Re: [Xen-devel] [PATCH v8 28/50] x86emul: support AVX512F floating point manipulation insns

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 29.05.19 at 14:51, <andrew.cooper3@xxxxxxxxxx> wrote:
> On 15/03/2019 10:56, Jan Beulich wrote:
>> @@ -9681,6 +9696,21 @@ x86_emulate(
>>          op_bytes = 4;
>>          goto simd_imm8_zmm;
>>  
>> +    case X86EMUL_OPC_EVEX_66(0x0f3a, 0x26): /* vgetmantp{s,d} 
>> $imm8,[xyz]mm/mem,[xyz]mm{k} */
>> +    case X86EMUL_OPC_EVEX_66(0x0f3a, 0x54): /* vfixupimmp{s,d} 
>> $imm8,[xyz]mm/mem,[xyz]mm,[xyz]mm{k} */
>> +        host_and_vcpu_must_have(avx512f);
>> +        if ( ea.type != OP_REG || !evex.brs )
>> +            avx512_vlen_check(false);
>> +        goto simd_imm8_zmm;
>> +
>> +    case X86EMUL_OPC_EVEX_66(0x0f3a, 0x27): /* vgetmants{s,d} 
>> $imm8,xmm/mem,xmm,xmm{k} */
>> +    case X86EMUL_OPC_EVEX_66(0x0f3a, 0x55): /* vfixupimms{s,d} 
>> $imm8,xmm/mem,xmm,xmm{k} */
>> +        host_and_vcpu_must_have(avx512f);
>> +        generate_exception_if(ea.type != OP_REG && evex.brs, EXC_UD);
> 
> Why the ea.type != OP_REG restriction?  These four instructions do take
> memory operands.

Did you perhaps read the && as || ? Scalar operations (not just the
ones here) don't support broadcast (with a memory operand), but
may support embedded rounding (or, like here, just SAE; with a
register operand). The exact same construct exists e.g. at the
simd_zmm_scalar_sae label (and the block of case labels ahead of
it actually also gets added to by the patch here).

Jan



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel