Support for AArch64 Scalable Matrix Extension in LLVM

1. Introduction

The AArch64 SME ACLE provides a number of attributes for users to control PSTATE.SM and PSTATE.ZA. The AArch64 SME ABI describes the requirements for calls between functions when at least one of those functions uses PSTATE.SM or PSTATE.ZA.

This document describes how the SME ACLE attributes map to LLVM IR attributes and how LLVM lowers these attributes to implement the rules and requirements of the ABI.

Below we describe the LLVM IR attributes and their relation to the C/C++ level ACLE attributes:

aarch64_pstate_sm_enabled
is used for functions with __attribute__((arm_streaming))
aarch64_pstate_sm_compatible
is used for functions with __attribute__((arm_streaming_compatible))
aarch64_pstate_sm_body
is used for functions with __attribute__((arm_locally_streaming)) and is only valid on function definitions (not declarations)
aarch64_pstate_za_new
is used for functions with __attribute__((arm_new_za))
aarch64_pstate_za_shared
is used for functions with __attribute__((arm_shared_za))
aarch64_pstate_za_preserved
is used for functions with __attribute__((arm_preserves_za))

Clang must ensure that the above attributes are added both to the function’s declaration/definition as well as to their call-sites. This is important for calls to attributed function pointers, where there is no definition or declaration available.

2. Handling PSTATE.SM

When changing PSTATE.SM the execution of FP/vector operations may be transferred to another processing element. This has three important implications:

  • The runtime SVE vector length may change.
  • The contents of FP/AdvSIMD/SVE registers are zeroed.
  • The set of allowable instructions changes.

This leads to certain restrictions on IR and optimizations. For example, it is undefined behaviour to share vector-length dependent state between functions that may operate with different values for PSTATE.SM. Front-ends must honour these restrictions when generating LLVM IR.

Even though the runtime SVE vector length may change, for the purpose of LLVM IR and almost all parts of CodeGen we can assume that the runtime value for vscale does not. If we let the compiler insert the appropriate smstart and smstop instructions around call boundaries, then the effects on SVE state can be mitigated. By limiting the state changes to a very brief window around the call we can control how the operations are scheduled and how live values remain preserved between state transitions.

In order to control PSTATE.SM at this level of granularity, we use function and callsite attributes rather than intrinsics.

Restrictions on attributes

  • It is undefined behaviour to pass or return (pointers to) scalable vector objects to/from functions which may use a different SVE vector length. This includes functions with a non-streaming interface, but marked with aarch64_pstate_sm_body.
  • It is not allowed for a function to be decorated with both aarch64_pstate_sm_compatible and aarch64_pstate_sm_enabled.
  • It is not allowed for a function to be decorated with both aarch64_pstate_za_new and aarch64_pstate_za_preserved.
  • It is not allowed for a function to be decorated with both aarch64_pstate_za_new and aarch64_pstate_za_shared.

These restrictions also apply in the higher level SME ACLE, which means we can emit diagnostics in Clang to signal users about incorrect behaviour.

Compiler inserted streaming-mode changes

The table below describes the transitions in PSTATE.SM the compiler has to account for when doing calls between functions with different attributes. In this table, we use the following abbreviations:

N
functions with a normal interface (PSTATE.SM=0 on entry, PSTATE.SM=0 on return)
S
functions with a Streaming interface (PSTATE.SM=1 on entry, PSTATE.SM=1 on return)
SC
functions with a Streaming-Compatible interface (PSTATE.SM can be either 0 or 1 on entry, and is unchanged on return).

Functions with __attribute__((arm_locally_streaming)) are excluded from this table because for the caller the attribute is synonymous to ‘streaming’, and for the callee it is merely an implementation detail that is explicitly not exposed to the caller.

