oota-llvm.git
19 years agoOoops, don't forget to clear this. The real inner loop is now:
Chris Lattner [Sat, 13 Aug 2005 07:42:01 +0000 (07:42 +0000)]
Ooops, don't forget to clear this.  The real inner loop is now:

.LBB_foo_3:     ; no_exit.1
        lfd f2, 0(r9)
        lfd f3, 8(r9)
        fmul f4, f1, f2
        fmadd f4, f0, f3, f4
        stfd f4, 8(r9)
        fmul f3, f1, f3
        fmsub f2, f0, f2, f3
        stfd f2, 0(r9)
        addi r9, r9, 16
        addi r8, r8, 1
        cmpw cr0, r8, r4
        ble .LBB_foo_3  ; no_exit.1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22782 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoRecursively scan scev expressions for common subexpressions. This allows us
Chris Lattner [Sat, 13 Aug 2005 07:27:18 +0000 (07:27 +0000)]
Recursively scan scev expressions for common subexpressions.  This allows us
to handle nested loops much better, for example, by being able to tell that
these two expressions:

{( 8 + ( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp 12)}<loopentry.1>

{(( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp12)}<loopentry.1>

Have the following common part that can be shared:
{(( 16 * ( 1 +  %Tmp11 +  %Tmp12)) +  %c_),+,( 16 *  %Tmp12)}<loopentry.1>

This allows us to codegen an important inner loop in 168.wupwise as:

.LBB_foo_4:     ; no_exit.1
        lfd f2, 16(r9)
        fmul f3, f0, f2
        fmul f2, f1, f2
        fadd f4, f3, f2
        stfd f4, 8(r9)
        fsub f2, f3, f2
        stfd f2, 16(r9)
        addi r8, r8, 1
        addi r9, r9, 16
        cmpw cr0, r8, r4
        ble .LBB_foo_4  ; no_exit.1

instead of:

.LBB_foo_3:     ; no_exit.1
        lfdx f2, r6, r9
        add r10, r6, r9
        lfd f3, 8(r10)
        fmul f4, f1, f2
        fmadd f4, f0, f3, f4
        stfd f4, 8(r10)
        fmul f3, f1, f3
        fmsub f2, f0, f2, f3
        stfdx f2, r6, r9
        addi r9, r9, 16
        addi r8, r8, 1
        cmpw cr0, r8, r4
        ble .LBB_foo_3  ; no_exit.1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22781 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoRemove an unncessary argument to SimplifySelectCC and add an additional
Nate Begeman [Sat, 13 Aug 2005 06:14:17 +0000 (06:14 +0000)]
Remove an unncessary argument to SimplifySelectCC and add an additional
assert when creating a select_cc node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22780 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix the fabs regression on x86 by abstracting the select_cc optimization
Nate Begeman [Sat, 13 Aug 2005 06:00:21 +0000 (06:00 +0000)]
Fix the fabs regression on x86 by abstracting the select_cc optimization
out into SimplifySelectCC.  This allows both ISD::SELECT and ISD::SELECT_CC
to use the same set of simplifying folds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22779 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoRemove support for 64b PPC, it's been broken for a long time. It'll be
Nate Begeman [Sat, 13 Aug 2005 05:59:16 +0000 (05:59 +0000)]
Remove support for 64b PPC, it's been broken for a long time.  It'll be
back once a DAG->DAG ISel exists.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22778 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix oversized GOT problem with gcc-4 on alpha
Andrew Lenharth [Sat, 13 Aug 2005 05:09:50 +0000 (05:09 +0000)]
Fix oversized GOT problem with gcc-4 on alpha

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22777 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoTeach SplitCriticalEdge to update LoopInfo if it is alive. This fixes
Chris Lattner [Sat, 13 Aug 2005 01:38:43 +0000 (01:38 +0000)]
Teach SplitCriticalEdge to update LoopInfo if it is alive.  This fixes
a problem in LoopStrengthReduction, where it would split critical edges
then confused itself with outdated loop information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22776 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoremove dead code. The exit block list is computed on demand, thus does not
Chris Lattner [Sat, 13 Aug 2005 01:30:36 +0000 (01:30 +0000)]
remove dead code.  The exit block list is computed on demand, thus does not
need to be updated.  This code is a relic from when it did.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22775 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoimplement a couple of simple shift foldings.
Chris Lattner [Fri, 12 Aug 2005 23:54:58 +0000 (23:54 +0000)]
implement a couple of simple shift foldings.

e.g.  (X & 7) >> 3   -> 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22774 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix for 2005-08-12-rlwimi-crash.ll. Make allowance for masks being shifted to
Jim Laskey [Fri, 12 Aug 2005 23:52:46 +0000 (23:52 +0000)]
Fix for 2005-08-12-rlwimi-crash.ll.  Make allowance for masks being shifted to
zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22773 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdded test cases to guarantee use of ORC and ANDC.
Jim Laskey [Fri, 12 Aug 2005 23:40:14 +0000 (23:40 +0000)]
Added test cases to guarantee use of ORC and ANDC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22772 91177308-0d34-0410-b5e6-96231b3b80d8

19 years ago1. This changes handles the cases of (~x)&y and x&(~y) yielding ANDC, and
Jim Laskey [Fri, 12 Aug 2005 23:38:02 +0000 (23:38 +0000)]
1. This changes handles the cases of (~x)&y and x&(~y) yielding ANDC, and
   (~x)|y and x|(~y) yielding ORC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22771 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agotestcase that crashed the ppc backend, distilled from crafty
Chris Lattner [Fri, 12 Aug 2005 23:34:03 +0000 (23:34 +0000)]
testcase that crashed the ppc backend, distilled from crafty

