move some methods, no other changes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23613 91177308-0d34-0410-b5e6-96231b3b80d8

minor microoptimizations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23612 91177308-0d34-0410-b5e6-96231b3b80d8

Use a map to cache the ModuleType information, so we can do logarithmic
lookups instead of linear time lookups.  This speeds up bc parsing of a
large file from

137.834u 118.256s 4:27.96
to
132.611u 114.436s 4:08.53

with a release build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23611 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor gathering node info and emission.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23610 91177308-0d34-0410-b5e6-96231b3b80d8

clean up this code a bit, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23609 91177308-0d34-0410-b5e6-96231b3b80d8

Speed up the asm printer a lot by not printing formatted LLVM asm output
for globals

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23608 91177308-0d34-0410-b5e6-96231b3b80d8

Break the body of the loop out into a new method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23606 91177308-0d34-0410-b5e6-96231b3b80d8

Fix case of path

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23605 91177308-0d34-0410-b5e6-96231b3b80d8

Make IVUseShouldUsePostIncValue more aggressive when the use is a PHI.  In
particular, it should realize that phi's use their values in the pred block
not the phi block itself.  This change turns our em3d loop from this:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r2, 0
        b LBB_test_6    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        or r2, r6, r6
        lwz r6, 0(r3)
        cmpw cr0, r6, r5
        beq cr0, LBB_test_6     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r2, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; endif.loopexit.loopexit_crit_edge
        addi r3, r2, 1
        blr
LBB_test_6:     ; loopexit
        or r3, r2, r2
        blr

into:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r2, 0
        b LBB_test_5    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        or r2, r6, r6
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        or r2, r6, r6
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; loopexit
        or r3, r2, r2
        blr

Unfortunately, this is actually worse code, because the register coallescer
is getting confused somehow.  If it were doing its job right, it could turn the
code into this:

_test:
        cmpwi cr0, r4, 0
        bgt cr0, LBB_test_2     ; entry.no_exit_crit_edge
LBB_test_1:     ; entry.loopexit_crit_edge
        li r6, 0
        b LBB_test_5    ; loopexit
LBB_test_2:     ; entry.no_exit_crit_edge
        li r6, 0
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit
LBB_test_5:     ; loopexit
        or r3, r6, r6
        blr

... which I'll work on next. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23604 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor some code into a function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23603 91177308-0d34-0410-b5e6-96231b3b80d8

This break is bogus and I have no idea why it was there.  Basically it prevents
memoizing code when IV's are used by phinodes outside of loops.  In a simple
example, we were getting this code before (note that r6 and r7 are isomorphic
IV's):

        li r6, 0
        or r7, r6, r6
LBB_test_3:     ; no_exit
        lwz r2, 0(r3)
        cmpw cr0, r2, r5
        or r2, r7, r7
        beq cr0, LBB_test_5     ; loopexit
LBB_test_4:     ; endif
        addi r2, r7, 1
        addi r7, r7, 1
        addi r3, r3, 4
        addi r6, r6, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit

Now we get:

        li r6, 0
LBB_test_3:     ; no_exit
        or r2, r6, r6
        lwz r6, 0(r3)
        cmpw cr0, r6, r5
        beq cr0, LBB_test_6     ; loopexit
LBB_test_4:     ; endif
        addi r3, r3, 4
        addi r6, r2, 1
        cmpw cr0, r6, r4
        blt cr0, LBB_test_3     ; no_exit

this was noticed in em3d.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23602 91177308-0d34-0410-b5e6-96231b3b80d8

when checking if we should move a split edge block outside of a loop,
check the presplit pred, not the post-split pred.  This was causing us
to make the wrong decision in some cases, leaving the critical edge block
in the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23601 91177308-0d34-0410-b5e6-96231b3b80d8

