In this course, you will learn to design the computer architecture of complex modern microprocessors.
Register Renaming with Pointers to IQ and ROB
Loading...
來自 Princeton University 的課程
Computer Architecture
267 個評分
從本節課中
Superscalar 4
This lecture covers the common methods used to improve the performance of out-of-order processors including register renaming and memory disambiguation.
與講師見面
David WentzlaffAssociate Professor
Electrical Engineering
Okay, so let's take a look at what we have to add to our pipelines.
So we have our in order fetch, out of order issue, out of order write back and
in order commit, plus that we had before. Note, it had variable length pipes.
It had a reorder buffer. It had a feature storer buffer.
It had a scoreboard, it had an instruction cue.
So it had all the, the structures we talked about last time.
And now we're going to add two more structures to it.
And were gonna modify the structures that are there slightly.
Now let's, let's talk about what this is gonna do.
So the first structure we're gonna add. Is a free list.
And the free list is gonna keep track of physical registers that we could go use.
So the physical registers will probably have more physical registers than we have
architectural registers. But you need to keep track of which ones
are free to be used because we are gonna basically be allocating deallocating from
the number of physical registers quickly while as we execute.
The other structure here, we call the rename table.
Sometimes this is called the rat. Which is the, sort of the intel
nomenclature for this; or actually the rat is either this table or the table we were
discussing in the Tomasulo algorithm variance of this but they're very similar.
And what this table does is it's going to map from architectural register to the
most up to date version in our physical register file.
So, it's gonna say, with instruction that's sitting here, at our decode stage,
where do we go find the value, cuz this gets complicated.
We're going to, we just renamed everything.
We have different names for everything. It's in some physical registers.
We need to go figure out where the value is.
And that's what this table, table does. We're also going to add two fields to your
buffer. I'll talk about that in a second, and
we're going to want to increase the size of the physical register file, so that we
can get more performance. If we have the same number of physical
registers as we have architectural registers, and we need to have at least
one physical register for each architectural register, we're not going to
get anymore performance from having a register renaming step to our pipe.
Okay so this is kind of for completeness where everything gets written in the pipe
in time. Two things I wanted to point out here, are
the free list gets updated at the front and also gets updated here at the end, and
the condition that you need to de-allocate a physical register, or a physical
register gets a little complicated, and, or we'll talk about that.
And the rename table, it gets red up here because that tells you that actually where
you get the value. It also gets updated here when we actually
emit an instruction down the pipe. And we also want to update some pending
bits when we get to the end of the pipe. So that it knows whether to go sort of
pickup from the physical register file or the architecture register file for roll
back issues. Okay, so let's jump into these data
structures and see what we add. Okay as I said we, we we're gonna add
stuff to the reorder buffer. Here, now our previous reorder buffer
looked very similar to this. We had some state where things were
pending, free, or finished where we said dash, dash represents free and f means
finished. It means the instruction got to the end of
the pipe and is waiting to commit. We got a bit that says, well it was after
a branch. You might have multiple of these branches
if you allow multiple branches in flight. A bit that says a store not.
A bit that says whether, it writes a register.
So this says the destination is valid and that's important for us to know because we
meet at the end of the pipe we need to know whether to actually commit some state
into the architecture register file. And we have a, a field here, which we had
before, which is the physical register file specifier.
So this tells us where to go read from. That's, that's all, that's all good.
But now we add some extra, extra bits here.
And the first one is a architectural register file specifier.
Okay, so this gets a little complicated. What, what are we thinking about here?
Why do we need this? When we get to the end of the pipe and we
are going to do commit, if we go back and look at this picture here, in the commit
stage, we take something from the physical register file, put it into and the rear
buffer drives this and says, okay copy that into the architecture register file,
when the commit occurs. Well now we've renamed everything.
So it's not an identity map from physical register number to architectural register
number. So we needed to know where you could
actually write in the architectural register file.
And that's what this does. It just tells us where to go write.
So, this is where we read from, this is where we write to.
When this instruction, let's say it's the most recent instruction here that's turned
to finish. It's going to go commit.
We read the value from here. We write it into the value pointed to by
here. And then we have one other field here,
which is a little bit odd. We have the previous physical register.