Combinations of calls for functions with different attributes
From To Before call After call After exception
N N      
N S SMSTART SMSTOP  
N SC      
S N SMSTOP SMSTART SMSTART
S S     SMSTART
S SC     SMSTART
SC N If PSTATE.SM before call is 1, then SMSTOP If PSTATE.SM before call is 1, then SMSTART If PSTATE.SM before call is 1, then SMSTART
SC S If PSTATE.SM before call is 0, then SMSTART If PSTATE.SM before call is 0, then SMSTOP If PSTATE.SM before call is 1, then SMSTART
SC SC     If PSTATE.SM before call is 1, then SMSTART

Because changing PSTATE.SM zeroes the FP/vector registers, it is best to emit the smstart and smstop instructions before register allocation, so that the register allocator can spill/reload registers around the mode change.

The compiler should also have sufficient information on which operations are part of the call/function’s arguments/result and which operations are part of the function’s body, so that it can place the mode changes in exactly the right position. The suitable place to do this seems to be SelectionDAG, where it lowers the call’s arguments/return values to implement the specified calling convention. SelectionDAG provides Chains and Glue to specify the order of operations and give preliminary control over the instruction’s scheduling.

Example of preserving state

When passing and returning a float value to/from a function that has a streaming interface from a function that has a normal interface, the call-site will need to ensure that the argument/result registers are preserved and that no other code is scheduled in between the smstart/smstop and the call.

define float @foo(float %f) nounwind {
  %res = call float @bar(float %f) "aarch64_pstate_sm_enabled"
  ret float %res
}

declare float @bar(float) "aarch64_pstate_sm_enabled"

The program needs to preserve the value of the floating point argument and return value in register s0:

foo:                                    // @foo
// %bb.0:
        stp     d15, d14, [sp, #-80]!           // 16-byte Folded Spill
        stp     d13, d12, [sp, #16]             // 16-byte Folded Spill
        stp     d11, d10, [sp, #32]             // 16-byte Folded Spill
        stp     d9, d8, [sp, #48]               // 16-byte Folded Spill
        str     x30, [sp, #64]                  // 8-byte Folded Spill
        str     s0, [sp, #76]                   // 4-byte Folded Spill
        smstart sm
        ldr     s0, [sp, #76]                   // 4-byte Folded Reload
        bl      bar
        str     s0, [sp, #76]                   // 4-byte Folded Spill
        smstop  sm
        ldp     d9, d8, [sp, #48]               // 16-byte Folded Reload
        ldp     d11, d10, [sp, #32]             // 16-byte Folded Reload
        ldp     d13, d12, [sp, #16]             // 16-byte Folded Reload
        ldr     s0, [sp, #76]                   // 4-byte Folded Reload
        ldr     x30, [sp, #64]                  // 8-byte Folded Reload
        ldp     d15, d14, [sp], #80             // 16-byte Folded Reload
        ret

Setting the correct register masks on the ISD nodes and inserting the smstart/smstop in the right places should ensure this is done correctly.

Instruction Selection Nodes

AArch64ISD::SMSTART Chain, [SM|ZA|Both], CurrentState, ExpectedState[, RegMask]
AArch64ISD::SMSTOP  Chain, [SM|ZA|Both], CurrentState, ExpectedState[, RegMask]

The SMSTART/SMSTOP nodes take CurrentState and ExpectedState operand for the case of a conditional SMSTART/SMSTOP. The instruction will only be executed if CurrentState != ExpectedState.

When CurrentState and ExpectedState can be evaluated at compile-time (i.e. they are both constants) then an unconditional smstart/smstop instruction is emitted. Otherwise the node is matched to a Pseudo instruction which expands to a compare/branch and a smstart/smstop. This is necessary to implement transitions from SC -> N and SC -> S.

Unchained Function calls

When a function with “aarch64_pstate_sm_enabled” calls a function that is not streaming compatible, the compiler has to insert a SMSTOP before the call and insert a SMSTOP after the call.

If the function that is called is an intrinsic with no side-effects which in turn is lowered to a function call (e.g. @llvm.cos()), then the call to @llvm.cos() is not part of any Chain; it can be scheduled freely.

