►
From YouTube: OMR Compiler Proposed Vector IL Changes 20180412
Description
Meeting agenda: https://github.com/eclipse/omr/issues/2403
Comments and questions may be left in the GitHub issue.
A
B
B
The
main
goal
is
to
exploit
intelligent
design
work
for
lang
pimentel,
which
already
available
it
was
possibly
in
the
future.
That
policy
also
have
bigger
size,
but
not
only
have
120
Intel
has
256
and
512,
so
this
can
be
used
for
arresting
the
exploitation
to
expand
or
other
to
do,
support
for
Panama
project
and
just
comment
like
we
would
like
to
be
able
to
use
them
those
different
sizes
in
one
compilation,
for
example,
Otto
Cindy,
we
should
be
able
to
generate
the
preparations
of
the
current
approach
you
have
now
is
well.
B
We
have
vector
types
and
the
operations,
of
course,
but
128-bit
length
is
implicitly
implied.
Data
types
right
now
describe
expanded
data
type
and
they
describe
vector
layout
precisely
example.
We
have
typed
vector
in
32
since
length
is
implied
128
bit,
so
it
means
it's
a
vector
of
181.
It
is
128
bit
wide
containing
4
3
be
integers.
So
we
have
full
information
encoded
in
type
right
now,
opcode
Nicole
at
type
list.
For
example,
we
add
the
cons
so
we're
missing
the
no
bottom.
B
There
there's
activity
to
know
the
content
of
the
vector
and
actual
exact
type,
but
tight
put.
You
can
be
derived
to
example,
when
we
say
get
type
on
the
node.
We
derive
it
from
either
from
children
from
simple
reference
or
it
can
be
cached
in
the
node
before
the
cost
doesn't
have
children
that
have
severe
form
so
type
that
recite
type
is
cached
on
the
decoder
tonton
the
world
itself.
So
this
is
current
approach.
Oh,
it
has
limitations,
since
only
one
length.
B
Of
course,
only
one
length
is
reported
types
the
dimension
described
precisely
so
we
won't,
if
you
want
to
add
new
types
with
new
support
like
phone
other
lengths,
so
we
would
have
neglected
int.
B
We
need
to
add
length
to
it
and
the
type
of
elements
or
the
more
vector
length
is
support.
The
more
of
these
types
are
going
to
be
adding
so
essentially
expanding
number
of
types
considerably,
and
the
high
price,
of
course,
are
the
goal
of
hang
ever
sub.
Goals
was
to
reduce
number
or
from
records,
but
the
disadvantages
that
incur
some
of
the
extra
AeroPro
because
for
what
I
described
just
before,
if
notice
there
on
the
construction
of
being
modified
incorrect,
I'd
capitalized,
so
that's
the
main
advantage.
B
B
So
what
is
beta
babies
basically
number
of
bigs
summer
vacation,
which
can
be
interpreted
in
any
way,
the
one
in
G
term
we
do
present
data
symbol,
references
and
similar
reference
have
data
type
associated
them,
so
complete
type
can
be
represented
as
triple
it's
a
full
file,
then,
whether
it's
the
integer
or
float
and
scale
of
example,
in
the
most
general
case,
like,
for
example,
128
in,
for
it
means
it
means
the
vector
of
size.
128-Bit
comes
consisting
over
or
integers,
which
means
in
each
integers
32-bit.
B
It's
like
full
information
about
the
type
if
the
three
parameters
so
just
are
like
a
mountain
duration
that
generally
what
happens
in
higher
in
languages
in
high-level
languages.
Programmers
describe
data
types
precisely
random
and
then
they
specify
overall
it
operation.
Then
compilers,
transform
or
little
decorations
into
this
best
ones,
equivocal,
adding
two
integers
to
flows
and
then,
as
soon
as
the
operation
is
present,
we
might
not
even
need
data
type
written
in
associated
with
data.
I
Gloria
know
what
to
do
with
my
data,
so
we
would
put
all
the
information
in
represent
anything
operation
themselves.
B
So
operation,
when
you
create
a
closed
doors
as
own
needs,
so
information
even
more
than
type
right.
We
need
to
do
the
null.
We
need
to
know
what
kind
of
data
we
will
get
and
what
to
do
is
that
so
look.
