Skip to content

trap: mstatus stack, WARL CSRs, misaligned-fetch and illegal-instruction faults, mstatus.FS - #487

Open
tinebp wants to merge 2 commits into
masterfrom
fix/trap-csr-mstatus
Open

tinebp wants to merge 2 commits into
masterfrom
fix/trap-csr-mstatus

Conversation

@tinebp

@tinebp tinebp commented Nov 19, 2026

Copy link
Copy Markdown
Collaborator

Summary

Two commits complete the M-mode trap path. The design doc is updated (docs/designs/trap_and_exception_foundation.md).

1. mstatus stack, WARL CSRs and misa (fixes #447 and #448)

  • Trap entry sets MPIE←MIE and MIE←0, and MPP reads 11.
  • mret restores MIE←MPIE and sets MPIE←1.
  • mtvec and mepc are WARL. misa puts MXL in the top bits at XLEN 64.
  • SimX, the debug module and dtm_test.py match.
  • The rv32mi/rv64mi riscv-tests run, with a documented skip list.

2. Fault producers and mstatus.FS (fixes #445 and #446)

  • Misaligned fetch (cause 0): a taken jump or branch to a target with bit 1 set, when C is off. Before, it fetched the misaligned PC in debug builds and silently dropped bits in release builds.

  • Illegal instruction (cause 2), through a new INST_BR_ILLEGAL uop:

    • unknown opcodes, including extensions built out;
    • reserved F/D encodings;
    • shamt[5] on RV32;
    • writes to read-only CSRs;
    • reserved compressed encodings;
    • reserved branch funct3 and unknown SYSTEM encodings;
    • FP instructions or FP CSR accesses while FS is Off.

    WFI runs as a no-op.

  • FS: resets to Initial. FP writes set Dirty at decode, and SD follows. FS is read-only zero without F.

  • A CSR access to mstatus holds its warp until the CSR unit takes it, so the FS check and the Dirty update stay in program order.

  • Trap CSRs are device state: a launch keeps them on both sides. SimX used to clear them at every launch.

  • mtval is 0 for every cause, which the spec allows. Writing the target there would cost about 330 LUT of per-warp muxes.

Impact and evidence

Impact class: R3 (ci/change_impact.py classify). Maintainer approval needed.

Contracts touched:

  • VX_decode_sched_if: adds fs_off and fs_dirty.
  • VX_sched_csr_if: adds csr_unlock.
  • sched_csr_wr_t
  • INST_BR_ILLEGAL replaces INST_BR_OTHER.
  • RTL and SimX decoders, moved together.
  • VX_types.toml: VX_CSR_MSTATUS_* bit positions.
evidence status detail
functional tests (simx / rtlsim / xrt) run rv32mi/rv64mi (mcsr, scall, sbreak, illegal, shamt, ma_fetch) on SimX and rtlsim at XLEN 32 and 64, with C off and with A+C. New trap_illegal passes on SimX and rtlsim at XLEN 32 and 64; on master it fails every probe. RVC unit test passes at XLEN 32 and 64. multikernel, reopen and module_reload pass on SimX. sgemm, diverge and sgemm_tcu pass.
model_parity not run sgemm instruction and cycle counts are unchanged
perf_gate not run no mstatus access in the perf kernels; sgemm counts unchanged
Beacon logic_gate run core (A+C, CI toolchain): PASS. Depth 15, unchanged. +68 LUT, +5 FF vs master.
synthesis not run Beacon depth unchanged
XLEN 64 build run rtlsim -Wall lint at XLEN 32 and 64 (standard matrix, C+A, 2 ALU lanes) and at XLEN 32 with F off; SimX with -DSIMX_FUNCTIONAL

New CI cases:

  • isa-mi: machine-mode suites with C off, the only build where ma_fetch reaches the trap.

  • trap-illegal

  • No test silenced. The csr riscv-test stays skipped for the documented reason: tests 14-15 need misa.U cleared.

  • No golden data changed.

  • RTL and SimX move together.

  • No toolchain pin change.

🤖 Generated with Claude Code

tinebp and others added 2 commits November 19, 2026 03:42
… misa MXL at XLEN 64

The synchronous trap path (ECALL/EBREAK/MRET) redirected the PC and saved
mepc/mcause but never touched mstatus, and the trap CSRs were plain XLEN
latches with no legalisation:
- trap entry did not do MPIE<-MIE, MIE<-0, and MRET did not restore MIE
  from MPIE, so the interrupt-enable stack that preemption relies on was
  missing (#447);
- mepc kept bit 1 with C disabled (and bit 0 always), and mtvec kept any
  MODE value, though only direct mode exists (#448).

Each warp now stores only the defined bits: mstatus MIE, MPIE and FS (FS is
a plain field for now), with MPP reading 11 since M is the only mode;
mtvec [XLEN-1:2], so MODE reads 0; mepc [XLEN-1:1] with C and [XLEN-1:2]
without. Trap entry and MRET update the stack; hardware still wins a
same-cycle software write. SimX matches bit for bit. The bit positions
are shared through VX_types.toml (VX_CSR_MSTATUS_*). Beacon on the core:
logic depth unchanged, 496 fewer flops (31 per warp).

At XLEN 64 misa placed MXL at bits 31:30 instead of XLEN-1:XLEN-2, in the
RTL, SimX and the SimX debug module; it now uses the top bits. The DTM test
read only DATA0 for misa and never wrote DATA1, which hid a stale high word
in its register writes; it now handles both at XLEN 64.

SimX gaps that blocked the machine-mode suite are closed: misa returned 0,
and mvendorid/marchid/mimpid aborted; they now match the RTL.

The ISA suites gain rv32mi/rv64mi (mcsr, scall, sbreak). Tests that need
fault producers, user mode, triggers, counters or PMP are listed in
ISA_SKIP with the reason. Passing on SimX and rtlsim at XLEN 32 and 64,
with C on and off; rv32ui/rv32uf unchanged.

Fixes #447, fixes #448.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cking

The trap path handled only ECALL/EBREAK. A taken jump or branch to a target
with bit 1 set (C off) did not trap: DEBUG builds fetched the misaligned PC
and release builds silently dropped the low bits (#445). mstatus.FS was not
tracked, so FP instructions ran with FS=Off and FS never became Dirty
(#446). Illegal encodings had no defined behaviour: undefined decode fields
in the RTL, an abort in SimX.

Misaligned fetch (cause 0): the branch unit takes target bit 1 from the
adder before PC truncation, so DEBUG and release behave the same; the
decision uses the last active lane, the lane the warp follows, and a
trapping jump writes no link register. mepc is the jump's PC (carried in
dest, as for ECALL); mtval is 0 for every cause, which the spec allows and
which avoids ~330 LUT of per-warp target muxes.

Illegal instruction (cause 2), a new INST_BR_ILLEGAL uop that stalls the
warp like ECALL: unknown major opcodes (including built-out extensions),
reserved F/D encodings, shamt[5] on RV32 and RV64 word shifts, writes to
read-only CSRs, reserved compressed encodings (the RTL decompressor now
expands them to the all-zero word; SimX marks the same set, and the rd=x0
HINT forms execute as no-ops on both), and FP instructions or FP-CSR
accesses under FS=Off, including the Vortex extensions that use FP
registers (packed/wide/matrix FP ld/st, TCU MMA, TFU, RTU GETWF/CB_RET),
conditional branches with the reserved funct3 2 or 3, and SYSTEM
encodings other than the CSR accesses, ECALL, EBREAK, URET, SRET, MRET
and WFI. WFI runs as a no-op (there is no interrupt to wait for: the RTL
decodes it as a jump to the next instruction), and no SYSTEM instruction
outside the CSR accesses writes rd.

FS: resets to Initial so kernels that never touch mstatus keep running,
and is read-only zero without F;
SD reads FS==Dirty; FP-state writers set Dirty at decode unless they trap.
An mstatus CSR access stalls its warp at decode until the CSR unit takes
it (csr_unlock on sched_csr_if), keeping the FS check and the Dirty update
in program order; SimX releases at the same point. misa writes are ignored
(WARL).

SimX applies trap, mret and PC-redirect effects only on an instruction's
last lane packet, as the RTL does; before, a split instruction raised a
trap once per packet and clobbered MPIE (also for ECALL/EBREAK/MRET). SimX
aborts with the cause, PC and instruction when an illegal or misaligned
fault is taken with mtvec=0 (a plain kernel has no handler); the RTL jumps
to mtvec as architected. The trap CSRs are device state on both sides: a
device reset clears them and a kernel launch keeps them (SimX used to
clear them at every launch, the RTL never did).

rv32mi/rv64mi now also run illegal, shamt and ma_fetch; csr stays skipped
(tests 14-15 need misa.U cleared or an unimplemented-CSR trap). CI runs
them on the default build with C off (isa-mi), the only build where
ma_fetch reaches the misaligned-fetch trap. A new regression, trap_illegal,
covers what no riscv-test that runs reaches: FP instructions, loads, CSR
accesses and moves under FS=Off, the reserved branch and SYSTEM encodings
and WFI (cause and mepc each), and FS going from Initial to Dirty with SD;
on the old code it fails every probe. The RVC unit test now expects the
rd=x0 HINTs to expand and checks the reserved points. Passing on
SimX and rtlsim at XLEN 32 and 64, default and A+C, rv32ui/uc with C on,
NUM_ALU_LANES=2 (rv32mi, sgemm, diverge), sgemm_tcu; sgemm instruction and
cycle counts unchanged. Beacon (core, A+C): depth unchanged (15, mean-50
11.6, 278 registers at 8+ levels), +68 LUT, +5 FF.

Fixes #445, fixes #446.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

1 participant