Add support for let expressions around an mdef.  This implements a new part
of Regression/TableGen/MultiClass.td.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30887 91177308-0d34-0410-b5e6-96231b3b80d8

add test for mdef in a let

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30886 91177308-0d34-0410-b5e6-96231b3b80d8

Clean-up of formatting and spelling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30885 91177308-0d34-0410-b5e6-96231b3b80d8

Typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30884 91177308-0d34-0410-b5e6-96231b3b80d8

Handle aliasing of loadext.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30883 91177308-0d34-0410-b5e6-96231b3b80d8

Explict processor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30882 91177308-0d34-0410-b5e6-96231b3b80d8

Reduce the amount of state in the lowering code and drop old pattern ISel functions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30881 91177308-0d34-0410-b5e6-96231b3b80d8

Fix regression in combiner alias analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30880 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure combiner alias analysis does it's thing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30879 91177308-0d34-0410-b5e6-96231b3b80d8

Naming consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30878 91177308-0d34-0410-b5e6-96231b3b80d8

Test case for the recently fixed FindModifiedNodeSlot bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30877 91177308-0d34-0410-b5e6-96231b3b80d8

Put code example inside of "doc_code" divisions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30876 91177308-0d34-0410-b5e6-96231b3b80d8

Use cute tblgen tricks to make zap handling more powerful.  Specifically,
when the dag combiner simplifies an and mask, notice this and allow those bits
to be missing from the zap mask.

This compiles Alpha/zapnot4.ll into:

        sll $16,3,$0
        zapnot $0,3,$0
        ret $31,($26),1

instead of:

        ldah $0,1($31)
        lda $0,-8($0)
        sll $16,3,$1
        and $1,$0,$0
        ret $31,($26),1

It would be *really* nice to replace the hunk of code in the
AlphaISelDAGToDAG.cpp file that matches (and (srl (x, C), c2) into
(SRL (ZAPNOTi)) with a similar pattern, but I've spent enough time poking
at alpha.  Make andrew will do this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30875 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase, this should turn into zap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30874 91177308-0d34-0410-b5e6-96231b3b80d8

Jimptables working again on alpha.

As a bonus, use the GOT node instead of the AlphaISD::GOT for internal stuff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30873 91177308-0d34-0410-b5e6-96231b3b80d8

Remove dead/redundant instructions.  These are handled by ZAPNOTi

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30872 91177308-0d34-0410-b5e6-96231b3b80d8

Special case tblgen generated code for patterns like (and X, 255) or (or X, 42).
The dag/inst combiners often 'simplify' the masked value based on whether
or not the bits are live or known zero/one.  This is good and dandy, but
often causes special case patterns to fail, such as alpha's CMPBGE pattern,
which looks like "(set GPRC:$RC, (setuge (and GPRC:$RA, 255), (and GPRC:$RB, 255)))".
Here the pattern for (and X, 255) should match actual dags like (and X, 254) if
the dag combiner proved that the missing bits are already zero (one for 'or').

For CodeGen/Alpha/cmpbge.ll:test2 for example, this results in:

        sll $16,1,$0
        cmpbge $0,$17,$0
        ret $31,($26),1

instead of:

        sll $16,1,$0
        and $0,254,$0
        and $17,255,$1
        cmpule $1,$0,$0
        ret $31,($26),1

... and requires no target-specific code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30871 91177308-0d34-0410-b5e6-96231b3b80d8

Both of these functions should turn into cmpbge instructions, even though
the second has an and of 254 not 255.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30870 91177308-0d34-0410-b5e6-96231b3b80d8

add two helper methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30869 91177308-0d34-0410-b5e6-96231b3b80d8

Split some code out into a new method.  The generated code is exactly
identical, this is just a refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30868 91177308-0d34-0410-b5e6-96231b3b80d8

This entry is done.  switched to the gcc way of doing things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30867 91177308-0d34-0410-b5e6-96231b3b80d8

FindModifiedNodeSlot needs to add LoadSDNode ivars to create proper SelectionDAGCSEMap ID.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30866 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for zap generation, from the alpha readme

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30865 91177308-0d34-0410-b5e6-96231b3b80d8

This has apparently been fixed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30864 91177308-0d34-0410-b5e6-96231b3b80d8

Make the bugpoint reduction heuristics more effective.  Patch submitted by
Domagoj Babic, thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30863 91177308-0d34-0410-b5e6-96231b3b80d8

uint <-> double conversion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30862 91177308-0d34-0410-b5e6-96231b3b80d8

Also update getNodeLabel for LoadSDNode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30861 91177308-0d34-0410-b5e6-96231b3b80d8

SDNode::dump should also print out extension type and VT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30860 91177308-0d34-0410-b5e6-96231b3b80d8

add fp sub

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30859 91177308-0d34-0410-b5e6-96231b3b80d8

add double <-> int conversion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30858 91177308-0d34-0410-b5e6-96231b3b80d8

