illegal instruction fault frcsr t0

[yf13]

Dear experts,

I got exception Illegal instruction fault(2) with mepc=80004d44 and mtval=0 when running M-mode program on QEMU RiscV target. The code fragment looks like:

$ riscv64-unknown-elf-objdump -dS sample | fgrep 80004d44: -C1
80004d42: ad6e                  fsd fs11,152(sp)
80004d44: 003022f3              frcsr t0
80004d48: de16                  sw  t0,60(sp)

Could someone kindly teach what this exception means here?

[aswaterman]

There are only two possible explanations: the M-mode software forgot to set mstatus.FS to a nonzero value, or the QEMU instance had floating-point disabled.

[allenjbaum]

Except epc == 4d44 points to the frcsr, therefore the fsd must have executed without exception (so mstatus.FS must be nonzero)

• but if fsd is at 4d42, then frcsr should be at 4d46 - so this is a bit weird

On Sat, Oct 5, 2024 at 11:37 PM Andrew Waterman @.***> wrote:

There are only two possible explanations: the M-mode software forgot to set mstatus.FS to a nonzero value, or the QEMU instance had floating-point disabled.

— Reply to this email directly, view it on GitHub https://github.com/riscv/riscv-isa-manual/issues/1674#issuecomment-2395315611, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHPXVJXROLOG52IR3WGKR2TZ2DLB5AVCNFSM6AAAAABPOBCRKOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGOJVGMYTKNRRGE . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[yf13]

many thanks @aswaterman and @allenjbaum. Here my FS field is 0, so it is case 1 above.

When saving context, we should:

• Read mstatus to t0 and check t0.FS field;
• Only save float regs (w/ fcsr) when t0.FS is Dirty, and update t0.FS field to Clean
• Save t0 as mstatus to stack;

When resuming context, we should:

• Load saved mstatus.FS to t0 from stack
• If t0.FS is Clean, load float regs (w/ fcsr) from stack
• Copy t0 to mstatus

I am wondering that mstatus.FS field may change implicitly when float registers are updated, so we should update mstatus after recovering the float registers?

[aswaterman]

The act of loading the float registers in the context-restore code will change FS to Dirty, so you should change FS to Clean after performing the loads.

[yf13]

The act of loading the float registers in the context-restore code will change FS to Dirty, so you should change FS to Clean after performing the loads.

Thanks. Then if FS was Initial upon context resuming, we shall do fmv.w.x ft0, zero alike to set float registers to zero, then set FS to Clean afterwards?

[aswaterman]

That’s the right idea, but in that case, you can set it to Initial rather than Clean.

[yf13]

thanks again. I'll continue check why FS became 0.