On 7/22/20 2:15 AM, frank.chang@sifive.com wrote:
> FIELD(VTYPE, VLMUL, 0, 2)
> FIELD(VTYPE, VSEW, 2, 3)
> -FIELD(VTYPE, VEDIV, 5, 2)
> -FIELD(VTYPE, RESERVED, 7, sizeof(target_ulong) * 8 - 9)
> +FIELD(VTYPE, VFLMUL, 5, 1)
> +FIELD(VTYPE, VEDIV, 8, 9)
> +FIELD(VTYPE, RESERVED, 10, sizeof(target_ulong) * 8 - 11)
> FIELD(VTYPE, VILL, sizeof(target_ulong) * 8 - 1, 1)
The ediv definition is wrong -- should be 8, 2.
OK, I will correct it.
> @@ -37,4 +38,10 @@ target_ulong fclass_d(uint64_t frs1);
> #define SEW32 2
> #define SEW64 3
>
> +/* table to convert fractional LMUL value */
> +static const float flmul_table[8] = {
> + 1, 2, 4, 8, /* LMUL */
> + -1, /* reserved */
> + 0.125, 0.25, 0.5 /* fractional LMUL */
> +};
> #endif
Don't define data in a header file; only declare it.
Fractional LMUL are used in cpu.h, translate.c and vector_helper.c.
I was trying to declare something which can be shared among these files to calculate the fractional LMUL value.
Perhaps it's better to declare it as the inline function which calculates fractional LMUL value in internals.h?
Or I can do the calculation explicitly at every place which requires the fractional LMUL value?
(only 4 places require this value by far.)
> @@ -60,6 +60,9 @@ typedef struct DisasContext {
> /* vector extension */
> bool vill;
> uint8_t lmul;
> + float flmul;
> + uint8_t eew;
> + float emul;
Why are you adding floating-point values to DisasContext?
flmul, eew and emul are required during rvv-0.9 vector load/store instructions.
Should I move these declarations to the vector load/store instructions patch to make it clearer?
> +static inline float vext_vflmul(uint32_t desc)
> +{
> + uint32_t lmul = FIELD_EX32(simd_data(desc), VDATA, LMUL);
> + return flmul_table[lmul];
> }
And in the helpers? Are you planning on some sort of path through int -> float
-> int for computation? That seems questionable.
desc only saves the raw LMUL bits.
(total 3 bits, I've packed the fractional LMUL bit together with two other LMUL bits in cpu_get_tb_cpu_state())
The helper here is to convert the 3-bits LMUL into the actual fractional number it represents.
r~
Frank Chang