[SOLVED] Encountering Illegal Instruction Error When Using QEMU to Simulate ARMv9 SME Program

Encountering Illegal Instruction Error When Using QEMU to Simulate ARMv9 SME Program

I am trying to use QEMU in user mode to simulate an ARMv9 SME program, but I encounter the following error during execution:

qemu: uncaught target signal 4 (Illegal instruction) - core dumped

Environment Information:

QEMU Version: qemu-9.0.2
Target Architecture: ARMv9

SME Code Snippet:

matmul_opt: // x0: M, x1: K, x2: N, x3: matLeft, x4: matRight, x5: matResult
  stp     x19, x20, [sp, #-48]!
  stp     x21, x22, [sp, #16]
  stp     x23, x24, [sp, #32]
  smstart
  // constants
  cntw    x6                      // SVLs
  mul     x22, x6, x1             // SVLs*K
  mul     x23, x6, x2             // SVLs*N
  add     x18, x23, x2            // SVLs*N + N
  add     x11, x4, x2, lsl #2     // Exit condition for N loop
  mov     x12, #0
  cntb    x6                      // SVLb
  mov     x14, #0
  ptrue   pn10.b                  // Predicate for SME2 VLx2 (a_ptr loads)
  whilelt pn8.s, x12, x0, vlx2    // tiles predicate (M dimension)
  sub     w6, w6, #8              // SVLb-8

.Loop_M:
  // Extract tile 0/1 and tile 2/3 predicates (M) from vlx2 predicate.
  pext    { p2.s, p3.s }, pn8[0]
  mov     x16, x4                 // b_base
  mov     x9, x5                  // c_base
  whilelt pn9.b, x16, x11, vlx2   // tiles predicate (N dimension)

.Loop_N:
  mov     x7, x3                  // a_ptr = a_base
  mov     x17, x16                // b_ptr = b_base
  mov     x10, x9                 // c_ptr0 = c_base

  // Extract tile 0/2 and tile 1/3 predicates (N) from vlx2 predicate.
  pext    { p0.b, p1.b }, pn9[0]

  add     x8, x3, x22, lsl #2     // a_base + SVLs*K FP32 elms (bytes)
  addvl   x15, x8, #-1            // Exit condition for K loop
  ld1w    {z1.s}, p2/z, [x7]      // Load 1st vector from a_ptr

  zero    {za}
  ld1w    {z2.s-z3.s}, pn9/z, [x17]  // Load 2 vectors from b_ptr

  fmopa   za0.s, p2/m, p0/m, z1.s, z2.s // ZA0+=1st a_ptr vec OP 1st b_ptr vec
  ld1w    {z5.s}, p3/z, [x7, x22, lsl #2]  // Load 2nd vector from a_ptr
  addvl   x7, x7, #1                       // a_ptr += SVLb (bytes)

.Loop_K:
  fmopa   za2.s, p3/m, p0/m, z5.s, z2.s // ZA2+=2nd a_ptr vec OP 1st b_ptr vec

  // ... (rest of the loop body)

  cmp     x7, x15
  b.mi    .Loop_K

  // ... (rest of the N loop body)

.Loop_store_ZA:
  // ... (rest of the store loop body)

  cmp     w13, w6
  b.mi    .Loop_store_ZA

  // ... (rest of the M loop body)

  smstop

  ldp     x23, x24, [sp, #32]
  ldp     x21, x22, [sp, #16]
  ldp     x19, x20, [sp], #48

  ret

Compile Command:

aarch64-none-elf-as -march=armv9.2-a -mcpu=cortex-a710+sme2 -o test_sme test_sme.S

Run Command:

qemu-aarch64 test_sme

While executing the above program, I receive the "Illegal instruction" error. What could be the reason for this? Does my version of QEMU support the SME2 instruction set that I'm using? How can I resolve this issue?

Any help would be greatly appreciated!

Solution

The instruction "ptrue pn10.b" is in the FEAT_SVE2p1 extension, which QEMU does not yet implement; similarly the form of the "while" insn you are using is FEAT_SVE2p1. So you get an "illegal instruction" error, the same as you would if you ran this binary on a real CPU that didn't implement FEAT_SVE2p1. If you want to run your code under QEMU you should stick to FEAT_SME/FEAT_SVE/FEAT_SVE2 for the moment. (FEAT_SVE2p1 and FEAT_SME2 are on the todo list but may be a while yet.)

The currently emulated set of architectural features are listed here: https://www.qemu.org/docs/master/system/arm/emulation.html