There
is
another
forum
to
look
at
your
operation.
Likes
example,
add
a
manna
type,
so
well,
I
think
it's
all
sort
of
you
know
obvious
I
think
that
when
we
have
a
mode
we
need
to
have
full
information
like
what
to
do.
We
need
to
do
vation,
vation
and
everything
used
to
do
there.
B
Whether
discussion
is
whether
sort
of
what
how
much
information
we
still
should
keep
update,
yeah
and
just
to
remind
like
ancient
operations
and
the
parameters
that
represented
us
knows
well.
Nodes
have
of
calls
and
the
describe
that
actual
duration
and
and
the
type
is
either
like.
You
know
it's
either
in
the
opcode
itself
recorded
or
as
like.
We
have
right
now
the
status,
what
courses
represented
first
type
derive
from
children,
but
we
quickly
reach
after
that
idea
of
driving
type
children.
B
C
C
B
Don't
page
5,
so
what
exactly
were
discussing
so
where
to
keep
information?
So
we
argued
that
beta
should
only
they
only
need
to
what
you
need
to
know
about
data
no,
usually
just
size
and
what
the
register
to
be
6.04
accomplish
an
operation
of
course,
as
I
mentioned,
should
have
everything
that
operate
the
size
will
give
complete
information
so
that
sort
of
what
we
and
winning
tool
and
then
the
only
question
is
that
was
relating
or
not
only
but
they're.
One
of
the
question
was
how
to
represent
that
information
program
below
of
Onan.
B
Don't
like
all
that
obvious
operation
information
on
amount
right,
they
could
just
flatten
it
like
who's
respect,
the
product
of
all
those
combinations,
or
we
can
split
it
into
two
like
home,
with
a
pair
like
opcode
and
David,
to
reduce
sort
of
number.
We
should
either
so
the
flattening
have
it
always
to
things
like
operation
and
type
electric
ample
in
the
LVM.
It's
both
houses
like
there
is
a
ad
and
then
it
has
like,
because
we've
go
ahead
and
it's
at
a
collector.
B
If
they
need
to
entertain
the
conversion
of
codes,
then
a
film
all
and
it's
like
all
types,
both
type
X
square,
but
then
now
sort
of
investigated.
More,
like
conversion
of
codecs,
of
my
last
note
on
that
which
five
actually
maybe
not
an
issue,
because
right
now
we
don't
create
like
order.
Cindy,
for
example,
doesn't
create
conversions,
so
that
makes
it
a
bit
better.
Fewer
few.
Of
course,.
B
A
B
That's
a
site
of
Java
below
also
the,
for
example,
insightiq
languages
that
isn't
took
happens
all
the
time
like
it.
Definitely
a
conversion
is
costing
like
we
were
there,
the
Baxter
over
the
number
for
integers.
We
want
to
look
at
it
as
a
vector
of
laws,
so
they
don't
do
anything
they
just
it's
just
a
cost,
but
you
need
to
know
code
for
that.
Okay,.
E
E
E
B
Just
gonna
do
yeah
I,
guess
I
was
saying
yeah
you're
right
you're,
the
other
case.
We
can
believe
the
rest
conversion
representative
in
the
white
coat
and
they
want
to
vectorize
a
loop
which
I
said:
don't
do
it
right?
What
syndrome
you're
right
here?
The
gazelle
think
we
don't
need
sort
of
those
casting
operations
languages
on
yet.
B
B
So
improper
noise
to
flattens,
of
example,
will
have
that
everything
according
up
codec,
so
it's
like
I
standing
for
integer
that
128
bead,
and
so
that
would
mean
now
heading
to
scalar.
Integer
is
not
vector
character,
then
I
want
to
take
eight,
for
it
means
is
Alec
underscore
for
it
in
the
scale
factor,
Sydney's
vector.
B
B
B
F
Let's
talk
about
a
bitin
of
the
worst-case
scenario,
so
you
can
have
a
vector
register
of
size,
128
between
16
bytes
right.
We
have
one
two,
four,
six,
eight,
sixteen
four
five.
Then
you
have
inter
inter
flow
that's
times
two
and
then
times
the
number
of
opcodes,
which
could
be
20,
so
gonna
introduce
100,
different
notes.
This.