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22770 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoWhen splitting critical edges, make sure not to leave the new block in the
Chris Lattner [Fri, 12 Aug 2005 22:22:17 +0000 (22:22 +0000)]
When splitting critical edges, make sure not to leave the new block in the
middle of the loop.  This turns a critical loop in gzip into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_8 ; loopentry.loopexit_crit_edge
.LBB_test_2:    ; shortcirc_next.0
        add r28, r3, r27
        lhz r28, 5(r28)
        add r26, r4, r27
        lhz r26, 5(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_7 ; shortcirc_next.0.loopexit_crit_edge
.LBB_test_3:    ; shortcirc_next.1
        add r28, r3, r27
        lhz r28, 7(r28)
        add r26, r4, r27
        lhz r26, 7(r26)
        cmpw cr0, r28, r26
        bne .LBB_test_6 ; shortcirc_next.1.loopexit_crit_edge
.LBB_test_4:    ; shortcirc_next.2
        add r28, r3, r27
        lhz r26, 9(r28)
        add r28, r4, r27
        lhz r25, 9(r28)
        addi r28, r27, 8
        cmpw cr7, r26, r25
        mfcr r26, 1
        rlwinm r26, r26, 31, 31, 31
        add r25, r8, r27
        cmpw cr7, r25, r7
        mfcr r25, 1
        rlwinm r25, r25, 29, 31, 31
        and. r26, r26, r25
        bne .LBB_test_1 ; loopentry

instead of this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_test_9   ; loopexit
.LBB_test_3:    ; shortcirc_next.0
        add r28, r3, r27
        lhz r28, 5(r28)
        add r26, r4, r27
        lhz r26, 5(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_5 ; shortcirc_next.1
.LBB_test_4:    ; shortcirc_next.0.loopexit_crit_edge
        add r2, r11, r27
        add r8, r12, r27
        b .LBB_test_9   ; loopexit
.LBB_test_5:    ; shortcirc_next.1
        add r28, r3, r27
        lhz r28, 7(r28)
        add r26, r4, r27
        lhz r26, 7(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_7 ; shortcirc_next.2
.LBB_test_6:    ; shortcirc_next.1.loopexit_crit_edge
        add r2, r9, r27
        add r8, r10, r27
        b .LBB_test_9   ; loopexit
.LBB_test_7:    ; shortcirc_next.2
        add r28, r3, r27
        lhz r26, 9(r28)
        add r28, r4, r27
        lhz r25, 9(r28)
        addi r28, r27, 8
        cmpw cr7, r26, r25
        mfcr r26, 1
        rlwinm r26, r26, 31, 31, 31
        add r25, r8, r27
        cmpw cr7, r25, r7
        mfcr r25, 1
        rlwinm r25, r25, 29, 31, 31
        and. r26, r26, r25
        bne .LBB_test_1 ; loopentry

Next up, improve the code for the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22769 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd a helper method
Chris Lattner [Fri, 12 Aug 2005 22:14:06 +0000 (22:14 +0000)]
Add a helper method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22768 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd a helper method
Chris Lattner [Fri, 12 Aug 2005 22:13:27 +0000 (22:13 +0000)]
add a helper method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22767 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix a FIXME: if we are inserting code for a PHI argument, split the critical
Chris Lattner [Fri, 12 Aug 2005 22:06:11 +0000 (22:06 +0000)]
Fix a FIXME: if we are inserting code for a PHI argument, split the critical
edge so that the code is not always executed for both operands.  This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops).  On gzip, for
example, we turn this ugly code:

.LBB_test_1:    ; loopentry
        add r27, r3, r28
        lhz r27, 3(r27)
        add r26, r4, r28
        lhz r26, 3(r26)
        add r25, r30, r28    ;; Only live if exiting the loop
        add r24, r29, r28    ;; Only live if exiting the loop
        cmpw cr0, r27, r26
        bne .LBB_test_5 ; loopexit

into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_test_9   ; loopexit
.LBB_test_2:    ; shortcirc_next.0
        ...
        blt .LBB_test_1

into this:

.LBB_test_1:    ; loopentry
        or r27, r28, r28
        add r28, r3, r27
        lhz r28, 3(r28)
        add r26, r4, r27
        lhz r26, 3(r26)
        cmpw cr0, r28, r26
        beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2:    ; loopentry.loopexit_crit_edge
        add r2, r30, r27
        add r8, r29, r27
        b .LBB_t_3:    ; shortcirc_next.0
.LBB_test_3:    ; shortcirc_next.0
        ...
        blt .LBB_test_1

Next step: get the block out of the loop so that the loop is all
fall-throughs again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22766 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoChange break critical edges to not remove, then insert, PHI node entries.
Chris Lattner [Fri, 12 Aug 2005 21:58:07 +0000 (21:58 +0000)]
Change break critical edges to not remove, then insert, PHI node entries.
Instead, just update the BB in-place.  This is both faster, and it prevents
split-critical-edges from shuffling the PHI argument list unneccesarily.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22765 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agomatch gcc's use of tabs, makes diffs easier
Andrew Lenharth [Fri, 12 Aug 2005 16:14:08 +0000 (16:14 +0000)]
match gcc's use of tabs, makes diffs easier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22764 91177308-0d34-0410-b5e6-96231b3b80d8

19 years ago.section cleanup, patch from Nicholas Riley
Andrew Lenharth [Fri, 12 Aug 2005 16:13:43 +0000 (16:13 +0000)]
.section cleanup, patch from Nicholas Riley

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22763 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFirst rev of Xcode 2.1 project
Chris Lattner [Thu, 11 Aug 2005 22:19:26 +0000 (22:19 +0000)]
First rev of Xcode 2.1 project

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22762 91177308-0d34-0410-b5e6-96231b3b80d8

19 years ago1. Added the function isOpcWithIntImmediate to simplify testing of operand with
Jim Laskey [Thu, 11 Aug 2005 21:59:23 +0000 (21:59 +0000)]
1. Added the function isOpcWithIntImmediate to simplify testing of operand with
specified opcode and an integer constant right operand.

2. Modified ISD::SHL, ISD::SRL, ISD::SRA to use rlwinm when applied after a mask.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22761 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoTidied up the use of dyn_cast<ConstantSDNode> by using isIntImmediate more.
Chris Lattner [Thu, 11 Aug 2005 17:56:50 +0000 (17:56 +0000)]
Tidied up the use of dyn_cast<ConstantSDNode> by using isIntImmediate more.
Patch by Jim Laskey.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22760 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoUse a more efficient method of creating integer and float virtual registers
Chris Lattner [Thu, 11 Aug 2005 17:15:31 +0000 (17:15 +0000)]
Use a more efficient method of creating integer and float virtual registers
(avoids an extra level of indirection in MakeReg).

  defined MakeIntReg using RegMap->createVirtualRegister(PPC32::GPRCRegisterClass)
  defined MakeFPReg using RegMap->createVirtualRegister(PPC32::FPRCRegisterClass)

  s/MakeReg(MVT::i32)/MakeIntReg/
  s/MakeReg(MVT::f64)/MakeFPReg/

Patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22759 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd a select_cc optimization for recognizing abs(int). This speeds up an
Nate Begeman [Thu, 11 Aug 2005 02:18:13 +0000 (02:18 +0000)]
Add a select_cc optimization for recognizing abs(int).  This speeds up an
integer MPEG encoding loop by a factor of two.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22758 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoSome SELECT_CC cleanups:
Nate Begeman [Thu, 11 Aug 2005 01:12:20 +0000 (01:12 +0000)]
Some SELECT_CC cleanups:
1. move assertions for node creation to getNode()
2. legalize the values returned in ExpandOp immediately
3. Move select_cc optimizations from SELECT's getNode() to SELECT_CC's,
   allowing them to be cleaned up significantly.

This paves the way to pick up additional optimizations on SELECT_CC, such
as sum-of-absolute-differences.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22757 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoMake SELECT illegal on PPC32, switch to using SELECT_CC, which more closely
Nate Begeman [Wed, 10 Aug 2005 20:52:09 +0000 (20:52 +0000)]
Make SELECT illegal on PPC32, switch to using SELECT_CC, which more closely
reflects what the hardware is capable of.  This significantly simplifies
the CC handling logic throughout the ISel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22756 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd new node, SELECT_CC. This node is for targets that don't natively
Nate Begeman [Wed, 10 Aug 2005 20:51:12 +0000 (20:51 +0000)]
Add new node, SELECT_CC.  This node is for targets that don't natively
implement SELECT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22755 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoChanges for PPC32ISelPattern.cpp
Chris Lattner [Wed, 10 Aug 2005 18:11:33 +0000 (18:11 +0000)]
Changes for PPC32ISelPattern.cpp
1. Clean up how SelectIntImmediateExpr handles use counts.
2. "Subtract from" was not clearing hi 16 bits.

Patch by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22754 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix an oversight that may be causing PR617.
Chris Lattner [Wed, 10 Aug 2005 17:37:53 +0000 (17:37 +0000)]
Fix an oversight that may be causing PR617.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22753 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonow that we handle non-constant strides, this testcase passes
Chris Lattner [Wed, 10 Aug 2005 17:17:45 +0000 (17:17 +0000)]
now that we handle non-constant strides, this testcase passes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22752 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoremove some trickiness that broke yacr2 and some other programs last night
Chris Lattner [Wed, 10 Aug 2005 17:15:20 +0000 (17:15 +0000)]
remove some trickiness that broke yacr2 and some other programs last night

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22751 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoChanged the XOR case to use the isOprNot predicate.
Chris Lattner [Wed, 10 Aug 2005 16:35:46 +0000 (16:35 +0000)]
Changed the XOR case to use the isOprNot predicate.

Patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22750 91177308-0d34-0410-b5e6-96231b3b80d8

19 years ago1. Refactored handling of integer immediate values for add, or, xor and sub.
Chris Lattner [Wed, 10 Aug 2005 16:34:52 +0000 (16:34 +0000)]
1. Refactored handling of integer immediate values for add, or, xor and sub.
  New routine: ISel::SelectIntImmediateExpr
  2. Now checking use counts of large constants.  If use count is > 2 then drop
  thru so that the constant gets loaded into a register.
  Source:

int %test1(int %a) {
entry:
       %tmp.1 = add int %a,      123456789      ; <int> [#uses=1]
       %tmp.2 = or  int %tmp.1,  123456789      ; <int> [#uses=1]
       %tmp.3 = xor int %tmp.2,  123456789      ; <int> [#uses=1]
       %tmp.4 = sub int %tmp.3, -123456789      ; <int> [#uses=1]
       ret int %tmp.4
}

Did Emit:

       .machine ppc970

       .text
       .align  2
       .globl  _test1
_test1:
.LBB_test1_0:   ; entry
       addi r2, r3, -13035
       addis r2, r2, 1884
       ori r2, r2, 52501
       oris r2, r2, 1883
       xori r2, r2, 52501
       xoris r2, r2, 1883
       addi r2, r2, 52501
       addis r3, r2, 1883
       blr

Now Emits:

       .machine ppc970

       .text
       .align  2
       .globl  _test1
_test1:
.LBB_test1_0:   ; entry
       lis r2, 1883
       ori r2, r2, 52501
       add r3, r3, r2
       or r3, r3, r2
       xor r3, r3, r2
       add r3, r3, r2
       blr

Patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22749 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agosorry!! this is temporary; for some reason the nasty constmul code seems to
Duraid Madina [Wed, 10 Aug 2005 12:38:57 +0000 (12:38 +0000)]
sorry!! this is temporary; for some reason the nasty constmul code seems to
be an infinite loop when using g++-4.0.1*, this kills the ia64 nightly
tester. A proper fix shall be forthcoming!!! thanks for not killing me. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22748 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix a bug compiling: select (i32 < i32), f32, f32
Chris Lattner [Wed, 10 Aug 2005 03:40:09 +0000 (03:40 +0000)]
Fix a bug compiling: select (i32 < i32), f32, f32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22747 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoMake loop-simplify produce better loops by turning PHI nodes like X = phi [X, Y]
Chris Lattner [Wed, 10 Aug 2005 02:07:32 +0000 (02:07 +0000)]
Make loop-simplify produce better loops by turning PHI nodes like X = phi [X, Y]
into just Y.  This often occurs when it seperates loops that have collapsed loop
headers.  This implements LoopSimplify/phi-node-simplify.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22746 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoNew testcase
Chris Lattner [Wed, 10 Aug 2005 02:06:35 +0000 (02:06 +0000)]
New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22745 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAllow indvar simplify to canonicalize ANY affine IV, not just affine IVs with
Chris Lattner [Wed, 10 Aug 2005 01:12:06 +0000 (01:12 +0000)]
Allow indvar simplify to canonicalize ANY affine IV, not just affine IVs with
constant stride.  This implements Transforms/IndVarsSimplify/variable-stride-ivs.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22744 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonew testcase
Chris Lattner [Wed, 10 Aug 2005 01:11:24 +0000 (01:11 +0000)]
new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22743 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix an obvious oops
Chris Lattner [Wed, 10 Aug 2005 00:59:40 +0000 (00:59 +0000)]
Fix an obvious oops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22742 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonew testcase we handle
Chris Lattner [Wed, 10 Aug 2005 00:48:11 +0000 (00:48 +0000)]
new testcase we handle

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22741 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoTeach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.
Chris Lattner [Wed, 10 Aug 2005 00:45:21 +0000 (00:45 +0000)]
Teach LSR to strength reduce IVs that have a loop-invariant but non-constant stride.

For code like this:

void foo(float *a, float *b, int n, int stride_a, int stride_b) {
  int i;
  for (i=0; i<n; i++)
      a[i*stride_a] = b[i*stride_b];
}

we now emit:

.LBB_foo2_2:    ; no_exit
        lfs f0, 0(r4)
        stfs f0, 0(r3)
        addi r7, r7, 1
        add r4, r2, r4
        add r3, r6, r3
        cmpw cr0, r7, r5
        blt .LBB_foo2_2 ; no_exit

instead of:

.LBB_foo_2:     ; no_exit
        mullw r8, r2, r7     ;; multiply!
        slwi r8, r8, 2
        lfsx f0, r4, r8
        mullw r8, r2, r6     ;; multiply!
        slwi r8, r8, 2
        stfsx f0, r3, r8
        addi r2, r2, 1
        cmpw cr0, r2, r5
        blt .LBB_foo_2  ; no_exit

loops with variable strides occur pretty often.  For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.

Now we can allow indvars to turn functions written like this:

void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
  int i, ai = 0, bi = 0;
  for (i=0; i<n; i++)
    {
      a[ai] = b[bi];
      ai += stride_a;
      bi += stride_b;
    }
}

into code like the above for better analysis.  With this patch, they generate
identical code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22740 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix Regression/Transforms/LoopStrengthReduce/phi_node_update_multiple_preds.ll
Chris Lattner [Wed, 10 Aug 2005 00:35:32 +0000 (00:35 +0000)]
Fix Regression/Transforms/LoopStrengthReduce/phi_node_update_multiple_preds.ll
by being more careful about updating PHI nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22739 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonew testcase
Chris Lattner [Wed, 10 Aug 2005 00:33:01 +0000 (00:33 +0000)]
new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22738 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix some 80 column violations.
Chris Lattner [Tue, 9 Aug 2005 23:39:36 +0000 (23:39 +0000)]
Fix some 80 column violations.

Once we compute the evolution for a GEP, tell SE about it.  This allows users
of the GEP to know it, if the users are not direct.  This allows us to compile
this testcase:

void fbSolidFillmmx(int w, unsigned char *d) {
    while (w >= 64) {
        *(unsigned long long *) (d +  0) = 0;
        *(unsigned long long *) (d +  8) = 0;
        *(unsigned long long *) (d + 16) = 0;
        *(unsigned long long *) (d + 24) = 0;
        *(unsigned long long *) (d + 32) = 0;
        *(unsigned long long *) (d + 40) = 0;
        *(unsigned long long *) (d + 48) = 0;
        *(unsigned long long *) (d + 56) = 0;
        w -= 64;
        d += 64;
    }
}

into:

.LBB_fbSolidFillmmx_2:  ; no_exit
        li r2, 0
        stw r2, 0(r4)
        stw r2, 4(r4)
        stw r2, 8(r4)
        stw r2, 12(r4)
        stw r2, 16(r4)
        stw r2, 20(r4)
        stw r2, 24(r4)
        stw r2, 28(r4)
        stw r2, 32(r4)
        stw r2, 36(r4)
        stw r2, 40(r4)
        stw r2, 44(r4)
        stw r2, 48(r4)
        stw r2, 52(r4)
        stw r2, 56(r4)
        stw r2, 60(r4)
        addi r4, r4, 64
        addi r3, r3, -64
        cmpwi cr0, r3, 63
        bgt .LBB_fbSolidFillmmx_2       ; no_exit

instead of:

.LBB_fbSolidFillmmx_2:  ; no_exit
        li r11, 0
        stw r11, 0(r4)
        stw r11, 4(r4)
        stwx r11, r10, r4
        add r12, r10, r4
        stw r11, 4(r12)
        stwx r11, r9, r4
        add r12, r9, r4
        stw r11, 4(r12)
        stwx r11, r8, r4
        add r12, r8, r4
        stw r11, 4(r12)
        stwx r11, r7, r4
        add r12, r7, r4
        stw r11, 4(r12)
        stwx r11, r6, r4
        add r12, r6, r4
        stw r11, 4(r12)
        stwx r11, r5, r4
        add r12, r5, r4
        stw r11, 4(r12)
        stwx r11, r2, r4
        add r12, r2, r4
        stw r11, 4(r12)
        addi r4, r4, 64
        addi r3, r3, -64
        cmpwi cr0, r3, 63
        bgt .LBB_fbSolidFillmmx_2       ; no_exit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22737 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoimplement two helper methods
Chris Lattner [Tue, 9 Aug 2005 23:36:33 +0000 (23:36 +0000)]
implement two helper methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22736 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd two helper methods
Chris Lattner [Tue, 9 Aug 2005 23:36:18 +0000 (23:36 +0000)]
add two helper methods

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22735 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix spelling, fix some broken canonicalizations by my last patch
Chris Lattner [Tue, 9 Aug 2005 23:09:05 +0000 (23:09 +0000)]
Fix spelling, fix some broken canonicalizations by my last patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22734 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoI can't believe I caught this before Misha! :)
Chris Lattner [Tue, 9 Aug 2005 23:08:53 +0000 (23:08 +0000)]
I can't believe I caught this before Misha! :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22733 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd a optimization note
Chris Lattner [Tue, 9 Aug 2005 22:30:57 +0000 (22:30 +0000)]
add a optimization note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22732 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd cc nodes to the AllNodes list so they show up in Graphviz output
Chris Lattner [Tue, 9 Aug 2005 20:40:02 +0000 (20:40 +0000)]
add cc nodes to the AllNodes list so they show up in Graphviz output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22731 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd testcases for new rlwinm cases handled, patch by Jim Laskey!
Chris Lattner [Tue, 9 Aug 2005 20:24:16 +0000 (20:24 +0000)]
Add testcases for new rlwinm cases handled, patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22730 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoUpdate the targets to the new SETCC/CondCodeSDNode interfaces.
Chris Lattner [Tue, 9 Aug 2005 20:21:10 +0000 (20:21 +0000)]
Update the targets to the new SETCC/CondCodeSDNode interfaces.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22729 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoEliminate the SetCCSDNode in favor of a CondCodeSDNode class. This pulls the
Chris Lattner [Tue, 9 Aug 2005 20:20:18 +0000 (20:20 +0000)]
Eliminate the SetCCSDNode in favor of a CondCodeSDNode class.  This pulls the
CC out of the SetCC operation, making SETCC a standard ternary operation and
CC's a standard DAG leaf.  This will make it possible for other node to use
CC's as operands in the future...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22728 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoMinor cleanup patch, no functionality changes. Written by Jim Laskey.
Chris Lattner [Tue, 9 Aug 2005 18:29:55 +0000 (18:29 +0000)]
Minor cleanup patch, no functionality changes.  Written by Jim Laskey.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22727 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix CodeGen/Generic/div-neg-power-2.ll, a regression from last night.
Chris Lattner [Tue, 9 Aug 2005 18:08:41 +0000 (18:08 +0000)]
Fix CodeGen/Generic/div-neg-power-2.ll, a regression from last night.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22726 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonew reg test for a failure last night on ppc/darwin
Chris Lattner [Tue, 9 Aug 2005 18:07:45 +0000 (18:07 +0000)]
new reg test for a failure last night on ppc/darwin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22725 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoSCEVAddExpr::get() of an empty list is invalid.
Chris Lattner [Tue, 9 Aug 2005 01:13:47 +0000 (01:13 +0000)]
SCEVAddExpr::get() of an empty list is invalid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22724 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoThis is now implemented
Chris Lattner [Tue, 9 Aug 2005 00:19:44 +0000 (00:19 +0000)]
This is now implemented

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22723 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoImplement: LoopStrengthReduce/share_ivs.ll
Chris Lattner [Tue, 9 Aug 2005 00:18:09 +0000 (00:18 +0000)]
Implement: LoopStrengthReduce/share_ivs.ll

Two changes:
  * Only insert one PHI node for each stride.  Other values are live in
    values.  This cannot introduce higher register pressure than the
    previous approach, and can take advantage of reg+reg addressing modes.
  * Factor common base values out of uses before moving values from the
    base to the immediate fields.  This improves codegen by starting the
    stride-specific PHI node out at a common place for each IV use.