Lowering of a Callsite creates a small chain of nodes which:

  • starts a call sequence
  • copies input values from virtual registers to physical registers specified by the ABI
  • executes a branch-and-link
  • stops the call sequence
  • copies the output values from their physical registers to virtual registers

When the callsite’s Chain is not used, only the result value from the chained sequence is used, but the Chain itself is discarded.

The SMSTART and SMSTOP ISD nodes return a Chain, but no real values, so when the SMSTART/SMSTOP nodes are part of a Chain that isn’t used, these nodes are not considered for scheduling and are removed from the DAG. In order to prevent these nodes from being removed, we need a way to ensure the results from the CopyFromReg can only be used after the SMSTART/SMSTOP has been executed.

We can use a CopyToReg -> CopyFromReg sequence for this, which moves the value to/from a virtual register and chains these nodes with the SMSTART/SMSTOP to make them part of the expression that calculates the result value. The resulting COPY nodes are removed by the register allocator.

The example below shows how this is used in a DAG that does not link together the result by a Chain, but rather by a value:

            t0: ch,glue = AArch64ISD::SMSTOP ...
          t1: ch,glue = ISD::CALL ....
        t2: res,ch,glue = CopyFromReg t1, ...
      t3: ch,glue = AArch64ISD::SMSTART t2:1, ....   <- this is now part of the expression that returns the result value.
    t4: ch = CopyToReg t3, Register:f64 %vreg, t2
  t5: res,ch = CopyFromReg t4, Register:f64 %vreg
t6: res = FADD t5, t9

We also need this for locally streaming functions, where an SMSTART needs to be inserted into the DAG at the start of the function.

Functions with __attribute__((arm_locally_streaming))

If a function is marked as arm_locally_streaming, then the runtime SVE vector length in the prologue/epilogue may be different from the vector length in the function’s body. This happens because we invoke smstart after setting up the stack-frame and similarly invoke smstop before deallocating the stack-frame.

To ensure we use the correct SVE vector length to allocate the locals with, we can use the streaming vector-length to allocate the stack-slots through the ADDSVL instruction, even when the CPU is not yet in streaming mode.

This only works for locals and not callee-save slots, since LLVM doesn’t support mixing two different scalable vector lengths in one stack frame. That means that the case where a function is marked arm_locally_streaming and needs to spill SVE callee-saves in the prologue is currently unsupported. However, it is unlikely for this to happen without user intervention, because arm_locally_streaming functions cannot take or return vector-length-dependent values. This would otherwise require forcing both the SVE PCS using ‘aarch64_sve_pcs’ combined with using arm_locally_streaming in order to encounter this problem. This combination can be prevented in Clang through emitting a diagnostic.

An example of how the prologue/epilogue would look for a function that is attributed with arm_locally_streaming:

#define N 64

void __attribute__((arm_streaming_compatible)) some_use(svfloat32_t *);

// Use a float argument type, to check the value isn't clobbered by smstart.
// Use a float return type to check the value isn't clobbered by smstop.
float __attribute__((noinline, arm_locally_streaming)) foo(float arg) {
  // Create local for SVE vector to check local is created with correct
  // size when not yet in streaming mode (ADDSVL).
  float array[N];
  svfloat32_t vector;

  some_use(&vector);
  svst1_f32(svptrue_b32(), &array[0], vector);
  return array[N - 1] + arg;
}

should use ADDSVL for allocating the stack space and should avoid clobbering the return/argument values.