This member can be const too

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23600 91177308-0d34-0410-b5e6-96231b3b80d8

put the right labels on the data

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23599 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a problem where the legalizer would run out of stack space on extremely
large basic blocks because it was purely recursive.  This switches it to an
iterative/recursive hybrid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23596 91177308-0d34-0410-b5e6-96231b3b80d8

silence a bogus warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23595 91177308-0d34-0410-b5e6-96231b3b80d8

silence some warnings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23594 91177308-0d34-0410-b5e6-96231b3b80d8

silence a warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23593 91177308-0d34-0410-b5e6-96231b3b80d8

add patterns for float binops and fma ops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23592 91177308-0d34-0410-b5e6-96231b3b80d8

Sort the cpu and features table, so that the alpha backend doesn't fail EVERY
compile with an assertion that the tables are not sorted!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23591 91177308-0d34-0410-b5e6-96231b3b80d8

Add assertions to the trivial scheduler to check that the value types match
up between defs and uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23590 91177308-0d34-0410-b5e6-96231b3b80d8

another solution to the fsel issue.  Instead of having 4 variants, just force
the comparison to be 64-bits.  This is fine because extensions from float
to double are free.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23589 91177308-0d34-0410-b5e6-96231b3b80d8

fsel can take a different FP type for the comparison and for the result.  As such
split the FSEL family into 4 things instead of just two.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23588 91177308-0d34-0410-b5e6-96231b3b80d8

fix an f32/f64 type mismatch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23587 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen CopyFromReg using the regclass that matches the valuetype of the
destination vreg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23586 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to change in ctor

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23585 91177308-0d34-0410-b5e6-96231b3b80d8

Emit the value type for each register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23584 91177308-0d34-0410-b5e6-96231b3b80d8

Expose the actual valuetype of each register class

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23583 91177308-0d34-0410-b5e6-96231b3b80d8

Minor tweak to the branch selector.  When emitting a two-way branch, and if
we're in a single-mbb loop, make sure to emit the backwards branch as the
conditional branch instead of the uncond branch.  For example, emit this:

LBBl29_z__44:
        stw r9, 0(r15)
        stw r9, 4(r15)
        stw r9, 8(r15)
        stw r9, 12(r15)
        addi r15, r15, 16
        addi r8, r8, 1
        cmpw cr0, r8, r28
        ble cr0, LBBl29_z__44
        b LBBl29_z__48                   *** NOT PART OF LOOP

Instead of:

LBBl29_z__44:
        stw r9, 0(r15)
        stw r9, 4(r15)
        stw r9, 8(r15)
        stw r9, 12(r15)
        addi r15, r15, 16
        addi r8, r8, 1
        cmpw cr0, r8, r28
        bgt cr0, LBBl29_z__48            *** PART OF LOOP!
        b LBBl29_z__44

The former sequence has one fewer dispatch group for the loop body.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23582 91177308-0d34-0410-b5e6-96231b3b80d8

like the comment says, enable this

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23581 91177308-0d34-0410-b5e6-96231b3b80d8

Add some very paranoid checking for operand/result reg class matchup
For instructions that define multiple results, use the right regclass
to define the result, not always the rc of result #0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23580 91177308-0d34-0410-b5e6-96231b3b80d8

Fix VC++ warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23579 91177308-0d34-0410-b5e6-96231b3b80d8

fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23578 91177308-0d34-0410-b5e6-96231b3b80d8

Modify the ppc backend to use two register classes for FP: F8RC and F4RC.
These are used to represent float and double values, and the two regclasses
contain the same physical registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23577 91177308-0d34-0410-b5e6-96231b3b80d8

Annotate nodes with their addresses if a graph requests it.

