add FCPYS and FCPYD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30995 91177308-0d34-0410-b5e6-96231b3b80d8

Use S_debug for dwarf info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30994 91177308-0d34-0410-b5e6-96231b3b80d8

Proper fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30993 91177308-0d34-0410-b5e6-96231b3b80d8

One more try.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30992 91177308-0d34-0410-b5e6-96231b3b80d8

Undo Chris' last patch, it caused a regression.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30991 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak x86-64 build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30990 91177308-0d34-0410-b5e6-96231b3b80d8

Be careful when looking through a vbit_convert. Optimizing this:
(vector_shuffle
  (vbitconvert (vbuildvector (copyfromreg v4f32), 1, v4f32), 4, f32),
  (undef, undef, undef, undef), (0, 0, 0, 0), 4, f32)
to the
  vbitconvert
is a very bad idea.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30989 91177308-0d34-0410-b5e6-96231b3b80d8

add fdivs e fdivd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30988 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::SHL_PARTS, ISD::SRA_PARTS and ISD::SRL_PARTS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30987 91177308-0d34-0410-b5e6-96231b3b80d8

Added a X86CompilationCallback variant which saves XMM argument registers for targets with SSE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30986 91177308-0d34-0410-b5e6-96231b3b80d8

Proper fix for rdar://problem/4770604   Thanks to Stuart Hastings!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30985 91177308-0d34-0410-b5e6-96231b3b80d8

Pass AliasAnalysis thru to DAGCombiner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30984 91177308-0d34-0410-b5e6-96231b3b80d8

add a ReplaceMBBInJumpTables method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30983 91177308-0d34-0410-b5e6-96231b3b80d8

Global name regression.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30982 91177308-0d34-0410-b5e6-96231b3b80d8

define the DFPBinOp class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30981 91177308-0d34-0410-b5e6-96231b3b80d8

add the FPBinOp class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30980 91177308-0d34-0410-b5e6-96231b3b80d8

define the Addr1BinOp class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30979 91177308-0d34-0410-b5e6-96231b3b80d8

define the IntBinOp class and use it to implement the multiply instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30978 91177308-0d34-0410-b5e6-96231b3b80d8

fix assembly syntax

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30977 91177308-0d34-0410-b5e6-96231b3b80d8

implement LDRB, LDRSB, LDRH and LDRSH

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30976 91177308-0d34-0410-b5e6-96231b3b80d8

implement smull and umull

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30975 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid getting into an infinite loop when -disable-x86-shuffle-opti is specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30974 91177308-0d34-0410-b5e6-96231b3b80d8

SelectScalarSSELoad should call CanBeFoldedBy as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30973 91177308-0d34-0410-b5e6-96231b3b80d8

Passing isel root and use operands to ComplexPattern functions, these should do the usual load folding checks as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30972 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30971 91177308-0d34-0410-b5e6-96231b3b80d8

Produce a useful error message for Regression/Verifier/2006-10-15-AddrLabel.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30970 91177308-0d34-0410-b5e6-96231b3b80d8

testcase that causes the asmparser to assert

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30969 91177308-0d34-0410-b5e6-96231b3b80d8

label* is not a valid type, add an assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30968 91177308-0d34-0410-b5e6-96231b3b80d8

fix a buggy check that accidentally disabled this xform

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30967 91177308-0d34-0410-b5e6-96231b3b80d8

clarify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30966 91177308-0d34-0410-b5e6-96231b3b80d8

Document the VAArg, and vector instruction opcodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30965 91177308-0d34-0410-b5e6-96231b3b80d8

Align stack size to DWORD boundary

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30964 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::BRCOND

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30963 91177308-0d34-0410-b5e6-96231b3b80d8

fix some fp condition codes
use non trapping comparison instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30962 91177308-0d34-0410-b5e6-96231b3b80d8

Tidy up after truncstore changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30961 91177308-0d34-0410-b5e6-96231b3b80d8

One cyclic dag a day is good for ya...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30960 91177308-0d34-0410-b5e6-96231b3b80d8

Debug tweak.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30959 91177308-0d34-0410-b5e6-96231b3b80d8

Corrected load folding check. We need to start from the root of the sub-dag
being matched and ensure there isn't a non-direct path to the load (i.e. a
path that goes out of the sub-dag.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30958 91177308-0d34-0410-b5e6-96231b3b80d8

Start checking from the root of the matched sub-tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30957 91177308-0d34-0410-b5e6-96231b3b80d8

When checking if a load can be folded, we check if there is any non-direct
way to reach the load via any nodes that would be folded. Start from the
root of the matched sub-tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30956 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure that the node returned by SimplifySetCC is added to the worklist
so that it can be deleted if unused.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30955 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for recent dag combiner patch.  Before the entry bb was:

_test:
        stwu r1, -80(r1)
        stw r1, 76(r1)
        mflr r11
        stw r11, 88(r1)
        rlwinm r2, r4, 0, 30, 31
        cmplwi cr7, r2, 2
        mfcr r2
        rlwinm r2, r2, 30, 31, 31
        cmplwi cr0, r2, 0
        bne cr0, LBB1_2 ;UnifiedReturnBlock

after it is:

_test:
        stwu r1, -80(r1)
        stw r1, 76(r1)
        mflr r11
        stw r11, 88(r1)
        rlwinm r2, r4, 0, 30, 31
        cmplwi cr0, r2, 2
        bgt cr0, LBB1_2 ;UnifiedReturnBlock

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30954 91177308-0d34-0410-b5e6-96231b3b80d8

fold setcc of a setcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30953 91177308-0d34-0410-b5e6-96231b3b80d8

When SimplifySetCC was moved to the DAGCombiner, it was never removed from
SelectionDAG and it has since bitrotted.  Remove the copy from SelectionDAG.
Next, remove the constant folding piece of DAGCombiner::SimplifySetCC into
a new FoldSetCC method which can be used by getNode() and SimplifySetCC.

This fixes obscure bugs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30952 91177308-0d34-0410-b5e6-96231b3b80d8

disable this pass for now, it's causing issues

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30951 91177308-0d34-0410-b5e6-96231b3b80d8

falling off the end of a function is ok with an unreachable instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30950 91177308-0d34-0410-b5e6-96231b3b80d8

testcase for recent selectiondag patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30949 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce the workload by not adding chain users to work list.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30948 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug where we incorrectly turned '(X & 0) == 0' into '(X & 0) >> -1',
which is undefined.  "0" isn't a power of 2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30947 91177308-0d34-0410-b5e6-96231b3b80d8

implement branch inspection/modification methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30946 91177308-0d34-0410-b5e6-96231b3b80d8

Merge ISD::TRUNCSTORE to ISD::STORE. Switch to using StoreSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30945 91177308-0d34-0410-b5e6-96231b3b80d8

it is easier to implement these when they are virtual

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30944 91177308-0d34-0410-b5e6-96231b3b80d8

allow branch reversal to fail

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30943 91177308-0d34-0410-b5e6-96231b3b80d8

Corrected formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30942 91177308-0d34-0410-b5e6-96231b3b80d8

add the branch folding pass as a late cleanup pass for all targets.  For now
it just deletes empty MBB's.  Soon it will do more :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30941 91177308-0d34-0410-b5e6-96231b3b80d8

replace the existing branch inspection/modification APIs with something more
useful and general.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30940 91177308-0d34-0410-b5e6-96231b3b80d8

disable some objectionable code, maybe we can bring this pass to life

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30939 91177308-0d34-0410-b5e6-96231b3b80d8

remove some dead code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30938 91177308-0d34-0410-b5e6-96231b3b80d8

add note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30937 91177308-0d34-0410-b5e6-96231b3b80d8

set isBarrier correctly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30936 91177308-0d34-0410-b5e6-96231b3b80d8

Correctly handle instruction separators.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30935 91177308-0d34-0410-b5e6-96231b3b80d8

Expose method and ivars for measuring inline asm length properly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30934 91177308-0d34-0410-b5e6-96231b3b80d8

Move some warnings to debug mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30933 91177308-0d34-0410-b5e6-96231b3b80d8

add FNEGS and FNEGD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30932 91177308-0d34-0410-b5e6-96231b3b80d8

Fix another dtor issue.  The function local statics in this function were
being destroyed at inconvenient times.  Switch to using non-local ManagedStatic
objects, which actually also speeds up ConstRules::get.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30931 91177308-0d34-0410-b5e6-96231b3b80d8

add SBCS and SUBS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30930 91177308-0d34-0410-b5e6-96231b3b80d8

implement calls to functions that return long

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30929 91177308-0d34-0410-b5e6-96231b3b80d8

implement unordered floating point compares

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30928 91177308-0d34-0410-b5e6-96231b3b80d8

Workaround for templates

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30927 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up dump.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30926 91177308-0d34-0410-b5e6-96231b3b80d8

avoid a ctor/dtor issue with the ProgramName global.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30925 91177308-0d34-0410-b5e6-96231b3b80d8

simplify trivial function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30924 91177308-0d34-0410-b5e6-96231b3b80d8

Really remove dead nodes from isel queue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30923 91177308-0d34-0410-b5e6-96231b3b80d8

shrink anon-ns and mark stuff static.  No functionality changes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30922 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30921 91177308-0d34-0410-b5e6-96231b3b80d8

Lower X%C into X/C+stuff.  This allows the 'division by a constant' logic to
apply to rems as well as divs.  This fixes PR945 and speeds up ReedSolomon
from 14.57s to 10.90s (which is now faster than gcc).

