Hi James,
I also think that refactoring the code generation part is a great idea. That
code is very complicated and difficult to maintain. I’ve wanted to rewrite that
code for a long time, but just have never got to it. There are quite a few edge
cases to handle (at least in the current code). I’ll take a deeper look at your
patch. The abstractions that you mention, Stage and Block, are good ones. I
think you’ll need to account for the cycle, or position within a Stage as well.
It is a complex problem with a lot different edge cases when dealing with Phis,
though I think they can be dealt with much better than the existing code.
Generating code for the 3 main parts, the prolog, kernel, and epilog, can be
challenging because each has a unique difference that made is hard to generalize
the code. For example, the prologs will not have an Phis, but the code
generation needs to be aware of when the Phis occur in the sequence in order to
get the correct remapped name. In the kernel, values may come from a previous
iteration of the kernel. In the epilog, values can come from the same epilog
block, the kernel, or a prolog block.
It would be great to be able to generate the compact representation that you
describe, so that a prolog/epilog do not need to be generated. Hexagon has a
special hardware loop instruction, spNloop, specifically for this. Although, the
compiler does not generate that instruction currently. Perhaps it would be
easier to generate this form first, and then convert it to a prolog(s), kernel,
epilog(s) sequence.
There are a lot of subtle rules when dealing with the Phi instructions depending
on when they are scheduled relative to uses. There are two different factors to
account for – the stage and cycle when the Phi is scheduled. For example, a use
of a Phi can be scheduled at an earlier cycle, but later stage. Several
different combinations can occur, and each use of the Phi can have a different
stage/cycle. Maintaining the map between new and old virtual registers depends
on the stage, cycle, and the current block (prolog, kernel, or epilog).
The epilogs are generated in reverse order due to the way the instructions need
to be laid out. The first epilog block after the kernel is going to have
instructions scheduled in the last stage only, then the next epilog block, if
needed, will have instructions from LastStage-1 first, then the LastStage.
Let’s say we have instructions x, y, z, that are scheduled in different
iterations/stage, x in stage0, y in stage1, and z in stage2, and x generates a
value used by y, which generates a value used by z. The generated code will look
like:
Prolog 0 (this block ends with a conditional jump to Epilog 0):
X
Prolog 1 (this block ends with a conditional jump to Epilog 1):
Y
X
Kernel:
Z, Y, X
Epilog 0:
Z
Epilog 1:
Y
Z
The instructions in the prolog and epilog are generate in original program order
within each stage, but in the last epilog stage, the instruction Y generates a
value used by the subsequent Z.
The reduceLoopCount function is needed for an architecture that has a hardware
loop instruction like Hexagon’s LOOP0. For each prolog generated, the loop
counter in the instruction needs to be decremented. This could have been done
only in the last prolog block, but it’s done every time a prolog block is
generated instead (initially, I thought that would be easier since the prolog
generation code wouldn’t need to know which block was the last one). But,
later, I added code to handle the case when the loop can be removed completely.
It seems reasonable to change this so that it’s updated only once, especially
since find the loop instruction can require searching through the CFG.
Loop carried Phis are hard. Initially, I tried handling the cases for adding new
Phis and existing Phis with one function. But, it turned out that adding new
Phis, when a instruction use is generated in an earlier cycle, but later stage,
doesn’t cause as many special cases as handling the different edge cases for an
existing Phis. There are less combinations when adding a new Phis.
Yes, you’re correct about the use of “iteration” vs. “stage” (I even used them
together earlier). The use of term stage comes from the paper, but I often use
iteration when talking to others about software pipelining, especially if they
have read the paper. I agree that we should be more consistent with using the
terms.
Let me think some more about the questions you’ve asked. I do think it’s
worthwhile to spend some time on creating a good abstraction and improving the
code. I appreciate that you’ve started to think/work on this problem, so I’d
like to help out to pursue this direction.
Thanks,
Brendon
From: James Molloy <james at jamesmolloy.co.uk>
Sent: Monday, July 15, 2019 11:05 AM
To: Jinsong Ji <jji at us.ibm.com>
Cc: Brendon Cahoon <bcahoon at quicinc.com>; Hal Finkel <hfinkel at
anl.gov>; LLVM Dev <llvm-dev at lists.llvm.org>
Subject: [EXT] Re: MachinePipeliner refactoring
Hi Jingsong,
Thanks for testing out the prototype! I'm not surprised there are errors in
that version; it's not 100% ready yet (mainly a demonstrator for the path
I'd like to take). I'm really trying to work out if it makes sense to
try and incrementally get from the current state of the world to that prototype
incrementally, or if it's just so far away that a full-on refactor (or two
code paths) is required. I suspect only Brendan really has the context to give
insight there!
James
On Mon, 15 Jul 2019 at 16:32, Jinsong Ji <jji at us.ibm.com<mailto:jji at
us.ibm.com>> wrote:
Hi James:
Personally, I like the idea of refactoring and more abstraction,
But unfortunately, I don't know enough about the edges cases either.
BTW: the prototype is still causing quite some Asseertions in PowerPC - some
nodes are not generated in correct order.
Best,
Jinsong Ji (纪金松), PhD.
XL/LLVM on Power Compiler Development
E-mail: jji at us.ibm.com<mailto:jji at us.ibm.com>
[Inactive hide details for James Molloy ---07/15/2019 06:16:22 AM---Hi Brendan
(and friends of MachinePipeliner, +llvm-dev for o]James Molloy ---07/15/2019
06:16:22 AM---Hi Brendan (and friends of MachinePipeliner, +llvm-dev for
openness), Over the past week or so I've
From: James Molloy <james at jamesmolloy.co.uk<mailto:james at
jamesmolloy.co.uk>>
To: LLVM Dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at
lists.llvm.org>>, jji at us.ibm.com<mailto:jji at us.ibm.com>,
bcahoon at quicinc.com<mailto:bcahoon at quicinc.com>, Hal Finkel
<hfinkel at anl.gov<mailto:hfinkel at anl.gov>>
Date: 07/15/2019 06:16 AM
Subject: [EXTERNAL] MachinePipeliner refactoring
________________________________
Hi Brendan (and friends of MachinePipeliner, +llvm-dev for openness),
Over the past week or so I've been attempting to extend the MachinePipeliner
to support different idioms of code generation. To make this a bit more
concrete, there are two areas where the currently generated code could be
improved depending on architecture:
1) The epilog blocks peel off the final iterations in reverse order. This
means that the overall execution of loop iterations isn't in a perfectly
pipelined order. For architectures that have hardware constructs that insist on
a first-in-first-out order (queues), the currently generated code cannot be
used.
2) For architectures (like Hexagon) that have dedicated predicate register
files, we can generate a compact representation of the loop by predicating
stages of the loop kernel independently. In this case we can either have a
prolog, epilog, or neither (wrapping the prolog and epilog inside the kernel by
using PHIs of predicates).
At the moment, a lot of the code generation helper code in MachinePipeliner is
tightly fit to its current code generation strategy ("If we're in the
epilog, to this, else do this"). I'm keen to try and make some of the
complex calculations it does, such as where PHIs should come from, more abstract
so they can be reused and composed.
https://reviews.llvm.org/D64665 is my current best-effort. This generates
perfect code for PowerPC, but causes a load of problems for Hexagon. It's
become apparent that I don't know enough about some of the edge cases in the
MachinePipeliner code to refactor this from scratch. I'm therefore looking
for direction in factoring in an incremental fashion.
I think there are a couple of areas that I'd like to change, and I'd
appreciate your ideas and opinions because I clearly don't know enough about
the edge cases here.
a) TII->reduceLoopCount() is hard to understand. Understanding the intended
semantics of this hook from the documentation, I've found, is hard. Its use
appears to be strongly fit to Hexagon (there is even a comment about the removal
of LOOP0 in the MachinePipeliner target agnostic code, which really
shouldn't be there). Why it's called multiple times I don't
understand (why can't we just call it once with the total number of
iterations to peel?).
b) Understanding how loop-carried PHIs are linked together is really hard.
There are two functions dedicated to this with many edge cases, which are linked
with the prolog and epilog schedule. It'd be great to somehow factor these
such that they are independent of the code generation strategy. Note that this
is really important for some of the code gen strategies I mention at the
beginning, because loop-carried PHIs in this case may actually end up being
selects or uses of predicated instructions.
c) There is a slight conflation of "iteration" and "stage"
in the documentation that makes it hard to follow what VRMap contains and the
invariants between functions.
My intent in D64665 was to create two abstractions: "Stage" and
"Block". Instead of referring to stages by index (VRMap), each Stage
would take a prior Stage as input. Stages are contained inside Blocks, which
handles predecessors and successors. I feel that arranging the code generation
in this CFG-like way will make the flow of data much easier to analyze. Of
course, a Stage doesn't just depend on a prior Stage - their loop carried
inputs could come from any other Stage (and while I think I understand how this
works, I clearly don't get all the edge cases).
What do you think of this abstraction? do you think it's doomed to failure
because it's too simplistic to cover all the cases?
Do you have any suggestions of areas where we can start to factor out without a
large-scale code breakage? I'm finding this hard to get my teeth into as the
proposed code structure is so different from its current form.
Thanks for any opinions or suggestions!
Cheers,
James
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190716/0b3688de/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.gif
Type: image/gif
Size: 105 bytes
Desc: image001.gif
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20190716/0b3688de/attachment-0001.gif>