As an example, we used to generate this for a loop in swim:

.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2:        ; no_exit.7.i
        lfd f0, 0(r8)
        stfd f0, 0(r3)
        lfd f0, 0(r6)
        stfd f0, 0(r7)
        lfd f0, 0(r2)
        stfd f0, 0(r5)
        addi r9, r9, 1
        addi r2, r2, 8
        addi r5, r5, 8
        addi r6, r6, 8
        addi r7, r7, 8
        addi r8, r8, 8
        addi r3, r3, 8
        cmpw cr0, r9, r4
        bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1

now we emit:

.LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_2:        ; no_exit.7.i
        lfdx f0, r8, r2
        stfdx f0, r9, r2
        lfdx f0, r5, r2
        stfdx f0, r7, r2
        lfdx f0, r3, r2
        stfdx f0, r6, r2
        addi r10, r10, 1
        addi r2, r2, 8
        cmpw cr0, r10, r4
        bgt .LBB_main_no_exit_2E_6_2E_i_no_exit_2E_7_2E_i_1

As another more dramatic example, we used to emit this:

.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2:       ; no_exit.1.i19
        lfd f0, 8(r21)
        lfd f4, 8(r3)
        lfd f5, 8(r27)
        lfd f6, 8(r22)
        lfd f7, 8(r5)
        lfd f8, 8(r6)
        lfd f9, 8(r30)
        lfd f10, 8(r11)
        lfd f11, 8(r12)
        fsub f10, f10, f11
        fadd f5, f4, f5
        fmul f5, f5, f1
        fadd f6, f6, f7
        fadd f6, f6, f8
        fadd f6, f6, f9
        fmadd f0, f5, f6, f0
        fnmsub f0, f10, f2, f0
        stfd f0, 8(r4)
        lfd f0, 8(r25)
        lfd f5, 8(r26)
        lfd f6, 8(r23)
        lfd f9, 8(r28)
        lfd f10, 8(r10)
        lfd f12, 8(r9)
        lfd f13, 8(r29)
        fsub f11, f13, f11
        fadd f4, f4, f5
        fmul f4, f4, f1
        fadd f5, f6, f9
        fadd f5, f5, f10
        fadd f5, f5, f12
        fnmsub f0, f4, f5, f0
        fnmsub f0, f11, f3, f0
        stfd f0, 8(r24)
        lfd f0, 8(r8)
        fsub f4, f7, f8
        fsub f5, f12, f10
        fnmsub f0, f5, f2, f0
        fnmsub f0, f4, f3, f0
        stfd f0, 8(r2)
        addi r20, r20, 1
        addi r2, r2, 8
        addi r8, r8, 8
        addi r10, r10, 8
        addi r12, r12, 8
        addi r6, r6, 8
        addi r29, r29, 8
        addi r28, r28, 8
        addi r26, r26, 8
        addi r25, r25, 8
        addi r24, r24, 8
        addi r5, r5, 8
        addi r23, r23, 8
        addi r22, r22, 8
        addi r3, r3, 8
        addi r9, r9, 8
        addi r11, r11, 8
        addi r30, r30, 8
        addi r27, r27, 8
        addi r21, r21, 8
        addi r4, r4, 8
        cmpw cr0, r20, r7
        bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1

we now emit:

.LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_2:       ; no_exit.1.i19
        lfdx f0, r21, r20
        lfdx f4, r3, r20
        lfdx f5, r27, r20
        lfdx f6, r22, r20
        lfdx f7, r5, r20
        lfdx f8, r6, r20
        lfdx f9, r30, r20
        lfdx f10, r11, r20
        lfdx f11, r12, r20
        fsub f10, f10, f11
        fadd f5, f4, f5
        fmul f5, f5, f1
        fadd f6, f6, f7
        fadd f6, f6, f8
        fadd f6, f6, f9
        fmadd f0, f5, f6, f0
        fnmsub f0, f10, f2, f0
        stfdx f0, r4, r20
        lfdx f0, r25, r20
        lfdx f5, r26, r20
        lfdx f6, r23, r20
        lfdx f9, r28, r20
        lfdx f10, r10, r20
        lfdx f12, r9, r20
        lfdx f13, r29, r20
        fsub f11, f13, f11
        fadd f4, f4, f5
        fmul f4, f4, f1
        fadd f5, f6, f9
        fadd f5, f5, f10
        fadd f5, f5, f12
        fnmsub f0, f4, f5, f0
        fnmsub f0, f11, f3, f0
        stfdx f0, r24, r20
        lfdx f0, r8, r20
        fsub f4, f7, f8
        fsub f5, f12, f10
        fnmsub f0, f5, f2, f0
        fnmsub f0, f4, f3, f0
        stfdx f0, r2, r20
        addi r19, r19, 1
        addi r20, r20, 8
        cmpw cr0, r19, r7
        bgt .LBB_main_L_90_no_exit_2E_0_2E_i16_no_exit_2E_1_2E_i19_1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22722 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoSuck the base value out of the UsersToProcess vector into the BasedUser
Chris Lattner [Mon, 8 Aug 2005 22:56:21 +0000 (22:56 +0000)]
Suck the base value out of the UsersToProcess vector into the BasedUser
class to simplify the code.  Fuse two loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22721 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoSplit MoveLoopVariantsToImediateField out from MoveImmediateValues. The
Chris Lattner [Mon, 8 Aug 2005 22:32:34 +0000 (22:32 +0000)]
Split MoveLoopVariantsToImediateField out from MoveImmediateValues.  The
first is a correctness thing, and the later is an optzn thing.  This also
is needed to support a future change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22720 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFactor out some common code, and be smarter about when to emit load hi/lo
Nate Begeman [Mon, 8 Aug 2005 22:22:56 +0000 (22:22 +0000)]
Factor out some common code, and be smarter about when to emit load hi/lo
code sequences.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22719 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoA testcase I don't want to break in the future
Chris Lattner [Mon, 8 Aug 2005 22:13:49 +0000 (22:13 +0000)]
A testcase I don't want to break in the future

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22718 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAllow tools with "consume after" options (like lli) to take more positional
Chris Lattner [Mon, 8 Aug 2005 21:57:27 +0000 (21:57 +0000)]
Allow tools with "consume after" options (like lli) to take more positional
opts than they take directly.  Thanks to John C for pointing this problem
out to me!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22717 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoRemove getImmediateForOpcode, which is now dead.
Chris Lattner [Mon, 8 Aug 2005 21:34:13 +0000 (21:34 +0000)]
Remove getImmediateForOpcode, which is now dead.

Patch by Jim Laskey.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22716 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd new immediate handling support for mul/div.
Chris Lattner [Mon, 8 Aug 2005 21:33:23 +0000 (21:33 +0000)]
Add new immediate handling support for mul/div.
Patch by Jim Laskey!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22715 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd support for OR/XOR/SUB immediates that are handled with the new immediate
Chris Lattner [Mon, 8 Aug 2005 21:30:29 +0000 (21:30 +0000)]
Add support for OR/XOR/SUB immediates that are handled with the new immediate
way.  This allows ORI/ORIS pairs, for example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22714 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoModify the ISD::AND opcode case to use new immediate constant predicates.
Chris Lattner [Mon, 8 Aug 2005 21:24:57 +0000 (21:24 +0000)]
Modify the ISD::AND opcode case to use new immediate constant predicates.
Includes wider support for rotate and mask cases.

Patch by Jim Laskey.

I've requested that Jim add new regression tests the newly handled cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22712 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoModify the ISD::ADD opcode case to use new immediate constant predicates.
Chris Lattner [Mon, 8 Aug 2005 21:21:03 +0000 (21:21 +0000)]
Modify the ISD::ADD opcode case to use new immediate constant predicates.
Includes support for 32-bit constants using addi/addis.

Patch by Jim Laskey.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22711 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoModify existing support functions to use new immediate constant predicates.
Chris Lattner [Mon, 8 Aug 2005 21:12:35 +0000 (21:12 +0000)]
Modify existing support functions to use new immediate constant predicates.
Patch by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22710 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd support predicates for future immediate constant changes.
Chris Lattner [Mon, 8 Aug 2005 21:10:27 +0000 (21:10 +0000)]
Add support predicates for future immediate constant changes.
Patch by Jim Laskey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22709 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoMove IsRunOfOnes to a more logical place and rename to a proper predicate form
Chris Lattner [Mon, 8 Aug 2005 21:08:09 +0000 (21:08 +0000)]
Move IsRunOfOnes to a more logical place and rename to a proper predicate form
(lowercase isXXX).

Patch by Jim Laskey.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22708 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix JIT encoding of ppc mfocrf instruction; the operands were reversed
Nate Begeman [Mon, 8 Aug 2005 20:04:52 +0000 (20:04 +0000)]
Fix JIT encoding of ppc mfocrf instruction; the operands were reversed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22707 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoUse the new 'moveBefore' method to simplify some code. Really, which is
Chris Lattner [Mon, 8 Aug 2005 19:11:57 +0000 (19:11 +0000)]
Use the new 'moveBefore' method to simplify some code.  Really, which is
easier to understand?  :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22706 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoReject command lines that have too many positional arguments passed (e.g.,
Chris Lattner [Mon, 8 Aug 2005 17:25:38 +0000 (17:25 +0000)]
Reject command lines that have too many positional arguments passed (e.g.,
'opt x y').  This fixes PR493.

Patch contributed by Owen Anderson!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22705 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoNot all constants are legal immediates in load/store instructions.
Chris Lattner [Mon, 8 Aug 2005 06:25:50 +0000 (06:25 +0000)]
Not all constants are legal immediates in load/store instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22704 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agonew testcase, not implemented yet
Chris Lattner [Mon, 8 Aug 2005 06:23:47 +0000 (06:23 +0000)]
new testcase, not implemented yet

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22703 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoImplement LoopStrengthReduce/share_code_in_preheader.ll by having one
Chris Lattner [Mon, 8 Aug 2005 05:47:49 +0000 (05:47 +0000)]
Implement LoopStrengthReduce/share_code_in_preheader.ll by having one
rewriter for all code inserted into the preheader, which is never flushed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22702 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoIt is better to not depend on CSE to share multiplies due to IV insertion.
Chris Lattner [Mon, 8 Aug 2005 05:46:51 +0000 (05:46 +0000)]
It is better to not depend on CSE to share multiplies due to IV insertion.
This testcase checks that only one mul is present in the output code, as it
should be.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22701 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoThese are both implemented by a recent LSR patch
Chris Lattner [Mon, 8 Aug 2005 05:29:51 +0000 (05:29 +0000)]
These are both implemented by a recent LSR patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22700 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoImplement a simple optimization for the termination condition of the loop.
Chris Lattner [Mon, 8 Aug 2005 05:28:22 +0000 (05:28 +0000)]
Implement a simple optimization for the termination condition of the loop.
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.