Fix another bug in extload promotion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30857 91177308-0d34-0410-b5e6-96231b3b80d8

compare doubles

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30856 91177308-0d34-0410-b5e6-96231b3b80d8

add some tests for floating point compare

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30855 91177308-0d34-0410-b5e6-96231b3b80d8

initial support for fp compares. Unordered compares not implemented yet

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30854 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug introduced by my LOAD/LOADX changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30853 91177308-0d34-0410-b5e6-96231b3b80d8

Add test case from PR940.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30852 91177308-0d34-0410-b5e6-96231b3b80d8

More isel time load folding checking for nodes that produce flag values.
See comment in CanBeFoldedBy() for detailed explanation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30851 91177308-0d34-0410-b5e6-96231b3b80d8

Comments; getChain(), getBasePtr(), etc. should return a SDOperand by value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30850 91177308-0d34-0410-b5e6-96231b3b80d8

shufps with load folding is better than movaps; movsd.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30849 91177308-0d34-0410-b5e6-96231b3b80d8

Don't go too crazy with these AddComplexity. Try matching shufps with load
folding first.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30848 91177308-0d34-0410-b5e6-96231b3b80d8

Don't convert to MOVLP if using shufps etc. may allow load folding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30847 91177308-0d34-0410-b5e6-96231b3b80d8

Use FindProgramByName instead of FindExecutable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30846 91177308-0d34-0410-b5e6-96231b3b80d8

Predicate function on the node should be matched before its childrean' matching
code. This is especially important now matching ISD::LOAD also requires a
Predicate_Load call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30845 91177308-0d34-0410-b5e6-96231b3b80d8

Reflects ISD::LOAD / ISD::LOADX / LoadSDNode changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30844 91177308-0d34-0410-b5e6-96231b3b80d8

Merging ISD::LOAD and ISD::LOADX. Added LoadSDNode to represent load nodes.
Chain and address ptr remains as operands. SrcValue, extending mode, extending
VT (or rather loaded VT before extension) are now instance variables of
LoadSDNode.

Introduce load / store addressing modes to represent pre- and post-indexed
load and store. Also added an additional operand offset that is only used in
post-indexed mode (i.e. base ptr += offset after load/store).

Added alignment info (not yet used) and isVolatile fields.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30843 91177308-0d34-0410-b5e6-96231b3b80d8

Do error checking.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30842 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a dead var noticed by Yorion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30841 91177308-0d34-0410-b5e6-96231b3b80d8

add some tests for floating point arithmetic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30840 91177308-0d34-0410-b5e6-96231b3b80d8

Fix build error in gcc 3.4 and make more this general

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30839 91177308-0d34-0410-b5e6-96231b3b80d8

Use GetTemporaryDirectory. Fix llvm.org/bugs/show_bug.cgi?id=894

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30838 91177308-0d34-0410-b5e6-96231b3b80d8

Fix usage example.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30837 91177308-0d34-0410-b5e6-96231b3b80d8

add float -> double and double -> float conversion

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30835 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR886:
The result of yyparse() was not being checked. When YYERROR or YYABORT is
called it causes yyparse() to return 1 to indicate the error. The code was
silently ignoring this situation because it previously expected either an
exception or a null ParserResult to indicate an error. The patch corrects
this situation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30834 91177308-0d34-0410-b5e6-96231b3b80d8

xfail this until andrew has a chance to implement the alpha TM method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30833 91177308-0d34-0410-b5e6-96231b3b80d8

add a test for adds adcs

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30832 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug pointed out by Zhongxing Xu

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30831 91177308-0d34-0410-b5e6-96231b3b80d8

add ADDS and ADCS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30830 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::SELECT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30829 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30828 91177308-0d34-0410-b5e6-96231b3b80d8

expand ISD::EXTLOAD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30827 91177308-0d34-0410-b5e6-96231b3b80d8

most ARM targets are little endian

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30826 91177308-0d34-0410-b5e6-96231b3b80d8

Implement SROA of unions with mixed pointers/integers in them.  This implements
PR892 and Transforms/ScalarRepl/union-pointer.ll:test2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30825 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR892

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30824 91177308-0d34-0410-b5e6-96231b3b80d8

Implement Transforms/ScalarRepl/union-pointer.ll:test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30823 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for SROA for stuff like "union { int*, float* }".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30822 91177308-0d34-0410-b5e6-96231b3b80d8

Eliminate more token factors by taking advantage of transitivity:
if TF depends on A and B, and A depends on B, TF just needs to depend on
A.  With Jim's alias-analysis stuff enabled, this compiles the testcase in
PR892 into:

__Z4test3Val:
        subl $44, %esp
        call L__Z3foov$stub
        movl %edx, 28(%esp)
        movl %eax, 32(%esp)
        movl %eax, 24(%esp)
        movl %edx, 36(%esp)
        movl 52(%esp), %ecx
        movl %ecx, 4(%esp)
        movl %eax, 8(%esp)
        movl %edx, 12(%esp)
        movl 48(%esp), %eax
        movl %eax, (%esp)
        call L__Z3bar3ValS_$stub
        addl $44, %esp
        ret

instead of:

__Z4test3Val:
        subl $44, %esp
        call L__Z3foov$stub
        movl %eax, 24(%esp)
        movl %edx, 28(%esp)
        movl 24(%esp), %eax
        movl %eax, 32(%esp)
        movl 28(%esp), %eax
        movl %eax, 36(%esp)
        movl 32(%esp), %eax
        movl 36(%esp), %ecx
        movl 52(%esp), %edx
        movl %edx, 4(%esp)
        movl %eax, 8(%esp)
        movl %ecx, 12(%esp)
        movl 48(%esp), %eax
        movl %eax, (%esp)
        call L__Z3bar3ValS_$stub
        addl $44, %esp
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30821 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR897

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30820 91177308-0d34-0410-b5e6-96231b3b80d8

Combiner alias analysis passes  Multisource (release-asserts.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30818 91177308-0d34-0410-b5e6-96231b3b80d8

completely disable folding of loads into scalar sse instructions and provide
a framework for doing it right.  This fixes
CodeGen/X86/2006-10-07-ScalarSSEMiscompile.ll.

Once X86DAGToDAGISel::SelectScalarSSELoad is implemented right, this task
will be done.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30817 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase we miscompile

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30816 91177308-0d34-0410-b5e6-96231b3b80d8

convert packed FP add/sub/mul/div to use a multiclass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30815 91177308-0d34-0410-b5e6-96231b3b80d8

one multiclass now defines all 8 variants of binary-scalar-sse-fp operations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30814 91177308-0d34-0410-b5e6-96231b3b80d8

Switch ADD/MUL/DIV/SUB scalarsse fp ops to a multiclass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30813 91177308-0d34-0410-b5e6-96231b3b80d8

Random acts of shrinkage

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30812 91177308-0d34-0410-b5e6-96231b3b80d8

Convert pand/por/pxor to use multiclass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30811 91177308-0d34-0410-b5e6-96231b3b80d8

Convert some more instructions over to use a new multiclass.

Fix a bug where the asmstring for PSUBQrm was wrong.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30810 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug where PADDQrm printed paddd instead of paddq.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30809 91177308-0d34-0410-b5e6-96231b3b80d8

Add multiclass for SSE2 instructions that correspond to simple binops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30808 91177308-0d34-0410-b5e6-96231b3b80d8

rename:
  PDI_binop_rm -> PDI_binop_rm_int
  PDI_binop_rmi -> PDI_binop_rmi_int

to make it clear that these are for use with intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30807 91177308-0d34-0410-b5e6-96231b3b80d8

Convert saturating PADD/PSUB's to use a multiclass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30806 91177308-0d34-0410-b5e6-96231b3b80d8

Convert PAVG*, PMADDWD, and PMUL* to use multiclasses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30805 91177308-0d34-0410-b5e6-96231b3b80d8

Fix typo in packsswb instr definition, where the load had the wrong type.
This allows us to use the multiclass for other packs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30804 91177308-0d34-0410-b5e6-96231b3b80d8

implement FUITOS and FUITOD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30803 91177308-0d34-0410-b5e6-96231b3b80d8

implement FLDD

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30802 91177308-0d34-0410-b5e6-96231b3b80d8

implement fadds, faddd, fmuls and fmuld

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30801 91177308-0d34-0410-b5e6-96231b3b80d8

handle pmin/pmax with multiclasses

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30800 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30799 91177308-0d34-0410-b5e6-96231b3b80d8

Bugfix: this allows multiclasses to have default arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30798 91177308-0d34-0410-b5e6-96231b3b80d8

simplify pack and shift intrinsics with multiclasses

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30797 91177308-0d34-0410-b5e6-96231b3b80d8

Use a multiclass to simplify 'SSE2 Integer comparison'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30796 91177308-0d34-0410-b5e6-96231b3b80d8

move class defns close to uses to make it easier to read

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30795 91177308-0d34-0410-b5e6-96231b3b80d8

simplify horizontal op definitions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30794 91177308-0d34-0410-b5e6-96231b3b80d8

remove more unneeded type info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30793 91177308-0d34-0410-b5e6-96231b3b80d8

remove unneeded definitions and type info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30792 91177308-0d34-0410-b5e6-96231b3b80d8

remove some unneeded type info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30791 91177308-0d34-0410-b5e6-96231b3b80d8

simplify patterns by merging in operand info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30790 91177308-0d34-0410-b5e6-96231b3b80d8

Factor operands into packed unary classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30789 91177308-0d34-0410-b5e6-96231b3b80d8

remove dead/duplicate instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30788 91177308-0d34-0410-b5e6-96231b3b80d8

Pull operand info up into parent class for scalar sse intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30787 91177308-0d34-0410-b5e6-96231b3b80d8

convert the sole sd unary intrinsic to a multiclass for consistency

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@30786 91177308-0d34-0410-b5e6-96231b3b80d8