Why, why would we need that? That doesn't make any sense.
It is the, what is, what is this doing? So this is something we actually read out
of the rename table at the front of the pipe.
And what it's going to tell us, is it's going to tell us.
Let's say this was register four. This is where we, the in-flight physical
register is, and this is the previous physical register that h-, held the value
of register four before we did the update. And the reason we need to know this, is
when we hit the end of the pipe, we need some way to de-allocate physical
registers, and we're going to use this to track that, and we'll give an example in a
minute. But what this is really going to do its
going to say; oh we wrote to the new value of register four, which means that the in
flight value, let's say it was register four, physical register 27.
And the new one is physical register 30 or something like that, need to deallocate
physical register 27 and we can do that when we reach the end of the pipe by
committing this instruction out of the reorder buffer and cleaning up all the
state. Okay.
A, a quick picture here of the, the rename table, the renaming table.
This is indexed by register. P tells us whether we have a write in
flight. So it knows that, that value is not in the
architectural register file. And p regulator tells us where in the
physical register file to go find the value.
And this is really important when a subsequent instruction is looking for that
value, shows up, and it wants to get that value before it hits the architecture
register file. It looks here, this tells us, tells us, oh
its pending, its gonna be here in a little bit.
Together with this and the scoreboard, you might even be able to bypass it early.
In, in, a, in a good day. In a bad day we have to wait for it to get
to the physical register file, but it's a lot better than having to go pick it out
of the architectural register file. And finally we have a free list.
And this is literally just a bit per physical register, which is very different
than a bit per architectural register. And this is going to have, let's say we
have big N physical registers. Or rather we have 256 physical registers,
and we have a bit saying whether that register has been deallocated and is ready
to be used, for future register renaming. Or whether the instruction, or whether
there's a instruction using that physical register, or it's waiting to commit to the
architectural register file, this will tell us that information in this table
here, and just a bit that says whether it's free or not, pretty simple.
Actually, before I go on here, I wanted to make a, make an interesting observation.
Where, where does this register renaming become really important?
Well, if we go look at something like the original Intel architecture, they had
eight registers. If you want to run high performance code,
you have to re-use those registers pretty quickly.
So they got register limited very quickly when they tried to build faster and faster
processors. So they had to introduce registry naming
quite early in the Intel imp-, micro-architecture implementations.
And they had many, many more physical registers than architectural registers
very quickly, cuz eight isn't gonna get you very far.
They, they can have, like about 100 in-flight instructions.
And, by definition you can't have that many inflight instructions, if you,
maintain right after write stalling, effectively cuz, you're, you're going to
have to rewrite some register. It's kind of like a, a, pigeon-hole
problem. If you have more than eight instructions,
at least one of those instructions is gonna cause a right after right,
dependency and you're gonna stall the pipe.
So, they're, they're not going to have more than say eight inflight instructions
pretty quickly, if they did not do register renaming.
So, they did register renaming pretty quickly, in their pipelines.
Okay, so this gets us to the, the I chart here.
Let's walk through, basic case, of what's in all these different tables as we
execute our basic, simple code here. On the top we have the four instructions,
two muls, and some adds. This was our original test case.
Note there is all these dependencies through register four we need to worry
about. There's both where you have to write,
write after read and write after write dependencies.
We're gonna execute it quickly here by pulling, as you can see, this add fires
early, or issues early the, the final add. And this is really driven by the
registering ini. So let's, let's take a look at what, what
happens here. We'll try to interpret this.
Here we have cycles. Cycles are also across the top of the
stage. We, we show what's in the decode issue,
write back and commit stage of the pipe. We leave out the execute stages cuz it's
too much to draw here and it's drawn at the top in a different form.
Let's first look at the renamed table. So.
We're at the rename table. Or, we're actually gonna say we only have,
for, for clueless here, let's say we only have seven architectural registers.
But we're going to have let's say. Ten physi, or, eleven physical registers.
So we're gonna have more physical registers than actual registers in this
example. We start off and we say, Okay, well,
register one. If you want to go find architectural
register one, the values in physical register zero.
And we could basically just, you know, we just come up with some allocation.
