[RFC] Intrinsics for Hardware Loops


Arm have recently announced the v8.1-M architecture specification for
our next generation microcontrollers. The architecture includes

vector extensions (MVE) and support for low-overhead branches (LoB),

which can be thought of a style of hardware loop. Hardware loops

aren’t new to LLVM, other backends (at least Hexagon and PPC that I

know of) also include support. These implementations insert the loop

controlling instructions at the MachineInstr level and I’d like to

propose that we add intrinsics to support this notion at the IR

level; primarily to be able to use scalar evolution to understand the

loops instead of having to implement a machine-level analysis for

each target.

I’ve posted an RFC with a prototype implementation in

https://reviews.llvm.org/D62132. It contains intrinsics that are

currently Arm specific, but I hope they’re general enough to be used

by all targets. The Arm v8.1-m architecture supports do-while and

while loops, but for conciseness, here, I’d like to just focus on

while loops. There’s two parts to this RFC: (1) the intrinsics

and (2) a prototype implementation in the Arm backend to enable

tail-predicated machine loops.

  1. LLVM IR Intrinsics

In the following definitions, I use the term ‘element’ to describe

the work performed by an IR loop that has not been vectorized or

unrolled by the compiler. This should be equivalent to the loop at

the source level.

void @llvm.arm.set.loop.iterations(i32)

  • Takes as a single operand, the number of iterations to be executed.

i32 @llvm.arm.set.loop.elements(i32, i32)

  • Takes two operands:

  • The total number of elements to be processed by the loop.

  • The maximum number of elements processed in one iteration of

the IR loop body.

  • Returns the number of iterations to be executed.


  • Takes as an operand, the number of elements that still need


  • Where ‘X’ denotes the vectorization factor, returns an array of i1

indicating which vector lanes are active for the current loop


i32 @llvm.arm.loop.end(i32, i32)

  • Takes two operands:

  • The number of elements that still need processing.

  • The maximum number of elements processed in one iteration of the

IR loop body.

The following gives an illustration of their intended usage:


%0 = call i32 @llvm.arm.set.loop.elements(i32 %N, i32 4)

%1 = icmp ne i32 %0, 0

br i1 %1, label %vector.ph, label %for.loopexit


br label %vector.body


%elts = phi i32 [ %N, %vector.ph ], [ %elts.rem, %vector.body ]

%active = call <4 x i1> @llvm.arm.get.active.mask(i32 %elts, i32 4)

%load = tail call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %addr, i32 4, <4 x i1> %active, <4 x i32> undef)

tail call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> %load, <4 x i32>* %addr.1, i32 4, <4 x i1> %active)

%elts.rem = call i32 @llvm.arm.loop.end(i32 %elts, i32 4)

%cmp = icmp sgt i32 %elts.rem, 0

br i1 %cmp, label %vector.body, label %for.loopexit


ret void

As the example shows, control-flow is still ultimately performed

through the icmp and br pair. There’s nothing connecting the

intrinsics to a given loop or any requirement that a set.loop.* call

needs to be paired with a loop.end call.

  1. Low-overhead loops in the Arm backend

Disclaimer: The prototype is barebones and reuses parts of NEON and

I’m currently targeting the Cortex-A72 which does not support this

feature! opt and llc build and the provided test case doesn’t cause a


The low-overhead branch extension can be combined with MVE to

generate vectorized loops in which the epilogue is executed within

the predicated vector body. The proposal is for this to be supported

through a series of pass:

  1. IR LoopPass to identify suitable loops and insert the intrinsics

proposed above.

  1. DAGToDAG ISel which makes the intrinsics, almost 1-1, to a pseduo


  1. A final MachineFunctionPass to expand the pseudo instructions.

To help / enable the lowering of of an i1 vector, the VPR register has

been added. This is a status register that contains the P0 predicate

and is also used to model the implicit predicates of tail-predicated


There are two main reasons why pseudo instructions are used instead

of generating MIs directly during ISel:

  1. They gives us a chance of later inspecting the whole loop and

confirm that it’s a good idea to generate such a loop. This is

trivial for scalar loops, but not really applicable for

tail-predicated loops.

  1. It allows us to separate the decrementing of the loop counter with

the instruction that branches back, which should help us recover if

LR gets spilt between these two pseudo ops.

For Armv8.1-M, the while.setup intrinsic is used to generate the wls

and wlstp instructions, while loop.end generates the le and letp

instructions. The active.mask can just be removed because the lane

predication is handled implicitly.

I’m not sure of the vectorizers limitations of generating vector

instructions that operate across lanes, such as reductions, when

generating a predicated loop but this needs to be considered.

I’d welcome any feedback here or on Phabricator and I’d especially like
to know if this would useful to current targets.


This proposal actually sounds very similar to what PPC currently does for counter-based loops. This solution tends to work well, in part because we can use SCEV to analyze loops at the IR level and generate trip-count expressions.

This seems like a generally reasonable approach. I have some hesitation about the potential separation of the control flow and the intrinsics (i.e. can we every confuse which loop they apply to?), but the basic notion seems reasonable. Particularly so as Hal points out that we already have something like this in PPC.  I’d suggest framing this as being an IR assist to backends rather than a canonical form or anything expected to be used by frontends though.

A couple of random comments; there’s no coherent message here, just a collection of thoughts.

  1. Your “loop.end” intrinsic is very confusingly named. I think you definitely need something different there name wise. Also, you fail to specify what the return value is.

  2. Your get.active.mask.X is a generally useful construct, but I think it can be represented via bitmath and a bitcast right? (i.e. does it have to be an intrinsic?)

  3. There seems to be a good amount of overlap with the SVE ideas. I’m not suggesting it needs to be reconciled, just pointing out many of the issues are common. (The more I see discussion of these topics, there more unsettled it all feels. Trying out a couple of experimental designs, and iterating until one wins is feeling more and more like the right approach.)


Hi Philip,

Yes, these constructs should really only be used by the compiler and probably always very late in the pipeline. To address your other points:

  1. Agreed. loop.end has now renamed to ‘loop.decrement’. I’ve also added ‘loop.decrement.reg’ which operates upon the updated loop counter, instead of some opaque system register.
  2. It could be handled by normal IR, the vectorizer currently splits out the equivalent when folding the epilogue into the loop body. The reason why we need an intrinsic is to work around the limitations of basic block isel. In our new architecture, the lane predication is implicit iff we can generate the hardware loop - but that doesn’t prevent other instructions, predicated on something other than the loop index, from being generated too. At ISel we can’t guarantee whether a predicate is loop index based or otherwise, so it has to be explicit coming into ISel.
  3. The main difference here is the same as (2). As I understand SVE, has bank of predicate registers that are explicitly accessed, whereas MVE has a status register that is used implicitly.

I’ll just note that I’m generally very skeptical of the argument in (2). Not actively objective, but every time this general line of thought comes up, I find the reasoning unconvincing.Â