On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:

_foo:
        li r2, 0
.LBB_foo_1:     ; no_exit
        li r5, 0
        stw r5, 0(r3)
        addi r2, r2, 1
        cmpw cr0, r2, r4
        bne .LBB_foo_1  ; no_exit
        blr

instead of:

_foo:
        li r2, 1                ;; IV starts at 1, not 0
.LBB_foo_1:     ; no_exit
        li r5, 0
        stw r5, 0(r3)
        addi r5, r2, 1
        cmpw cr0, r2, r4
        or r2, r5, r5           ;; Reg-reg copy, extra live range
        bne .LBB_foo_1  ; no_exit
        blr

This implements LoopStrengthReduce/exit_compare_live_range.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22699 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd new helper function
Chris Lattner [Mon, 8 Aug 2005 05:21:50 +0000 (05:21 +0000)]
add new helper function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22698 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd a new helper method
Chris Lattner [Mon, 8 Aug 2005 05:21:33 +0000 (05:21 +0000)]
add a new helper method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22697 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoHandle 64-bit constant exprs on 64-bit targets.
Chris Lattner [Mon, 8 Aug 2005 04:26:32 +0000 (04:26 +0000)]
Handle 64-bit constant exprs on 64-bit targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22696 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAll stats are "Number of ..."
Chris Lattner [Sun, 7 Aug 2005 20:02:04 +0000 (20:02 +0000)]
All stats are "Number of ..."

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22694 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoAdd some simple folds that occur in bitfield cases. Fix a minor bug in
Chris Lattner [Sun, 7 Aug 2005 07:03:10 +0000 (07:03 +0000)]
Add some simple folds that occur in bitfield cases.  Fix a minor bug in
isHighOnes, where it would consider 0 to have high ones.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22693 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoFix typoCVS: ----------------------------------------------------------------------
Chris Lattner [Sun, 7 Aug 2005 07:00:52 +0000 (07:00 +0000)]
Fix typoCVS: ----------------------------------------------------------------------

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22692 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd a small simplification that can be exposed after promotion/expansion
Chris Lattner [Sun, 7 Aug 2005 05:00:44 +0000 (05:00 +0000)]
add a small simplification that can be exposed after promotion/expansion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22691 91177308-0d34-0410-b5e6-96231b3b80d8

19 years ago* Use the new PHINode::hasConstantValue method to simplify some code
Chris Lattner [Sun, 7 Aug 2005 04:27:41 +0000 (04:27 +0000)]
* Use the new PHINode::hasConstantValue method to simplify some code
* Teach this code to move allocas out of the loop when tail call eliminating
  a call marked 'tail'.  This implements TailCallElim/move_alloca_for_tail_call.ll
* Do not perform this transformation if a call is marked 'tail' and if there
  are allocas that we cannot move out of the loop in #2.  Doing so would increase
  the stack usage of the function.  This implements fixes
  PR615 and TailCallElim/dont-tce-tail-marked-call.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22690 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoNew testcases for PR615
Chris Lattner [Sun, 7 Aug 2005 04:25:39 +0000 (04:25 +0000)]
New testcases for PR615

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22689 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoConsolidate the GPOpt stuff to all use the Subtarget, instead of still
Chris Lattner [Fri, 5 Aug 2005 22:05:03 +0000 (22:05 +0000)]
Consolidate the GPOpt stuff to all use the Subtarget, instead of still
depending on the command line option.  Now the command line option just
sets the subtarget as appropriate.  G5 opts will now default to on on
G5-enabled nightly testers among other machines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22688 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadjust to change in getSubtarget() api
Chris Lattner [Fri, 5 Aug 2005 21:54:27 +0000 (21:54 +0000)]
adjust to change in getSubtarget() api

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22687 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoSince getSubtarget() always provides a const Subtarget, dont' require the user
Chris Lattner [Fri, 5 Aug 2005 21:53:21 +0000 (21:53 +0000)]
Since getSubtarget() always provides a const Subtarget, dont' require the user
to pass it in.  Also, since it always returns a non-null pointer, make it
return a reference instead for easier use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22686 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoEnable gp optimizations by default when available, even when a target triple
Chris Lattner [Fri, 5 Aug 2005 21:25:13 +0000 (21:25 +0000)]
Enable gp optimizations by default when available, even when a target triple
is available, since the target triple doesn't specify whether to use gpopts
or not.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22685 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoteach TestRunner about prcontext
Chris Lattner [Fri, 5 Aug 2005 19:48:29 +0000 (19:48 +0000)]
teach TestRunner about prcontext

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22683 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agotwo simple testcases loopreduce should handle but does not yet currently
Chris Lattner [Fri, 5 Aug 2005 19:47:39 +0000 (19:47 +0000)]
two simple testcases loopreduce should handle but does not yet currently

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22682 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoadd a note
Chris Lattner [Fri, 5 Aug 2005 19:18:32 +0000 (19:18 +0000)]
add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22681 91177308-0d34-0410-b5e6-96231b3b80d8

19 years agoChange FindEarliestCallSeqEnd (used by libcall insertion) to use a set to
Chris Lattner [Fri, 5 Aug 2005 18:10:27 +0000 (18:10 +0000)]
Change FindEarliestCallSeqEnd (used by libcall insertion) to use a set to
avoid revisiting nodes more than once.  This eliminates a source of
potentially exponential behavior.  For a small function in 191.fma3d
(hexah_stress_divergence_), this speeds up isel from taking > 20mins to
taking 0.07s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22680 91177308-0d34-0410-b5e6-96231b3b80d8