Single precision instruction on double precision register value
philippuk opened this issue · 1 comments
philippuk commented
Assembly code to produce this issue and the output of different RISC-V implementation
fclass.s s0, fa3
----------------------------------------
Reg_name Spike RVEMU FORVIS SAIL
x8/s0/fp 200 4 200 200
fadd.s fa6, fs9, fa3
----------------------------------------
Reg_name Spike RVEMU FORVIS SAIL
f16/fa6 nan 0 ffffffff7fc00000 ffffffff7fc00000
fnmadd.s ft9, fa7, ft0, fs10
----------------------------------------
Reg_name Spike RVEMU FORVIS SAIL
f29/ft9 nan -0 ffffffff7fc00000 ffffffff7fc00000
What happened?
Using the reset vector and the assembly listed above, I obtained different register output for the single precision instruction on double precision register value comparing with other RISC-V Simulator. For example, the first test cases fclass.s incorrectly identify the value as negative subnormal number rather than quiet NaN. The following two test cases also demonstrate the cases where the instructions did not handle the double precision value correctly by returning NaN value.
Reset vector
fscsr zero
li x1, 0
li x2, 0
li x3, 0
li x4, 0
li x5, 0
li x6, 0
li x7, 0
li x8, 0
li x9, 0
li x10, 0
li x11, 0
li x12, 0
li x13, 0
li x14, 0
li x15, 0
li x16, 0
li x17, 0
li x18, 0
li x19, 0
li x20, 0
li x21, 0
li x22, 0
li x23, 0
li x24, 0
li x25, 0
li x26, 0
li x27, 0
li x28, 0
li x29, 0
li x30, 0
li x31, 0
fcvt.d.w f0, x0
fcvt.d.w f1, x0
fcvt.d.w f2, x0
fcvt.d.w f3, x0
fcvt.d.w f4, x0
fcvt.d.w f5, x0
fcvt.d.w f6, x0
fcvt.d.w f7, x0
fcvt.d.w f8, x0
fcvt.d.w f9, x0
fcvt.d.w f10, x0
fcvt.d.w f11, x0
fcvt.d.w f12, x0
fcvt.d.w f13, x0
fcvt.d.w f14, x0
fcvt.d.w f15, x0
fcvt.d.w f16, x0
fcvt.d.w f17, x0
fcvt.d.w f18, x0
fcvt.d.w f19, x0
fcvt.d.w f20, x0
fcvt.d.w f21, x0
fcvt.d.w f22, x0
fcvt.d.w f23, x0
fcvt.d.w f24, x0
fcvt.d.w f25, x0
fcvt.d.w f26, x0
fcvt.d.w f27, x0
fcvt.d.w f28, x0
fcvt.d.w f29, x0
fcvt.d.w f30, x0
fcvt.d.w f31, x0