Re: [PATCH 03/11] softfloat: Introduce float_flag_inorm_denormal

[Richard Henderson]

On 6/7/21 10:19 AM, Alex Bennée wrote:
> > If you've got a better ordering of operations for this, do tell.
>
> What I really want is to know which instructions translate into the if
> (s->flush_inputs_to_zero) and verifying that is only checked once. Maybe
> I'm just suspicious of compilers ability to optimise things away...




>   Dump of assembler code for function float32_mul:
>      0x0000000000895d60 <+0>:     movzbl 0x1(%rdx),%eax
>      0x0000000000895d64 <+4>:     test   $0x10,%al
>      0x0000000000895d66 <+6>:     je     0x895e30 <float32_mul+208>

s->float_exception_flags & float_flag_inexact

>      0x0000000000895d6c <+12>:    cmpb   $0x0,(%rdx)
>      0x0000000000895d6f <+15>:    jne    0x895e30 <float32_mul+208>

s->float_rounding_mode == float_round_nearest_even

>      0x0000000000895d75 <+21>:    test   $0x7f800000,%edi
>      0x0000000000895d7b <+27>:    jne    0x895da0 <float32_mul+64>
>      0x0000000000895d7d <+29>:    test   $0x7fffffff,%edi
>      0x0000000000895d83 <+35>:    je     0x895da0 <float32_mul+64>

float32_is_denormal

>      0x0000000000895d85 <+37>:    cmpb   $0x0,0x5(%rdx)
>      0x0000000000895d89 <+41>:    je     0x895e60 <float32_mul+256>

s->flush_inputs_to_zero

>      0x0000000000895d8f <+47>:    or     $0x40,%eax
>      0x0000000000895d92 <+50>:    and    $0x80000000,%edi
>      0x0000000000895d98 <+56>:    mov    %al,0x1(%rdx)

flush-to-zero and set iflush_denormal

>      0x0000000000895da0 <+64>:    test   $0x7f800000,%esi
>      0x0000000000895da6 <+70>:    jne    0x895dd0 <float32_mul+112>
>      0x0000000000895da8 <+72>:    test   $0x7fffffff,%esi
>      0x0000000000895dae <+78>:    je     0x895dd0 <float32_mul+112>

float32_is_denormal (second operand)

>      0x0000000000895db0 <+80>:    cmpb   $0x0,0x5(%rdx)
>      0x0000000000895db4 <+84>:    movzbl 0x1(%rdx),%eax
>      0x0000000000895db8 <+88>:    je     0x895e50 <float32_mul+240>
>      0x0000000000895dbe <+94>:    or     $0x40,%eax
>      0x0000000000895dc1 <+97>:    and    $0x80000000,%esi

s->flush_inputs_to_zero,
flush-to-zero,
set iflush_denormal.

...

>      0x0000000000895e50 <+240>:   or     $0x20,%eax
>      0x0000000000895e53 <+243>:   mov    %al,0x1(%rdx)
>      0x0000000000895e56 <+246>:   jmpq   0x895dd0 <float32_mul+112>

set inorm_denormal (second operand)

>      0x0000000000895e60 <+256>:   or     $0x20,%eax
>      0x0000000000895e63 <+259>:   mov    %al,0x1(%rdx)
>      0x0000000000895e66 <+262>:   jmpq   0x895da0 <float32_mul+64>

set inorm_denormal (first operand)

There do seem to be 3 reads/writes to exception_flags for float_raise.


r~