Re: [RFC PATCH v2 2/2] RISC-V: Add support for the Zfa extension

[Jan Beulich <jbeulich@suse.com>]

On 30.03.2023 12:54, Jan Beulich via Binutils wrote:
> On 30.03.2023 12:30, Christoph Müllner wrote:
>> On Mon, Mar 27, 2023 at 11:54 AM Jan Beulich <jbeulich@suse.com> wrote:
>>>
>>> On 27.03.2023 10:53, Kito Cheng wrote:
>>>> Wait, I mean the hex floating point format defined in C99/C++17, not
>>>> the raw hex value.
>>>> so something like 0x1p-16 (0.0000152587890625), 0x1p-2 (0.25) 0x1p+0,
>>>> -0x1p+0 could be used for fli.* instruction.
>>>>
>>>> You could use printf with %a to get those values.
>>>>
>>>> https://gcc.gnu.org/onlinedocs/gcc/Hex-Floats.html
>>>> https://developer.arm.com/documentation/dui0375/latest/Compiler-Coding-Practices/Hexadecimal-floating-point-numbers-in-C99
>>>
>>> Sure, my (secondary) suggestion ...
>>>
>>>> On Mon, Mar 27, 2023 at 4:39 PM Jan Beulich via Binutils
>>>> <binutils@sourceware.org> wrote:
>>>>>
>>>>> On 27.03.2023 10:01, Christoph Muellner wrote:
>>>>>> --- a/opcodes/riscv-opc.c
>>>>>> +++ b/opcodes/riscv-opc.c
>>>>>> @@ -110,6 +110,16 @@ const char * const riscv_vma[2] =
>>>>>>    "mu", "ma"
>>>>>>  };
>>>>>>
>>>>>> +/* The FLI.[HSDQ] value constants.  */
>>>>>> +const char * const riscv_fli_value[32] =
>>>>>> +{
>>>>>> +  "-1.0", "min", "0.0000152587890625", "0.000030517578125",
>>>>>> +  "0.00390625", "0.0078125", "0.0625", "0.125",
>>>>>> +  "0.25", "0.3125", "0.375", "0.4375", "0.5", "0.625", "0.75", "0.875",
>>>>>> +  "1.0", "1.25", "1.5", "1.75", "2.0", "2.5", "3.0", "4.0",
>>>>>> +  "8.0", "16.0", "128.0", "256.0", "32768.0", "65536.0",  "inf", "nan",
>>>>>> +};
>>>>>
>>>>> Especially for values like 1.0x2^^-n (entries 2 and onwards) I question
>>>>> the spelled out numbers to be the most suitable ones usability wise. At
>>>>> least some alternative spelling (e.g. 2.e-16) ought to be recognized as
>>>>> well. But since there are meany reasonable spellings (leading 0 omitted
>>>>> in 0.<fraction> or trailing zero omitted in <num>.0), I guess I'd prefer
>>>>> if values were actually parsed as a floating point number (e.g. via
>>>>> ieee_md_atof()), and then matched against values stored in the table.
>>>>> One might further consider to also permit the 2nd form accepted
>>>>> elsewhere, see read.c:parse_one_float().
>>>
>>> ... here wasn't meant to collide with yours. What you're asking for is
>>> covered by my primary suggestion (to actually parse the values), extended
>>> by the need to actually recognize C99 hex float in the parser then (leaving
>>> aside for now whether that's feasible in the first place).
>>
>> Thanks for all the suggestions!
>>
>> I worked my way through this and I believe that the following would be
>> a reasonable solution:
>> * constants min, inf and nan must be symbols (as stated in the specification)
>> * other constants are parsed by scanf("%f") as floats and compared
>> (float compare in C) against the numeric constants in the table
>> * output in the disassembly uses symbols for min/inf/nan and %a (hex
>> FP literals) for other constants
>>
>> So we support every format that '%f' accepts including hex FP literals
>> (e.g. -0x1p0, 0x1p+0, ...) and normal FP constants (e.g.
>> 0.0000152587890625, 25E-4).
> 
> How's this going to work with a cross-assembler run on an architecture
> supporting a floating point format other than IEEE 754's? Hence why I
> suggested using ieee_md_atof() instead.

Hmm, I see riscv_fli_numval[] has "float" as the base type (which I
didn't really expect), so this ought to work as long as the initializers
used can be represented exactly in whatever arch's floating point format.

Jan