, r5, -32
        srw r3, r2, r3
        slw r4, r2, r5
        slw r2, r2, r6
        or r3, r4, r3
        or r3, r3, r2
        blr

Now we produce:

_test:
        li r2, 1
        addi r3, r5, -32
        subfic r4, r5, 32
        slw r3, r2, r3
        srw r4, r2, r4
        or r3, r4, r3
        slw r4, r2, r5
        blr
_test2:
        li r2, -1
        subfic r3, r5, 32
        addi r6, r5, -32
        srw r3, r2, r3
        slw r4, r2, r5
        slw r2, r2, r6
        or r3, r4, r3
        or r3, r3, r2
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30479 91177308-0d34-0410-b5e6-96231b3b80d8

Fold extract_element(cst) to cst

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30478 91177308-0d34-0410-b5e6-96231b3b80d8

Minor speedup for legalize by avoiding some malloc traffic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30477 91177308-0d34-0410-b5e6-96231b3b80d8

If multiple predicates are listed, they must all pass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30476 91177308-0d34-0410-b5e6-96231b3b80d8

Enable dejagnu tests for predicate simplifier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30475 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30474 91177308-0d34-0410-b5e6-96231b3b80d8

There!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30473 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Regression/TableGen/2006-09-18-LargeInt.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30472 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30471 91177308-0d34-0410-b5e6-96231b3b80d8

Allow  i32 UDIV, SDIV, UREM, SREM to be expanded into libcalls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30470 91177308-0d34-0410-b5e6-96231b3b80d8