And the circles mean that it's not pending.
It's not in flight in the pipe. That's just sort of the base case.
Everything is, the, the pipe has been relaxed.
Everything is, is allocated, and we just drew a basic allocation here at the
beginning. Now as we go to execute, some interesting
stuff starts to happen. The first thing that is going to happen
is, we are actually going to, here, issue this instruction, which writes to register
one. We need to rename this, at this point.
Register one will have to be named as something else.
So in this table here, if we look, we register, we rename register one to
physical register seven. Okay, that sounds good.
What happens next. Well we next sit here, and we try to
execute this instruction here, It says mall, and it goes to try to read register
one. When it goes to read register one though,
we can go look at the rename table and say, oh well that's actually in flight,
and it's in physical register seven. So if we go look over here, we can draw
this and say, oh that value is actually in physical register seven, and it's
currently not ready, maybe, and, But P4, the other input, register five, to do,
okay, yeah, register five got renamed to P4, is ready.
So it's ready to go. Okay, let's, one of the other interesting
things that happens here is we can see that as we go to allocate this, we have to
remove it from the free list. So this list here is the list of all the
free registers. We start off with four free registers and
we sort of narrow it down as we start to do rights.
At some point we run out. So I want to make an important note about
this is that when we run out of physical registers we're going to have to stall the
pipe, because we can't do any more renaming.
We can't issue more instructions at that point.
So that's, that's really, that's really important to realize that when you build
your machines you have to have enough physical registers that you don't run out
very often. Now, it's possible that you could still
run out. So let's say you have hundreds of in
flight instructions. And you only have, let's say, 64 physical
registers. You might still run out, But the
probability of that happening, my, might be relatively low.
And the, your utilization, and, you know, you sort of bake into this your CPI.
Your CPI may not be less than one, or may not be low.
So, you know, the probability of that actually happening.
You may not worry about it too much. Another cute little story here is there's
actually been some interesting bugs in processors around the free list.
So there were some alpha processors that actually leaked free list entries in their
register file. So what happened was if you ran a certain
piece of code for a long enough period of time all of a sudden this processor just
ground to a halt cause it was not able to allocate more physical registers and it
ran out. And ends up with fewer physical registers,
architectural registers, and the machine just stopped.
And this was a, sort of well-known bug in, in some of the early Alpha, I think this
was actually in the first, out of, I want to think where was this, I think this was
in the, 21264 had this problem. They, they fixed it.
And, they pulled those chips off the shelf.
And, you know, that's a, that's a really bad thing to, have happen, in your
processor. How embarrassing.
But as I said, if you run out, you're really not going to be able to issue more.
But in this case we made sure we had enough.
So we're not actually going to see any stalls.
And let's look at how things get on the free list here.
Cause that's a little bit interesting. In our reorder buffer, I said we had extra
fields. If you recall, we had the previous
physical register that this was allocated into.
So, if we go look, at this instruction, which is the first instruction, go to
execute that mull, R1 was in P0. So when that instruction commits, we
actually put p0 onto the free list. And we're going to look at a case in a
second why you can't do it earlier. Because it seems like you should be able
to basically de-allocate physical registers earlier.
You know, no one's probably gonna be reading that value.
Why can't you just, you know, get rid of it early?
But we'll look at in a second that a test case that, that, that's, that's a problem
with. Let's see any, any other fun insights
here? That's, that's about it, what I wanted to
get across from, from this diagram. As, as the code continues on we end up
with more and more free physical registers.
One thing one thing I did wanna just to walk through this, understand this a
little bit. Let's say we have this instruction here
which is our one, two, three, four. It's our last instruction that we execute.
Let's go see what it's doing here. So writes architecture register four, so
let's just store that in the reorder buffer cuz we don't know where to go to
the right. We had allocated p10 to that, and we did
that right here, when we actually issued it.
So he pulled it off the free list. And the previous thing that it wrote was
P8, so when in, that ultimately commits, P8 is gonna end up back in our free list.
So that's a, that's a nice little thing, the circles just show when the values are
no longer pending, so they are actually not in the pipes anymore.
And you can see that continuing here, this instruction here, which is the second
multiply. When it commits it's going to free up P3.
So P3 ends up on the list. P5, P5 ends up on the list.
P8, P8 ends up on the list. And then if when we see true read after
writes, for example right here we need to make sure to pick up that correct value.
We do that by looking up in the rename table.
So let's go find that in this chart here. So instruction two.
Is let's see what it's doing here. So, that's gonna be right here.
It's waiting on the eighth to be become ready, in order to issue.
So it's signaling an instruction queue, and this is gonna stall.
It's gonna stall all the way out to here, or to stall to right there.
And that's when it comes out of the instruction.
Okay. So let's look at freeing up physical
registers, and what is a good policy for freeing up physical registers.
So we're gonna have a different piece of code here we're going to look at.
It's gonna be just a bunch of ads. And we're gonna look at.
This code has some. Read after write dependency is in it.
Namely R1 there. And, let's say, we try to go execute this.
Well we, we're gonna, here's some execution order.
We're gonna look at, oh sorry, that one there so, I meant point out for the read
after write. A write after read dependency here.
So let's look at some execution order and see what happens, let's say we allocate
physical register zero, at the beginning somewhere for register one.
And then when we do the commit, we free it up in our, free list.
Well, lo and behold, another instruction, in time, comes along here, and allocates
in the physical register zero. And it goes and writes to it.
And we, like, free it up there. This instruction here which we'd renamed
and earlier we had renamed our one for and we go to try to read this value, goes to,
do the read and it looks in physical register zero.
And I guessed the wrong value. Ooh.
Yeah, we don't, we don't want that. So what's a, what's a good policy here?
Let's say instead, we don't, free up a physical register until someone else goes
to write that physical register. Or our subsequent instruction goes to
write that physical register. Because then we know that, that physical
register is in use, or could be in use by other readers of that value.
So if we look at this case here, let's say we, write, physical register zero.
And then we allocate a different physical register.
Right? We allocate physical register two for this
right here. Register eight.
And then we de-allocate when we go to overwrite register one.
So by doing that, we know when this R1 gets written, that no one else can
possibly use that physical register, that is after this instruction in program word,
because we overwrote it, so the value is no longer visible.
So that's, that's pretty, pretty nice. So that's, that's the a, a, a very good
heuristic or very good way to get this correct; cuz you could just keep the
physical register live until you rewrite the physical, or your rewrite the
architectural register that physical register maps to.
And at that point you can remove it from the number of allocated physical
registers, and put in on the free list. If you do it early, with the out of order
execution pipeline, you know, bad, bad things can happen.
You can go read the wrong values. Okay, so this brings us to a couple
optimizations on register renaming. The biggest one here is you can try to
combine the architectural register file and the physical register file to save
space, and the insight here is, if you go to try to combine these two things you can
store the architectural register value and the physical register value in the same
physical storage location. If that physical register's no longer
pending. So if there's nothing in flight to it and
you don't have to roll back, if you're just going to roll back to the same value
anyway, why, why keep extra space for this?
One, one change you need to do here is, so you're going to remove the architectural
register file. Which you still basically need to know
when you go to do a rollback of some speculative, say you take an interrupt, or
you take a branch miss-predict, you still need to know where to go rollback out of
and we're gonna do that by let's say having a second renaming table here, which
allows us to keep track of just the architectural state.
So we have a speculative renaming table, then we have an architectural renaming
table. It just has pointers in it, instead of
actual values, at the end of the pipe. And what's also nice here, is instead of
copying values, we don't actually have to move something out of the physical
register file into the architectural register file.
Instead we just have to update a pointer in a table now.
And we did the copy, to potentially also make rollback easier, cause we have to up
date pointers now instead of actually copying an enter register file, which can
take awhile or requires lots of ports or something else.
So you, you can have a little table there to do this remapping for you.
And as I say you can typically get away with less space than having for the same
performance than if you were to having two separate structures.
When it downsizes, you, you might need to have more Depending on how you implement
this. You're architectural register file, and
your physical register file are now together.
It may be bigger. So your registered default access might be
a little slower. Something like that could be, could be a
down side, versus having it in two separate partition structures.
