* Unbreak optimized build (noticed by Eric van Riet Paap)
* Comment #endif clauses for readability

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22646 91177308-0d34-0410-b5e6-96231b3b80d8

Add Subtarget support to PowerPC.  Next up, using it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22644 91177308-0d34-0410-b5e6-96231b3b80d8

refactor some code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22643 91177308-0d34-0410-b5e6-96231b3b80d8

this is not implemented by lsr yet

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22642 91177308-0d34-0410-b5e6-96231b3b80d8

invert to if's to make the logic simpler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22641 91177308-0d34-0410-b5e6-96231b3b80d8

When processing outer loops and we find uses of an IV in inner loops, make
sure to handle the use, just don't recurse into it.

This permits us to generate this code for a simple nested loop case:

.LBB_foo_0:     ; entry
        stwu r1, -48(r1)
        stw r29, 44(r1)
        stw r30, 40(r1)
        mflr r11
        stw r11, 56(r1)
        lis r2, ha16(L_A$non_lazy_ptr)
        lwz r30, lo16(L_A$non_lazy_ptr)(r2)
        li r29, 1
.LBB_foo_1:     ; no_exit.0
        bl L_bar$stub
        li r2, 1
        or r3, r30, r30
.LBB_foo_2:     ; no_exit.1
        lfd f0, 8(r3)
        stfd f0, 0(r3)
        addi r4, r2, 1
        addi r3, r3, 8
        cmpwi cr0, r2, 100
        or r2, r4, r4
        bne .LBB_foo_2  ; no_exit.1
.LBB_foo_3:     ; loopexit.1
        addi r30, r30, 800
        addi r2, r29, 1
        cmpwi cr0, r29, 100
        or r29, r2, r2
        bne .LBB_foo_1  ; no_exit.0
.LBB_foo_4:     ; return
        lwz r11, 56(r1)
        mtlr r11
        lwz r30, 40(r1)
        lwz r29, 44(r1)
        lwz r1, 0(r1)
        blr

instead of this:

_foo:
.LBB_foo_0:     ; entry
        stwu r1, -48(r1)
        stw r28, 44(r1)                   ;; uses an extra register.
        stw r29, 40(r1)
        stw r30, 36(r1)
        mflr r11
        stw r11, 56(r1)
        li r30, 1
        li r29, 0
        or r28, r29, r29
.LBB_foo_1:     ; no_exit.0
        bl L_bar$stub
        mulli r2, r28, 800           ;; unstrength-reduced multiply
        lis r3, ha16(L_A$non_lazy_ptr)   ;; loop invariant address computation
        lwz r3, lo16(L_A$non_lazy_ptr)(r3)
        add r2, r2, r3
        mulli r4, r29, 800           ;; unstrength-reduced multiply
        addi r3, r3, 8
        add r3, r4, r3
        li r4, 1
.LBB_foo_2:     ; no_exit.1
        lfd f0, 0(r3)
        stfd f0, 0(r2)
        addi r5, r4, 1
        addi r2, r2, 8                 ;; multiple stride 8 IV's
        addi r3, r3, 8
        cmpwi cr0, r4, 100
        or r4, r5, r5
        bne .LBB_foo_2  ; no_exit.1
.LBB_foo_3:     ; loopexit.1
        addi r28, r28, 1               ;;; Many IV's with stride 1
        addi r29, r29, 1
        addi r2, r30, 1
        cmpwi cr0, r30, 100
        or r30, r2, r2
        bne .LBB_foo_1  ; no_exit.0
.LBB_foo_4:     ; return
        lwz r11, 56(r1)
        mtlr r11
        lwz r30, 36(r1)
        lwz r29, 40(r1)
        lwz r28, 44(r1)
        lwz r1, 0(r1)
        blr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22640 91177308-0d34-0410-b5e6-96231b3b80d8

Teach loop-reduce to see into nested loops, to pull out immediate values
pushed down by SCEV.

