[PATCH] x86/APX: convert runtime special case to build-time one

[Jan Beulich]

On 17.01.2025 03:51, Jiang, Haochen wrote:
>> From: Jan Beulich <jbeulich@suse.com>
>> Sent: Thursday, January 16, 2025 6:26 PM
>>
>> --- a/opcodes/i386-gen.c
>> +++ b/opcodes/i386-gen.c
>> @@ -1039,6 +1039,20 @@ process_i386_cpu_flag (FILE *table, char
>>  	  all[Cpu64].value = 1;
>>
>>        output_cpu_flags(table, all, ARRAY_SIZE (all), -1, comma, indent, lineno);
>> +
>> +      /* For APX_F extension of multiple cpuid enabled insns, we cannot use
>> +	 APX_F(cpuid_A&cpuid_B) in the opcode table, as the result would fail
>> +	 to be parsed.  Furthermore, the result also wouldn't be quite valid.
>> +	 However, the assembler's cpu_flags_match() will simply propagate
>> "any"
>> +	 to "all", zapping "any" afterwards altogether.  IOW in this situation
>> +	 both masks have "&&" meaning.  Set the missing flag here.  */
>> +      if (all[CpuAMX_TRANSPOSE].value && all[CpuAMX_MOVRS].value)
>> +	{
>> +	  if (!any[CpuAPX_F].value || !any[CpuAMX_MOVRS].value)
>> +	    fail ("%s: %d: internal error: APX_F=%d AMX_MOVRS=%d\n",
>> +		  filename, lineno, any[CpuAPX_F].value,
>> any[CpuAMX_MOVRS].value);
> 
> I am not sure if this error is too strict and will cause issue in the future. But
> LGTM for now since I don't see problem in the foreseeable future.

It's intentionally as strict as it can be right now. If and when there are
problems with it, it can and should be refined / relaxed. But we want to
be sure to at least be aware of the situation, to properly evaluate what
our options are (without waiting for a bug report by a user).

Jan