_Z3foof:                                // @_Z3foof
// %bb.0:                               // %entry
        stp     d15, d14, [sp, #-96]!           // 16-byte Folded Spill
        stp     d13, d12, [sp, #16]             // 16-byte Folded Spill
        stp     d11, d10, [sp, #32]             // 16-byte Folded Spill
        stp     d9, d8, [sp, #48]               // 16-byte Folded Spill
        stp     x29, x30, [sp, #64]             // 16-byte Folded Spill
        add     x29, sp, #64
        str     x28, [sp, #80]                  // 8-byte Folded Spill
        addsvl  sp, sp, #-1
        sub     sp, sp, #256
        str     s0, [x29, #28]                  // 4-byte Folded Spill
        smstart sm
        sub     x0, x29, #64
        addsvl  x0, x0, #-1
        bl      _Z10some_usePu13__SVFloat32_t
        sub     x8, x29, #64
        ptrue   p0.s
        ld1w    { z0.s }, p0/z, [x8, #-1, mul vl]
        ldr     s1, [x29, #28]                  // 4-byte Folded Reload
        st1w    { z0.s }, p0, [sp]
        ldr     s0, [sp, #252]
        fadd    s0, s0, s1
        str     s0, [x29, #28]                  // 4-byte Folded Spill
        smstop  sm
        ldr     s0, [x29, #28]                  // 4-byte Folded Reload
        addsvl  sp, sp, #1
        add     sp, sp, #256
        ldp     x29, x30, [sp, #64]             // 16-byte Folded Reload
        ldp     d9, d8, [sp, #48]               // 16-byte Folded Reload
        ldp     d11, d10, [sp, #32]             // 16-byte Folded Reload
        ldp     d13, d12, [sp, #16]             // 16-byte Folded Reload
        ldr     x28, [sp, #80]                  // 8-byte Folded Reload
        ldp     d15, d14, [sp], #96             // 16-byte Folded Reload
        ret

Preventing the use of illegal instructions in Streaming Mode

  • When executing a program in streaming-mode (PSTATE.SM=1) a subset of SVE/SVE2 instructions and most AdvSIMD/NEON instructions are invalid.
  • When executing a program in normal mode (PSTATE.SM=0), a subset of SME instructions are invalid.
  • Streaming-compatible functions must only use instructions that are valid when either PSTATE.SM=0 or PSTATE.SM=1.

The value of PSTATE.SM is not controlled by the feature flags, but rather by the function attributes. This means that we can compile for ‘+sme’ and the compiler will code-generate any instructions, even if they are not legal under the requested streaming mode. The compiler needs to use the function attributes to ensure the compiler doesn’t do transformations under the assumption that certain operations are available at runtime.

We made a conscious choice not to model this with feature flags, because we still want to support inline-asm in either mode (with the user placing smstart/smstop manually), and this became rather complicated to implement at the individual instruction level (see D120261 and D121208) because of limitations in TableGen.

As a first step, this means we’ll disable vectorization (LoopVectorize/SLP) entirely when the a function has either of the aarch64_pstate_sm_enabled, aarch64_pstate_sm_body or aarch64_pstate_sm_compatible attributes, in order to avoid the use of vector instructions.

Later on we’ll aim to relax these restrictions to enable scalable auto-vectorization with a subset of streaming-compatible instructions, but that requires changes to the CostModel, Legalization and SelectionDAG lowering.

We will also emit diagnostics in Clang to prevent the use of non-streaming(-compatible) operations, e.g. through ACLE intrinsics, when a function is decorated with the streaming mode attributes.

Other things to consider

  • Inlining must be disabled when the call-site needs to toggle PSTATE.SM or when the callee’s function body is executed in a different streaming mode than its caller. This is needed because function calls are the boundaries for streaming mode changes.
  • Tail call optimization must be disabled when the call-site needs to toggle PSTATE.SM, such that the caller can restore the original value of PSTATE.SM.

3. Handling PSTATE.ZA

In contrast to PSTATE.SM, enabling PSTATE.ZA does not affect the SVE vector length and also doesn’t clobber FP/AdvSIMD/SVE registers. This means it is safe to toggle PSTATE.ZA using intrinsics. This also makes it simpler to setup a lazy-save mechanism for calls to private-ZA functions (i.e. functions that may either directly or indirectly clobber ZA state).

For this purpose, we’ll introduce a new LLVM IR pass that is run just before SelectionDAG.