In a nested loop case, this allows us to emit this:

        lis r3, ha16(L_A$non_lazy_ptr)
        lwz r3, lo16(L_A$non_lazy_ptr)(r3)
        add r2, r2, r3
        li r3, 1
.LBB_foo_2:     ; no_exit.1
        lfd f0, 8(r2)        ;; Uses offset of 8 instead of 0
        stfd f0, 0(r2)
        addi r4, r3, 1
        addi r2, r2, 8
        cmpwi cr0, r3, 100
        or r3, r4, r4
        bne .LBB_foo_2  ; no_exit.1

instead of this:

        lis r3, ha16(L_A$non_lazy_ptr)
        lwz r3, lo16(L_A$non_lazy_ptr)(r3)
        add r2, r2, r3
        addi r3, r3, 8
        li r4, 1
.LBB_foo_2:     ; no_exit.1
        lfd f0, 0(r3)
        stfd f0, 0(r2)
        addi r5, r4, 1
        addi r2, r2, 8
        addi r3, r3, 8
        cmpwi cr0, r4, 100
        or r4, r5, r5
        bne .LBB_foo_2  ; no_exit.1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22639 91177308-0d34-0410-b5e6-96231b3b80d8

improve debug output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22638 91177308-0d34-0410-b5e6-96231b3b80d8

Scalar SSE: load +0.0 -> xorps/xorpd
Scalar SSE: a < b ? c : 0.0 -> cmpss, andps
Scalar SSE: float -> i16 needs to be promoted

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22637 91177308-0d34-0410-b5e6-96231b3b80d8

this now passes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22636 91177308-0d34-0410-b5e6-96231b3b80d8

Move from Stage 0 to Stage 1.

Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).

This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:

        or r30, r3, r3
.LBB_test1_1:   ; Loop
        li r2, 0
        stw r2, 0(r30)
        stw r2, 4(r30)
        bl L_pred$stub
        addi r30, r30, 8
        cmplwi cr0, r3, 0
        bne .LBB_test1_1        ; Loop

instead of this code:

        or r30, r3, r3
        or r29, r3, r3
.LBB_test1_1:   ; Loop
        li r2, 0
        stw r2, 0(r29)
        stw r2, 4(r30)
        bl L_pred$stub
        addi r30, r30, 8        ;; Two iv's with step of 8
        addi r29, r29, 8
        cmplwi cr0, r3, 0
        bne .LBB_test1_1        ; Loop

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22635 91177308-0d34-0410-b5e6-96231b3b80d8

Alpha ABI specifies stack is always 16 byte alligned, and gcc does it, so I will too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22634 91177308-0d34-0410-b5e6-96231b3b80d8

Rename IVUse to IVUsersOfOneStride, use a struct instead of a pair to
unify some parallel vectors and get field names more descriptive than
"first" and "second".  This isn't lisp afterall :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22633 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a nasty dangling pointer issue.  The ScalarEvolution pass would keep a
map from instruction* to SCEVHandles.  When we delete instructions, we have
to tell it about it.  We would run into nasty cases where new instructions
were reallocated at old instruction addresses and get the old map values.
Bad bad bad :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22632 91177308-0d34-0410-b5e6-96231b3b80d8

Fix this to test the BE we care about

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22631 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an obvious bug in the Log2 stuff that broke SingleSource/UnitTests/2005-05-12-Int64ToFP
last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22630 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR611, codegen'ing SREM of FP operands to fmod or fmodf instead of
the sequence used for integer ops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22629 91177308-0d34-0410-b5e6-96231b3b80d8

The correct fix for PR612, which also fixes
Transforms/LowerInvoke/2005-08-03-InvokeWithPHIUse.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22628 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR612

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22627 91177308-0d34-0410-b5e6-96231b3b80d8

When inserting code, make sure not to insert it before PHI nodes.  This
fixes PR612 and Transforms/LowerInvoke/2005-08-03-InvokeWithPHI.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22626 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR612

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22625 91177308-0d34-0410-b5e6-96231b3b80d8