It compiles CodeGen/X86/rem.ll into:

_test1:
        subl $4, %esp
        movl %esi, (%esp)
        movl $2155905153, %ecx
        movl 8(%esp), %esi
        movl %esi, %eax
        imull %ecx
        addl %esi, %edx
        movl %edx, %eax
        shrl $31, %eax
        sarl $7, %edx
        addl %eax, %edx
        imull $255, %edx, %eax
        subl %eax, %esi
        movl %esi, %eax
        movl (%esp), %esi
        addl $4, %esp
        ret
_test2:
        movl 4(%esp), %eax
        movl %eax, %ecx
        sarl $31, %ecx
        shrl $24, %ecx
        addl %eax, %ecx
        andl $4294967040, %ecx
        subl %ecx, %eax
        ret
_test3:
        subl $4, %esp
        movl %esi, (%esp)
        movl $2155905153, %ecx
        movl 8(%esp), %esi
        movl %esi, %eax
        mull %ecx
        shrl $7, %edx
        imull $255, %edx, %eax
        subl %eax, %esi
        movl %esi, %eax
        movl (%esp), %esi
        addl $4, %esp
        ret

instead of div/idiv instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30920 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30919 91177308-0d34-0410-b5e6-96231b3b80d8

Add another cyclic dag test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30918 91177308-0d34-0410-b5e6-96231b3b80d8

Use RemoveDeadNode to kill dead node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30917 91177308-0d34-0410-b5e6-96231b3b80d8

Add RemoveDeadNode to remove a dead node and its (potentially) dead operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30916 91177308-0d34-0410-b5e6-96231b3b80d8

add a minor dag combine noticed when looking at PR945

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30915 91177308-0d34-0410-b5e6-96231b3b80d8

Doh. This wasn't causing problems by luck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30914 91177308-0d34-0410-b5e6-96231b3b80d8

Some X86ISD::CMP were created with wrong ValueType's.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30913 91177308-0d34-0410-b5e6-96231b3b80d8

Fix massive resource leaks in the bytecode reader.  Reading a bytecode file
with ParseBytecodeFile used to leak both a ModuleProvider (and related
bytecode parser stuff attached to it) AND a file descriptor, which was
never closed.  This prevented gccld/llvm-ld/llvm-link from linking together
apps with more that ~252 .bc files on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30912 91177308-0d34-0410-b5e6-96231b3b80d8

mark call adjustments as modifying the SP

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30911 91177308-0d34-0410-b5e6-96231b3b80d8

adjcallstack up/down clobbers the sp

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30910 91177308-0d34-0410-b5e6-96231b3b80d8

adjcallstackup/down clobbers the stack pointer

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30909 91177308-0d34-0410-b5e6-96231b3b80d8

mark adjcallstack up/down as clobbering and using the SP

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30908 91177308-0d34-0410-b5e6-96231b3b80d8

Move the Imp tblgen class from the X86 backend to common code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30907 91177308-0d34-0410-b5e6-96231b3b80d8

restore my previous patch, now that the X86 backend bug has been fixed:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20061009/038518.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30906 91177308-0d34-0410-b5e6-96231b3b80d8

Mark ADJCALLSTACKUP/DOWN as clobbering ESP so that virtregmap will notice
that it can't assume ESP is unmodified across the instrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30905 91177308-0d34-0410-b5e6-96231b3b80d8

D'oh - need to use the rigth kind of store.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30903 91177308-0d34-0410-b5e6-96231b3b80d8

Backing out Chris' last commit. It's breaking llvm-gcc bootstrapping.

It's turning:
        movl -24(%ebp), %esp
        subl $16, %esp
        movl -24(%ebp), %ecx
into
        movl -24(%ebp), %esp
        subl $16, %esp
        movl %esp, (%esp)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30902 91177308-0d34-0410-b5e6-96231b3b80d8

fix compilation failure of smg2000

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30900 91177308-0d34-0410-b5e6-96231b3b80d8

If we see a load from a stack slot into a physreg, consider it as providing
the stack slot.  This fixes PR943.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30898 91177308-0d34-0410-b5e6-96231b3b80d8

Prior load folding check on chain operand was too strict. It requires the
chain operand to point to the load being folded. Now we relax this, traversing
up the chain, if it doesn't reach the load, then it's ok. We will create a
TokenFactor (of all the chain operands and the load's chain) to capture all
the control flow dependencies.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30897 91177308-0d34-0410-b5e6-96231b3b80d8

Replace custom dispatch code with two uses of InstVisitor. Improves
compile-time performance.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30896 91177308-0d34-0410-b5e6-96231b3b80d8

Update tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30895 91177308-0d34-0410-b5e6-96231b3b80d8

Fold "zero extending vector loads" now that evan added the chain manip stuff.
This compiles both tests in X86/vec_ss_load_fold.ll into:

_test1:
        movss 4(%esp), %xmm0
        subss LCPI1_0, %xmm0
        mulss LCPI1_1, %xmm0
        minss LCPI1_2, %xmm0
        xorps %xmm1, %xmm1
        maxss %xmm1, %xmm0
        cvttss2si %xmm0, %eax
        andl $65535, %eax
        ret

instead of:

_test1:
        movss LCPI1_0, %xmm0
        movss 4(%esp), %xmm1
        subss %xmm0, %xmm1
        movss LCPI1_1, %xmm0
        mulss %xmm0, %xmm1
        movss LCPI1_2, %xmm0
        minss %xmm0, %xmm1
        xorps %xmm0, %xmm0
        maxss %xmm0, %xmm1
        cvttss2si %xmm1, %eax
        andl $65535, %eax
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30894 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30893 91177308-0d34-0410-b5e6-96231b3b80d8