B
B
B
B
Another
thing
like
Rome:
after
long
discussion,
we
concluded
that
boxer
type
is
similar
to
the
maggrip,
a
gradual
aggregate
type.
We
have
like
physical
TR
act
contrite
or
we
have
a
pointer
type,
agile
like
which
doesn't
tell
us
exactly
what
exactly
post
or
for
aggregate
it
does
this
kind
of
mail.
So
why
would
be
in
a
type
in
care
type?
Why
would
we
precisely
try
to
precisely
describe
with
this
layout?
B
We
usually
need
to
know
size
and
whether
it
belongs
to
exploit
like
medium
integer
fix
point
of
floating-point
Ridge.
So
so,
because
of
that,
we
decided
to
have
type
sex
about
into
128
or
FP
128,
so,
for
example,
into
128
will
represent,
be
sort
of
the
don't
care.
What
is
inside,
it
can
be
just
one
integer
of
128
bit
integer
or
we
can
have
two
longs
for
into
the
16
lives.
B
B
So
going
back
to
the
discussion
on
data
and
operation,
so
data
will
have
all
type
represented,
but
not
all
information
kind
of
describing
the
data,
but
not
in
the
type
of
button
symbol.
It
will
display,
like
part
of
information
and
type
weather
like
each
ones,
including
even
22
before
and
but
whether
it's
elect
or
not,
will
be
for,
like
a
modify,
unstable
insights.
The
world
and
again
argument
is
that
it's
not
that
also
that
we
need
to
know
whether
it's
absolutely
we
don't
need
no
like
visible
register
location,
but
otherwise
we
usually
turn
without
effect.
B
B
B
B
A
B
D
F
B
F
B
B
A
If
this
were
to
go
ahead
in
terms
of
actually
how
you
would
go
about
thinking,
this
change,
I
mean.
Presumably
you
could
add
new
op
codes
to
the
OP
code
table,
not
a
problem,
but
in
terms
of
actually
making
all
the
changes
to
the
optimizer
aisle
gen
code
generators.
That
kind
of
thing
is
that
something
that
has
to
happen
in
what
Big
Bang
or
is
that
something
that
can
be
staged
at
all?
C
B
B
Hotel
change,
of
course,
change.
I
call
evaluators,
can
reshuffle
existing
ones
input
agenda.
We
are
not
try
to
add
all
the
size
of
this.
For
my
content,
hypocritical
now,
yes,
just
transform
what
we
have
now
do
you
know
code,
which
essentially
would
mean
molecular,
meaning
will
not
be
even
reading
new
one
animal
course
evading
is
infeasible
the
dread
94b.
Why.
B
B
It's
electable,
sometimes
they
ask
you
whatever,
like,
for
example,
Jerry
register
0.
It
has
consumed
approximately
certain
physical
impact,
so
right
now
Africa's
it
back
to
type,
but
not
many
places
then
go
to,
but
that's
the
whole
set
of
ideas
going
that
usually
like
like
when
we
transform
that
type
information
to
operations
we
cook
the
aisle
like
Walker,
generate.
Of
course,
awesome.
Dakotas
indicates
that's
all
code,
it's
great
from
exist
right,
like
we
have
over
half
like
I,
add
and
we
create
vector
head.
So
at
that
point
we
will
now
available
to
create.
B
E
Doesn't
that
sorry,
would
it
make
sense
as
part
of
phasing
and
then
to
try
to
hide
getting
the
type
stride
and
vectors
of
it
hide
those
behind
a
query
that
has
access
to
all
of
the
places
where
we
might
put
this
with
it?
We
have
described
the
proposal
that
we
think
we've
settled
on
as
a
reasonable
trade
off
point,
but
this
is
the
second
such
proposal
for
how
to
do
this.
So
I
guess
I'm
wondering
if
maybe
now
is
the
time
to
introduce
an
API.
E
E
B
E
E
B
B
E
B
E
F
B
B
B
B
F
E
You're
never
going
to
get
an
answer
coding
them
just
adding
them
soft,
oh
yeah,
I!
Guess
the
main
question
is:
what's
the
cult?
What's
the
bad
consequence
of
adding
a
large
chunk
of
opposite
of
the
optical
table
and
I?