This is Jim's feature implemented so that graphs could 'opt-in' and get
this behavior.  This is currently used by selection dags.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23576 91177308-0d34-0410-b5e6-96231b3b80d8

add a method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23575 91177308-0d34-0410-b5e6-96231b3b80d8

typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23574 91177308-0d34-0410-b5e6-96231b3b80d8

1. Simplify the gathering of node groups.

2. Printing node groups when displaying nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23573 91177308-0d34-0410-b5e6-96231b3b80d8

Should be using flag and not chain.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23572 91177308-0d34-0410-b5e6-96231b3b80d8

Remove some now-dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23571 91177308-0d34-0410-b5e6-96231b3b80d8

more specific tests of subtarget stuff

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23570 91177308-0d34-0410-b5e6-96231b3b80d8

subtarget support for CIX and FIX extentions (the only 2 I care about right now)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23569 91177308-0d34-0410-b5e6-96231b3b80d8

Reverting change moving to selection dag graph.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23568 91177308-0d34-0410-b5e6-96231b3b80d8

Added allnodes_size for scheduling support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23567 91177308-0d34-0410-b5e6-96231b3b80d8

1. Made things node-centric (from operand).

2. Added node groups to handle flagged nodes.

3. Started weaning simple scheduling off existing emitter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23566 91177308-0d34-0410-b5e6-96231b3b80d8

Add the node name (thus the address) to node label.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23565 91177308-0d34-0410-b5e6-96231b3b80d8

Rename MRegisterDesc -> TargetRegisterDesc for consistency

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23564 91177308-0d34-0410-b5e6-96231b3b80d8

Update the discussion of TargetRegisterDesc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23563 91177308-0d34-0410-b5e6-96231b3b80d8

remove some more initializers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23562 91177308-0d34-0410-b5e6-96231b3b80d8

trim down the target info structs now that we have a preferred spill register class for each callee save register

Why is V9 maintaining these tables manually? ugh!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23561 91177308-0d34-0410-b5e6-96231b3b80d8

trim down the target info structs now that we have a preferred spill register class for each callee save register

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23560 91177308-0d34-0410-b5e6-96231b3b80d8

now that we have a reg class to spill with, get this info from the regclass

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23559 91177308-0d34-0410-b5e6-96231b3b80d8

constant fold these calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23558 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code using the new regclass info passed in

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23557 91177308-0d34-0410-b5e6-96231b3b80d8

Now that we have getCalleeSaveRegClasses() info, use it to pass the register
class into the spill/reload methods.  Targets can now rely on that argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23556 91177308-0d34-0410-b5e6-96231b3b80d8

expose a new virtual method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23555 91177308-0d34-0410-b5e6-96231b3b80d8

stub out a virtual method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23554 91177308-0d34-0410-b5e6-96231b3b80d8

Compute a preferred spill register class for each callee-save register

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23553 91177308-0d34-0410-b5e6-96231b3b80d8

CR registers are not used by this "target"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23552 91177308-0d34-0410-b5e6-96231b3b80d8

these registers don't belong to any register classes, so don't mark them
as callee save.  They can never be generated by the compiler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23551 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23550 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23549 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor this a bit to move ParsingTemplateArgs to only apply to classes,
not defs.

Implement support for forward definitions of classes.  This implements
TableGen/ForwardRef.td.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23548 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test that you can forward ref a class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23547 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23546 91177308-0d34-0410-b5e6-96231b3b80d8

Generate a parse error instead of a checked exception if template args are
used on a def.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23545 91177308-0d34-0410-b5e6-96231b3b80d8

Now that self referential classes are supported, get rid of a work-around.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23544 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23543 91177308-0d34-0410-b5e6-96231b3b80d8

Refactor the grammar a bit to implement TableGen/ForwardRef.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23542 91177308-0d34-0410-b5e6-96231b3b80d8

Check that we can refer to the same class we are defining.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23541 91177308-0d34-0410-b5e6-96231b3b80d8

allow regs to be in multiple reg classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23540 91177308-0d34-0410-b5e6-96231b3b80d8

pass extra args

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23539 91177308-0d34-0410-b5e6-96231b3b80d8

these methods get extra args

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23538 91177308-0d34-0410-b5e6-96231b3b80d8

Pass extra regclasses into spilling code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23537 91177308-0d34-0410-b5e6-96231b3b80d8

