[PATCH, alpha]: Add earlyclobber to sqrtt/sqrtf insns.

[Uros Bizjak]

On Fri, Apr 14, 2017 at 3:55 PM, Richard Henderson <rth@twiddle.net> wrote:
> On 04/14/2017 05:30 AM, Uros Bizjak wrote:
>>
>> Add earlyclobber to sqrtt/sqrtf insns.
>>
>> When using software completions, we have to prevent assembler to match
>> input and output operands of sqrtt/sqrtf insn. Add earlyclobber to
>> output operand to avoid unwanted operand matching.
>>
>> 2017-04-14  Uros Bizjak  <ubizjak@gmail.com>
>>
>>     * sysdeps/alpha/fpu/math_private.h (__ieee754_sqrt): Add
>>     earlyclobber to output operand of sqrt insn.
>>     (__ieee754_sqrtf): Ditto.
>>
>> diff --git a/sysdeps/alpha/fpu/math_private.h
>> b/sysdeps/alpha/fpu/math_private.h
>> index 9e06e25..1e97c86 100644
>> --- a/sysdeps/alpha/fpu/math_private.h
>> +++ b/sysdeps/alpha/fpu/math_private.h
>> @@ -27,9 +27,9 @@ __ieee754_sqrt (double d)
>>  {
>>    double ret;
>>  # ifdef _IEEE_FP_INEXACT
>> -  asm ("sqrtt/suid %1,%0" : "=f"(ret) : "f"(d));
>> +  asm ("sqrtt/suid %1,%0" : "=&f"(ret) : "f"(d));
>>  # else
>> -  asm ("sqrtt/sud %1,%0" : "=f"(ret) : "f"(d));
>> +  asm ("sqrtt/sud %1,%0" : "=&f"(ret) : "f"(d));
>>  # endif
>
>
> Hmm.  This is surprising because any host that has sqrtt also has exact
> traps, and so trap shadows and recovery of the input shouldn't be an issue.

I have tried with a sqrtt/sud $f0, $f0, where the value in $f0 was
0x0....10. I can confirm that the returned result was 0.

> That said, do we actually still support a gcc that doesn't have these as
> builtins?  That would be more ideal than inline asm.

Compiling

double test_sqrt (double d)
{
        return __builtin_sqrt (d);
}

with -O2 -mcpu=ev6 resulted in:

test_sqrt:
        .frame $30,16,$26,0
        .mask 0x4000000,-16
$LFB0:
        .cfi_startproc
        ldah $29,0($27)         !gpdisp!1
        lda $29,0($29)          !gpdisp!1
$test_sqrt..ng:
        sqrtt $f16,$f0
        lda $30,-16($30)
        .cfi_def_cfa_offset 16
        stq $26,0($30)
        .cfi_offset 26, -16
        .prologue 1
        cmpteq $f0,$f0,$f10
        fbeq $f10,$L4
$L2:
        ldq $26,0($30)
        lda $30,16($30)
        .cfi_remember_state
        .cfi_restore 26
        .cfi_def_cfa_offset 0
        ret $31,($26),1
$L4:
        .cfi_restore_state
        ldq $27,sqrt($29)               !literal!2
        jsr $26,($27),sqrt              !lituse_jsr!2
        ldah $29,0($26)         !gpdisp!3
        lda $29,0($29)          !gpdisp!3
        br $31,$L2

__builtin_sqrt expands with a call to sqrt(), this could possibly
create a recursive call loop if the builtin is used in libc to
implement sqrt.

Uros.