In this course, you will learn to design the computer architecture of complex modern microprocessors.
Register Renaming Introduction
Loading...
来自 Princeton University 的课程
Computer Architecture
267 个评分
从本节课中
Superscalar 4
This lecture covers the common methods used to improve the performance of out-of-order processors including register renaming and memory disambiguation.
与讲师见面
David WentzlaffAssociate Professor
Electrical Engineering
Okay. So that brings up the, the question of
register renaming. What is register renaming?
Hopefully some of you skimmed the Thomas UIlo algorithm paper that I assigned, cuz
we're going to be discussing that and the motivation for that work.
Okay. So, what is learning our performance in
these pipelines, these out of order pipelines that we've discussed so far?
Okay. A couple things.
Write after write and write after read dependencies.
Let's, let's talk about these. In write after write dependence, you're
going to write to one register, then you're going to write to the register
again, and in our pipelines we've talked about so far, you're basically gonna stall
the pipe while you're waiting for the first right to commit.
Because we're not able to handle multiple rights in, in the pipeline at the same
time. But these are not fundamental
dependencies. So computer architects put on our thinking
cap and we came up with ways to break these dependencies.
A read after write, that, that also is true for a write after a read.
So, if you do write to let's says register four after a read from register four,
well, there should be nothing wrong with that.
But, if you try to execute the instructions out of order, then you need
to think about that. A read after write is a true dependence,
because you actually need the data, you need the value to go execute the
subsequent instruction. And we're going to call write after read,
write after write and write after read dependencies name dependencies, and we're
going to call a read after write a true dependence that we can't break.
Okay. So let's, let's look at a some example
code here, and see what can go wrong if you just ignore all the name dependencies.
Like I said, they're not true dependencies, so maybe we just don't need
them. Okay, so we have, a different code
sequence here, we have a mul, mul, then two add immediates.
Let's, let's identify some important things in here.
First let's identify the true read after write dependencies.
So I, I put some circles and some arrows here.
So let's take a look here. This writes register one in the mul.
The second mall reads the results of that. Okay?
Let's see here, this add reads register four in the previous instruction and
writes register four, so that's a true dependence.
We can't break those. We make talk at the end of class at the
end of the term, maybe about some ways to break those, but they get pretty, pretty
crazy. So, let's, let's look at the, write after
write dependence. So the first write after write dependence
is here. We're writing register four, and then we
write register four again, but an out-of-order processor, if we try to break
all these dependencies, we can see that we're actually writing to register four
here, and then we write register four here.
And that's, like, out of order. Whoa, what just happened here?
Well, we said we just broke the dependency.
We're not gonna stall the front of the pipe on this.
So if you to execute this on one of our out of order issue pipes that we've,
looked at so far. Let's say this is executing on the in
order fetch, out of order issue, out of order execute, and, right back in out, in
order commit pipe. We can see that we will write to register
four in time before we wrote to register four here.
Now that's going to cause some major problems.
We just wrote the wrong value. Oops.
And the other one here is a write after read dependence.
So here we have a read of register four. And here we have a write of register four.
And because this add got pulled so early in the, in the execution order, we
actually wrote before this instruction had a chance to read the value.
And the reason this instruction got delayed was cuz it was also dependent on a
true dependency here. But it's dependent on two things.
But all of a sudden we wrote register four with the value from this add, and then we
went and read it, and we read the wrong value.
So, we can't just go change and break right after write dependencies and right
after read dependencies very easily. We need to think a little bit harder about
this. One, one last interesting thing that
happens here. This is, this is kind of fun.
We do commit in order in this pipe. But look what happens to register four.
We wrote register four, here, then we write register four again.
And then we commit from physical register four to architectural register four.
So we just committed the wrong state, also, to the architectural register file.
So we're having lots of, lots of problems with here.
It's not just, basic things. So what's the solution to this?
Well solution, as we can start thinking about how to add more registers, so at the
top here is the same example I had in the previous slide so nothing, nothing new
there but I wanted to compare this to, let's say our conservatively stalling
pipeline from performance perspective first.
So here we have our in order, our in order fetch out of order issue out of order
write back and in order commit pipe. This is our most advanced one from last
time. But, it conservatively stalls on write
after write and write after read dependencies, and that's drawn here of
these arrows. So we can't even issue this instruction
until we know that, let's say these two instructions here, which have something to
do with register four, commit. And then we can go to issue this.
Now this might be a little bit conservative.
This might be even a little bit over conservative.
It might be possible to pull this back one or two cycles, maybe to, sort of to the
point where this instructions does the write back.
But one of the challenges there is, you don't necessarily, you can't necessarily
track that very easily inside of your reorder buffer, unless you have something
there that scans for this exact case. It's not going to save you that much
performance either. What, what I'm trying to get across here
though is that the performance of this instruction sequence is actually worse.
It takes longer than our incorrect, but we'll call ideal case here on top.
So let's, let's do one little change to the instruction sequence, highlighted in
red here, and see what happens to our execution.
So we took this ad which wrote to register for and changed it.
We added another register usage here and now we write to register eight.
And lo and behold, it breaks all the write after write dependencies, and it breaks
all the write after read dependencies. And all of a sudden, we get to, our
idealized performance, this, exact, this is the exact same case as that.
But we require another register. Hm.
Well, please add infinite numbers of registers.
So, what is, what's, what's the con of adding infinite numbers of registers?
Anyone have any ideas? So, if we're trying to use more registers
here we might use up all of our architecture registers.
Can we just add more architecture registers to our instruction set?
Yeah so, so it takes up space. So if we look at our registers, we could
have a larger name space for our registers, but if we have 32 registers,
that takes five bits. If we have 128 registers, that's gonna
take, you know, seven bits. And if we have infinite, that will take
infinite number of bits. But what we're gonna talk about in today's
lecture is how to do this in hardware such that you have some more registers in your
physical register space, but not more registers in your architecture register
space. And this is, I should point out this is
not only a register problem. This could also happen with memory.
If you name your memory inappropriately if you have a very small amounts of memory
and you try to reuse it very aggressively, you can get name, naming problems in that
also. But for today we're going to mostly be
focusing on registering. And so I'll define registry name as, well
you're gonna change the naming of the registers in hardware to eliminate these
write after write name dependencies and the write after read dependencies.
Okay, so we're gonna be talking about two major schemes, and there, they are mirrors
of each other and have slightly different, hardware requirements.
What their lot, mostly when you think of them, they're just sort of logically duals
of each other and there's different ways of thinking of the same problem.
So the first scheme we're gonna be looking at is we are going to add pointers at, in
our instruction queue and reorder buffer. To allow us to have different register
names in them and not actually just have architectural register names in those data
structures. The other option, which is, if you, if
anyone actually read the Thomas Ullo algorithm paper, is to actually store the
actual value, the data value, in these data structures.
In the reorder buffer and in the instruction queue.
And, and they look, they look very similar, if you sort of think about it,
and they're gonna have the same performance.
I'll, I'll, I'll give you the sort of end of the novel at the beginning here.
They're gonna have the same performance, they're, they're, they're doing the same,
doing the same thing but they're, they're slightly different in the mechanics
perspective. And, and we are gonna start off by looking
at this first one here, we are gonna have pointers in the instruction queue in the
reorder buffer, Mainly, because we already have pointers in our design, that we
looked at the last time, this in order issue, out of order, or similarly, in
order fetch, out of order issue, out of order executing right back and in order
commit design.