Change this code ot pass register classes into the stack slot spiller/reloader
code.  PrologEpilogInserter hasn't been updated yet though, so targets cannot
use this info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23536 91177308-0d34-0410-b5e6-96231b3b80d8

Change these methods to take RC's

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23535 91177308-0d34-0410-b5e6-96231b3b80d8

Use the 32-bit version for now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23534 91177308-0d34-0410-b5e6-96231b3b80d8

Add a bunch of patterns for F64 FP ops, add some more integer ops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23533 91177308-0d34-0410-b5e6-96231b3b80d8

Remove code for patterns that are autogenerated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23532 91177308-0d34-0410-b5e6-96231b3b80d8

begining alpha subtarget support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23531 91177308-0d34-0410-b5e6-96231b3b80d8

tblgen autogens this pattern now

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23530 91177308-0d34-0410-b5e6-96231b3b80d8

Teach tablegen to reassociate operators when possible.  This allows it to
find all of teh pattern matches for EQV from one definition

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23529 91177308-0d34-0410-b5e6-96231b3b80d8

copy and paste error

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23528 91177308-0d34-0410-b5e6-96231b3b80d8

now that tblgen is smarter, this pattern is not needed.  Also, tblgen
now inverts commuted versions of ANDC/ORC with the current .td file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23527 91177308-0d34-0410-b5e6-96231b3b80d8

Teach tblgen to build permutations of instructions, so that the target author
doesn't have to specify them manually.  It currently handles associativity,
e.g. knowing that (X*Y)+Z  also matches  X+(Y*Z)  and will be extended in
the future.

It is smart enough to not introduce duplicate patterns or patterns that can
never match.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23526 91177308-0d34-0410-b5e6-96231b3b80d8

Insert stores after phi nodes in the normal dest.  This fixes
LowerInvoke/2005-08-03-InvokeWithPHI.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23525 91177308-0d34-0410-b5e6-96231b3b80d8

consistency with other cases, no functionality change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23524 91177308-0d34-0410-b5e6-96231b3b80d8

Make the JIT default to the DAG isel instead of the pattern isel, like LLC.
The Pattern isel has some strange memory corruption issues going on. :(

This should have been converted over anyway, but it got forgotten somehow
when switching to the dag isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23523 91177308-0d34-0410-b5e6-96231b3b80d8

Fix two bugs in my patch earlier today that broke int->fp conversion on X86.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23522 91177308-0d34-0410-b5e6-96231b3b80d8

Fold isascii into a simple comparison.  This speeds up 197.parser by 7.4%,
bringing the LLC time down to the CBE time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23521 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for isascii

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23520 91177308-0d34-0410-b5e6-96231b3b80d8

remove a bunch of unneeded stuff, or self evident comments

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23519 91177308-0d34-0410-b5e6-96231b3b80d8

add a new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23518 91177308-0d34-0410-b5e6-96231b3b80d8

Implement a couple of memcmp folds from the todo list

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23517 91177308-0d34-0410-b5e6-96231b3b80d8

Silence VC++ redeclaration warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23516 91177308-0d34-0410-b5e6-96231b3b80d8

Never rely on ReplaceAllUsesWith when selecting, use CodeGenMap instead.
ReplaceAllUsesWith does not replace scalars SDOperand floating around on
the stack, permitting things to be selected multiple times.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23515 91177308-0d34-0410-b5e6-96231b3b80d8

Codegen ADD X, IMM -> addis/addi if needed.
This implements PowerPC/fold-li.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23514 91177308-0d34-0410-b5e6-96231b3b80d8

add a testcase for a feature we regressed on because noone wrote the test! :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23513 91177308-0d34-0410-b5e6-96231b3b80d8

Autogen MUL, move FP cases together

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23512 91177308-0d34-0410-b5e6-96231b3b80d8

disentangle FP from INT versions of div/mul

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23511 91177308-0d34-0410-b5e6-96231b3b80d8