rv64i_m ebreak.S - DUT and Ref run different test programs

[Peter R Herrmann]

I am running an RV64IMZicsr_Zifencei DUT against SAIL and I am struggling to understand why the elf's generated for the Ref and DUT differ on this single instruction:

00000000800008fc <chk_Mtval>:
    800008fc:   00f57513            andi    a0,a0,15        # a0 contains mcause (0x3 for ebreak)
    80000900:   00000437            lui	    s0,0x0          # s0 = 0
    80000904:   00040413            mv      s0,s0           # - 
    80000908:   00b41413            slli    s0,s0,0xb       # - 
    8000090c:   00040413            mv      s0,s0           # - 
    80000910:   00b41413            slli    s0,s0,0xb       # - 
    80000914:   02c40413            addi    s0,s0,44        # s0 = 0x2c
    80000918:   00a41413            slli    s0,s0,0xa       # s0 = 0x2c << 0xa = 0xB000
    8000091c:   0fb40413            addi    s0,s0,251       # s0 = 0xB0FB       <- This instruction's immediate is different between DUT (0x0FB) and REF (0x0F3)
    80000920:   00a45433            srl     s0,s0,a0        # s0 = 0xB0FB >> 0x3 = 0x161F (DUT) or 0x161E (Ref)
    80000924:   03f41413            slli    s0,s0,0x3f      # s0 = 0x161F << 0x3f = 0x8000000000000000(DUT) or 0x0 (Ref)
    80000928:   04045c63            bgez    s0,80000980 <sv_Mtval>

This section of code checks mcause against a mask, 0xB0FB or 0xB0F3.

• If the mask bit associated with the given mcause value is set (i.e 3rd bit for mcause = 3), then it proceeds to check if mtval holds the PC for the exception thrown, writing the offset from the beginning of the test to the signature file.
• If the same bit is clear, then the test assumes that mtval contains zero and writes that to the signature.

This behavior seems to correspond to the mtval_update config in the riscv-config, but changing this value has no effect.

Essentially what I am running into is that the Ref test executable seems to always assume that mtval=0 after EBREAK (without checking), while the DUT executable expects mtval=PC after EBREAK.

If I modify the DUT's elf so that this mask is the same as the instruction in the Ref's elf, the test passes, otherwise the test fails. Likewise, If I modify the Ref's elf file so that the mask matches the DUT, the test passes.

• If both tests have 8000091c: 0fb40413, the signatures match
• If both tests have 8000091c: 0f340413, the signatures match
• If the tests have different values here (specifically if bit 3 is different) the signatures will not match

It does not seem correct to me that the DUT and Ref could possibly have different expectations during the test like this.

My isa.yaml:

hart_ids: [0]
hart0:
  ISA: RV64IMZicsr_Zifencei
  physical_addr_sz: 56
  User_Spec_Version: '2.3'
  supported_xlen: [64]
  hw_data_misaligned_support: false

and my platform.yaml:

mtime:
  implemented: false
mtimecmp:
  implemented: false
nmi:
  label: nmi_vector
reset:
  label: reset_vector

And here is the source for my project if needed:
rv-application-rough-RV64IMZicsr_Zifencei.zip