riscv-software-src / riscv-pk

RISC-V Proxy Kernel
Other
570 stars 304 forks source link

GCC 15 + RVV cause illegal instruction sequence in PK #330

Open Maor545 opened 3 weeks ago

Maor545 commented 3 weeks ago

When compiling PK with the latest GCC v15 with architecture RV64GCV the compiler uses vector instructions for uart init before mstatus.VS is properly set which causes illegal instruction error in spike.

See the instruction sequence:

_install/bin/spike --isa=rv64gcv -d pk hello
(spike)
core   0: 0x0000000000001000 (0x00000297) auipc   t0, 0x0
(spike)
core   0: 0x0000000000001004 (0x02028593) addi    a1, t0, 32
(spike)
core   0: 0x0000000000001008 (0xf1402573) csrr    a0, mhartid
(spike)
core   0: 0x000000000000100c (0x0182b283) ld      t0, 24(t0)
(spike)
core   0: 0x0000000000001010 (0x00028067) jr      t0
(spike)
core   0: >>>>  MEM_START
core   0: 0x0000000080000000 (0x1f80006f) j       pc + 0x1f8
(spike)
core   0: >>>>  do_reset
core   0: 0x00000000800001f8 (0x00000093) li      ra, 0
(spike)
core   0: 0x00000000800001fc (0x00000113) li      sp, 0
(spike)
core   0: 0x0000000080000200 (0x00000193) li      gp, 0
(spike)
core   0: 0x0000000080000204 (0x00000213) li      tp, 0
(spike)
core   0: 0x0000000080000208 (0x00000293) li      t0, 0
(spike)
core   0: 0x000000008000020c (0x00000313) li      t1, 0
(spike)
core   0: 0x0000000080000210 (0x00000393) li      t2, 0
(spike)
core   0: 0x0000000080000214 (0x00000413) li      s0, 0
(spike)
core   0: 0x0000000080000218 (0x00000493) li      s1, 0
(spike)
core   0: 0x000000008000021c (0x00000613) li      a2, 0
(spike)
core   0: 0x0000000080000220 (0x00000693) li      a3, 0
(spike)
core   0: 0x0000000080000224 (0x00000713) li      a4, 0
(spike)
core   0: 0x0000000080000228 (0x00000793) li      a5, 0
(spike)
core   0: 0x000000008000022c (0x00000813) li      a6, 0
(spike)
core   0: 0x0000000080000230 (0x00000893) li      a7, 0
(spike)
core   0: 0x0000000080000234 (0x00000913) li      s2, 0
(spike)
core   0: 0x0000000080000238 (0x00000993) li      s3, 0
(spike)
core   0: 0x000000008000023c (0x00000a13) li      s4, 0
(spike)
core   0: 0x0000000080000240 (0x00000a93) li      s5, 0
(spike)
core   0: 0x0000000080000244 (0x00000b13) li      s6, 0
(spike)
core   0: 0x0000000080000248 (0x00000b93) li      s7, 0
(spike)
core   0: 0x000000008000024c (0x00000c13) li      s8, 0
(spike)
core   0: 0x0000000080000250 (0x00000c93) li      s9, 0
(spike)
core   0: 0x0000000080000254 (0x00000d13) li      s10, 0
(spike)
core   0: 0x0000000080000258 (0x00000d93) li      s11, 0
(spike)
core   0: 0x000000008000025c (0x00000e13) li      t3, 0
(spike)
core   0: 0x0000000080000260 (0x00000e93) li      t4, 0
(spike)
core   0: 0x0000000080000264 (0x00000f13) li      t5, 0
(spike)
core   0: 0x0000000080000268 (0x00000f93) li      t6, 0
(spike)
core   0: 0x000000008000026c (0x34001073) csrw    mscratch, zero
(spike)
core   0: 0x0000000080000270 (0x00000297) auipc   t0, 0x0
(spike)
core   0: 0x0000000080000274 (0xd9428293) addi    t0, t0, -620
(spike)
core   0: 0x0000000080000278 (0x30529073) csrw    mtvec, t0
(spike)
core   0: 0x000000008000027c (0x30502373) csrr    t1, mtvec
(spike)
core   0: 0x0000000080000280 (0x00629063) bne     t0, t1, pc + 0
(spike)
core   0: 0x0000000080000284 (0x00012117) auipc   sp, 0x12
(spike)
core   0: 0x0000000080000288 (0xc3c10113) addi    sp, sp, -964
(spike)
core   0: 0x000000008000028c (0xf14026f3) csrr    a3, mhartid
(spike)
core   0: 0x0000000080000290 (0x00c69613) slli    a2, a3, 12
(spike)
core   0: 0x0000000080000294 (0x00c10133) add     sp, sp, a2
(spike)
core   0: 0x0000000080000298 (0x00069463) bnez    a3, pc + 8
(spike)
core   0: 0x000000008000029c (0x3da0306f) j       pc + 0x33da
(spike)
core   0: >>>>  init_first_hart
core   0: 0x0000000080003676 (0x00007179) c.addi16sp sp, -48
(spike)
core   0: 0x0000000080003678 (0x0000852e) c.mv    a0, a1
(spike)
core   0: 0x000000008000367a (0x0000f406) c.sdsp  ra, 40(sp)
(spike)
core   0: 0x000000008000367c (0x0000f022) c.sdsp  s0, 32(sp)
(spike)
core   0: 0x000000008000367e (0x0000ec26) c.sdsp  s1, 24(sp)
(spike)
core   0: 0x0000000080003680 (0x00001800) c.addi4spn s0, sp, 48
(spike)
core   0: 0x0000000080003682 (0x0000e84a) c.sdsp  s2, 16(sp)
(spike)
core   0: 0x0000000080003684 (0x0000e44e) c.sdsp  s3, 8(sp)
(spike)
core   0: 0x0000000080003686 (0x000084ae) c.mv    s1, a1
(spike)
core   0: 0x0000000080003688 (0x5fc000ef) jal     pc + 0x5fc
(spike)
core   0: >>>>  query_uart
core   0: 0x0000000080003c84 (0xcc747057) vsetivli zero, 8, e8, mf2, ta, ma
core   0: exception trap_illegal_instruction, epc 0x0000000080003c84
core   0:           tval 0x00000000cc747057

I assume gcc uses vector calling convention now, since compiling with -fno-tree-vectorize does not fix the issue. The PK make script needs to set proper compiler flags for disabling vectorisation and vector calling convention (which I am not aware of) or needs to set mstatus ASAP before any compiler generated code gets executed.

What I did as a temporary fix in minit.c:

void init_first_hart(uintptr_t hartid, uintptr_t dtb)
{
// set mstatus first
  mstatus_init();
  // Confirm console as early as possible
  query_uart(dtb);
  query_uart16550(dtb);
  query_uart_litex(dtb);
  query_htif(dtb);
  ...

and removed the line

if (supports_extension('V'))
aswaterman commented 3 weeks ago

I think we'll have to avoid compiling with v in the -march string. But if you find a better way, I'm all ears.