[llvm-dev] Simulation of load-store forwarding with MI scheduler on AArch64

[Evgeny Leviant via llvm-dev]

Hi list,

 

Is it possible to simulate load to store forwarding on aarch64 with MI scheduling model on AArch64?

For instance $x0 data latency in the example below should be 1 cycle

 

ldr $x0, [$x1]

str $x0, [$x2]

 

But it should be 4 cycles if we have another instruction:

 

ldr $x0, [$x1]

add $x0, $x0, 4

 

For ALU instructions it’s possible to use either ReadAdvance or SchedReadAdvance, but I don’t see how

to do this with WriteLD or WriteST. Is there some workaround?

[Andrew Trick via llvm-dev]

The main purpose of ReadAdvance is pipeline forwarding.

I think you can just want a read resource in your subtarget like this:

  def ReadAdr : SchedReadAdvance<3, [WriteLD]>

Briefly glancing at the AArch64 target I see this for stores:

  Sched<[WriteST]>;

So it doesn't look like there's any existing name for the store’s address operand. You could add a general ReadAdr SchedRead resource

in AArch64Schedule.td. Then you would need to change the ReadAdr line in your subtarget to an override:

  def : ReadAdvance<ReadAdr, 3, [WriteLD]>

Or instead you can just add a rule in your subtarget listing the opcodes or using a regex, and using the ReadAdr resource that you defined in the same file.

  def : InstRW<[WriteST, ReadAdr], (instregex "ST(someregex)$")>;

Being careful about store-pair and vector stores.

Then you always want to debug your target’s llvm-tblgen command by adding a flag

-debug-only=subtarget-emitter

And even trace the schedule for some simple cases with -debug-only=machine-scheduler

I haven't actually done any of this in several years, someone with more recent experience may have better tips.

-Andy

[Evgeny Leviant via llvm-dev]

Thanks for prompt response, Andy

 

This will work for cases when address is not modified. However this doesn’t seem to work for pre/post increment load stores.

Consider data to address forwarding:

 

$x0 = ldr x0, [x1]

$x0, $x2 = ldr x2, [x0, 16]!

 

The second instruction will have it’s own latency for address modification ($x0 register). So I don’t see how we can use ReadAdr stuff

here. May be forwarding is not supposed to work in such cases for ARM cpus? Cortex-A55 software optimization guide says this:

 

“load data from a limited set of load instructions can be forwarded from the beginning of the wr pipeline stage to either the load or store AGU base operand”

 

However nothing is said about pre/post indexed forms.

 

From: Andrew Trick
Sent: 15 сентября 2020 г. 7:04
To: Evgeny Leviant
Cc: llvm...@lists.llvm.org
Subject: [EXTERNAL] Re: [llvm-dev] Simulation of load-store forwarding with MI scheduler on AArch64

 

CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.  If you suspect potential phishing or spam email, report it to Repor...@accesssoftek.com

[Evgeny Leviant via llvm-dev]

Sorry, it seems I have figured out the answer myself:

 

Instruction

$x0, $x2 = LDRXpre $x0, 1

 

will have 4 arguments, so it seems possible to assign both SchedRead and SchedWrite  for $x0 and the result sched

list for LDRXpre would be:

 

[WriteAdr, WriteLD, ReadAdr]

 

Strange that AArch64InstrFormats.td doesn’t implement this

 

Sent from Mail for Windows 10

[Andrew Trick via llvm-dev]

On Sep 15, 2020, at 5:01 AM, Evgeny Leviant <elev...@accesssoftek.com> wrote:

Sorry, it seems I have figured out the answer myself:

 

Instruction

$x0, $x2 = LDRXpre $x0, 1

 

will have 4 arguments, so it seems possible to assign both SchedRead and SchedWrite  for $x0 and the result sched

list for LDRXpre would be:

 

[WriteAdr, WriteLD, ReadAdr]

 

Strange that AArch64InstrFormats.td doesn’t implement this

Looking at AArch64InstrFormats.td, it has the writeback operand in a different order. 

Sched<[WriteLD, WriteAdr]>;

If that’s incorrect, it might be worth fixing.

Otherwise, that looks like the right answer. The address writeback is not really load forwarding. It's a totally separate scheduling resource with its own latency.

-Andy