Think
the
the
biggest
downside
which
I
don't
personally
consider
to
be
a
massive
downside,
is
the
number
of
static
tables
that
are
XY
off
code
will
have
to
grow
by
and
a
robot
code
that
you've
added
but
I
think
we
looked
in
that
quickly.
E
B
B
F
Think
for
the
implementation,
we
can
create
a
pass
at
some
point,
maybe
even
after
optimizer,
which
transforms
the
new
into
the
old
or
vice
versa.
Temporarily,
we
got.
B
F
B
F
F
E
B
F
E
A
Of
the
problems
we
had
with
the
vector
tests
was
that
it
was
running
on
some
machine
to
support
service,
there's
no
way
in
the
testing
framework
right
now
to
determine
what
the
capabilities
of
the
target
that
you're
running
on
so
just
assumes.
You
know.
I1Z,
therefore,
I
can
use
everything
that
I
know
about
Z,
where
whether
that
target
machine
supports
it
or
not.
So
that's
something
we
have
to
fix
in
the
testing
framework
includes
wider
testing
of
vector
stuff
right.
E
E
G
F
G
Up
the
bit,
it
might
be
a
might
not
be
the
hardware
itself,
but
the
hypervisor
layer
right.
Okay,
just
you
should
all
go
check.
Okay,.
E
A
Was
open,
yes,
so
I
think
you're.
Maybe
the
next
step
here
is
to
actually
create
an
issue
in
the
Omar
that
basically
describes
their
everything
that
you
have
here,
and
you
know
that
that
may
invite
more
discussion
as
well,
but
I
think
that's
been,
maybe
ultimately
the
precursor
for
a
pull
request
or
a
series
of
pull
requests
that
start
to
make
whatever
changes
you
want
to
make
your
but
I'm
just
getting
all
this
content
out
into
an
issue
so
that
it's
it's
there.
G
That
question
is
going
back
to
the
OP
code,
name
itself,
you're,
looking
at
it
that,
despite
the
I
128
at
right
now,
I
mean
it
is
that's
the
point
where
you
know
there's
a
little
bit
of
confusion
in
terms
of
how
we
interpret
I
I
mean
I.
Think
in
this
context
you
mentioned
it
was
referencing.
Is
it
being
operations?
G
Sir
I
guess
my
question
here
is
you
know
instead
of
this
model
would
would
it
be
more
intuitive?
If,
if
we
went,
you
know
stuck
with
our
existing
types?
Let's
say
for
the
second
example
where
yeah
I
128
underscore
for
right.
You
know
that's
adding
for
32-bit
integers
right,
and
so
you
know,
instead
of
actually
explicitly
encoding,
128
directly
end
up
called
itself.
We
have
an
I,
add
under
4,
for
and
and
that
I
will
be
treated
as
32-bit
integer
like
the
traditional
way.
We
interpret
eyes.
G
G
How
would
I
was
my
hundred
miles
I
think
here
by
the
think
there
would
be.
You
know,
you
know
with
the
two
examples.
The
last
two
examples
right
with
underscore
for
an
under
or
to
an
alternate
wave
with
the
first
one
I
128
at
underscore
for
would
be
I,
add
underscore
for
and
then
the
last
one,
whereas
underscore
two
OPA
l
@
underscore
yeah.
B
A
F
B
B
B
B
B
B
B
E
C
A
E
F
B
E
B
B
F
B
B
B
B
E
F
B
E
E
E
B
E
E
D
B
E
B
Logical
sort
of
conceptual
it's
officially
too
hopeful
from
basic
type
and,
to
some
degree
this
what
the
data,
just
because
certain
my
lipslide
kind
a
little
bit
to
to
some
degree
redundantly
like
for
service
form,
yeah
right
now
say
not
all
combinations
are
supported
like
we
only
is
the
fourth
basic
Texas.
It
will
never
be
like,
for
example,
I
don't
know,
250
speaks
underscore,
they
don't
have
complete.
B
A
A
Okay,
so,
like
she
said,
she's
going
to
be
writing
up
these
design
idea
into
into
an
issue
and
I
think
we
can
carry
on
some
discussion
there.
So
thanks
gia
and
thanks
everyone
for
for
attending
and
we'll
see
you
again
next
time,
thanks
guys.