Add a couple rlwinm tests for bitfield clears

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22624 91177308-0d34-0410-b5e6-96231b3b80d8

Update rlwimi tests to catch all the cases we care about

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22623 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase  that used to crash simplifycfg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22622 91177308-0d34-0410-b5e6-96231b3b80d8

Fix Transforms/SimplifyCFG/2005-08-03-PHIFactorCrash.ll, a problem that
occurred while bugpointing another testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22621 91177308-0d34-0410-b5e6-96231b3b80d8

add support for Graphviz when viewing CFGs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22620 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar: apostrophe-s ('s) is possessive, not plural; also iff vs. if.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22619 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap comments to 80 cols, fix code sequence for CountLeadingZeros_64 on
non-ppc GCC 4.0 machines.  Patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22618 91177308-0d34-0410-b5e6-96231b3b80d8

minor capitalization thing, patch by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22617 91177308-0d34-0410-b5e6-96231b3b80d8

Fix another bug in the clz patch that caused miscompilations when !gcc4

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22616 91177308-0d34-0410-b5e6-96231b3b80d8

Finally, add the required constraint checks to fix  Transforms/SimplifyCFG/2005-08-01-PHIUpdateFail.ll
the right way

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22615 91177308-0d34-0410-b5e6-96231b3b80d8

un-xfail this

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22614 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify some code, add the correct pred checks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22613 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor code out of PropagatePredecessorsForPHIs, turning it into a pure function with no side-effects

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22612 91177308-0d34-0410-b5e6-96231b3b80d8

use splice instead of remove/insert to avoid some symtab operations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22611 91177308-0d34-0410-b5e6-96231b3b80d8

move two functions up in the file, use SafeToMergeTerminators to eliminate
some duplicated code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22610 91177308-0d34-0410-b5e6-96231b3b80d8

Rip some code out of the main SimplifyCFG function into a subfunction and
call it from the only place it is live.  No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22609 91177308-0d34-0410-b5e6-96231b3b80d8

This fails for now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22608 91177308-0d34-0410-b5e6-96231b3b80d8

Disable this patch:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050801/027345.html

This breaks real programs and only fixes an obscure regression testcase.  A
real fix is in development.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22606 91177308-0d34-0410-b5e6-96231b3b80d8

Change a place to use an arbitrary value instead of null, when possible

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22605 91177308-0d34-0410-b5e6-96231b3b80d8

If the user interrupts bugpoint, don't extract loops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22603 91177308-0d34-0410-b5e6-96231b3b80d8

Add a note, people are responsible for requesting that they be added to the
file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22602 91177308-0d34-0410-b5e6-96231b3b80d8

Pass -export-dynamic to gcc when compiling with -native and the link is
performed with -export-dynamic (aka. -disable-internalize).

Patch by Nicholas Riley!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22601 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the non-gcc 4.0 paths for countleadingzeros

Patch fixed by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22598 91177308-0d34-0410-b5e6-96231b3b80d8

Fix inverted conditionals

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22597 91177308-0d34-0410-b5e6-96231b3b80d8

one more hunk that got dropped

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22596 91177308-0d34-0410-b5e6-96231b3b80d8

This hunk accidentally got dropped.  Patch by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22595 91177308-0d34-0410-b5e6-96231b3b80d8

Update to use the new MathExtras.h support for log2 computation.
Patch contributed by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22594 91177308-0d34-0410-b5e6-96231b3b80d8

Fix the non-gcc 4.0 path to compile

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22593 91177308-0d34-0410-b5e6-96231b3b80d8

Update to use the new MathExtras.h support for log2 computation.
Patch contributed by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22592 91177308-0d34-0410-b5e6-96231b3b80d8

Consolidate all of the various log2 computing functions into MathExtras.h.
Also, provide accelerated implementations when building with GCC.
Patch contributed by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22591 91177308-0d34-0410-b5e6-96231b3b80d8

Update a doc, patch contributed by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22590 91177308-0d34-0410-b5e6-96231b3b80d8

add a pass name to make debugging dumps nicer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22588 91177308-0d34-0410-b5e6-96231b3b80d8

Fix grammar: it's == "it is".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22587 91177308-0d34-0410-b5e6-96231b3b80d8

Like the comment says, do not insert cast instructions before phi nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22586 91177308-0d34-0410-b5e6-96231b3b80d8

It's dangerous coding on Mondays.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22585 91177308-0d34-0410-b5e6-96231b3b80d8

This code was very close, but not quite right.  It did not take into
consideration the case where a reference in an unreachable block could
occur.  This fixes Transforms/SimplifyCFG/2005-08-01-PHIUpdateFail.ll,
something I ran into while bugpoint'ing another pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22584 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase that caused simplifycfg to crash

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22583 91177308-0d34-0410-b5e6-96231b3b80d8

Implement SetInterruptFunction for Windows.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22582 91177308-0d34-0410-b5e6-96231b3b80d8

add a comment, make a check more lenient

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22581 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify for loop, clear a per-loop map after processing each loop

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22580 91177308-0d34-0410-b5e6-96231b3b80d8

When the user hits ctrl-c, bugpoint should attempt to stop reduction as
quickly as possible and output what it has so far.  If they hit it twice,
bugpoint is killed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22579 91177308-0d34-0410-b5e6-96231b3b80d8

Implement sys::SetInterruptFunction on Unix, stub it on win32 so that the
build will not fail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22578 91177308-0d34-0410-b5e6-96231b3b80d8

add a new function proto

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22577 91177308-0d34-0410-b5e6-96231b3b80d8

Add a comment
Make LSR ignore GEP's that have loop variant base values, as we currently
cannot codegen them

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22576 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an iterator invalidation problem

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22575 91177308-0d34-0410-b5e6-96231b3b80d8

200.sixtrack prints FP numbers with a very strange notation that uses D
instead of E for exponentials (e.g. 1.234D-43).  Add support for this
notation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22574 91177308-0d34-0410-b5e6-96231b3b80d8

make is the standard name, not gmake

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22573 91177308-0d34-0410-b5e6-96231b3b80d8

Mark these as V9 specific

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22572 91177308-0d34-0410-b5e6-96231b3b80d8

update function codes to reflect /su flags that have been added since this was written

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22571 91177308-0d34-0410-b5e6-96231b3b80d8

Fix casts from long to sbyte on ppc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22570 91177308-0d34-0410-b5e6-96231b3b80d8

use llabs not abs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22569 91177308-0d34-0410-b5e6-96231b3b80d8

one cannot allocate a global, until one is done initializing the global pointers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22568 91177308-0d34-0410-b5e6-96231b3b80d8

temporarily fail these two tests until LSR evolves to support them

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22567 91177308-0d34-0410-b5e6-96231b3b80d8

ConstantInt::get only works for arguments < 128.

SimplifyLibCalls probably has to be audited to make sure it does not make
this mistake elsewhere.  Also, if this code knows that the type will be
unsigned, obviously one arm of this is dead.

Reid, can you take a look into this further?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22566 91177308-0d34-0410-b5e6-96231b3b80d8

Keep tabs and trailing spaces out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22565 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VC++ build problems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22564 91177308-0d34-0410-b5e6-96231b3b80d8

fix float->long conversions on x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22563 91177308-0d34-0410-b5e6-96231b3b80d8

this tests every fp-to-integer conversion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22562 91177308-0d34-0410-b5e6-96231b3b80d8

fix a typeo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22561 91177308-0d34-0410-b5e6-96231b3b80d8

Ack, typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22560 91177308-0d34-0410-b5e6-96231b3b80d8

Change the fp to integer code to not perform 2-byte stores followed by
1 byte loads and other operations.  This is bad for store-forwarding on
common CPUs.  We now do this:

fnstcw WORD PTR [%ESP]
mov %AX, WORD PTR [%ESP]

instead of:

fnstcw WORD PTR [%ESP]
mov %AL, BYTE PTR [%ESP + 1]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22559 91177308-0d34-0410-b5e6-96231b3b80d8

Commit a new LoopStrengthReduce pass that can use scalar evolutions and
target data to decide which loop induction variables to strength reduce
and how to do so.  This work is mostly by Chris Lattner, with tweaks by
me to get it working on some of MultiSource.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22558 91177308-0d34-0410-b5e6-96231b3b80d8

Break SCEVExpander out of IndVarSimplify into its own .h/.cpp file so that
other passes may use it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22557 91177308-0d34-0410-b5e6-96231b3b80d8

Use a custom expander for all FP to int conversions, as the X86 only has
FP-to-int-in-memory: this exposes the load from the stored slot to the
selection dag, allowing it to be folded into other operaions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22556 91177308-0d34-0410-b5e6-96231b3b80d8

Allow targets to have custom expanders for FP_TO_*INT conversions where
both the src and dest values are legal

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22555 91177308-0d34-0410-b5e6-96231b3b80d8

support near allocations for the JIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22554 91177308-0d34-0410-b5e6-96231b3b80d8

turn off GOT on archs that didn't use it (not that it appeard to harm them much with it on)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22553 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a FIXME: move a bunch of cruft for handling FP_TO_*INT operations
that the X86 does not support to the legalizer.  This allows it to be better
optimized, etc, and will help with SSE support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22551 91177308-0d34-0410-b5e6-96231b3b80d8

Don't forget to diddle with the control word when performing an FISTP64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22550 91177308-0d34-0410-b5e6-96231b3b80d8

Use a custom expander to compile this:

long %test4(double %X) {
        %tmp.1 = cast double %X to long         ; <long> [#uses=1]
        ret long %tmp.1
}

to this:

_test4:
        sub %ESP, 12
        fld QWORD PTR [%ESP + 16]
        fistp QWORD PTR [%ESP]
        mov %EDX, DWORD PTR [%ESP + 4]
        mov %EAX, DWORD PTR [%ESP]
        add %ESP, 12
        ret

instead of this:

_test4:
        sub %ESP, 28
        fld QWORD PTR [%ESP + 32]
        fstp QWORD PTR [%ESP]
        call ___fixdfdi
        add %ESP, 28
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22549 91177308-0d34-0410-b5e6-96231b3b80d8

Allow targets to define custom expanders for FP_TO_*INT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22548 91177308-0d34-0410-b5e6-96231b3b80d8

allow a target to request that unknown FP_TO_*INT conversion be promoted to
a larger integer destination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22547 91177308-0d34-0410-b5e6-96231b3b80d8

instead of having all conversions be handled by one case value, and then have
subcases inside, break things out earlier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22546 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase for PR610

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22545 91177308-0d34-0410-b5e6-96231b3b80d8

support bsr, and more .td simplification

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22543 91177308-0d34-0410-b5e6-96231b3b80d8

new is not a valid default anywhere, so make this pure virtual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22542 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a problem in getDirectoryContents where sub-directory names were
appended to a path string that didn't end in a slash, yielding invalid
path names.

Path contribute by Nicholas Riley.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22539 91177308-0d34-0410-b5e6-96231b3b80d8

get lazy JITing working.  Some of shootout runs now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22538 91177308-0d34-0410-b5e6-96231b3b80d8

Like constants, globals on some platforms are GOT relative.  This means they have to be allocated
near the GOT, which new doesn't do.  So break out the allocate into a new function.

Also move GOT index handling into JITResolver.  This lets it update the mapping when a Lazy
function is JITed.  It doesn't managed the table, just the mapping.  Note that this is
still non-ideal, as any function that takes a function address should also take a GOT
index, but that is a lot of changes.  The relocation resolve process updates any GOT entry
it sees is out of date.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22537 91177308-0d34-0410-b5e6-96231b3b80d8